Real engine output / synthetic inputs
Find the symbol
through the noise.
NeuroControlKit stores a symbol as a 1,024-component vector. This run flips an increasing number of its signs, then asks the shipped scalar, packed CPU and Metal code to recover the closest stored symbol.
Recorded Swift execution. These controls explore 18 results captured from the current engine. They do not execute Swift or Metal in your browser. This is a numerical demonstration, not a native AssistantApp screenshot, brain measurement or rehabilitation result.
Input preview · first 128 of 1,024 components
Upper grid: original target. Lower grid: corrupted input. Mint = +1; dark = −1; amber outline = flipped component. These values come directly from the captured arrays.
What this run demonstrates
All 18 recorded cases returned the same winning key and score as the scalar reference. At 50% sign flips, the intended symbol is no longer recovered; agreement between implementations does not imply successful recognition.
| Flipped | Scalar reference | Accelerated result | Target recovered |
|---|
8 shipped compute checks passed — inspect details
| Check | Result | Observed detail |
|---|
Several shipped checks use 1,200 rows, below the current Metal threshold. The separate 4,096-row demonstration above records six actual GPU executions and their CPU/scalar comparison.