What problem does it solve? Reproducing a machine learning paper requires proving that every planned experiment actually executes end-to-end before committing to expensive full-length runs. This Skill automates stage 4 of a paper-reproduction pipeline: it executes the mock commands written by earlier stages, classifies each failure, fixes code bugs one git commit at a time, and delegates environment errors to the environment-setup skill. ## Core Features & Use Cases - Error-driven fix loop: Runs each mock command in the prepared conda env, classifies failures as environment, code, or asset errors, and iterates until the run passes or a stop rule fires. - Git-tracked repairs: Creates a run-validation branch where every logical fix is a separate, revertable commit with a structured message tracing back to the error. - Honest validation: Requires exit code 0 plus evidence of real work (finite losses, logged metrics, output artifacts) before declaring a mock run passed. - Use Case: After scoping, downloading, and environment setup are done for a cloned paper repo, invoke this Skill to validate all planned experiments, producing a run-validation.md report with validated commands, a fix log, and notes for the full runs. ## Quick Start Run stage 4 of the reproduction pipeline on this repo: execute the mock commands from the scoping report and fix whatever breaks until they pass, tracking every fix in git.
