Experiment

Experiment: the one evaluation loop (backtrader’s Cerebro).

Both trust levels share this orchestrator — the Verifier is this plus submission policy. Two execution modes over identical settlement logic:

event

reset → per origin: observe → predict → step → settle → analyze. Always correct; required for online-learning models.

vectorized

For models with supports_batch (FeaturePredictor): the feature matrix for all origins is materialized in one shot, the model predicts once as a batch, and the precomputed actions are replayed through the same environment. Orders of magnitude fewer model calls; settlement, validation and rewards are byte-identical because the env does them in both modes.

mode="auto" picks vectorized when the model supports it.

class emflow.run.experiment.Experiment(problem, model, analyzers: str | Sequence[Analyzer] | None = 'default', split: str = 'validation')

Bases: object

run(mode: str = 'auto') → Result