Quality & performance evidence

Benchmarks without invented measurements.

Quality uses the reviewed Northstar audit. Trace coverage comes from the verified Langfuse export. Historical response latency is marked unavailable because the original trace began after generation.

Baseline quality86.7%52 of 60 passed

Fixed quality93.3%+6.7 percentage points

Langfuse coverage120/120completed case observations

Timing samples—not captured in historical run

Quality by failure mode

The same cases before and after remediation.

Pass rate is shown on a shared 0–100% scale.

Out-of-scope

Before70%

After80%

Stale information

Before100%

After100%

Injection

Before100%

After100%

Ambiguity

Before70%

After90%

Numerical precision

Before100%

After100%

Tone

Before80%

After90%

Evidence coverage

What the run preserved.

Model responses120

Langfuse observations120

Semantic judgments120

Manual reviews12

Manual review is intentionally a targeted 12-case adjudication, not a missing-data condition.

Langfuse timing

Not available for this run.

Historical limitation

The original observation began after model generation, so its duration measures trace recording rather than response latency.

RetrievalTimer added for the next run
Model generationTimer added for the next run
Rule gradingTimer added for the next run
Total caseBackdated into the Langfuse observation

Next valid live run

The evaluator now records each stage with a monotonic timer and adds those durations to Langfuse metadata. The benchmark page will switch automatically from “not captured” to p50 and p95 results.

Inspect the evidence ledger