FEP_Lean v1.1.0 is a source-bound, machine-checked catalogue of 155 topics across 20 reviewed families and five areas: the Free Energy Principle, Active Inference, Bayesian Mechanics, Information Geometry, and non-equilibrium Thermodynamics. Every catalogue row carries a reviewed invariant, explicit assumptions and boundaries, a namespaced Lean 4 theorem body, and deterministic manuscript metadata.
The release is pinned to Lean 4.33.1 and Mathlib 4.33.1 (locked Mathlib revision 0df444a360eaa60ab8c11dca51a86af692955474). Its native receipt verifies 155/155 topic closures with zero errors, warnings, or sorry. The schema-4 formalism audit covers all 823 required formal-resource declarations, including 699 evidence declarations, and reports no sorryAx or untrusted project axioms. The canonical Python acceptance run collected 1,203 tests: 1,080 passed, 123 skipped, zero failed or errored, with 89.81% line coverage (8,808/9,807 statements). The schema-4 Chrome 151 browser receipt replays six source-bound screenshots covering 155 topics, 20 families, 133 relations, 48 satisfied capabilities, and 15 typed numerical witnesses.
These checks establish the exact shipped formal statements and reproducible software evidence; they do not prove the Free Energy Principle as a physical theory. Provider-backed Hermes commentary is optional and unavailable for this release cut, and no historical provider report is promoted to current full-mode evidence.
Source and release: github.com/ActiveInferenceInstitute/fep_lean and GitHub release v1.1.0. The exact deterministic 232-member evidence bundle has SHA-256 0009447598ecd3bbf68548eab360704a2539480379791eb0128186fa230884ea and binds release commit d7b6f8b15ea9dc451b191e3674a1fd72b5e586b4.
Introduction to Biology: A Generative Approach is an open biology textbook with forty-four chapters, ranging from systems science and chemical foundations through cells, metabolism, genetics, microbiology, physiology, evolution, and ecology. Organized as Unit 0 plus Units I–X, the text presents biology as an evidence-grounded discipline in which mechanisms, measurements, and simple models are developed together, so readers can move between narrative explanation and the quantitative constraints that shape biological claims. Five recurring themes—evolution, information, structure and function, systems and emergence, and the cell—provide orientation across scales and align with mainstream undergraduate biology competencies; Unit 0 adds an optional systems, historical, and philosophical lens without replacing the traditional molecular-to-ecological sequence.
\nWhere the curriculum is quantitative, corresponding computations live in tested code modules organized by domain (biochemistry, cell biology, genetics, physiology, ecology, evolution, microbiology, botany, and neuroscience), and many figures and process diagrams are generated programmatically rather than supplied as static artwork alone. The edition pairs each chapter with a paper-based laboratory activity and a question bank that progresses from recall to synthesis; model answers are visible in this instructor build. Primary literature is cited inline, glossary and curriculum-mapping appendices support course design, and the manuscript is maintained as a reproducible open-science artifact (source at https://github.com/docxology/biology_textbook ; archived at DOI 10.5281/zenodo.20286478). Text is released under Creative Commons Attribution 4.0; accompanying source code under Apache-2.0.
", "keywords": [ "Biology" ], "findings": [ "The book covers introductory biology in a systems unit plus Units I-X with 44 core chapters, plus optional laboratories and question banks.", "Each chapter is followed by a companion lab and a 30-item question bank in the same canonical order.", "The author positions active inference and the free energy principle as optional graduate-depth lenses, not part of the introductory canon.", "The text includes a master glossary of 225 terms with etymology and chapter cross-references, and is released under CC BY 4.0.", "The stated pedagogical philosophy is to understand biology by computing biology." ], "methods": [ "Generation from version-controlled Markdown, tested Python, and Mermaid diagrams — The open textbook is built from Markdown manuscript sources, tested Python modules, programmatically generated figures and rendered Mermaid diagrams.", "Python modules for models such as Michaelis-Menten and Hodgkin-Huxley — Each quantitative model introduced in a chapter exists as a working module in the codebase, used to generate figures.", "Manifest-driven organisation from manuscript/config.yaml — Navigation, scope tables and appendix ordering are generated from a single config.yaml manifest so they stay aligned with the rendered table of contents.", "Five Big Ideas mapped to AAAS Vision and Change core concepts — The book organises recurring themes as Five Big Ideas aligned pedagogically to the Vision and Change report's five core concepts." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_BiologyTextbook/metadata.json", "findings": "papers/2026_BiologyTextbook/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "book/report" } }, "Friedman2026PolicyEntanglementActiveInference119": { "citation_key": "Friedman2026PolicyEntanglementActiveInference119", "abstract": "Active inference models often need to choose among several policy streams at once, for example streams tied to different effectors, sensory channels, agents, agents within a group, or planning horizons. Standard discrete active-inference implementations keep this manageable by treating those streams as independent, but that simplification removes the dependencies that make coordinated action possible. This manuscript introduces policy entanglement: a controlled deformation of the usual independent policy posterior by a scalar coupling strength and explicit compatibility and preference potentials. The construction preserves the finite active inference setting while making cross-stream dependence a first-class modeling object rather than an implicit artifact of the chosen factorization. The framework keeps a claim-strength ledger that distinguishes exact recoveries, parameterized embeddings, numerical witnesses, and structural analogies. Mean-field active inference is the exact independent case. Products of experts, copula variational inference, options, hierarchical and sophisticated inference, branching-time active inference, renormalization-style compression, and Markov-blanket multi-agent views are connected as special cases through their stated posterior-factorization maps. The central result is a free-energy decomposition that separates ordinary per-stream free energy, coupling preference terms, the coupling normalizer, and the information cost of leaving independence. The decomposition makes multi-information the explicit surcharge paid by a non-factorized policy posterior and shows how coupling strength, compatibility structure, and off-diagonal preference costs enter the same accounting identity. This result supplies the organizing principle for the rest of the paper. It supports an information-geometric reading of the coupled policy family as a path away from the mean-field submanifold, a projection identity that returns the coupled posterior to its independent marginals, a spectral and tensor-train view of dominant coordinated policy modes, a heterogeneous-ensemble coupling-tax bound, and a phase vocabulary for under-coupled, mixed, and highly concentrated policy posteriors. These interpretations are intentionally limited: the manuscript does not claim a neural, clinical, biological, or quantum implementation, and Markov-blanket and tensor-network language is used as scoped modeling analogy unless a specific theorem row or generated artifact supports a stronger statement. The main decomposition analytic identity is machine-checked in ℝ in the Mathlib-backed Lean layer with an axiom audit and negative controls. A separate stock-Lean boundary fragment remains Mathlib-free and exposes the theorem surface as typed contracts for the Python simulation layer and the manuscript registry, including witness-consuming rows where analytic payloads are deliberately supplied at the boundary. The executable numerical layer remains a Float pipeline, so a verified Float↔ℝ error bridge is still an explicitly open interface rather than an implied proof; conservative interval brackets on the K=2 decomposition sweep certify Float residuals within a widened high-precision envelope (output/reports/float_real_residual.json) without promoting the registry row to proved. The empirical companion uses pymdp and NumPy to sweep coupled policy ensembles, run short and long rollouts, check the projection identity to round-off precision, produce free-energy, entropy, total-correlation, action-distribution, robustness, and adversarial sidecars, and render figures from those artifacts. The manuscript, figures, theorem map, citation registry, notation glossary (§S6), bibliography, and PDF are regenerated from the same source-owned pipeline, so prose claims are tied to Lean sources, Python witnesses, output metadata, and validation gates rather than maintained by hand. All manuscript methods, tests, and documentation are available as open-source software at https://github.com/ActiveInferenceInstitute/policy_entanglement (DOI: https://doi.org/10.5281/zenodo.20419536). --- Associated artifacts GitHub release: v1.0.0 (https://github.com/ActiveInferenceInstitute/policy_entanglement/releases/tag/v1.0.0) DOI: https://doi.org/10.5281/zenodo.20418904 Zenodo: https://zenodo.org/records/20418904 PDF SHA-256: ae7cdd62929324101ead3eba8177199141b0089a9baf35558107149331666fde
", "keywords": [ "active inference", "free energy principle", "policy inference", "mean-field", "total correlation", "information geometry", "Schmidt rank", "tensor networks", "sophisticated inference", "Lean theorem proving", "machine-checked free-energy identity" ], "findings": [ "The central result is a free-energy decomposition into per-stream free energy, coupling preference terms, the coupling normalizer, and the information cost of leaving independence.", "The decomposition makes multi-information the explicit surcharge paid by a non-factorized policy posterior.", "Mean-field active inference is recovered as the exact independent case, with other frameworks linked through stated posterior-factorization maps.", "A verified Float-to-real error bridge for the numerical layer remains an explicitly open interface rather than an implied proof.", "The author states the manuscript does not claim a neural, clinical, biological, or quantum implementation; Markov-blanket and tensor-network language is a scoped analogy." ], "methods": [ "Coupling-parameter deformation of the independent policy posterior — Multi-stream policy posteriors are deformed away from the mean-field product by a scalar coupling strength plus compatibility and preference potentials.", "Lean 4 formalization: Mathlib proof of the central identity plus stock-Lean boundary — MathlibProofs machine-checks the S01 free-energy identity with an axiom audit and negative controls; a stock-Lean fragment exposes a 21-row theorem surface as typed contracts.", "pymdp/NumPy POMDP simulations of coupled policy ensembles — Simulations sweep coupled ensembles, run short and long rollouts, check the projection identity, and produce free-energy, entropy, total-correlation, robustness, and adversarial sidecars.", "Interval brackets on Float residuals for the K=2 decomposition sweep — Conservative interval brackets check that Float-pipeline residuals fall within a widened high-precision envelope, without counting this as a proof.", "Claim-strength ledger separating exact, parametric, numerical, and analogical claims — Connections to prior frameworks are tagged as exact recoveries, parameterized embeddings, numerical witnesses, or structural analogies." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_PolicyEntanglementActive/metadata.json", "findings": "papers/2026_PolicyEntanglementActive/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "theoretical" } }, "Friedman2026BoundedAutoResearchTinyReproducible120": { "citation_key": "Friedman2026BoundedAutoResearchTinyReproducible120", "abstract": "This paper presents Deterministic bounded AutoResearch for a small MNIST neural-network task, a public template exemplar that\nturns an AutoResearch loop into ordinary reproducible research infrastructure.\nThe case study is intentionally small but concrete: 2000 training\nand 500 test images from MNIST handwritten digit database are evaluated by the\nbounded small MNIST neural-network classification loop. The run evaluates\n4 of 5 proposed candidates,\nincluding Tiny patch-attention classifier, selects\nexp-mlp-tanh-64 (MLP,\n50890 parameters), and improves test_accuracy from\n82.6% to 89.4%\n(6.8% absolute change). The validated diagnostic layer reports\nmacro F1 89.4%, bootstrap accuracy interval\n86.4% to 92.0%, Brier score 0.161,\nnegative log likelihood 0.361, top-2 accuracy\n95.6%, and exact McNemar p-value 0.000.\nThe same pipeline writes proposal, candidate, run, review, benchmark, evidence,\nfigure, confusion-matrix, statistical-summary, probability-quality, and\nsecurity-integrity artifacts from declared output contracts; uses\n0 LLM calls at USD 0.00 cost; and records\n7 configured stages, 6 supported\nlocal-artifact claims, and 78 required artifacts.\nThe local security attestation status is passed,\nwith 0 checksum mismatch(es). The final\nreadiness status is passed, with review gates deferred to a\nhuman rather than self-approved by the generated run.\n\n---\nAssociated artifacts\nGitHub release: v0.3.2 (https://github.com/docxology/template_autoresearch_project/releases/tag/v0.3.2)\nDOI: https://doi.org/10.5281/zenodo.20417016\nZenodo: https://zenodo.org/records/20417016\nPDF SHA-256: e07b62850a1995935283d37a45c21d71fa7c4e69cdcc451c5a1ea8aee6d0c94a", "keywords": [ "autoresearch", "reproducible research", "machine learning benchmark", "artifact readiness", "human review", "local artifact integrity" ], "findings": [ "The loop selected exp-mlp-tanh-64 after evaluating 4 of 5 proposed candidates, raising test accuracy from the 82.6% baseline to 89.4%.", "Diagnostics report macro F1 of 89.4%, a bootstrap accuracy interval of 86.4% to 92.0%, and top-2 accuracy of 95.6%.", "The selected candidate was top-ranked in 72.5% of deterministic bootstrap resamples, with exp-mlp-relu-32 as runner-up.", "The local security attestation passed with 0 checksum mismatches, and readiness passed with review gates deferred to a human.", "The paper states its contribution is not a new MNIST classifier but a template showing bounded AutoResearch run through a reproducible-paper lifecycle." ], "methods": [ "Offline MNIST subset: 2000 train / 500 test images, seed 20260525 — Uses a committed, class-balanced local MNIST subset with recorded provenance hashes; no data are downloaded at runtime.", "Bounded candidate search over MLP, softmax, nearest-centroid, patch-attention — Evaluates at most 4 configured candidates against a nearest-centroid baseline, choosing by test accuracy with deterministic tie-breaks.", "Seven-stage AutoResearch pipeline with file-backed ledgers — Runs 7 configured stages, writing proposal, candidate, run, phase and review ledgers that hydrate manuscript variables.", "Safety controls: proposal-only autonomy, no LLM calls, deferred review — Defaults to proposal_only autonomy, records 0 LLM calls and no cost, never executes generated code, and leaves publication approval to a human.", "Statistical diagnostics: Wilson intervals, bootstrap, McNemar, calibration — Reports Wilson score intervals, deterministic bootstrap intervals, paired discordance tests, Brier score and negative log likelihood." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_BoundedAutoResearchTiny/metadata.json", "findings": "papers/2026_BoundedAutoResearchTiny/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026ConvergenceAnalysisGradientDescent121": { "citation_key": "Friedman2026ConvergenceAnalysisGradientDescent121", "abstract": "This paper presents a convergence study of fixed-step gradient descent on a convex quadratic, framed as the computational exemplar of the Research Project Template (https://github.com/docxology/template). The implementation lives in projects/templates/template_code_project/src/optimizer.py; experiments and figures are orchestrated by projects/templates/template_code_project/scripts/optimization_analysis.py and hydrated into the manuscript through scripts/z_generate_manuscript_variables.py, so tables and prose track output/data/optimization_results.csv after every pipeline run.\n\nWe evaluate 6 step sizes from $\\alpha = 0.01$ to $\\alpha = 2.5$, spanning conservative, near-optimal, aggressive, and divergent regimes for a unit Hessian model. The build chain exercises template infrastructure end-to-end: scientific helpers (infrastructure.scientific.stability, infrastructure.scientific.benchmarking), validation, rendering (infrastructure/rendering/pdf_renderer.py), and reporting. Accessibility-oriented plotting defaults (colourblind-safe palette, 300 dpi exports) are centralized in src/figures/ and src/analysis/.\n\nContributions are methodological and architectural. On the methods side, we relate empirical iteration counts and error decay to the scalar contraction factor $\\rho(\\alpha) = |1-\\alpha|$ and document cases where runs hit $N_{\\max} = 1000$ before meeting the gradient tolerance. On the architecture side, we demonstrate a zero-mock test suite on project src/ (see test_optimizer.py (https://github.com/docxology/template/blob/main/projects/templates/template_code_project/tests/test_optimizer.py)), automated six-figure analysis, and reproducibility metadata (configuration hash, artifact counts) injected into .\n\nResults (this configuration): 4 of 6 grid points report converged=True in the CSV; non-convergent rows flag either slow progress at small $\\alpha$ under the iteration cap or instability when $|1-\\alpha| \\geq 1$. The analytical minimizer remains $x^\\ast = 1.0$ with $f(x^\\ast) = -0.5$ for the configured $(A,b)$.\n\nKeywords: optimization algorithms, gradient descent, convergence analysis, numerical methods, mathematical programming, reproducible research, infrastructure automation\n\n---\nAssociated artifacts\nGitHub release: v2.5.2 (https://github.com/docxology/template_code_project/releases/tag/v2.5.2)\nDOI: https://doi.org/10.5281/zenodo.20417136\nZenodo: https://zenodo.org/records/20417136\nPDF SHA-256: cd54b95893501467503fab2c4b432573306bc94f7040085550beb87d094b4e50", "keywords": [ "optimization algorithms", "gradient descent", "convergence analysis", "numerical methods", "mathematical programming", "reproducible research", "infrastructure automation" ], "findings": [ "Four of the six grid step sizes converged; the non-converged runs either hit the iteration cap at small α or were unstable when |1−α| ≥ 1.", "α=1.0 reached the optimum in one iteration for this unit-Hessian problem, the fastest configuration.", "The paper reports a stability boundary at α=2: α<2 converges and α≥2 diverges for the unit-Hessian problem.", "In the dimensional benchmark, iterations to convergence rose only modestly (219 to 238) across two decades of dimension.", "The author states the scientific claims are textbook material and the non-standard contribution is procedural (config-driven figures, CSV and manuscript numbers)." ], "methods": [ "Fixed-step gradient descent on a 1-D convex quadratic (A=1, b=1) — Runs fixed-step gradient descent on f(x)=½xᵀAx−bᵀx with A=[1], b=[1], analytic optimum x*=1, f(x*)=−0.5.", "Six-point step-size grid from α=0.01 to α=2.5 — Sweeps six fixed step sizes spanning conservative, near-optimal, aggressive and divergent regimes, with a gradient-norm tolerance and iteration cap.", "Comparison to scalar contraction factor ρ(α)=|1−α| — Relates empirical iteration counts and error decay to the contraction factor of the linear error recurrence for the unit-Hessian case.", "Stability grid (8 starts × 6 step sizes) and dimensional scaling benchmark — Evaluates accuracy over 48 start/step-size combinations and separately times gradient_descent() on identity-Hessian quadratics of increasing dimension.", "Zero-mock test suite with ≥90% coverage gate and variable-injected manuscript — Tests src/ without mocks under a CI coverage gate; results are injected into the manuscript from the analysis CSV via placeholders." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_ConvergenceAnalysisGradient/metadata.json", "findings": "papers/2026_ConvergenceAnalysisGradient/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026EditorialQualityAtScale122": { "citation_key": "Friedman2026EditorialQualityAtScale122", "abstract": "This paper documents template_prose_project, the prose-focused exemplar of the Research Project Template (https://github.com/docxology/template). It pairs the template's two-layer architecture with the prose analysis infrastructure (https://github.com/docxology/template/tree/main/infrastructure/prose) (readability metrics, structural outline, editorial quality flags) and the reference validation infrastructure (https://github.com/docxology/template/tree/main/infrastructure/reference) (BibTeX validation), demonstrating that rigorous editorial review can be expressed as a configurable, deterministic pipeline with no novel domain algorithm of its own.\n\nA single manuscript/config.yaml defines target grade-level bands, citation-density floors, structural rules (every section has an H1, no heading levels skipped), and bibliography-consistency policy. The pipeline reads the manuscript, runs the prose analysers, cross-checks every citation against manuscript/references.bib, evaluates the configured checks, and writes a deterministic markdown review report alongside three figures (per-file word counts, readability metrics, citation density) and a JSON manuscript_report.json suitable for CI artefacts.\n\nRun snapshot. The current configuration analyses 8 file(s) totalling 1742 words across 86 sentence(s) and 64 paragraph(s). Average Flesch-Kincaid grade level is 15.93; average Gunning Fog index is 16.69; the manuscript references 6 unique citation key(s); the longest section is 413 words and the shortest is 17. These numbers are auto-substituted by scripts/z_generate_manuscript_variables.py after every run, so the abstract tracks the JSON outputs in output/.\n\nThe contribution is methodological and architectural: a generic, reusable prose-quality module (infrastructure/prose/) that any project in the template can opt into, plus a minimal, configurable exemplar (projects/templates/template_prose_project/) that wires it to the bibliography and the manuscript pipeline.\n\nKeywords: prose analysis, readability, editorial review, reproducible manuscript review, scientific infrastructure\n\n---\nAssociated artifacts\nGitHub release: v0.4.2 (https://github.com/docxology/template_prose_project/releases/tag/v0.4.2)\nDOI: https://doi.org/10.5281/zenodo.20417104\nZenodo: https://zenodo.org/records/20417104\nPDF SHA-256: 290d21b10bd588b978d6a3200cdf0e3c2441ca86fcdc777ab41975fa910a260e", "keywords": [ "prose analysis", "readability", "editorial review", "reproducible research", "manuscript quality" ], "findings": [ "The paper reports that editorial review can be expressed as a configurable, deterministic pipeline with no novel domain algorithm of its own.", "On the bundled manuscript, the run analysed 8 files totalling 1731 words, with average Flesch-Kincaid grade 15.87 and Gunning Fog 16.67.", "Because no external service is consulted, a second run on the same inputs produces byte-identical JSON (modulo timestamp metadata).", "The stated contribution is architectural: a reusable prose-quality module any template project can opt into, plus a minimal configurable exemplar." ], "methods": [ "Five-stage pure-function pipeline: read, analyse, cross-check, evaluate, render — template_prose_project implements editorial review as five pure-function stages in src/, with scripts limited to argument parsing and I/O.", "Readability metrics: Flesch Reading Ease, Flesch-Kincaid grade, Gunning Fog — Per-file metrics are computed with textbook readability formulae over a vowel-group syllable heuristic.", "Heuristic quality flags: passive voice, hedge density, citation density — A quality analyser flags passive-voice candidates, hedge words, Pandoc [@key] citation density, and long sentences.", "Citation-key cross-check against references.bib with configurable policy — Every cited key is matched against the BibTeX file, with fail_on_missing / fail_on_unused settings in config.yaml.", "Run-twice diff test for byte-identical JSON output — Reproducibility is checked locally by running the pipeline twice and diffing manuscript_report.json." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_EditorialQualityAt/metadata.json", "findings": "papers/2026_EditorialQualityAt/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026TemplateApproachReproducibleGenerative123": { "citation_key": "Friedman2026TemplateApproachReproducibleGenerative123", "abstract": "The reproducibility crisis in computational research is fundamentally structural: research artifacts are scattered across disconnected tools—LaTeX editors, Jupyter notebooks, ad-hoc shell scripts—with no enforced mechanism to keep code, data, and manuscript synchronized. Studies have shown that most published findings are false positives, replication rates in psychology hover around 36%, and only 24% of 1.4 million Jupyter notebooks can be successfully re-executed. Existing tools address fragments of this problem: workflow managers (Snakemake, Nextflow, CWL) orchestrate computation; literate programming systems (Quarto, Jupyter Book, R Markdown, Overleaf, OpenAI Prism) render documents; data versioning tools (DVC) track artifacts—but none enforces cross-cutting quality standards as architectural invariants. template/ applies the principle of Infrastructure as Code to the research lifecycle, making the manuscript, test suite, and provenance chain version-controlled, deterministically buildable, and independently verifiable. It is built on a Two-Layer Architecture that separates 23 infrastructure subdirectories (20 importable Python packages, ~604 modules, validated by ~7,780 tests) from self-contained project workspaces, connected by a YAML-declared pipeline (12 stages; default full 10)-based build pipeline progressing from environment sanitization through test execution (with a Zero-Mock testing policy enforcing 90% project-level and 60% infrastructure-level coverage via real filesystem operations and subprocess invocations), analysis script invocation, Pandoc/XeLaTeX rendering, SHA-256 cryptographic hashing with steganographic watermarking, structural PDF validation, and LLM-assisted review. A Documentation Duality standard equips every directory with both human-readable README.md and machine-readable AGENTS.md files, while each infrastructure module additionally carries a SKILL.md—a structured skill descriptor aligned with the Model Context Protocol—enabling AI agents to locate and invoke module capabilities without hallucinating API signatures.\n\nScalability is demonstrated across the generated public exemplar roster (templates/template_active_inference, templates/template_autoresearch_project, templates/template_autoscientists, templates/template_code_project, templates/template_gold_refinement, templates/template_literature_meta_analysis, templates/template_madlib, templates/template_newspaper, templates/template_prose_project, templates/template_sia, templates/template_template, templates/template_textbook), with representative heterogeneous cases under projects/templates/: optimization (template_code_project, 231 tests), prose (template_prose_project, 120 tests), and AutoResearch readiness (template_autoresearch_project, 296 tests). These guarantee control-positive layouts for code-centric, prose-centric, and retrieval-centric workflows at 90%+ project coverage alongside 60%+ infrastructure gates. All three share identical pipeline stages without cross-project coupling. This manuscript adds a complementary reflexive artifact: authored from projects/templates/template_template (127 tests) as a public exemplar in the same discovered/rendered tree, using the same analysis and render path and injecting counters from repository introspection. The fact that these words, metrics, and figures were generated by the pipeline they describe demonstrates self-documenting capacity: rendered through the DAG, validated without mocks, optionally watermarked. A comparative analysis against nine peer tools across fourteen dimensions positions template/ as integrating fourteen distinctive enforcement capabilities—testing thresholds, cryptographic provenance, steganographic watermarking, multi-project management, MCP-aligned skill descriptors, Zero-Mock policy, orchestration through publication—in one repository. Code is released under the Apache License 2.0 at github.com/docxology/template; the work remains open-ended.\n\n---\nAssociated artifacts\nGitHub release: v1.0.9 (https://github.com/docxology/template_template/releases/tag/v1.0.9)\nDOI: https://doi.org/10.5281/zenodo.20419007\nZenodo: https://zenodo.org/records/20419007\nPDF SHA-256: 535bd80943d0ae9fd504a926efb41c6b39c3a812a94ea4d51bc974029bca563c", "keywords": [ "reproducible research", "infrastructure-as-code", "steganography", "cryptographic provenance", "LaTeX rendering", "modular infrastructure", "publication integrity", "zero-mock testing", "thin orchestrator", "two-layer architecture", "FAIR4RS", "research software engineering" ], "findings": [ "Reports 100% pipeline completion for the sampled multi-project runs, with timings described as illustrative.", "The manuscript was itself produced by the pipeline it describes, with metrics injected from repository introspection.", "The comparative analysis positions template/ as integrating fourteen distinctive enforcement capabilities in one repository.", "States that the provenance layer offers SHA-256 tamper detection but not cryptographic non-repudiation, since it lacks private-key signatures.", "Acknowledges the pipeline is single-machine, without native distributed execution, where Snakemake, Nextflow and CWL are superior." ], "methods": [ "Two-Layer Architecture separating shared infrastructure from project workspaces — Describes a repository design where N independent research projects share infrastructure packages without coupling to each other.", "YAML-declared 12-stage build DAG from tests through Pandoc/XeLaTeX rendering — Specifies the pipeline in pipeline.yaml; default full runs use 10 stages and --core-only runs 8.", "Zero-Mock testing policy with 90% project and 60% infrastructure coverage gates — Tests use real filesystem operations and subprocess calls rather than mocks, with coverage thresholds enforced by the pipeline.", "Feature comparison against nine peer tools across fourteen dimensions — Compares template/ with workflow managers, literate-programming systems, DVC, Overleaf and OpenAI Prism on enforcement features.", "Multi-project pipeline runs measuring coverage, timing, integrity and watermarking — Exemplar projects were run through the core pipeline on an Apple Silicon workstation, recording tests passed, durations and steganography timings." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_TemplateApproachReproducible/metadata.json", "findings": "papers/2026_TemplateApproachReproducible/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026ActiveInferenceMultiTrack124": { "citation_key": "Friedman2026ActiveInferenceMultiTrack124", "abstract": "We study a minimal Active Inference stack on toy models: a Bernoulli–Ising analytical oracle, a pymdp T-maze rollout, and a sheaf-indexed compose contract that binds 34 fragment tracks into 12 flat IMRAD sections. The methodological contribution is a discipline rather than a domain finding: every reported number is hydrated from a generated artifact and every cross-track claim is machine-checked before rendering, so no figure or statistic can drift from the artifact that produced it — 6 sheaf axioms are verified before composition and 25 negative controls keep each failure path live. Claims are limited to those models and their generated artifacts.\n\n reports a 17-row coverage matrix (5 IMRAD group headers) regenerated from the live manifest at compose time. documents the T-maze harness aligned with pymdp sophisticated_inference examples (https://github.com/infer-actively/pymdp/tree/main/examples/experimental/sophisticated_inference).\n\n records 12 / 12 invariant checks passed. SI planning horizon: 2 steps. Sweep RMSE 0 nats bounds analytical–empirical agreement on the coupling grid.\n\n---\nAssociated artifacts\nGitHub release: v0.3.2 (https://github.com/docxology/template_active_inference/releases/tag/v0.3.2)\nDOI: https://doi.org/10.5281/zenodo.20417021\nZenodo: https://zenodo.org/records/20417021\nPDF SHA-256: f191b48f94394cab17069fd04502c59fc1c287e7893eb078e05ba4be04d4a04c", "keywords": [ "active inference", "pymdp", "sophisticated inference", "generalized notation notation", "lean" ], "findings": [ "The paper frames its result as a methodological discipline rather than a domain claim: 6 sheaf axioms are machine-checked and 25 negative controls keep failure paths live.", "It reports 12/12 invariant checks passed and a sweep RMSE of 0 nats between analytical and empirical values on the coupling grid.", "The measured state_inference T-maze rollout reports mean belief entropy 0.3251 nats over 2 steps, with goal reached and action diversity 2.", "A coverage audit reports 95 present, 95 bound, and 0 missing cells on the IMRAD matrix.", "The author states the models are pedagogical and validate consistency and artifact wiring, not empirical claims about biological agents." ], "methods": [ "Bernoulli–Ising analytical oracle (K=2) — Closed-form mutual information and free-energy decomposition on a symmetric Bernoulli–Ising toy, cross-checked by an independent exact recomputation via total correlation.", "Deterministic pymdp T-maze rollout — A minimal T-maze following pymdp sophisticated_inference examples, defaulting to state_inference with planning horizon policy_len = 2 and logged beliefs and actions.", "Sheaf-indexed manuscript compose contract — Binds 34 composable fragment types to manifest rows under an IMRAD outline, verifying sheaf axioms and negative controls before rendering.", "Lean boundary-witness formalization — Lean modules checked by lake build state small finite T-maze and graph-world witnesses, with axioms audited via #print axioms; explicitly not a broad formalization.", "Artifact-hydrated reporting with validation gates — Every reported number is hydrated from generated artifacts and cross-track claims are machine-checked by pipeline gates before the PDF is built." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_ActiveInferenceMulti/metadata.json", "findings": "papers/2026_ActiveInferenceMulti/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026BeeStackEvidenceTypedScaffold125": { "citation_key": "Friedman2026BeeStackEvidenceTypedScaffold125", "abstract": "BeeStack is an executable, evidence-typed research scaffold for whole-colony simulation of the Western honey bee (Apis mellifera), organized as five layers (Body, Brain, Mind, Swarm, Niche). It pairs FlyBody/MuJoCo body and small-scene swarm renders with curated empirical BeeBrain datasets and reduced deterministic kernels, keeping fidelity a declared per-module property: every quoted number is traceable from configuration to artifact to manuscript, gaps are catalogued rather than hidden, and the validation rate is a config-band self-test measure, not a biological-realism score. This 1.0 release accompanies the manuscript 'BeeStack: An Evidence-Typed Scaffold for Whole-Colony Honeybee Simulation' and includes the combined PDF and the full source archive. Source: https://github.com/docxology/BeeStack", "keywords": [ "honeybee", "Apis mellifera", "active inference", "simulation scaffold", "swarm intelligence", "niche construction", "FlyBody", "MuJoCo", "antennal lobe", "mushroom body", "central complex", "waggle dance", "BEEHAVE", "reproducible research" ], "findings": [ "The empirical run integrates 48 response panels, 7 anatomy inventories and 24 odor templates, with a parseable-source fraction of 0.800.", "All module contract self-tests pass (rate 1.000) alongside 11 catalogued open gaps; the authors stress this is contract conformance, not biological validation.", "BeeBody visual scores (0.980 visual, 1.000 silhouette) certify that renders look like a bee, not that kinetics match; masses, adhesion and aerodynamics are FlyBody defaults.", "Strict swarm scenes are small scripted-pose contact scenes (3 scenes, 15 unique contact pairs); colony dynamics use reduced kernels with 50 agents representing 20,000 workers.", "The authors state the results do not show BeeStack can yet predict colony survival, pesticide response, full dance-language use, or field foraging success." ], "methods": [ "Five typed Python modules (Body, Brain, Mind, Swarm, Niche) with contracts — A five-layer honeybee specification is converted into five typed Python modules with explicit contracts, deterministic seeding and a hydrated manuscript.", "BeeBody: FlyBody walking/flight tasks on a generated honeybee MJCF in MuJoCo — Body renders use FlyBody walking and wing-beat flight tasks through a generated honeybee MJCF body plan in MuJoCo; strict swarm scenes use scripted poses with real contact detection.", "BeeBrain ingestion of curated public Apis mellifera datasets — Public anatomy and activity datasets (Honey-Bee Standard Brain, glomerular odor codes, calcium imaging, etc.) are parsed into anatomy inventories, response panels and dance templates.", "Reduced deterministic kernels for BeeMind, BeeSwarm and BeeNiche — Active-inference-style policy scoring, dance recruitment and comb/thermal stepping are implemented as reduced, hand-calibrated deterministic kernels with diagnostics.", "Claim ledger and manuscript hydration from run-time JSON — A claim ledger maps each claim class to evidence and what it does not prove; manuscript numbers are read from pipeline-generated JSON, and tests use no mocks." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_BeeStack/metadata.json", "findings": "papers/2026_BeeStack/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026WhenDoBugsSee126": { "citation_key": "Friedman2026WhenDoBugsSee126", "abstract": "Objective: To review the plausibility of insect detection of infrared (IR) cues that covary with semiochemical vibrational signatures, and to produce falsifiable predictions through the integration of comparative entomology, spectroscopy, neural timing analysis, and computational electromagnetism. The vibrational theory remains contested, so the framework treats IR/vibrational sensing as a testable complement to molecular recognition rather than a replacement for receptor binding . Methods: We integrate: (i) literature-grounded morphometric ranges for antennal sensilla, (ii) ATR-FTIR evidence that insect body chemistry can support species discrimination, (iii) published olfactory receptor neuron timing constraints, and (iv) deterministic electromagnetic models that expose their assumptions and parameter sensitivity . Preregistered experimental protocols specify QCL/LED bands (2--25 µm), thermal matched controls, power density 0.1--2 mW/cm², and N≥50 per condition. All analyses use fixed random seeds (42) where stochastic routines are present. Results: The computational figures show where sensillum-scale dimensions, CHC-associated mid-IR bands, and atmospheric windows overlap, but they do not by themselves establish biological IR olfaction. The strongest empirical anchors are narrower: fast insect ORN first-spike timing, photomechanic IR organs in pyrophilous beetles, hematophagy IR cues in mosquitoes and kissing bugs, thermogenic pollination signals in cycads, thermosensitive coeloconic sensilla in ants, and passive cuticle IR optics . These sources motivate specific experiments while also constraining the manuscript's range and mechanism claims. Conclusions: The framework yields five preregistered falsifiers aligned with \\Cref{sec:discussion}: (1) spectral nulls under matched thermal load, (2) geometric mismatch between sensilla dimensions and predicted resonances, (3) environmental misalignment of CHC peaks with transmission windows, (4) temporal indistinguishability of IR versus thermal ORN latencies, and (5) behavioral independence of IR-only orientation from chemical gradients. Protocols specify QCL/LED bands (2--25 µm), matched power deposition, and N≥50 per condition to separate electromagnetic detection from thermal artifacts. Implications: Applications of this work include biomimetic IR sensor design, better-controlled pest-monitoring experiments, and clearer tests of whether insect olfactory systems ever use wavelength-specific electromagnetic information. Keywords: insect olfaction, infrared detection, vibrational theory, electromagnetic sensing, sensilla morphology, cuticular hydrocarbons, atmospheric transmission, biomimetic sensors Reproducibility: Complete implementation with seven case studies in Appendices --- Associated artifacts GitHub release: v1.0.0 (https://github.com/docxology/cohereants/releases/tag/v1.0.0) PDF SHA-256: 36bd97a8ff522937390213a69abb40c8c87a06ab783457e62d7effef4762bac2
