Question 1

Is the zero-error result real, or a simulation that cannot fail?

Accepted Answer

The runs are simulated decisions, not clinical outcomes, and the scenarios were written by the author to have clear answers. What the simulation can fail is the protocol: two TLA+ configurations are deliberately kept as invariant violations so the model checker still has something to catch, and the severe-chaos tier does produce failures of a different kind — 46 rounds escalated instead of being decided.

Question 2

What does the stability gate actually buy me?

Accepted Answer

It keeps only votes whose margin exceeds the oracle's measured noise, so five agents cannot decide on a difference smaller than the model's own jitter. The author shows the same set of votes deciding under the identity floor and escalating under the paraphrase floor, and reports one underdetermined record where the kernel escalated 86 times out of 120 and decided 34 — the gate filters noise, it does not tell you the question was unanswerable.

Question 3

What does one Jev call cost here, and does batching cost latency?

Accepted Answer

The README does not quote a price. It does characterise the billing meter as deterministic and linear in input characters — input_tokens = 0.151 × chars + 281, with zero variance across 110 repeated requests — and reports latency as flat in question count, around 97 ms for one question and 98 ms for 38, so asking five paraphrases at once does not cost you time.

Question 4

Do I need the live API and the whole toolchain to try this?

Accepted Answer

For the full verification chain, yes: a TypeSafe key, Java plus tla2tools.jar for TLC, Node for AsyncAPI validation, and cargo for the Rust tests and the simulation example. The film pipeline is heavier, wanting ffmpeg, Playwright Chromium and an ElevenLabs key for narration. The kernel itself does not depend on the story directory.

Question 5

What does the TLA+ layer actually prove?

Accepted Answer

Four spec modules across 11 configurations, and the quorum bound was derived rather than assumed: a 24-configuration TLC sweep confirmed safe iff 2Q > N and Q > 2f. Two configs are kept as deliberate failures so the invariants keep their teeth, and the reported run at 5 agents, quorum 3 and 2 crashes explored 1,049,750 distinct states with 0 errors.

Question 6

Is the oracle itself deterministic?

Accepted Answer

No, and the repo shows it rather than hiding it. Over 1,490 hash-verified calls to jev-1.13.0, the identical request returned 0.03, 0.03, 0.03, 0.04, 0.04. The measured noise floors are 0.042 for identity, 0.059 for reordering the questions and 0.073 for the paraphrase cohort, which is what the stability gate is calibrated against.

Question 7

How do I reproduce it, and under what license?

Accepted Answer

MIT, with the ElevenLabs narration voice documented separately. The chain is ./consensus/verify.sh, running TLC, AsyncAPI validation and cargo test in about three minutes, then a Python audit and a cargo example run. The pharmacy scenarios are synthetic and written to be unambiguous: the claims are about the protocol under chaos, not about clinical judgment.

Jev Labs

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community