jev-labs
TLA+-verified consensus kernel around Jev that may escalate but never returns a confident wrong verdict, tested on 1,680 pharmacy cases.
Overview
jev-labs wraps Jev in a consensus kernel and tests it the way you would test a database: model-checked in TLA+, contract-generated into Rust, and run through 1,680 simulated pharmacy decisions (for example whether a drug is controlled or conflicts with a documented allergy) under seeded chaos against the live API. The invariant is not 'always right' but that the kernel may escalate to a human and may never return a confident wrong verdict. It consults repeatedly, decides only when at least three stable votes agree, and escalates when the budget is spent without a stable quorum.
The README reports zero wrong verdicts in 1,080 golden rounds across no, realistic and severe chaos, bounding the true violation rate below 0.28 percent at 95 percent confidence, while noting this does not prove it is zero. Severe chaos raised escalations from 5.0 to 18.0 percent, the intended direction. The TLC model check explored about 1.05 million states with no errors. A four-minute film reads its numbers from captured data.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this drug a controlled substance?(inferred) | — |
| Yes / no | Does this prescription conflict with a documented allergy?(inferred) | — |
How it works
- Build state from a prescription record
- Consult Jev repeatedly across agents
- Decide on three stable agreeing votes; otherwise escalate when budget is spent
- Verify invariants with TLA+ and chaos testing
Strengths
- Formal verification of the protocol
- No confident wrong verdicts observed
- Escalation increases as evidence degrades
Caveats
- Simulated pharmacy cases
- Zero errors bounds but does not prove the rate
At a glance
- Type
- Project
- Status
- research
- Language
- Rust
- Runtime
- Consensus kernel with TLA+ specs
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Never confidently wrong: a TLA+-verified consensus kernel around TypeSafe's Jev, run through 1,680 chaos-tested pharmacy decisions with zero wrong verdicts. Film, code, and every captured call.
logicrw/awesome-jev-projectsNever confidently wrong: a TLA+-verified consensus kernel around TypeSafe's Jev, run through 1,680 chaos-tested pharmacy decisions with zero wrong verdicts. Film, code, and every captured call.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Pharmacy safety questions, aggregated by a consensus kernel.
Can it be wrong?
It may escalate, but in 1,080 golden rounds it returned no confident wrong verdict.
How is it verified?
Model-checked in TLA+ and chaos-tested against the live API.
Related projects
Medplum demo where a bot uses Jev to compare a provider's visit note with a discharge summary and flag agreement or potential conflict.
AppprototypeTypeScriptToolkit for Jev decision contracts: compiled policies, egress controls, tracing, stability checks and certified decision SLOs.
ProjectbetaTypeScriptPre-registered study testing Jev's calibration claims on clinical diagnosis cases against Claude Opus 5 and Opus 5.5.
ProjectresearchPythonSpanish oncology decision-support prototype that walks ESMO guideline trees and uses Jev only for judgment calls.
ProjectresearchPythonResearch implementation of Jev-style typed decisions over clinical evidence: supported, contradicted or unresolved.
ProjectresearchPythonTwo frozen open models answer yes/no per exam option; their agreement drives an auto-release gate and a conformal candidate set.
ServiceresearchPython