jev-proof
Java tool that checks whether a sponsored video delivered its brief, with Jev judging subtitle evidence and code gating the result.
Overview
jev-proof is one of four apps in jev-suite, a set of Java decision-quality tools built on Jev with a shared kernel. It answers whether a sponsored video actually delivered the sponsor's brief, by fetching subtitles (via yt-dlp), feeding Jev sanitised evidence, and asking per-fact questions. Thresholds, vetoes and routing stay in plain Java.
The shared kernel provides a TypeSafe client with retry handling, a three-way NoulGate (confident pass, confident fail, everything else to human review), and a sanitiser that defends against prompt injection in external text. Design rules are: Jev judges and code decides; degrade toward caution (unreachable model means review, never auto-pass); calibration runners refuse mock data; and no evidence means fail closed. The suite README reports for jev-proof v0.1.0 a 60-sample calibration with the gated pipeline correct on 68 of 68 decisions, 12 designed abstentions and 0 of 10 injection flips; a list summary cites a later 90-sample run at 0.922 gated accuracy. Sibling apps cover job fit, edit fidelity and rental listings.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does the sponsored segment satisfy this acceptance rule?(inferred) | yes · no |
| Choice | How is this required fact covered in the video?(inferred) | — |
How it works
- Fetch subtitles and extract the sponsored segment
- Sanitise external text against injection
- Ask Jev a Noul and Choice per required fact
- Gate results into pass, fail or human review
- Report acceptance with evidence
Strengths
- Conservative gating with human review for uncertain cases
- Injection defence tested in calibration
- Evals refuse to run on mock data
Caveats
- Small calibration samples
- Depends on subtitle availability and yt-dlp
- Requires JDK 17+, Maven and a TypeSafe key
Getting started
With JDK 17+ and Maven (and yt-dlp for jev-proof), run mvn package in the jev-suite repository.At a glance
- Type
- Project
- Status
- beta
- Language
- Java
- Runtime
- JVM CLI
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Creator sponsorship: verifies each sponsored ad segment in video subtitles against acceptance rules with one Jev `Noul`+`Choice` call per fact while deterministic code keeps the confidence gate; 90-sample calibration reports 0.922 gated accuracy and 0/15 injection flips.
yibie/awesome-jev
Frequently asked questions
What does Jev decide?
Whether each required element of a sponsorship brief appears in the video's subtitles.
What happens if Jev is unreachable?
Results go to review; the system never auto-passes.
How accurate is it?
The suite README reports 68 of 68 gated decisions correct in a 60-sample calibration, with 12 abstentions.
Related projects
CLI and CI gate that asks Jev one yes/no question per Markdown requirement and fails the build when code drifts from the spec.
ProjectbetaTypeScriptPi package that gates pi-subagents dispatches with one Jev call against configurable rule sets and adds a jev_ask tool.
PluginbetaTypeScriptHKUST policy-tree guard with a 2B open classifier that routes each agent request or action to a leaf and classifies safety.
ProjectresearchPythonTLA+-verified consensus kernel around Jev that may escalate but never returns a confident wrong verdict, tested on 1,680 pharmacy cases.
ProjectresearchRustGovernance proxy between Jev or local Laya and agents that turns probabilities into allow, ask, deny, keep, truncate or drop actions.
MCP serverbetaRepository for a public X demo of Jev, apparently checking ads before they run; the README has only a title.
Appprototype