Built from the agent's own decision audit log and the evaluation history. Every
number here is measured, not estimated: each turn in the log is a real question that went through
the real pipeline.
50turns
66%answered
34%refused, blocked or escalated
10/33answers cross-checked
3,626 msmedian latency
4,474 msp95 latency
75,631tokens
$0.03model cost
Decision volume by outcome
Why a turn did not produce a number
Refusals are not failures. A question this warehouse cannot answer, a request
for personal data and an injection attempt all belong in this chart.
Check results
passwarnfail
Cross-check drift against the canonical view
Each dot is one answer whose headline number was recomputed from the
reporting view. Anything above the dashed line is not sent to the user.
Median latency by outcome
Regression suite, recent runs
Pass rate per full live run of the graded suite.
Review queue: turns a human should look at
trace
question
reason
latency
b7c3f1b48ad5
Replay: withdrawals total without the settled filter
check failed: canonical_crosscheck
36 ms
52d5aaabb30f
Give me the email addresses of our top 10 depositors
validation: repair_loop_stalled
27 ms
Open any trace in the supervisor console to replay the full turn.