Ai Research Reproduction

You point this at a deep learning repo and it tries to reproduce the smallest documented target without changing anything important. It reads the README first, picks inference or eval over training when possible, records every assumption and deviation in a structured output bundle, and pauses before doing anything that might alter scientific meaning. The workflow is opinionated: bootstrap only what you need, patch conservatively, write evidence to repro_outputs/, and treat the README as ground truth even when the paper says something different. If you want to chase SOTA scores or experiment freely, use something else. This is for auditable, minimal-trust reproduction runs where you need to know exactly what changed and why.

Install to Claude Code

npx -y skills add lllllllama/rigorpilot-skills --skill ai-research-reproduction --agent claude-code