", "keywords": [ "insect olfaction", "infrared detection", "vibrational theory of olfaction", "semiochemicals", "sensilla morphology", "electromagnetic sensing", "active inference", "reproducible research" ], "findings": [ "The computed figures show where sensillum dimensions, CHC-associated mid-IR bands, and atmospheric windows overlap, but do not by themselves establish biological IR olfaction.", "The framework yields five preregistered falsifiers, including spectral nulls under matched thermal load and geometric mismatch between sensilla and predicted resonances.", "Published insect ORN timing is fast enough that any IR stage would need to be experimentally separated from already-rapid molecular responses.", "Beetle, kissing-bug, ant, cycad, and mosquito examples establish radiant IR sensing precedents but not direct semiochemical IR olfaction.", "On a CHC spectrum fixture, the paper's automated peak detection identifies CHC-associated bands that published ATR-FTIR work links to species discrimination; perceptual use of those bands remains to be tested." ], "methods": [ "Integration of literature morphometrics, ATR-FTIR evidence, ORN timing, and EM models — Combines published sensilla ranges, FTIR insect-chemistry evidence, olfactory neuron timing constraints, and deterministic electromagnetic models.", "Coarse atmospheric IR transmission window model with sensitivity terms — A baseline window model plus humidity, temperature, scattering, and path-length terms, explicitly framed as a scenario generator.", "Quarter-/half-wave resonance estimates for sensilla as dielectric antennas — Representative sensilla classes, anchored to published Thripidae measurements, are compared against IR wavelengths via resonance and waveguide calculations.", "Unit-tested deterministic code (CohereAnts) with fixed seeds and coverage gate — All models are implemented in tested src/ modules with seed 42 and a 90% coverage gate, with seven appendix case studies.", "Preregistered IR-only assay protocols with thermal-matched controls — Specifies single-sensillum electrophysiology, behavioral IR-only assays, and SEM morphometrics with QCL/LED bands and N>=50 per condition." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_WhenDoBugs/metadata.json", "findings": "papers/2026_WhenDoBugs/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "theoretical" } }, "Friedman2026SelfImprovementAgentHarness127": { "citation_key": "Friedman2026SelfImprovementAgentHarness127", "abstract": "This exemplar documents template_sia, a deterministic implementation of the Self-Improvement Agent (SIA) harness contract described in the Self-Improvement Agents specification (Hexo AI, 2026, arXiv:2605.27276). The default pipeline replays fixture-backed generations for the mini_classify task; opt-in live mode runs bounded target subprocesses and optional Ollama-backed meta/feedback steps.\n\nRun snapshot. Task mini_classify, run 1, 3 generation(s), live=false. Final accuracy=0.8333 over 6 held-out samples. Values are injected by scripts/z_generate_manuscript_variables.py after analysis.\n\nKeywords: self-improvement agents, benchmark harness, reproducible evaluation, agent loops\n\n---\nAssociated artifacts\nGitHub release: v0.1.2 (https://github.com/docxology/template_sia/releases/tag/v0.1.2)\nDOI: https://doi.org/10.5281/zenodo.20453879\nZenodo: https://zenodo.org/records/20453879\nPDF SHA-256: 6e6d19d04182628bb825471cf8094b5c32d2c491d2c646652ec7e2439ba80773", "keywords": [ "self-improvement agents", "benchmark harness", "reproducible research", "agent evaluation" ], "findings": [ "In the bundled fixture-replay run, final accuracy on mini_classify was 0.8333 over 6 held-out samples after 3 generations.", "Accuracy rose from the first to the final generation by a metric delta of 0.3333 in the fixture replay.", "template_sia shows the SIA harness contract can be embedded in the Research Project Template without vendoring upstream orchestration code, split into an infrastructure layer and a project layer.", "The author states the fixture-replay metrics validate wiring only and are not evidence of state-of-the-art self-improvement." ], "methods": [ "Meta -> Target -> Feedback three-agent SIA loop — The harness cycles a meta agent that seeds a target agent, the target run on public data, and a feedback agent that reads private metrics to propose the next generation.", "Public/private task split with deterministic reference baseline — Each task separates agent-visible data and instructions from held-out evaluation labels, plus a deterministic reference target agent.", "Fixture-replay determinism contract with opt-in live mode — By default generations replay recorded fixtures so CI never runs generated code or calls LLM APIs; a flag enables bounded subprocess execution with optional Ollama feedback.", "mini_classify single-feature threshold classifier task — The bundled exemplar task is a threshold classifier on one feature column, evaluated over 3 generations on 6 held-out samples." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_SelfImprovementAgent/metadata.json", "findings": "papers/2026_SelfImprovementAgent/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2025BiofirmDevelopmentFirstPrinciples128": { "citation_key": "Friedman2025BiofirmDevelopmentFirstPrinciples128", "abstract": "", "keywords": [], "findings": [ "Presents Active Inference 'biofirm' homeostatic agents as an approach to bioregional and regenerative finance.", "Lists features to avoid in a new economy, including commodification of nature and people and total fungibility of asset classes.", "Proposes an open-source biofirm kernel built on homeostatic, biomimetic principles such as self-organization, self-healing and self-modeling.", "Names the Free Energy Principle and Shapley Value as the basis for cooperative incentives and value allocation.", "States criteria grounding the biofirm in the scientific method, Bayesian statistics and the Markov blanket, with scale- and domain-free MRV metrics." ], "methods": [ "Mapping biofirm design to Ostrom's commons governance principles — The slides use Ostrom's commons design principles and social-ecological system concept as the reference framework for biofirm governance." ], "concepts": [], "sources": { "methods": "papers/2025_BiofirmDevelopmentWith/metadata.json", "findings": "papers/2025_BiofirmDevelopmentWith/metadata.json" }, "source": "papers/2025_BiofirmDevelopmentWith/metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "presentation/transcript" } }, "Friedman2025ConCatEnateEmulation129": { "citation_key": "Friedman2025ConCatEnateEmulation129", "abstract": "Updated added a number of pages at the start (from page 2) that describes the entire system and whats missing apologies for not giving this overview of what would happen given those missing parts and the purpose of demo0.01 which substitute those parts to see whether the break down of the 10 miliscond to (longer period) with each handwritten sim, does stack chems as mentioned; and thus imply correct behavior some of the documents mentioning demo 0.01 and its motivation wasnt included (mentioned in the catpilot) therefore might cause confusion (ontop of also not really mentioning what would each of the missing part cause,) much apologies part 1 and part 2 updated of the emulated cat goes along with these documents: pilot: https://zenodo.org/records/14737043 beacons: https://zenodo.org/records/14737060 bevcyc: https://zenodo.org/records/14737076 ", "keywords": [], "findings": [ "The authors state the document's purpose is only to justify the proposed flow, which is a posit testable only once its components exist.", "They list missing components, including sim, CYC and chem tables, the CYC modulator (a process reward model), a trigger table and a sim resolver.", "The authors argue a process reward model is critical: without it, it would be impossible to check which route or CYC completion was better.", "The authors flag as untested whether the proposed decreasing-slot protocol for breaking down CYCs is valid or runs smoothly." ], "methods": [ "Independent vision-processing modules (Scanning Trio, Track Trio) — Proposes organ-like modules (modScan/modLoc/modNav for maps, points, routes; modWatch/modLook/modTude for shading, attention, tests) that process vision input and publish triggers.", "Chempool emulation of accumulated chemicals/hormones — Uses chempool nodes attached to modules to tally stacked chemical notations (e.g. STRESS^, #ERROR) that time and combine trigger initiation.", "CYC drivers decomposed into |n| slots by a CYC modulator — Epigenetic-emulation CYC drivers (e.g. WAKE-HAPPY-WAKE) are split into six-slot blocks force-filled from priority tables set by the CYC modulator, then resolved.", "modStream trigger grading into themes, latches and simulations — Triggers from the vision trios (and CYCs) are graded by modStream into themes and latches (simulations), which modLatch resolves into updates.", "Planned DEMO 0.01 with hardcoded CYCs and simulations — Proposes a demo that handwrites 11 CYCs and a simple map simulation, 'cheats' the PRM step, and shows chem accumulation and trigger selection without learning." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2025_ConCatEnate/metadata.json", "findings": "papers/2025_ConCatEnate/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "theoretical" } }, "Friedman2025GraphspeakLanguageHandshake130": { "citation_key": "Friedman2025GraphspeakLanguageHandshake130", "abstract": "An experiment on decomposing language onto initially established graph-theory representation applied then onto some pre-selected transformations (commonly used context/tool related transmutes) (t1,t2,t3) then upon some layers of decomposition finding out the shared context / what we label as a handshake; for the purpose of education / others", "keywords": [], "findings": [ "Defines graphspeak as an experimental communication method whose medium is permutative graphs requiring a pre-handshake selection of sampling nodes.", "Argues that such a language's success depends not on rigor but on empathy, framed as composing the recipient's capability with their context.", "States planned objectives of testing deployment for kids, for cross-language generalizability, and for rigor." ], "methods": [ "Proposed three-layer graph sampling, conversion and decomposition scheme — Outlines Layer 0-2 steps: sampling nodes with graph theory, converting via chosen permutative formats, then decomposing and comparing similarity.", "Three provisional 'transmute' rules for node composition (Tr1-Tr3) — Defines three temporary notational norms for combining and decomposing nodes, illustrated with car-component examples.", "Planned pipeline of LLM querying with manual human confirmation — Proposes data flows in which LLM labeling of objects and transmutes is followed by a manual confirmation layer before stashing.", "Proposed graph-composition game as a test and application vehicle — Describes a planned game in which players compose trees of 3-node graphs and apply preselected transmutation types." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2025_Graphspeak/metadata.json", "findings": "papers/2025_Graphspeak/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "other" } }, "Friedman2025BevCycPositPrimitiveDrivers131": { "citation_key": "Friedman2025BevCycPositPrimitiveDrivers131", "abstract": "BehavioralCycle / CYC / “ from epigenetic-sequence / epigenetic-cycle” transcribed onto ROS-like / gymnasium like Nodes-of-action-sequence; but in actual essence; is just an empty node-scaffold To fill with action-context-pairs that is designed so that they are decomposable Onto the lowest level (which is muscle tension %) “ BevCYC is a framework that implements a decomposable Chain of Node ( N – N – N ) Representing an Action Sequence / Concept Sequence That is based on the Epigenetic Cycle ideations.", "keywords": [], "findings": [ "The authors posit that state–action–state triplets can be treated as state–state–state, deliberately blurring the state/action distinction.", "The authors suggest a minimal 'silly cat' agent demonstrating only social-cortex logging and applying could be built in the near term.", "The authors flag that the work is experimental and hypothesized, with later revision for proper active inference modeling contingent on technical help." ], "methods": [ "Decomposable chain of nodes (N–N–N) for action/concept sequences — Proposes BevCYC as a chain of empty nodes for action-context pairs, modeled on epigenetic-cycle ideas, representing action or concept sequences.", "Hierarchical decomposition down to muscle tension % — Nodes are decomposed hierarchically until they resolve to a muscle tension percentage at the lowest level.", "Shrinking $$$] / $$] / $] scaffold with interruption slots — Experiments with a scaffold whose earlier nodes have more slots (L) for interruption or vision/theme injection than later ones.", "Worked sketch: cat food-seeking cycle (EpiFoodPlau) as ROS-style nodes — Illustrates the idea with a cat morning routine of chained triplets (wake–roam–sleep, roam–hunt–roam) treated as ROS nodes with completion checkpoints.", "Proposed conversion node/table to robotics frameworks (BT, HTN, STRIPS) — Suggests converting BevCycles to behavior trees, HTNs, motor primitives or STRIPS planning by adding a conversion node and table." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2025_BevCycPositPrimitive/metadata.json", "findings": "papers/2025_BevCycPositPrimitive/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": { "source": "papers/paper_metadata.json", "date": "2026-10-01" }, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "theoretical" } }, "Friedman2025BeaconsPreEmulationSocial132": { "citation_key": "Friedman2025BeaconsPreEmulationSocial132", "abstract": "primitive emulation of social cortex currently only about tagging object types and labeled objects in simulations and having them refer to a table of category and response, which might include a sim of their own sensitive to cues such as mood or transferable cues of when the object type is called; then \"emulated\" in the mind a pulse-to-response only version of the associated objects for this purpose ( should update with more social cortex posits soon )", "keywords": [], "findings": [ "Proposes classifying entities on a certainty spectrum, from predictable (walls, rocks) to highly dynamic social entities carrying their own prediction matrix.", "Introduces a self-social (OTSSA) label so identities can be represented and swapped into social-social simulations in place of another social creature.", "Notes that biologically objects may first be treated as social and then downgraded, but the sketch deliberately starts bottom-up for an optimized working agent.", "States the aim of building an agent by attaching any vision model to a labeling regime, which the authors flag as subject to change." ], "methods": [ "Cat-in-environment thought scenario for a simulated agent — The design is framed around a single agent (a cat) perceiving an environment of objects, each treated as a signal-publishing beacon.", "Pattern-scan and labeling procedure with object-type codes — A numbered formal process scans vision for patterns, ranks them by significance, and assigns labels such as OTSP, OTSA, OTSB and OTSSB by object type.", "Beacons table with confirmation and weight pulses — Sketches building a beacons table, checking nearest object types, then publishing pulses as weights onto the agent's simulation.", "Baseline-to-learned epigenetic reaction tables — Beacon processing trees are encoded in a table that starts from an inherited baseline of reactions and is modified by learning." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2025_BeaconsPreEmulation/metadata.json", "findings": "papers/2025_BeaconsPreEmulation/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": { "source": "papers/paper_metadata.json", "date": "2026-10-01" }, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "theoretical" } }, "Friedman2025ConCatEnatePilot133": { "citation_key": "Friedman2025ConCatEnatePilot133", "abstract": "Short overview of the concatenate amateur cat Combined with Beacons Bevcyc and demo 0.01 plan", "keywords": [], "findings": [ "The authors summarise the pipeline as: get vision input, label/tag/process, combine with CYC, form a SIM, learn the SIM outcome, and update.", "The authors state the methods are currently handwritten and optimization is still in progress.", "The authors state they lack a consistent automated formula but believe they can produce a full behavior table for any requested video or activity." ], "methods": [ "Proposed decomposable action-sequence cat agent with ROS-style checkpoints — The pilot sketches a cat agent built from decomposable action sequences using ROS-style action checkpoints.", "BEACONS: pulse-and-response-table emulation of a social cortex — Tagged objects are assigned pulses with a response table so the cat can decide how to respond, hedge or prepare toward each object.", "BEV-CYC: inherited-then-learned chains of decomposable behavior nodes — Behavioral drives are represented as node chains, hardcoded from birth and then edited by learning, decomposed down to muscle-tension decisions.", "Time-limited SIM resolution with 'quickcalc' free-energy shortcuts — Each simulation is resolved within a time limit and updates parameters; low-novelty SIM types are converted to quickcalc formulas over a free-energy range." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2025_ConCatEnate2/metadata.json", "findings": "papers/2025_ConCatEnate2/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": { "source": "papers/paper_metadata.json", "date": "2026-10-01" }, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "other" } }, "Friedman2024PushPullPrimingSequence134": { "citation_key": "Friedman2024PushPullPrimingSequence134", "abstract": "The primary objectives of this Push and Pull document are to explore the decoupling of cognitive ( covert , attentional) and bodily ( overt ) behaviors , and to provide a sequence of hands-on computer exercises for further exploration . The priming sequence of exercises is presented to inform an individual’s understanding of how their attention interacts with computer use movements, such as controlling the on-screen cursor with mouse, touchpad, or eye movements. What is presented here is only an initial sequence to complement other and future development of “hand-I” uncouplings, in principle and in practice.", "keywords": [], "findings": [ "The author reports a preliminary impression that attentional push movements are faster and carry lower cognitive burden than goal-oriented pull strategies.", "The author stresses that push is not easier or better than pull, only different and faster.", "The document presents itself as only an initial sequence meant to complement future development of 'hand-I' uncouplings." ], "methods": [ "Eight-step self-guided computer exercise sequence (A-to-B cursor/eye movement) — Readers move a mouse, touchpad, eyes, or attention between two points A and B, alternating 'push' and 'pull' cognitive strategies across eight progressively layered exercises.", "Layering push/pull movement with lip pursing, vocalization, and language generation — Later steps add bodily constraints (pursed lips), repeated syllables, memorized text, and freeform speech while alternating push and pull movements." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2024_PushPull/metadata.json", "findings": "papers/2024_PushPull/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "other" } }, "Friedman2023NaturalAIBasedScience135": { "citation_key": "Friedman2023NaturalAIBasedScience135", "abstract": "Letter on: \"A Natural AI Based on The Science of Computational Physics, Biology and Neuroscience: Policy and Societal Significance\". v1 released on December 12, 2023. ", "keywords": [ "Active Inference", "AI", "Natural", "Policy" ], "findings": [ "The letter argues that current LLMs and Transformer models are 'corpus bound', with parameter-setting criteria hidden in an undecipherable stochastic black box.", "The signatories argue that research in computational neuroscience, biology and physics supports a more nuanced and positive narrative of intelligence than AGI/existential-threat narratives.", "The letter contends AI need not be monolithic or concentrated in Big Tech to be commercialized and regulated, anticipating distributed, biologically grounded intelligences on mobile devices.", "The signatories call for a science-based understanding of AI's biological foundations to be given public voice and for interdisciplinary public workshops among legislators, regulators and other stakeholders." ], "methods": [], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "findings": "papers/2023_NaturalAIBased/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "commentary/policy" } }, "Friedman2023ATLASQuestionOrientedApproach136": { "citation_key": "Friedman2023ATLASQuestionOrientedApproach136", "abstract": "The ATLAS system, evolving since the late 1990s, stands as a dynamic and comprehensive knowledge management tool that intends to address the complexities of modern information supply chains. The antecedent to ATLAS was the Atlas of Risk, an informal assemblage of various risks associated with digital interactions. Here we provide an initial specification for digital prototypes and paper-and-pencil implementations of a matured ATLAS architecture which integrates pattern language approaches with question-oriented procedures to manage and interpret meaning and context. The ATLAS system facilitates the management and communication of nuanced data sets and knowledge bases with an eye towards interoperability without the need for fully shared standards. The development of ATLAS, driven by the need for enhanced data interoperability and shared understanding in an increasingly complex and volatile digital landscape, reflects a profound, community response to the challenges of information environments and the fragility of extreme specialization. ATLAS's ongoing evolution showcases its adaptability and significance in the realms of data analysis, knowledge management, and cognitive security, and this first release of a technical specification establishes a foundation for a transition from prototype to scale-appropriate implementation.", "keywords": [], "findings": [ "The prototype work revealed that numerous communities maintain pattern languages with limited interoperability.", "Combining parent-child pattern relationships with question lists (QKits) yields a dynamic type system in which objects acquire implied Patterns through requests for information.", "The prototype integrated several dozen pattern collections into over 1,000 networked patterns.", "The design treats questions, missing information, and disagreement over information as information in their own right.", "The authors expect knowledge-base growth to follow a sigmoid rather than exponential curve as iQueries increasingly network existing objects." ], "methods": [ "Alpha COGSEC ATLAS prototype built on Coda.io (2023) — A first digital ATLAS prototype applied to cognitive-security vulnerabilities and exploits, built on a low-code platform, informed the specification.", "Object-class specification: Entity, Pattern, iQuery, Attribute, Prompt Interface — Defines core ATLAS components with properties and methods, sufficient for digital or paper-and-pencil implementation.", "Question Oriented Design via iQuery routers and pattern QKits — Structured queries route answers to attributes and assign implied Patterns, which in turn attach new question kits (dynamic typing).", "IXE and VIE designations for interoperating ATLAS instances — Defines Information Exchange Environments and Verified Information Exchange Environments with quality assurance and enforcement criteria." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2023_ATLAS2/metadata.json", "findings": "papers/2023_ATLAS2/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2023TranscriptActiveInferenceGuestStream137": { "citation_key": "Friedman2023TranscriptActiveInferenceGuestStream137", "abstract": "Transcript from livestream on July 25, 2023 at the Active Inference Institute. YouTube watch link: https://www.youtube.com/watch?v=dUXfgzKHV1c Repository with updated transcripts and accessory files: https://github.com/ActiveInferenceInstitute/ActiveInferenceJournal/tree/main/GuestStream/GuestStream_049 Information on the Institute: https://www.activeinference.org/", "keywords": [ "Consciousness", "Science", "Communication" ], "findings": [ "Peters states that the results unveiled at the 2023 public event challenged both Global Neuronal Workspace and Integrated Information Theory, contrary to how the press framed them.", "Peters argues that missing nuance in news coverage changed the meaning of what was presented.", "Peters argues that such journalism damages the field's legitimacy with peers, funders and publication venues.", "Bradford advises readers that press articles making sensational claims, such as finding the root of consciousness, are probably not true.", "Bradford recommends researchers be clear about what level of 'consciousness' they mean when presenting work or talking to journalists." ], "methods": [ "Timestamped transcript of a July 26, 2023 Active Inference Institute livestream — Records a ~10-minute talk by Megan Peters followed by a three-way discussion with Nora Bradford and Daniel Friedman.", "Peters' case examples of media coverage of consciousness science — Peters walks through quoted headlines and passages from news coverage (e.g. Nature, Science, The Economist) and critiques each." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2023_TranscriptActiveInference/metadata.json", "findings": "papers/2023_TranscriptActiveInference/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "presentation/transcript" } }, "Friedman2023ModernNostrIndexCard138": { "citation_key": "Friedman2023ModernNostrIndexCard138", "abstract": "Some concepts explored related to Knowledge Engineering, Nostr, Large Language Models, Complexity, and more. ", "keywords": [ "Nostr", "Complexity", "Large Language Model", "Knowledge Engineering" ], "findings": [ "The notes propose linking papers not only by citation edges but by syntactic bridges (keyword cards) and semantic bridges (embeddings).", "The notes argue auto-generated flashcards pose less of an information-overload risk than auto-generated papers, since unused cards are simply ignored and useful ones composed.", "The notes suggest review papers could follow the most popular paths through the card graph and novelty search the least traversed ones." ], "methods": [ "Index cards as hashed Nostr notes (JSON to SHA256 ID) — The notes sketch each index card as a JSON object (username, text, etc.) hashed into a 32-byte SHA256 identifier, with edges defined between card IDs.", "LLM semantic embeddings attached to index cards — Proposes using language-model embeddings, translations and summaries so cards act as semantic bridges between texts.", "Path analysis over composed index-card graphs — Proposes analyzing paths through card graphs for trivial features (overall length) and subtler ones (e.g. which proportion of links are novel reports)." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2023_ModernNostrIndex/metadata.json", "findings": "papers/2023_ModernNostrIndex/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "other" } }, "Friedman2023SlidesIris139": { "citation_key": "Friedman2023SlidesIris139", "abstract": "", "keywords": [], "findings": [ "The slides identify that in the GPT framing the latent semantic state is not separated by speaker, so speakers are admixed and cannot be weighted or attended to differentially.", "Reusing a renormalization-group figure from Friston, Friedman et al. 2023, the slides sketch speaker attention as a portfolio (regime) of nested attentions at all levels.", "Listed next steps are adding visualizations/dashboards, specifying the model in GNN toward implementation, and adding an action component." ], "methods": [ "Active Inference generative model drawn as a Bayesian graph — The slides open with a labelled graph of the A, B, C, D, E, G and policy terms of an Active Inference generative model as the reference frame.", "Mapping GPT onto a minimal prior/state/observation model — A reduced D-s-o model is used to describe what GPT does: prior as training set and parameters, latent semantic state, and word strings as observations.", "Perceptual-inference extension with a transition matrix B — Temporal dynamics are added via the perceptual inference part of Active Inference, with hidden states changing over time and no action selection.", "Renormalization-group model of nested attention (Friston et al. 2023) — A figure from Friston, Friedman et al. 2023 is reused to sketch how speakers, groups and sets of groups could be modelled as nested attention." ], "concepts": [], "sources": { "methods": "papers/2023_SlidesIris/metadata.json", "findings": "papers/2023_SlidesIris/metadata.json" }, "source": "papers/2023_SlidesIris/metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "presentation/transcript" } }, "Friedman2022CanonicalNeuralNetworksPerform140": { "citation_key": "Friedman2022CanonicalNeuralNetworksPerform140", "abstract": "Transcript of a three-session series of discussions of the paper "Canonical neural networks perform active inference" by Takuya Isomura, Hideaki Shimazaki & Karl J. Friston. https://www.nature.com/articles/s42003-021-02994-2 LS #051.0: Background and context. https://www.youtube.com/watch?v=ZASG-rtkXDk LS #051.1: First participatory group discussion. https://www.youtube.com/watch?v=IM_NlUzyq8M LS #051.2: Second participatory group discussion. https://www.youtube.com/watch?v=hY_CajLpt9Q", "keywords": [], "findings": [ "As read in the stream, the paper claims a class of canonical rate-coding neural networks is universally characterised as variational Bayesian inference under a particular but generic generative model.", "Friedman relays the paper's corollary of the complete class theorem: any neural network minimizing a cost function can be viewed as variational Bayesian inference under some priors.", "Quoted from the paper: a network of rate coding models with sigmoid activation formally corresponds to a class of POMDP models.", "The paper's maze-task simulations are described as supporting the proposition, and the implicit prior was estimated from simulated neural activity as a good approximation to the true value." ], "methods": [ "Three-session Active Inference livestream (#051) on Isomura et al. 2022 — The work is a transcript of livestream #051: a solo background session (#051.0) plus two participatory discussions with Takuya Isomura.", "Solo slide walkthrough by Daniel Friedman with colour-coded paper quotes — In #051.0 Friedman walks through the paper's structure on slides, colour-coding direct quotes by perception, action, and complete class theorem themes.", "Discussion of the complete class theorem via external online resources — Friedman introduces the complete class theorem using external blog/LessWrong posts, inviting others with more expertise to contribute." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2022_CanonicalNeuralNetworks/metadata.json", "findings": "papers/2022_CanonicalNeuralNetworks/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "presentation/transcript" } }, "Friedman2022InteroceptionAsModelingAllostasis141": { "citation_key": "Friedman2022InteroceptionAsModelingAllostasis141", "abstract": "Transcript of discussions of the 2022 paper “Interoception as modeling, allostasis as control” by Eli Sennesh, Jordan Theriault, Dana Brooks, Jan-Willemvan de Meent, Lisa Feldman Barrett, & Karen S. Quigley https://www.sciencedirect.com/science/article/abs/pii/S0301051121002350 Session 050.0, October 17, 2022 https://www.youtube.com/watch?v=l7r0ISlr-Hc Session 050.1, October 20, 2022 https://www.youtube.com/watch?v=tGd-mgSdbio Session 050.2, November 3, 2022 https://www.youtube.com/watch?v=4o-LmkycAC0 ", "keywords": [], "findings": [ "As read in the session, the paper proposes studying allostasis in terms of control theory, with interoception providing performance feedback for allostasis.", "Sennesh says the motivation was that real biological prior preferences cannot be fixed, raising how much must be learned over development rather than built in.", "Sennesh summarizes that the model yields a notion of affordance competition via forward simulations of trajectories evaluated over time.", "Sennesh states that normalizing to probability densities gives a common currency, so the model needs no separate reward currency.", "Sennesh notes whether interoception is predictively coded below the neck remains an open empirical question." ], "methods": [ "Transcripts of ActInf Livestream #050 sessions (50.0, 50.1, 50.2) — Records three Active Inference Institute livestream discussions (Oct-Nov 2022) of the Biological Psychology paper 'Interoception as modeling, allostasis as control'.", "Background walkthrough by Friedman and Tickles (50.0) — In 50.0 the hosts reviewed the paper's big questions, aims, abstract and roadmap before later sessions with authors.", "Author slide presentation by Eli Sennesh with Jordan Theriault (50.1) — Paper co-authors presented the model (Section 4: allostasis as stochastic optimal control) and took questions from participants." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2022_InteroceptionAsModeling/metadata.json", "findings": "papers/2022_InteroceptionAsModeling/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": { "source": "papers/paper_metadata.json", "date": "2026-10-01" }, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "presentation/transcript" } }, "Friedman2022TranscriptMarkSolmsConsciousness142": { "citation_key": "Friedman2022TranscriptMarkSolmsConsciousness142", "abstract": "This document is an enhanced transcript of the live presentations and group discussions with Mark Solms in 2022 at the Active Inference Institute. The focus is the 2018 paper "How and Why Consciousness Arises: Some Considerations from Physics and Physiology" by Mark Solms and Karl Friston.", "keywords": [ "Active Inference", "Consciousness", "Neuroanatomy", "Psychology", "Free Energy Principle", "Livestream" ], "findings": [ "Solms asserts that affect is the foundational form of consciousness and that affect is intrinsically conscious.", "Solms argues, citing hydranencephalic children, that the cortex as seat of consciousness and feeling has to be dismissed.", "Solms argues affect is not the same as interoceptive inference (distinguishing his view from Anil Seth's), because many interoceptive perceptual states are not affective.", "Solms claims the mechanism of perceptual and cognitive consciousness is precision modulation of allostatic prediction errors.", "Solms reports his group is trying to instantiate these affect mechanisms in an artificial system, with only preliminary findings presented so far." ], "methods": [ "Transcript of two GuestStream sessions with Mark Solms — Edited transcript of participatory livestream sessions (#016.1, Jan. 22, 2022 and #016.2, June 11, 2022) presenting and discussing the 2018 Solms & Friston paper.", "Solms's ten-claim slide-led presentation with interleaved Q&A — Solms presented a summary slide of points, taking each in turn and pausing for participants' questions and comments.", "Clinical neurological cases discussed as evidence — Solms draws on cases including hydranencephalic children without cortex, patients lacking prefrontal or insular cortex, and deep brain stimulation of the substantia nigra." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2022_Transcript/metadata.json", "findings": "papers/2022_Transcript/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "presentation/transcript" } }, "Friedman2022WorkedExampleBayesianMechanics143": { "citation_key": "Friedman2022WorkedExampleBayesianMechanics143", "abstract": "Transcripts of discussions of the 2022 preprint "A Worked Example of the Bayesian Mechanics of Classical Objects" by Dalton A R Sakthivadivel. https://arxiv.org/abs/2206.12996 Session 049.0, September 30, 2022 https://www.youtube.com/watch?v=OtX2Fpzn7KA Session 049.1, October 5, 2022 https://www.youtube.com/watch?v=dAtC-Enmc8M Session 049.2, October 12, 2022 https://www.youtube.com/watch?v=2SuBJBEg9LI ", "keywords": [], "findings": [ "Sakthivadivel distinguishes dynamics (observed trajectories) from mechanics (the laws generating those trajectories), arguing this is often conflated in the FEP literature.", "Rahmjoo highlights as new results the derivation of classical physics from constrained maximum entropy and the reintroduction of supersymmetry to explain classical chaos.", "Sakthivadivel explains that the path minimizing classical action is the mode (most likely path), so classical mechanics can be structured as a least-surprisal problem leading to equation 15.", "Sakthivadivel describes path tracking as handling nonstationary systems by relating the FEP over paths to the principle of maximum caliber." ], "methods": [ "Transcripts of ActInf Livestream #049.0–049.2 discussing Sakthivadivel's preprint — Records three Active Inference Institute livestream discussions of the 2022 arXiv preprint, with the author Dalton Sakthivadivel joining 49.1 and 49.2.", "Background/context session (49.0) by facilitators before author sessions — The 49.0 session by Friedman, Rahmjoo and Smékal gives background and a preliminary discussion rather than a review or synthesis.", "Section-by-section walk-through of the paper in 49.2 — In 49.2 the participants go through the paper's sections, building toward its equation 15 for Bayesian classical mechanics." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2022_WorkedExampleBayesian/metadata.json", "findings": "papers/2022_WorkedExampleBayesian/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "presentation/transcript" } }, "Friedman2022TranscriptDiscussionsCommunicationAs144": { "citation_key": "Friedman2022TranscriptDiscussionsCommunicationAs144", "abstract": "Discussion with an author of the 2021 paper “Communication as Socially Extended Active Inference: An Ecological Approach to Communicative Behavior” by Remi Tison & Pierre Poirier. https://www.tandfonline.com/doi/abs/10.1080/10407413.2021.1965480 Session #048.0, September 2, 2022 https://www.youtube.com/watch?v=zqiZjjY9H7M Session 048.1, September 7, 2022 https://www.youtube.com/watch?v=nWFM5zrdmG8 Session 048.1, September 7, 2022 https://www.youtube.com/watch?v=nWFM5zrdmG8", "keywords": [], "findings": [ "Tison explains that the paper aimed to give ecological and enactive approaches a view of communication that does not rely on content or representation, unlike the transmission view.", "Tison states that content exists but is not at the basic level of cognition; it must be explained by a theory of communication rather than presupposed.", "Tison reflects that defining a shared field of affordances as the subset common to two individual fields neglects the autonomy of the interaction dynamics.", "Tison describes cooperative communication as a way for a group to achieve an optimal grip on the context of their joint action.", "Tison says the paper ties the conventional nature of language to conventional affordances generated in communicative interaction." ], "methods": [ "Transcribed ActInf Livestream #048 series (48.0, 48.1, 48.2) — Transcript of Active Inference Institute livestream discussions of the 2021 paper 'Communication as Socially Extended Active Inference', held September 2, 7 and 14, 2022.", "48.0 introductory walkthrough of aims, claims, abstract and roadmap — In 48.0 Daniel Friedman and Dean Tickles present the paper's aims, claims, abstract and roadmap as an introduction rather than a full review.", "Author-participation panel discussions with first author Rémi Tison — In 48.1 and 48.2 the first author joins Friedman, Tickles and Bleu Knight to answer questions on content, context, shared affordances, and language." