Found first: a primary source the press has not covered yet.
Mostafa 'Neo' Mohsenvand at MIT Media Lab has published a paper introducing Sketchpad, a system in which a Qwen3.5-0.8B model with a 10.2-million-parameter LoRA adapter performs exact arithmetic on inputs of arbitrary length. A Lean 4 proof establishes that the model is exact on all inputs if and only if it is exact on a finite set of 2,446 views; all 2,446 views pass on the deployment device. The system scored 200 of 200 on a pre-registered test suite and 220 of 220 on sums up to 1,000 digits and products up to 20 digits; GPT-6.1 Sol, the strongest of ten comparison models, scored 217 of 220.
What the source says
The adapter is a rank-16 LoRA on 150 projection matrices, adding 10.2 million parameters, 1.2% of the base model. Each forward pass reads a three-cell view of external working memory and emits one step. Operations certified are signed addition and subtraction, long multiplication and division, square roots, and greatest common divisors at arbitrary length. The Lean 4 proof uses Mathlib 4.35.0 without sorry or additional axioms; the guarantee tolerates bounded numerical perturbation and extends to compositions. Verification runs on an Apple M4 using ONNX Runtime Web on WebGPU in 16-bit precision, with a minimum margin of 3.02 logits across all 2,446 views.
Why it matters
Neural networks are normally evaluated on test sets, leaving open the question of behavior on unseen inputs. This paper reduces the infinite input space for structured arithmetic to a finite certificate, making the neural component verifiable the same way a formal program is. A model this small, running in a browser, can carry a proof of exactness rather than a benchmark score, which changes what reliability means for arithmetic in deployed systems. The certification approach extends to composed operations and tolerates numerical noise, so it is not limited to the single-step case.