Configuration

eigenhelm uses .eigenhelm.toml for project-level configuration. Run eh init to generate a starter file.

Threshold hierarchy

Settings are resolved in priority order:

1. CLI flags (--accept-threshold, --reject-threshold, --strict, --lenient)
2. Config file (.eigenhelm.toml)
3. Model calibration (empirical p25/p75 from training corpus)
4. Hardcoded defaults (accept=0.4, reject=0.6)

Full config reference

# .eigenhelm.toml

# Path to a pre-trained eigenspace model (.npz)
# model = "models/eigenspace.npz"

# Default language override for all files
# language = "python"

# Global accept/reject thresholds
[thresholds]
accept = 0.4
reject = 0.6

# Treat all "warn" decisions as "reject" (useful for strict CI)
# strict = false

# Path-specific threshold overrides (last-match-wins)
[[paths]]
glob = "src/core/**"
[paths.thresholds]
accept = 0.3
reject = 0.5

[[paths]]
glob = "vendor/**"
[paths.thresholds]
accept = 0.6
reject = 0.9

# Map non-standard extensions to language keys
[language_overrides]
".jsx" = "javascript"
".tsx" = "typescript"

Models

eigenhelm ships with bundled models for common use cases:

Model	Languages	Use case
general-polyglot-v1.npz	Python, JS, TS, Go, Rust	General-purpose polyglot evaluation
lang-python.npz	Python	Python-specific evaluation
lang-javascript.npz	JavaScript	JavaScript-specific evaluation
lang-typescript.npz	TypeScript	TypeScript-specific evaluation
lang-go.npz	Go	Go-specific evaluation
lang-rust.npz	Rust	Rust-specific evaluation

When no model is specified, eigenhelm uses the bundled polyglot model.

Strict and lenient modes

# Strict: marginal → reject (exit code 1)
eh evaluate src/ --strict

# Lenient: marginal → accept (exit code 0)
eh evaluate src/ --lenient

In CI, --strict is recommended to prevent marginal code from merging.