reliability · T1 · grounding
Answer from context failed
rel_ground_001
Score
0%
Turns
1
1 useful · 0 useless
Interventions
0
harness hints (HHI)
Tool calls
0/0
matched expected trajectory
Errors
0/0
recovered / total
Latency
828ms
Tokens
83
0 reasoning
Cost
$0.00000
hallucination_unsupported — Claim not grounded in context (extrinsic) · expected 3, got 5
Trajectory timeline
Turn-by-turn replay · useless turns annotated with the classifier reason
- #0modelusefulresponse
Prompt / output inspector
instruction
Using ONLY the context, how many known issues does release 2.4.0 have? Reply with the number only.
context (120 chars)
Release notes — Kestrel 2.4.0 shipped on 12 May with three known issues (two cosmetic, one performance). 2.3.9 had five.
final output
5
expected
3
Spans
1 recorded · tokens & cost per span
- generate828ms82 in · 0 cached · 1 out · $0.00000 · ttft 783ms
Events
- task_started{"trial":0,"taskId":"rel_ground_001","scaffold":null}
- task_finished{"score":0,"passed":false,"taskId":"rel_ground_001","failureLabel":"hallucination_unsupported"}
Scorer metadata
{
"toolLog": [],
"scaffold": null
}Reliabilitynumericdifficulty 2/5