# E-023 — The Phase 1 gate refuted the local latency budget

- Date: 2026-08-03
- Status: complete executable calibration; latency budget not approved.

## Trajectory and evidence

Round `R-F1-CALIBRATION-GATE` unified the golden corpus, the detector benchmark, and the CLI process benchmark into `npm run calibration:gate`. The runner executes three independent passes and writes `synthetic/results/f1-calibration-gate-v1.json`, even when the final decision is `fail`.

The frozen corpus passed all three runs: 100 of 100 scenarios, with zero failures. However, the frozen budget was not met: `syntax-js` exceeded the 200 ms p95 across independent passes, and the cold CLI process exceeded the 400 ms p95 by a wide margin. In the versioned run, the CLI p95 values were 1303.558 ms, 2328.511 ms, and 2701.709 ms.

## Verdict and limits

The result is a local refutation of the Phase 1 budget, not a refutation of the detectors: the corpus only covers known rules, and the benchmark measures this machine and this Node version. The gate was not relaxed or relabeled as warm. The artifact retains thresholds, passes, and the `fail` decision so that a future optimization can be compared against the same rule.

There is no causal implication about agents, task quality, or token savings. This episode only demonstrates that the calibration instrument can already reject an unsupported performance claim.

## Follow-up

Separate process spawn cost from detector cost, implement or evaluate a persistent-process architecture for hooks, and repeat the same gate without modifying the corpus or the thresholds.

## Subsequent correction — pilot decision

On 2026-08-03, it was decided to temporarily accept this unapproved budget for local use and dogfooding. The exception is a product decision, not a measurement one: hooks remain fail-open, and the `calibration:gate` command continues to return `fail`. It does not authorize declaring Phase 1 promoted, nor using the performance numbers as a commercial or scientific claim.

The review condition is explicit: before promoting Phase 1, or if perceived latency forces hooks to be disabled during the pilot, E-023 is revisited and the same frozen gate is repeated.