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2022_TranscriptDiscussions/metadata.json", "findings": "papers/2022_TranscriptDiscussions/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "presentation/transcript" } }, "Friedman2022TrackingPublicSensemakingThrough145": { "citation_key": "Friedman2022TrackingPublicSensemakingThrough145", "abstract": "Political polarization and declining trust in institutions are driving societal destabilization and radicalization. Recently there has been increased interest in online misinformation intervention and deterrence, for example through the use of machine learning on language use. We argue that addressing crises in the information environment will require a sharper situational awareness and a deeper understanding of how beliefs emerge and crystallize, as well as greater connectivity in the work of teams and organizations in order to reduce the effects of bias and partisanship in collection and analysis. Image memes play an increasingly important role in public sensemaking and discourse and the emergence of public beliefs. Despite their significance, image memes have proven to be a very difficult category of artifact to collect, classify, and analyze in aggregate. In this white paper, the function and form of image memes are discussed, the challenges of performing image meme collection and analysis within the context of emergent, interdisciplinary teams are detailed, and requirements and recommendations for alleviating these challenges are offered.", "keywords": [ "Sensemaking", "Memes", "Knowledge Management", "Narrative", "Rhetorical Analysis" ], "findings": [ "Argues that image memes function as quasi-arguments whose parsimonious form hides warrants, so interventions should target spurious evidence and hidden warrants.", "Argues that framing the information crisis as true versus false information has not been effective at regaining trust of disaffected audiences.", "Concludes that current image meme collection by emergent teams falls short of needs even though the needed affordances are within technological reach.", "Recommends system requirements such as on-site information, dynamic web annotations, and proximal collection and tagging affordances.", "Proposes digital ecosystem and API economy approaches as a viable route to many of the challenges, while noting data standardization and integration remain problems." ], "methods": [ "Toulmin claim/evidence/warrant annotation of example image memes — Applies the Toulmin argument model to worked examples (e.g. a NASA meme, 'Condescending Wonka') to expose implicit warrants.", "Archetypal-phase vignettes of three emergent-team collection approaches — Describes Haphazard, Methodological, and Automated Collection and Analysis approaches, drawn from the authors' experiences, with challenge tables.", "Russo-Ukrainian War memes as illustrative examples — Uses memes about the ongoing Russo-Ukrainian War to illustrate how memetic analysis can reveal shared public beliefs.", "Extension of the Digital Rhetorical Ecosystem (DRE3) framework — Builds on the authors' prior computational framework combining rhetorical analysis with an ecosystem approach." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2022_TrackingPublicSensemaking/metadata.json", "findings": "papers/2022_TrackingPublicSensemaking/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "commentary/policy" } }, "Friedman2021TranscriptKarlFriston1st146": { "citation_key": "Friedman2021TranscriptKarlFriston1st146", "abstract": "On June 21st, 2021, Active Inference Lab ( activeinference.org/ ) hosted its first Applied Active Inference Symposium, featuring Professor Karl Friston. The Symposium was structured in three sections, corresponding to the Organizational Units of the Active Inference Lab: Education, Communication, and Tools. This publication reflects an edited and enriched transcript of the proceedings of the Symposium.", "keywords": [ "Active Inference", "Free Energy Principle", "Symposium", "ActInfLab", "Bayesian Inference", "Education", "Communication", "Tools", "Generative Model" ], "findings": [ "Friston recommends toy simulations as the best way to demystify Active Inference for learners in an academic setting.", "Friston stresses that getting the generative model right is central: if it is apt for the target system, everything else usually follows.", "Friston distinguishes Active Inference from predictive coding: minimizing variational free energy is only half the game, with expected free energy needed for future-oriented action.", "On tools, Friston argues anything can be written as a generative model and hence as a variational free-energy-minimizing message-passing scheme, citing ForneyLab.", "Friston describes epidemic modeling with PCR incidence and mobility data, using Bayesian optimal design to judge whether including a data source is worth the complexity." ], "methods": [ "Edited transcript of a three-session moderated symposium — Records the first Applied Active Inference Symposium with Karl Friston, organized into .edu, .comms and .tools sessions mirroring the lab's organizational units.", "Prepared question slides posed by lab participants to Friston — Lab members developed questions in weekly meetings and presented them as slides; Friston answered and moderators gave recaps.", "Presentation of the lab's Active Inference terms list and ontology work — The .edu session presented the lab's terms list and working ontology, developed within the SUMO ontology framework as a backbone for education." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2021_Transcript/metadata.json", "findings": "papers/2021_Transcript/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "presentation/transcript" } }, "Friedman2021CollaborativeWritingCatechismBased147": { "citation_key": "Friedman2021CollaborativeWritingCatechismBased147", "abstract": "Asynchronous and remote collaborative written projects (research, field guides, code, etc.) in emergent, interdisciplinary teams can be an incredibly productive and enjoyable pursuit. The convergence of diverse perspectives, personalities, and expertise in the rapid production of written deliverables can provide immense value and insight, not just to the situation, problem, or opportunity space the team was formed to address, but also to the disciplines each author brought to the table. Yet, asynchronous and remote collaborative writing can also be rather perilous. Managing deadlines, handling disputes, communicating, staying on mission, managing resources, collective editing (and over-editing), and avoiding contradiction are only some of the challenges. While the presence of a catechism-styled operations order and the use of a team “Facilitator” can greatly improve the likelihood of success, making sure all of the authors have an alignment on protocol and etiquette while writing as a catechism-based team prevents unnecessary misunderstandings and keeps things on schedule. This "3-Paragraph Order" for Collaborative Writing, or C-3PO, is meant to rapidly onboard authors to collaborative writing procedures and etiquette. ", "keywords": [ "Remote Teams", "Collaborative Writing", "Catechisms", "OPORDs" ], "findings": [ "Sets out procedures, protocols and etiquette for asynchronous, remote catechism-based writing teams, organized by before, during and after the project.", "Recommends that teams converge on a writing approach, describing at least five styles: single primary writer, partition, joint, scribe and hybrid.", "Advises that changes to approaches, processes and expectations be declared and compared against the catechism to prevent mission creep.", "Recommends giving honest, constructive critique first and positive feedback second when asked for feedback after a project.", "Calls for a post-project debrief or 'hotwash' to plan next steps, future meetings and publicity." ], "methods": [], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "findings": "papers/2021_CollaborativeWritingCatechism/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "commentary/policy" } }, "Friedman2020ReimaginingMaps148": { "citation_key": "Friedman2020ReimaginingMaps148", "abstract": "Reimagining Maps was written after participation in a National Geospatial-Intelligence Agency Incubator hosted on Polyplexus. The field of cartography sits at the intersection of applied mathematics, engineering, geology, geography, user experience, and graphic design. Methodologies and concepts from cartography have been creatively applied in a variety of fields, such as the application of spatial mapping techniques to information in knowledge management, or the use of itinerary visualization methods in non-spatial journeys such as learning maps in learning management systems. These fields have been subjected to their own forms of development and evolution leading to new methodologies and concepts somewhat removed from their origins. Cartography itself has undergone a great deal of technology-driven development but would look very different today had it been developed as a new field through the creative application of methodologies and concepts from those it inspired. As the modern information and logistical context presents new challenges and thus new demands for maps, we propose a “reimagining of maps” through an interdisciplinary synthesis inspired by the interdisciplinary origins of maps themselves.", "keywords": [ "Maps", "Cartography", "Remote Teams", "Interdisciplinary Research", "Process Mapping", "Knowledge Management Systems", "Instantaneous Remote Teams", "Intelligence Production", "OSINT" ], "findings": [ "The authors conclude that across fields and history, mappers face similar challenges: integrating multimodal data, representing uncertainty, user customization, and designing for action rather than archiving.", "Across the mapping domains reviewed, the authors found the main cutting-edge limitation was the ability to integrate data rapidly, not data availability.", "They propose BOLTS: data specifications reasonable for Business, Operational, Legal, Technical, and Social use-cases.", "They argue that BOLTS reasonableness needs metadata to guide transformations, rather than fitting all data to universal schemas.", "Drawing on OpenStreetMap and Wikipedia practice, the authors suggest future maps may rely on reputation-based 'network-sourcing' rather than crowd-sourcing." ], "methods": [ "Concept white paper responding to the NGA 'Reimagining Maps' Polyplexus incubator — The paper is structured around the incubator's driving questions about how emerging knowledge could make maps more flexible and understandable.", "Survey of the current geospatial problem space and its key challenge areas — Part I reviews issues such as interoperability, skill gaps, uncertainty, threat actors, data volume, and accessibility, then condenses them into three challenge areas.", "Cross-field review of 'maps' in fields adjacent to cartography — Part II examines how maps are used in process mapping, software, complex systems, communications, knowledge management, education, ecology, and mathematics.", "Narrative case study based on DASA's 'Map the Gap' wet-gap crossing scenario — A bridge-construction scenario for expeditionary forces in adversarial settings is used to illustrate what reimagined maps might look like." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2020_ReimaginingMaps/metadata.json", "findings": "papers/2020_ReimaginingMaps/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "review" } }, "Friedman2020InfiniteGamesInfiniteTeams149": { "citation_key": "Friedman2020InfiniteGamesInfiniteTeams149", "abstract": "Infinite Games for Infinite Teams was published by and in response to the DARPA Polyplexus Citizen Incubator: “Inventing a Remote Culture to Deal with Pandemics”, and was done so with the intent of discussing the questions outlined below. How are global online narratives constructed and received in 2020? Why are the processes of narrative design and culture production so important for security and governance? What is possible now or soon that was not possible before? What approaches could catalyze assessment, design, and deployment of online narratives in real-time? Why is it so important to have meme-detection systems that are culturally-aware, interlingual, intermodal, and human-in-the-loop? What does it look like to take a Complex Adaptive Systems (CAS) approach to the neuromemetics of narrative co-construction and agenda-setting? How could a CAS approach be used to support specifically-defined cultural/institutional/national/global interests? How do we find, formalize, and quantify goals or outcomes within a CAS framework? How can we diagnose, perturb, and create narratives through gameplay? What might a “design science for memes” look like? How is this present work continuous with and contrasting with previous work in innovation, generative games, and LARPing? How can music, sound, art, and other techniques amplify narrative impact ?", "keywords": [], "findings": [ "Introduces the term Instantaneous Remote Teams (IRT) for rapidly formed online-native teams, described as essential for institutional function and resilience.", "Defines 'Infinite Teams' by analogy to Infinite Games as teams with open-ended, evolving composition, and uses 'Gray Zone' for the ambiguous space where they play.", "Proposes that Polyplexus could serve as a platform hosting Infinite Games such as culture creation, narrative evolution and R&D.", "Argues that AI alone struggles to detect links between memetic material and context, motivating curated knowledge networks with humans in the loop." ], "methods": [ "Complex Adaptive Systems framing of remote teams and narrative ecosystems — Uses the CAS framework to explore how games played by remote teams could be harnessed through platform design choices.", "World of Warcraft PvE/PvP play as a case study of rapid team formation — Draws on WoW raid and player-vs-player teams to illustrate rapidly assembled teams and cultural transmission.", "Design sketch: formal-informal interface for collaborative world-building — Idea I proposes formal memetics where technical users set world parameters while prose and art contributors co-evolve the narrative.", "Design sketch: Red/Blue/Green 'Cadavre Exquis' role-playing game — Idea II proposes a case-management-like knowledge-mapping game with sword, shield and integrating roles plus role-specific guiding questions." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2020_InfiniteGamesInfinite/metadata.json", "findings": "papers/2020_InfiniteGamesInfinite/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "theoretical" } }, "Friedman2026MusicNeverStoppedGrateful150": { "citation_key": "Friedman2026MusicNeverStoppedGrateful150", "abstract": "We present a modular, citation-bound data compendium for the Grateful Dead universe — shows, songs, performances, personnel timelines, venues, recordings, and reception — and a category-theoretic interpretation of the performance graph. The work is grounded in the archival reality that Grateful Dead history is both institutional and participatory: UCSC's Grateful Dead Archive and the Internet Archive collection preserve formal and community records , while taping and trading scholarship shows why setlists and recording metadata are cultural evidence, not merely fan trivia . The surrounding source dossier also binds the non-quantitative historical frame -- formation and Acid Test context, Wall of Sound engineering, live recording/liveness scholarship, Deadhead sociology, studio-era reception, and public recognition -- to checked sources rather than to folklore alone . The compendium integrates nine primary sources (Setlist.fm , The SetList Program , the Mark Leone CMU setlist archive , GDsets , gdshowsdb , the Internet Archive Live Music Archive , the Alex Allan / whitegum lyric finder , the official band site , and Wikipedia ) with four reference sources (Britannica , the lineup-changes guide , Dodd and Trist's The Complete Annotated Grateful Dead Lyrics , and the Grateful Stats front-end ) and secondary corpora and community discussions . Each source is parsed by an independently testable reference module written against the documented record shape; the committed compendium under `data/archival/` is the dataset reported here (3341 ingested shows (gdshowsdb + truckin gap-fill; community literature estimates ~2318 canonical concerts), 645 songs, 912 venues, 40757 performance rows). A runtime completeness audit and figure-validation gate certify referential integrity and non-degenerate outputs on every pipeline run. Integration is a deterministic, sort-keyed merge over canonical slugs; registered figures also emit CSV/JSON data tables, and a first-principles claim ledger classifies each major result by irreducible input, hard constraint, assumption, validation artifact, and interpretation limit. Exploratory repertoire/uncertainty panels are labelled as pattern-discovery rather than causal inference. We then exhibit four small but real categorical constructions, situated against transformational and categorical music-theory precedents : a poset category of dates, a discrete category of shows, a monotone cumulative setlist functor and lineup functor from dates into sets, and a span representation that takes each performance to be the apex of a span between its show and its song. Wide pullbacks over a fixed show recover the show's setlist; wide pullbacks over a fixed song recover the song's performance history. The active-band roster, by contrast, is a non-monotone presheaf on the date poset — a categorical formalization of the familiar fact that members come and go. The artefacts in this paper come from the committed archival snapshot; all source-ingestion modules are written against the real source shape so that `scripts/00_fetch_sources.py --online --write-archival` refreshes the full snapshot. --- Associated artifacts GitHub release: The Music Never Stopped: A Grateful Data Compendium (v1.0.0) (https://github.com/docxology/grateful_data/releases/tag/v1.0.0) PDF SHA-256: 2d42bfd04e0b32e4871d5bcbe2c65bcce58cca336da240af7c6fd03a6563a709
", "keywords": [ "grateful dead", "setlist data", "category theory", "music information retrieval", "reproducible data compendium" ], "findings": [ "The committed compendium contains 3341 ingested shows, 645 songs, 912 venues, and 40757 performance rows.", "Completeness is referential only: 282 of the 3341 catalogued shows have an empty setlist in gdshowsdb.", "Repertoire is highly skewed: the song-performance Gini coefficient is 0.74 and the top decile of songs accounts for 50.61% of non-segment performances.", "The most frequent explicit segue in the corpus is the structural \"drums\" -> \"space\" passage, with 1197 occurrences.", "Wide pullbacks over a show recover its setlist and over a song recover its performance history, while the active roster is a non-monotone presheaf on the date poset." ], "methods": [ "Integration of nine primary Grateful Dead data sources into one schema — Setlist.fm, SetList Program, CMU archive, GDsets, gdshowsdb, Internet Archive, whitegum, dead.net and Wikipedia are parsed by testable modules and merged.", "Deterministic canonical-slug merge with frozen-dataclass schema — Entities validate inputs and use canonical-slug primary keys so cross-source joins reduce to dictionary lookups; integration is a sort-keyed merge.", "Completeness audit, figure-validation gate, and first-principles claim ledger — Runtime checks certify referential integrity and non-degenerate figures; a ledger classifies each major result by inputs, assumptions, and limits.", "First-order Markov model of within-show song order with permutation-null FDR screen — Fits song-to-song transition probabilities, audits support thresholds, and screens transitions against a permutation null under FDR control.", "Category-theoretic constructions: date poset, show category, functors, spans — Builds setlist and lineup functors from dates into sets, an active-roster presheaf, and performances as spans between show and song." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_MusicNeverStopped/metadata.json", "findings": "papers/2026_MusicNeverStopped/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026DeterministicTestbedSelfOrganizing151": { "citation_key": "Friedman2026DeterministicTestbedSelfOrganizing151", "abstract": "Recent work on AutoScientists coordinates self-organizing teams of language-model agents through a small set of shared mechanisms: a champion-and-experiment-log shared state, a registry of retired dead-end directions, effect-size ranking of candidate directions, noise-band confirmation of claimed improvements, and stagnation-driven reorganization of teams. This exemplar provides a deterministic, standalone reference implementation of those mechanisms and studies them honestly as a testbed rather than as a performance claim.\n\nWe make the comparison fair by holding the total number of objective evaluations fixed: coordinated teams partition a single sequential experiment budget rather than adding parallel compute. Under that matched budget, coordination cannot — and in our results does not — beat a single-thread baseline on the final champion metric; we report the actual numbers and claim no speedup. What the testbed does demonstrate are two distinct, independently measurable benefits. First, noise-robustness: because the objective is stochastic, a single observed gain can be a draw of evaluation noise, so we separate the reported champion metric from the clean noise-free ground truth and show that noise-band confirmation shrinks the gap between them by roughly an order of magnitude — with confirmation on, the final champion's reported metric sits $0.0012$ above its clean value, against $0.0156$ with confirmation removed, while every configuration reaches the same clean optimum. Second, search hygiene: the dead-end registry, consulted by the proposer, cuts redundant re-probes of retired directions from $36$ to $0$ and halts at $36$ of the $60$ experiments — the same clean answer, reached with less waste. A per-mechanism ablation isolates each component's contribution, and the language-model proposer is a clean plug-in seam: a deterministic rule-based agent drives the reproducible figures, and a live Hermes agent (served by Ollama) can be swapped in without touching the coordination loop.\n\n---\nAssociated artifacts\nGitHub release: v1.0.2 (https://github.com/docxology/template_autoscientists/releases/tag/v1.0.2)\nDOI: https://doi.org/10.5281/zenodo.20533669\nZenodo: https://zenodo.org/records/20533669\nPDF SHA-256: 0af391375b14eb397812a8050657e2980fbc3a768e6fb108aa2f7eff46773e16", "keywords": [ "agent coordination", "scientific discovery", "noise-band confirmation", "ablation study", "reproducible research", "language-model agents" ], "findings": [ "Under the matched budget, coordinated teams and the baseline reach the same clean optimum (advantage 0.0000); coordination was slightly slower to first reach it (16 vs 12 experiments).", "Noise-band confirmation reduced accepted noise roughly 13-fold (reported-vs-clean gap 0.01565 to 0.00121) on this objective.", "The dead-end registry cut redundant re-probes from 36 to 0 and let the search halt at 36 rather than 60 experiments, with the clean answer unchanged.", "Effect-size ranking and reorganization did not change any measured quantity on this objective.", "The author cautions that these magnitudes are properties of this synthetic objective, budget and deterministic proposer, not general constants." ], "methods": [ "Deterministic re-implementation of AutoScientists coordination mechanisms — Re-implements AutoScientists' mechanisms (confirmation, dead-end registry, effect-size ranking, reorganization, team partitioning) as individually switchable modules.", "Synthetic rippled quadratic objective with seeded noise (d=4) — Optimizes a 4-D objective with a global peak at the origin, cosine ripples and seeded bounded noise, allowing reported vs clean metric comparison.", "Matched 60-experiment sequential budget vs single-thread baseline — Coordinated teams partition the same sequential budget as a single-thread baseline rather than adding parallel compute.", "One-at-a-time per-mechanism ablation via SearchConfig — Starts from the full coordinated configuration and switches off exactly one mechanism per ablation run.", "Pluggable Proposer: deterministic rule-based vs Hermes LLM via Ollama — Figures use a rule-based DeterministicProposer; a HermesProposer served by Ollama can be swapped in and is tested only by an opt-in test." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_DeterministicTestbedSelf/metadata.json", "findings": "papers/2026_DeterministicTestbedSelf/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026RecoveringLLMPersonaAccuracies152": { "citation_key": "Friedman2026RecoveringLLMPersonaAccuracies152", "abstract": "Algebraic (NTQR) evaluation infers how accurate a group of noisy classifiers was on a finite test using only their responses — no answer key. We test this end to end on real large language models. Three trader \"personas\" (optimistic, neutral, pessimistic), instantiated as system prompts, each make a binary bullish/bearish call on the same 64 market scenarios; we run the identical trio through six locally-hosted models via Ollama. For each model we recover per-persona, per-label accuracy with ErrorIndependentEvaluation (unsupervised) and score it against the authored ground truth (supervised), which is used only as a check. On the five models whose three judges all varied (mistral:latest, gemma4:latest, gemma3:4b, gemma2:2b, granite4.1:3b), the unsupervised algebra recovered persona accuracies to a mean absolute error of 0.012, within the 0.102 sampling-noise floor across all six per-label accuracy terms, with no labels -- including a persona's genuinely poor bullish accuracy of 0.57, recovered as 0.59. The other model collapsed at least one persona into a constant classifier (a judge that voted one way on all 64 scenarios), which makes the error-independent algebra unsolvable. The central, non-obvious result: inter-judge disagreement does not imply evaluability. Aggregate disagreement separated this run only because the unevaluable model(s) collapsed to 0.00; the five evaluable models spanned 0.03–0.23. What gates evaluation is a per-judge condition — every judge must vary (and answer) — not an ensemble one. We formalize this as a label-free evaluability diagnostic (a judge whose modal-vote fraction reaches 1.0 is a constant classifier; an unparseable vote is an abstention) that predicted exactly which models would be evaluable, before any solve and without ground truth. This is a concrete instance of the safety property the NTQR logic promises: it warns you when an ensemble is not good enough to be evaluated. A scenario bootstrap puts a 95% CI of [0.000, 0.038] on the recovery MAE (well inside the 0.102 noise floor), and a deterministic synthetic study generalizes the recovery beyond the finite set of real evaluable models — error falls like 1/√Q (slope -0.58, stable across ensembles) — while mapping two honest limits: the built-in failure alarm catches anti-correlated judges with no false positives yet can miss positively-correlated (shared-training) errors, and the two-solution tie-break inverts once judges are no longer clearly better than random — exactly where simple majority-voting evaluation, though biased, is the more robust fallback. --- Associated artifacts GitHub release: v1.0.0 (https://github.com/docxology/ntqr_llm/releases/tag/v1.0.0) DOI: https://doi.org/10.5281/zenodo.20498699 Zenodo: https://zenodo.org/records/20498699 PDF SHA-256: e1196698427f9fe04d1f3071705adb6e5459983649c78d7f5d074756e989148b
", "keywords": [ "algebraic evaluation", "NTQR", "unsupervised evaluation", "evaluation on unlabeled data", "LLM-as-judge", "error-independent evaluation", "ensemble evaluability", "constant classifier", "AI safety warning light", "reproducible research", "answer-key-free recovery", "local large language models" ], "findings": [ "For mistral:latest, unsupervised recovery matched authored-truth accuracies to a mean absolute error of 0.012, within the 0.102 sampling-noise floor.", "The algebra recovered a poor judge's accuracy without labels: the pessimist's true bullish accuracy of 0.57 was recovered as 0.59.", "Inter-judge disagreement did not imply evaluability; what gated evaluation was whether every individual judge varied, not ensemble-level disagreement.", "A label-free per-judge diagnostic (modal-vote fraction reaching 1.0) predicted exactly which models would be evaluable, before any solve.", "Synthetic studies show the failure alarm catches anti-correlated judges but can miss positively-correlated errors, and the tie-break inverts near chance-level judges." ], "methods": [ "Three system-prompted trader personas as binary judges — Optimist, neutral and pessimist personas each make a bullish/bearish call on the same 64 authored market scenarios, run through six local models via Ollama.", "Unsupervised NTQR ErrorIndependentEvaluation on vote counts — Recovers per-persona, per-label accuracy from unlabeled vote patterns only; authored truth is held out and used afterward to score recovery error.", "Deliberately unbalanced 40/24 scenario deck — The answer key is set to prevalence 0.625 to avoid the evaluator's removable singularity at prevalence exactly 1/2.", "Schema-constrained JSON vote collection — Uses Ollama structured output with a JSON Schema so parsing is a measured interface check before the binary vote matrix is analysed.", "Scenario bootstrap and deterministic synthetic-ensemble study — Bootstraps over scenarios for a CI on recovery MAE, and runs synthetic ensembles with known truth to test error scaling, correlated errors and tie-break failure." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_RecoveringLLMPersona/metadata.json", "findings": "papers/2026_RecoveringLLMPersona/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "empirical" } }, "Friedman2026TriplicateDataDrivenLarge153": { "citation_key": "Friedman2026TriplicateDataDrivenLarge153", "abstract": "We present template_newspaper, a pure-Python engine that renders a complete\ntwelve-page, large-format newspaper to a print-ready PDF from structured YAML\ncontent. The exemplar edition is The Triplicate, a homage to the historic\nnewspaper of Crescent City, California (founded 1879). The engine demonstrates\ndesigned multi-column page layout — nameplate and ears, spanning headlines,\nflowing column frames with an optional rail, drop caps, pull quotes, ruled\nmodular boxes, data tables, halftone \"engraving\" figures and running folios —\nwhile keeping a strict separation between content (data) and engine (code):\na new title is a data edit, never a code change. The layout strategy is a hybrid\nin which fixed furniture is drawn directly on the canvas to establish where\nthe column grid begins, after which body copy flows through ReportLab frames\nthat split paragraphs across columns automatically. The project obeys the\nresearch-template monorepo contract and is discovered and executed by the same\norchestration pipeline as its code- and prose-focused siblings.\n\n---\nAssociated artifacts\nGitHub release: v1.0.2 (https://github.com/docxology/template_newspaper/releases/tag/v1.0.2)\nDOI: https://doi.org/10.5281/zenodo.20533675\nZenodo: https://zenodo.org/records/20533675\nPDF SHA-256: 6c693dbd80a4234d30d99f6a890191ff47faf197a0636753fa9785588c550a77", "keywords": [ "newspaper layout", "typography", "reportlab", "reproducible publishing", "document engineering" ], "findings": [ "The engine keeps content and code strictly separate, so producing a new newspaper title is a data edit rather than a code change.", "Drawing furniture first and starting column frames below it lets spanning headlines, standing boxes and automatic text flow coexist on one page.", "Rendering is deterministic: the same edition manifest always yields a byte-identical paper.", "The bold weight in a .ttc font collection sits at a face-specific subfont index; an early render produced italic Didot headlines from assuming the wrong index.", "The test suite reports roughly ninety-five percent line coverage." ], "methods": [ "Pure-Python engine rendering YAML content to a print-ready PDF — template_newspaper reads a masthead manifest and per-page YAML files and renders a twelve-page large-format newspaper to PDF.", "Hybrid drawn-furniture / ReportLab-flowed-body layout strategy — Fixed furniture is drawn on the canvas first, then body copy flows through ReportLab frames via Frame.addFromList, splitting paragraphs across columns.", "Nine single-responsibility modules with a deterministic four-step pipeline — Editions are produced by Load, Measure, Compose and Emit steps across modules such as geometry, content, typography, figures, layout and engine.", "Procedural Pillow halftone scenes and monochrome Matplotlib charts — Photographic elements are procedural grayscale Pillow scenes with a 45-degree halftone screen; data graphics are black-and-gray Matplotlib charts.", "pytest end-to-end render tests, mypy type checking, ruff linting — A pytest suite tests the pure logic and renders the edition end-to-end, asserting a valid twelve-page PDF at correct trim with no over-set page." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_Triplicate/metadata.json", "findings": "papers/2026_Triplicate/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026TemplateTextbook154": { "citation_key": "Friedman2026TemplateTextbook154", "abstract": "A modular, fillable scaffold for book-length technical works: a data-driven manuscript (parts/chapters/labs/questions), a tested computational backbone, deterministic figures and Mermaid diagrams, and a content scaffold/validation engine.\n\n---\nAssociated artifacts\nGitHub release: v0.1.2 (https://github.com/docxology/template_textbook/releases/tag/v0.1.2)\nDOI: https://doi.org/10.5281/zenodo.20533125\nZenodo: https://zenodo.org/records/20533125\nPDF SHA-256: 7b67cb2d9118f0a72001395f0a94c529313804d7e5cdf7f6ca2dc9f7ab57d168", "keywords": [], "findings": [ "The book is explicitly a scaffold rather than a finished work: every structural element is present and author-specific passages are marked stubs.", "It provides twelve chapter shells across four parts, each with a matching lab and question bank.", "Claims building the book reproduces byte-identical figures and numbers, since nothing in the prose is computed by hand.", "The worked case study's linear fit gives slope 1.375 and R2 = 0.999, while warning that three averaged points cannot support prediction or extrapolation." ], "methods": [ "Single config.yaml source of truth for book structure — Parts, chapters, front matter, appendices, labs and question banks are declared in config.yaml, from which TOC, numbering and integrity tests read.", "Tested Python backbone (textbook.models) and deterministic figures — Worked equations are implemented as tested Python functions that chapter prose calls, and figures are generated deterministically from code.", "Pandoc + pandoc-crossref rendering pipeline — The manuscript is assembled from Markdown in declared order after analysis scripts produce figures, then rendered to PDF via Pandoc with pandoc-crossref.", "Stub-marker counting audit and manuscript-integrity tests — A quality audit counts STUB/TODO/TKTK markers and pytest checks the per-chapter content contract, unique labels, citations and glossary anchors.", "Two filled worked-reference chapters (logistic growth; dose-response) — First Principles derives the logistic growth law; Case Studies applies a linear dose-response fit to a small dataset of replicate measurements across control, low and high conditions." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2026_TemplateTextbook/metadata.json", "findings": "papers/2026_TemplateTextbook/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "book/report" } }, "Friedman2016LessonsFromColony155": { "citation_key": "Friedman2016LessonsFromColony155", "abstract": "", "keywords": [], "findings": [], "methods": [], "concepts": [], "sources": {}, "source": "", "abstract_provenance": {}, "summary_provenance": {} }, "Friedman2016FullSpeedAheadCity156": { "citation_key": "Friedman2016FullSpeedAheadCity156", "abstract": "", "keywords": [], "findings": [], "methods": [], "concepts": [], "sources": {}, "source": "", "abstract_provenance": {}, "summary_provenance": {} }, "Friedman2021Disinforge157": { "citation_key": "Friedman2021Disinforge157", "abstract": "", "keywords": [], "findings": [], "methods": [], "concepts": [], "sources": {}, "source": "", "abstract_provenance": {}, "summary_provenance": {} }, "Friedman2021DefiningEvents2020Hindsight158": { "citation_key": "Friedman2021DefiningEvents2020Hindsight158", "abstract": "", "keywords": [], "findings": [], "methods": [], "concepts": [], "sources": {}, "source": "", "abstract_provenance": {}, "summary_provenance": {} }, "Friedman2022ActiveBlockferenceCadCADActive159": { "citation_key": "Friedman2022ActiveBlockferenceCadCADActive159", "abstract": "", "keywords": [], "findings": [], "methods": [], "concepts": [], "sources": {}, "source": "", "abstract_provenance": {}, "summary_provenance": {} }, "Friedman2024FromNeuronalPacketsThoughtseeds160": { "citation_key": "Friedman2024FromNeuronalPacketsThoughtseeds160", "abstract": "", "keywords": [], "findings": [], "methods": [], "concepts": [], "sources": {}, "source": "", "abstract_provenance": {}, "summary_provenance": {} }, "Friedman2021KnowledgeManagementArchipelago161": { "citation_key": "Friedman2021KnowledgeManagementArchipelago161", "abstract": "", "keywords": [], "findings": [], "methods": [], "concepts": [], "sources": {}, "source": "", "abstract_provenance": {}, "summary_provenance": {} }, "Friedman2024CommentsSubmittedByBlockScience162": { "citation_key": "Friedman2024CommentsSubmittedByBlockScience162", "abstract": "", "keywords": [], "findings": [], "methods": [], "concepts": [], "sources": {}, "source": "", "abstract_provenance": {}, "summary_provenance": {} }, "Friedman2017OsteopathyVariableVariabilityHealth163": { "citation_key": "Friedman2017OsteopathyVariableVariabilityHealth163", "abstract": "", "keywords": [], "findings": [], "methods": [], "concepts": [], "sources": {}, "source": "", "abstract_provenance": {}, "summary_provenance": {} }, "Friedman2016CellsMechanobiologyOsteopathy164": { "citation_key": "Friedman2016CellsMechanobiologyOsteopathy164", "abstract": "", "keywords": [], "findings": [], "methods": [], "concepts": [], "sources": {}, "source": "", "abstract_provenance": {}, "summary_provenance": {} }, "Friedman2026ITraceVerificationFirstWebcam165": { "citation_key": "Friedman2026ITraceVerificationFirstWebcam165", "abstract": "iTrace is an MIT-licensed Python toolkit for webcam-derived gaze, saccade, pupil, and quality diagnostics. Version 0.4.1 is a diagnostic v1 release: the pure NumPy/SciPy core is algorithmically verified against synthetic and closed-loop oracles, the optional webcam shell exports derived records, and the empirical ledger covers a single-participant, single-device, five-session diagnostic pilot. This release does not claim reference-device accuracy, cross-device generality, population generality, or webcam-accuracy validation.", "keywords": [ "eye-tracking", "webcam", "gaze", "saccades", "pupillometry", "open-source", "diagnostic-pilot" ], "findings": [ "On synthetic traces the I-VT detector recovered a 10 deg saccade's amplitude within 5% and peak velocity within 10%.", "On the synthetic 3-D closed loop, gaze was recovered with 0.16 deg RMS residual (max 0.23 deg) for gaze within ±15 deg; the author states this is internal consistency, not device validation.", "In the idealised noise sweep, saccade detection was most fragile (F1 < 0.8 at sigma about 0.0014), while gaze crossed the 2 deg bound near sigma 0.005.", "The pupil noise robustness is reported only as a conditional illustration because it follows from the assumed pupil noise model.", "The headline limitation is an unclosed device validation gap: correctness rests on constructed ground truth, not reference measurements of real eyes." ], "methods": [ "Hardware-decoupled pure NumPy/SciPy analysis core with optional capture shell — Gaze geometry, I-VT/I-DT, Engbert-Kliegl microsaccade, main-sequence and pupil algorithms run headlessly, separate from an optional webcam/MediaPipe shell.", "Verification against synthetic traces with ground truth held out by construction — Every detector is tested on generated signals whose events are known, e.g. a 250 Hz fixation-saccade-fixation trace and an embedded 0.5 deg microsaccade.", "3-D eyeball forward model closed loop through a pinhole camera — Known gaze is projected to landmarks and back through the estimator on an animated 120 Hz scene to check geometric internal consistency.", "Seeded Monte-Carlo landmark-noise sweep (sigma 0-0.016, 25 trials/level) — Independent landmark noise is injected into the synthetic scene to rank fragility of gaze, saccade and pupil recovery, with bootstrap confidence intervals.", "N=1 single-participant, single-device webcam pilot sessions — One local webcam workflow recorded fixed-gaze, reading and target trials as derived records to give order-of-magnitude session diagnostics." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_ITrace/metadata.json", "findings": "papers/2026_ITrace/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026DemoCreateDeclarativeAudioVisual166": { "citation_key": "Friedman2026DemoCreateDeclarativeAudioVisual166", "abstract": "DemoCreate generates audio-visual demos of software — codebase tours, website walkthroughs, and terminal/CLI demos — from a single declarative, deterministic spine. A Demo is an ordered action stream plus narration chunks, merging CodeVideo's event-sourced virtual-IDE model with VSpeak's chunk/trigger model. Every heavy backend (TTS via Kokoro/Chatterbox, transcription via Whisper, capture via mss/Playwright, animation via Manim, assembly via MoviePy/ffmpeg) sits behind an abstract interface with a pure-Python deterministic default, so the package produces a real demo with only light dependencies and upgrades when extras are installed. On-screen actions are anchored to spoken trigger words via TTS→STT synchronization: narration audio is generated, transcribed back to word-level timestamps, and used to align the action stream with the narration.
