How zhangcy122/openjev works
OpenJevPro is a Python decision engine for AI agents, exposing typed categorical, binary, and ordinal decisions. An agent submits context and candidates to the client, which scores choices through an inference backend and returns calibrated results with confidence and abstention information. Optional paths add deliberation and crystallization for ambiguous queries, or guarded local/cloud routing with circuit-breaker fallback. The README also documents local Laya inference and comparative evaluation; those harness capabilities are represented, while source excerpts do not establish every internal connection. Evidence is limited to the README and sampled source.
Main entry points: client.py Decision Client · gateway.py Hybrid Gateway · flywheel.py Flywheel Orchestrator · mock.py Offline Mock Engine
Components 19
Decision API
- Decision Client
openjevpro/client.py - Typed Results
openjevpro/schemas.py - Inference Backend
- Temperature Calibration
openjevpro/calibrator.py - Offline Mock Engine
openjevpro/mock.py
Deliberation
- Flywheel Orchestrator
openjevpro/flywheel.py - Fast Decision Engine
- Reasoning Model
- Crystallization Operator
openjevpro/flywheel.py - Crystallization Store
openjevpro/flywheel.py
Hybrid Routing
- Hybrid Gateway
openjevpro/gateway.py - Local Engine
- Cloud Engine
- Safety Guard
openjevpro/guard.py - Circuit Breaker
openjevpro/gateway.py
Engines And Evaluation
- Evaluation Harness
openjevpro/harness.py - Laya Engine
openjevpro/harness.py - Laya Inference Service
Other components
- AI Agent
Connections 19
- AI Agent → Decision Client: submits decision
- Decision Client → Inference Backend: requests scores
- Inference Backend → Decision Client: returns scores
- Decision Client → Typed Results: returns typed result
- Decision Client → Offline Mock Engine: uses offline mode
- Offline Mock Engine → Temperature Calibration: calibrates probabilities
- Decision Client → Flywheel Orchestrator: creates flywheel
- Flywheel Orchestrator → Fast Decision Engine: evaluates fast path
- Flywheel Orchestrator → Reasoning Model: escalates ambiguity
- Flywheel Orchestrator → Crystallization Operator: distills exploration
- Crystallization Operator → Crystallization Store: updates criteria
- Flywheel Orchestrator → Crystallization Store: merges criteria
- Hybrid Gateway → Local Engine: evaluates locally
- Hybrid Gateway → Safety Guard: guards probabilities
- Hybrid Gateway → Circuit Breaker: checks circuit
- Hybrid Gateway → Cloud Engine: escalates queries
- Hybrid Gateway → Local Engine: falls back
- Evaluation Harness → Laya Engine: evaluates with engine
- Laya Engine → Laya Inference Service: requests inference