research

Building an Auditable AI: A Complete Walkthrough

Published: April 6, 2026 · Dev preview — pending ICLR 2027 submission

From Proofs to Programs to... Text?

Published: April 6, 2026 · Dev preview — pending ICLR 2027 submission

Attention, But Make It Type-Safe

Self-attention lets every token see every other token — and that's both its power and its failure mode. We constrain transformers with categorical reachability so attention edges that violate the type system can never form. A dev-preview write-up of work targeting a workshop submission.