You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
One-time scanning certifies a moment, not a corpus — that is a named dead end. Rules improve, corpora get re-tiered, and sources turn out dirtier than believed, so revalidation is a recurring gate: new rules find old leaks. The cadence design has three moving parts:
Sampled human audit per round: a small random sample of each tier hand-checked, to estimate the scanners' miss rate instead of assuming it. "No detector catches everything" is every tool's own disclaimer; the audit sample turns that caveat into a number.
Allowlist expiry: allowlist entries carry reasons and expiry; a stale allowlist is how leaks get grandfathered.
Scope
This design takes #6021's recurrence recommendation (one-off vs standing gate) as its input and specifies the standing form: triggers, per-layer scope per trigger, audit sample sizes, and where the cadence plugs into the ingestion-operations design (#6019) as a standing gate.
Done looks like
A cadence spec: trigger table, scope per trigger, audit method with miss-rate estimate, allowlist lifecycle.
Not already done: no revalidation-cadence spec exists in the repo — code search for revalidation returns only unrelated hits (opencode-claimant migrations, id web assets). The standing-gate design has not been written.
Input verified present: the cited research source ~/workspace/scratch/atlas-ingestion-brief-2026-11-04/prior-art-research.md exists in the workspace (same atlas brief set used by the sibling follow-ups).
Scoped deliverable (single design pass): the cadence spec the body names — trigger table (rule / inventory-pack / recognizer-set changes), per-layer scope per trigger (deterministic layer in full; NER layer per its task(atlas): classification revalidation scan — verify nothing sensitive sits at the wrong tier #6021 scoping), audit method with sample sizes and a miss-rate estimate, and the allowlist lifecycle (reason + expiry on every entry) — recorded on this issue or in a doc/ record linked from it.
Activity
developing-today commentedon Nov 5, 2026
CODE-1716
dezren39 commentedon Nov 5, 2026
🤖 Agent note · done
Triage findings (verified fresh 2026-11-05)
revalidationreturns only unrelated hits (opencode-claimant migrations, id web assets). The standing-gate design has not been written.~/workspace/scratch/atlas-ingestion-brief-2026-11-04/prior-art-research.mdexists in the workspace (same atlas brief set used by the sibling follow-ups).doc/record linked from it.