What problem does it solve? Research experiments often lose their commands, metrics, and logs between sessions, making results impossible to rerun, verify, or hand off. This Skill enforces a contract-first experiment protocol so every run leaves behind reproducible commands, budgets, metrics, and a verifiable evidence pack. ## Core Features & Use Cases - Experiment Contracts: Declares hypothesis, baseline, metric direction, early/final budgets, success thresholds, and failure interpretations before execution in a JSON contract file. - Evidence Pack Lifecycle: Initializes, finalizes, and strictly verifies evidence packs via the ds_lite_evidence.py CLI, capturing stdout, stderr, metrics.json, and sanitized environment descriptions. - Long-Task and tmux Governance: Plans user-provisioned tmux capacity, verifies server fingerprints, and reconciles external task attempts so long-running GPU or cluster jobs survive session handoffs. - Use Case: A researcher comparing two model variants defines the contract, runs the baseline and candidate under a budget cap, finalizes the evidence pack, and hands the verified manifest to the review skill so the next session can audit exactly what ran. ## Quick Start Use ds-lite-experiment to define the contract for my baseline comparison, run it under a fixed budget, and package the metrics and logs as a verified evidence pack.
