jev-workflow
Toolkit for Jev decision contracts: compiled policies, egress controls, tracing, stability checks and certified decision SLOs.
Overview
jev-workflow provides evidence-bound decision SLOs, contracts, adversarial testing, flight recording, temporal stability and privacy controls for Jev. Its premise is that the useful contract is not that the model usually answers correctly, but that exact operations happen in code, data leaving the process is declared and inspectable, irrelevant input changes do not silently change an action, every decision can be traced and replayed, noisy repeated decisions cannot make an actuator flap, uncertain results lead to a bounded fallback, and automated actions consume a finite error budget backed by held-out evidence.
A YAML policy compiles into a fingerprinted artifact. The certify command learns an action threshold on a calibration split of labelled historical decisions and evaluates it once on a fixed holdout, issuing a checksummed certificate only if a one-sided Clopper-Pearson risk bound, a coverage target and per-slice checks (with Bonferroni correction) pass. At runtime, input goes through preflight and an egress policy, Jev, bounded routing and a certificate gate that permits or sends to review. RiskCard renders certificates as HTML reports and badges in French, English or Spanish. The README calls it an early release, tested but not independently security-audited.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Other | Application-defined actions (e.g. auto_approve, approval_required) gated by a certified threshold on Jev's score | — |
How it works
- Compile a YAML policy into a fingerprinted artifact
- Certify a threshold on calibration and holdout data with risk bounds
- At runtime: preflight, egress policy, Jev, bounded routing
- Certificate gate permits automation or sends to review; decisions recorded
Strengths
- Statistical risk bounds before automating
- Egress controls declare what data leaves
- Replayable traces and anti-flapping stability
- Multilingual RiskCard reports
Caveats
- Early release, not security-audited
- Requires labelled historical decisions for certification
Getting started
npm run demo:riskcard (synthetic, no Jev or network).At a glance
- Type
- Project
- Status
- beta
- Language
- TypeScript
- Runtime
- Node CLI / library
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Decision contracts, adversarial testing, tracing, stability, and privacy…
wh000wh000/awesome-jev-live
Frequently asked questions
What does certify do?
Learns a threshold on calibration data and issues a certificate only if a Clopper-Pearson risk bound and coverage target pass on a holdout.
Is the fingerprint a signature?
No, the README says it is an integrity checksum, not a cryptographic signature.
Can I try it without Jev?
Yes, the RiskCard demo produces synthetic reports with no network.
Related projects
Nautilus Compass agent memory layer whose jev-trust middleware logs Jev decisions and measures calibration in your domain.
ServicebetaPythonDecision-forensics library that explains a Jev answer by ablating parts of the input and measuring how far the probability moves.
ProjectbetaPythonRecords the full distribution of every Jev answer by state, question and model, so thresholds can move over history with no new calls.
LibrarybetaPythonMonitoring proxy and client that records Jev answer probabilities and alerts on version changes, drift, edge crowding or accuracy drops.
ProxybetaPythonTLA+-verified consensus kernel around Jev that may escalate but never returns a confident wrong verdict, tested on 1,680 pharmacy cases.
ProjectresearchRustOffline Python toolkit that measures sensitivity, Fisher information and information loss in the probabilities of Jev-style agent systems.
ProjectresearchPython