What happened: Dung argues three theses. First, a test of machine consciousness can be legitimate even if an AI could be built ad hoc to pass it. Second, such AI can be identified by borrowing the philosophy-of-science criteria for an ad hoc hypothesis — his naturalness constraint: "Passing a test of machine consciousness only provides good grounds for believing that the machine is conscious[] if the machine is natural with respect to that test." Third, given the first two, the strongest tests of animal consciousness serve as valid positive tests of machine consciousness. The constraint reaches corpus absorption directly: an AI "trained on explicit information about human responses to questions about consciousness… would be ad hoc in respect to the ACT", and Schneider's boxing requirement is treated as a partial instance of it. His positive proposal deliberately avoids language processing, and is built on Birch (2022)'s own theory-light strategy — the facilitation hypothesis — together with double-dissociation paradigms after Ben-Haim et al. (2021). He states the limit himself: the tests indicate "the presence, not the absence, of consciousness", and they presuppose sensory processing or a functional analogue.
Diagnostic reading: Dung defends in a peer-reviewed venue the conclusion Birch rejects, and does it with Birch's own instrument — the theory-light strategy Birch developed for animal consciousness in 2022, extended to machines. The two are engaged on one problem rather than passing each other: the naturalness constraint reaches corpus absorption itself, counting a system trained on explicit information about human consciousness-responses as ad hoc with respect to the test. What the pairing exposes is a shape the project had been meeting one instance at a time. Faced with criterion-aware performance, each party looks for a channel the gaming does not reach — architectural indicators for Birch, non-linguistic capacity clusters and double dissociation for Dung, non-behavioural converging measures in the clinical literature — and each new channel inherits the framework-dependence of the one before it. Birch's own §4.3 → §4.4 is the demonstration: he moves to architectural indicators and meets the Janus problem on arrival. That is ARG-03's regress arriving from three directions in the literature, none of them the project's, and it indicates the substrate-generation axiom operating at the level of method rather than at the level of claim. Dung's limits travel with him: the tests indicate presence and not absence, and they presuppose sensory processing or a functional analogue, so they do not reach text-only systems as they stand.
Held and read; see working/CANON_UPDATE_SOURCE_LEDGER_2026-08-25.md §4.4, which corrects an earlier misreading of this paper in place. Received May 2023, so it predates the LLM-specific framing of the gaming problem.