What TrueLoop does not do

A short boundary for problems that require a different tool.

TrueLoop does not solve scheduling, routing, portfolio, QUBO, or any objective delivered only as a single score. Those problems use a different information interface, and the envelope declines them before evaluation.

It does not make an individual device operation or simulator call faster. It changes how many acquisitions a supported control job needs.

It does not run larger simulations. The simulator's own limits are untouched. It can control more parameters of the simulation you already run when the component-level response contract fits.

It does not claim quantum advantage.

It does not claim to outperform accurate model-based control. When a current exact model exists and can be trusted through the operating window, use it.

It does not turn a cell-scoped boundary into a universal constant. A matching registered decline or a directly measured out-of-contract condition can decline; unmatched evidence remains boundary.

The boundary of the contract is an information boundary: acquisition rate versus drift, information per read, and signal per control. The envelope states it in checkable terms.

The supported optimization surface is response matching: a feasible vector target, an exposed component response, timely coverage, and operation inside the tested coupling boundary.

For QPU scaling, per-qubit floors compound across the device. To preserve a fixed whole-device fidelity target as qubit count grows, the allowable component-response floor must fall approximately as one over the square root of n, and the readout must resolve that smaller floor.

It does not improve argmax consumers. Ranking and best-index outputs are insensitive to systematic drift (measured +0.7 points against +34.0 for a numeric consumer on the same hardware).

It does not improve a first-order variational optimiser's converged answer. The loop re-tunes its parameters and absorbs coherent drift; energies and forces at a variational minimum are second-order protected. The gradient's own fidelity is what the layer improves, and that matters when the gradient is the deliverable.

It corrects systematic, observable, actuator-invertible drift only. Stochastic decoherence is non-unitary and outside any calibration layer. It does not predict drift, since it holds no model to extrapolate.

It pays most on structured circuits. Repeated identical operations accumulate coherent error in amplitude; randomised or scrambling sequences average it out, and the measured advantage there is small.

layer = StabilityContract(key, n=len(x0), mode="regulation", target=target)
x = layer.start(x0)
for _ in range(rounds):
    x = layer.step(measure(x), target=target)
proof = layer.end()