RESEARCH / INDEX
Questions I could only answer in code.
Model behaviour, privacy, and the tooling that falls out of both. Each entry links to the work itself: a result with its numbers, a repository with its code, or project notes with what I got wrong.
01 / FEATURED RESULT
Qwen3.5-0.8B-NKC
Editing a small model's existing weights to add physics and depth abilities and repair thinking-mode looping, with full results, method and limits.
READ THE RESULT02 / STUDIES AND TOOLS
Everything else
- STUDY2025-09
Blasphemous
Tooling for studying how transformer models represent refusal, and what happens when projection or LoRA interventions are measured against a fixed evaluation.
STUDY2026ZKYC
Zero-knowledge identity: a Circom circuit and browser-side Groth16 proofs combined with a vision model that checks document authenticity.
TOOL2026-09Code Cartographer
A Go CLI that maps a codebase for agents: 22 languages, symbol and source search, dependency impact, and token-budgeted context.
TOOL2026-05Open Vision
A local-first video-analysis CLI that returns timestamped visual and audio evidence an agent can use, instead of a prose summary.
03 / CITED
Where the work travelled
Apr 2025
My analysis of GPT-4o's behaviour was cited as a source in a thread that reached 32M views; Elon Musk replied “Yikes” inside it.
Surface. Thanks for diving.