| name | replicate |
| description | Scope, plan, and coach a paper replication project, producing a replication report under research/replication-projects/. Use for "/replicate <paper>", "help me reproduce this result", or reviewing replication progress. |
Replication Coach
Coach a replication of the paper named in the arguments, from scoping through write-up. The deliverable is a project directory under research/replication-projects/ containing code, an experiment log, and a report built from templates/replication-report.md.
Phase A — Scope before code
- Convert the paper into a reproducibility checklist: required artifacts (code/data/weights), expected metrics with tolerances, hardware/compute budget, and explicit risk points (undocumented preprocessing, tuning sensitivity, randomness).
- Pick the minimum decisive experiment — the smallest run that tests the paper's central claim. Replicating every table is Phase-2 scope creep; one core result reproduced cleanly beats five reproduced sloppily.
- Define success and failure criteria numerically before running anything (e.g. "within 1.5 points of reported accuracy on the same split"). Write these into the report skeleton now so they can't drift after results arrive.
Phase B — Execution coaching
- Insist on: pinned dependencies, seeded runs, config files over hardcoded constants, and an append-only experiment log (date, commit, config, result, surprise).
- When results diverge from the paper, work the standard diagnosis ladder with the user: data/preprocessing diff → metric definition diff → hyperparameters → implementation bug → the paper is wrong. Most divergences live in the first two rungs.
- A failed replication with a documented cause is a successful project. Never massage toward the published number; document the gap and the best hypothesis for it.
Phase C — Write-up
- Fill completely, including negative results and abandoned paths.
