VERIFYBusiness workflows
Jev Traffic Accident Analysis
燐夜 Lava@orz99
Jev processed an AI-written traffic accident analysis in Traditional Chinese, generating 15 typed legal decisions from 5 findings and 3 questions with high accuracy and low cost.
Tested @typesafeai's Jev on an 11-page, AI-written traffic-accident analysis in Traditional Chinese.
5 legal findings × 3 questions = 15 typed decisions, in 1 API call:
• 831 ms end-to-end
• ~$0.0003 (7,645 input tokens; output is free)
• 15/15 answers passed schema validation
The statutes, report excerpts, questions and option criteria were all in Traditional Chinese. Where Jev said the facts were insufficient, the report had already said so itself.
Jev flagged one finding as a possible mismatch, with only 0.25 confidence. I think it's a fair catch. The §101 text the report attached is about overtaking, but the finding cites it for a lane change. That's my read, not a lawyer's. At 0.25 I'd send it to a human rather than auto-reject the report.
Next I want a human-labeled set that includes cases where the cited law clearly doesn't fit, to check whether Jev is right or just cautious. cc @CompleteSkeptic