Establishes verified research baselines through attach, import, reproduce, or repair routes.
npx skills add https://github.com/OpenLAIR/dr-claw --skill ds-baseline
Please help me install this Agent Skill. Skill: ds-baseline Source: https://github.com/OpenLAIR/dr-claw/tree/main/skills/ds-baseline Command: npx skills add https://github.com/OpenLAIR/dr-claw --skill ds-baseline
Research experiments need a trustworthy reference point before any comparison is meaningful, but reproducing or validating baselines is error-prone, often producing fabricated metrics, unverified runs, or incomparable results. This Skill enforces a disciplined workflow for attaching, importing, reproducing, or repairing a baseline so downstream experiment and writing stages can compare against it fairly.
Ask the agent to establish and verify a baseline for the current quest by reproducing the specified paper's repository and confirming the comparison metrics.
High-intent search queries and answers about installing and using this skill.
Read the source paper and repository first, then create PLAN.md and CHECKLIST.md defining the route, command path, and acceptance condition. Set up the Python environment with uv, run one bounded smoke test, then launch the real run and verify the metrics against the paper before acceptance.
Prefer attach or import when a verified reusable baseline already exists or a baseline package is available, since reuse is cheaper than reproduction. Reproduce only when no credible baseline exists, and repair only when an existing baseline has a bounded, identifiable failure.
A baseline is comparable when the task identity, dataset version, split contract, evaluation path, metric keys, and source commit are explicit and recorded in metric_contract.json. Verification classifies the result as verified_match, verified_close, verified_diverged, or broken before acceptance.
Failures are classified as missing_source, missing_code, environment_infeasible, command_unknown, run_failed, or verification_failed. One autonomous fix per failure class is the normal bound; recurring failures should be recorded as blocked with a next-step decision rather than retried endlessly.
Yes, but only through an explicit waiver recorded via artifact.waive_baseline with a clear reason. Without a confirmed or waived baseline gate, comparison-heavy downstream stages like experiment should not proceed.
ResearAIStandardize a reference object for quest evaluation with plan and checklist requirements.
theamDesign rigorous Limina research experiments with method-validity gates and decisive thresholds.
Rycen7822Retrieve and verify arXiv/OpenReview research papers for quest evidence.
louzhengshuaiEstablish a baseline comparator and metrics for quests.
