Comparison·support-triage-160
support-triageproduction + evals · last 7 days

How is the agent doing?

Production quality and evaluation quality are different questions. This page never blends them into one number — an agent can pass every test and still fail customers.

84.1%
Eval pass rate
+1.7 vs prompt-v7
78.9%
Production pass rate
−5.2 below evals
87.2%
Grader agreement
+15.8 since v1 rubric
19
Open disagreements
7 awaiting review
! Production is running 5.2 points below the eval set. The gap means the eval set does not yet contain the cases customers actually send — not that the agent got worse.

Recent activity

prompt-v8 evaluated against support-triage-160 — 11 regressions found12 min ago
Alert: tone clause agreement fell to 63% across the last 40 graded cases1 h ago
Rubric v4 published — added the jurisdiction clause2 days ago
smoke-12 flagged as saturated — 98.3% pass for six weeks2 days ago
M. Duarte resolved 12 disagreements in the review queue3 days ago

Needs a decision

hold11 regressions on prompt-v8Review before shipping. Four are in the retention flow.
rubricTone clause is ambiguousGraders disagree on 37% of cases that touch it.
hygienesmoke-12 no longer discriminatesRetire it or add harder cases.