Jev

What it's good at

• Strong zero-shot: works on a new question without any training.
• Better calibrated out of the box than Laya on the published typed-decisions benchmark (ECE 0.144 vs 0.213).
• Nothing to host. No GPU, no model files, no cold starts.
• Large label sets hold up: 0.870 on Banking77's 72 labels in a published comparison.

Limits and caveats

• Closed weights, so no fine-tuning and no running it on your own hardware.
• Your data goes to a third-party API.
• The 193.6× faster and 444.6× cheaper headline comes from workflows TypeSafe's own team built. TypeSafe itself says the gains likely sit at the high end of real-world results.
• “Can't hallucinate” means it can't answer outside your schema. It can still be confidently wrong inside it.

Directory entries using Jev