The Physical Grounding Fault: Why Software Audit Sidecars Fail in Embodied Agents
Year: 2026
Type: preprint
Abstract: False completion in large language model (LLM) software agents can be audited, in the digital regime, by a four-valued verdict and a software sidecar that records whether a claimed effect was delivered [8]. The same pattern does not transfer to embodied agents. A software exception can roll back a file write; it cannot roll back a crash. Harness-written records are not evidentiary [5]. Vision–language–action (VLA) stacks collapse modular contracts into a stochastic control surface [6]. Independent hardware can take refusal off the host for access governance [2]. None of these results, taken alone, is a theorem about the sidecar. This paper treats the software audit sidecar itself as the failing object. We formalise three structural faults of physical grounding. Fault I (shared-scheduler refusal): when the sidecar shares the controller’s scheduler and fault domain, refusal remains an exception and is unreachable on a physical timescale; NVIDIA Sentry shows that leaving the host is feasible, but does not show that a co-scheduled sidecar is a Simplex shield. Fault II (physical-oracle substitution): the physical world is nearly a free outcome oracle [11], yet the sidecar replaces that oracle with a homologous shadow (logs, self-reports, hashes), so temporal witness is colluded away. Fault III (assurance-coverage collapse): a Lean 4 development can close the discrete four-valued lattice, while worst-case execution time (WCET), allocation freedom, interrupts, and sensor timing drive the fraction of discharged physical objectives toward zero—a collapse of coverage, not of the VLA module graph. Prior work established the ingredients. This paper supplies the unified critique: a sidecar that already works for digital agents fails structurally once the effect is a state of a body. The architecture that remains is a split between a slow record layer and a fast enable layer whose default, on silence, is a predetermined safe state. The discussion is restricted to civil special-use robots and unmanned aircraft (inspection, firefighting, logistics). It is not a first discovery of watchdogs or of stack collapse; it is a unified framework and formal critique of sidecar failure on a physical timescale.
Author: Liao, Zhizhi (廖至智)
Language: English
Subfield: Safety, Risk, Reliability and Quality
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225798189
Cites:
Cited by:
Related to: