03 / Deeper foundations
Better ways to think.
Better reasons to trust.
More capable systems need more than a polished interface. I am also working on how intelligence communicates, how knowledge can be checked, and how software can make stronger promises.
ABIOS explores structured, compact communication between agents. Tokenpage studies how tokenization shapes that exchange; Nullmodem explores communication below ordinary text; Scry investigates questions about training and reasoning.
These are separate research tracks feeding the wider system, not established improvements in general intelligence. Synthetic benchmarks and early studies need evidence beyond their initial test conditions. The razi-cells program (colonies of small agent harnesses under a measured selection rule) is on record in the open Weft repository, including the Weft papers.
An open research effort across mathematics, physics, and astronomy. The aim is a growing body of work that people and agents can check, extend, and cite: interactive explanations, papers, and executable claim checkers.
My incentre-lattice work began with a drawing and grew into questions about geometry, rigidity, and folding. Much of the framework is known; priority remains unresolved. Proofs and replayable checks are published, but the work has not been peer reviewed.
Most computation, proofs, checking, and writing in that project were carried out under my direction by Claude models from Anthropic and OpenAI models. Contribution record
Programme & early modules A Foundation-0 programme to rebuild widely used software libraries in Lean, with public specifications, machine-checkable proofs, and comparisons against the originals. The ambition includes a path for people to learn verification by contributing real modules under review.
The tar-extractor pilot and a growing set of library projects are underway. Each project must earn its own verification claim. A proof covers its stated specification and assumptions, not every property of a deployed system.