Governance measures and scoring
Read the measure name before interpreting a score. Quality, risk, instrumentation, and governance rates answer different questions. Check the application, period, source, and scoring profile when comparing values.
Measures in the current model
| Measure | What it summarizes |
|---|---|
| Quality score | Outcome, coherence, error, risk exposure, duration, token efficiency, and confidence signals |
| Instrumentation score | Coverage of expected SDK fields for the application profile |
| Risk score | Recorded failure, risk, guardrail, grounding, and confidence contributors |
| Governance rates | Oversight, grounding, guardrail, and safety evidence observed in scope |
Profiles can define weights, thresholds, and expected rates for an application. Tenant preferences and platform defaults provide the remaining values.
A score summarizes recorded evidence. It is not a compliance result or a judgment about whether an AI output was correct.
Check applicability
Some fields do not apply to every application. For example, decision, risk, outcome, and oversight coverage may be marked not applicable when the application never produces AI decisions.
Confirm the application profile before treating a missing field as a gap.
Check source coverage
Scores depend on the fields your integration records. Missing adapter support, partial instrumentation, or a filtered date range can change the result.
Open the related sessions and provenance details before acting on a score change. Distinguish declared values from values derived by the platform.
Use scorecard links
Governance scorecard tiles open filtered records for the selected measure. Available tiles depend on the user’s role.
- Developers review guardrail reporting and integration evidence.
- Governance Engineers review guardrail and grounding results.
- Compliance Officers review oversight, grounding, and safety evidence.
- Administrators can review the complete configured scorecard.
Investigate a change
- Confirm the application, profile, date range, and filters.
- Open the records behind the changed measure.
- Check instrumentation coverage and provenance.
- Separate evidence gaps from recorded policy or behavior results.
- Record the owner and required follow-up.