Run started 2026-10-18 08:50 UTC
· completed 09:00 UTC
Headline Metrics
A case passes when (Correctness or String in Response) and Relevance and Tools pass.
Scorers that had nothing to judge are ignored; cases where no scorer produced a result are
counted as no data and excluded from every pass rate.
First Token Time
Time to the first response token (TTFT) — how long before the user sees anything.
mean
90.03s
p50
87.29s
p90
116.40s
p99
132.65s
min
34.18s
max
134.46s
std
28.79s
First Insight Time
Time to the agent's first update, such as a tool call or a reasoning step.
mean
6.17s
p50
4.78s
p90
9.09s
p99
10.29s
min
3.05s
max
10.42s
std
2.48s
Case Duration
Wall-clock time the eval harness measured for the whole case, including scoring and gRPC overhead. Always at least the agent's own time_total.
mean
103.07s
p50
95.74s
p90
144.03s
p99
146.79s
min
39.77s
max
147.10s
std
33.50s
LLM Time
Total time inside LLM spans, summed across every LLM call in the case.
mean
57.00s
p50
57.89s
p90
75.16s
p99
83.72s
min
28.72s
max
84.67s
std
17.47s
Tool Time
Total time executing tools — database queries, API calls and the like.
mean
38.30s
p50
49.29s
p90
57.49s
p99
66.76s
min
9.32s
max
67.79s
std
20.97s
Other Time
Agent time outside the LLM and tools: orchestration overhead.
mean
0.37s
p50
0.00s
p90
1.35s
p99
2.12s
min
0.00s
max
2.21s
std
0.79s
LLM Time Share
LLM time as a percentage of the agent's total span duration. This is the spread of per-case percentages, not the run-level ratio.
mean
63.3%
p50
61.5%
p90
81.1%
p99
82.3%
min
40.1%
max
82.5%
std
15.9%
Tool Time Share
Tool time as a percentage of the agent's total span duration. This is the spread of per-case percentages, not the run-level ratio.
mean
38.4%
p50
38.9%
p90
52.6%
p99
56.8%
min
22.2%
max
57.3%
std
13.4%
Tool Calls
Number of tool spans the agent produced for the case.
mean
11.33
p50
13.00
p90
13.20
p99
13.92
min
7.00
max
14.00
std
2.45
sum
102.00
Total Tokens
LLM tokens consumed across every call in the case (prompt plus completion).