JournalDAY 51 / X

FIELD NOTE / X

A QA agent without a veto is a report generator.

The complete written thought and the evidence behind it. The video edition will follow its public release.

Journal September 25, 2026 · X target November 17, 2026

A QA agent without a veto is a report generator.

Video caption

A QA agent without a veto is a report generator. Pass, fail and unverified are different. An exception must expire. #EricFieldNotes

Full written post / accessibility read

If the coding worker can merge while the QA agent says evidence is missing, the QA agent has no operational authority. Let it search for failures, but let a separately controlled gate stop promotion when required evidence fails or never arrives. A human may override only through a named, bounded decision.

A test that ran and found no violation can pass. A detected violation fails. A missing, stale or empty report is unverified. Argo Rollouts models successful, failed and inconclusive analysis as distinct paths; the policy still needs meaningful metrics, cohort and owner. Do not turn absence into zero failures.

Run required checks outside the worker's edit authority, on the exact commit and build digest. Require fresh, non-empty results and a negative control that fails when the entitlement rule is removed. GitHub rulesets can require an expected check source; that prevents a convenient replacement badge, not a weak test.

My rule: QA can stop a release for failed or absent required evidence; only a named owner may record scope, reason, compensation and expiry for an exception. Do this because otherwise agent-generated confidence becomes permission by default, and a missing test silently becomes green.

#EricFieldNotes

Four-beat scene transcript

1. A QA agent without a veto is a report generator.

If the coding worker can merge while the QA agent says evidence is missing, the QA agent has no operational authority. Let it search for failures, but let a separately controlled gate stop promotion when required evidence fails or never arrives. A human may override only through a named, bounded decision.

Visual: A red finding must change the release path.

2. Separate three outcomes.

A test that ran and found no violation can pass. A detected violation fails. A missing, stale or empty report is unverified. Argo Rollouts models successful, failed and inconclusive analysis as distinct paths; the policy still needs meaningful metrics, cohort and owner. Do not turn absence into zero failures.

Visual: Pass, fail and unverified are different.

3. Bind the veto to the artifact.

Run required checks outside the worker's edit authority, on the exact commit and build digest. Require fresh, non-empty results and a negative control that fails when the entitlement rule is removed. GitHub rulesets can require an expected check source; that prevents a convenient replacement badge, not a weak test.

Visual: A green badge needs trustworthy provenance.

4. Give QA teeth and overrides a receipt.

My rule: QA can stop a release for failed or absent required evidence; only a named owner may record scope, reason, compensation and expiry for an exception. Do this because otherwise agent-generated confidence becomes permission by default, and a missing test silently becomes green.

Visual: An exception must expire.

Research and claim limits

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.

More notes from the work ↗