JournalDAY 51 / INSTAGRAM

FIELD NOTE / INSTAGRAM

Missing evidence must never paint the dashboard green.

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

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

Missing evidence must never paint the dashboard green.

Video caption

Missing evidence must never paint the dashboard green.

Pass, fail and unverified.

The worker must not grade its own permission.

Record who can override it and for how long.

#EricFieldNotes

Full written post / accessibility read

An agent says the feature is ready. The suite it ran is green, but the required customer-entitlement postcondition did not run. If your dashboard treats no result as zero failures, it will approve the exact question nobody answered.

Pass means fresh evidence for the exact build and required cohort. Fail means an observed contradiction. Unverified means skipped test, stale report, absent target-state readback or too little canary exposure. Each requires a different action; unverified should pause for a named review, not disappear from the denominator.

Store the entitlement fixture and expected tenant outcomes beyond the coding agent's write scope. Run it on the exact artifact and require a non-empty report. Deliberately remove the entitlement guard in an isolated mutant; a correct check must fail before the real release earns a green status.

The rule: a QA agent may find evidence, but a separately owned gate enforces fail or unverified states. An accountable person can authorize a bounded exception with reason, cohort, expiry and rollback. Do this because a warning without release authority is only decoration.

#EricFieldNotes

Four-beat scene transcript

1. Missing evidence must never paint the dashboard green.

An agent says the feature is ready. The suite it ran is green, but the required customer-entitlement postcondition did not run. If your dashboard treats no result as zero failures, it will approve the exact question nobody answered.

Visual: A blank report is a separate state.

2. Use three colors with explicit meaning.

Pass means fresh evidence for the exact build and required cohort. Fail means an observed contradiction. Unverified means skipped test, stale report, absent target-state readback or too little canary exposure. Each requires a different action; unverified should pause for a named review, not disappear from the denominator.

Visual: Pass, fail and unverified.

3. Make the check independent.

Store the entitlement fixture and expected tenant outcomes beyond the coding agent's write scope. Run it on the exact artifact and require a non-empty report. Deliberately remove the entitlement guard in an isolated mutant; a correct check must fail before the real release earns a green status.

Visual: The worker must not grade its own permission.

4. Give QA a stop button.

The rule: a QA agent may find evidence, but a separately owned gate enforces fail or unverified states. An accountable person can authorize a bounded exception with reason, cohort, expiry and rollback. Do this because a warning without release authority is only decoration.

Visual: Record who can override it and for how long.

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 ↗