resident weights
depth-eight simulated stack
Analog AI accelerators
Keep physical weights accurate.
Keep the dense transform in physics without giving the scaling advantage back to calibration.
top-1 agreement
versus 65.7% free-running
structured channels
75.4% agreement
control arithmetic
at a 10 Hz cycle
Integration
measure(x) is your per-weight or per-column monitor readout.
The forward pass stays in the accelerator. No retraining enters the fast loop.
Open the five-line quickstartQUALIFY THE INTERFACE
Can the programmed weights or columns be monitored while inference runs?
Run compatibility checksWhere it applies
01Photonic matrix cores02Analog in-memory arrays03Neuromorphic weight banks04Wavefront processors05Crossbar accelerators06On-device inference
Evidence
Every number keeps its boundary.
Controlled simulation. Physical transfer remains unverified.