Manual-review reapplication repeat offendersRisk & Fraud
How each model performed on this task. Every chart states its N, and the reward carries a 95% confidence interval once there is more than one trial to measure.
Core metrics
Check Pass %
Mean share of the task's checks passing, averaged over trials. N = 10 trials per model.
98%
GPT-5.5
78%
Claude Opus 4.8
76%
Gemini 3.1 Pro
32%
Gemini 3.5 Flash
Extended metrics
Median Time
Wall-clock minutes, median across trials. N = 10 trials per model.
12m
Gemini 3.5 Flash
4m
GPT-5.5
3m
Gemini 3.1 Pro
2m
Claude Opus 4.8
Tool Calls
Tool invocations, median across trials — counted from the trajectories. N = 10 trials per model.
38
Gemini 3.5 Flash
30
GPT-5.5
19
Claude Opus 4.8
18
Gemini 3.1 Pro
Verifier × model pass rate
How often each verifier passed, per model. Green is passing, red is failing — a red band across a row is a verifier nothing clears.
| Verifier | GPT-5.5n=10 | Claude Opus 4.8n=10 | Gemini 3.1 Pron=10 | Gemini 3.5 Flashn=10 |
|---|---|---|---|---|
| checks (13) | ||||
| answer_cohort_severity_metrics | 100% | 90% | 100% | 50% |
| answer_repeat_offender_count | 90% | 10% | 40% | 50% |
| deliverable_jira_ticket_filed | 100% | 100% | 40% | 20% |
| deliverable_slack_channel_exists | 100% | 100% | 100% | 20% |
| deliverable_slack_message_content | 100% | 90% | 100% | 20% |
| answer_cohort_manual_review_apps | 90% | 10% | 40% | 50% |
| answer_total_outstanding_doc_requests | 90% | 10% | 40% | 50% |
| floor_all_connectors | 100% | 100% | 100% | 20% |
| floor_read_source_ticket | 100% | 100% | 100% | 100% |
| floor_sql_investigation | 100% | 100% | 100% | 100% |
| floor_write_back | 100% | 100% | 100% | 20% |
| trace_all_connectors_and_trap | 90% | 10% | 40% | 20% |
| trace_person_level_cohort_method | 20% | 0% | 0% | 0% |