0. Executive summary — the seven things that matter for Darlin

The only paper that actually tries to give you a checklist is Butlin, Long et al. 2023 (arXiv 2308.08708). It gives 14 "indicator properties" across five theories (RPT, GWT, HOT, AST, PP) plus agency/embodiment, under an explicit computational functionalism working assumption. Critically, the authors themselves say satisfying the indicators *"would not mean that such an AI system would definitely be conscious"* — the output is a credence update, not a verdict.
Theories that tie consciousness to causal/physical structure (IIT, RPT, and physical-structure readings of GWT) are the ones the falsifiability literature attacks hardest. Doerig et al.'s unfolding argument concludes such a link "is either false or outside the realm of science"; Hanson & Walker show IIT is *"simultaneously falsified at the finite-state automaton (FSA) level and unfalsifiable at the combinatorial state automaton (CSA) level."* Do not build a test suite on Φ.
The one theory that survived an adversarial empirical test least badly is GNWT, and even it took hits. The Cogitate consortium's Nature 2025 paper (Ferrante et al., doi 10.1038/s41586-025-08888-1) reports results that *"align with some predictions of both IIT and GNWT, while substantially challenging key tenets of both theories."* Nobody won.
**The single most defensible measurable construct in the whole literature is Type-2 metacognitive sensitivity (meta-d'), not "consciousness".** 2025–2026 work has moved meta-d' from humans to CNNs/VLMs/LLMs explicitly as the gold-standard AI metacognition metric. It is well-defined, has a null, and is dissociable from raw accuracy.
**The only causal (not correlational) test protocol that has been demonstrated on an artificial agent is lesion/ablation with a double dissociation.** Phua (arXiv 2512.19155) shows a "no-rewire Self-Model lesion abolishes metacognitive calibration while preserving first-order task performance, yielding a synthetic blindsight analogue", and a workspace lesion produces graded collapse in access markers. ReCoN-Ipsundrum (arXiv 2602.23232) does the same with lesion AUC drops. Copy this protocol.
**For "self-model", the literature converges on a minimum bar that is not first-person report: to count, a self-model must be (a) an explicit internal prediction of the agent's own body/policy/attention, (b) causally used in control (not an add-on), and (c) selectively necessary — lesioning it must break self-referential tasks while sparing first-order performance. Cheap add-on confidence heads fail** this bar on the evidence: Xie (arXiv 2604.11914) found auxiliary self-monitoring modules collapsed to near-constant output (confidence std < 0.006) and had no significant effect on decisions.
There are serious, published arguments that the whole question is untestable and should be abandoned in favour of moral-status/precautionary reasoning (Schwitzgebel arXiv 2510.09858; Garrido-Merchán arXiv 2405.07340; Long et al. arXiv 2411.00986). Any Darlin claim should be pre-registered as a *claim about measurable functional properties*, never as "Darlin is conscious".