Insurance submission triage: four extraction prompts over the 27-case golden set, two attribution judges validated against 54 hand labels. https://github.com/kyle-bartlett/exchange-triage @ main
A suite names its cases, its variants and its binary checks. The baseline and candidate it declares are what the gate compares.
No suite registered yet. The agent registers weigh/*.suite.ts from the repo the first time it pulls.
Newest first. A running row fills in live on its own page.
| run | suite | status | pairs | gate | pass-all | tokens | requested |
|---|---|---|---|---|---|---|---|
| 2026-10-19-prompts | exchange-triage | finished | 108/108 | pass | 26% to 48% | +6.2% | 2026-10-20 19:20 |