generate/ | situation grid, writer, diversity selection, agent runners and adapters |
score/ | conduct checks, judges, quality ranking, selection for SFT and RL, the trust and delta reports |
verify/ | verifiers: programmatic rewards (MathEqual, CodeExec, JSONSchema, …) that honor the judge contract |
ingest/ | trace loading, OpenTelemetry rows (gen_ai.usage.* sums into row["usage"]), platform push and pull |
world/ | the mock tool environment (WorldOptions in sandbox.py) |
run/ | the engine behind simulate(): knob resolution (config.py), spec loading (spec.py), row helpers (rows.py), and the scheduler itself (engine.py: inputs, build, loop, finish) |
simulation.py, data.py, export.py | the simulate() entry point, its result object, and training export |
schema.py, schemas/ | the typed row (Task, Rollout, Judgment, Marker) and the row-v1.json wire contract |
defaults.py | every default the engine and the reports use, each with the reason it is what it is |
environment.py | export_environment and load_environment: a run as an installable RL environment |
training.py | hosted training runs, training_run, serve, reward_model |
monitor.py | HackMonitor, the during-training reward-hacking watch |