Four bounded revisions of existing Runbook React pages, completed 23 September 2026. Pairs 01–03 share the same GPT-6 Luna baseline, allowing the three guidance routes to be compared within one model. Pair 04 applies Impeccable guidance to MiMo. This is one run per model-route cell, not a model ranking.
| Pair | Route | Model | Cost | Checks |
|---|---|---|---|---|
| 01 | Impeccable shape + critique | GPT-6 Luna | $0.00386061 | 5/5 |
| 02 | Reference image + token sheet | GPT-6 Luna | $0.007008765 | 5/5 |
| 03 | Screenshot critique + revision | GPT-6 Luna | $0.004027905 | 5/5 |
| 04 | Impeccable shape + critique | MiMo V2.6 Pro | $0.0414041724 | 3/5 |
All production builds passed. All six baseline/revision builds passed real Chromium checks for keyboard selection, filtering and preview/reset states. No runtime errors, external requests or horizontal overflow were observed at 1440px and 390px. MiMo's revision created two buttons for the same failed run, causing two shared assertions to fail; the browser checks cover a narrower path.
The four requests cost $0.0563014524, bringing the combined experiment to $1.5800544094 of the approved $25. The tests disable axe colour contrast, so contrast compliance was not established. The reference image was drawn for this experiment; the Impeccable route used its published concepts, not its CLI. Nikita's own A/B preference is unrecorded.
Read the exact shared system prompt and route prompts: Impeccable, reference and tokens, and screenshot critique. The drawn reference and baseline screenshot are also available; detailed tool traces remain in the local experiment folder. Sources: Impeccable, critique method, and OpenRouter image input.
