The failure mode
Language models are confident arithmetic liars. Asked to do mathematics unaided, they produce a proof-shaped paragraph with a hole in step four, written in the same assured tone as the steps that were correct.
How Okita works it
We never ask the model to be the calculator. It reads the problem, structures the derivation, and hands the actual work to tools — computer algebra, numerical libraries, and, where the stakes justify it, a proof assistant that will refuse an invalid step. What comes back is a result with the work shown and the verification attached. Across a longer program, the same system tracks notation, prior results, and which methods have already failed.
What you leave with
- Tool-routed derivations using computer algebra and numerical solvers
- Machine-checked proof steps where correctness has to be demonstrable
- Notation, result, and dead-end tracking across a long research program
- Connections between methods and results from separate subfields
- Local deployment for unpublished, confidential, or export-controlled work
Where the system stops
Anything that matters gets verified by a tool that can fail it. A result without a check attached is a draft, and we label it that way.