Repository navigation
Reproducible companion to Reading Between the Weights.
This release includes frozen classifiers, per-adapter results, 326 locally produced adapter artifacts (315 family-study adapters plus 11 norm/scaling artifacts), exact synthetic training data and recorded evaluation evidence. The blog's main comparison remains 21 references + 98 tests; auxiliary splits are documented separately.
Verified from the release weights:
- Original 98 tests: update size 49, archived-paper features 68, broader spectral 85, cosine 98 correct.
- Recipe transfer: broader spectral 93/168; cosine 168/168, with unchanged fits/references.
- Norms for the nine original adapters and two scaled copies match the archived measurements.
Start with uv sync --locked, uv run rbw fetch --bundle family, and uv run rbw reproduce --study family. See the README for setup and the other bundles. SHA256SUMS covers the archives; metadata/releases.json also covers every extracted file.
The code, synthetic data and locally trained adapters use Apache 2.0. Base weights and third-party Mole adapter weights are not redistributed. These are controlled known-family classification results on one base checkpoint, not a general maliciousness detector. The shared synthetic-data generator and lack of unknown-family rejection limit generalization.