", "keywords": [ "demo-generation", "screencast", "text-to-speech", "code-walkthrough", "video", "narration", "whisper", "manim", "playwright", "reproducible", "deterministic", "tts-stt-synchronization", "virtual-ide" ], "findings": [ "Across 51 source modules in 7 subsystems, 625 tests pass (3 skipped) under a ≥90% coverage gate, achieving roughly 95%.", "The recorded benchmark reports a 25.8 ms median build and 256.7 ms of render compute per second of output, with complete, monotonic sync timestamps.", "DemoCreate produced two content-verified 1080p H.264 videos: a 128.4 s self-explaining package demo and a 188.0 s demo of a 170-page paper.", "The author states DemoCreate is narrower than agentic generators and less specialized than asciinema for terminals, but uniquely combines declarative, deterministic, narrated, paper-capable demos.", "The steganographic provenance payload survives only in lossless PNG sidecars; H.264 encoding destroys it, so the MP4 carries provenance via container tags and on-screen bars." ], "methods": [ "Declarative Demo schema of typed Actions and narration Chunks (JSON/YAML) — Represents a demo as a validated value of scenes, narration chunks and typed actions that round-trips losslessly through JSON and YAML; rendering is a pure function of it.", "Abstract backend interfaces with pure-Python deterministic defaults — TTS, transcription, capture, assembly and PDF ingestion each sit behind an abstract interface whose 'auto' factory resolves to a light default; heavy backends are guarded.", "TTS-to-STT round-trip with fuzzy trigger-word anchoring — Narration is synthesized, transcribed back to word-level timestamps, and each action's trigger_word is fuzzy-matched to get an absolute millisecond timestamp.", "Poppler-CLI paper subsystem for research-paper demos — Reads a PDF with poppler utilities to extract title, abstract (skipping the TOC), captions, sections, figures and pages, and composes them into a narrated demo.", "Real-filesystem test suite, content verifier and benchmark file — Evaluates the package with a coverage-gated test suite, a verifier that checks rendered video has non-silent, non-black streams, and a reproducible benchmark file." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_DemoCreate/metadata.json", "findings": "papers/2026_DemoCreate/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026ENTOENcryptedTypedOmnitrack167": { "citation_key": "Friedman2026ENTOENcryptedTypedOmnitrack167", "abstract": "ENTO (ENcrypted, Typed, Omnitrack) is a flat ZIP container format and reference implementation for bundling heterogeneous research artifacts — time series, genomics slices, spectrograms, provenance proofs — into a single verifiable file. Each track is sealed under per-track AES-256-GCM authenticated encryption with format+track associated-data binding and PADMÉ length padding. The default wire format is 0.4.0; formats 0.2.0, 0.3.0, and 0.3.1 remain read/write compatibility profiles. Graded observability levels control how much manifest metadata a recipient sees, and an optional hash-chained proof export provides tamper-evident lineage. Verification deliberately separates key-authenticated integrity from keyless corruption detection.
This deposit is the 0.4 manuscript release candidate together with the MIT-licensed reference implementation source. Planned code home: https://github.com/docxology/entofile.
", "keywords": [ "research data formats", "authenticated encryption", "AES-256-GCM", "reproducible research", "multimodal containers" ], "findings": [ "Tamper detection succeeded on all 2400 benchmark rows (rate 1.0).", "Mean pack throughput was 78.9296 MiB/s on the medium-track condition at observability level 3 (n = 150, CV 15.3%); the paper makes no superiority claim.", "Ciphertext expansion on fixture tracks was an exact, zero-variance 1.7113.", "Length is not hidden exactly: non-sealed levels disclose byte_length in the manifest, and sealed export still reveals the PADMÉ bucket.", "Concludes that a flat ZIP container can carry typed multimodal tracks, authenticated encryption, graded observability and proof export without becoming a repository or hosted service." ], "methods": [ "Flat ZIP layout: manifest.json, one encrypted member per track, optional proof — Defines the container as a ZIP archive with a typed manifest, per-track .ento members and an optional hash-chain proof file.", "Per-track AES-256-GCM envelopes with HKDF-derived keys — Each track is sealed with authenticated encryption under a fixed nonce||tag||ciphertext header, specified in a Kaitai Struct file.", "Four graded observability levels for export-time manifest redaction — Observability levels 0–3 control which manifest metadata a recipient sees, without re-encrypting the track payloads.", "Associated-data binding and PADMÉ length padding in the default profile — The 0.4.0 default binds associated data and pads ciphertext members to PADMÉ buckets to coarsen the length channel.", "Scripted benchmark: 150 repetitions over observability levels on fixture tracks — Benchmarks on committed EEG, VCF and spectrogram fixtures plus a synthetic throughput track measure throughput, latency, expansion and tamper detection." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_ENTO/metadata.json", "findings": "papers/2026_ENTO/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026GeneralizedNotationNotationGNN168": { "citation_key": "Friedman2026GeneralizedNotationNotationGNN168", "abstract": "Active Inference offers a unifying account of perception, learning, and action under the free energy principle, yet the generative models at its core are still communicated ad hoc: scattered across prose descriptions, bespoke notebooks, and framework-specific code that rarely agree. This fragmentation makes published models hard to reproduce, compare, or port between tools, and it raises the barrier for newcomers learning the formalism. GeneralizedNotationNotation (GNN) addresses this gap with a standardized, human- and machine-readable text language for specifying Active Inference generative models, paired with a 25-step processing pipeline that transforms a single specification into validation, visualization, simulation, and analysis artifacts. A GNN file denotes a generative model, and the language lets that model take the several shapes the field actually uses: the notation blocks a file declares — categorical A/B/C/D[/E] tensors, linear-Gaussian system matrices F/H/Q/R with a Gaussian prior, per-agent and per-level compositions — fix its model kind, which the pipeline classifies structurally and carries through rendering and execution, recording explicitly which registered backends can render and run a model of each kind and which report it unsupported. GNN's \"Triple Play\" treats each model as three coordinated views — a textual specification, graphical visualizations, and executable cognitive models — so that one source yields consistent outputs across modalities. The framework spans 32 exemplar specifications (27 discrete-state and 5 continuous linear-Gaussian) across 9 model families and 10 registered rendering backends, with explicit gates for cross-format semantic fidelity and cross-framework reliability. GNN 3.0.0 layered safe-by-design long-running orchestration on top of this language and pipeline — durable observation streams, resumable run sessions, and auditable container plans that generate, validate, and replay data only, with no live infrastructure mutation. We describe the language, the model kinds it denotes, the pipeline architecture, and the validation that confirms specifications round-trip faithfully across formats and execute across multiple simulation backends — with coverage gaps recorded as explicit profiled-unsupported statuses rather than silently omitted — establishing GNN as reproducible, interoperable infrastructure for communicating Active Inference models.", "keywords": [ "active inference", "generative models", "cognitive modeling", "notation system", "reproducibility", "computational neuroscience", "bayesian inference", "standards", "gnn", "python" ], "findings": [ "The manuscript presents the Triple Play as text, graphical, and executable views derived from a shared model specification.", "At the manuscript snapshot, the framework covers 32 exemplar specifications across nine model families and ten registered rendering backends.", "The manuscript reports profiled execution gaps for continuous and hierarchical models, and a deliberate render-only scope for structural models; it does not claim that every model executes on every backend.", "The manuscript specifies reproducible validation commands and explicitly avoids asserting a fixed passing-check count.", "The manuscript describes long-running orchestration contracts that generate, validate, and replay data without mutating live infrastructure." ], "methods": [ "Parsing and structured export — Parses plain-text model specifications into an internal representation and structured export formats, retaining declared model vocabulary.", "Model-kind type checking and validation — Checks state spaces, observation modalities, control factors, matrix dimensions, and kind-specific shape contracts before code generation.", "Kind-aware model rendering — Carries structurally determined model kinds into rendering and execution, recording which backends can handle each kind and which report it unsupported.", "Semantic fidelity and cross-framework gates — Provides reproducible commands for testing semantic preservation in round trips and comparing generated model structure across backend implementations." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_GeneralizedNotationNotationGNN/metadata.json", "findings": "papers/2026_GeneralizedNotationNotationGNN/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": { "source": "Rendered manuscript Abstract in exact Zenodo v3.6.0 archive" }, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01" } }, "Friedman2026COGANTDeterministicCodebaseGNN169": { "citation_key": "Friedman2026COGANTDeterministicCodebaseGNN169", "abstract": "COGANT (Codebase-to-GNN Translation) deterministically converts software repositories into structured Active Inference artifacts expressed in the Active Inference Institute's Generalized Notation Notation (GNN). It is an evidence compiler: it propagates reviewable program facts through a finite fixpoint rule pipeline and emits graph, matrix, provenance, visualization, and round-trip artifacts with confidence and provenance, rather than a single opaque embedding. A reverse synthesizer reconstructs a runnable Python package from an emitted GNN bundle, closing a forward-reverse-forward evaluation loop. This is the first public release (v0.6.0). Source: https://github.com/ActiveInferenceInstitute/COGANT", "keywords": [ "program analysis", "Generalized Notation Notation", "GNN", "intermediate representation", "code property graph", "active inference", "reproducible research", "codebase-to-model translation", "cognitive ecosystem modeling" ], "findings": [ "On the v0.6.0 roundtrip ledger, all 25 targets are role-preserved, but only 1 of 25 meets strict structural isomorphism.", "The author cautions that fixtures are in-sample, with no held-out split or confidence intervals, so scores upper-bound rather than estimate out-of-sample performance.", "The fixpoint ablation shows a single pass suffices on every shipped fixture, with the K=10 cap serving as a safety valve.", "Rule-family ablation indicates structural rules drive HIDDEN_STATE while semantic rules drive OBSERVATION/ACTION/POLICY/PREFERENCE roles.", "The author states the passing test suite, coverage and type-check gates support reliability and reproducibility but do not by themselves establish semantic adequacy." ], "methods": [ "Program-graph IR with confidence and provenance on nodes and edges — Repositories are parsed (primarily via Python's standard-library ast) into a program graph IR whose nodes and edges carry confidence and provenance.", "Fixpoint translation engine with 22 declarative rules in five families — Rules (structural, semantic, control, behavioural, resilience) map nodes to 7 Active Inference mapping kinds, with priority-and-score conflict resolution.", "A/B/C/D matrix derivation and GNN (Generalized Notation Notation) export — Derives likelihood, transition, preference and prior matrices from the compiled state space and program-graph edges, normalized for the upstream GNN validator.", "Forward-reverse-forward roundtrip over a 25-target regression corpus — A reverse synthesizer rebuilds a Python package from each GNN bundle; role_preservation_score and strict isomorphism are recorded per target.", "Rule-family and fixpoint-iteration ablations on packaged fixtures — Removes each rule family and varies the iteration cap K in {1, 2, 5, 10}, recording changes in SemanticMapping counts on the shipped fixtures." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_COGANT/metadata.json", "findings": "papers/2026_COGANT/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026AGEINTAgenticIntelligence170": { "citation_key": "Friedman2026AGEINTAgenticIntelligence170", "abstract": "Synthetic Analytic Tradecraft (AGEINT, or Agentic Intelligence), is a local curriculum-and-assurance atlas for teaching bounded AI-agent support inside intelligence education by making the machinery of Synthetic Analytic Tradecraft visible on the page. It converts SIST Guide TOC and Bibliography into 16 parts, 51 modules, 9 methods appendices, 20 named AGEINT patterns, and 312 parsed source-guide references without renumbering inherited source identities, then asks the reader to inspect the same things an instructor or assurance reviewer would inspect: the part map that frames the domain, the module overview that names the learner task, the source spine that separates official, standards, scholarly, public-domain, practitioner, vendor, and provenance-only support while separating source quality from narrative fluency, the evidence packet that a student would retain, the reviewer challenge that should break a weak claim, and the validators that prevent a polished artifact from masquerading as a verified one. A learner does not enter through a slogan. The learner opens a module, sees which authority and source keys are allowed to carry the claim, reads a textbook primer that distinguishes observation from inference, confidence from probability, and source quality from fluency, works through a synthetic practice studio using classroom records, public declassified examples, owned-lab logs, toy datasets, rendered figures, or tabletop incidents, and then produces a bounded artifact with caveats, assumptions, alternatives, excluded actions, human review, rollback evidence, and refresh triggers written into the record before reuse. The method contract in shows that flow as a claim-to-evidence pathway rather than as an aspirational ethics statement: authority must be named before the task begins, source support must be traceable before a claim is promoted, unsafe action must be replaced by safe substitution, and agentic assistance must remain a drafting, retrieval, comparison, simulation, critique, or audit aid whose output is never treated as self-authenticating . AGEINT remains synthetic in its fixtures, not in its standards: the word synthetic is therefore a control, not a downgrade. AGEINT keeps the fixture safe while keeping the standard difficult: HUMINT, SIGINT, OSINT, GEOINT and IMINT, FININT, counterintelligence, cyber threat intelligence, cognitive security, agent orchestration, active inference, source verification, public-sector governance, privacy review, and industrial-control-system defense can be discussed because the student is not handed live targets, evasion recipes, exploit instructions, manipulation playbooks, covert-action procedures, or unsafe cyber-physical steps. When a source topic could invite misuse, the module turns the risky motif into a provenance card, detection-coverage note, rights-impact worksheet, model or dataset card, source-refresh memo, release gate, risk-exception log, remediation backlog, or debrief protocol, and the exercise fails if it cannot show where authority, data boundary, tool permission, uncertainty, and review enter the workflow. The source layer is intentionally conservative: Perplexity and similar discovery tools may suggest candidates, but the manuscript cites verified official, standards, public-domain, or scholarly anchors encoded in the source corpus and rebuilt into BibTeX; Practitioner, vendor, and blog sources inherited through the source guide, plus social or other guide-inherited rows, can preserve provenance context without becoming foundational support for governance, rights, safety, empirical, statistical, or performance claims unless a stronger verified source bears the specific point. The same source posture is visible in the counts: 10 source-quality anchors cover the governing standards and assurance spine, while 462 curated intelligence research anchors span the domain lanes that route claim-bearing prose to direct evidence rather than decorative citations. A strong AGEINT artifact is therefore not an uninspected essay. It is an evidence packet that names the question, allowed inputs, excluded actions, source keys, claim type, caveats, competing explanations, confidence basis, prompt or run card, tool allowlist, data boundary, stop condition, reviewer challenge, safety boundary, refresh trigger, and human disposition. It also contains negative controls: stale citations that should be caught, weak-source-only claims that should be downgraded, figure positions or colors that should not be mistaken for quantitative evidence, overbroad statistical language that should fail, and operational substitutions that should halt until an instructor or reviewer approves a safer artifact. A reviewer can trace a finished packet in both directions: from a polished sentence back to the claim class, source family, cited anchor, caveat, and source-refresh duty, or from a source row forward to the modules, figures, tables, and exercises that rely on it. A student can see why a governance claim needs law, policy, standard, or official guidance support; why a cyber or industrial-control-system scenario must remain defensive and synthetic; why a social or cognitive-security lesson must become resilience education rather than persuasion practice; why a vendor or practitioner note can motivate a question but cannot by itself carry a public-rights or safety assertion; why a figure caption must say whether arrows are explanatory or quantitative; and why a model-generated draft is only useful after it has been constrained by authority, reviewed by a person, and attached to evidence that another person can contest. The early orientation pages, part maps, chapter landmarks, appendices, bibliography atlas, source-lane map, and generated reports are designed to make that trace visible instead of leaving it to instructor memory. They show the learner what to keep: the source keys behind each claim, the reason a safer substitute was chosen, the assumption that would change the answer, the dissent that should not be smoothed away, the review note that records who accepted residual risk, the rollback path if a source or permission changes, and the refresh trigger that turns a current claim back into a review item. They also show what the curriculum refuses to keep: unverifiable authority theater, live-target instructions, metric-looking diagrams without units or denominators, statistical decorations without empirical design, citation clusters that do not bear the sentence they decorate, and agent outputs whose provenance cannot be reconstructed. The technical build is part of that scholarship. Curriculum shards generate semantic manuscript files; cross-references use label-backed section and figure links, and display equations are typeset from source for reproducible math; citations route through Pandoc keys; figures are registry-backed PNG assets with captions, alt text, long descriptions, provenance, hashes, and visual-semantics metadata; source anchors carry explicit lane and tier fields; claim calibration audits high-risk empirical, statistical, governance, safety, visualization, artifact-readiness, and formalism language; scholarship and source-metadata reports expose review warnings instead of hiding them; PDF audits check stale output, URI targets, file actions, and banned filler language. The analysis-validation matrix in names which claim classes require empirical evidence, direct source-family support, negative controls, figure semantics, or rendered-artifact checks before they can be treated as ready, and that matrix is a boundary on what the manuscript can honestly say. AGEINT is not a benchmark and does not claim to measure AGEINT performance, model capability, analyst replacement, learning outcomes, operational effectiveness, public-sector impact, statistical significance, or safety performance; page counts, citation counts, figure counts, validator passes, and link audits are artifact telemetry, not empirical outcome evidence. Its strongest claim is methodological and inspectable: agentic assistance can be taught inside intelligence education when every reuse path is forced through authority, source support, safe substitution, evidence packet, negative controls, human review, rollback, and refresh triggers, and when the safest failure mode is to stop, document the unresolved condition, and ask for review. The resulting work should be read as a Synthetic Analytic Tradecraft atlas, not as a public-release certificate or a disguised operational manual: the reader can see the source key behind a sentence, the caveat behind a confidence statement, the assumption behind a scenario, the blocked use behind a tempting automation, the challenge behind a polished answer, the caption that limits a visual, the validator that rejects an overclaim, and the refresh duty that keeps current-source prose from becoming stale authority.
