How gitchw/lct works
Jev-LCT is a low-latency System-One decision engine: clients submit choice, noul, or score questions; the API delegates inference to a recurrent transformer, which returns decisions with trajectory-derived confidence. Training and evaluation support the model lifecycle, while the Snake showcase exercises decisions in a real-time game loop. README capabilities include adaptive early exit, calibration, Jev-compatible serving, benchmarks, and Snake analysis. Some internal wiring and the browser dashboard are not visible in the sampled source, so uncertain links are omitted.
Main entry points: engine.py Inference Engine · snake/server.py Game API Server · client.py Game Runner · api/server.py Jev REST API
Components 24
Decision Serving
- Python SDK
lct/api/client.py - Jev REST API
lct/api/server.py - Request Schemas
lct/api/schema.py - Inference CLI
lct/cli.py
Recurrent Inference
- Inference Engine
lct/inference/engine.py - Looped Transformer
lct/modeling/looped_transformer.py - Adaptive Early Exit
lct/modeling/early_exit.py - Candidate Readout
lct/modeling/readout.py - Confidence Trajectory
lct/modeling/trajectory.py - Confidence Mapping
lct/calibration/mapper.py - Calibration Metrics
lct/calibration/metrics.py
Training Evaluation
- Model Trainer
lct/training/trainer.py - Training Losses
lct/training/losses.py - Benchmark Data
lct/data/benchmarks.py - Benchmark Evaluator
lct/evaluation/evaluator.py - Results Visualizer
lct/evaluation/visualizer.py
Snake Showcase
- Game Runner
snake/client.py - Game Rules
snake/engine.py - Spatial Analysis
snake/analysis.py - Decision Prompt
snake/prompt.py - Model Adapter
snake/models_adapter.py - Game API Server
snake/server.py - Terminal UI
snake/ui_terminal.py
Other components
- Decision Client
Connections 24
- Decision Client → Python SDK: submits query
- Python SDK → Jev REST API: sends HTTP request
- Jev REST API → Request Schemas: validates payload
- Jev REST API → Inference Engine: dispatches inference
- Inference CLI → Inference Engine: runs inference
- Inference Engine → Looped Transformer: runs recurrent model
- Inference Engine → Adaptive Early Exit: checks loop exit
- Inference Engine → Candidate Readout: selects candidate readout
- Inference Engine → Confidence Mapping: maps confidence
- Looped Transformer → Confidence Trajectory: records loop dynamics
- Confidence Mapping → Calibration Metrics: evaluates calibration
- Model Trainer → Training Losses: computes training loss
- Model Trainer → Looped Transformer: trains model
- Benchmark Data → Benchmark Evaluator: supplies benchmark data
- Benchmark Evaluator → Inference Engine: evaluates predictions
- Benchmark Evaluator → Results Visualizer: provides results
- Game Runner → Game Rules: advances game
- Game Runner → Spatial Analysis: analyzes game state
- Game Runner → Game API Server: requests move
- Game API Server → Spatial Analysis: analyzes game state
- Game API Server → Decision Prompt: builds decision prompt
- Game API Server → Model Adapter: selects decision model
- Model Adapter → Inference Engine: runs LCT inference
- Terminal UI → Game Runner: runs game client