How mithalouni/system-one-open works
System One is an open Jev-style decision model: callers provide state and typed questions, and receive calibrated choice, score, or yes/no results from one model pass. The repository also includes dataset preparation, model training, evaluation, live demos, and game-policy examples. The graph follows documented flows and sampled call sites; several demo and evaluation internals are represented at their indexed module paths because their full implementations were not sampled.
Main entry points: serve.py HTTP decision service · engine.py Decision inference · train.py Model training · evaluate.py Model evaluation
Components 18
Decision Data
- Public task loaders
s1/data_real.py - Synthetic task generators
s1/data_synth.py - Dataset mixture builder
build_data.py
Model Runtime
- Typed prompt schema
s1/schema.py - Decision inference
s1/engine.py
Serving & Demos
- HTTP decision service
serve.py - Demo presets
s1/demo_presets.py - Main demo page
s1/demo_page.py - Race, email, viral demos
s1/demo_pages2.py - Drive, home demos
s1/demo_pages3.py - Demo question sets
s1/demo_race.py - Game policy runner
games.py - Doom policy simulator
s1/doom_sim.py
Training & Evaluation
- Model training
train.py - Model evaluation
evaluate.py - Calibration metrics
s1/metrics.py - TypeSafe benchmark data
typesafe_eval.py
Other components
- API caller
Connections 25
- Public task loaders → Typed prompt schema: creates examples
s1/data_real.py - Synthetic task generators → Typed prompt schema: creates examples
s1/data_synth.py - Public task loaders → Dataset mixture builder: supplies tasks
build_data.py - Synthetic task generators → Dataset mixture builder: supplies tasks
build_data.py - Dataset mixture builder → Model training: prepares mixture
README.md - Decision inference → Typed prompt schema: builds prompts
s1/engine.py - API caller → HTTP decision service: sends request
serve.py - HTTP decision service → Decision inference: calls decide
serve.py - HTTP decision service → Demo presets: loads presets
serve.py - HTTP decision service → Main demo page: serves page
serve.py - HTTP decision service → Race, email, viral demos: serves pages
serve.py - HTTP decision service → Drive, home demos: serves pages
serve.py - HTTP decision service → Demo question sets: loads questions
serve.py - Race, email, viral demos → Drive, home demos: shares styles
s1/demo_pages3.py - Race, email, viral demos → HTTP decision service: posts race request
s1/demo_pages2.py - Drive, home demos → HTTP decision service: posts decisions
s1/demo_pages3.py - Game policy runner → Decision inference: requests decisions
games.py - Doom policy simulator → Synthetic task generators: uses task states
s1/doom_sim.py - Model training → Typed prompt schema: loads examples
train.py - Model training → Decision inference: trains model
train.py - Model training → Calibration metrics: measures calibration
train.py - Model evaluation → Decision inference: scores decisions
evaluate.py - Model evaluation → Calibration metrics: computes metrics
evaluate.py - Model evaluation → Doom policy simulator: evaluates policy
evaluate.py - TypeSafe benchmark data → Model evaluation: provides benchmark
README.md