recallgate

Overview ¶

Command recallgate runs the P8 recall regression gate (issue #162): it measures recall@10 / nDCG@10 of the production fusion semantics (position- only RRF k=60) over the golden deterministic corpus and compares against the pinned thresholds.

Fully offline and deterministic: no network, no live embedding provider, no wall-clock influence on the verdict. Exit code 0 = thresholds met, 1 = regression, 2 = operational failure (bad thresholds, harness error).

Flags:

-thresholds path   thresholds JSON (default: embedded, must equal
                   bench/recall/thresholds.json — enforced by a test)
-json path         optional path to write the full JSON report; when it
                   matches a checked-in golden file, the report is the
                   baseline artifact (bench/recall/golden-report.json)