What was tested

Global neuronal workspace theory (GNWT) holds that a content becomes conscious when it is broadcast across a network of long-range connections centred on prefrontal and parietal cortex. The broadcast starts with "ignition", a sudden, self-sustaining burst of activity that makes the content available to many specialised processes at once. Integrated information theory (IIT) starts from the properties of experience and infers what a physical substrate must be like to have them. It holds that consciousness is identical to the cause-effect structure of a system with maximal integrated information, and it locates the main such system in human brains in posterior cortex, a "hot zone" spanning parts of the parietal, temporal and occipital lobes.

The theories disagree about a great deal, but most of what they disagree about can't be measured directly. You can't measure integrated information in a human brain; the calculation is intractable for systems of that size. You can't observe a workspace as such. What you can measure is activity, and the patterns of activity each theory leads you to expect. The art of an adversarial test is to find patterns that one theory requires and the other forbids.

The design

The consortium's design is set out in a protocol published in PLOS ONE and an earlier Science perspective by Lucia Melloni, Liad Mudrik, Michael Pitts and Christof Koch. The final study used 256 participants across three methods: functional magnetic resonance imaging (fMRI), magnetoencephalography (MEG) and intracranial electroencephalography (iEEG) in epilepsy patients, Meera's kind of patient. Participants saw faces, objects, letters and false fonts, with 20 identities per category, in three orientations, for 0.5, 1.0 or 1.5 seconds. They performed a target-detection task, and which category counted as a target varied across blocks, so that the same stimulus was sometimes relevant to the task and sometimes not. Orientation was never relevant. The trick was that a stimulus on screen for 1.5 seconds is consciously seen for about 1.5 seconds, whether or not you are attending to its category, and the theories disagree about how the brain represents that persistence.

For fMRI and MEG, analyses were tuned on a third of the data and then run, as preregistered, on the remaining two-thirds. The iEEG data didn't use that split. The theories' leaders, Stanislas Dehaene for GNWT and Giulio Tononi with Christof Koch and Melanie Boly for IIT, agreed in advance on three predictions, on what would count as passing and failing each, and on how each outcome would be interpreted.

The three predictions and what happened

The published results can be summarised prediction by prediction.

Prediction 1, decoding conscious content. IIT predicted that content would be decodable mainly from posterior cortex, and that adding prefrontal cortex (PFC) wouldn't improve decoding. GNWT predicted that PFC would carry the conscious content. Both were true in part. In iEEG, posterior regions decoded faces versus objects at more than 95 per cent for roughly the stimulus duration; PFC decoded at about 70 per cent, in a window of roughly 0.2 to 0.4 seconds. fMRI showed a similar ordering, about 75 per cent posterior and 60 per cent in PFC. Face orientation was decoded posteriorly but was absent from PFC in iEEG and fMRI. Adding PFC to posterior regions didn't improve decoding, and sometimes made it worse.

Prediction 2, maintenance. IIT predicted that posterior cortex would show sustained activity tracking how long the stimulus lasted. GNWT predicted brief ignition in PFC about 0.3 to 0.5 seconds after the stimulus appeared, and again after it disappeared, with the content held in a silent state in between. In posterior iEEG, 25 of 657 electrodes tracked duration. In PFC, 99 electrodes showed non-selective responses at onset and 24 showed category-selective onset responses, but none of 655 PFC electrodes showed the predicted profile of onset plus offset ignition.

Prediction 3, connectivity. IIT predicted sustained synchronisation in the gamma band within posterior cortex, between category-selective areas and early visual areas V1 and V2, for as long as the stimulus was seen. GNWT predicted brief, late synchronisation between category-selective areas and PFC. The preregistered measure, pairwise phase consistency, found posterior synchronisation that was early, brief (under 0.75 seconds), confined to 2 to 25 hertz and mostly explained by the stimulus-evoked response. It found no content-selective synchronisation between category areas and PFC in the relevant window. The paper's verdict was that these results "supported neither theory".

The authors identified one headline challenge for each theory. For IIT, "a lack of sustained synchronization within the posterior cortex contradicts the claim that network connectivity specifies consciousness". For GNWT, the challenge was "the general lack of ignition at stimulus offset and limited representation of certain conscious dimensions in the prefrontal cortex". Their overall assessment, from the summary figure, was that IIT's results mixed "a passed prediction" with "a failure (maximum integrated information)", and GNWT's a "partly challenged" prediction with a "partly supported one".

