Cost and latency are quality gates
A route that answers well but blows the budget is not production-ready. Cost per successful task and p95 latency belong next to groundedness and refusal on the same scoreboard.
What to write down
- Budget per successful task (and what counts as success)
- p95 latency for the user-visible path
- What happens when the budget trips: degrade, queue, or kill switch
Re-run after model, tool, or retrieval changes. Cheap models that hallucinate more are not a saving.
Arviser packaging
Cost or latency sits in the six audit areas. Shared metrics and CI gates pages show how to keep those numbers next to quality, not in a separate spreadsheet nobody opens.
A 24h Technical PoC Spike writes p95 latency and unit cost into the pass marks before build. For an on-device latency budget example (~15 ms WASM, 8.2 MB), see Open proof.
Next: Shared metrics · CI gates · Kill switch · 24h PoC Spike · Open proof · Request scope