668 warnings, and what each one became
Every warned event has exactly one outcome, and the four below add back to the total: 275 + 222 + 111 + 60 = 668 of 668 warned. A funnel whose stages do not tie has an unknown denominator, and every rate hanging off it is then unfalsifiable.22% in week 5 to 11% in week 10
Overridden ÷ warned, per week, both counted off event fields. The fall of 11 points dates from the v3 tuning in week 6 — the version that moved no cell — and the latest week sits under the 15% target. Weeks before the first warn rung had no warnings to override, so the series starts where the ladder did.The tuning changed no policy cell — it changed the detectors, and the registry version is recorded on the policy version that shipped it. That is what makes the fall attributable rather than coincidental: nothing about the rung moved on the same day.
Three adoption rates, three definitions, three sets of counts
Two of these read 41% and one reads 68%, and they are not the same measure: redirect acceptance is the latest two weeks, the safe transform is one role group over the window, and the pilot-to-date redirect share is every warned event of all ten weeks. The first review round found the 68% printed over the third one’s counts. Each card below carries the numerator and the denominator its own rate was divided out of, with the window inside the label — a percentage on its own is a claim.Every figure above counts events, not people. The funnel counts submissions — a person who is warned twice is warned twice — and the first build called those counts “users”, which is what let a two-week rate over warned events be read as a share of staff.
2 behavioural signals in the log
Repeat attempts and after-hours activity, each written by the detector that fired, each carrying the events it read and its own threshold. Every one of these detectors is asserted silent against a healthy twin of this bank, because a false accusation about a named person discredits every true alarm beside it.V. Sokolov (IT contractor, IT) submitted one API Keys / Creds value to Claude.ai (personal) in weeks 5, 6, 8 — 3 attempts inside 30 days, decided monitor then block.
Attempts3First to last21 daysThreshold3The 3 events behind itL. Tran (Operations specialist, Retail & Branch Ops) made an upload to Gemini Advanced at 23:10 on 2026-08-08, outside the bank's working time. The submission was held while classification ran.
Events outside working hours1WeekWeek 9Open the chainCompliance detection testing
§4.2.3’s authorized-exception path. An override by an exempt group is a policy outcome, not a policy bypass: the group is named on the decision, the rule that softened it is written down, and every use is logged and reviewable — which is the difference between an exception and a hole.Coverage testing needs to reach the detector it is testing. The group is warned rather than blocked, and every use is logged and reviewed — an exception that is a control, not a hole.
| Who | Week | Group | Justification, as written to the event | |
|---|---|---|---|---|
| R. Castellano | Week 7 | compliance-testing | “Authorized detection-coverage test CT-2291. Decommissioned sandbox credential, no customer data and no live secret. Reviewed with the IT director.” | Chain |
68% acceptance, and the version it justifies
The recommendation is computed rather than asserted: the case is the funnel, and the proposal beside it is what the case is for. Nothing here advances anything — v6 is a proposal, and it says so wherever it is drawn.Redirect acceptance: 68% — 126 redirected of 185 warned events, weeks 9 and 10. Safe-transform adoption, Lending: 41% — 55 transformed of 135 warned Lending events, trailing 30 days. Neither of those is the third rate on this page: Redirect share, pilot to date: 41% — 275 redirected of 668 warned events, weeks 1 to 10. The case for v6 is the first of the three, and the window it was counted over is in its own label.
Redirect acceptance68%redirected (weeks 9 and 10)126warned events (weeks 9 and 10)185Proposed: advance Retail & Branch Ops to warn on member PII and SSN / tax ID. Not in force, and unapproved — the approver is not the proposer, and has not signed.
The stage definitions above — what counts as redirected, when a warning becomes abandoned — are this demo’s, recorded in the product definition rather than blessed by an institution. A pilot makes them for real, and a funnel whose definitions are hidden is a funnel nobody can argue with.
Every stage above drills to the events behind it, and every one of those events carries the same provenance chain the examiner reads: the overridden warnings, in the log .