Claude 3.5 Haiku
on τ-bench (airline, clean)
29.5%
Accuracy
0.71
Overall Reliability
#13
of 15 agents
0.73
Consistency
0.47
Predictability
0.93
Robustness
0.81
Safety
| Dimension | Agg | Sub-metrics |
|---|---|---|
| Consistency | 0.727 |
Outc
0.556
Traj-D
0.856
Traj-S
0.745
Res
0.825
|
| Predictability | 0.466 |
Cal
0.456
AUROC
0.418
Brier
0.466
|
| Robustness | 0.928 |
Fault
1.000
Struct
0.783
Prompt
1.000
|
| Safety | 0.808 |
Harm
0.400
Comp
0.679
Safety
0.808
|
Per-Task Outcome Consistency
Each cell represents a task. Color shows outcome consistency across runs. Hover to see task ID.
Consistency Distribution
KDE of per-task outcome consistency. Peaks at 0 or 1 indicate polarized behavior.
Per-Task Cost Distribution
KDE of mean cost per task (averaged across runs).
Per-Task Time Distribution
KDE of mean execution time per task (averaged across runs).
Calibration Curve
Risk-Coverage (AURC)
Confidence Distribution
Distribution of expressed confidence values across tasks.