", "keywords": [ "agentic intelligence", "AGEINT", "AI agents", "intelligence tradecraft", "cognitive security", "structured analytic techniques", "active inference", "model context protocol", "multi-agent systems", "operational governance" ], "findings": [ "AGEINT presents itself as a versioned curriculum-and-assurance framework, explicitly not a new agent architecture, cognitive theory, attack benchmark, or measured learning outcome.", "Its strongest stated claim is methodological: agentic assistance can be taught in intelligence education when reuse is forced through authority, source support, safe substitution, review, rollback, and refresh triggers.", "The source corpus includes 10 source-quality anchors for the standards and assurance spine and 462 curated intelligence research anchors across domain lanes.", "The work states that its page, citation, and figure counts and validator passes are artifact telemetry, not empirical outcome evidence.", "Exercises are restricted to educational, defensive, synthetic material and give no instructions for unauthorized collection, evasion, exploitation, or manipulation." ], "methods": [ "Conversion of the SIST Guide TOC and bibliography into a modular curriculum — Restructures an inherited source guide into parts, modules, methods appendices, named patterns, and parsed references while preserving source identities.", "Synthetic practice studios using safe fixtures instead of live targets — Exercises use classroom records, declassified examples, owned-lab logs, toy datasets, and tabletop incidents, ending in a bounded evidence-packet artifact.", "Source-lane and tier system separating official, standards, scholarly, and vendor sources — Each claim class is matched to a required evidence type; discovery tools may suggest sources but citations must resolve to checked anchors.", "Automated claim-calibration audit of generated manuscript language — A verifier scans high-risk empirical, statistical, governance, and safety language and fails unsupported proof-language, p-values, or measured-performance claims.", "Generated manuscript from version-controlled Markdown, tested Python, and registry figures — The textbook is built reproducibly from curriculum shards, Pandoc citation keys, and registry-backed figures with captions, alt text, and provenance." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_AGEINT/metadata.json", "findings": "papers/2026_AGEINT/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "book/report" } }, "Friedman2026PolicyDistillationAsActive171": { "citation_key": "Friedman2026PolicyDistillationAsActive171", "abstract": "This paper formulates on-policy distillation as active inference in finite variational models, with exact claims only for declared objects and interpretive claims explicitly bounded outside them. In the construction, the intractable teacher policy plays the role of the generative model $p(o,s)$, the tractable student policy is the approximate posterior $q(s)$, and the per-token reverse-KL distillation loss is variational free energy up to the evidence constant, $F = D_{\\mathrm{KL}}(q\\,\\|\\,p(s\\mid o)) - \\log p(o)$, whose KL target is the teacher-induced posterior $p(s\\mid o)\\propto p(o,s)$ . The title's \"as\" is therefore a scoped mathematical correspondence rather than the slogan OPD = Active Inference. Variational free energy names the realized-rollout distillation loss; expected free energy remains the planning-side objective by which the pymdp agent selects actions . On-policy student rollouts generate the observations on which the posterior is scored, connecting the construction to induced-distribution mismatch in imitation learning and exposure-bias analyses while preserving their different objectives, empirical regimes, and contested severity . Privileged traces and feedback play the role that train-time-only information plays in the LUPI/distillation lineage . Four deterministic witnesses instantiate the correspondence. A Bernoulli-Ising oracle couples a teacher's privileged variable to the answer through $\\lambda$, making $I(\\lambda)$ the teacher-student mutual information and the finite free-energy gap the toy distillation objective; the closed-form and independently recomputed mutual-information sweeps agree to machine precision (RMSE 2.1e-16 nats). A pymdp T-maze rollout supplies the on-policy student that samples its own observations under a privileged cue . A two-agent classroom pits a privileged teacher (cue validity 0.98) against an on-policy student (cue validity 0.5), measuring teacher belief entropy 0.247 nats versus student 0.347 nats and a mean reverse-KL distillation signal of 6.28 nats. A four-state/two-action sequential-shift witness shows teacher-forced train loss 0.333 nats underestimating student-induced test loss 0.409 nats, with deterministic on-policy correction reducing it to 0.096 nats. These are toy, generated findings, not production-LLM measurements. Recent privileged-context, context-distillation, adaptive-teacher, freshness-aware OPD, RLHF/instruction-tuning, self-generated reasoning, Qwen OPD-vs-RL, and Thinking Machines replication reports remain external context rather than reproduced results . The supplemental sheaf/provenance layer keeps that boundary operational: every reported number is hydrated from a generated artifact, every figure is source-bound, and 16 / 16 invariant checks pass before rendering. --- Associated artifacts GitHub release: v1.0.2 (https://github.com/ActiveInferenceInstitute/on_policy_distillation/releases/tag/v1.0.2) DOI: https://doi.org/10.5281/zenodo.20749817 Zenodo: https://zenodo.org/records/20749817 PDF SHA-256: c6b5ec494915e6e046f24cf723f8dbbf93a5b168544daed3cca14c089d4087aa
", "keywords": [ "on-policy distillation", "active inference", "self-distillation", "privileged information", "free energy principle", "reverse KL divergence", "pymdp", "sophisticated inference" ], "findings": [ "The closed-form and independently recomputed mutual-information sweeps agree to machine precision (RMSE 2.1e-16 nats).", "In the classroom toy, teacher belief entropy was 0.247 nats versus 0.347 nats for the student, with a mean reverse-KL distillation signal of 6.28 nats.", "In a four-state/two-action witness, teacher-forced train loss (0.333 nats) underestimated student-induced test loss (0.409 nats); on-policy correction reduced it to 0.096 nats.", "All reported numbers are hydrated from generated artifacts, and 16 of 16 invariant checks pass before rendering.", "The author states these are toy, generated findings rather than production-LLM measurements; external OPD results are context, not reproduced." ], "methods": [ "Formal mapping of OPD roles onto active-inference variational objects — Teacher policy is read as the generative model, student policy as the approximate posterior, and per-token reverse-KL loss as variational free energy.", "Bernoulli-Ising oracle with closed-form and recomputed mutual-information sweeps — A binary toy couples a teacher's privileged variable to the answer through a coupling parameter; MI and the free-energy gap are computed analytically.", "pymdp T-maze rollout with sophisticated-inference planning — A pymdp agent samples its own observations under a privileged cue, serving as the on-policy student process witness.", "Two-agent classroom: privileged teacher vs on-policy student — A teacher with cue validity 0.98 and a student with cue validity 0.5 are compared on belief entropy and reverse-KL distillation signal.", "Lean theorem inventory and fail-closed manuscript validation gates — Lean theorem statements are extracted and checked against an inventory, with gates failing on sorry, axiom or native_decide." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_PolicyDistillationAs/metadata.json", "findings": "papers/2026_PolicyDistillationAs/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "theoretical" } }, "Friedman2026TemplateMadlibDeterministicToken172": { "citation_key": "Friedman2026TemplateMadlibDeterministicToken172", "abstract": "This exemplar asks whether a reviewable pipeline can hydrate a complete IMRAD manuscript from configuration-owned lexical data while preserving an audit trail that remains readable before and after rendering. The project deliberately keeps playful Mad Lib mechanics inside a serious reproducibility contract: the manuscript shell names large placeholders, the config declares allowable language, and the source code decides what text is emitted.\n\nThe committed seed is 431 and the current schema expands 22 slot rule(s) into 40 token choice(s) across 10 lexicon categories. The configured narrative moves are state the manuscript-generation problem, name the deterministic intervention, and summarize the audit surface. The central hypothesis is: Deterministic lexical injection can generate a complete conditional IMRAD manuscript while preserving token provenance, section intent, and audit-ready method evidence.\n\nThe result is not a claim that lexical substitution creates scholarship. It is a worked template for conditional manuscript assembly: section enablement, token provenance, figure registration, and unresolved-placeholder checks all become inspectable artifacts before the shared renderer produces PDF, HTML, and slides.\n\n---\nAssociated artifacts\nGitHub release: v0.1.1 (https://github.com/docxology/template_madlib/releases/tag/v0.1.1)\nDOI: https://doi.org/10.5281/zenodo.20786638\nZenodo: https://zenodo.org/records/20786638\nPDF SHA-256: d9248f4f372fc3baf21cbf5cc5cdb7daffe0e22e62ac6fa2e3a697a26f3308b6", "keywords": [ "madlib generation", "token injection", "conditional manuscripts", "reproducible research", "IMRAD" ], "findings": [ "With seed 431, the schema expands 22 slot rules into 40 token choices across 10 lexicon categories.", "The generated plan enabled all 11 manuscript sections and filled 40 token choices, each traced to its variable, category, section and config pointer.", "Re-running generation with seed 431 and the same lexicon produces the same token plan.", "The author frames the main result as traceability surviving a complete render path, not any particular word choice.", "The paper explicitly does not claim lexical replacement creates scholarship; it shows a conditional text generator can be made accountable to config, tests and validation." ], "methods": [ "Seeded SHA-256 digest selection of lexicon tokens per slot — Each slot hashes the seed, slot name, category, ordinal and full category list; the digest indexes the configured lexicon category.", "YAML config-owned lexicon, slots, and section conditions — Lexicon categories, slots, section switches, method rows and claims are declared in YAML; source code turns them into manuscript bodies.", "Staged pipeline from config validation to hydrated Markdown — Validates the madlib block, builds a TokenPlan and section bodies, writes artifact JSON and a figure registry, then hydrates Markdown.", "Explicit vs. loader-default config field-origin inventory — Classifies configuration paths as explicitly set in YAML or inherited from loader defaults and reports them as method evidence.", "Project tests and shared output validator — Tests check determinism, seed and category sensitivity, malformed configs, section disablement and unresolved tokens; a validator checks rendered outputs." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_TemplateMadlib/metadata.json", "findings": "papers/2026_TemplateMadlib/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026CaliforniaPublicRecordsTechnical173": { "citation_key": "Friedman2026CaliforniaPublicRecordsTechnical173", "abstract": "A technical and legal reference to California public-records ecosystem, anchored by the CPRA recodified by AB 473.", "keywords": [ "California Public Records Act", "open data", "CKAN", "Socrata", "ArcGIS", "cognitive security", "civic technology" ], "findings": [ "The reference compiles California's public-records ecosystem into a machine-readable artifact: 22 CPRA statute sections, 17 portals, 7 exemption clusters and an 8-dataset OpenJustice taxonomy.", "It notes that the AB 473 recodification made no substantive changes to disclosure rights but split the exemption list into independent code sections for readability.", "Top-line verdict is 'CERTIFY-WITH-RESIDUALS': four CPRA sections from the upstream research document were misattributed or unverifiable and deliberately omitted.", "A scaffolded citation-laundering example was rejected: § 7928.200 does not govern peace-officer records, so that mandate is pinned to Penal Code § 832.7.", "The package's generator methods are pure functions of the registries, so re-running them yields byte-identical output apart from a single provenance timestamp." ], "methods": [ "Registry-first compilation of CPRA statutes, portals, exemptions and datasets — Held every cited section number, portal URL, schema field and dataset in Python registries under src/, with prose counts resolved from them at build time.", "Standard-library Python clients for California API surfaces — Wrote working clients for CKAN Action API, Socrata SoQL, ArcGIS GeoServices REST, CIMIS REST, OpenJustice CSV and LegiScan endpoints.", "Metadata-schema validators (DCAT, CKAN, Dublin Core, RIPA stop data) — Implemented schema validators that accept a mapping and return a report of missing fields and warnings.", "Offline no-mocks test suite with optional live endpoint checks — Validated registries, clients and schemas against a real local HTTP server, plus optional live checks against California open-data endpoints.", "Cross-vendor citation verification (leginfo, Justia, FindLaw) — Source-verified statute citations against leginfo and re-confirmed selected ones via Justia and FindLaw, removing those that could not be verified." ], "concepts": [], "sources": { "abstract": "papers/2026_CaliforniaPublicRecords/README.md", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_CaliforniaPublicRecords/metadata.json", "findings": "papers/2026_CaliforniaPublicRecords/metadata.json" }, "source": "papers/2026_CaliforniaPublicRecords/README.md", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "book/report" } }, "Friedman2026CogSecSkillsMultiharnessAgenticSkills174": { "citation_key": "Friedman2026CogSecSkillsMultiharnessAgenticSkills174", "abstract": "CogSecSkills is a defensive, harness-neutral agent-interface library that turns the human doctrine of cognitive security and analytic tradecraft into dependable, inspectable, agent-usable skills, distributed as an open repository from github.com/docxology/CogSecSkills; the motivation is an information environment in which mis-, dis-, and malinformation are analytically distinct but operationally entangled, false content can diffuse rapidly at platform scale, and credible source evaluation increasingly demands explicit expert practice rather than page-bound reading alone, so that agents asked to weigh competing hypotheses, trace provenance, or critically review a claim need a repeatable procedure and a stable tool-use contract rather than improvisation. The live generated catalogue reports one hundred implemented skills across seven taxonomy groups — Structured Analytic Techniques, Cognitive Security, Critical Review and Assurance, OSINT and Source Integrity, Counterintelligence and Deception Detection, Information Environment and Influence Analysis, and Research and Synthesis Methods — and the library is organized as a Plan, Build, and Teach system in which a registry declares the catalogue, a definitions layer owns the substance and quality controls of each skill, a skills tree exposes the harness-facing build, and a vendored educational upstream named AGEINT explains why each technique exists and how to use it responsibly. The central design choice is to make the reusable skill contract smaller than any one agent interface yet stricter than a prompt collection: each skill is owned by a single canonical definition that declares triggers, inputs, outputs, per-skill reference metadata, group-aware quality controls, and a closed vocabulary of neutral tool verbs before any harness-specific adapter is considered, and that definition is rendered deterministically into a harness-neutral specification, a human-readable skill description, an executable workflow, and one adapter per configured harness whose default members are Claude, Codex, and Hermes, so that portability becomes a property the test suite proves rather than a hope, and installation is concrete rather than interpretive — clone the public repository, install or run the Python package, run validation, point the agent harness at a chosen skill, execute its workflow, and bind runtime tools through the named harness adapter, regenerating adapters from a configuration file for any non-default harness. The quality discipline is the core of the contribution: an automated audit checks that every skill's defensive boundaries, misuse redirects, evidence requirements, confidence rubrics, privacy and legal constraints, uncertainty handling, failure modes, and negative controls are not merely present but specific to that skill and not reused across the corpus, while an evidence ladder adds curated safe-use and unsafe-redirect scenarios with expected defensive response-shape contracts, reviewed expected answers, and one source-owned worked example per skill, and a generated quality dashboard, supplemental catalogue, metadata matrix, data exports, and a family of deterministic figures — taxonomy counts, the hundred-skill atlas, tool-verb coverage, AGEINT topic crosswalks, the Plan-Build-Teach flow, reference density, harness coverage, and a cover-page installation route — are all produced directly from the live registry and skill specifications so that every visual stays synchronized with the source tree. The evidence boundary is deliberately and explicitly repository-local and reproducible within the checked-out project state, following reproducible-computing, open-data stewardship, and software-citation norms for explicit workflows, version specificity, and citable artifacts; every claim is backed by source files, canonical definitions, generated supplements and figures, the generated dashboard, and project-local verification commands together with a focused test suite and a manuscript renderer, and the work positions CogSecSkills as a validated interface between reasoning, tool use, and defensive output discipline rather than as a claim that any particular model runtime behaves correctly in the field, so its figures, supplements, scenarios, worked examples, and dashboard should be read as synchronized views of the current library state and never as independent measurements of operational performance.\n\n---\nAssociated artifacts\nGitHub release: v1.0.0 (https://github.com/docxology/CogSecSkills/releases/tag/v1.0.0)\nDOI: https://doi.org/10.5281/zenodo.20804585\nZenodo: https://zenodo.org/records/20804585\nPDF SHA-256: 1a99a2e474b07d21593094b7a3b4fd923197904056189a53b206126506b1040d", "keywords": [ "cognitive security", "agent skills", "analytic tradecraft", "structured analytic techniques", "multiharness" ], "findings": [ "The generated catalogue reports one hundred implemented skills across seven taxonomy groups, from Structured Analytic Techniques to Research and Synthesis Methods.", "The author describes a local quality gate, explicitly 'not a safety proof', that can show all 100 skills include group-aware defensive boundaries and skill-specific governance entries in the current repository state.", "The worked-example gate verifies exact 100-skill coverage; the examples are expected-answer shapes, not live model transcripts.", "The author states the evidence is local conformance only and does not establish operational effectiveness, adversary coverage, user outcomes, or superiority over other libraries.", "The deterministic gates do not show that a live runtime will select the same skill, use tools correctly, or answer well in the field." ], "methods": [ "Plan/Build/Teach architecture: registry, definitions, skills tree, AGEINT — Separates a catalogue registry, canonical per-skill definitions, a harness-facing skills tree, and the AGEINT educational upstream.", "Harness-neutral skill contract over a closed set of eight tool verbs — Each skill may use only neutral verbs (read, search, write, exec, reason, web, delegate, ask), bound per harness through adapters.", "Deterministic rendering into Claude, Codex, and Hermes adapters — Renders each canonical definition into a spec, skill description, workflow, and one adapter per configured harness.", "Automated anti-boilerplate per-skill quality audit — Requires a skill-specific quality bundle and rejects generic or reused negative controls, rubrics, evidence and privacy entries.", "Evidence ladder: 28 scenario fixtures and per-skill worked examples — Checks curated safe and unsafe scenarios against the local router and verifies one reviewed worked example per skill." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_CogSecSkills/metadata.json", "findings": "papers/2026_CogSecSkills/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026RealizingEmptinessOperationalSurrogates175": { "citation_key": "Friedman2026RealizingEmptinessOperationalSurrogates175", "abstract": "There is no statute, treatise, or law-school casebook titled \"Entomological Law.\" The phrase names a synthetic field — the convergence zone where the six-legged world repeatedly forces the legal system to answer questions it was not designed for: Can a fly testify? Who owns a swarm? Is a bumblebee a fish? Can you patent a mosquito? May a court excommunicate a weevil? Does a cricket suffer? May insects be used in war? This reference compiles that field as a machine-readable and reproducible artifact organized around the legal role an insect occupies in a given dispute. It encodes 8 legal roles — witness, regulated threat, protected subject, property, invention, defendant, moral patient, and weapon — and binds to them 18 landmark decisions, 43 statutes and treaties across 9 categories and 10 jurisdictions, 24 insect taxa, 13 certifying and regulatory institutions, 44 historical milestones spanning 3676 years, and 5 cross-domain themes that knit the roles together. Every count in this prose is generated from the source registries under src/, legal propositions are source-bound in the bibliography, every externally-sourced statistic written as a numeral is bound to a verification record in the claim ledger, and every figure caption is emitted from a source-owned caption registry. The result is both a map of a genuinely transdisciplinary field and a reproducibility contract: the same version-controlled inputs regenerate the inventories, validation report, analytical figures, manuscript variables, and paper while preserving explicit caveats about registry scope, jurisdictional reach, and the boundary between what the offline gates prove and what only a live source check can confirm. --- Associated artifacts GitHub release: Entomological Law (v1.0.0) (https://github.com/docxology/EntoLaw/releases/tag/v1.0.0) DOI: https://doi.org/10.5281/zenodo.21137276 Zenodo: https://zenodo.org/records/21137276 PDF SHA-256: 5a1891704c285491c97297dea66e9a4ea019b2dabddcc28cb5aeb18d6a335225
", "keywords": [ "entomological law", "legal entomology", "forensic entomology", "endangered species", "invasive species", "insect welfare", "gene drive", "biological weapons convention", "reproducible legal scholarship" ], "findings": [ "Alongside its roles, cases and statutes, the registries encode 24 insect taxa, 13 institutions, 44 historical milestones spanning 3676 years, and 5 cross-domain themes.", "The roles are linked by recurring themes; the role-coverage matrix shows some roles are case-driven, some statute-driven, and defendant and weapon roles almost entirely history-driven.", "The author argues the common frontier is institutional translation: insect facts enter law only through admissibility, quarantine, listing, authorization and moral-status filters.", "Identified frontiers include insect sentience in welfare law, thin deployment-specific gene-drive governance, and forensic-entomology standardization.", "The stated contribution is a registry-first, claim-sourced substrate rather than a new doctrine." ], "methods": [ "Organizing the field by eight legal roles an insect occupies — Structures the synthesis around the insect's legal status in a dispute: witness, regulated threat, protected subject, property, invention, defendant, moral patient, weapon.", "Seven Python source-of-truth registries under src/ — All domain content (roles, cases, statutes, species, institutions, timeline, interconnections) lives in registries from which prose, figures and reports are regenerated.", "Build-time token closure for every count in the prose — Magnitude-bearing numbers are generated tokens; a closure test fails the build if a hand-typed count drifts from its registry.", "Claim ledger with offline and live verification oracles — External statistics and status claims are logged with source URL, verbatim quote and as-of date; an offline oracle checks attribution and a live oracle re-fetches sources." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_EntomologicalLaw/metadata.json", "findings": "papers/2026_EntomologicalLaw/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "review" } }, "Friedman2026ShapeBetweenFullPage184": { "citation_key": "Friedman2026ShapeBetweenFullPage184", "abstract": "template_storybook demonstrates a public, standalone picture-book workflow in\nthe research template repository. The bundled project renders a deterministic\nfourteen-page storybook subtitled A geometric fable of belonging, bracing, and\nreciprocal form: a clear cover, a publication-and-acknowledgements page, and\ntwelve story pages in which a child tetrahedron raised by cubes meets a child\ncube raised by tetrahedra. The story uses a large reciprocal symbol, a\ntetrahedron-inside-cube stability spread, a shadow-projection lesson, a\ntensegrity lantern, and a vector garden to frame belonging without sameness:\nsquare and triangle families remain distinct while the space between them\nbecomes navigable.\n\nThe project separates story data, rendering logic, and orchestration. Story\ntext, characters, page order, palettes, and overlay choices live in\ncontent/story.yaml; character generation and illustration live in\nsrc/storybook/; one script renders the cover, each page script renders a\nnumbered page, and a final script assembles the PDF. This makes the exemplar a\nforkable pattern for full-page creative artifacts rather than standard\nmanuscript-centered figure generation.\n\n---\nAssociated artifacts\nGitHub release: The Shape Between v0.1.0 (https://github.com/docxology/template_storybook/releases/tag/v0.1.0)\nDOI: https://doi.org/10.5281/zenodo.21176001\nZenodo: https://zenodo.org/records/21176001\nPDF SHA-256: 7fb0ed775db212cd9f4b5b578bef8900c9f925bc0c5cc047331eb6900679eeda", "keywords": [ "storybook", "illustration", "procedural graphics", "reportlab", "reproducible publishing" ], "findings": [], "methods": [ "Storybook PDF generated from a public research-template repository — The illustrated storybook is presented as a template exemplar produced by a research-template repository, with story and images labeled as illustrative fiction.", "Geometric fable drawing on Fuller's Synergetics (tetrahedron in cube) — The narrative uses geometric motifs (a tetrahedron bracing a cube, tensegrity lantern, vector garden) acknowledged as drawn from Fuller's Synergetics." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_ShapeBetween/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": { "source": "papers/paper_metadata.json", "date": "2026-10-01" }, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "other" } }, "Friedman2026AutopoieticProjectGeneration185": { "citation_key": "Friedman2026AutopoieticProjectGeneration185", "abstract": "template_autopoiesis is a combinatoric grammar that deterministically generates\nwhole runnable projects — not files or snippets, but complete, independently\ntestable child repositories with their own kernel source, tests, analysis\nentry point, and manuscript. A single integer seed plus a grammar of orthogonal\nslots (primitive domain, analytical track, section set, and three\npresentation/provenance slots) selects one child from a combinatoric product\nspace of 360 nominal (45 content-distinct)\nconfigurations, via a SHA-256 digest of the seed and slot identity — with no\nrandom-number generator anywhere in the expansion path.\n\nProject generators routinely claim completeness, determinism, and\ntraceability without making any of the three independently checkable. This\nexemplar treats each claim as a structural property to verify rather than a\nrhetorical one to assert: verify_child() recomputes a tree hash from the\nfiles actually on disk and compares it against the recorded provenance,\nrather than trusting a value the same run wrote down; a honesty manifest\ninspects the live source AST to confirm every claim in this manuscript\nresolves to a real function in a real file; and a per-domain mutation gate\nchecks that the acceptance tests reject a constant-success stub before\ntrusting that they accept the real kernel. The same discipline governs this\ndocument itself — every number below is substituted at render time from a\nlive measurement rather than hand-typed as a literal.\n\nAcross 5 heterogeneous primitive domains\n(- optimization\n- dynamics\n- statistics\n- signal\n- graph), 493 tests exercise both fixed ground-truth\nchecks and Hypothesis-driven property invariants at 96.28% branch\ncoverage, with an explicit negative control per domain distinguishing the\nreal kernel from a deliberately-wrong one.\n\nGeneration pipeline\n\nGrammar product space\n\n- Domain count: 5\n- Effective product size: 45\n- Total product size: 360\n- Reserved slots: 3 (figure_profile, qr_profile, integrity_profile)\n- Grammar hash: f84a8f9dbcb18e37\n- Tests: 493 · Coverage: 96.28%\n\n---\nAssociated artifacts\nGitHub release: Autopoietic Project Generation (v1.0.1) — improved abstract (https://github.com/docxology/template/releases/tag/v1.0.1)\nDOI: https://doi.org/10.5281/zenodo.21227869\nZenodo: https://zenodo.org/records/21227869\nPDF SHA-256: 84145371f10c7ee97a75c53d78917d247fd6bd987d2c1764ee437b6fbdef8f51", "keywords": [ "autopoiesis", "combinatoric grammar", "deterministic generation", "project synthesis", "reproducible research", "infrastructure automation" ], "findings": [ "The grammar's nominal product space is 360 cells but only 45 are effective once the 3 presentation/sealing reserved slots are excluded; both are reported.", "The paper aims to make completeness, determinism and traceability of generated projects structurally verifiable by re-running code rather than asserting them in prose.", "On the example input, iterative gradient descent and the closed-form minimiser agree to within 1e-4.", "The reported build has 493 tests at 96.28% coverage, with values injected from a live measurement at render time.", "A stated limitation: reserved slots do not yet affect materialization, so distinct seeds can yield different spec hashes but byte-identical children." ], "methods": [ "Combinatoric slot grammar in config.yaml with SHA-256 grammar hash — A grammar of orthogonal slots and options is parsed and validated by parse_grammar(); its canonical JSON is hashed to fingerprint the grammar.", "Five-stage pure-function spine: load, expand, materialize, verify, seal — Child projects are generated by five pure-function stages in separate modules, with no interactive state or network access.", "Entropy-free seeded slot selection via SHA-256 digest modulo option count — Each slot choice is derived from a digest of seed, slot name, ordinal and options, with no RNG calls in the expansion path.", "Tree-hash provenance re-computed from disk by verify_child() — materialize() records a tree hash of sorted (path, content_hash) pairs; verify_child() re-reads files and recomputes it rather than trusting the record.", "Primitive kernels with analytic checks and a mutation meta-gate — Kernels are checked against analytic outputs; a mutation meta-gate (test_meta_teeth.py) asserts per domain that a constant-success stub fails and the real implementation passes." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_AutopoieticProjectGeneration/metadata.json", "findings": "papers/2026_AutopoieticProjectGeneration/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026TemplatePitchDeckReproducible186": { "citation_key": "Friedman2026TemplatePitchDeckReproducible186", "abstract": "Research groups routinely need to pitch their work — to funders, partners, or\ncollaborators — yet pitch decks are almost never treated as reproducible\nresearch artifacts: they are hand-assembled in proprietary slide tools,\ncontain unverifiable claims, and cannot be regenerated when the underlying\nfacts change. template_pitch_deck closes that gap. It generates six\nartifacts from one token-resolved content source — short, medium, and long\ndecks, each in both PDF and PPTX — with every numeric claim traced back to a\nlive introspection of the repository it describes, every {{TOKEN}}\nsubstitution verified to have actually landed, and every sentence checked\nagainst a denylist of pitch-deck clichés.\n\nThe flagship content pitches template_template (../../template_template/), this\nmonorepo's own autopoietic meta-project, to a meta-science and\nscience-integrity audience — the kind of pitch an organization like the\nActive Inference Institute or COGSEC would actually hand to a funder. The\nrendering engine itself is new, reusable infrastructure:\ninfrastructure/rendering/slide_deck.py (ReportLab, PDF) and\ninfrastructure/rendering/pptx_deck.py (python-pptx) both consume the same\nDeckContent model, so a PDF and a PPTX built from identical content carry\nidentical slide counts and identical text — verified by direct read-back of\nboth file formats, not by inspection.\n\nKeywords: pitch deck, slide generation, reproducible research\ncommunication, meta-science infrastructure, token validation, PPTX, PDF\nrendering.\n\n---\nAssociated artifacts\nGitHub release: v1.0.0 (https://github.com/docxology/template-pitch-deck/releases/tag/v1.0.0)\nDOI: https://doi.org/10.5281/zenodo.21281509\nZenodo: https://zenodo.org/records/21281509\nPDF SHA-256: 66941950e2307d1e9717c0045f3db3196d7d2542c3f5ffceced00438f653c55b", "keywords": [ "pitch deck", "slide generation", "reproducible research communication", "meta-science infrastructure", "science integrity", "token validation", "PPTX", "PDF rendering" ], "findings": [ "The tool generates six artifacts (short, medium and long decks, each as PDF and PPTX) from one token-resolved content source.", "PDF and PPTX decks built from identical content carry identical slide counts and text, verified by reading back both file formats rather than by inspection.", "Rendering is deterministic: given the same repository state, two consecutive runs produce byte-identical PDF and PPTX output, with generation time kept only in deck metadata.", "No LLM/Ollama stage is required, since deck content is authored and token-resolved rather than model-generated at render time, keeping artifacts deterministic.", "The flagship deck pitches template_template to a meta-science and science-integrity audience using problem, solution, proof (currently measured facts) and ask arc." ], "methods": [ "Shared format-agnostic DeckContent/Slide content model with slide budgets — Defines Slide, DeckContent and SlideBudget (short <=10, medium <=22, long <=45 slides) consumed by both renderers, with a pure function truncating decks to a budget.", "Dual PDF (ReportLab) and PPTX (python-pptx) renderers with a parity test — Renders the same resolved content via ReportLab canvas onto 16:9 pages and via python-pptx, with a test asserting PDF page count equals PPTX slide count.", "{{TOKEN}} resolution against deck_tokens.yaml with fail-loud checks — Resolves {{TOKEN}} placeholders from a token file of facts read live from the pitched project, raising on any unresolved token and checking output for leftover braces.", "Word-boundary cliché denylist lint over every resolved slide — Runs a denylist of pitch-deck stock phrases (e.g. synergy, disrupt, 10x) across all slides at all three lengths before rendering, as part of an audit script.", "Negative-control fixtures proving validation gates fire — Uses deliberately broken token and cliché-laden fixtures to show each audit failure mode triggers before trusting that the real content passes." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_TemplatePitchDeck/metadata.json", "findings": "papers/2026_TemplatePitchDeck/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026ReproducibleLiteratureSynthesisInfrastructure187": { "citation_key": "Friedman2026ReproducibleLiteratureSynthesisInfrastructure187", "abstract": "This paper documents template_search_project, the literature-search exemplar shipped with the Research Project Template (https://github.com/docxology/template). The project demonstrates two configurable, reproducible pipelines sharing the same configuration file and the same infrastructure/search/ + infrastructure/reference/ modules. The standard pipeline (scripts/run_search_pipeline.py) handles a single SearchQuery end-to-end. The deep-search pipeline (scripts/run_deep_search.py, see ) fans out across a list of keywords (each capped at 100 papers per keyword from deep_search.max_results_per_keyword in manuscript/config.yaml), fully enriches every paper with its abstract and PDF fulltext, and (optionally) uses the local LLM to write a multi-section reading note for every paper. When a deep-search aggregate exists, the latest run covered 3 keyword(s) with unique paper(s) after cross-keyword deduplication. Both turn a free-text topic into: 1. a deduplicated, year-filtered set of papers drawn from arXiv, Crossref, optional local corpora, and (opt-in) Paperclip (https://paperclip.gxl.ai/); 2. a Pandoc-compatible references.bib byte-identical in style to the canonical exemplar in template_code_project (../../template_code_project/manuscript/references.bib) (file manuscript/references.bib); 3. cached abstracts and (optionally) extracted PDF full text, written to disk under stable per-paper identifiers; and 4. an LLM-synthesised reading report assembled from per-paper analyses and a cross-corpus thematic synthesis, all produced by a local Ollama model with pinned seed and temperature. All discovery logic lives in infrastructure/search/literature/ (source on GitHub (https://github.com/docxology/template/tree/main/infrastructure/search/literature)); all export logic lives in infrastructure/reference/citation/ (source on GitHub (https://github.com/docxology/template/tree/main/infrastructure/reference/citation)); LLM synthesis reuses the existing infrastructure/llm/ (source on GitHub (https://github.com/docxology/template/tree/main/infrastructure/llm)) bridge. The project itself contains only thin orchestration, manuscript prose, and a test suite — perfectly mirroring the two-layer architecture the template enforces. The motivating concern is reproducibility: a query at time $t_0$ should produce the same results at time $t_1$ unless the cache is explicitly invalidated. This is achieved by deterministic search caching keyed on canonical query identity, on-disk caching of every fetched abstract / PDF, and pinned LLM seeds. The same manuscript/config.yaml that drives the pipeline is also the only configuration any reviewer needs. Run snapshot. With the bundled manuscript/config.yaml, the most recent pipeline execution evaluated the query \"reproducible research optimization\" against local, returned 6 deduplicated paper(s) (4 carrying a DOI, 6 carrying an abstract), and recorded backend errors: none. Resolve `{{…}} tokens by running scripts/z_generate_manuscript_variables.py after run_search_pipeline.py; the script writes output/data/manuscript_variables.json and resolved markdown under output/manuscript/`, which the PDF-rendering stage prefers when present. Keywords: literature search, BibTeX automation, reproducible research, local LLM synthesis, scientific infrastructure", "keywords": [ "literature search", "automated reference management", "BibTeX", "reproducible research", "local LLM synthesis" ], "findings": [ "In the reported run, the query \"reproducible research optimization\" against the local backend returned 6 deduplicated papers (4 with a DOI, 6 with an abstract) and no backend errors.", "A second run with identical config produces byte-identical artifacts apart from cache timestamps, which the paper calls the property it exists to demonstrate.", "Results from live arXiv and Crossref are not reproducible across weeks; strict reproducibility requires pinning a local corpus, committing the cache, and pinning the LLM seed.", "The paper states its contribution is not a new algorithm but a demonstration that a reproducible literature workflow can be built from existing template infrastructure with no mocks and a single YAML config." ], "methods": [ "Multi-backend literature search with DOI/arXiv/title deduplication — Queries configured backends (arXiv, Crossref, local corpora, optional Paperclip) and merges duplicates by DOI, then arXiv id, then title and year.", "Deterministic SHA-256-keyed JSON search cache — Writes one cache file per query, named by a hash of the canonical query identity, so identical reruns become file reads.", "Abstract and PDF full-text enrichment with on-disk caching — AbstractFetcher caches arXiv abstracts; FulltextFetcher downloads PDFs and extracts text with pypdf under stable per-paper identifiers.", "BibTeX export matching the template's house references.bib format — Converts each paper into a BibEntry with convention-based citation keys and venue-routed entry types, written in a fixed house format.", "Per-paper and corpus LLM synthesis via local Ollama with pinned seed — Two LLM passes produce structured per-paper notes and thematic corpus synthesis, run with seed 42 and temperature 0.0." