One number with provenance. The eval issues each gold-set question exactly as the recall hook does — hybrid mode, term extraction, ×2 over-fetch, the admissibility filter, temporal bounds — and scores exact-path membership at k=5 against gold-set-v3.json (64 scored questions, 20 easy negatives, 10 near-miss negatives). Baselines embed a corpus fingerprint (document count, embedded-in-scope count, gold-set hash, pin date); --compare refuses across fingerprints unless --drifted-ok says the drift is understood, and a gold-set hash mismatch is never overridable. R@5 is the product metric — the hook injects five — and R@1 rides along as the ordering signal.

Printed on every report, so nobody can pre-register a bar the instrument cannot resolve without having seen it refuse the idea:

scored n	smallest clean gain (6 flips)	Wilson 95% width near 75%
64 (today)	+9.4pp	~21pp
100	+6.0pp	~17pp
125	+4.8pp	~15pp
250	+2.4pp	~11pp

Six flips is derived from the exact paired test (two-sided binomial: 2·(½)^k < 0.05 first at k=6), not chosen. A true 5pp improvement at n=64 best-cases p=0.25 — invisible. Per-target rank probes remain the sensitive instrument for sub-MDE mechanisms: they measure movement exactly, and they are how three alias strategies were shown to move nothing at all.

A RULE file (template: scripts/health/results/RULE-TEMPLATE.md) is committed before any run, and it must contain:

1. A power-checked bar. coin_pass_probability(bar, n) ≤ 0.05, computed by the eval's own function and quoted in the file. Both of this arc's pre-flight bars — floorless rerank's ≥5-of-9 and fusion's ≥3-of-5 — sit at exactly 0.50 by this arithmetic: coins with paperwork. The floorless refutation is withdrawn as unproven because the bar was its only evidence; fusion's stands only on its deterministic per-question trace.
2. A positive control. Something in the run that fails if the instrument is dead, named in advance. The standing controls (schema assertion, planted canary, score spread) cover the eval itself; a probe touching anything else supplies its own.
3. Per-question diffs. A net "+2" once hid a +5/−3 split; totals are not results.
4. A written prediction. Checked against the outcome either way, in the same file.

The field-name null, twice. expected read where the fixture says expected_note_paths parses every expectation to an empty list, and a run reports a clean "0 of N" that is really N comparisons against nothing. It shipped one refutation and reproduced on the next rung's first attempt. The schema control now aborts it (exit 4).

The keyword-soup null. The cross-encoder rung fed both models concatenated terms rather than natural questions; the technique looked flat because the query representation was. Re-measured with real phrasing before the verdict was trusted — which is what clause 2 generalizes.

None are load-bearing here, deliberately — the scoring path is deterministic. Any future judged metric must first measure operator agreement on a sampled set and publish that number beside its own. The completeness-v1 scorer is the standing reason: its gutted-note check was green in both directions while it agreed with the operator at ρ = 0.11, because the check exercised a stage downstream of the defect.

The rules above govern the offline gold-set ladder. Live traffic adds these. Each one exists because breaking it produced a wrong number once.

rule	why	breaking it cost
No raw prompt text on disk	the judge's account of a gap quotes the request by construction	a repository holding the operator's prompts
Deterministic comparator runs and is reported apart	it fired on 7 of 3,004 injected notes — showing that is the point	a floor mistaken for a second opinion
Suppress the comparison when the floor never fires	the disagreement then equals the judged rate exactly	one number printed twice as though two signals agreed
No rate without the instrument's resolution beside it	the judge drifts ~5 points across identical runs	a precision claim the instrument cannot support
Always-valid interval, not fixed-horizon	the number is read whenever someone wonders	over 800 peeked streams, a Wald interval was wrong on 43–46% against a sequence's 0.0%
Measure at the change, paired and interleaved	the judge's ~5-point drift is common-mode and cancels in a difference	two rates a month apart, each carrying the full drift
A sample too small says so	~45 judged turns clear the bar, 274 reach ±0.10	[0%, 54%] rendered as if it were a range

κ before publication does not apply here, and the exemption is recorded rather than waived. No independent human labels are obtainable from this data — unaided ones proved error-prone, aided ones are not independent — so the online row carries a permanent UNVALIDATED marker with its reason attached, instead of a κ it cannot compute.

CI-Gates · the gate script scripts/check-retrieval-regression.sh · results record scripts/health/results/goldv3/NOTES.md · design agentm-hybrid-retrieval.