Present the kitchen test
FIVE-MINUTE EVIDENCE SEGMENT / PROTOCOL V2
The Service House kitchen test compares three treatments. It currently has no completed controlled three-arm result in the inspected evidence.
Before the session
- Open the public kitchen guide. It works without the separate harness.
- Confirm the expected result from K14, K20 and K25. Know which screen is a requirement, which is calibration and which is an actual candidate result.
- If demonstrating a local reference fixture, verify it in advance and label it “reference calibration.” Do not substitute it for Naked, Lintel or Swarming.
- Use a reviewed, sanitized result export only after a real cohort exists. A stale “running” record is not current runtime evidence.
The five-minute stage path
| Time | Action | Say |
|---|---|---|
| 0:00–1:30 | Show the three treatments and fixed brief. | “Same product contract. The workflow treatment changes.” |
| 1:30–3:30 | Trace exact cents, cancel-once and reload-once. Optional: a clearly labelled calibration fixture. | “These are the behaviors we inspect in every candidate.” |
| 3:30–5:00 | Show the separate measures and current evidence status. | “The controlled comparison is not complete. There is no winner to announce.” |
What changed in v2
The controller serves the app. The candidate works in public/, follows the shared runtime, then submits once with BENCHMARK-SUBMIT.mjs. Native journal time and frozen content define the observation boundary. Candidate-owned npm start is the older runtime, not this protocol.
Keep the historical proofs separate
The saved profile, cold-recovery and regression examples remain demonstrations of specific helpers at their recorded revision. They are not a new whole-harness evaluation or results from this kitchen test.
If a live view fails
Spend at most 30 seconds recovering. Return to the public requirement and evidence guide. Do not change a score, reset a clock, silently repair a frozen candidate or launch a paid model run to fill time.