What Cogitate got right

Before the diagnosis, a fair account of the achievement, because what follows could be read as a criticism of the people who ran the study, and it isn't. Three features of Cogitate are new in consciousness science, and each of them is the reason a diagnosis is possible at all.

First, the predictions were written down before the data existed, by the theories' own leaders, in a public protocol. Most disputes in this field are fought over results that each side interprets after the fact, with predictions reconstructed to fit. Here, anyone can open the PLOS ONE protocol and compare what was promised with what was found. Every argument in this paper depends on that record. Without it, there would be no way to show that offset ignition was demoted, because there would be no earlier position to demote it from.

Second, the study used three methods with different strengths. fMRI locates activity well and times it badly; MEG times it well and locates it less well; intracranial recordings do both, but only where clinicians have put electrodes. Agreement across the three counts for more than any one, and disagreement is informative too. When orientation was decodable from prefrontal sources in MEG at about 35 per cent but absent in iEEG and fMRI, the authors could flag possible signal leakage from posterior sources instead of reporting a prefrontal finding. That kind of cross-check is only available to a study built on several methods at once.

Third, the fMRI and MEG analyses were tuned on a third of the data and then run, unchanged, on the remaining two-thirds. This guards against the most common way honest researchers fool themselves: trying analyses until one works. It means the positive results that did emerge, such as the strong posterior decoding, are unlikely to be artefacts of analytic flexibility.

So the study did the hard parts. It got two rival camps to agree on what they were testing, it measured the brain in three ways, and it protected its analyses against wishful choices. What it couldn't do was stop the camps from disagreeing afterwards about what the agreed predictions had been worth. That is a smaller gap than the ones it closed, and it is the one this paper is about. The authors themselves describe their aim as inviting readers to weigh the evidence, not issuing a verdict, and their closing call for a framework that accounts for "prediction centrality" shows they saw the gap before any critic did.

How each side kept its theory

The paper's supplementary material includes statements from the theory leaders, and the GNWT side later published a separate commentary. Reading them together shows the mechanism by which a challenge stops short of refutation. Neither side denied the data. Each side relocated the failure.

The IIT reply

The IIT statement, by Boly, Koch and Tononi, says the results "corroborate IIT's overall claim that posterior cortical areas are sufficient for consciousness, and neither the involvement of PFC nor global broadcasting are necessary." On the failed synchrony prediction, it offers several explanations. The prediction "may be wrong with respect to the frequency range"; synchrony may occur at lower frequencies, and indeed content-specific synchrony between category areas and V1/V2 did appear in the 2 to 25 hertz range for the first three-quarters of a second. The failure "may also stem from technical limitations": there were "only 12 iEEG electrodes in V1/V2 (against 472 in PFC)". The final iEEG analysis "did not strictly follow pre-registered analyses methods for selecting ROIs", using a more lenient selection of regions of interest that "decreased the relevance of the synchrony analysis to test IIT". And micro-saccades, small involuntary eye movements, became more frequent after half a second, which on IIT would shift which V1/V2 units contribute to the experience, and so dilute any measured synchrony.

Each of these is a reasonable scientific point. Together they show where the prediction lived. IIT's core claim is about causal relations among the units of a maximally integrated substrate. The step from causal relations to gamma-band phase synchrony between electrodes is an auxiliary hypothesis: the statement itself says relations "are likely to result in increased synchrony among active units, hence the prediction". When the prediction failed, the auxiliary absorbed the blow.

The GNWT reply

The GNWT side made the same move, at a different joint. A commentary in Neuroscience of Consciousness by Lionel Naccache, Claire Sergent, Stanislas Dehaene and colleagues argues that "while ignition after stimulus onset is indeed a core prediction of GNWT", "offset ignition is not". Offset ignition is expected "only" if the participant consciously attends to the offset, and in the key task-irrelevant conditions, with long and variable durations, "a systematic awareness of stimulus offset seems unlikely". The commentary notes that "none of the simulations of the model" and "none of the experimental recordings" had revealed a definite offset ignition. It also argues that the study, by using only clearly visible stimuli, "could not test the most central predictions of the GNWT, which tackle the critical contrast between conscious and non-conscious processing", and that the posterior decoding counted for IIT "could have arisen from purely non-conscious processes".

