Answers are grounded in this paper and cite their sources.
Embedded runtime controllers decide from—and are evaluated on—self-reported signals. We contribute a claim-admission audit: each channel is specified by its requested and observed signals, a consistency predicate with an explicit tolerance, and an exclusion rule naming the claims that become inadmissible when the predicate fails (request against readback, timing composition, request against enabled state, plausibility and attribution), applied to an evolutionary controller on a Jetson Nano. On three diagnostic missions the checks refuted three channels: the reported CPU frequency stayed at 1479 MHz at every requested cap, a composed end-to-end latency violated its own definition, and the thermal supervisor denied GPU preference in 94–99% of preferring records under restrictive profiles. The readback the first check calls for located the frequency cause (unprivileged, unchecked sysfs writes); two further missions located the latency cause (capture and processing stamps on clocks drifting apart at 30 µ s/s). On three final missions on one binary carrying the write/readback and clock-basis corrections (80834 control periods) the written cap equals its policy readback in every period, while the reported selected frequency exceeds two off-table caps in 3.8% of periods; the latency channel passes composition on one clock but, with 99.9% of records stamped at dequeue, is a dequeue-to-completion latency, not sensor-to-output; and GPU denial is absent under the inactive profile, occurs once under the moderate and reaches 8999 of 11168 preferring periods under the aggressive, exactly at the supervisor’s predicate. The audit runs one scan per check.
No takes yet. Share an insight, caveat, or question.
Rodriguez et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: