Building behavioral auditing and alignment tools for LLMs. Try the demo →
rho-eval — Drop-in behavioral audit for any LLM. Measures 8 dimensions, no internet required. Apple Silicon MLX + CUDA + CPU.
pip install rho-eval
# Audit any model
rho-eval Qwen/Qwen2.5-7B-Instruct --behaviors all
# One-command behavioral repair
rho-surgery Qwen/Qwen2.5-7B-Instruct -o ./repaired-7b/- Rho-Guided SFT — Post-training repair of calibration damage in LLMs. DOI: 10.5281/zenodo.18854943
- Grassmann Geometry of Behavioral Entanglement — Surgery compresses subspaces, doesn't rotate them. DOI: 10.5281/zenodo.18865861
- Behavioral Phase Transitions — Geometric scaffolding precedes behavioral emergence. DOI: 10.5281/zenodo.18865198
- Confidence Cartography — Teacher-forced probability as a false-belief sensor. DOI: 10.5281/zenodo.18703505
- CF90 — Knowledge-preserving SVD compression for LLMs. DOI: 10.5281/zenodo.18718545
- Contrastive Pretraining Teaches Format Generation, Not Behavioral Knowledge — 5% injection breaks the behavioral wall at 7M. DOI: 10.5281/zenodo.18870555
- Small Language Models Already Know More Than They Can Say — The 41% universal constant and the generation bottleneck. DOI: 10.5281/zenodo.18895248
| Repo | What it does |
|---|---|
| knowledge-fidelity | Behavioral auditing + alignment toolkit. PyPI. |
| confidence-cartography | Teacher-forced confidence as a false-belief sensor. |
| intelligent-svd | Knowledge-preserving SVD compression for LLMs. |