Blog · 8 min

Give your if-statements probabilities

RLCD trains Jev probabilities to match empirical rates. You set thresholds in code for auto-act, review, and block.

calibrationconfidenceroutingharness

An LLM confidence score matches no measured rate. Jev probabilities match empirical rates. Higher confidence means higher accuracy. You branch on the number. You auto-act above a threshold and send low scores to review. You keep the policy in code. Jev supplies calibrated inputs to it. See typesafe.ai/blog/introducing-system-one-models-and-jev and langchain.com/blog/building-a-harness-with-jev.

LangChain shows the pattern in code. ModelRouterMiddleware has Jev pick the cheapest capable model per request from the latest user message. It picks fast Luna for lookups and powerful Sol for architecture calls. AutoModeMiddleware checks tool calls for risky decisions before they run. Coding harnesses kept that classifier step closed-source. LangChain includes it for your agent. You keep probabilities and confidence in agent state for later branching. You keep the planner and tool execution in the harness. You give Jev the decision boundary.

Calibration tuner · Noul

P(true) is 0.88, so the zone is auto-act

RLCD trains probabilities to match empirical rates. 0.88 proves correct in about 88 of 100 runs. You keep the threshold in code. Auto-act at 0.85 and above. Human review from 0.50 to 0.85. Block below 0.50. 0.50 means equal odds.

Production cookbooks use one rule. Auto-act at 0.85 and above. Human review from 0.50 to 0.85. Block below 0.50. Tighten the gates for irreversible actions. You write one policy per consequence level instead of one threshold per prompt. When you reprioritize, you change a coefficient in code instead of rewriting a prompt.

computeFit simulator · same engine as the evaluator

Atomicity: 1.00

Decision-centric: 1.00

Bounded output: 1.00

Generation penalty: 0.00

native. Fit is 0.95. Bands are 0.72 and above, 0.45 to 0.71, 0.25 to 0.44, and below 0.25.

VERDICT: NATIVEprimitive · choiceconfidence · 0.90

Native. Pure Jev decision layer

fit 0.95 · band ≥0.72

Fitness signals

atomicity 1.00 · splits100%
decision-centric 1.00 · acts100%
bounded-output 1.00 · fixed answers100%
text-state 1.00 · fits in text100%
latency-pressure 0.80 · high volume80%
generation-penalty 0.00 · prose cost0%
reasoning-penalty 0.10 · reasoning cost10%
precision-penalty 0.00 · precision cost0%

Noul 0.5 means equal odds.

Economics

100ms vs 2,500ms (25x)

$0.004 vs $0.50 per 100k

Estimated savings ≈ $5.95 per year at 100k per month

You send high-throughput judgment to Jev. You call the frontier model for prose.

Jev evaluation request lifecycle — One request in order. The browser submits text. The server wraps it in { problem_statement }. The Gateway routes typesafe-ai/jev. Choice, Score, and Noul run in parallel in about 100ms. computeFit composes the verdict. You get a typed response with an LZ-compressed playground URL. Ten rubric questions run in one pass.

This simulator runs the same computeFit as the evaluator. Weights: atomicity 0.30, decision-centric 0.25, bounded output 0.20, text-state 0.15, latency pressure 0.10. Generation, reasoning, or precision penalties discount the total up to 80%, with hard overrides when deterministic code suffices or when the input is not text. Drag generation up. The verdict drops from native to hybrid. Jev answers ten literal workflow properties. You apply the verdict in TypeScript code.

Try it on your own problem.

Paste the fuzzy version and get a verdict: Jev answers the typed questions, your code computes the fit.

Evaluate your problem