When the Gate Fires: Calibration-Gated Agentic Discovery, and the Structure-Based Methods It Disqualified

Agentic discovery pipelines are judged by the shortlists they produce. We report a pipeline that makes refusal an explicit output through retrieval controls, a seven-module authority split, and a calibration gate for methods used to rank unmeasured compounds. Applied to a 51,479-pair neurological target–disease database, it selected SCN8A, a gain-of-function epilepsy target. Two site transfers reach AUC 0.731 and 0.577 against 27 measured actives and 41 measured inactives, with pose controls of 1.57 and 2.82 Å. SCN8A is re-evaluated retrospectively under the rule reported here; its original primary-condition pass is retained in the run history. At the declared 0.70 floor, no structural evaluation with retained scores qualifies. At 0.60, the primary SCN8A evaluation qualifies statistically; this does not independently validate the transferred receptor. On ABL1, top-ranked redocking recovers the crystallographic pose, but pooled ranking is disqualified at AUC 0.621. On a matched split, a 2D-similarity baseline reaches 0.936 versus docking's 0.602 and qualifies. Four benchmark-audit findings leave the re-evaluated SCN8A verdicts unchanged. A historical ABL1 AUC of 0.976 over 551 actives and 12 inactives is disclosed without a verifiable qualification verdict because its interval, null and scores are unavailable. The false-refusal rate remains unmeasured. We propose comparison through withdrawal logs, calibration verdicts and provenance records; the structured ledger is specified but has not recorded a live run.