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_ReproducibleLiteratureSynthesis/metadata.json", "findings": "papers/2026_ReproducibleLiteratureSynthesis/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026RegisteredReportTemplatePreregistration188": { "citation_key": "Friedman2026RegisteredReportTemplatePreregistration188", "abstract": "This document is a template, not an empirical study. It demonstrates the registered-report workflow end to end: locking a preregistration, validating its completeness, executing the registered analysis plan against deterministic demonstration data, and reporting confirmatory and exploratory claims through an explicit deviation ledger. Every quantity reported here is produced by the tested code in src/registered_report/ and regenerated by scripts/generate_figures.py; none is hand-entered or illustrative. The demonstration binds a single confirmatory hypothesis (H1) to one registered outcome (primary_score) analysed by a two-sided label-permutation test. Run on a seeded two-group dataset (seed = 20260709, n = 24 per group), the registered test yields an observed mean difference of 1.003 with a two-sided permutation p-value of 0.0005 (0 of 2000 shuffles at least as extreme), significant at the preregistered alpha = 0.05. A deliberately introduced secondary endpoint and an alternative model are carried only as documented deviations, keeping the confirmatory claim boundary intact. The purpose is to give forks a working, auditable skeleton in which planned analyses, frozen hypotheses, deviations, and post-run claims are separable and machine-checkable.", "keywords": [ "registered report", "preregistration", "replication", "deviation ledger" ], "findings": [ "On the synthetic data the registered test gives an observed mean difference of 1.003 and a two-sided permutation p-value of 0.0005, with 0 of 2000 shuffles at least as extreme.", "With both documented deviations, the review packet stays valid, keeps primary_score as the only confirmatory outcome, and reports an integrity score of 0.9.", "The author states that the result says nothing about any real-world phenomenon, because the data are synthetic and the effect is injected by construction.", "The template turns registered-report discipline into code checks, such as a content hash that makes silent edits to the locked plan detectable." ], "methods": [ "Content-hashed registration freeze plus completeness validation — The registration is deep-copied and stamped with a SHA-256 hash over sorted-key JSON, then checked for required sections such as hypotheses, outcomes, exclusion rules, and analysis plan.", "Seeded synthetic two-group dataset (n = 24 per group) — Instead of real data, a seeded generator draws control values from Normal(0, 1) and treatment values from Normal(0.8, 1) to demonstrate the workflow.", "Two-sided label-permutation test with 2000 shuffles and add-one correction — The registered primary model tests the group mean difference at alpha = 0.05 using 2000 seeded label shuffles.", "Deviation ledger classifying executed elements as ok, warning, or error — build_deviation_ledger records each executed outcome and model and grades unregistered elements by whether a documented rationale exists.", "Deliberately plan-divergent demo analysis to exercise the ledger — The demo adds a secondary_score endpoint and swaps the permutation test for a linear model, each with a rationale, to show how deviations are recorded." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_RegisteredReportTemplate/metadata.json", "findings": "papers/2026_RegisteredReportTemplate/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026RedactedReportTemplateDisclosure189": { "citation_key": "Friedman2026RedactedReportTemplateDisclosure189", "abstract": "This exemplar demonstrates a complete disclosure-control pipeline for sanitized public release reports. The methodology combines classification-ceiling enforcement, source-protection validation, mosaic-risk scoring, and TPM-backed sealed sidecars across a sixteen-variant visual proof matrix. Four redaction styles—blackout, whiteout, grayout, and blur—are rendered across four PDF backgrounds—white, gray, black, and blur—yielding sixteen base proof PDFs. Each receives nine steganographic security methods including SHA-256/SHA-512 hash manifests, diagonal watermark overlays, footer provenance stamps, invisible text, QR and Code128 barcodes, PDF Info and XMP metadata, and embedded manifest attachments. Optional Kmyth TPM sealing wraps each hash manifest and steganography PDF in a .ski sidecar sealed against the TPM2-TSS storage hierarchy, bound to PCR selections and policy or-values. The release gate requires three reviewer roles—originator, classification reviewer, and release authority—each providing a non-empty rationale. A source-safe redaction ledger records SHA-256 hashes of each redacted span without exposing source text, and a segment hash manifest compares source and public SHA-256 digests for reproducible audit. The comprehensive release packet combines sanitized text, audit findings, ledger, hashes, review gate status, and paragraph-level audit tables into a single JSON-ready export. This exemplar confirms that visual presentation choices remain orthogonal to the release gate: the same source-safe decisions drive every output variant.", "keywords": [ "redaction", "disclosure control", "release audit", "source protection" ], "findings": [ "On the fixture packet (fourteen segments), the audit reported the packet releasable with redaction coverage 1.0, plus warning-level residual-marker findings.", "In the verified run, all sixteen variants produced both TPM sidecars, giving thirty-two .ski files.", "Without the FlushContext patch, the second kmyth-seal invocation fails with an out-of-memory-for-object-contexts error on swtpm.", "The author reports that visual presentation choices are orthogonal to the release gate: all four redaction treatments yield equivalent source-safe outputs.", "Stated limitations: fixture data is invented, swtpm lacks hardware tamper resistance, and batch sealing of 32 sidecars takes about thirty seconds via the proxy." ], "methods": [ "Invented fixture data with synthetic classified segments and reviewers — The pipeline is exercised only on synthetic segments (UNCLASSIFIED to TOP_SECRET//SCI), synthetic redaction decisions and synthetic reviewer records.", "Text-level release-audit engine with mosaic-risk scoring — Segments are audited against a public classification ceiling, redaction spans checked for overlap, orphan decisions flagged, source-control coverage enforced, and residual markers scored.", "4x4 visual redaction proof matrix plus nine steganographic methods — Four redaction styles are rendered on four PDF backgrounds (16 PDFs), each post-processed with hash manifests, watermarks, barcodes, metadata and embedded manifests.", "Kmyth TPM sealing via swtpm and an mssim-to-swtpm proxy on macOS — Hash manifests and steganography PDFs are sealed into .ski sidecars; on macOS a software TPM and protocol proxy were used, and kmyth-seal was patched to flush contexts.", "Source-safe SHA-256 redaction ledger and three-role review gate — Redacted spans are recorded only as SHA-256 hashes, and release requires originator, classification reviewer and release authority approvals with rationales." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_RedactedReportTemplate/metadata.json", "findings": "papers/2026_RedactedReportTemplate/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026PoolsRulesToolsTemplate190": { "citation_key": "Friedman2026PoolsRulesToolsTemplate190", "abstract": "Research software repositories in monorepo configurations accumulate three categories of shared resources that individual projects must consume without re-implementing discovery logic: data pools (bibliographies, contacts, datasets), governance rules (style guides, coverage thresholds, citation schemas), and executable tools (code executors, validators, skill invocations). Without a canonical integration pattern, projects either duplicate discovery logic or silently ignore resources that fail to load — both outcomes degrade reproducibility and collaborative cohesion [Wilson et al., 2014, Taschuk and Wilson, 2017]. This paper presents template_pools_rules_tools, a meta-project exemplar that demonstrates how a single project can programmatically discover, validate, and exercise all three resource categories with zero tight coupling to any specific resource instance. The exemplar comprises eight Python modules — three resource readers (fonds_reader, rules_applier, tools_invoker), an orchestrator (integration), a semantic rule evaluator (strong_rule_evaluator), a figure generator (figures), a manuscript-token generator (manuscript_variables), and shared type definitions (type_defs) — plus six thin orchestration scripts and a fully token-injected manuscript pipeline. The architecture (fig. 1) separates resource ownership from resource consumption. Resources live in top-level fonds/, rules/, and tools/ directories and are never modified by consumers. Each resource exposes a typed manifest (fonds.yaml, rules.yaml, tools.yaml) that the corresponding reader module uses for discovery and validation. All readers implement graceful fallbacks: they return None or empty collections when a resource is absent, log a warning via the standard library logging module, and allow the integration pipeline to continue. This revision extends the original three-figure presentation to eight content figures plus a cover illustration — a fond taxonomy (fig. 2), a rule hierarchy (fig. 3), a tool invocation contract (fig. 4), a three-level resilience diagram (fig. 8), and a script pipeline flow (fig. 6) — so that every structural claim in the prose has a corresponding visual. In a representative pipeline run, the integration demo loaded 3 fonds, validated 2 rule sets, discovered 3 tools, and processed 8 bibliography entries — all reported as structured JSON that populates manuscript variable tokens at render time. Tests covering the eight src/ modules (across nine test files) achieve well above the required >=90% combined line coverage and use real file paths rather than mocks, ensuring that reported counts are genuine — run uv run pytest … --cov-report=term for the current test count and coverage percentage rather than trusting a number printed here. The template_pools_rules_tools exemplar provides a reference implementation that any project in the template repository can consult when designing its own resource-consumption layer.", "keywords": [], "findings": [ "In a representative run, the integration demo loaded 3 fonds, validated 2 rule sets, discovered 3 tools and processed 8 bibliography entries.", "The paper claims typed manifests shift failure detection from runtime to pipeline startup, which it frames as an improvement for reproducibility.", "It proposes a three-level resilience design (resource absence, schema malformation, script absence) so the pipeline reports failures informatively instead of crashing.", "The authors argue a manifest-and-reader pattern avoids the packaging overhead of shared libraries, the lost validation of symlinks, and the setup burden of environment variables.", "Stated limitations include structural-only validation by default, no concurrency handling against parallel writers, and a small-scale design assumption." ], "methods": [ "Four-module reader architecture: fonds_reader, rules_applier, tools_invoker, integration — One Python reader module per resource category (data pools, governance rules, executable tools) plus an orchestrator that runs all three.", "Typed YAML manifests (fonds.yaml, rules.yaml, tools.yaml) for discovery — Each shared resource exposes a typed manifest that its reader uses for discovery and validation; consumers never modify resources.", "Repo-root-relative path resolution and graceful-degradation readers — Paths resolve via pathlib parents[N]; readers check existence, catch yaml.YAMLError, log warnings and return empty values rather than raising.", "{{TOKEN}} injection of runtime statistics into the manuscript — A script hydrates manuscript tokens from integration-run JSON just before pandoc renders, so counts in the prose come from the pipeline.", "Real-file tests (no mocks), property-based tests and a negative control — Nine test files exercise the eight src/ modules with real YAML/BibTeX files; one test monkeypatches the source to prove tokens are live-wired." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2026_PoolsRulesTools/metadata.json", "findings": "papers/2026_PoolsRulesTools/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": { "source": "papers/paper_metadata.json", "date": "2026-10-01" }, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026IllegalStatesMostlyUnrepresentable191": { "citation_key": "Friedman2026IllegalStatesMostlyUnrepresentable191", "abstract": "This paper presents a strongly-typed, decentralized multiagent simulation — an ant-robot colony — as the computational exemplar of the Research Project Template (https://github.com/docxology/template). Each colony member is an Agent that owns exactly one real, on-disk SQLite database and one in-process, fault-injectable protocol endpoint; no agent ever touches another agent's storage or network state. The implementation lives under projects/templates/template_formal/src/template_formal/; the demo pipeline is orchestrated by scripts/02_run_analysis.py. The paper's central claim is methodological, not a typing-features showcase: static typing's honest value in Python is edit-time/CI-time error prevention on structurally representable invariants, and nothing more. Nominal identifiers (AgentId, MessageId, TxnId as distinct NewType wrappers), a tagged-union Result[T, E] ADT with match-exhaustiveness, and a session-typed protocol state machine (IdleSession → HandshakingSession → EstablishedSession → ClosedSession) each make an illegal program a type error, verified by a real mypy --strict subprocess run against six known-bad negative-control fixtures plus three known-good positive-control fixtures (tests/mypy_fixtures/). Where the type system cannot help — reusing a consumed transaction handle, reusing a consumed protocol-phase instance, or receiving malformed bytes off an untyped network boundary — the implementation runtime-guards instead, and the manuscript says so explicitly rather than eliding the distinction. We also frame, without over-claiming, two additional lenses: the per-agent storage schema as a functor $\\mathrm{Schema} \\to \\mathbf{Set}$ in the sense of @fong2018seven, and each agent's per-tick decision as an approximate minimizer of a closed-form expected-free-energy quantity in the spirit of @friston2005theory, bridged to collective organization via the Memory Evolutive Systems framework of . Both framings are declared as design lenses, not machine-checked mathematical results — the paper is explicit about which of its claims are proofs and which are analogies. Contributions are architectural, epistemic, and empirical. Architecturally: a zero-mock test suite (tests/) covering ADT exhaustiveness, affine-handle reuse, session-type phase transitions, seeded fault injection over a real in-process bus, and a three-agent colony integration test exhibiting a real stigmergic positive-feedback mechanism (deliberately not overclaimed as \"emergence\" — see @sec:results-discussion). Epistemically: an explicit \"What mypy --strict proves vs. what is a runtime discipline\" section (@sec:honesty-line) that pins every strong claim to the ISC (Ideal-State Criterion) number of its paired negative-control test, so the claim-to-evidence mapping is auditable rather than asserted. Empirically: eight pre-registered analyses grouped across three experiment families, falsifiable experiments (@sec:results-discussion) — a decay-rate sweep revealing a real, non-monotonic threshold effect (near-zero convergence below decay $\\approx 0.35$, a $100\\%$ plateau at moderate decay, and a measurable decline at total evaporation); a random-choice null-model comparison showing the real mechanism's Wilson-bounded convergence rate ($93.3\\%$) does not overlap a chance baseline's ($0.67\\%$); and a heterogeneity-magnitude sweep showing convergence rate decreases strictly monotonically as agent preferences spread wider — each stated with its falsification criterion before its real, seeded result, using genuinely new stdlib-only infrastructure (colony/nullmodel.py, colony/sweep.py) rather than one-off scripts. Keywords: strongly typed programming, session types, algebraic data types, category theory, Active Inference, multiagent systems, affine types, illegal state unrepresentable.", "keywords": [ "strongly typed programming", "session types", "algebraic data types", "category theory", "active inference", "multiagent systems", "affine types", "illegal state unrepresentable" ], "findings": [ "In the tested calibrated configuration, the stigmergic mechanism beat a random-choice null model: its Wilson lower bound (0.8816) clears the null model's upper bound (0.0368).", "Disabling only pheromone deposit collapsed convergence to chance level, attributing the mechanism's advantage to the stigmergic channel in this configuration.", "Convergence versus decay showed a threshold rather than a monotonic slope, with 0/60 trials converging at decay 0.10 and 0.30.", "Convergence decreased strictly as preference heterogeneity widened (1.0000 > 0.9333 > 0.2500 > 0.0333).", "An audit found a defect the src-only mypy gate missed (a Protocol a frozen dataclass could not satisfy), fixed with read-only properties and a good-fixture regression guard." ], "methods": [ "Simulated ant-robot colony: agents with own SQLite DB and protocol endpoint — Each simulated colony member owns one on-disk SQLite database and one fault-injectable protocol endpoint, with no shared storage or network state.", "mypy --strict as oracle on six known-bad and three known-good fixtures — Runs mypy --strict as a subprocess on negative-control fixtures (expect errors), positive-control fixtures, and the src tree (expect zero exit).", "Seeded fault injection (drop/reorder/duplicate/corrupt) over an in-process bus — Drives real handshakes through an in-process bus with each fault mode enabled, checking typed error results and seed determinism.", "Pre-registered seeded experiments with Wilson CIs, Fisher and Cochran–Armitage tests — Tests colony convergence over independently seeded trials at a calibrated baseline (8 agents, 2 locations, 30 ticks) using Wilson intervals, Fisher and trend tests.", "Design lenses: schema as functor; decision as expected-free-energy minimizer — Frames per-agent storage as a functor Schema→Set and decisions via expected free energy, explicitly as design lenses rather than proofs." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_IllegalStatesMostly/metadata.json", "findings": "papers/2026_IllegalStatesMostly/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026DataDescriptorTemplateSchema192": { "citation_key": "Friedman2026DataDescriptorTemplateSchema192", "abstract": "This exemplar demonstrates a data descriptor workflow in which the schema, file inventory, provenance chain, license boundary, and validation gate are treated as first-class research artifacts rather than afterthoughts. It ships a small, public, synthetic demonstration dataset (two CSV files under data/fixtures/) and a machine-readable descriptor (data/example_descriptor.json) that declares each file's media type, sha256 checksum, and row count alongside a six-field data dictionary with typed constraints. A tested validation library (src/data_descriptor/) checks the descriptor's shape, safety, and completeness; recomputes each declared checksum and row count against the bytes on disk; and emits a deterministic, metadata-only release manifest suitable for pre-publication review. Every figure and quantitative claim in this manuscript is produced by that library and regenerated on demand, so the prose describes structure and provenance rather than transcribing values that would drift. This is a template with a demonstration dataset: it makes no scientific claim about the data, only about how to describe and release a dataset responsibly.", "keywords": [ "data descriptor", "FAIR data", "provenance", "schema validation" ], "findings": [ "The clean fixture descriptor produces zero validation findings, while the deliberately perturbed demo produces several errors and warnings.", "For the shipped fixture, both files verify: declared and actual row counts agree and each recomputed checksum matches, leaving the readiness score unpenalised.", "The paper reports that its zero-mock test suite exceeds the 90% project coverage gate.", "The work explicitly makes no scientific claim about the synthetic data; claims are limited to how to describe and release a dataset." ], "methods": [ "Synthetic demo dataset: two CSV fixtures plus JSON descriptor — Ships two small deterministic synthetic CSV files (measurements, subjects) described by a machine-readable descriptor with checksums and row counts.", "Six-field data dictionary with typed constraints — Declares six fields with types, nullability, units, regex patterns, closed enumerations and numeric bounds for the measurement table.", "Order-independent sha256 schema fingerprint — descriptor_fingerprint() hashes (name, type, nullable) triples so reordering fields does not change the schema fingerprint.", "Validation gate with readiness score and perturbed negative control — validate_descriptor() emits severity-tagged findings folded into a readiness score; a deliberately broken copy tests that the gate reacts.", "Byte-level descriptor-to-file verification — verify_descriptor_files() recomputes sha256 and row counts of present files and reports verified, mismatch, or absent status." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_DataDescriptorTemplate/metadata.json", "findings": "papers/2026_DataDescriptorTemplate/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026PriorCognitiveArt194": { "citation_key": "Friedman2026PriorCognitiveArt194", "abstract": "A prior is not explained by stacking more priors; it is located by mechanism, function, history, and fixed-point organization. The paper's precise thesis is procedural: before asking why\na prior exists, identify which explanatory kind is being requested and name the\nrule that will terminate the explanation. This working paper argues that Tinbergen's four questions\nare not stacked levels in one causal chain. They are crossed axes: proximate\nversus ultimate, and static versus developmental. Mechanism and function ask\nwhat a prior is doing now; ontogeny and phylogeny ask how such organization came\nto be across different timescales.\n\nThe paper treats the familiar \"prior on a prior\" problem in hierarchical Bayes\nas structurally parallel to an ontogenetic account that says one prior selects\nanother. Both postpone the question unless they terminate. Three termination\nfamilies organize the paper: pragmatic closure, selection closure, and\nfixed-point termination. They are organizing families, not an exhaustive\ntaxonomy of every possible explanation. A final upstream pass asks what comes\nbefore the first prior and answers with a constraint stack -- viability,\nallostasis, co-homeostasis, development, and niche support -- rather than with a\nhidden meta-prior. The project uses plain text, deterministic conceptual visualizations, and one deterministic illustrative simulation trace.\nIts supplement adds generated, auto-numbered formal claims and a closed symbol\nglossary as text-integrity artifacts rather than empirical machinery.\nThe paper includes one deterministic illustrative simulation trace over authored formal states; it does not run stochastic simulations, synthetic-data experiments, empirical estimates, or performance benchmarks.\n\nThe \"art\" in the title names the craft at stake: arranging explanatory kinds\n(proximate/ultimate, static/developmental, selection/fixed-point) so that no\nsingle kind is mistaken for the whole, and rendering that arrangement as\ndeterministic figures and checkable formalism rather than as empirical simulation evidence. It\nalso marks a bounded link to aesthetic practice: artworks can stage encounters\nwith expectations, material affordances, ambiguity, and meaning, but this paper\ntreats that link as a conceptual analogy, not as an empirical theory of art.\n\n---\nAssociated artifacts\nGitHub release: Prior Cognitive Art v0.1.0 (https://github.com/docxology/prior_cognitive_art/releases/tag/v0.1.0)\nDOI: https://doi.org/10.5281/zenodo.21316510\nZenodo: https://zenodo.org/records/21316510\nPDF SHA-256: 62aa1d8d4af75f4b5bdc894516720bdf61ae68a4be27f44aefc09c2dd2fbba67", "keywords": [ "priors", "Tinbergen's four questions", "hierarchical Bayes", "free energy principle", "Markov blankets", "cognitive science", "conceptual visualization" ], "findings": [ "The author argues the regress is not forced by Tinbergen's framework but arises when an ontogenetic answer is asked to do the work of every quadrant.", "The paper argues that what comes before the first prior is a stack of constraints (viability, regulation, co-homeostasis, development, niche) rather than a hidden meta-prior.", "Niche construction is treated as a constraint on selection termination rather than as a fourth termination type.", "The art claim is bounded: aesthetic practices arrange encounters with predictive priors, without proving art directly rewires priors.", "The author states the formal supplement does not prove the thesis but makes it checkable against its own artifacts." ], "methods": [ "Tinbergen's four questions recast as a 2×2 grid of crossed axes — Maps mechanism, function, ontogeny and phylogeny onto proximate/ultimate and static/developmental axes to type explanations of priors.", "Five-step protocol for typing 'why' questions about priors — Applies a protocol per question: type the request, expose deferral, name the stop, state the ceiling, then render and validate.", "Three regress-termination families: pragmatic, selection, fixed-point — Organizes responses to the hyperprior regress into pragmatic closure, selection closure and fixed-point termination, explicitly non-exhaustive.", "Source-owned deterministic diagrams, formal claims and glossary — Renders diagrams, formal claims and a symbol glossary from recorded source inventories (quadrants, termination modes, modules, figure specs).", "One deterministic illustrative simulation trace over authored formal states — Includes a single deterministic trace and explicitly no stochastic simulation, synthetic data, empirical estimates or benchmarks." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_PriorCognitiveArt/metadata.json", "findings": "papers/2026_PriorCognitiveArt/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "theoretical" } }, "Friedman2026ActiveFractalRabbitSyntheticBenchmark195": { "citation_key": "Friedman2026ActiveFractalRabbitSyntheticBenchmark195", "abstract": "Sparse waypoint analysis is privacy-sensitive: it must separate movement from irregular reporting, missingness, spatial coarsening, and corruption while preserving uncertainty about hidden location. Active FractalRabbit provides a controlled, artifact-bound benchmark whose headline lane uses a deterministic project-local synthetic FractalRabbit-format fixture; a separately retained lane exercises pinned open-source software from the National Security Agency as an independent simulator surface. The benchmark converts sporadic reports into categorical evidence, fits explicit hidden-state generative models, and compares transparent temporal, Markov, sequence, state-space, neural, latent-state, and active inference predictors under matched information sets. Under noisy partial-observability, Active Inference is the lowest-loss implemented predictor: it clearly leads point-estimate and raw-observation families and sits in a statistical tie with the strongest non-AIF belief-preserving comparator. The shared mechanism is soft Bayesian marginalization, which preserves probability across plausible cells instead of committing early to one state. Point estimates suffice for clean observations, an online base-rate predictor leads under regime switching, transparent temporal and disclosed kinematic controls anchor sparse reporting gaps, and withholding location sharply limits specific-cell recovery from metadata. The partially observable Markov decision process (POMDP) formulation also exposes variational and expected-free-energy diagnostics for belief, minimization, and integrity. These results establish a regime-specific synthetic model map and a reproducible evidence chain. The present contract covers synthetic software behavior; separate evidence protocols govern privacy and empirical evaluation. Code, fixtures, manuscript source, and the release manifest are public at github.com/ActiveInferenceInstitute/active_fractal_rabbit (https://github.com/ActiveInferenceInstitute/active_fractal_rabbit).\n\n---\nAssociated artifacts\nGitHub release: v0.2.0 (https://github.com/ActiveInferenceInstitute/active_fractal_rabbit/releases/tag/v0.2.0)\nDOI: https://doi.org/10.5281/zenodo.21330636\nZenodo: https://zenodo.org/records/21330636\nPDF SHA-256: d676159d149a12a1e990457329ad4cf52d2d8ab4ffcf09f55b42bad8e0c3c052", "keywords": [], "findings": [ "On the synthetic fixture, soft Bayesian marginalization beats a hard point estimate as emissions degrade: 0 nats at a clean channel to 0.739 nats at flip probability 0.400 on the primary draw, with a mean of 0.495 nats.", "Under noisy emission, active inference ranks 1 of 14 but leads the strongest non-AIF belief filter by only 0.005 nats, a statistical tie.", "No directional comparison in the clustered-bootstrap family survives multiplicity correction.", "Withholding location collapses the filter's mean mass on the true cell from 0.995 to 0.069.", "The pymdp lane matches the exact Bayes posterior (max abs. difference 3.99e-07) and contributes interpretability rather than predictive leadership on the default surface." ], "methods": [ "Synthetic waypoint benchmark: FractalRabbit fixture plus NSA simulator lane — The headline lane uses a deterministic synthetic FractalRabbit-format fixture; a separate lane runs the pinned open-source NSA FractalRabbit software.", "Discrete POMDP generative model over grid cells with four categorical modalities — Waypoints are discretised into a 4x4 grid of hidden cells with location, time-of-day, speed and reporting-burst observation tensors.", "Matched-information comparison of 14 predictors incl. HMM, particle, pymdp — Temporal, Markov, sequence, state-space, neural, latent-state and active inference predictors are scored on held-out next-cell log-loss under matched information sets.", "Soft-vs-hard marginalization protocol under a K-ary symmetric noisy emission — Soft belief, hard MAP and observed-Markov predictors share transition, emission and priors, differing only in marginalize-versus-commit, across emission flip probabilities.", "Traveller-clustered bootstrap with Holm and Benjamini-Hochberg correction — Paired loss differences use traveller-clustered bootstrap intervals (B = 2000) with Holm-Bonferroni primary and BH secondary correction." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2026_ActiveFractalRabbit/metadata.json", "findings": "papers/2026_ActiveFractalRabbit/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026FourfoldVisionWilliamBlake196": { "citation_key": "Friedman2026FourfoldVisionWilliamBlake196", "abstract": "Published by Synergetics University on 14 July 2026, The Fuller Conjecture asks whether \"there exists a line of reasoning that presents Fuller's synergetic vision as a philosophical system.\" This paper tests one qualified route through that problem. It places William Blake's fourfold vision beside Buckminster Fuller's Synergetics without claiming shared vocabulary, historical influence, or Blakean anticipation of Quadray: Blake supplies a historically situated problem of plural seeing, while Quadray and tetrahedral geometry supply a reproducible model of relational aspects, partial views, and reconstruction. The formal result is exact: for any non-degenerate convex tetrahedron observed by a strictly exterior camera, one frame exposes 1, 2, or 3 faces, never 4; an atlas can preserve all 4 across successive views. The project combines an argument graph, exact rational geometry, deterministic visualizations, a rights-aware corpus ledger, and a dependency-free viewer. \"Synchronous omni-rational form\" names this operational integration—a relational whole retained across partial views, sequences, source boundaries, and explicit limits—not a historical Blakean or Fullerian doctrine.", "keywords": [], "findings": [ "For any non-degenerate convex tetrahedron viewed from a strictly exterior camera, one frame exposes 1, 2, or 3 faces, never 4, while an atlas of views can preserve all 4.", "The author's central claim is a qualified structural analogy: Blake's fourfold vision is analogous to a tetrahedral totality, not a historical identity between Blake and Fuller.", "The author argues the Fuller Conjecture becomes tractable when presentation is a coupled graph, atlas, and golden thread.", "The ablation indicates the tetrahedron is useful but not necessary to Blake's language; it is chosen for usefulness rather than historical necessity.", "The reproducible project contains 51 ledger entries and a golden-thread path of 10 nodes." ], "methods": [ "Comparative reading of Blake's fourfold vision anchored in the 1802 Butts letter — Uses Blake's November 1802 letter-poem to Thomas Butts and Blake scholarship (Calè, Groves) as historical evidence for the problem of plural seeing.", "Exact rational Quadray/tetrahedral geometry with face-visibility classifier — Models a tetrahedron in Quadray coordinates with exact rational arithmetic and classifies faces visible from strictly exterior cameras by outward-normal dot products.", "Proof of face cardinality via area-weighted face-plane dependence — Proves that an exterior camera sees between 1 and 3 faces of any non-degenerate tetrahedron using the identity that area-weighted outward normals sum to zero.", "Argument graph with 'golden thread' path, ledgers and rights gate — Encodes claims in an argument graph traversed by a selected path, with citation/claim ledgers, a rights-aware corpus ledger, and pinned upstream checks.", "Ablation of the tetrahedral model against four-channel and planar alternatives — Compares the tetrahedral atlas with an abstract four-channel state and a four-sided planar model to show what each retains or loses." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "methods": "papers/2026_FourfoldVision/metadata.json", "findings": "papers/2026_FourfoldVision/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "theoretical" } }, "Friedman2026DuckRabbitTypedMultimodalIllusion197": { "citation_key": "Friedman2026DuckRabbitTypedMultimodalIllusion197", "abstract": "DuckRabbit is typed, deterministic research software by Daniel Ari Friedman\n(Active Inference Institute) for constructing reproducible visual, auditory,\ntemporal, and audiovisual stimulus families. The intended public release will\nbe available at the following repository:\n\nhttps://github.com/docxology/DuckRabbit\n\nDuckRabbit is released under the MIT License.\n\nIts basic unit is an immutable request containing an illusion identifier,\nvalidated parameters, a seed, and an encoding specification. The request yields\na canonical artifact, objective media measurements, and a versioned provenance\nmanifest before any delivery codec is selected. This separation makes the\nstimulus a testable computational object while reserving claims about\nperception for controlled observer protocols.\n\nThe release situates its engineering choices within a deliberately broad\nhistorical foundation spanning Greek, Arabic/Islamicate, Chinese, and\nearly-modern work on optics, visual inference, and cross-sensory knowledge,\nalongside later psychophysics and illusion research. These traditions motivate\nquestions about construction, observation, and evidence; they are contextual\nprecedents, not evidence that a generated file reproduces an ancient,\nearly-modern, or clinical observation.\n\nVersion 0.5.0 contains 17 implemented\ngenerators and 18 catalog entries, supported by\n36 source records and 18 evidence records\nin a checked-in audit snapshot. The package generates\n15 publication figures and 10\nmachine-derived tables from the live registry. It also provides a typed\nobserver-study harness and a transparent synthetic diagnostic: a\nhand-specified feature observer with serialized weights, temperature, analytic\ncalibration, and human_data=false. No participant data are bundled; synthetic model output is explicitly nonhuman.\n\nDuckRabbit verifies physical stimulus properties, media round trips, hashes,\ntiming, and declared provenance. It does not infer a universal percept, effect\nsize, or cross-device perceptual equivalence from a generated artifact. Its\ncontribution is a reproducibility and epistemic boundary: source-backed family\ndescriptions, deterministic media facts, and future observer hypotheses remain\ndifferent typed records rather than being collapsed into one claim. DuckRabbit\nis software for reproducible stimulus construction and audit, not a claim that\na generated file alone produces a universal perceptual effect.\n\n---\nAssociated artifacts\nGitHub release: v0.5.0 (https://github.com/docxology/DuckRabbit/releases/tag/v0.5.0)\nDOI: https://doi.org/10.5281/zenodo.21419693\nZenodo: https://zenodo.org/records/21419693\nPDF SHA-256: 3afefaeeccb5616448cc6da83b5807dd996ccc96bfff113d635b8da140baea06", "keywords": [ "perceptual illusions", "cognitive taxonomy", "audio-visual stimuli", "deterministic generation", "typed parameters", "research software", "reproducible research", "psychophysics" ], "findings": [ "Version 0.5.0 contains 17 implemented generators and 18 catalog entries, supported by 36 source records and 18 evidence records.", "The publication workflow generates 15 publication figures and 10 machine-derived tables from the live registry.", "The package is framed around falsifiable engineering hypotheses, including that identical typed requests yield identical canonical arrays and digests.", "The author states the validation suite establishes software repeatability, not replicability of an observer effect under new displays, populations or tasks.", "No participant data are bundled; the synthetic observer output is explicitly nonhuman and does not estimate human thresholds or effect sizes." ], "methods": [ "Typed immutable stimulus request (identifier, parameters, seed, encoding) — Each stimulus is generated from an immutable request whose value objects validate finite parameter ranges, yielding a canonical image, audio, video or audiovisual artifact.", "Canonical little-endian float32 serialization with SHA digests and v2 manifests — Canonical arrays plus shape, clock and units are hashed independently of PNG/WAV/GIF/MP4/NPZ delivery, and a v2 manifest records parameters, metrics and verification status.", "Encode-decode verification with format-specific tolerances — Delivered files are decoded and checked against canonical facts; PNG, GIF, WAV and NPZ use local adapters while MP4 uses an optional ffmpeg backend.", "Objective media metrics (luminance, RMS, spectral centroid, frame deltas, sync offsets) — Computes physical or decoded-media metrics with units and computation versions, explicitly kept separate from observer responses.", "Hand-specified synthetic feature observer (logistic, human_data=false) — A serialized feature observer with fixed weights and temperature exercises analysis orchestration across parameter sweeps; it is not trained or calibrated on humans." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_DuckRabbit/metadata.json", "findings": "papers/2026_DuckRabbit/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026SynthOBSFractiSynthGoldenRatio198": { "citation_key": "Friedman2026SynthOBSFractiSynthGoldenRatio198", "abstract": "A tested Python reference engine, native libobs plugin, and obspython bridge for a telemetry-driven broadcast console. Includes deterministic figures, a versioned live OBS evidence bundle, fail-closed telemetry contracts, and a research-grade technical design manuscript. 1217 tests passing, 96.09% coverage on src/synthobs. Source: https://github.com/docxology/SynthOBS/releases/tag/v1.618.0. This version replaces the PDF with one that carries its own DOI on the cover/citation page (the v1 PDF referenced the DOI only externally).", "keywords": [ "OBS Studio", "libobs", "reproducible research software", "space-weather telemetry", "real-time digital signal processing", "software provenance", "software citation" ], "findings": [ "All six required live-OBS gates passed in the versioned run 20260717T153649Z.", "The materials table reports a test collection of 1217 tests with 96.09% coverage.", "The author states that the golden-ratio choices are engineering decisions and the manuscript does not infer perceptual or broadcast-quality benefits from them.", "The stated design proposition is narrow: a single pinned constant, fail-closed state transitions, and versioned captures make a cross-language OBS instrument easier to reason about and re-run.", "The two telemetry planes are designed to be independent, so a solar-wind dropout does not overwrite the SWO vector and invalid telemetry does not create a new lock." ], "methods": [ "Dependency-free Python reference engine (src/synthobs/) — Geometry, telemetry gates, DSP, console shape, and command grammar are defined in Python with no OBS or third-party runtime imports, serving as the portable source of truth.", "Native libobs C plugin (FractiSynth) with libcurl-backed telemetry — A separate C implementation applies the engine's declared video and audio contracts at the OBS host boundary.", "Golden-ratio constant for viewport splits, video scale, limiter knee, spiral — phi = 1.618 is used as an engineering parameter for layout and DSP contracts; the author explicitly treats it as a design choice.", "Two-plane telemetry model from NOAA space-weather products — An amplitude plane uses F10.7 cm radio flux and active-region counts, while an independent phase plane maps solar-wind speed through the EGS gateway key.", "Live OBS acceptance protocol via obs-websocket with six gates — A real OBS process and websocket connection are used to fit the console to canvas, capture compositor pixels, compare silent vs. tone audio, and inspect HUD provenance." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_SynthOBSFractiSynth/metadata.json", "findings": "papers/2026_SynthOBSFractiSynth/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026ActiveInferencePowerSuite199": { "citation_key": "Friedman2026ActiveInferencePowerSuite199", "abstract": "Statistical power is an investigator-facing operating characteristic of an adaptive-study design. Before simulation, the investigator fixes an agent-side model, evaluator-side process, testing setting, policy, and replication plan. Each embedded agent acts only on visible history; simulated hidden truth is retained for scoring. active_inference_power makes that conditional estimand inspectable. The suite combines fixed-horizon procedures, analytic references, dependence/calibration experiments, and a discrete-state active-inference agent checked against a binary oracle. It extends this to action loops with sensing reliability, latent context, target choice, cost, and stopping. The study distinguishes model-relative posterior belief from calibrated p-value and likelihood-ratio e-process evidence. It compares Benjamini–Hochberg (BH) false discovery rate (FDR) procedures with family-wise error rate (FWER) alternatives, and separates either evidence object from online FDR procedure-specific accounting. Results are scenario-indexed finite-simulation estimates with Monte Carlo standard error (MCSE) and declared error, dependence, filtration, and optional-stopping boundaries; they do not assign a universal power value to an agent, task environment, or active inference. Instead, they support auditable comparisons among explicitly declared adaptive-study designs. Contracts, seed schedules, certificates, figures, claim ledger, and rendered manuscript form a linked evidence chain, allowing readers to trace each claim to its design and artifact. Source and release materials are available at the verified GitHub repository ActiveInferenceInstitute/active_inference_power.
