CompanyBench

Duplicate-funding R01 lock cohort auditRisk & Fraud

15 verifiers · 4 LLM-graded

EnvironmentsJiraSlackZeta3 SQL

TaxonomyRisk & FraudRisk & 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.

  1. 98%

    Claude Opus 4.8

  2. 95%

    GPT-5.5

  3. 92%

    Gemini 3.1 Pro

  4. 90%

    Gemini 3.5 Flash

Extended metrics

Median Time

Wall-clock minutes, median across trials. N = 10 trials per model.

  1. 5m

    Gemini 3.5 Flash

  2. 3m

    Gemini 3.1 Pro

  3. 2m

    Claude Opus 4.8

  4. 2m

    GPT-5.5

Tool Calls

Tool invocations, median across trials — counted from the trajectories. N = 10 trials per model.

  1. 33

    Gemini 3.5 Flash

  2. 26

    Gemini 3.1 Pro

  3. 14

    Claude Opus 4.8

  4. 12

    GPT-5.5

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.

VerifierGPT-5.5n=10Claude Opus 4.8n=10Gemini 3.1 Pron=10Gemini 3.5 Flashn=10
checks (15)
cp_bug_lockout_count80%100%100%100%
cp_bug_lockout_exposure_usd80%100%100%100%
cp_cohort_count100%100%100%100%
cp_locked_count100%100%100%100%
cp_returned_exposure_usd100%100%100%100%
cp_sunset_locked_count100%100%100%100%
deliverable_jira_ticket_filed100%100%100%80%
deliverable_slack_channel_exists100%100%100%80%
deliverable_slack_message_content100%100%100%80%
reconstructs_double_funding_lock_cohort90%80%20%60%
cp_bug_lockout_account_ids80%100%100%100%
floor_read_source_ticket100%100%100%100%
floor_write_back100%100%100%80%
process_floor100%100%100%100%
trace_all_connectors100%100%100%80%

Failure modes on this task

17 tagged verifier failures, by the stage they were assigned to. Higher is worse.

Failure stageGPT-5.5Claude Opus 4.8Gemini 3.1 ProGemini 3.5 Flash
Answer1 mode3284