world-model-harness

Iteratively improve the RubricJudge (or any LLM scorer) against a hand-labeled dataset - run the judge, identify false positives/negatives, diagnose why each failed, propose one experiment (prompt, model, or context) per failure class, and prove the fix without regressing controls. Use when the user distrusts judge scores, asks to improve/calibrate/tune the judge, or a new corpus type needs judge coverage.

Mandatory pre-merge gate for every PR. Runs /code-review --fix at an effort level scaled to the PR's breadth, resolves every review comment (Cursor, Greptile, humans), and verifies the diff complies with AGENTS.md. Use whenever the user says a PR is ready to merge, asks to merge, or invokes /ready-for-merge.