AI inner-loop applications

The stability layer beneath a physical forward pass.

In a hybrid AI system, the digital host owns training, semantics, routing, and control flow. A photonic or in-memory analog coprocessor can own the dense forward transform, with weights retained in physical state and only vectors crossing the interface.

TrueLoop belongs between those layers when each drifting physical weight or structured column response can be monitored and corrected before its reading becomes stale.

The measured result that carries this page is numeric validity: on 96 qubits of IBM hardware, a real image-classification pipeline embedded through the maintained analog stage kept 80.9% of its numeric outputs valid against 46.8% free-running, +34.0 points. That is the consumer a maintained substrate helps: one that needs the number. On the same run the argmax consumer moved +0.7 points, because a ranking is insensitive to systematic drift. Lead with the number you actually consume.

In the simulated campaign, a depth-eight stack with 8.4 million resident weights retained 78.6% top-1 agreement under elementwise drift, versus 65.7% free-running and 70.1% with scheduled recalibration. The structured 16,384-channel geometry retained 75.4%, versus 61.3% and 66.9%.

This is evidence for the maintenance layer, not a claim of end-to-end application accuracy or a demonstrated physical analog-AI product. A native device pilot and a workload executed across both substrates remain open proofs.