| name | code-reproduction |
| description | Use when reproducing a paper, repo, benchmark, or reported result and you need an explicit plan for environment setup, execution, discrepancy logging, and final reproduction status. |
Purpose
Reproduce reported results without overstating success and without losing track of discrepancies.
When To Use
- implementing or rerunning a paper's released code
- checking whether a baseline can be reproduced before extending it
- validating a public repo before building new work on top of it
- preparing an artifact report or reproducibility note
Inputs
- target paper, repo, or benchmark claim
- environment requirements and hardware constraints
- available scripts, checkpoints, configs, and datasets
Outputs
- reproduction plan and execution order
- environment summary and setup log
- discrepancy report between expected and observed behavior
- reproduction status: exact, partial, failed, or blocked
- list of missing assumptions or undocumented steps
Workflow
- Extract the exact claims to reproduce before running anything.
- Reconstruct the environment and document every assumption.
- Run the minimum viable path first, then expand to full results.
- Compare observed outputs against reported numbers or behaviors.
- Log mismatches with likely causes instead of hiding them.
Constraints
- distinguish exact replication from faithful reimplementation
- never claim success without a concrete comparison target
- treat undocumented setup steps as findings, not personal fixes
- keep a clear boundary between upstream behavior and your modifications
