ai-research-reproduction

The saved excerpt is a snapshot from when this skill was listed. The external source remains the complete and most current version.

---
name: ai-research-reproduction
description: Rigor Reproduce compatible skill slug for README-first deep learning repository reproduction. Use when the user wants an end-to-end, minimal-trustworthy flow that reads the repository first, selects the smallest documented inference or evaluation target, coordinates intake, setup, trusted execution, optional trusted training, optional repository analysis, and optional paper-gap resolution, enforces conservative patch rules, records evidence assumptions deviations and human decision points, and writes the standardized `repro_outputs/` bundle. Do not use for paper summary, generic environment setup, isolated repo scanning, standalone command execution, silent protocol changes, score chasing, or broad research assistance outside repository-grounded reproduction.
---

# 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.

For first-use problems, run `scripts/doctor.py` with the intended Python (read-only; optional `--repo` and `--require-module`).
The deterministic entrypoint is `scripts/orchestrate_repro.py` with a self-contained `_bundled/` runtime, so this skill works when installed alone;
separately installed companion skills remain optional reusable entrypoints.
Use the entrypoint and `--help` for routine runs; inspect its implementation when a concrete blocker or safety question requires it.
Executed commands persist lifecycle state, append-only events, and full streamed
stdout/stderr under `repro_outputs/_runtime/<run_id>/`. 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/run_agent.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.
5. Run `analyze-project` only when structure, insertion points, or suspicious
   implementation patterns need read-only clarification.
6. 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 authori