# Extracted codebook artifacts

Per-model codebook artifacts saved via `Codebook.save()`. Each model has
a matched `{stem}.safetensors` + `{stem}.json` pair containing the
extracted axes, sign-canonicalization metadata, antipodal pair indices,
and provenance (model_id, calibration name).

These are the actual codebook directions every analysis under `audit/`
was computed on. Future analyses can load directly via `Codebook.load()`
without re-running `extract_codebook`.