In 2023, we argued that language models could help draft formal specifications and proof candidates, with a proof assistant checking the results. That division of work has held up in our experience. So has the paper’s emphasis on domain expertise and readable formal statements: people still need to understand and approve the claim being checked.

We have been building on those ideas systematically. Codex Scribe brings AI-assisted formalization together with Agda checking and deterministic readback. MLTTDB connects formal types to operational records, while our Agda libraries provide reusable definitions and proofs. Each gives us a more practical way to put the architecture to work, and each has taught us where it needs refinement.

What surprised us most is how far the capabilities available to us have advanced by 2026. They have exceeded what we expected when we wrote this paper, making its central idea more relevant and more realistic to build: let AI propose, let formal methods check, and keep the meaning open to human review. It is encouraging to reread these early plans and see how much more we can now attempt.