I have spent nearly 30 years working in consequential systems that were already moving and not yet well defined. The domains have included mortgage and credit operations, automated underwriting, technical ventures, crypto trade operations, manufacturing and distribution, campaigns, institutions, and now LLM-mediated research.
My role is usually to learn the system quickly enough to find the question that matters. Sometimes the official objective is too broad to build. Sometimes the technical answer is correct but cannot be used by the institution. Sometimes the process is organized around a constraint nobody has examined. I work with the people who know each part, form a working model, and turn the result into something that can be tested and operated.
Since late 2024, I have used teams of LLMs to expand the scale and range of work I can attempt. This has made one problem impossible to ignore: models can produce implementation and analysis faster than I can determine what the accumulated work actually warrants. A model can make an unsurfaced change to a shared artifact; later models inherit it as if it were authoritative. A formal or deterministic check can pass while the surrounding claim exceeds what was checked.
My current research asks where useful control belongs: in the model, the harness, the work environment, or the claim and authorization systems around the output. Claim Fidelity is one concrete prototype in that program.
I work through LLMs much as I have worked through technical specialists for decades, but the research is mine. I originated every line of inquiry in Claim Fidelity, designed the experiments and controls, read and challenged the outputs, introduced falsifiers and counterexamples, issued explicit course corrections, and decided what survived. The models were the implementation and analysis layer. They produced the code and much of the draft analysis; I use LLMs to implement rather than claiming traditional line-by-line authorship of the Lean/Python code.
How to read this site
Work contains research and operating history in compact case form. Writing contains essays, technical notes, and working arguments. If you want the fastest technical read, start with Proofs, claims, and oversight. If you want the broader record, start with Work.
The technical material is available but not forced into the first impression. It’s there when you want to go deeper. The site is also intentionally readable by language models: LLM interface, /llms.txt, and /llms-full.txt exist so discovery can be delegated without guesswork.
To get in touch, see Contact.