researcher-agentsmd: Instructions file for Codex

AGENTS.md — Research Execution Rules

These rules apply to any agent working in this repository. They are meant to protect research validity, reproducibility, and auditability.

1. Core Principle

A result is not accepted because it looks good.

A result may support a research claim only when it is traceable, reproducible, audited, and robust enough for the claim being made.

If another agent cannot reconstruct how a result was produced from saved artifacts, then that result cannot support a research claim.

2. Result Status

Every reported result must be assigned one status:

• EXPLORATORY: informal observation; useful for direction only.
• DIAGNOSTIC: useful for debugging or understanding behavior; not claim-supporting.
• EXISTENCE_PROOF: shows something happened once; not evidence of a stable recipe.
• SINGLE_RUN: produced by one traceable run; not yet robust.
• REPLICATED: reproduced across appropriate seeds, splits, or equivalent robustness checks.
• ACCEPTED: audited replicated result that may support a stated claim.
• LOCKED: accepted result written into a consolidated report; changes require correction or retraction.
• FAILED: did not meet its predeclared success criteria.
• RETRACTED: previously stated claim invalidated by later evidence.
• PROVENANCE_ORPHAN: artifact exists, but its generation path is incomplete or inconsistent.

Do not present EXPLORATORY, DIAGNOSTIC, EXISTENCE_PROOF, or PROVENANCE_ORPHAN results as accepted claims.

3. Forbidden Actions

Agents must not:

• produce claim-supporting model artifacts from inline commands, notebooks, REPL sessions, pasted scripts, or unsaved temporary code;
• overwrite existing model artifacts or reuse an old output directory for a new run;
• copy a model artifact and describe the copy as a fresh training run;
• silently change metrics, data splits, training objectives, or model-selection rules;
• use test data or answer labels to select a model artifact unless the claim explicitly permits it;
• expand the experiment scope beyond the approved plan;
• describe a single successful run as stable or general;
• hide failed runs, failed seeds, or negative controls.