How kyegomez/open-jev works
Open Jev is a small PyTorch research implementation of typed, parallel decisions over structured program state, rather than a text-generating application. A caller supplies state and typed questions; Jev encodes state once, lets isolated query slots read the shared representation, and returns Noul, Choice, or Score answers. The same module exposes soft-target RLCD calibration and consistency training. The README documents structural embeddings, evidential confidence, and cacheable state encoding; sampled excerpts are partial, so uncertain internal call wiring is omitted. All runtime components reside in open_jev/main.py.
Main entry points: main.py Jev model
Components 17
State encoding
- Hash tokenizer
open_jev/main.py - State encoder
open_jev/main.py - Shared state cache
open_jev/main.py - Text encoder
open_jev/main.py
Typed inference
- Jev model
open_jev/main.py - Query slots
open_jev/main.py - Readout blocks
open_jev/main.py - Noul head
open_jev/main.py - Choice head
open_jev/main.py - Score head
open_jev/main.py - Confidence head
open_jev/main.py
Calibration training
- RLCD loss
open_jev/main.py
Other components
- Caller
- State and questions
- Typed answers
- Soft targets
- Augmented states
Connections 24
- Caller → State and questions: provides
- State and questions → Jev model: passes
- Jev model → Hash tokenizer: encodes text
- Jev model → State encoder: encodes state
- State encoder → Shared state cache: produces cache
- Jev model → Text encoder: encodes questions
- Jev model → Query slots: initializes slots
- Query slots → Shared state cache: cross-attends
- Jev model → Readout blocks: updates slots
- Readout blocks → Noul head: scores truth
- Readout blocks → Choice head: scores options
- Readout blocks → Score head: scores levels
- Readout blocks → Confidence head: estimates confidence
- Noul head → Typed answers: returns probability
- Choice head → Typed answers: returns distribution
- Score head → Typed answers: returns ordinal result
- Confidence head → Typed answers: returns confidence
- Typed answers → Caller: returns
- Caller → RLCD loss: evaluates model
- RLCD loss → Jev model: queries model
- Caller → Soft targets: supplies distributions
- Soft targets → RLCD loss: guides loss
- Caller → Augmented states: supplies variants
- Augmented states → RLCD loss: supports consistency