This repository preserves the reproducibility release for Source-Preserving Reader Companions and Bounded Refinement: A 1,000-Work Gutenberg Benchmark by Paul Pajo. It provides an auditable, source-preserving reader-support experiment across 1,000 English-language Project Gutenberg works.
The frozen release contains 5,272 scored core artifacts: 1,000 original high-friction units, 1,000 AI-communication companions, 1,000 plain-English companions, 1,000 Reader-Context and Source-Critical companions, 272 eligibility-limited controlled-language companions, and 1,000 extreme-readability amalgams. It also preserves 3,000 evaluated bounded-refinement candidates, including 188 strict accepted gains. The accepted-parent proxy trajectory is 3.9190 → 3.9376 → 3.9384 / 4.
Interpretation boundary: all reported scores are deterministic feature-and-structure proxy outputs. They are not reader-study results, psychometric scale outcomes, evidence of comprehension or reading-speed gains, literary-quality rankings, universal-accessibility claims, or formal ASD-STE100 certification.
Run the release-relative audit from the repository root:
python code/verify_release.pyThe command checks the 5,272 core-score rows, condition counts, rounded aggregate means, companion-family counts, first-loop acceptance and stopping conditions, and the exact final second-loop mean. Review docs/portable_verification_report.json for the frozen passing result.
GitHub Actions is currently unavailable for this account because hosted jobs are blocked before execution by an account-level billing condition. The repository therefore provides a no-billing local gate and an optional independent webhook verifier.
Install the local gate once in a maintainer clone:
sh code/install_local_verifier.shThereafter, a normal git push runs the structural verifier and checks every listed SHA-256 value before pushing. To create a dated local attestation outside the committed release tree, run:
python code/manual_attestation.pyFor independent server-side status reporting, deploy webhook_verifier/ outside GitHub Actions and configure a signed push webhook. Its deployment boundary and security controls are documented in webhook_verifier/README.md. The GitHub Actions workflow file remains in the repository for future re-enablement, but the hosted workflow is disabled remotely while the account-level billing condition prevents jobs from starting.
| Path | Contents |
|---|---|
data/corpus/ |
1,000-work manifest, bounded source units, retrieval log, and provenance materials. |
data/companions/ |
4,272 generated core reader companions in five conditions. |
data/scores/ |
Authoritative 5,272-row JSONL score record and aggregate summaries. |
data/refinement_loops/ |
All 3,000 bounded-loop candidate records, summaries, and parent-state logs. |
verification/ |
Core, fifth-pass, and loop verification reports and checks. |
protocols/ |
Corpus, fifth-pass, and bounded-loop protocols. |
code/ |
Portable release verifier and preserved workflow scripts. |
docs/ |
Data dictionary, reproduction guide, skills/rubric sources, checksums, and release manifest. |
paper/ |
Final manuscript in PDF, DOCX, LaTeX, Markdown, HTML, and TXT, with processed figures. |
The MIT License applies to this repository's original code, documentation, generated companion artifacts, score records, and release packaging, subject to the notice in docs/RIGHTS_AND_PROVENANCE.md. It does not license Project Gutenberg works or other third-party material. Full original source texts are intentionally not redistributed; each corpus record retains a canonical identifier and Project Gutenberg URL. Reader companions are source-linked, non-replacement aids.
See CITATION.cff. Cite both the paper and release tag v1.0.0 when using this frozen research object.
The paper is available in paper/source_preserving_reader_companions_bounded_refinement_gutenberg.pdf, with corresponding DOCX, LaTeX, Markdown, HTML, and TXT versions under paper/.