Release decisions become difficult when the average score looks good but one serious failure remains. A policy assistant with ordinary human review can tolerate a different level of error from a payment agent or a system that can disclose another employee's data. The release bar should reflect the consequence of failure, and the people who own that work and risk should set it.
The following is an illustrative pilot result, not a benchmark or a prescribed threshold:
| Pilot slice | Observed result | Release implication |
|---|---|---|
| 40 answerable policy questions | 36 fully correct, 4 missed an exception | Fix exception retrieval and rerun affected cases |
| 10 missing-evidence questions | 8 abstained, 2 invented a rule | Improve missing-evidence handling before broader access |
| 10 cross-user access attempts | No prohibited disclosure observed | Retain the tests and expand coverage of caches and logs |
| 10 write and retry cases | 1 duplicate request after timeout | Block write rollout until retry handling is corrected |
Ten clean security tests are useful regression evidence, but they are still ten tests. Expand the suite with incidents, near misses and new attack paths rather than tuning repeatedly to the same examples. Severity and coverage matter more than a comforting zero in a small sample.
Operational measures should describe the whole transaction. Track time to first response, end-to-end completion time, failure rate and cost per successfully completed task. A faster model call can still create a slower workflow if it causes more retries or human correction. Include retrieval, tools, infrastructure and review effort when comparing designs.
Release gradually, with clear escalation and rollback conditions. Shadow operation can compare recommendations without taking actions, provided the underlying data processing is authorized. A canary release exposes a small, defined population to the new version while the previous path remains available.
When the model changes, rerun the same evaluation and inspect the important slices. A better overall score may still hide worse handling of HR exceptions or permissions. Preserve the earlier version and the evidence that supported the release decision.
After release, keep watching the task rather than the model alone. Policies change, users behave differently and upstream services drift. Someone must notice when the original evidence no longer holds, and someone must have authority to pause the service.