What happened: A DeepMind-led group of fourteen proposes to replace the “cacophony” of eighteen theories with a single question. The move is stated as a bracketing: separating the hard problem from the “mapping problem” “allows the deepest metaphysical disagreements to be set aside, with only views that deny any discoverable psychophysical mapping excluded, on methodological rather than metaphysical grounds. Granting that experience supervenes on a system’s organisation, the tractable question becomes at which grain of description that supervenience base sits.” Marr’s three levels become five; each theory is placed at a level; indicators are attached; a Bayesian model combines theoretical credence across levels with indicator evidence. The illustrative runs pass “the face validity checks of a human and a thermostat”, show “that evidence of fine-grained organisation carries more weight than surface evidence”, and find “that the verdict for current LLMs is driven as much by where theoretical credence is placed as by how the evidence is read: under different stipulated readings and credence distributions, assessments range from below 0.01 to roughly 0.8.” The authors add that this “shows sensitivity to assumptions and should not be interpreted as an empirically established range”, and offer the framework as supporting “a structured agnosticism, in which theoretical commitments are made explicit, credences are updated as evidence accumulates, and assessments take the form of aggregated probabilities rather than verdicts.” Shevlin’s specificity problem is acknowledged as recurring “at every level of the hierarchy and inside every indicator”.
Diagnostic reading: The project’s spine asks what dissolves, what appears and what is addressable when a starting point is chosen, and here the indicator programme has run that question on itself and published the answer. The founding sentence does two things at once: it sets the hard problem aside, and it grants “that experience supervenes on a system’s organisation” — which is the materialist axiom FND-02 v5.8’s dual reductio examines, written into the method as its one assumption and labelled methodological. FWK-01 v2.4’s two-axis placement locates the result before any credence is assigned: every theory in the hierarchy sits on one side of the ontological axis, so the “cacophony” the framework orders is a disagreement among positions that share a premise, and the spread it reports is the spread available within that premise. That is the finding, and it is the authors’ own: two orders of magnitude of verdict produced by where credence is placed, with the evidence held fixed. The project would say the framing question has shown up as a sensitivity parameter, which is the only way an instrument built inside one framework can register it. FWK-03 v1.5’s criteria-movement watch finds the honest version of the move here — behavioural indicators down-weighted on stated theoretical grounds rather than retired after the fact — and ARG-02 v1.7’s boundary trilemma is reproduced in miniature by the face-validity anchors: human at one, thermostat at zero, and everything of interest decided by what sits between on a scale the two anchors were chosen to define. Symmetry is partial. The framework is applied to humans and flies as well as to language models, which is more than most of the literature does, but the human case is the calibration rather than a candidate, so the question ARG-01 v1.5 asks — whether the same instrument, run on us, would return the same spread for the same reason — is not put. The entry’s judgement is not that the framework fails; it is that it succeeds in making explicit what the project has argued is there to make explicit, and that “structured agnosticism” is the name the programme gives to arriving at the project’s starting observation by the long road. Stable on the direction axis: no criterion has moved; what has moved is the location of the uncertainty, from the evidence to the premise, and the authors put it there themselves. [Derived]
Preprint (arXiv:2609.35618, cs.AI, 28–29 September 2026; c. 150pp; fourteen authors including Chandaria, Rosas, Seth, Shevlin, Shanahan, Hutter, Legg and Frith), read from the arXiv HTML 2026-10-15 — abstract, §1.4, §3, §7.4 and the conclusion in full, the rest by section — quotations verified. The assessments are the authors' own illustrations with stipulated inputs and they say so: the spread 'should not be interpreted as an empirically established range for the probability of consciousness.' Seth is at his fourth appearance as a subject in this log and here is one of fourteen; the subject is the framework. Nothing here is evidence for or against any system experiencing anything.