Evals & ops

Every extractor change runs against a fixed labeled review set before shipping. The gate optimizes for grounding and compliance precision, not headline F1 or cost.

Pipeline runs, 8 weeks

695

321 conversations processed

Total model spend

$3.54

$0.011 per call

Current intent F1

0.87

extract@v2.3 on 240 labeled calls

Unsupported-claim rate

2.1%

claims without a matching evidence span

Regression gate: why the cheaper model was rejected

The v3.0 candidate beat production on intent F1 and cost, and still failed the gate: its unsupported compliance claims nearly tripled. Grounding is the binding constraint because compliance flags trigger coaching and audits.

Metricextract@v2.3 (production)extract@v3.0 (candidate)Verdict
Intent F10.870.89candidate better
Evidence grounding0.930.86production better
Unsupported compliance claims2.1%5.8%production better, gating
Cost per call$0.014$0.010candidate better
Latency p501800ms1100mscandidate better

Decision: reject v3.0. Follow-up candidate v3.1 adds a mandatory critic gate on compliance and escalation fields and is in an expanded 500-call eval now.

Eval history

extract@v3.1candidate2026-07-06 · 240 labeled calls
Intent F1
0.88
Grounding
0.91
Unsupported
2.6%
Critic disagree
8.0%
Cost / call
$0.012
Latency p50
1300ms

Small model with mandatory critic gate on compliance and escalation fields. Close to v2.3 quality at 14% lower cost; running expanded 500-call eval before decision.

extract@v3.0rejected2026-06-29 · 240 labeled calls
Intent F1
0.89
Grounding
0.86
Unsupported
5.8%
Critic disagree
14.0%
Cost / call
$0.010
Latency p50
1100ms

REJECTED despite better intent F1 and 29% lower cost: unsupported compliance claims nearly tripled (2.1% to 5.8%). Compliance precision is the binding constraint, not cost.

extract@v2.3shipped2026-06-08 · 240 labeled calls
Intent F1
0.87
Grounding
0.93
Unsupported
2.1%
Critic disagree
6.0%
Cost / call
$0.014
Latency p50
1800ms

Added mandatory evidence-span citation to the prompt; unsupported claims fell by more than half.

extract@v1.0shipped2026-05-18 · 240 labeled calls
Intent F1
0.81
Grounding
0.88
Unsupported
4.9%
Critic disagree
11.0%
Cost / call
$0.019
Latency p50
2900ms

Initial production model over the 240-call labeled review set.

Pipeline cost & latency by agent

Pipeline agentRunsAvg latencyRetriedTotal cost
ConversationExtractor3211764ms10$3.21
EvidenceValidator531034ms0$0.33
TranscriptNormalizer32134ms0$0.00