Overview
Source code for Skillarum: Conditionally Reproducible Website-to-Agent-Skill Compilation. Skillarum turns selected public website pages into portable SKILL.md documents for agent harnesses such as Codex, Claude Code, and Hermes, in five inspectable stages (acquire, prepare, process, parse, render) with provenance manifests and an evidence-gated research package. Release 0.3.0 added index-derived profiles (one skill per page listed on a site's own index; 528 Math 4 Wisdom Research pages), block-preserving extraction, authored target descriptions, and math4wisdom-cube, a versioned investigation skill with separate source, mathematics, interpretation, and test ledgers, audited by three independent AI investigators whose verbatim reports are included. Release 0.3.1 is a correction release: math4wisdom-cube 0.3.0 withdraws an investigator claim that 0.2.0 adopted without recomputation (1+3+3+1 is not the generic grading pattern) and adds a check for the corrected statement; the changelog's account of investigator independence and two documentation statements are corrected.
Overview source: Curated paper metadata.
Methods and contributions
Read the source for the full argument, qualifications, and evidence.
Findings and contributions
- 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.
Key concepts
- website extraction
- agent skills
- provenance
- reproducibility
- prompt injection
- static compilation
Summary sources: Paper metadata and evidence · Learning guide · Extracted source text.
Document downloads
PDFs are hosted here; supporting sources may use pinned archives.
- Skillarum_combined.pdf PDF · 3.46 MiB · extracted-text source
Citation
Citation metadata follows the unified bibliography.
Catalog details and resources
Related works
Ranked by recorded methods, topics, contributors and associated software, with contributor roles and domain-only matches labelled. Metadata overlap is not a citation or a scientific equivalence.