AGENTS.md

Instructions for AI coding and research agents

This repository is FNE–AXIOMESH v0.2.0, with the Agent and Expert Support edition of its distribution. The scientific kernel remains a small exact finite-field reference implementation. Read README.md, STATUS.json, docs/CLAIMS.json, and docs/IMPLEMENTATION.md before interpreting results or changing semantics.

Start here

1. Use Python 3.10+; install requirements.txt in a virtual environment.
2. Run python -m pytest -q and python validate_release.py.
3. Run python agent_interface.py --request examples/agent_recover.json.
4. Read docs/AGENT_SUPPORT.md for JSON requests and result interpretation.
5. Read docs/EXPERT_REVIEW.md for a proof-to-code review map.

Research contracts

• Action words are chronological: [a, b] applies a, then b. Thus the ambient matrix is T_b @ T_a.
• The declared query, finite prime field, operator family, and trusted surviving measurements determine exactness. Do not generalize certificates to arbitrary neural representations, noisy sensors, unknown model families, or corrupt code.
• Preserve ambiguous and inconsistent outcomes. A missing distinction is not identified by choosing a plausible completion or sampling an LLM response.
• Exact acquisition debt assumes unrestricted scalar linear measurements. Restricted sensors require a separate attainability analysis.
• Tether length counts additional field symbols. Count matrices, maps, metadata, training history, compilation, sensing, search, and decoding costs separately.
• A conventional pooled learner ties the principal prediction benchmark. A rank ratio is not an intelligence multiplier. The five integer seed labels are pseudorandom seeds, not five calendar dates.
• The prior-art report is a bounded inherited search. Do not turn it or old subjective scores into a world-first, 100%-novel, or peer-reviewed claim.

Working protocol

Use the data-only agent_interface.py for external proposals inside its declared limits. It validates small matrices and does not execute submitted code. A larger or nonlinear proposal needs an explicit new backend and validity contract. The interface supplies no oracle: suggested measurement rows still need real evidence before an agent may append their values to memory.

Run exploratory experiments with --output local_runs/<run-id> so the bundled scientific evidence is preserved. Agent guidance is documentation; the repo does not launch background agents, allocate GPU jobs, or submit messages to experts.

For edits to the adapter, test malformed inputs, budget exhaustion, and honest abstention as well as successful execution. For scientific changes, document assumptions, add a meaningful counterexample/regression check, update the claim ledger and evaluation protocol, and run the relevant suite. Distinguish a rerun of published evidence from a new claim.

After intended edits, refresh checksums with python validate_release.py --refresh-manifest, then verify them. Do not refresh the manifest to conceal unexpected changes. Public upload is an explicit action via the supplied uploader; local review commands do not publish. Preserve the owner's existing license status.

Reporting

Return concrete file/function references, exact commands, environment, assumptions, results, negative results, and unresolved issues. A review by an AI is an automated review, not independent human peer review. Use the status vocabulary in the claim ledger and the report template in docs/REVIEW_REPORT_TEMPLATE.md.