Intent · CHG-2026-014916

change_id: CHG-2026-014916
risk_class: R3
autonomy_tier: A2
controls: ['TRC-01', 'CHG-04', 'HUM-14']
data_classification: internal
originator: olaf@freundcloud.com
agent_identity: svc-agent-platform
model_route: gateway/tier-frontier
supersedes: None

Problem

Stage 6 re-raises an ongoing excursion every night. It does not ask whether the thing it found is already being tracked; it draws the same conclusion from the same data and files it again.

The evidence is unambiguous. CHG-2026-014911, 014912 and 014915 — issues #7, #8 and #10 — each recorded 6.926 sigma on post_deploy_error_rate. Identical readings of an unchanged fixture, filed as three separate findings, each with its own change id, its own branch and its own issue. All three were dismissed together as one finding counted three times.

Both proposing paths do it, which is easy to miss: diagnose at 2 sigma and propose at 3 sigma each create an issue unconditionally.

Why this matters more than tidiness. A finding that appears every night stops being read. The response is a filter rule, and then the one night the reading changes materially, nobody notices — the detector is still perfect and the control has failed anyway. §8.5 names this as ASI09, human-agent trust exploitation: the failure is not a hostile agent, it is a competent one trusted into the background. A duplicate-generating detector manufactures that outcome directly.

It also puts a number on the cost. Left alone, this produces roughly 365 issues, 365 branches and 365 change ids a year for one unfixed problem, in a repository whose §6.2 claim is that the change id is a reliable join key.

Who is affected

• Anyone on the receiving end of Stage 6, who currently gets a duplicate every morning.
• The change id sequence, which burns an id per night per ongoing anomaly.
• The argument on the showcase site that the loop closes cleanly without a human in the invocation path. It closes; it just closes repeatedly onto the same spot.

Success criteria

1. An excursion that is already tracked by an open Stage 6 finding does not create a second one, at either tier.
2. When the tracked finding is closed and the metric breaches again, a new finding is created. Suppression must not become silence.
3. The suppression is visible: a run that suppressed something says so, and names what it deferred to.
4. The decision is deterministic. No model decides whether two findings are the same.

Out of scope

• Changing the detection bands or the fixture. The detector is right; what it does with being right is what changes.
• Auto-closing a finding when the metric recovers. That is a triage decision with a human in it, and inferring “resolved” from a metric returning to baseline is exactly the kind of conclusion this repository does not let a machine draw unsupervised.
• Deduplicating across different metrics. Two metrics breaching together are two findings, and deciding they share a cause is diagnosis, not detection.

Constraints

• Detection stays deterministic and model-free (§4, principle 4).
• Suppression must fail open. If the lookup cannot run, the finding gets raised. A duplicate is an annoyance; a suppressed real anomaly is the failure mode being introduced.