limen.inference

Canonical docs

Trainer

In the research projects

The Python project promotes evaluated fitted objects without fitting them again. The Compiled project uses limen assess to reconstruct selected recorded rounds and qualified folds through Trainer. The Sensor and Cohort references distinguish retained fitted state from reconstruction under a recorded definition.

What this package owns

Owns the reconstruction-and-inference path: replaying a selected permutation from a completed experiment, validating it against the original log, and wrapping it as a live Sensor. Also owns convenient Sensor.from_model and advanced Sensor.from_python construction without a Manifest. Defaults handle ordinary fitted prediction; optional functions retain custom preparation, prediction and output control. Caller declarations are distinct from compiled evidence. Does not own experiment search execution (owned by limen.experiment) or the reference model implementations it reuses (owned by limen.sfd.reference_architecture).

Key entry points

Entry point	Use case	Notes
`Trainer`	Replay selected permutations from a completed YAML experiment	Re-runs `prepare_data()`/`run_model()` with the original split and compares metrics to the log
`Sensor`	Native inference over compiled or live Python fitted state	`from_model` wraps fitted models in one line; `from_python` accepts arbitrary state/history/callbacks
`BarPrediction`	Result of a single-bar prediction	Carries `prediction`, `probability`, and a `reason`
`ReconstructionError`	Raised when metric validation deviates beyond tolerance	Signals a permutation could not be faithfully rebuilt

Adjacent modules

limen.experiment produces the experiment logs and artifacts Trainer reconstructs. It does not re-export inference names; import them from limen.inference or top-level limen.
limen.sfd.reference_architecture.ReferenceModel is the compiled Sensor model contract; Python construction does not require inheritance.
limen.yaml.compiler.CompiledSFD supplies the compiled manifest used during reconstruction.

Quick orientation

Module or family	Responsibility
`trainer.py`	Trainer: reconstruct and validate a permutation
`sensor.py`	Sensor entry points and BarPrediction re-export
`_python_sensor.py`	Python construction contract and shared result
`_sensor_from_model.py`	Fitted-model defaults and optional hooks
`errors.py`	ReconstructionError

text · README.md

1inference/2├── README.md3├── __init__.py4├── _fold_metrics.py5├── _fold_sensor.py6├── _fold_trainer.py7├── _python_sensor.py8├── _sensor_from_model.py9├── _trainer_records.py10├── errors.py11├── sensor.py12└── trainer.py

Merged source PR · ea0b4cc8Package layout · Limen 6.0.1 · current package validation · package source

Things to know

Compiled rule rounds return native selected RuleBasedStrategy Sensors. Their raw-history, feature-row, temporal and causality semantics are documented once in Trainer. Qualified folds retain their existing opt-in evidence requirements.

Trainer validates by re-running the manifest and comparing metrics to the original log; a deviation past tolerance raises ReconstructionError rather than returning a mismatched model.
Metric comparison uses a float tolerance (with a wider tolerance for stochastic models), and a fixed set of bookkeeping columns (id, execution_time, …) is skipped during comparison.
BarPrediction.reason explains a None prediction ('warm-up', 'inside-training-window', 'null-features', 'sensor-error') — a null prediction is expected during warm-up, not an error.
Python callback/adapter failures raise; invalid aligned values become sensor-error. Use Cohort.from_python(selected_sensors) for live composition; compiled set_members rejects Python-origin members. See the Python contract and recorded workflow.