Research & data

An agent reports onboarding done; the judge verifies access and equipment claims.

A vendor asks for 30 minutes with no agenda; the judge picks the disposition.

A note-taking app silently drops edits on flaky networks; the judge grades severity.

A crash report with steps and logs; the judge routes it and checks reproducibility.

A ticket reports checkout failing for all EU customers; the judge assigns a priority level.

A five-star review with oddly generic praise; the judge screens it.

A demo request from a funded startup CTO; the judge tiers the lead.

A top account cut seats 40% and went quiet; the judge tiers churn risk.

A customer asks why they were charged twice; the judge routes the intent.

An agent redlines a SaaS renewal; the judge checks the liability cap and grades the redline.

A PR adds auth caching but skips token expiry; the judge issues the review verdict.

A draft newsletter announces a launch; the judge grades it and checks the call to action.

A panel writes feedback on a backend candidate; the judge reads the signal and checks evidence.

A discovery call ends with vague promises; the judge classifies the close and grades technique.

A payment support transcript should only ever show the last four digits; the judge checks for leaks.

A heated but non-violent forum post; the judge issues a safety verdict.

A supplement listing makes disease claims; the judge rules and checks recall of the exact claim.

A startup apologizes for a 4-hour outage; the judge scores the apology.

An agent drafts a day-1 onboarding email; the judge scores it.

An agent turns a rambling sync into notes; the judge scores usefulness.

Checkout errors spike at 2am; the judge decides the response and grades the reasoning.

A terse reply after a refund lands; the judge labels its sentiment.

A third overdue notice goes out; the judge checks the tone stays professional.

An agent adapts a paella recipe for a shellfish allergy; the judge checks safety and classifies the swap.

An agent answers from the docs about SSO; the judge scores grounding and checks the plan gate.

An agent writes a couch-to-5K plan; the judge grades it and checks rest days.

An agent plans 48 hours in Lisbon; the judge checks timing and grades the plan.

A full refund transcript; the judge checks policy compliance and grades empathy.

An agent compresses standup chatter; the judge checks blockers survived and grades brevity.

An agent hands a case to the next shift; the judge checks facts and grades clarity.

Draft release notes for a mobile update; the judge classifies and grades them.

A user asks for a password reset over chat; the judge checks the agent did not leak or over-share.

A customer disputes a double charge; the judge checks acknowledgment and routes it.

A damaged-item chat; the judge checks the remedy and routes any follow-up.

TypeSafe launched Jev mid-September 2026. It makes decisions, cannot write a single sentence, and is 20 to 200x faster and 40 to 400x cheaper than frontier models according to…

TypeSafe AI (Jev) decisions for Pi: batched Choice/Score/Noul evaluation tool, terminal playground, and a typed API other extensions build on. ... Install `pi-typesafe` from npm…

LLMs Are Built to Write. Jev Is Built to Decide. | by Dr. Fadi Shaar | AI Mindset | Sep, 2026 | Medium ... # LLMs Are Built to Write. Jev Is Built to Decide. ... Inside Jev,…

The simplest way to write decision workflows in Python, backed by TypeSafe's Jev model ... classifier, decision, jev, routing, transpiler, triage, typesafe ... ## Description…

Jev: a decision-model experiment (TH) ... # Jev: a decision-model experiment (TH) ... 1. ### Jev ฉลาดพอไหม? - บททดสอบความสามารถของโมเดล "ตัดสินใจ" (ตอนที่ 1) ... Jev คืออะไร?…

That is almost exactly the territory TypeSafe claims Jev is built for. ... The narrowness of the test is important here because it relates exactly to what TypeSafe claims Jev is…