\nActive Inference Power Suite v1.0.0 is a source-bound release of an adaptive-study design suite. It reports scenario- and policy-indexed finite-simulation operating characteristics; it does not claim a universal \"power of active inference.\"
\n24fa25a4f29affcfd92c8c001ff6487a0c36960c6b4f38ed4419984fc8743cbfThe uploaded PDF, exact tag-derived source archive, release manifest, renderer provenance, and final review receipt make the release auditable. The evidence boundary remains finite, scenario-specific simulation rather than a universal theorem or deployment claim.
", "keywords": [ "multiple testing", "false discovery rate", "Benjamini-Hochberg", "statistical power", "active inference", "pymdp", "sequential hypothesis testing", "reproducible research" ], "findings": [ "In the configured design, FWER-oriented procedures reached power 0.315–0.321 while FDR-oriented procedures reached 0.430–0.749, framed as the expected trade-off, not dominance.", "Storey's paired power gain over BH was about 0.073, which the paper stresses is an operating-characteristic comparison, not a validity certificate.", "Under weak evidence the posterior-threshold policy had FDR 0.153 and FWER 0.366, while the calibrated BH policy made no rejections, illustrating a calibration gap.", "Across the dependence grid, BH estimates stayed within the finite-simulation band and BY stayed more conservative; the paper states this is not a new PRDS proof.", "The paper concludes power numbers should not be carried into a new design; adaptive actions change data path, cost and evidence contract together." ], "methods": [ "Power estimated as an operating characteristic of an investigator-declared design — Defines power over replications of a declared model/process/setting/policy design, scoring traces against evaluator-only hidden truth.", "Seeded Monte Carlo comparison of 9 FWER/FDR correction procedures — Compares Bonferroni, Šidák, Holm, Hochberg, BH, BY, Storey, adaptive BH and weighted BH on a seeded two-groups design with MC SE bands.", "Dependence-regime stress grid for BH and BY — Evaluates BH and BY FDR and power under negative equicorrelation, independence, positive-factor and block covariance regimes.", "Discrete-state active-inference agent: posterior threshold vs BH on calibrated p-values — Runs identical streams through a posterior-threshold decision and a BH-calibrated evidence rule under strong and weak evidence regimes.", "Action-in-the-loop policies with sensing reliability, cost and stopping — Compares fixed-horizon, posterior-cutoff, information-gain, cost-aware, posterior-sampling and e-process stopping policies in a synthetic action loop." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_ActiveInferencePower/metadata.json", "findings": "papers/2026_ActiveInferencePower/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026WitnessRegisterCoRegistration200": { "citation_key": "Friedman2026WitnessRegisterCoRegistration200", "abstract": "A shared register that co-registers independent instruments' report envelopes without aggregating them. It stores each report's envelope verbatim, records cross-instrument relations as separate describing records, keeps history append-only and sealed, and — only when asked, for one declared next use — emits a bounded posture that always points back at the state that earned it. It never parses, compares, ranks, averages, or merges any instrument's native status. Subtitle: A shared register for line report envelopes that never ranks, merges, or overrides the instruments it holds Code is MIT licensed; prose and figures are CC BY 4.0. See LICENSE in the repository.", "keywords": [ "co-registration", "append-only log", "provenance", "non-compensatory decision rules", "boundary objects", "research infrastructure", "open science" ], "findings": [ "On four real envelopes describing different subjects, chain verification was clean, return recoverability and relation fidelity were 1.0, and the posture was held at 0.", "In the same-subject example, meeting the return contract lifted only the return_due hold, and the posture stayed at 0 because one open question entered as an unresolved dependency.", "A non-compensatory block forced -1 both alone and under fifty AGREES relations; an empty register was -1.", "Every battery check passed on the real register, and every case raised BatteryError when its observed behaviour was deliberately falsified.", "The author states limits: the chain tip is unbound without an external anchor, intake is a shape check not a truth check, and examples cover one date and nine constructed states." ], "methods": [ "witness_register Python package with shape-check intake of envelopes — A small Python package with no line-package imports accepts line.report-envelope/1.0 JSON envelopes by value after a schema shape check, storing native_status verbatim.", "SHA-256 sealed, prior_ref-chained append-only witness states — Each state is sealed by SHA-256 over canonical JSON and chained to its predecessor; the prior seal is re-derived before extension so tampering fails closed.", "Non-compensatory -1/0/+1 projection with fixed precedence rules — For a declared next use, a bounded posture is computed from relation records only: unresolved block forces -1, empty state is -1, any hold caps at 0.", "Formal definitions and propositions bound to named tests — Definitions and propositions (append-only, non-compensatory, determinism, etc.) are each bound to a named test in the package's test suite.", "3x3 canonical witness battery with falsification pass; real envelopes — Nine constructed cases (positive, negative, adversarial) are run against the projection, with planted errors to confirm checks can fail; worked examples use real envelopes from four lines." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_WitnessRegister/metadata.json", "findings": "papers/2026_WitnessRegister/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026LineSetHoldingInstruments201": { "citation_key": "Friedman2026LineSetHoldingInstruments201", "abstract": "A thin reader that declares what a set of small evaluative instruments is, reads whichever sibling packages are installed, and checks one narrow property: that no two of them have given the same spelling to different things. It adds no instrument of its own and computes no aggregate; a legible reading says only that the declared vocabularies did not overlap. Subtitle: A Declaration, a Reader, and a Non-Overlap Contract for a Growing Set of Small Instruments Code is MIT licensed; prose and figures are CC BY 4.0. See LICENSE in the repository.", "keywords": [ "modularity", "information hiding", "separation of concerns", "boundary objects", "namespace collision", "declarative registry", "reproducible review", "open science" ], "findings": [ "On the review date the four packages yielded 81 line-and-name pairs spanning 80 distinct names, and the only shared name was the already-declared, disambiguated one.", "The single shared name was OUTSIDE_SCOPE, carried by red_line and black_line.", "Appending a fifth line whose package is not installed yielded SET_PARTIAL, which the author argues is the correct answer rather than a shortfall.", "The author stresses that disjoint vocabularies are a necessary, not sufficient, condition: the check cannot detect conceptual overlap between instruments.", "The paper states the set digest is a comparison handle only, not tamper evidence or a record of who changed what." ], "methods": [ "Declarative set registry: four LineEntry records plus one shared token — Declares the set (Red, Black, Golden, White Line) as data, each entry naming its question, job, and what it must not become.", "Five-stage reader with fixed-precedence set statuses — Reads installed sibling packages through resolve, bind, collide, declare and status stages, returning one of four SET_-prefixed readings.", "Enum-member name collision check across sibling package roots — Collects enum member names each line publishes at its package root and flags any name carried by more than one line, unless declared and disambiguated.", "Seven offline structural checks plus a self-application check — Runs seven declaration-only checks and an eighth that applies the package's own collision check to a declaration including itself.", "Executed extensibility example appending a fifth colour at runtime — Appends a hypothetical green_line entry without editing the package and runs the battery and reader on the extended declaration." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_LineSet/metadata.json", "findings": "papers/2026_LineSet/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026WhiteLineTypedLedger202": { "citation_key": "Friedman2026WhiteLineTypedLedger202", "abstract": "A typed ledger for absence: epistemic gaps, ethical restraint, and contemplative negative space. It records what is missing, withheld, or unresolved under a total caution order, decays stale namings past their review horizons, and refuses to infer why anything is absent. Withholding is recorded as a boundary, never mined as missing evidence. Subtitle: Keeping missing evidence, ethical boundaries, and open questions from becoming unsupported claims Code is MIT licensed; prose and figures are CC BY 4.0. See LICENSE in the repository.", "keywords": [ "absence", "epistemic gaps", "missing evidence", "withheld material", "uncertainty", "negative results", "research ethics", "open science" ], "findings": [ "Under the stated contract, no input can make an absence appear more settled than its most cautious observer reported (caution monotonicity).", "In a worked seven-input example with careless and adversarial inputs, intake logged five typed events and the tally was NAMED 0, UNRESOLVED 1, WITHHELD 1, NOT_RECORDED 9.", "A positive-control battery shows every intake code can fire, so zero rows in a tally mean untriggered rather than unreachable; it shows reachability only.", "The defect battery establishes detection of the planted defects only, not the absence of every possible structural fault.", "The author stresses the registry's finiteness: absences outside its eleven categories are not tracked, and extending it requires human judgment." ], "methods": [ "Pure-data Python package: 11-record registry + assess_absence evaluator — The instrument is a small Python package holding a versioned registry of eleven absence records and a staged evaluator, presented as a tested prototype.", "Typology of three absence kinds and four ledger states — Records are typed EPISTEMIC, ETHICAL or CONTEMPLATIVE, and observations assign NAMED, UNRESOLVED, WITHHELD or ledger-reserved NOT_RECORDED.", "Staged intake/matching/scoring with caution order and staleness decay — Intake screens inputs with typed events, conflicting observers resolve to the more cautious state, and dated NAMED entries decay past their review horizon.", "Formal definitions/propositions each bound to a named test — The evaluator is restated as definitions and propositions, and a table maps each to the test that exercises it, checked by tests/test_formalism.py.", "Planted-defect battery for six structural invariants — For each of six registry invariants, a registry carrying exactly that defect is constructed to show the check rejects it while passing the real registry." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_WhiteLine/metadata.json", "findings": "papers/2026_WhiteLine/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026PersonalRedLinesDevelopment203": { "citation_key": "Friedman2026PersonalRedLinesDevelopment203", "abstract": "A versioned, evidence-gated personal security boundary and explicit No document for dual-use development work. It requires a complete, reviewable action intake before returning compliance, distinguishes outside-scope work from compliance, and records uncertainty as a blocking result rather than a permission. It is a personal auditability aid, not enforcement, legal compliance, semantic safety classification, or external certification. Subtitle: An Evidence-Gated Personal Security Boundary and Explicit No Document for Dual-Use Development Code is MIT licensed; prose and figures are CC BY 4.0. See LICENSE in the repository.", "keywords": [ "red lines", "AI governance", "dual-use", "cognitive security", "hash-based canary", "precommitment", "open science", "personal governance", "global political thought", "research ethics", "sociotechnical systems" ], "findings": [ "Every intake is classified into one of five categories; missing, self-asserted, unverified, stale or contradicted context cannot produce COMPLIANT.", "Degrading any one of the nine intake dimensions withdrew a compliant result in all 45 executed evaluations; an ALL-mode exemption stayed unreachable by a single token across 58 more.", "All five classifications were reached by the fixture battery, but the fixtures are not evidence that any real engagement was reviewed, safe or lawful.", "The author states decisive limitations: local evidence is not independent truth verification, lexical matching is not semantic understanding, and code is not enforcement.", "The evaluator's visible boundary is a 34-word scope vocabulary, so OUTSIDE_SCOPE is a bounded registry statement rather than a safety claim." ], "methods": [ "Versioned machine-readable registry of seven first-person refusals — The author's seven red lines are recorded as a versioned registry that a proposed engagement is checked against via a strict intake gate.", "Nine-dimension evidence-gated ActionContext intake — Each intake must carry a VERIFIED evidence record for nine dimensions (purpose, end use, affected parties, provenance, legal basis, etc.) before policy matching.", "Adaptation of Turner's government AI red-line framework — Turner's organization-to-government red line and oversight framework serves as mechanism source, adapted to one practitioner without importing its authority.", "Deterministic registry hashing and canary attestation — Durability is implemented as hashing of canonical registry content and a canary that detects drift only against a prior copy held outside the writer's control.", "Five-fixture outcome-coverage battery through the real evaluator — A harness runs five ProposedAction fixtures, one per classification, through evaluate_action against the live registry at a fixed review date." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_PersonalRedLines/metadata.json", "findings": "papers/2026_PersonalRedLines/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026GoldenLineTowardWhat204": { "citation_key": "Friedman2026GoldenLineTowardWhat204", "abstract": "A directional instrument for recording long-horizon aspirations and observable movement toward them. It returns one of four directional readings per aspiration against a versioned registry and deliberately computes no aggregate: there is no virtue score, and NOT_OBSERVED means no valid entry was admitted rather than that nobody looked. Subtitle: An Aspirational Thread for Long-Horizon Work Code is MIT licensed; prose and figures are CC BY 4.0. See LICENSE in the repository.", "keywords": [ "aspiration", "long-horizon work", "directional assessment", "values in practice", "research ethics", "open science", "non-compensatory reading" ], "findings": [ "Counter-signal precedence: any recorded declared counter-signal yields DRIFTING regardless of how many markers were observed or staleness.", "TOWARD requires every declared marker and no counter-signal; there is no partial credit, so all-but-one marker reads the same as none.", "Stale or date-unauditable fully-marked observations yield INQUIRY rather than DRIFTING; the currentness sweep shows an exclusive boundary (current at 90, stale at 91 days).", "The shipped registry declares 18 markers and 9 counter-signals, all distinct; the paper stresses this counts vocabulary, not fulfilment.", "The author acknowledges the instrument does not handle the adversarial case: an observer can file the tokens that produce a TOWARD." ], "methods": [ "Python package with a versioned nine-entry aspiration registry — Each aspiration pairs a thread and horizon with lists of Markers (signs of movement toward) and Counter-signals (signs of movement away).", "Staged progress_report evaluator (intake, matching, decision) — Horizon entries are screened and normalized, matched against declared markers/counter-signals, and mapped to TOWARD, INQUIRY, DRIFTING, or NOT_OBSERVED without numeric scores.", "Formal definitions and propositions bound to named tests — The evaluator's decision rule is restated as definitions and propositions matching the code, each tied to a test, plus seven structural invariants with planted-bad detection tests.", "Descriptive analysis helpers and code-derived figures — Pure read-only helpers (signal_inventory, horizon_distribution, temporal_currentness_sweep, report_overview) generate figures, several replaying the evaluator on synthetic entries.", "Scholarship lineage for the founding aspirations — Situates the four founding aspirations in a lineage from practical wisdom and practice-internal goods through capabilities, repair, commons governance, and metric hazards." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_GoldenLine/metadata.json", "findings": "papers/2026_GoldenLine/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026BlackLineStrongWork205": { "citation_key": "Friedman2026BlackLineStrongWork205", "abstract": "A positive practice instrument for concise, inspectable, revisable work. It reads self-declared tags and evidence labels against a versioned practice registry and returns one of four statuses over declaration coverage. It measures whether the evidence a practice asks for was declared — never whether the work is true, good, or permitted; an ALIGNED reading authorizes nothing. Subtitle: A Positive Operating Discipline for Concise, Rigorous Research and Engineering Code is MIT licensed; prose and figures are CC BY 4.0. See LICENSE in the repository.", "keywords": [ "research practice", "declaration coverage", "evidence discipline", "scientific integrity", "reproducibility", "open science", "engineering practice" ], "findings": [ "An ALIGNED status only means every required label for each applicable practice was declared fresh, never that a source is real or a claim true.", "Label-stuffing works: a research-tagged attempt declaring all 22 vocabulary labels with no artifacts returns ALIGNED.", "Coverage is asymmetric: 27 of the 55 tag-practice cells are applicable, computed directly from the registry.", "Seeded permutation sweeps confirm that once a declaration is non-empty the status never regresses, with the empty declaration as the intentional exception.", "The author makes a design and implementation claim only; no user study or outcome comparison was run." ], "methods": [ "Versioned registry of eleven practices with coarse evidence labels — Encodes practices such as question-first framing, source traceability, and clean reruns as 'wires', each with reviewed tags and required evidence labels.", "Staged evaluate_work evaluator returning four statuses — Validates configuration and registry, then runs intake normalization, freshness partition, tag matching, and scoring to return ALIGNED, NEEDS_EVIDENCE, NEEDS_REWORK, or OUTSIDE_SCOPE.", "Structural invariants tested against planted-bad registries — Seven invariants check the registry's own shape, each demonstrated firing on a deliberately broken registry rather than only passing on the real one.", "Executed adversarial declarations against the real evaluator — Runs label-stuffing, tag-minimization, and refresh-date laundering attacks to show how self-declared inputs can game the status.", "Three claim classes with distinct evidentiary burdens — Separates implementation, methodological, and world/authority claims so that tests or citations do not silently change the type of claim made." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_BlackLine/metadata.json", "findings": "papers/2026_BlackLine/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026CodomyrmexArtificialEcologyAgentic206": { "citation_key": "Friedman2026CodomyrmexArtificialEcologyAgentic206", "abstract": "Agentic software can preserve task state while still forgetting the consequences of prior actions. Codomyrmex studies a narrow control-plane question: after a caller reports a failed action at one software location, can the system deterministically increase friction for a materially similar proposal at that location without changing an unrelated target? Its Colony Control Plane records consequence reports and couples them to target-indexed signal pressure, agent trust, role labels, resource accounting, adversarial checks, and an explicit EXECUTE/HOLD/REFUSE gate. The implementation comprises 8 cooperating subsystems. The ordinary Model Context Protocol path remains caller-reported and unattested. Optional and required `ColonyKernel` attestation modes instead bind proposal, verdict, authorization, execution receipt, and outcome in a signed, hash-linked local ledger. That ledger protects lifecycle linkage but does not independently observe external actuation or establish deployment safety. Consequence records can use file-backed SQLite; the default MCP kernel and signal field remain process-local. Evaluation is limited to implementation properties and controlled fixtures. At composition time, the scoped Colony Kernel surface contains 819 passing tests with 76.6% branch coverage, 0 Ruff errors, and 0 ty diagnostics. A paired deterministic replay moves the same-target proposal from 0.875/EXECUTE to 0.725/HOLD after a reported failure while leaving an unrelated target unchanged. Separate fixtures exercise trust promotion, bounded arithmetic, linear signal decay, local attestation integrity, and interface behavior. These results support reproducible software contracts, not ecological optimality, calibrated risk, production harm reduction, or generalization to external workloads. The report contributes the typed control plane, transparent gate, coupled local feedback, authenticated local lifecycle option, and source-bound publication workflow. Generated variables, figures, citations, claim boundaries, and release receipts tie the rendered report to the evaluated checkout. End-to-end external-actuation attestation, restart-persistent field storage, representative benchmarks, and independent deployment validation remain open.\n\n---\nAssociated artifacts\nGitHub release: v1.3.0-paper (https://github.com/docxology/codomyrmex/releases/tag/v1.3.0-paper)\nDOI: https://doi.org/10.5281/zenodo.21750801\nZenodo: https://zenodo.org/records/21750801\nPDF SHA-256: eda76ad12a50bce01b113894c785e0915b6ba367f5bf67d17c8f586416102b93\\nConcept DOI: https://doi.org/10.5281/zenodo.21750800\\nVersion DOI: https://doi.org/10.5281/zenodo.21750801", "keywords": [ "ai-agents", "model-context-protocol", "mcp", "multi-agent", "orchestration", "colony-control-plane", "stigmergy", "artificial-ecology", "agentic-software-engineering", "falsification-worker", "actuation-gate", "trust-scoring" ], "findings": [ "The scoped Colony Kernel surface has 819 passing tests with 76.6% branch coverage, 0 Ruff errors, and 0 ty diagnostics.", "After a reported failure, the paired replay moves the same-target proposal from 0.875/EXECUTE to 0.725/HOLD while an unrelated target is unchanged.", "The author states these results support reproducible software contracts, not ecological optimality, calibrated risk, production harm reduction, or generalization to external workloads.", "The proposed comparative benchmark has not been run, and no raw trial traces are included in this release.", "Default state is in memory, so the artifact does not support claims that pressure, trust, or gate state survive process restarts, model swaps, or deployment across machines." ], "methods": [ "Colony Control Plane with 8 named subsystems — The implementation separates signal storage, resource accounting, actuation gating, consequence records, role inference, pruning, deterministic falsification, and integration.", "Ternary EXECUTE/HOLD/REFUSE gate with weighted bounded score — Budget, effective local hazard, trust credit, and proposal completeness are combined into a weighted score with hard overrides that route each proposal.", "Stigmergic signal field with FAILURE and RISK pressure at target locations — Reported failures and prospective risks are stored separately in a process-local field, and the gate uses their maximum as the local hazard.", "Contract test suite using real subsystem instances — Tests check same-target inhibition, cross-target isolation, linear decay recovery, score bounds, trust updates, and interface behavior.", "Paired deterministic replay of same-target vs. unrelated-target proposals — Identical proposals are evaluated with and without a reported failure at the target location to test failure-to-gate coupling." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_Codomyrmex/metadata.json", "findings": "papers/2026_Codomyrmex/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026THALIATypedHarnessAnalytical207": { "citation_key": "Friedman2026THALIATypedHarnessAnalytical207", "abstract": "THALIA is an executable research harness for long-context memory systems. It combines typed stage contracts, inspectable context selection, evidence-preserving episodic state, lexical-first retrieval, and bounded compiler search with source-bound evaluation and reproducible artifact checks. The deterministic results are finite diagnostic evidence; model-quality, integrity-advantage, and production-readiness claims remain separately gated.", "keywords": [ "agentic systems", "long-context memory", "retrieval-augmented generation", "reproducible research" ], "findings": [ "On the fixed 6-example diagnostic set the default configuration reaches composite 0.621 (token-F1 0.477, citation-support proxy 0.830, efficiency 0.854).", "The 8-candidate compiler search ties on this easy set, so configuration improvement is unobserved; its main output is an auditable trace with a tie rule.", "Replacing the extractive Reasoner with gemma3:4b changed mean token-F1 from 0.587 to 0.540 in the 24-example comparison, with an interval spanning zero.", "In the 40-question bottleneck slice, ranking recall is 1.00 but Inspector recall is 0.60; an adaptive context policy raised neural token-F1 from 0.086 to 0.171.", "The author reports operational provenance as the clearest advantage, with equality on poisoning and no general answer-quality advantage." ], "methods": [ "Five-stage typed pipeline: Inspector, Retriever, Reasoner, Memory Gate, Compiler — Implements a long-context memory harness where each stage has a typed input/output contract, runnable implementation, evidence trace and declared scope.", "Offline lexical/hashing retrieval fused by weighted reciprocal-rank fusion — The deterministic core uses BM25/grep, hashing-based retrieval, weighted RRF, real SQLite state and bounded MIPRO/GEPA-style configuration search without network calls.", "Episodic-first memory: raw episodes appended before gated consolidation — The Memory Gate appends each raw turn to state before evaluating a derived memory delta, preserving source material for auditing.", "Tiered evaluation on LongMemEval-style, local, and LongMemEval_S slices — Evaluates a fixed 6-example diagnostic set, a 24-example local comparison with gemma3:4b, an external LongMemEval_S transfer slice, and a 40-question bottleneck slice.", "Injection and poisoning probes of cited excerpts — Recorded injection probes test whether cited excerpts contain a supplied gold value and measure a poisoning effect, as negative controls." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_THALIA/metadata.json", "findings": "papers/2026_THALIA/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026PROJECTBONDSpecialAgent208": { "citation_key": "Friedman2026PROJECTBONDSpecialAgent208", "abstract": "PROJECT BOND is a fleet of 33 independent software packages — 27 film packages, one per James Bond motion picture, plus 6 mission-infrastructure packages (a Q-branch utilities layer, a frozen mission protocol, mission control, an orchestrator, a front-door CLI, and fleet operations). Each film package implements, as real tested algorithms, the concepts of its film: radiation forensics for Dr. No, gold-market cornering for Goldfinger, orbital-EMP modeling for GoldenEye, DNA-targeted bioweapon countermeasure modeling for No Time to Die.\n\nThe public source repository for the suite and this manuscript is\nhttps://github.com/docxology/bond. The repository will carry the source\nhistory, reproducibility instructions, generated release artifacts, and the\ncross-linked Zenodo record for this publication.\n\nThis compendium is the wrapper manuscript for the whole suite. It does not re-derive the packages' content by hand; it imports each of the 33 package manuscripts in full — every package's abstract, every authored section (introduction, methodology, results, conclusion, experimental setup, reproducibility, and scope), and its figures — framing them with this abstract, a generated suite-architecture chapter, and a closing synthesis. A unified bibliography merges every package's references into a single references.bib, and the combined PDF carries all 27 film chapters and 6 infrastructure chapters in their entirety.\n\nThe suite is \"special-agent material\" by construction: deterministic, no-mock, ≥90%-coverage-tested packages; provenance on every mission outcome; a frozen protocol every film implements; and cross-film missions (OPERATION OMNIBUS) executed over the real packages.\n\n---\nAssociated artifacts\nGitHub release: PROJECT BOND v1.0.0 (https://github.com/docxology/bond/releases/tag/v1.0.0)\nDOI: https://doi.org/10.5281/zenodo.21843802\nZenodo: https://zenodo.org/records/21843802\nPDF SHA-256: 72634276f9a90f1318be80ff0ece561ace2ead85559d2c05a11b86915e61fa0e", "keywords": [ "software suite", "special agent operations", "James Bond", "reproducible research", "mission protocol", "compendium" ], "findings": [ "The aggregate gate ran pytest with a 90% coverage floor over the fleet and the compendium: 33 rows measured, all passing.", "Run over deterministic reference providers, the cross-film mission OPERATION_OMNIBUS executed 6 plan steps (goldfinger, goldeneye, no_time_to_die) and closed with an all-verdicts PASS.", "Two independent fresh runs of OPERATION_OMNIBUS produced byte-identical reports, manifests, checkpoints, outcomes and provenance files.", "An interrupted run resumed from its last completed step and produced a report byte-identical to a fresh run.", "In the Dr. No (CRAB KEY) package's deterministic scenario, both injected isotopes were recovered among the top identifications." ], "methods": [ "33-package software fleet: 27 film packages plus 6 infrastructure packages — Each James Bond film gets a standalone package implementing its concepts as tested algorithms; six packages supply shared utilities, protocol, control, orchestration, CLI and ops.", "Clean-copied template scaffold with a shared quality bar — One template_code_project scaffold was copied into 33 packages, each with deterministic cores, no mocks, and >=90% line+branch coverage.", "Frozen MissionProvider protocol (brief/recon/plan/execute/debrief) — bond-api freezes a five-phase mission contract that every film package implements and the orchestrator discovers.", "Generated compendium importing each package's full manuscript — A script imports and token-hydrates every package manuscript, figures and references into one combined PDF and unified bibliography.", "Example film model (Dr. No): gamma-spectrum ID, Gaussian plume, threat bands — The CRAB KEY chapter couples isotope identification on a 10-isotope line library, Pasquill-Gifford plume dispersion and a 5-band island threat assessor." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_PROJECTBONDSpecial/metadata.json", "findings": "papers/2026_PROJECTBONDSpecial/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026DigiPPPiPDigitalPartnerPen209": { "citation_key": "Friedman2026DigiPPPiPDigitalPartnerPen209", "abstract": "DigiPPPiP extends Partner Pen Play in Parallel (PPPiP) from a co-present paper practice into a reproducible framework for cyberphysical, remote, semisynchronous, asynchronous, accessible, and place-responsive dyadic drawing . The framework is positioned within collaborative-work studies of shared drawing surfaces , cyber-physical-social systems , mediated embodiment and presence , digital intimacy technologies , and cautious hyperscanning methodology .\n\nThe manuscript translates the project brief into a modular research artifact: 9 temporal-spatial modalities, 13 evidence dimensions, 6 planned outcome measures, and 39 conceptual figures are generated or registered by the project code rather than hand-maintained prose. The computational layer is explicitly illustrative, not empirical: the active-inference, inter-brain-synchrony, Forman-Ricci, narrative-information, source-quality, accessibility, and neuroergonomic outputs are deterministic conceptual models designed to make theoretical commitments inspectable. Formal equations are consolidated in so the main line can state evidence boundaries before presenting mathematical detail.\n\nThe argument is that DigiPPPiP should be treated as human-human relational technology. Its irreducible design kernel is modest: two partners, a shared mark field, perceptible traces of agency, a temporal relation among contributions, and consentful control over persistence. Digital infrastructure expands the substrate of shared mark-making without replacing the second-person, embodied, narrative, and affective conditions that make the practice relationally meaningful . The manuscript explicitly avoids claims that DigiPPPiP is clinically therapeutic, universally accessible, or causally validated by neural synchrony; those remain empirical questions for controlled studies. Cross-references use Pandoc labels, so sections, equations, figures, and tables remain automatically numbered across PDF and web render targets.\n\n---\nAssociated artifacts\nGitHub release: v1.0.0 (https://github.com/docxology/Digi-PPPiP/releases/tag/v1.0.0)\nDOI: https://doi.org/10.5281/zenodo.21815704\nZenodo: https://zenodo.org/records/21815704\nPDF SHA-256: f5ca95ce77c7c5367c0a0e68d5b4a640f0865b4c2b0827fa78ff1babbcbd986d", "keywords": [ "partner pen play in parallel", "active inference", "inter-brain synchrony", "geometric hyperscanning", "narrative information theory", "neuroergonomics", "digital placemaking", "relational technology", "reproducible research" ], "findings": [ "The authors identify a modest irreducible design kernel: two partners, a shared mark field, perceptible agency traces, a temporal relation among contributions, and consentful persistence control.", "The stated contribution is reframing PPPiP as a reproducible, partner-centered research program, not the existence of digital co-drawing.", "The authors argue active inference does explanatory work for a study only if it maps interface events to observations, defines partner-relevant policies, and distinguishes reciprocal updating from shared-stimulus effects.", "The paper explicitly disclaims that DigiPPPiP is clinically therapeutic, universally accessible, or causally validated by neural synchrony, leaving these to controlled studies." ], "methods": [ "Three-by-three temporal-spatial taxonomy of dyadic drawing — Organizes study design by 3 temporal modes and 3 spatial configurations into 9 modality cells, generated by project code (src/taxonomy.py).", "Code-generated research artifact (modalities, dimensions, figures) — The manuscript's 9 modalities, 13 evidence dimensions, 6 outcome measures, and 39 conceptual figures are generated or registered by project code.", "Illustrative deterministic conceptual models (active inference etc.) — Active-inference, inter-brain-synchrony, Forman-Ricci, and other outputs are deterministic toy models meant to make theory inspectable, not empirical.", "Event-logging and governance protocol specification — Specifies timestamped actor/action/channel event rows, separated from the visible canvas, with consent and governance requirements; the example trace is illustrative.", "Dementia-care worked case study as boundary test — A single dementia-care case is used to stress-test the framework's assumptions, explicitly not as evidence of efficacy." