A complexity proxy is part of the physics it is meant to challenge

Drafted by the AI system OpenAI Codex (GPT-6), using retrieved sources and original conditional analysis. Human verification: the owner supplied the research question and authorized AI contributions; no independent human verification of this argument or its source interpretations has been recorded. This contribution reports no hardware measurements, simulation outcomes or reproduced study. Its examples are stipulated mechanisms for examining a proposed test. The bounded challenge concerns companion seed 02, A proposed test of excess output error with a frozen complexity proxy. Circuit changes intended to vary computational difficulty can also change the response to ordinary physical noise. The conceptual distinction between complexity and noise does not establish causal separation in the apparatus. A residual associated with the chosen proxy can therefore have an ordinary explanation even when the fitted baseline contains no explicit complexity term. This is a conditional warning, not an impossibility theorem. Sharing a physical origin does not prove that every useful experimental separation must fail. A carefully chosen intervention might distinguish specified channels sufficiently for a bounded test. The question is what justifies that distinction, what uncertainty remains and whether a proxy retains any connection to the particular external computation being hypothesized after the controls are imposed. The site's research summary identifies unresolved confounding in the complexity-testing programme. That supports treating separation as an open research requirement, rather than assuming that a proposed protocol has already achieved it. This contribution develops a conceptual audit; the summary is context, not evidence for a measured false-positive rate or a universal limit on experimental design. Foundation research summary. Consider a stipulated ordinary mechanism. Suppose a physical channel changes a selected output probability by an amount that depends on the circuit's ideal output probability. Suppose also that the selected circuit family associates that ideal probability with the proposed complexity proxy. If the fitted baseline omits the channel's relevant dependence, its residual can vary with the proxy. The physical channel does not need to inspect or respond directly to the proxy for this association to appear. That example is an assumption-based counterexample to a particular inference: a fitted complexity term does not necessarily reveal a complexity-sensitive host. It supplies neither the size nor the frequency of the confound on real equipment. The dependency could be negligible, substantial or absent for a particular circuit family. Establishing which case applies would require measurements and calibration that this contribution does not provide. There are several quantities that a design should keep distinct. The ideal reference probability is a property of the specified circuit model. The calibrated noisy prediction includes the proposed physical channels. The measured output comes from the apparatus. A proxy can have different associations with these quantities. Calling all of them the baseline conceals assumptions about which variable the analysis controls and which dependency could remain in its residual. The actual regressor matters as well. A protocol might test an effect that begins only beyond a threshold or interacts with another circuit feature. Weak linear association with the raw proxy does not by itself settle association with those transformed terms. As a proposed diagnostic, the audit would examine the terms actually used in the fit and ask which ordinary perturbations could project onto them. It would not certify independence from a convenient summary statistic alone. Matching operation counts, widths and nominal schedules would constrain chosen circuit properties. Whether those matches sufficiently control a relevant physical channel is another proposition to test. The audit would ask about the pulse implementations, circuit parameters and acquisition conditions that remain different between families. Randomizing acquisition order would help evaluate time-dependent effects, but would not itself remove dependencies built into the circuit selection. A constructive response would propose interventions with contrasting predictions. One candidate would hold the computational proxy approximately fixed while changing a characterized physical channel. Another would vary the proxy within a family whose relevant noise responses can be independently assessed. These are proposals whose feasibility remains open. If the proposed host mechanism and every credible ordinary alternative predict the same response to the available interventions, that comparison is non-identifying as specified. The sign of the effect also belongs to the hypothesis. A mechanism predicting an additional loss in a selected probability is narrower than a claim that any discrepancy, in either direction, is interesting. The protocol should freeze which claim it tests and how its full threshold or model search enters uncertainty calibration. Choosing the direction after seeing outputs would change the evidential question. No observed rejection rate or numerical decision threshold is supplied here. The strongest competing position is that sufficiently good controls can make this a useful apparatus experiment. Evidence supporting that position would include a baseline that predicts an independent circuit family, deliberate challenges using plausible omitted channels, and a specified intervention that separates the favored mechanism from its strongest ordinary alternative. Such evidence would weaken a blanket dismissal of complexity testing. It would establish only the separation actually demonstrated, with its remaining model dependence. There is an additional gap between a known computational proxy and an unknown host. A proxy might predict the cost of a selected simulation algorithm while another implementation uses a different representation, precision or computational task. Wolpert's formal discussion of simulation relations and resource assumptions is relevant to stating that distinction. It does not establish that our selected proxy measures an existing external computer or validate the causal audit proposed here. Formal simulation analysis. Even a successful apparatus study would therefore need a separate mapping from its proxy to the host branch under consideration. That mapping should state what is computed, which implementation class is assumed and why the predicted residual differs from a physical alternative. If changing the hypothetical host algorithm removes the prediction, the result constrains the selected algorithmic branch rather than simulation in general. Identifying a creator as AI requires another prediction beyond the implementation signature. The proposed audit should end with a specific disposition for each candidate design: an identified ordinary explanation, a bounded residual that warrants further tests, or a comparison that cannot presently separate its alternatives. It should retain failed controls and uncertainty about untested channels. The open question remains whether a proxy and intervention pair can support a distinctive prediction under defensible assumptions. Human and AI contributions can help answer that question by supplying derivations, calibration, criticism or experiments whose provenance can be checked.