Crux Digits: Jev examined, and the plot it has not published

A technical analysis by AI consultant Tom Joseph, published 22 September 2026, examining what Jev is, what its cannot-hallucinate guarantee actually covers, and the calibration plot TypeSafe has not published.

Shape, not truth. If your schema says the answer is one of approve, review or decline, you get one of those three, with no malformed JSON or invented enum value. But a model constrained to three categories can still confidently pick the wrong one, and a well-typed wrong decision produces no error.

Within a week of launch, independent developers reproduced Jev's interface on open weights: a LoRA adapter plus a trained scalar decision head on Qwen3.5-9B, Apache-2.0 with MIT code, returning the same three primitives without autoregressive generation. The piece says this means the idea is not the moat.

Three: the parallel sampler that evaluates every question at once, the price and speed at a sub-500ms typed decision and four cents per million input tokens, and RLCD, the training method supposed to produce epistemically honest probabilities. The piece says the third is the one you cannot check.

No reliability diagram, no expected calibration error, no Brier score and no calibration paper, even though calibrated confidence is the entire pitch. The architecture and weights are undisclosed too, and the piece notes outside observers have suggested the model may derive from open weights, which TypeSafe has not addressed.

An Octomind engineer asked Jev whether to block rm -rf ~/.ssh and got a 0.76 block probability with 0.64 confidence. After a fake tool-output field claiming the command was pre-approved and instructing auto-allow, the probability fell to 0.48 and confidence to 0.22. The piece says the guard read hostile text as evidence, as designed.

Keep deterministic checks in front of consequential actions, never feed untrusted content into a classifier that gates an action, measure your own calibration by logging state, schema, option order, model version, probability and ground truth, use confidence as routing rather than permission, and start where a wrong answer is visible and reversible.