How to Use This Demo

The ecosystem is gated. Sign in as the owner for full access or as the in-scope viewer to browse. This app is viewer-visible for the catalog, prompts and run log; promoting a version is owner-only.

OPERATE shows the dual-mode model catalog (family, context window, mode badge, active/candidate/retired status). Select a prompt to see its versions side by side — bodies, eval scores, and the active marker. Run eval on the selected prompt for a fixture-harness score with the mode badge. The run log lists recent runs with prompt, model, mode, latency, eval and ok.

Owners promote or roll back the active version, inline or from CONFIGURE. Viewers see the controls disabled. The Audit tab records each promotion.

Run eval is the metered step. Flip the mode switch and open "Behind the scenes" to see the trace (load · evaluate), the mode badge, and the cost. Cloud runs `claude-haiku-4-5` cost-capped and fail-closed; OSS is recorded at $0. The catalog, prompt bodies and run log are deterministic and free. Instrumentation is inspector-only.