michef 0.0.1

MI-Chef: an audit suite for mechanistic interpretability - testing attribution graph faithfulness by serving corpora to cross-layer transcoders. v0.0.1 is an early-development release for an active research project; the audit suite ships with the paper.

MI-Chef

An audit suite for mechanistic interpretability: testing attribution graph faithfulness by serving corpora to cross-layer transcoders.

michef v0.0.1 is an early-development release reserving the package name for an active research project. The full audit suite ships alongside the accompanying paper. If you landed here early: the API below is the roadmap, not yet the product.

Why

Attribution graphs — the causal stories produced by circuit tracing — are never computed on a model directly. They are computed through a replacement model (a cross-layer transcoder, CLT) trained on a corpus the researcher chooses. Whether the choice of corpus changes the story has never been measured. MI-Chef measures it, and packages the instruments so you can audit your own interpreters before trusting their testimony.

Roadmap (ships with the paper)

• michef.audit — the product: circuit stability score, four-level agreement metrics (feature / subspace / graph / narrative), seed and paraphrase noise floors, Procrustes gauge controls, anti-phantom validation battery.
• michef.pantry — loaders for the corpus-controlled CLT grid (HuggingFace).
• michef.serve — corpus-to-CLT recipes (thin wrapper over CLT-Forge; consumes, never reimplements).
• michef.taste — side-by-side attribution-graph comparison dashboards.

Integrates with circuit-tracer and CLT-Forge checkpoints.

Status

Active research project targeting v0.1 with the paper release. Watch this space.

License

MIT