The direction law, measured: what two quantum processors settled
For a year the honest caveat under every number on this site was the same: established in simulation, hardware validation pending. That caveat just retired. Two commercial quantum processors, one pre-registered ablation, one 107-channel closed loop, and the boundary of the whole framework drawn from both sides.
What was actually at stake
The runtime’s premise is simple to state: treat the device’s writable configuration as the retained state of the computation and update it directly from one parallel measurement per cycle. Everything else on this site follows from that premise, which means everything on this site inherited one risk: does the loop survive contact with real hardware, real sampling noise, real cloud latency, real billing?
It survived twice, on two different vendors’ machines.
107 channels, twelve measurements, zero excuses
On a 108-qubit superconducting processor we held the excitation probabilities of 107 physical channels to a hidden measured target under injected drift we disclosed in advance. One 150-shot parallel acquisition per cycle. The runtime converged in twelve acquisitions, in both of two runs. Finite differences, given the identical budget on the same device and target, probed 11 of the 108 evaluations its first gradient requires: zero completed updates, and its error never improved. That is not a tuning gap. Its own arithmetic disqualifies it before the first step.
The invoice matters as much as the convergence. From billing metadata across thirty-plus jobs: one runtime update cost one task, roughly 20 credits, roughly 100 milliseconds of machine execution. One finite-difference update at this size costs 108 tasks and roughly 2,180 credits. And the cloud loop itself carried a median 48.9 seconds of wall time per 100-millisecond cycle of physics, a measured 489-fold overhead that is the entire case for running this control co-located with the device.
The ablation that isolates the mechanism
Convergence alone does not prove the retained configuration is what does the work. So on a second vendor’s processor we ran the cleanest test we could design, and pre-registered it: two controllers identical in every component, except one carries the configuration between cycles and the other re-draws it. Five hypotheses committed before any hardware submission. Five passed. The reset arm failed to descend on every instance; the retained arm descended on every instance; the pooled error ratio was 7.4 against a committed bar of 2.0. Retention is not a detail of the implementation. It is the mechanism.
Both sides of the boundary
The same program mapped where the loop cannot go, because a regime map with no red cells is marketing. Feedback that arrives as a single scalar score provably starves at scale, and we measured the decay curve. Unstructured spin glasses trap below the classical ceiling and we publish the gap. A landscape your laptop can evaluate for free belongs to your laptop. And an annealing solver actively resists mid-solve stabilization; correcting it makes it worse, which we also publish. The winning rows are only believable because the losing rows sit next to them, with the same error bars.
Run it yourself
Everything above that runs without special hardware access is in public repositories, with pre-stated verification bars and every result JSON, runnable to 4,096 channels on a free evaluation key. The research page has the links, the evidence ladder, and the manuscript. If a number does not hold on your machine, we want to know.