trace.

Fraud detection.

Ready 09:30:00

0/ 124

cases routed

0 in flight

0—

auto-cleared

0

open for review

By priority

0

escalated

Model conflicts

—

median time to route

Clear — Review —
About this data
  • 124 synthetic test cases with recorded model outputs.
  • Timing is simulated at measured medians: 128 ms per decision on four workers, 1,535 ms per SFT verdict on one worker.
  • Scores are not calibrated for kinds of cases the model has not seen.

Live routing

Scoring Clear Review Audit sample

Decision scores P(review), log scale

SFT risk levels

Top reasons

    Review queue

    Priority Case Type Order value Signals P(review) Waiting Status

    Cases appear here as the replay routes them.

    How it works

    How it was built →
    1. 01

      Decision model

      Qwen3.5-4B with a LoRA adapter. Scores ok against review in one pass.

      Median latency
      128 ms
      Test AUROC
      1.0
      Workers
      4
    2. 02

      SFT model

      Same base, second adapter. Returns a risk level and reason codes.

      Median latency
      1,535 ms
      Test match
      124 / 124
      Workers
      1
    3. 03

      Routing policy

      Clear below 2% P(review). Audit 2% anyway. Escalate conflicts. Never auto-reject.

      Policy
      research-v1
      Audit
      2%
      Runtime
      Elixir