What problem does it solve? Reproducing a machine learning paper's results end to end involves many fragile steps — scoping experiments, downloading datasets, building environments, validating runs, launching full training, and auditing outcomes — and losing track of any one of them invalidates the whole effort. This Skill drives that entire pipeline in order, gating each stage on the previous stage's written report so the work survives context loss and can be verified by a human afterward. ## Core Features & Use Cases - Sequential stage orchestration: Invokes experiment-scoping, resource-download, environment-setup, run-validation, run-experiment, baseline-reproduction, result-analysis, and reproduction-audit one at a time, never concurrently. - Report-based state tracking: Reads existing reports under .paper-reproduction/ to resume where a previous run stopped instead of re-running completed stages. - Cost gating and honest stopping: Pauses for user approval before GPU-intensive full runs and baseline reproduction, and always runs the final audit even when an earlier stage fails. - Use Case: You clone a paper's repository, drop the PDF inside, and say "reproduce this paper" — the pipeline scopes the experiments, fetches assets, builds the conda environment, validates with mock runs, launches full runs, compares metrics against the paper, and delivers an audit verdict. ## Quick Start Clone the paper's codebase, place the paper's PDF inside it, and ask the agent to reproduce this paper end to end.
