Overview
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.
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).
records 12 / 12 invariant checks passed. SI planning horizon: 2 steps. Sweep RMSE 0 nats bounds analytical–empirical agreement on the coupling grid.
---
Associated artifacts
GitHub release: v0.3.2 (https://github.com/docxology/template_active_inference/releases/tag/v0.3.2)
DOI: https://doi.org/10.5281/zenodo.20417021
Zenodo: https://zenodo.org/records/20417021
PDF SHA-256: f191b48f94394cab17069fd04502c59fc1c287e7893eb078e05ba4be04d4a04c
Overview source: Curated paper metadata.
Methods and contributions
Read the source for the full argument, qualifications, and evidence.
Findings and contributions
- 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.
Summary sources: Paper metadata and evidence · Extracted source text.
PDF downloads
Archived files available directly from this site.
- Friedman_2026_Active_158b2fe2.pdf PDF · 3.83 MiB
- Friedman_2026_Active_713452dd.pdf PDF · 0.59 MiB
- Friedman_2026_Active_f191b48f.pdf PDF · 3.91 MiB · extracted-text source
These files may be different versions or companion documents. The archive does not identify a latest edition.
Citation
Citation metadata follows the unified bibliography.
Catalog details and resources
Related in Computational
Other catalogued works in the same domain.