limen.experiment.reducer

Canonical docs

Reducers And Feedback

In the research projects

The search scenarios in the Python project and the Compiled project each run a separate four-candidate grid with a three-round BudgetReducer. Their audit and checkpoint records establish the actual trim; they do not claim that pruning improves model quality.

What this package owns

Owns the reducer strategies that consume in-flight experiment results and mutate the search — trimming to budget, pruning correlated or saturated regions, and refocusing on promising areas — plus the registry and shared filter/action constants. Does not own the parameter domain or the search strategies it steers.

Key entry points

Entry point	Use case	Notes
`REDUCER_REGISTRY`	Name-to-reducer lookup	Resolves reducers named in a manifest
`BudgetReducer`	Cap the remaining permutation budget	Trim via `TRIM_RANDOM` or `TRIM_WORST_FIRST`
`CorrelationReducer`	Remove wrong-direction numeric values or suggest low-impact keeps	Uses bootstrap correlation and sign stability
`FocusReducer`	Refocus the domain on promising regions	Narrows around strong rounds
`SaturationReducer`	Down-sample saturated values	Uses reversible named filters
`SanityReducer`	Guard against degenerate configurations	Filters obviously bad combinations
`PruningStrategy`	Base pruning behavior	Emits `ACTION_SUGGEST`-style actions

Adjacent modules

limen.experiment.param_domain is the domain reducers mutate.
limen.experiment.param_search strategies observe those mutations and change what they yield next.
limen.experiment (UEL) invokes reducers between rounds with accumulated results.

Quick orientation

Module or family	Responsibility
`budget_reducer.py`	BudgetReducer (+ TRIM_* constants)
`correlation_reducer.py`	CorrelationReducer
`focus_reducer.py`	FocusReducer
`saturation_reducer.py`	SaturationReducer
`sanity_reducer.py`	SanityReducer
`pruning_strategy.py`	PruningStrategy (+ ACTION_SUGGEST)
`filter_types.py`	FILTER_* constants
`registry.py`	REDUCER_REGISTRY

text · README.md

1reducer/2├── README.md3├── __init__.py4├── _forest_analysis.py5├── budget_reducer.py6├── correlation_reducer.py7├── filter_types.py8├── focus_reducer.py9├── pruning_strategy.py10├── registry.py11├── sanity_reducer.py12└── saturation_reducer.py

Merged source PR · ada04783Package layout · Limen 6.0.1 · current package validation · package source

Things to know

Reducers inspect the log and MSQ but return declarative interventions without mutating either. FeedbackController dispatches those interventions to MSQ, then notifies the active search strategy.
filter_types.py defines the FILTER_* constants (FILTER_KEEP_VALUES, FILTER_KEEP_BETWEEN, FILTER_EXCLUDE_VALUE, FILTER_SAMPLE) shared across reducers.
Reducers are optional in CLI and Python/MSQ runs, configured through YAML or the UEL constructor.