CS-OS testbed

Real customer-service data, for a real component.

The always-on testbed the CS-OS cockpit fetches from. A commit-safe real ticket corpus mapped to the closed 15-intent taxonomy, the component's actual precomputed decisions on a cross-taxonomy subset, and the live eval KPIs. Kept in its own repo so the data never mixes into the component we ship.

JSON endpoints (CORS-open, GET)

The decisions.json outputs are what the component actually returned (real Haiku + Sonnet, temperature 0), not the gold labels. They are produced by the same logic the cockpit's live /api/triage runs, so precomputed and live agree.

Licensing

Bitext is CDLA-Sharing (paraphrased), authored cases fill gaps, rjac is MIT. Every row is our own transformed text: no raw third-party rows are reproduced. The CC BY-NC-SA Twitter corpus is inspected locally only and is never included. Full audit trail in PROVENANCE.md.