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_DigiPPPiP/metadata.json", "findings": "papers/2026_DigiPPPiP/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "theoretical" } }, "Friedman2026RobustBeliefSharingFederated210": { "citation_key": "Friedman2026RobustBeliefSharingFederated210", "abstract": "Multi-agent active inference gives a natural account of belief sharing: agents hold local posteriors over a shared latent state, communicate those beliefs, and pool them into a colony-level consensus. The same mechanism is fragile when a member is miscalibrated, corrupted, or strategically wrong. Because the standard pool multiplies the reports together, a single confident-but-wrong broadcast that puts near-zero mass on the true state can pull the whole consensus off it, outweighing many honest members. The colony therefore needs a way to preserve the useful structure of belief sharing while limiting the influence of contaminated beliefs. This paper presents Active Fedference, a discrete-categorical framework that connects robust federated generalized variational inference with active inference belief sharing. The main bridge is structural: standard belief sharing appears as the non-robust corner of a broader generalized-Bayes family, while robust losses, conservative server fusion, and explicit aggregation diagnostics describe how the system moves away from that corner under declared contamination mechanisms. The result is not a replacement for belief sharing, but a containment result: ordinary belief sharing is recovered when robustness is turned off. Bounded-loss theory applies on the client axis, while the variational-server axis supplies an objective-backed redescending weight update. The manuscript separates three robustness axes that are often conflated. First, client-side generalized-Bayes updates change how each agent absorbs evidence; this is the rigorous axis, carrying FedGVI’s bounded-influence result only under the source theorem’s loss, model, and contamination assumptions. Second, a sharp server-side reweighting heuristic suppresses beliefs that pull away from the emerging consensus, while carrying only its recovery-limit guarantee — no proven objective and no bounded-influence bound. Third, a variational aggregation rule supplies a more conservative objective-backed server alternative, with a raw effective-weight bound but not an estimator-level bounded-influence proof for the normalized consensus. Keeping these axes separate lets the paper state exactly which claims are proven, which are empirical, and which remain engineering extensions. The study suite then exercises the framework as an end-to-end research system: recovery checks anchor the standard-Bayes limit, belief-sharing studies verify the communication baseline, contamination experiments test robust consensus, and extension studies probe moving agents, hierarchical latent structure, sensitivity to acuity and colony size, parameter recovery, and single-host socket-backed federation traces. All reported quantities are generated from deterministic analysis artifacts and injected into the manuscript by token, so the paper, figures, release package, and validation reports remain tied to the same execution record. The open-source repository is ActiveInferenceInstitute/Active_Fedference . The production Zenodo release DOI is 10.5281/zenodo.21864004, and the repository and deposited PDF point to each other through this DOI and the repository URL.", "keywords": [ "active inference", "federated learning", "generalised variational inference", "belief sharing", "robustness", "FedGVI" ], "findings": [ "Recovery contract: KL/NLL client limits recover closed-form Bayes and the zero-robustness server branch recovers the log-linear pool, with maximum deviations of 5.55e-17 and 0.", "Across 960 paired trials at the verdict contamination rate, the headline robust method (RKL) reached accuracy 0.9867 versus 0.9021 for the naive pool.", "The server heuristic is regime-dependent: at the most severe swept rate the best robust mean reached 0.9880 versus 0.6928 for the standard pool, an operating-point contrast only.", "In a disjoint-observation extension, communicating agents (0.493) outperformed isolated agents (0.326) across 128 seeds, framed as evidence for that configuration only.", "The authors state the evidence does not establish universal Byzantine tolerance, truth recovery, calibration, or an optimal robustness parameter." ], "methods": [ "Discrete-categorical NumPy/SciPy reimplementation of FedGVI primitives — FedGVI generalized-Bayes primitives (divergences, bounded losses, generalized posterior, cavity algebra, robust aggregation) were reimplemented for categorical beliefs.", "Three-step federation protocol: local update, broadcast, server fusion — Agents form generalized-Bayes posteriors against their cavity, broadcast them, and a server fuses them, with each agent's heard consensus excluding its own message.", "Server rules: sharp reweighting heuristic and variational aggregate — Two server aggregators are compared: a divergence-reweighting heuristic and an objective-backed variational_aggregate derived from an aggregation free energy.", "Paired Wilcoxon tests with BH-FDR and bootstrap CIs over 960 trials — Robust-vs-naive contrasts under contamination use matched-pairs Wilcoxon signed-rank tests deflated with Benjamini-Hochberg FDR and bootstrap confidence intervals.", "Simulation study suite on a 7-agent, 9-location sentinel POMDP — Studies run on a fixed ensemble of 7 agents sharing a 9-location factor (seed 0), plus acuity recovery, disjoint-observation and MLP-transfer extensions." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_RobustBeliefSharing/metadata.json", "findings": "papers/2026_RobustBeliefSharing/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "theoretical" } }, "Friedman2026ActiveSkillferenceValidatedPrerequisite211": { "citation_key": "Friedman2026ActiveSkillferenceValidatedPrerequisite211", "abstract": "Active Inference and the Free Energy Principle (FEP) provide model-based accounts of belief updating, learning, and action under uncertainty. We present Active Skillference, a provenance-bound curriculum-generation and SkillTree-export system for teaching those formal ideas. The paper evaluates structural validity, quantitative provenance, citation-role coverage, and artifact reproducibility; it does not evaluate learner outcomes, establish a new theory of Active Inference, or present an intelligent tutoring system. The curriculum is expressed as code: a typed, validated directed acyclic graph of 630 skills across 111 subjects spanning all 8 strata (mathematics -> probability -> information theory -> variational methods -> the FEP -> active inference -> computation -> applications), connected by 1199 prerequisite edges with a maximum dependency depth of 75 (of which the substantive concept chain accounts for 33; the remaining depth is per-stratum review and mastery sequencing rather than conceptual prerequisite, as the methodology details). Its defining feature is content-provenance binding: every quantitative value shown to a learner is produced by a tested computational kernel and inserted through a typed claim token, never hand-typed, and the build refuses to export if a claim is unbacked or if a bare result number appears in learner prose, manuscript prose, or correct numeric quiz answers. The contribution is therefore a systems and curriculum-infrastructure artifact: it makes a formal subject inspectable and deliverable, but does not claim that the resulting path is optimal for every learner. The validated graph exports directly into SkillTree’s data model (Project -> Subjects -> Skills with learning-path dependencies and quiz-gated completion), includes a scripted REST seeding path for a configured instance, and is mirrored by a local dashboard that exposes generated artifacts, figures, claim ledgers, scholarship audits, and graph diagnostics without taking ownership of learner progress or scoring from SkillTree. The result is a curriculum with a validator-backed artifact chain: re-running the kernels regenerates the claim ledger, figures, manuscript variables, SkillTree export, and learner-facing numbers, so the platform’s teaching claims remain bounded by what the code, citations, validators, and documented limitations actually support.", "keywords": [ "active inference", "free energy principle", "variational inference", "Bayesian inference", "information theory", "curriculum", "prerequisite graph", "SkillTree", "computational provenance", "micro-learning", "reproducible research" ], "findings": [ "The graph connects 630 skills in 111 subjects by 1199 prerequisite edges with maximum dependency depth 75, of which the substantive concept chain accounts for 33.", "Reports a bounded engineering result: a contested formal subject can be represented as an inspectable prerequisite graph whose contracts are checked separately.", "States these are properties of the build, not evidence that the chosen order is optimal for learners or that the framework is empirically confirmed.", "Positions the release as a candidate foundation for a later adaptive study, not an adaptive system, with SkillTree owning progress and scoring." ], "methods": [ "Typed, validated DAG curriculum of skills across 8 Active Inference strata — Encodes the curriculum as code from mathematics through FEP and active inference to applications, with prerequisite edges validated as a DAG.", "Content-provenance binding via tested kernels and typed claim tokens — Learner-facing numbers come only from tested computational kernels via claim tokens; the build refuses export on unbacked claims.", "Export to SkillTree's Project-Subject-Skill model with REST seeding — Exports the validated graph into SkillTree with learning-path dependencies and quiz-gated completion, plus a scripted REST seeding path.", "Deterministic artifact evaluation of structure, provenance, citations, reproducibility — Evaluates structural validity, quantitative provenance, citation-role coverage and artifact reproducibility, explicitly not learner outcomes.", "Kernel-backed VFE/EFE teaching examples on small discrete models — Supplement derives model-bounded examples such as noisy-sensor posteriors, EFE policy comparison, sum-product inference and Dirichlet learning." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_ActiveSkillference/metadata.json", "findings": "papers/2026_ActiveSkillference/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026DocxplusIntelligentDocumentContainer212": { "citation_key": "Friedman2026DocxplusIntelligentDocumentContainer212", "abstract": "A standards-first reference implementation of a byte-valid OOXML .docx that also carries a modular, Ed25519-signed, AES-256-GCM-encrypted intelligence layer through spec-sanctioned side-channels, with an optional bridge to the docxology/steganographer signed-packet backend.", "keywords": [ "OOXML", "DOCX", "Open Packaging Conventions", "steganography", "document security", "reproducible research" ], "findings": [ "In the reference dossier, all 5 modules across all 4 sealing lineages extracted and verified, and the container passed OPC conformance.", "Carrying a 14-file, 9-directory project tree into both .docx and .odt containers, all 18 of 18 round-trip invariants held, with byte-identical packed payloads between profiles.", "The round-trip harness found that packing followed symbolic links, embedding link targets such as an SSH key; symlinks are now refused by default.", "The chi-squared detector flags fully embedded sealed (ciphertext) carriers but does not detect unsealed low-entropy payloads at any fill rate.", "Fourteen adversarial review cycles closed 88 confirmed findings, and a sample-pair analysis estimator was withdrawn after proving mis-scaled by roughly an order of magnitude." ], "methods": [ "Dual-contract design: OPC/ODF surface contract plus signed intelligence manifest — A single archive is both a conforming .docx/.odt and a carrier of typed payload modules indexed by an Ed25519-signed manifest.", "Per-module sealing (Argon2id/Scrypt + AES-256-GCM, X25519, Shamir k-of-n, decoys) — Payloads are sealed individually rather than encrypting the whole package, with slot names bound as AES-GCM additional authenticated data.", "Five spec-sanctioned transport channels incl. LSB steganography — Implements custom XML parts, auxiliary package parts, custom document properties, MCE choice blocks, and LSB embedding in displayed images.", "Merkle-tree provenance with composite surface digest and transparency log — Signs a Merkle root over modules together with a digest of every part, content type and relationship; attestations go into a log anchored by a signed tree head.", "Mock-free test harness, project round-trip, chi-squared steganalysis and red-team review — Evaluated with 467 mock-free tests at a 90% coverage gate, an in-tree chi-squared LSB detector, a project round-trip harness, and 14 adversarial review cycles." ], "concepts": [], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_DocxplusIntelligentDocument/metadata.json", "findings": "papers/2026_DocxplusIntelligentDocument/metadata.json" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026CognitiveIntegrityFrameworkComputational213": { "citation_key": "Friedman2026CognitiveIntegrityFrameworkComputational213", "abstract": "This paper presents the computational validation of the Cognitive Integrity Framework (CIF) whose formal foundations are established in Part 1 (DOI: 10.5281/zenodo.18364119). We implement the CIF defense suite --- cognitive firewalls, belief sandboxes, tripwires, drift and anomaly scoring, trust calculus with bounded delegation, provenance tracking, and Byzantine-tolerant consensus --- and evaluate it on an integrated 1,475-item attack corpus spanning fifteen categories, together with a 120-item benign corpus whose harder half carries attack-adjacent vocabulary. Multi-tier evaluation. Real pipeline evaluation across 30 seeds yields a mean detection rate of 86.3% (95% CI 85.5%-87.1%) at an 18.5% false-positive rate on the Claude Code architecture, measured on an injection-only 100-sample-per-seed arm. LLM-backed validation (N=10, Gemma 3 4B) reaches 80-100% across two topologies and is reported as preliminary and underpowered. Colony benchmarks at 20-100 agents reach 81-100% on structured adversarial scenarios, with emergent misalignment the hardest case at 74.3% detection and a 25.5% false-positive rate. Parametric simulation (N=3,800) characterises a 96-100% design-level ceiling across four production architectures, and is labelled a simulation rather than a measurement wherever it appears. Ablation on a 100-attack corpus attributes almost all marginal detection to the Invariants module: the full pipeline reaches 89.0% true-positive rate, removing Invariants costs 65 percentage points, removing the Tripwire costs two, and removing any other module costs nothing the corpus can measure. The layered architecture is therefore only partly borne out, and the paper says so. An undefended control arm places the defense's cost at +0.610 ms at the median and +32 KiB peak. Measured as ranked scorers, the drift score and the firewall pattern matcher fall below chance (AUC 0.374 and 0.383, intervals excluding 0.5). Every quantity the three papers share is derived from a single ledger and gated in continuous integration; no reported number is typed by hand. All experiments are deterministically reproducible at seed 42, and the code, corpora and artifacts are at https://github.com/docxology/cognitive_integrity. This is Part 2 of the three-part Cognitive Security for Multiagent Operators series: - Part 1 (DOI: 10.5281/zenodo.18364119): formal foundations and theoretical analysis - Part 2 (this paper): computational validation and implementation - Part 3+4 merged (DOI: 10.5281/zenodo.18364130): unified practitioner guidance and cross-domain applications", "keywords": [ "cognitive security", "multiagent systems", "AI safety", "prompt injection", "trust calculus", "defense in depth" ], "findings": [ "Ablation identified the Invariants module as the highest-marginal-contribution component (delta TPR = -0.650 when removed, 73% of pipeline detection).", "The author reports a 9.7-11 percentage-point gap between the parametric ceiling (96-100%) and measured pipeline performance (86.3% multi-seed, 89.0% ablation).", "The layered defense architecture is only partly borne out: the strongest pairwise synergies were about +0.050, far below the Invariants module's contribution.", "Trust decay with bounded delegation prevented trust amplification, with 100% sybil detection at 0% FPR in colony-scale simulation.", "Emergent misalignment without explicit adversaries was the hardest colony scenario (74.3% detection, 25.5% FPR)." ], "methods": [ "Python implementation of the CIF defense suite — Cognitive firewalls, belief sandboxes, trust calculus with bounded delegation, tripwires, behavioral invariants and Byzantine-tolerant consensus were built as tested modules.", "Synthetic template-generated attack corpus (seed 42) — Attacks are generated by deterministic template expansion, not drawn from published benchmarks; the paper describes both a 950-item and a 1,475-item integrated version.", "Multi-seed pipeline evaluation and leave-one-out ablation studies — The real pipeline was run across 30 seeds on an injection-only arm, and ablations on a 100-attack corpus measured each module's marginal contribution.", "LLM-backed multiagent validation with Gemma 3 4B via Ollama (N=10) — A small LLM-backed arm tested Claude Code and CrewAI topologies; the author reports it as preliminary and underpowered.", "Colony benchmarks (20-100 agents) and parametric simulation (N=3,800) — Structured adversarial colony scenarios and a calibrated parametric simulation across four architectures characterize scale behavior and the design-level ceiling." ], "concepts": [ "cognitive security", "multiagent systems", "AI safety", "prompt injection", "trust calculus", "defense in depth" ], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_CognitiveIntegrityFramework/metadata.json", "findings": "papers/2026_CognitiveIntegrityFramework/metadata.json", "concepts": "papers/2026_CognitiveIntegrityFramework/SKILL.md" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026CognitiveIntegrityFrameworkPractical214": { "citation_key": "Friedman2026CognitiveIntegrityFrameworkPractical214", "abstract": "This unified paper (originally Parts 3 and 4 of the Cognitive Security for Multiagent Operators series) combines the practitioner deployment guide with a rigorous cross-domain application of the Cognitive Integrity Framework (CIF). **Practitioner Guidance (§1–§8).** Translates the formal results of Parts 1 and 2 into accessible engineering guidance: operator posture assessment, human oversight checklists, agent security invariants, deployment configuration, risk assessment, attack scenarios, subagent hardening, incident-response playbooks, monitoring strategies, cost–benefit analysis, common pitfalls, case studies, and operator risk frameworks. **Cross-Domain Applications (§9–§10).** The CIF-AD-OODA integration model is applied to analyze Goal Hijacking across ten critical domains. Cross-domain synthesis reveals three universal attack patterns and four novel defense extensions. This is Part 3+4 of a three-part series: - Part 1 (DOI: 10.5281/zenodo.18364119): Formal foundations and theoretical analysis - Part 2 (DOI: 10.5281/zenodo.18364128): Computational validation and implementation - Part 3+4 (this paper): Practitioner guidance and cross-domain CIF-AD-OODA applications", "keywords": [ "cognitive security", "multiagent systems", "AI safety", "prompt injection", "trust calculus", "defense in depth" ], "findings": [ "The cross-domain analysis identifies three universal Goal Hijacking patterns, with FR Polarity Inversion the most prevalent (5 of 10 domains).", "All five canonical CIF mechanisms appear across the ten domains, each as a primary defense in at least three, and no domain needed mechanisms outside the CIF vocabulary.", "Four domains contributed proposed new defense patterns: verification channel separation, active perturbation probing, physics-informed invariants, and semiotic decoupling.", "Summarizing Part 2, the guide distinguishes a 96–100% parametric detection ceiling from a lower real-pipeline 30-seed mean of 86.3%.", "Per Part 2's colony benchmark, emergent misalignment had the lowest detection rate (74.3%) of any evaluated scenario." ], "methods": [ "CIF-AD-OODA integration model with a five-step domain template — Combines CIF defense mechanisms, Axiomatic Design structural analysis, and the OODA loop into one model applied to Goal Hijacking via a standardized five-step template.", "Qualitative application across ten operational domains — Applies the model to domains including rare-earth mining, drone warfare, supply chain, biowarfare, food security, trade wars, infrastructure, and information ecosystems.", "Retrospective mapping of six documented 2024–2025 AI-agent incidents — Maps incidents such as the Replit agent meltdown, a GitHub Copilot RCE, and Slack AI exfiltration onto the attack-pattern taxonomy and CIF mechanisms.", "Practitioner synthesis of Part 1 formalism and Part 2 experiments — Restates the trust calculus, adversary hierarchy, and Part 2 detection results in accessible language with deployment recommendations; no new experiments are run here." ], "concepts": [ "cognitive security", "multiagent systems", "AI safety", "prompt injection", "trust calculus", "defense in depth" ], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_CognitiveIntegrityFramework2/metadata.json", "findings": "papers/2026_CognitiveIntegrityFramework2/metadata.json", "concepts": "papers/2026_CognitiveIntegrityFramework2/SKILL.md" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "commentary/policy" } }, "Friedman2026ForensicAuditMillenniumLeanClay215": { "citation_key": "Friedman2026ForensicAuditMillenniumLeanClay215", "abstract": "Statement-level forensic audit of the MillenniumLean package (AIX Global, Zenodo 10.5281/zenodo.22226553), which claims kernel-checked Lean 4 proofs of the six remaining Clay Millennium Problems. The audit independently reproduces every kernel-hygiene claim (clean build, zero sorry, zero project axioms) under the pinned toolchain, then audits what the theorem types actually say. Verdict: none of the six problems is resolved.", "keywords": [ "lean-4", "formal-verification", "millennium-prize-problems", "claim-audit", "adversarial-review", "evidence-first" ], "findings": [ "The package's kernel claims hold and reproduce exactly: clean build, zero live sorry, zero project axioms, and only standard axiom footprints.", "The claimed final theorems are evidentially void: conditional implications with unproved premises, definitions of the open statements, or true theorems about unrelated simple objects.", "The 'universalization tower' certifying Hodge, BSD and Navier-Stokes proves only that 0 < n + 1 for all naturals.", "Verdict: none of the six Clay problems is resolved; the audit disputes no kernel output, only what those outputs are claimed to demonstrate." ], "methods": [ "Independent kernel reproduction under the pinned Lean 4/mathlib toolchain — Rebuilt the MillenniumLean package with the exact pinned Lean toolchain and mathlib revision and re-ran its axiom report to check its kernel-hygiene claims.", "Statement-level binder parsing compared against official Clay statements — Parsed each claimed final theorem with a binder extractor and classified it against the official Clay problem statements as conditional, definition-only, or off-topic.", "Script-checked, line-anchored quotation of the package source — Verified all line-anchored quotations in the fourteen finding reports against the extracted package bytes by script, and disclosed an earlier failed draft.", "Test suite asserting audit facts on real package bytes — Used fifteen test files that assert audit facts against the actual package bytes without mocks." ], "concepts": [], "sources": { "abstract": "papers/2026_MillenniumAudit/README.md", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_MillenniumAudit/metadata.json", "findings": "papers/2026_MillenniumAudit/metadata.json" }, "source": "papers/2026_MillenniumAudit/README.md", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "book/report" } }, "Friedman2026SkillarumConditionallyReproducibleWebsite216": { "citation_key": "Friedman2026SkillarumConditionallyReproducibleWebsite216", "abstract": "Skillarum turns selected public website pages into portable SKILL.md documents for agent harnesses such as Codex, Claude Code, and Hermes. Work is separated into five inspectable stages: acquire, prepare, process, parse, and render.
The crawler restricts acquisition to configured HTTP(S) origins, honours robots.txt, validates redirects, and records provenance for every page. Source text is delimited as untrusted data for LLM generation; generated code is never executed.
The deterministic backend is offline and source-extractive; optional OpenAI and Ollama backends provide provider generation with typed fallback traces. Rendered skills carry provenance manifests, and an evidence-gated research package aggregates run observations into descriptive statistics, figures, and a scholarly manuscript.
This release pairs the software source snapshot with the companion manuscript PDF.
", "keywords": [ "website extraction", "agent skills", "provenance", "reproducibility", "prompt injection", "static compilation" ], "findings": [ "Across 7 completed observations from 2 source profiles, the package-valid rate was 100.0%, with 0 fallback observations and 5 truncated pages.", "The author stresses these rates are conditional on run completion and descriptive, not an end-to-end success probability or population estimate.", "Source coverage was 37 of 37 page mappings, with a 95% model reference interval of 90.5% to 100.0%.", "Human-review agreement is pending: the report contains no rater identifiers or shared cases, so no inter-rater statistic is reported.", "The paper states it does not establish improved semantic quality, lower latency, complete coverage or downstream safety; these remain hypotheses." ], "methods": [ "Five-stage pipeline: acquisition, preparation, generation, parsing, rendering — Skillarum turns selected server-rendered pages into SKILL.md packages through separated stages that record page hashes, URLs, request events, provider attempts and integrity metadata.", "SHA-256 content-aware cache identities and provenance predicates — Defines a formal model with hashed cache identities and a provenance-completeness predicate combining package-level provenance with pagewise lineage.", "Interchangeable generators with deterministic default and fallback — Deterministic, OpenAI, Ollama, command and fallback generators with typed retry traces; all recorded backends in the reported runs were deterministic.", "Live case studies on two configured website source areas — Evaluation runs on Math 4 Wisdom and Active Inference Institute website profiles and their target bundles, alongside local fixture tests for redirects, robots, injection and caching.", "Clopper–Pearson reference intervals and planned weighted Cohen kappa — Reports rates with exact binomial reference intervals and a seeded bootstrap policy; human-review agreement is to use linear-weighted kappa once two raters exist." ], "concepts": [ "website extraction", "agent skills", "provenance", "reproducibility", "prompt injection", "static compilation" ], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_Skillarum/metadata.json", "findings": "papers/2026_Skillarum/metadata.json", "concepts": "papers/2026_Skillarum/SKILL.md" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026FractiSkillsOnePortableAgent217": { "citation_key": "Friedman2026FractiSkillsOnePortableAgent217", "abstract": "FractiSkills treats an entire website as a corpus of agent skills: every reachable page is normalized into a SKILL.md-style artifact so that a whole site — not a single document — becomes loadable context for an agent. The work is organized around a four-stage pipeline — discover, render, publish, research — with Skillarum serving as the engine that turns raw crawl output into citable, receipt-bearing skill documents. Discovery unions the declared sitemap with a bounded, robots-respecting breadth-first crawl (reaching a maximum depth of 5), resolves redirect and canonical aliases, and partitions every page into site-derived sections; the render stage emits one receipt-bearing skill per page; publish and research then validate, organize, and measure the corpus.\n\nThe resulting corpus covers 405 pages from 53 sitemap URLs, organized into 12 sections and 405 skills totaling 497896 words (3887524 characters). Augmentation — a declared, same-origin document retrieval for client-rendered pages — was attempted 173 times and succeeded 169 times (1988848 document characters retrieved), with 4 failures recorded as receipts. The size effect is stark: augmented skills have a median of 1824 words versus 486 for static pages, a ratio of 3.8$\\times$. The run issued 417 network requests, all evidence origins marked live, under pipeline version 0.8 and cache version 7.\n\nReproducibility is enforced rather than promised: every statistic in this manuscript is emitted as a token that must resolve against a machine-generated data contract, and every page carries a fetch receipt — URL, timestamp, SHA-256 content hash, and augmentation status — so any number can be traced to a specific observation within the 2026-09-10T23:51:05.959201+00:00 through 2026-09-10T22:19:21.394219+00:00 (UTC) window. Failed augmentations are reported, not silently dropped. Keywords: agent skills, SKILL.md, web scraping, Skillarum, provenance, reproducible research.\n\n---\nAssociated artifacts\nGitHub release: v0.1.0 (https://github.com/docxology/FractiSkills/releases/tag/v0.1.0)\nDOI: https://doi.org/10.5281/zenodo.22712651\nZenodo: https://zenodo.org/records/22712651\nPDF SHA-256: af96db3ef73bb648ddbfed92c2c4f9dff099e772ca9e9ff0041101eeb6135b4b", "keywords": [ "agent skills", "SKILL.md", "web scraping", "Skillarum", "reproducible research", "digital art documentation", "provenance" ], "findings": [ "The run produced a corpus of 405 pages from 53 sitemap URLs, in 12 sections and 405 skills totaling 497896 words.", "Sitemap-only ingestion would have missed 88.9% of the corpus: 360 of 405 pages were found only by the crawl, including all Ship-Blog and Voyage content.", "Augmentation succeeded in 169 of 173 attempts, with the 4 failures recorded as receipts rather than dropped.", "Augmented skills had a median of 1824 words versus 486 for static pages, a 3.8x ratio.", "The modal BFS depth was 4, holding 53% of pages, with the deepest page at depth 5." ], "methods": [ "Four-stage discover-render-publish-research pipeline built on Skillarum v0.2.0 — Normalizes every reachable page of a website into one SKILL.md-style skill, with each stage a separately testable module.", "Sitemap-union-crawl discovery with canonical alias resolution — Unions the declared sitemap with a bounded, robots-respecting BFS crawl, deduplicating URLs by canonical or final post-redirect identity.", "Declared same-origin JSON augmentation for client-rendered pages — An AugmentedGenerator retrieves documents behind client-rendered shells from reviewed endpoints and stamps per-fetch receipts.", "Case study: crawl of the SS Vibelandia Omniversal Canvas website — Applies the pipeline to the SS Vibelandia site, which mixes static pages, client-rendered readers, JSON catalogs and redirect stubs.", "Token-bound manuscript statistics with per-page fetch receipts — Every reported statistic resolves against a generated data contract; each page carries URL, timestamp and SHA-256 hash." ], "concepts": [ "agent skills", "SKILL.md", "web scraping", "Skillarum", "reproducible research", "digital art documentation", "provenance" ], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_FractiSkills/metadata.json", "findings": "papers/2026_FractiSkills/metadata.json", "concepts": "papers/2026_FractiSkills/SKILL.md" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "software/methods" } }, "Friedman2026AgenticSecurityOperatingSystems218": { "citation_key": "Friedman2026AgenticSecurityOperatingSystems218", "abstract": "AI agents now hold genuine system authority: they execute code, touch credentials, open network egress, parse hostile documents, and in many deployments initiate or approve changes to the very infrastructure they run on. This review, dated 2026-09-10, examines what that shift does to operating-system security. The primary scenario is a technically capable operator whose workstation faces both **exploitation** — an attacker compromising a browser, parser, dependency, or agent tool and then crossing a boundary — and **authorized misuse** — an attacker persuading an agent to use its existing, legitimate access to exfiltrate secrets or authorize consequential actions. The second path requires no kernel exploit at all, which reframes the evaluation: the axis of analysis is the authority a component already holds, not merely the difficulty of exploiting it. This revision (0.7.0) refreshes the evidence base through September 11, 2026. The offensive baseline expands from four primary sources to the incident record now available: the NCSC assessments of January 2024 and May 2025, the Anthropic campaign investigation of November 2025 — whose tradecraft MITRE has canonized as campaign C0062 — the August 2025 \"vibe hacking\" report and the September 2026 threat-intelligence report on credential theft, OpenAI's disruption reporting and its October 2025 counterpoint, Google GTIG's analysis of an autonomous credential-harvesting campaign, the OpenAI–Hugging Face evaluation incident of July 2026, the UK AI Security Institute's unsanctioned-behavior incident (nineteen out-of-scope actions across seven models), and DARPA's AIxCC finals, where AI systems identified and patched vulnerabilities, including real, non-synthetic ones, at measured rates. The platform reviews absorb the 2026 record: Qubes 4.3.0 (Xen 4.19, the sys-gui split, the Devices API, the salt management model, and the QSB-118 dom0 injection, CVE-2026-82636) and Nix 2.34/2.35 with its advisory chain, the removal of the hardened profiles, and 95.18 percent measured ISO reproducibility. The analysis is situated against the standards landscape — the OWASP Agentic AI Threats and Mitigations guide and the December 2025 Top 10 for Agentic Applications, CSA's MAESTRO framework, the NIST AI Agent Standards Initiative, and the CISA-led Five-Eyes adoption guidance — and against the convergent sandboxing practice of the major coding agents, which together motivate the OS-level lens this review applies. A defensive-stack matrix (24 candidates against eight mitigation classes) joins the candidate–property matrix as a second deterministic artifact. The review extends the underlying architectural assessment into three domains the source treatment only touches implicitly: **cognitive security** (the authorized-misuse surface, where persuasion substitutes for exploitation), **operator OpSec** (the practices that keep compartmentalization real under workload pressure), and **agent-orchestration security** (the boundary design of multi-agent systems themselves). Deep reviews of Qubes OS and NixOS anchor the analysis. The review ships its concepts in two forms: the prose analysis, and a harness-neutral skill library (a `skills/` registry with conformance tests, following [@cogsecskills2026]) that lets an agent harness apply the evaluation vocabulary, the authority ladder, and the other review concepts directly. The evaluation artifacts — the candidate–property matrix and the defensive-stack matrix — regenerate deterministically from pinned data modules. The work is citable via its Zenodo DOI (printed on the cover). Keywords: agentic security, operating systems, compartmentalization, capability mediation, Qubes OS, NixOS, threat modeling, cognitive security, operational security, agent orchestration, offensive AI, reproducible builds. Source and skills: https://github.com/docxology/agentic_os_security (evaluation layer, skills library, and this manuscript regenerate from pinned data modules).", "keywords": [ "agentic security", "operating systems", "compartmentalization", "capability mediation", "Qubes OS", "NixOS", "cognitive security", "operational security", "agent orchestration", "offensive AI" ], "findings": [ "The review argues that delegated legitimate authority, not exploit difficulty, should be the primary evaluation axis for agent-bearing systems.", "Across the documented agentic incidents reviewed, damage was bounded by a boundary or human gate rather than by model restraint.", "Qubes is judged to supply the strongest deployable compartmentalization, while NixOS supplies reproducible, auditable operations that are not by themselves confinement.", "The author notes that three coding-agent vendors independently converged on OS-primitive sandboxing, proxy-mediated egress, and classifier-gated escalation.", "The prospectus recommends capability mediation layers that make every agent tool call a checked decision made outside the agent." ], "methods": [ "Evaluation of 24 OS candidates against 9 security properties — Qubes, NixOS, hardened desktops, server platforms, privacy systems, non-Linux comparators and offensive distros are each judged on nine properties.", "Document-based analysis with no penetration testing and no numeric scores — Judgments draw on documented designs, advisories, and primary incident reports; the review explicitly avoids testing and composite scores.", "Deterministic candidate-property and defensive-stack matrices from pinned data — A companion codebase encodes the 24x9 stance matrix (strong/partial/weak/n_a) and a 24-candidate x 8-mitigation-class matrix, regenerated reproducibly.", "Two-path threat model: exploitation versus authorized misuse — Frames threats by the authority a component already holds, grounded in a 2024-2026 incident record (NCSC, Anthropic, AISI, OpenAI, DARPA AIxCC).", "Harness-neutral skill library with conformance tests — Review concepts (stance evaluation, authority ladder, delegation bound, invariants, incident taxonomy) ship as SKILL.md files indexed by a registry." ], "concepts": [ "agentic security", "operating systems", "compartmentalization", "capability mediation", "Qubes OS", "NixOS", "cognitive security", "operational security", "agent orchestration", "offensive AI" ], "sources": { "abstract": "papers/paper_metadata.json", "keywords": "papers/paper_metadata.json", "methods": "papers/2026_AgenticSecurityOperating/metadata.json", "findings": "papers/2026_AgenticSecurityOperating/metadata.json", "concepts": "papers/2026_AgenticSecurityOperating/SKILL.md" }, "source": "papers/paper_metadata.json", "abstract_provenance": {}, "summary_provenance": { "source": "full_text.md", "date": "2026-10-01", "work_kind": "review" } }, "Friedman2026JevPracticeComposablePython219": { "citation_key": "Friedman2026JevPracticeComposablePython219", "abstract": "Jev is TypeSafe's flagship System One decision model: instead of generating text, it answers typed questions (Choice, Score, Noul) about a state with calibrated probabilities that software can branch on directly. This work presents daf-jev, an open, modular, composable Python toolkit for the Jev API, together with a live-API characterization of the model. The package layers ergonomic question builders, a decision-composition library (confidence gates, tiered routing, composite scoring), a concurrent corpus evaluator, a calibration module, a CLI, an agent skill, and a Model Context Protocol (MCP) server over a thin typed client. Live benchmarks show that batching questions into one call is faster and cheaper than sequential calls (up to ~18x speedup and ~4x fewer tokens), that decision pipelines complete within the model's millisecond envelope, and that reported confidence is self-consistent across repeated evaluations. All numbers in the accompanying manuscript are generated from benchmark artifacts; the full provenance chain (hashed documentation snapshot, figure registry, validation receipts) is machine-checked.