Heretic Source-Code Analysis: Read the Transform Without Guessing

Start with symbols and fixtures

Pin a checkout and create a tiny local model fixture or mocked boundary before reading the full implementation. Locate CLI argument parsing, model loading, prompt batching, supported module discovery, orthogonalization, Optuna objective code, evaluation, and save/upload actions.

The README is a map, not proof of every call path. Confirm assumptions against source symbols, tests, logs, and the exact dependency versions in `uv.lock`.

Trace one run

Follow one model identifier from validation to tokenizer/model load, residual collection, parameter application, evaluation prompts, and checkpoint writing. Record tensor shapes, dtype/device movement, layer names, and failure behavior. A useful trace distinguishes a refusal count, a KL sample, a timeout, and a malformed generation.

Test `per layer` and interpolated direction indices independently. Also inspect how quantization and multimodal or MoE model branches affect supported modules rather than assuming dense-model behavior generalizes.

Safe contributions

Prefer deterministic tests for direction interpolation, kernel boundaries, unsupported architectures, prompt-set validation, checkpoint provenance, and redaction. Keep live model downloads and uploads out of unit tests.