Configuration

Configuration

EnsembleKit is configured entirely via environment variables. All values have defaults that produce the published benchmark results; no .env file is required to run.

Environment variables

Variable	Default	Description
ENSEMBLEKIT_COMBINER	full	Combiner to use: average, weighted, robust, full, single
ENSEMBLEKIT_REGIME	homogeneous	Regime: homogeneous, het_competence, corrupted
ENSEMBLEKIT_LABELS	(auto)	Override path for Bayes label file (optional)
ENSEMBLEKIT_SAMPLES	1600	Number of samples per run
ENSEMBLEKIT_RHO	0.0	Inter-learner error correlation for diversity sweep (0.0 to 1.0)
ENSEMBLEKIT_SEED	42	Base random seed (harness iterates over 16 seeds from this base)
ENSEMBLEKIT_BACKEND	numpy	Compute backend: numpy (default, offline)

.env.example

# EnsembleKit configuration
# Copy to .env and edit as needed. All values are optional (defaults shown).

ENSEMBLEKIT_COMBINER=full
ENSEMBLEKIT_REGIME=homogeneous
# ENSEMBLEKIT_LABELS=       # leave unset to auto-generate Bayes labels
ENSEMBLEKIT_SAMPLES=1600
ENSEMBLEKIT_RHO=0.0
ENSEMBLEKIT_SEED=42
ENSEMBLEKIT_BACKEND=numpy

Selecting a combiner

# Run one combiner on one regime
ENSEMBLEKIT_COMBINER=weighted ENSEMBLEKIT_REGIME=het_competence ensemblekit compare

# Or use the CLI flag
ensemblekit compare --combiner weighted --regime het_competence

Selecting a regime

# List available regimes and their descriptions
ensemblekit regimes

# Compare all combiners on one regime
ensemblekit compare --regime corrupted

Diversity sweep

# Run the diversity sweep with the default rho grid
ensemblekit diversity

# Or set a single rho value for a one-shot run
ENSEMBLEKIT_RHO=0.6 ensemblekit compare --regime homogeneous

Reproducibility

Fix ENSEMBLEKIT_SEED and ENSEMBLEKIT_SAMPLES to reproduce exact numbers. The published table uses SEED=42, SAMPLES=1600, iterated over 16 seeds (42 through 57). The harness handles this automatically; ENSEMBLEKIT_SEED sets only the single-run seed for ensemblekit compare.