Maintenance-Coupled Inquiry

A preregistered computational experiment testing whether component-aligned maintenance signals alter evidence allocation, causal diagnosis, and replicated restoration under finite inquiry budgets.

Repository status: Public reproducibility release
Version: 1.0.0
Paper status: Manuscript prepared; not yet published
License status: Code Apache-2.0; data, manuscript, documentation, and figures CC BY 4.0
Latest release: v1.0.0
日本語: README_ja.md

Scope and claim boundary

This repository reports a functional computational experiment in a constrained causal microworld. It supports the following limited findings:

• component-aligned maintenance signals changed evidence allocation;
• actual_need outperformed a yoked_need control that preserved total need while disrupting component alignment;
• replicated restoration improved under actual need;
• evidence-trace differences remained under a common decoder that received neither condition identity nor need vectors; and
• actual need passed the prespecified neutral-world false-repair safety gate.

It does not establish constitutive autonomy, endogenous normativity, constitutive death, biological individuality, a general autonomous scientist, or human-level cognition. Need variables, viable ranges, damage, repair, intervention vocabulary, and stopping rules remain designer specified. See CLAIM_BOUNDARY.md.

Main confirmatory results

Values below were cross-checked against the frozen CSV outputs and integrity audit, not transcribed from memory.

Measure	Frozen result
Confirmatory seeds	72
Primary source runs	576
Ablation source runs	432
Exact replay	288/288
Target sensing difference, actual − yoked	+0.1173
95% stratified paired bootstrap CI	[+0.0722, +0.1632]
Prespecified SESOI	+0.08
Actual replicated restoration	41/72
Yoked replicated restoration	20/72
Paired difference	+0.2917
Neutral-world actual false repair	0/72
One-sided 95% exact upper bound	0.0408

The actual condition did not succeed in every world; replicated restoration was 41/72. The supported claim is a probabilistic improvement over the yoked control, not perfect or general causal competence. See RESULT_SUMMARY.md.

Repository map

Path	Role
manuscript/	Markdown manuscript (public-link canonical copy), DOCX snapshot, and machine-readable tables
supplement/	Supplementary Markdown and DOCX snapshot
figures/	Released paper and supplement figures in PDF and PNG
protocol/	Frozen confirmatory protocol and freeze-readiness evidence
src/	Byte-preserved frozen v0.3 Python package
scripts/	Frozen runners, archived post-run analysis, and public verification helpers
tests/	Frozen unit tests
manifests/	Seed, yoke, chunk, endpoint, SESOI, activation, and freeze manifests
data/processed/	Merged primary and ablation confirmatory data
data/analysis/	Endpoint, common-decoder, ablation, verdict, and summary outputs
data/execution_receipts/	All 120 immutable execution receipts
docs/development/	Development-phase and repository-audit records
legacy/	Historical v0.14 pre-unification baseline, not the paper implementation

See FILE_MAP.md for canonical-source and generated-output relationships.

Quick start

The frozen package declares Python 3.11 or later and has no runtime third-party dependencies. Pytest is required for the test suite.

git clone https://github.com/Louis-Takeuchi/Maintenance-Coupled-Inquiry.git
cd Maintenance-Coupled-Inquiry
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -r requirements.txt
pytest -q
python scripts/verify_public_results.py

The frozen suite contains 45 tests. The verification helper is read-only and checks the published run counts and headline values against the CSV outputs.

Reproducing the analysis

To regenerate the principal endpoint, common-decoder, and ablation tables into an ignored build directory without overwriting the frozen outputs:

python scripts/reproduce_processed_analysis.py --output build/reproduced-analysis
python scripts/compare_reproduced_analysis.py build/reproduced-analysis

The inputs are the merged files in data/processed/. The archived original post-run scripts are retained byte-for-byte in scripts/archived_postrun/; they contain the original /mnt/data/paperB-unified-v1 container path and are preserved as provenance rather than silently rewritten.

No figure-generation source script was present in the supplied public package. The released PDF/PNG figures are available in figures/, but exact figure regeneration from code cannot currently be claimed. This is recorded as an open reproducibility limitation.

Analysis-only reproduction reads existing merged summaries and runs the fixed 50,000-replicate bootstrap. Full confirmatory reproduction instead executes 1,008 source runs plus 288 replay cases and requires the frozen activation, chunk, seed, and yoke manifests. Do not edit those manifests. Review REPRODUCIBILITY.md before attempting a full run.

Data availability

Git tracks the merged primary and ablation run summaries, endpoint tables, analysis outputs, integrity audits, and execution receipts. The v1.0.0 Release provides the verified all-in-one package and cleaned legacy archive together with their checksum files. A separate core ZIP is not included and must not be reconstructed from unverified assumptions.

Full step-level traces are not included in the main Git history or current validated package. A previous full-trace ZIP failed archive-integrity checking and was excluded. The frozen code and manifests required to regenerate traces are provided; a fresh validated trace deposit is still required before claiming that every raw step trace is public.

Historical versions

v0.14 is preserved as a historical pre-unification baseline. It did not generate the final confirmatory results and must not be treated as the canonical implementation of the paper.

Versions v0.1–v0.13 are development provenance, not the main implementation. Their cleaned archive is a Release asset rather than part of the main tree. See LEGACY_VERSIONS.md.

Citation

See CITATION.cff and the manuscript title in manuscript/. The paper is unpublished, and no journal volume, issue, pages, article DOI, or archive DOI has been assigned here. ORCID, affiliation, DOI, and submission metadata remain future additions documented in PUBLICATION_OPEN_FIELDS.md.

AI assistance disclosure

OpenAI ChatGPT assisted with English translation, structural editing, and document formatting. The author reviewed and verified the scientific claims, calculations, references, and final wording and remains responsible for the content. This repository does not claim that AI autonomously made research decisions or ran the confirmatory experiment without human oversight.

License

Material	License
Code (src/, scripts/, tests/, Python configuration, CI)	Apache License 2.0
Data, manifests, CSV outputs, and execution receipts	CC BY 4.0
Manuscript, supplement, documentation, figures, README, and protocol documents	CC BY 4.0

Third-party materials remain subject to their own rights and licenses. See NOTICE.md for the scope statement.