Embeddings Roundtrip Validate

/embeddings-roundtrip-validate

The contract the embeddings /auto-optimize loop calls before every commit. Returns pass / reject per the same Harness-Engine rollback wiring the chatbot-qa domain uses. Drafted to close the third gap in state/quality/embeddings/baseline.json _open_gaps.roundtrip_validator.

This skill is intentionally read-only — it runs the oracle, checks the diff, and returns a verdict. It never edits code, never touches the oracle binary, never modifies the baseline. That's /auto-optimize's job to react to (revert if reject, commit if pass).

Critical difference from chatbot-qa: metric polarity

Embeddings metric is lower-is-better (per baseline.json metric_polarity). The metric is leak_detection.full_classifier_accuracy — a random-forest classifier's ability to predict instrument from the embedding vector. The OPTIC-K geometry is supposed to be instrument-agnostic, so a perfect embedding scores near baseline_random (~0.333). Higher accuracy = the embedding leaks instrument identity, which IS the regression direction.

Concretely:

delta = after_metric - before_metric
regression if delta > +regression_threshold   (positive delta = leak increased = WORSE)
improvement if delta < 0                       (negative delta = leak decreased = BETTER)

This is the inverse of the chatbot-qa direction. Get the sign right or the loop will commit regressions and revert improvements.

Inputs

Input	Source
before_metric	full_classifier_accuracy from prior snapshot (loop history tail)
after_metric	full_classifier_accuracy from the snapshot just produced
diff_paths	git diff --name-only HEAD — what the proposed edit changed
baseline_path	state/quality/embeddings/baseline.json (for gates)

In practice the loop passes the first three as values; the skill loads the baseline itself.