What problem does it solve? Running full-scale machine learning experiments after validation is error-prone: commands get re-derived incorrectly, failed runs trigger ad-hoc fixes that invalidate the reproduction, and crashed sessions lose the record of completed runs. This Skill scales validated mock commands back to paper-faithful settings, executes every run sequentially, and leaves an auditable report on disk. ## Core Features & Use Cases - Manifest-driven execution: Derives full commands from the validated stage-4 commands plus stage-1's downscaling notes and sweep tables, writing the complete run manifest before launching anything. - Sequential monitored runs: Launches each run through the run-visible helper in a watchable terminal, with poll cadence scaled to the estimated wall time, and records failures as findings rather than entering fix loops. - Resume and evidence tracking: Scans prior run directories to skip already-complete runs, judges completion by log and artifact evidence (not just exit code), and writes a run.json per run plus a structured run-experiment.md report. - Use Case: After stages 1-4 of a paper reproduction pipeline finish, invoke this Skill to run a 12-run hyperparameter sweep overnight; if the session dies after run 5, re-invoking it resumes from run 6 without redoing finished work. ## Quick Start Run stage 5 of the reproduction pipeline on this repo: launch the full-length validated experiments sequentially and record everything in .paper-reproduction/run-experiment.md.
