Building tools that find where AI models are wrong and fix them.
- Correcting Suppressed Log-Probabilities in Language Models with Post-Transformer Adapters (2026) A 786K-parameter post-transformer adapter corrects censorship-suppressed log-probabilities across Qwen3-4B, 8B, and 14B. Includes documentation of a silent gradient bug in Apple MLX. Code · Weights
| Repo | What it does |
|---|---|
| qwen-adapter-correction | Post-transformer adapters for correcting suppressed log-probabilities in language models. |
| sqg-lean-proofs | Lean 4 + mathlib formalization of Fourier-space identities for the inviscid Surface Quasi-Geostrophic equation, with a conditional regularity roadmap. 20k+ lines, zero sorry, no axioms beyond mathlib. |
| Package | What it does |
|---|---|
| noethersolve | 300+ verified computational tools for AI agents via MCP. |
| rho-eval | Behavioral auditing toolkit for LLMs. |