examplePublic example environment
Synthetic data · no credentials, API, or model requests.

LLM Eval Control Plane

Release evidence

Compare a candidate against an immutable baseline and trace every policy decision to bounded evidence.

Projectpublic-example
Offline deterministic evaluationSynthetic latency and usage

Gates passed

3 / 41 gate requires review

Release decision

BlockedPolicy prevented regression

Newly failing cases

1Selected gate projection

Selected gate coverage

8 / 8Scored / attempted cases

Decision · 10 Nov 2026 · 19:03 UTC · release/offline

Release blocked

Blocked by 1 of 4 configured gates.

Baselinefake/release · revision 1Candidatefake/release · revision 2Datasetrelease-gate/offline · revision 1

First review

Refusal correctnesssafety.refusal_correct

Inspect the blocked gate, then follow its case-level score transitions.

Slice lens

Delta is always candidate − baseline

Decision evidence

Gate ledger

4 shown · failed first

Bounded aggregates

Distribution comparison

safety.refusal_correct · safety/refusal
Score and candidate-minus-baseline distribution
StatisticBaselineCandidateDelta
Coverage8/88/88/8
MEAN1.0000.875-0.125
P501.0001.0000.000
P951.0001.0000.000
Operational distribution; no case-level values
MeasurementBaseline p50 / p95Candidate p50 / p95Coverage
LatencySuppressed for privacy / Suppressed for privacySuppressed for privacy / Suppressed for privacy8/8 ↔ 8/8
Input unitsSuppressed for privacy / Suppressed for privacySuppressed for privacy / Suppressed for privacy8/8 ↔ 8/8
Output unitsSuppressed for privacy / Suppressed for privacySuppressed for privacy / Suppressed for privacy8/8 ↔ 8/8
Total unitsSuppressed for privacy / Suppressed for privacySuppressed for privacy / Suppressed for privacy8/8 ↔ 8/8

Small score samples are descriptive only. Operational quantiles below 20 measurements are suppressed for privacy. Counts remain visible; raw samples are never transferred.