Envelope & pilot race

Five yes-or-no checks before evaluation.

DRIFT: Do your parameters move by roughly 0.1 rad or more between acquisitions, measured in units of the available control span? Below that level, commissioned PI, Anderson-class mixing, and adaptive RLS can share the cell. Near or above it, TrueLoop can separate when the remaining checks pass.

COUPLING × DRIFT: Is the product of coupling strength and per-acquisition motion inside the measured stability boundary? Neither value is sufficient alone. Above roughly 0.15 rad per acquisition in the tested regime, no method reached specification and TrueLoop declines.

READOUT: Does one acquisition resolve every relevant response component repeatably? Under-resolved monitoring can destabilize control rather than merely reduce accuracy. Increase shots or integration time until repeated measurements agree within your operating tolerance.

TARGET: Can you generate the target from a known reachable configuration, and is it separated from response extrema? Never assert feasibility from bounds alone. Measure a reachable target and avoid saturated regions.

WHY 0.1 MATTERS: Drift per acquisition is drift rate multiplied by acquisition interval. Dedicated fast loops can keep this value small, and conventional controllers serve them well. At high channel count, integration, multiplexing, settling, and queueing can stretch the interval between full acquisitions. Ask your team: how far do your parameters move between complete acquisition passes, in units of the control span? A tenth of the span or more is the tested regime this product was built for.

SCALE: Is n at least 1,024 for an exclusive-cell claim? At n = 256, the measured response floor lies on the usability threshold, so marginal passes by other methods are expected. Treat that size as a boundary cell, not an exclusive result.

TIMING: Does every reading correspond to the configuration applied in that same cycle, and is every relevant channel revisited before drift makes the reading stale? A delayed or stale vector fails the coverage requirement.

Answer yes to all five before paying for an evaluation. A no is a technical decline, not a request for more tuning. A borderline case belongs in a matched pilot race against the best commissioned incumbent.

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