Update from Vitals World

20 Sep — a model that "cannot hallucinate" launched this week; our ward already couldn't

TypeSafe AI's Jev returns typed decisions with calibrated probabilities instead of prose. Honest comparison: Jev's guarantee is about shape — it can only answer from declared options, and can still pick the wrong one. Vitals World's core has no model: a deterministic physiology engine, every shift a tape, replayed from the chain to the same state — by our ward, or by anyone who runs the open engine — and checkable against the chain. Nothing to hallucinate because nothing is guessing.

The one place a Jev-style typed check would help us is the guard on the patient's voice, which today matches phrases and could miss a paraphrased slip. Plan and evaluation bar written (recall on paraphrase, false-positive rate, calibration, latency). Not built — step 0 is a person reading fifty replies by eye.

Built today, staging revisions 00062 → 00067:

• Boot 41.6 s → 3.3 s: the tape repair ran before the port opened.

• Ward pass 100 s → ~2 s quiet: it listed every patient's chain history three times per pass against a rate-limited public RPC. Now: ask the chain's own shift count first; list only what moved; share the result; budget the pass.

• Three bugs found by reading: a boot that could delete the only copy of a tape it was about to restore; a refresh gate one panic would have left stuck for the life of the process; a pass that trusted a half-read history for a chain write. Each fixed with a test.

Lesson, three times in two days: a rule that lives only in a comment has a shelf life. The fix each time was a test, not a better sentence.

Devnet · free · replayable by anyone — world.vitals.academy