Research prototype
Procrustes Bridge
Test whether a rotation aligns two models' internal representations.
An experimental pipeline for extracting hidden states, fitting an orthogonal alignment, and comparing cross-model decoding against controls. The result is an experiment, including when the hypothesis fails.
What you can do with it
Who it's for: Interpretability researchers studying representation alignment between language models.
A controlled alignment experiment with held-out decoding metrics and baseline comparisons.
Your first useful result
- 01
Inspect the configuration
Read the pipeline and config source to identify model IDs, shared hidden dimensions, dataset, layers, and output location.
- 02
Plan a held-out comparison
Keep fitting and evaluation prompts separate. Include random rotation, shuffled pairing, and identity controls before interpreting a learned alignment.
- 03
Inspect metrics before making a claim
Compare top-k accuracy, reciprocal rank, and gold-token probability. A successful matrix fit alone does not demonstrate transferable meaning.
Know the limits
- The public repository has no README onboarding path, and its package build backend needs correction before a reliable installation guide can be offered.
- Full model extraction and cross-model quality results have not been reproduced as part of this site review.