Neural Referee

Hardware sometimes computes a wrong value without any error; Meta studied these silent data corruptions across its fleet (Dixit et al., 2021). As computation moves inside models, the same question applies there: when a step goes wrong, which step and which part of the model? Researchers now compile real computers into transformer weights (Percepta's WebAssembly transformer, QwenPU's RISC-V CPU) but only check the final output. Neural Referee is the verifier: Muse Spark 1.3 (Meta Model API) writes the program and its own tests, the program runs inside a 7-layer transformer whose weights are a computer, and the referee bisects hash chains of the execution (about 11 comparisons for a 3,000-step run) and rechecks one step layer by layer. It returns a receipt, or the first wrong step and the faulty sublayer; Muse reads the verdict, re-runs or repairs, and only delivers answers with a passing receipt. Measured on CPU: bit flips in activations or attention memory during execution were located 30 of 30 times, where comparing weight files located 0 of 30; weight corruption 55 of 55; no false alarms on clean programs up to 44,332 steps; one hidden error caught where the final answer was correct but step 1,909 of 3,107 was wrong. Stress test with 8 bit flips per run on 12 tasks: an unchecked pipeline delivered 4 wrong answers; the Muse agent loop delivered 0 wrong (10 correct with receipts, 2 with no compilable program).