limen.experiment
Define manifests and run basic or adaptive parameter search.
Canonical docs
What this package owns
Owns Manifest, MLManifest, RuleBasedManifest, the Universal Experiment Loop, parameter-search strategies, mutable search state, reducers, feedback, and checkpointing.
Does not own inference reconstruction (limen.inference), model architectures, indicators, features, or metric helpers.
Key entry points
| Entry point | Use case | Notes |
|---|---|---|
Manifest, MLManifest, RuleBasedManifest, CalibrationConfig | Declarative experiment pipelines and calibration configuration | Exported at the package root |
UniversalExperimentLoop | Run an SFD across parameter permutations | Exported at the package root |
GridStrategy, RandomStrategy, STRATEGY_REGISTRY | Built-in advanced-search strategy surface | Exported at the package root |
Adjacent modules
limen.sfdsupplies the experiment configuration that this package runs.limen.datasupplies raw data when a manifest declares a data source.limen.loganalyzes completed runs.limen.cohortbuilds decoder cohorts on top of finished experiment results.limen.inferenceownsTrainer,Sensor, andReconstructionError; the top-levellimenpackage lazily re-exports those names.
Quick orientation
experiment/
├── experiment_core.py # UniversalExperimentLoop
├── manifest_core.py # Manifest and builder components
├── checkpoint_manager.py # Persist and restore advanced-run state
├── feedback_controller.py # Mid-run adaptive callbacks
├── msq.py # Mutable Search Queue
├── param_domain.py # Mutable parameter domain
├── reducer/
│ └── pruning_strategy.py # Pruning interfaces and implementations
└── param_search/ # SearchStrategy, GridStrategy, RandomStrategy
Things to know
- Manifest-driven SFDs run prep every round:
prep_each_roundauto-resolves toTrueat its default, and explicitFalseis rejected. - The basic
run()path uses legacyParamSpace. The advanced path addsSearchStrategy,ParamDomain,MSQ, checkpoints, and feedback hooks. - When
experiment_diris set, Limen writes all major artifacts under one directory, including checkpoint, audit, round-data, and results files. - Results are appended incrementally during a run, so experiment files should be treated as durable artifacts rather than temporary output.