Deployments & Versions
Promote improvements into production with a measured before/after, and roll back in one click. Nothing ships without an eval result behind it.
Production is on v1.4, no candidate pending
Test an improvement on the Improvements page to create a candidate.
Deployment history
Every promotion with its real before/after.
v1.3v1.4liveM. Okafor · 5d ago
Few-shot from fire-rating corrections · eval +2.3 pts, fire-rating slice +17 pts.
Accuracy
91.8%94.1%▲2.3 pts
Cost
$0.0094$0.0089▼5.3%
p95
1980ms1840ms▼140.0ms
v1.2v1.3supersededM. Okafor · 2w ago
Citation guardrail · block compliant verdicts without current-edition citation.
Accuracy
89.1%91.8%▲2.7 pts
Cost
$0.0096$0.0094▼2.1%
p95
2010ms1980ms▼30.0ms
v1.1v1.2supersededRyshe Delivery · 4w ago
Re-indexed standards to current editions; top-k 6 + query rewrite.
Accuracy
84.6%89.1%▲4.5 pts
Cost
$0.0101$0.0096▼5.0%
p95
2120ms2010ms▼110.0ms
v1.0v1.1supersededRyshe Delivery · 6w ago
Clause-by-clause prompt + few-shot. First measured eval baseline.
Accuracy
81.2%84.6%▲3.4 pts
Cost
$0.0103$0.0101▼1.9%
p95
2160ms2120ms▼40.0ms