Jev

What Jev is

Jev is TypeSafe’s flagship System One model. It evaluates a state against typed questions and returns structured answers your code can branch on. It does not write replies, produce code, or explain its reasoning.

• One request. Every question runs in parallel, in isolation, against the same state.
• Adding questions barely changes latency. Questions do not rot each other’s context.
• Default model name jev-latest currently resolves to jev-1.13.0. Pin the version if you tune thresholds.

What it will not do

• Not a replacement for Claude Code, Cursor, Codex, or Hermes. No streaming text, no tool calls, no file edits.
• No copy, no quiz prose, no headlines. It gave up string generation.
• No images, audio, or video. Pre-process those into text first.
• Not fine-tuned per customer. Same weights for every account. Domain rules go in state, instructions, and criteria.
• English is the primary training language. Other languages work less evenly. Watch confidence.
• Noul answers have no confidence field. Choice and Score do.

The three primitives

• Choice — one option from a list you define. Returns choice, probabilities, confidence.
• Score — a position on a rubric you define. Returns score, probabilities, confidence.
• Noul — is this statement true? Returns noul from 0 to 1.

Mix all three in one call. If a question needs extended reasoning or several independent factors, split it. Combine the numbers in your code.

Confidence is not probability

probabilities is the spread across options or levels. confidence collapses that shape to 0–1 so you can threshold without doing the math. A flat spread means low confidence.

• High: act.
• Medium: confirm, or flag for review.
• Low: do not act. Human, or a reasoning model.

Thresholds scale with risk. A wrong screen is not a wrong payout. Start conservative and test on your own data. TypeSafe’s sample floors (0.5, 0.9) are examples, not a standard.

Limits that are on the models page

• Price: $42 per billion input tokens ($0.042 / million). Output tokens are free.
• Rate limits, as published: 250,000 tokens/sec and 1,200 requests/min. TypeSafe says these are moving while demand is high.
• Context: 64k tokens per request. 32k for the state plus the longest question.
• Not trained on customer requests. Zero data retention is an enterprise term, not the default.

Speed is not an SLA here. The launch post says 70–500ms end to end. The use-case map says about 150ms for real-time. Treat both as their claims, not a contract.

Fit on this stack

This block is our read of live products, not a TypeSafe claim.

• Use: Hermes skill routing and inbound guardrails. LawnLeveling lead vs spam vs not-leveling, with low confidence sent to a person. Qualitative overlay on arbitrage (same SKU, condition, policy) while fee math stays in code.
• Skip: Gut Gut Goose scoring (already deterministic). Product Ladder, 80/20, and RFM formulas. Ad and quiz copy. Cloud explorer catalogs.