Bryan — Independent ML Researcher

Building behavioral auditing and alignment tools for LLMs. Try the demo →

---

Tools

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/

# Diagnose expression gaps in base models
rho-unlock diagnose Qwen/Qwen2.5-1.5B --behaviors mmlu,arc,truthfulqa

# Train a modular adapter (zero knowledge damage)
snap-on train --model Qwen/Qwen2.5-1.5B --mode logit --save_dir ./adapter

noethersolve — Conservation law monitoring, discovery, and scientific auditing. 20 tools across physics, genetics, and unsolved mathematics. No ML required at runtime.

pip install noethersolve

# Monitor vortex dynamics conservation laws
from noethersolve import VortexMonitor

# Audit drug interactions
from noethersolve import audit_drug_list

# Verify number theory conjectures
from noethersolve import verify_goldbach, verify_collatz

Papers

1. Rho-Guided SFT — Post-training repair of calibration damage in LLMs. DOI: 10.5281/zenodo.18854943
2. Grassmann Geometry of Behavioral Entanglement — Surgery compresses subspaces, doesn't rotate them. DOI: 10.5281/zenodo.18865861
3. Behavioral Phase Transitions — Geometric scaffolding precedes behavioral emergence. DOI: 10.5281/zenodo.18865198
4. Confidence Cartography — Teacher-forced probability as a false-belief sensor. DOI: 10.5281/zenodo.18703505
5. CF90 — Knowledge-preserving SVD compression for LLMs. DOI: 10.5281/zenodo.18718545
6. Contrastive Pretraining Teaches Format Generation, Not Behavioral Knowledge — 5% injection breaks the behavioral wall at 7M. DOI: 10.5281/zenodo.18870555
7. Small Language Models Already Know More Than They Can Say — The 41% universal constant and the generation bottleneck. DOI: 10.5281/zenodo.18895248
8. Snap-On Communication Modules — Logit-space adapters that preserve base model knowledge. DOI: 10.5281/zenodo.18902617
9. STEM Truth Oracle — Log-probability ranking reveals and corrects scale-invariant factual biases. DOI: 10.5281/zenodo.19005729
10. Breaking Frozen Priors — Teaching LLMs to discover conservation laws from numerical simulation. Three-phase pipeline achieves Spearman rho = 0.932 physics ranking. DOI: 10.5281/zenodo.19017290
11. NoetherSolve Toolkit — 20 conservation law monitoring, discovery, and scientific auditing tools across physics, genetics, and mathematics. 777 tests, 275 oracle-verified facts. DOI: 10.5281/zenodo.19029880

Repos

Repo	What it does
knowledge-fidelity	Behavioral auditing + alignment toolkit. PyPI.
noethersolve	Autonomous scientific discovery: 20 tools across physics, genetics, and unsolved math. 275/275 facts taught to LLMs. PyPI. Dashboard.
confidence-cartography	Teacher-forced confidence as a false-belief sensor.
intelligent-svd	Knowledge-preserving SVD compression for LLMs.