ai-research-reproduction

Purpose

Guide README-first deep learning reproduction toward a minimal trustworthy run with auditable evidence. Reproduction is not "make it run by changing anything"; faithfully read the README, environment, weights, datasets, and documented commands, then record results and deviations. Start with references/agent-operating-principles.md; load references/research-rigor-principles.md and references/deep-learning-experiment-principles.md when scientific meaning or experiment details are at stake.

The deterministic entrypoint is scripts/orchestraterepro.py. It includes a self-contained bundled/ runtime, so this skill works when installed alone; separately installed companion skills remain optional reusable entrypoints. Executed commands persist lifecycle state, append-only events, and full streamed stdout/stderr under reprooutputs/runtime/<runid>/. A CANCEL file in the active run directory requests process-tree cancellation. For recovery, queues or model gates, read references/runtime-and-model-adapter.md; for the optional model/tool loop, read references/agent-runner.md and use scripts/runagent.py.

Fit

Use this skill when all are true:

• The target is an AI code repository with a README, scripts, configs, or

documented commands.

• The request spans multiple trusted phases such as intake, setup, execution,

training verification, analysis, paper-gap resolution, and reporting.

• The desired result is a small reproducible target, not broad experimentation.

Do not use this skill for paper summaries, generic environment setup, isolated repo scanning, standalone command execution, open-ended research design, or explicit candidate-only exploration.

Trusted Target Selection

Choose the smallest target that can honestly demonstrate repository-grounded reproduction:

1. documented inference
2. documented evaluation
3. documented training startup or partial verification
4. full training only after explicit user confirmation

Treat README guidance as the primary reproduction intent. Use repository files to clarify the README, not to silently replace it. When the README and paper conflict, record the conflict and use paper-context-resolver only for the narrow reproduction-critical gap.

Workflow

1. Read the README and nearby repo signals.
2. Run the bundled repo-intake-and-plan stage to extract commands and targets.
3. Select and justify the minimum trustworthy target.
4. Run env-and-assets-bootstrap only for target-specific environment,

checkpoint, dataset, and cache assumptions.

1. Run analyze-project only when structure, insertion points, or suspicious

implementation patterns need read-only clarification.

1. Use minimal-run-and-audit for documented inference, evaluation, smoke, or sanity execution. Keep direct execution as the default; native shell syntax requires explicit review and authorization.
2. Use run-train instead when the selected trusted target is training startup,

short-run verification, full kickoff, or resume.

1. Pause for human review before fuller training claims or any change that could

alter dataset, split, checkpoint, preprocessing, metric, loss, model semantics, or result interpretation.

1. Award result-match only when explicit expected metrics are compared under a recorded tolerance; observed metrics alone prove execution, not reproduction. Then write the standardized outputs and a concise final note in the user's language when practical.

Patch Boundary

Prefer no repository edits. If edits are needed, keep them conservative and auditable:

• Try command-line arguments, environment variables, path fixes, dependency

version fixes, or dependency-file fixes before code changes.

• Reproduction fixes are allowed when needed, but they must not be hidden. State

what changed, why it was necessary, whether it changes scientific meaning, and whether it affects comparability with the paper, README, or baseline.

• Avoid changing model architecture, core inference semantics, training logic,

loss functions, or experiment meaning.

• If repository files must change, create a branch named

repro/YYYY-MM-DD-short-task, keep verified patch commits sparse, and record README-fidelity impact in PATCHES.md.

See references/patch-policy.md.

Outputs

Always target repro_outputs/:

SUMMARY.md
COMMANDS.md
LOG.md
SCIENTIFIC_CHANGELOG.md
COMPARABILITY_REPORT.md
status.json
ANNOTATED_README.md   # original README + colored per-section agent-action annotations
PATCHES.md   # only if patches were applied

Use the templates under assets/ and the field rules in references/output-spec.md.

• Put the shortest high-value summary in SUMMARY.md.
• Put copyable commands in COMMANDS.md.
• Put process evidence, assumptions, failures, and decisions in LOG.md.
• Put scientific meaning and change effects in SCIENTIFIC_CHANGELOG.md.
• Put comparison anchors and protocol deviations in COMPARABILITY_REPORT.md.
• Put durable machine-readable state in status.json.
• Put branch, commit, validation, and README-fidelity impact in PATCHES.md when needed.
• Put the researcher's at-a-glance view in ANNOTATED_README.md: the README replayed byte-for-byte—including its image, GIF, video, and HTML markup—with exactly one marked color annotation after every heading block. Never extract a text-only surrogate. Generation must pass the built-in strip/check round trip before the file is kept.
• For original relative media/file context, use --source-adjacent-readme to also write RIGORPILOT_README.md beside the source README; inspect the reported path/status and never replace an unrelated existing file. See references/output-spec.md.
• Distinguish verified facts from inferred guesses.

Reference Loading

• Load references/language-policy.md when writing human-readable outputs.
• Load references/research-rigor-principles.md before making comparability, contribution, or research-result claims.
• Load references/deep-learning-experiment-principles.md when dataset, split, metric, checkpoint, training, or evaluation details matter.
• Consult ~/.rigorpilot/PERSONAL_RIGOR.md if present, under references/continuous-learning-policy.md (advisory only; core wins).
• Failed and later-resolved runs are auto-recorded as lessons via shared/scripts/lessonsstore.py (RIGORPILOTLESSONS=0 disables).
• Load references/research-safety-principles.md before protocol-sensitive

decisions.

• Load references/patch-policy.md before modifying repository files.
• Keep specialized logic in sub-skills, scripts, templates, or references rather

than expanding this entrypoint.