Jev

Jev accepts text, strings, JSON-like state, and arrays, then answers typed questions. Choice selects among options, Score places an item on a scale, and Noul returns probabilities for reusable labels. It does not generate user-facing prose, code, chain-of-thought, images, audio, or video.

• Use deterministic code to turn a probability into an action, and keep the threshold visible and testable.
• Escalate uncertain or high-impact cases to a person or a reasoning model instead of forcing a binary answer.
• Calibration is measured over groups of predictions. A 0.9 score is not proof that one specific decision is correct.
• Pin a dated model version when a calibrated threshold matters; the jev-latest alias can move.

As checked on September 22, TypeSafe documents jev-1.13.0 as the current model and jev-latest as its moving alias. The direct API is text-only and uses POST /v1/systemone. TypeSafe lists a 64,000-token overall request limit, a 32,000-token state budget plus the longest question, and a $0.042 per million input-token rate with no output-token charge.

• The rate reflects a model that returns compact probabilities, not free generated answers.
• Published rate limits are dynamic. Validate the active account limit instead of designing from a snapshot.
• TypeSafe says customer requests and responses are not used for training; zero-data-retention is documented as an enterprise option, not a default promise for every account.
• CJK text is supported, but TypeSafe describes English as the primary and best-supported language. Test your own language and label distribution.

The supplied Awesome Jev thread links fourteen public repositories. They are more useful as a map of decision-shaped jobs than as proof that every project is mature or faster than an alternative.

• Browser control: jev-ultrafast chooses browser actions and elements; agent-desktop chooses actions from an accessibility tree.
• MCP access: typesafe-mcp and jev-mcp expose Jev decisions to Claude, Codex, and other MCP-capable clients.
• Developer utilities: semdecide adds semantic choice, scoring, filtering, and guards to Unix pipelines; jev-codex-router routes coding tasks by estimated difficulty.
• Context and review: winnow filters low-value tool results; jev-review stages code-review decisions in a dashboard.
• Repository navigation: Blink performs decision-guided codebase search; neo4jev chooses graph edges while traversing Neo4j data.
• Experiments: typesafe-ai-playground explores the API, OneVOneJev drives a game agent, and safer-with-jev proxies model choices through Neon infrastructure.
• Bounded domain agents: prism-liquidity-agent presents advisory liquidity decisions rather than direct autonomous trading.

Marcel Pociot demonstrated a browser extension that asks Jev whether an X post matches a natural-language hide rule, then collapses matching posts. This is a concrete content-firewall pattern: use a fast classifier at the edge, keep the rule readable, and let the interface expose what was hidden.

• A fast demo does not prove negligible latency, cost, or false positives for every feed and language.
• Content filters need an undo or reveal path so a mistaken decision is visible and reversible.
• Browser extensions must still respect the target platform, user privacy, and extension-permission boundaries.

fast-jev-compaction uses Jev to decide which older tool calls and results should stay, be truncated, or be dropped before a coding-agent request. Its own README preserves user and assistant messages, pins the newest items, pairs calls with results, and falls back when the reduction is insufficient or a request fails. That is narrower than claims that it safely compresses all agent memory.

• The repository explicitly says probability is not proof that a tool result is safe to delete.
• Removing an error or observation can cause a later agent step to repeat work or lose the reason behind a prior decision.
• Changing the middle of a prompt can invalidate downstream prefix-cache reuse; the price impact is provider- and cache-policy-specific.
• Measure task success, repeated tool calls, retained corrections, latency, and total billed tokens together. A smaller prompt alone is not a successful compaction result.
• For durable memory, extract verified decisions and constraints into an owned record instead of treating deletion as memory formation.

Start with one reversible decision that already has labels and a measurable cost of error. Compare Jev with a deterministic rule and the current model path on the same examples. Choose thresholds only after inspecting calibration, abstentions, subgroup errors, and the behavior of the fallback path.

• Record the exact Jev version, question schema, labels, threshold, language, and evaluation set.
• Prefer ranking, routing, filtering, moderation triage, or tool selection over tasks that require a novel explanation.
• Keep consequential actions behind validation or human approval even when the probability is high.
• Re-test after an alias, prompt schema, input distribution, or downstream policy changes.