What problem does it solve? After an ML experiment is audited, the diagnostic findings sit in a report without a clear path to action. This Skill reads the audit digest and turns each flagged check into a concrete, deduplicated backlog row so the next experiment idea comes from evidence rather than guesswork. ## Core Features & Use Cases - Digest-driven sourcing: Reads scratch/audit/<stem>/audit.md as text and converts every issue or tip row in the ## Checks summary into one backlog candidate, without re-opening the skore Project. - Documentation-grounded mitigations: Follows each check's documentation_url via WebFetch so the proposed experiment idea reflects the official skore remediation guidance rather than model memory. - Deduplication and safe fallbacks: Drops candidates whose audit:<stem>:checks.<code> citation already exists in JOURNAL.md, and returns an explicit empty result when the digest is clean or unreadable. - Use Case: After the 01_baseline classifier audit flags an SKD003 check, ask the agent to mine the report; it returns backlog rows citing the check plus a one-paragraph summary, and the parent skill writes them into the journal for you to promote via B<N>. ## Quick Start Ask the agent to mine the latest skore audit digest and fill the experiment backlog from its diagnostic findings.
