JevOps
Python kernel of Jev Choice/Score/Noul gates, caches and loops used to admit or reject Lean proof candidates in research pipelines.
Overview
JevOps is a TypeSafe/Jev kernel split out of the Lean Refactor Arena and related papers. It treats Jev as a gate, not a generator or a proof authority: optional refactoring and autoencoder modules propose Lean candidates, Jev ranks and filters them, and only the consuming Lean/Lake verifier can admit them. The verifier is injected by the caller.
The package contains many modules: Choice, Score and Noul projectors with a fixture client, a Jev beam and ranking picker, an outer action loop with stall and stop handling, an oracle for ordering candidates, layered caches with context budgets, success and failure memory, a goal/subgoal board and a guarded legal adapter. A roadmap describes a proposed workflow for formalising parts of the U.S. Code with source locks, and the README is explicit that this is a plan, not a claim of formalisation. It is a research codebase for people building verifier-backed search loops, licensed AGPL-3.0.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which candidate (e.g. Lean edit) should be tried next? (Jev beam)(inferred) | — |
| Yes / no | Should this candidate step be admitted for verification?(inferred) | — |
| Score | How promising is this candidate?(inferred) | — |
How it works
- Generator modules propose Lean candidates
- Jev projectors rank and filter candidates in a beam
- Outer loop routes the next action and detects stalls
- Caller-supplied Lean/Lake verifier admits or rejects
- Memory records successes, failures and blacklists
Strengths
- Clear separation: model gates, verifier decides
- Injectable verifier and fixture client for offline tests
- Caching and context budgeting built in
Caveats
- Research code with many experimental modules
- Legal formalisation work is a proposed plan, not a result
- Native backend qualification still pending per the README
At a glance
- Type
- Library
- Status
- research
- Language
- Python
- Runtime
- library
- Has code
- yes
- Has tests
- yes
- License
- AGPL-3.0
Use cases and integrations
How curated lists describe it
Jev is a \*\*gate\*\*, not a generator. This package does \*\*not\*\* write Lean. Lake (or another oracle) lives in the implementation that \*uses\* the kernel.
logicrw/awesome-jev-projects
Frequently asked questions
Does JevOps write proofs?
No. It gates and ranks candidates; generation is done by other modules and admission by a Lean/Lake verifier.
What is the legal roadmap?
A proposed plan to apply the kernel to source-grounded U.S. law formalisation, explicitly not a completed result.
Can I test without the API?
Yes, a FixtureClient replays Jev answers for offline tests.
Related projects
Explainable day-trip planner where an LLM drafts the itinerary and Jev screens places, compares transport and reviews each step.
AppprototypeTypeScriptResearch implementation of Jev-style typed decisions over clinical evidence: supported, contradicted or unresolved.
ProjectresearchPythonInference-time loop where a System One judge like Jev scores a model's thought paths and picks which reasoning branch to keep.
ProjectresearchPythonPython harness stress-testing Jev's confidence, its accuracy as a RAGTruth hallucination judge, and its use as a RAG reranker.
DemoresearchPython- OpenDecision★ 58
Local decision engine on zero-shot ModernBERT and SigLIP2 answering Choice, Noul, Score and Relation over text, documents and images.
ServicebetaPython Open-weight 0.6B decision model on a Qwen3 backbone: state and typed questions in, a probability for every option out, nothing decoded.
ModelresearchPython