Verifiable AI: no hallucinated math

Why language models hallucinate numbers

A large language model does not calculate. It predicts plausible text — and a plausible-looking number is exactly as easy to predict as a correct one. On general questions the error rate is tolerable. On a hypersonic heating estimate, an orbital insertion delta-v, or a bolt-group load distribution, "plausible" is indistinguishable from "wrong" until someone checks by hand. That failure mode does not shrink as models grow; it just becomes more confident.

What verifiable AI requires

The fix is architectural, not a bigger model. Verifiable AI separates the two jobs language models conflate:

• Reasoning — understanding the engineering question, selecting the governing equations, framing assumptions. Language models are genuinely good at this.
• Computation — evaluating those equations. This must be done by a deterministic engine that executes the math step by step, the same way every time.

The result is an answer with a paper trail: the formula sequence chosen, the assumptions declared, each substitution shown, and the validity range of the method stated. An engineer — or a certification authority — can follow the derivation and reject it if an assumption does not hold. Trust is replaced by inspection.

How AETHER implements it

In AETHER, ARITHMOS converts an engineering question into an explicit formula plan, then hands every numeric step to a deterministic sequence engine — the model never invents a number. The output shows the derivation, the assumptions, and the bounds inside which the method is valid. TELOS applies the same discipline to requirements (a requirement is either verifiable or it is rewritten until it is), and THEMIS ties compliance claims back to the clauses they satisfy. The principle is uniform: every claim carries its evidence.

Why this matters beyond correctness

Certification and mission assurance run on reproducibility. A design decision made today must be defensible in an audit years from now — after the model that suggested it has been retired. Deterministic computation makes the result independent of the model version: re-run the sequence, get the same number. That property is also what makes verifiable AI compatible with sovereign deployment — an air-gapped programme can validate results without phoning home to anyone.

Verifiable AI is what allows engineering organisations to adopt AI without lowering their standard of evidence. See it working in the AETHER console, or explore programme-level deployment.

Published · 6 July 2026 · All technical notes