Preset

Eval lab · eval_scored · version 3

Thresholds

Quality barLatency p95 (ms)Max cost / request (USD)

Eval backfill follows a cost-first policy.
A served response has to score at least 0.80 and finish inside 86,400,000 ms at p95.
Anything priced above $0.0100 a request is outside this policy.
Right-size is off, so Parezen keeps the model the app asked for.
Cascade is off, so there is no cheap-first attempt.
If the chosen model fails, fallbacks run in order: google/gemini-2.5-flash.
Output is capped at 400 tokens and must be structured, with reasoning effort none.
The route is marked async, so eligible work can take the provider's batch price.
Exact-match response cache is opt-in and on for this route.
The near-duplicate (semantic) cache runs only on routes you list as safe for it.
Automation may apply changes this policy already allows.

This sample has no cascade escalations. When one happens, its cost is included, not hidden.
Exact-match cache is on. This sample has no exact-match hits yet, so the projection does not invent one.
The near-duplicate (semantic) cache is left out of this projection; it runs only on routes you list.