The transparency record: latest calibration snapshot per domain plus EVERY launch-gate evaluation ever run, verdicts included — failing runs are part of the audit trail, not something to hide.

A run whose report names a family in calibration.UNTRUSTED_RESOLUTION_FAMILIES is published as a run and withheld as a measurement. The row keeps when it ran, its domain, family, as-of date, horizon, forecaster version and verdict; every statistic in its report is JSON null (never 0, never an omitted key) and the row carries a withheld block with reason: "resolutions_untrusted" saying why. reasons goes with them, as a whole list, because those strings quote the same figures in prose. withheld is null on every row still published, and rows for trusted families — the rollup_threshold news-tone record included — are returned exactly as stored.

That the runs stay visible is deliberate and is the honest half of the story: we ran our own backtest, found the GH Archive series under those resolutions had collapsed, and withdrew the numbers. Deleting the rows would hide the withdrawal too.

Withheld per ROW, not per latest row: proof._gate_withheld reads only the newest evaluation (ORDER BY id DESC LIMIT 1), and the withdrawn runs are history. This matches /v1/risk, /v1/proof and the calibration page exactly — one endpoint must not publish a number the next one withholds.