Dehaene repeated the point in his Collège de France lectures: offset ignition "is NOT a core prediction of GNW; it depends on where attention is." In the same lecture he presented the onset results as confirming GNWT: in prefrontal and cingulate cortex, "activity is transient and independent of duration… exactly as predicted by GNW".

Duhem's problem, twice

The structure of these replies has a name. Pierre Duhem argued that a physical hypothesis never meets an experiment alone; it always meets it in company with assumptions about instruments, background conditions and other theories. W. V. O. Quine generalised the point. If a prediction fails, logic tells you that something in the bundle is false, but not which thing. An earlier paper in this series used Duhem to explain why disputes about evidence keep reopening. Here I want to use him more narrowly, to locate the joints at which the Cogitate predictions could bend.

The lesson of the two planets isn't that rescuing a theory with an auxiliary is cheating. Neptune shows it can be the best science there is. The lesson is that a rescue earns its keep only when the auxiliary makes a new, independent prediction that comes true. Neptune was a rescue that pointed at a place in the sky. Vulcan was a rescue that pointed at a place in the sky and found nothing there. The difference was settled by looking.

Apply that test to the Cogitate replies. The IIT reply proposes that synchrony occurs at lower frequencies, and that denser sampling of V1/V2 would reveal sustained gamma synchrony. Those are testable: they point at a place in the sky. The GNWT reply proposes that offset ignition appears when participants attend to offsets, and Dehaene's lecture reports a dual-task experiment designed to test whether stimulus offsets occupy the workspace. That is also testable. So neither reply is a Vulcan yet. Each is a Neptune that hasn't been looked for.

Where each prediction could bend

Each of the three Cogitate predictions can be drawn as a chain: core claim, bridge assumption, measured quantity, analysis choice, result. The joints in each chain are where a failure can be absorbed. For IIT's synchrony prediction, there are at least four: whether causal relations show up as synchrony at all, whether in the gamma band, whether the regions selected were the right ones, and whether 12 electrodes in V1/V2 could detect the effect. For GNWT's offset prediction, there are at least three: whether ignition at offset is core or conditional on attention, whether the task made offsets conscious events, and whether ignition would show in high-gamma activity on the sampled PFC electrodes.

Notice that the chains differ in one revealing way. IIT's joints are mostly about measurement: band, sampling, region selection. GNWT's main joint is about the theory itself: whether a prediction the theory's leader signed is core. I'll argue below that the second kind of relocation is more costly, because it changes what the theory says rather than how it is measured. But I'll also argue that both were made possible by the same gap in the protocol.

Common reasons and the wheel of nine

The second reason the results didn't decide comes from the design's choice of stimuli. Every stimulus was clearly visible. That was deliberate: the aim was to study how a conscious percept is maintained over time, and a clearly visible stimulus presented for 1.5 seconds gives you a percept with a known duration. But it has a cost the GNWT commentary named. Without a contrast between conscious and unconscious processing of the same stimulus, a signal that tracks the stimulus could reflect conscious experience or unconscious processing of it. So a successful posterior decoding result is what IIT predicts, and it is also what GNWT predicts for the unconscious stages of processing that precede broadcast. The evidence is present whether IIT is right or wrong.

Dignāga had a classification for exactly this situation, and it is more precise than anything in the modern discussion. In his theory of inference, a good reason (hetu) must satisfy three conditions, the trairūpya: it must belong to the subject under discussion, it must be present in similar cases (sapakṣa), where the property to be proved is known to hold, and it must be absent from dissimilar cases (vipakṣa), where the property is known to be absent. His Wheel of Reasons, the hetucakra, combines three possibilities for each: present in all, some or none of the similar cases, and present in all, some or none of the dissimilar cases. That gives nine types of reason. As Britannica summarises it, two of the nine yield valid conclusions, two are contradictory and the remaining five leave the matter uncertain.

A study of the wheel in the Austrian Academy's series lays out the classification. The valid reasons are the second type, present in all similar cases and absent in all dissimilar ones, and the eighth, present in some similar cases and absent in all dissimilar ones. What matters is absence from every dissimilar case. The first type, present in all similar and all dissimilar cases, is inconclusive because it is "common" (sādhāraṇa): it shows up whether or not the property holds. The fifth, present in neither, is inconclusive because it is "uncommon" (asādhāraṇa). The fourth and sixth are contradictory: they show up only where the property is absent.

Translate this into a test of theories. The "property to be proved" is "IIT is true" (or GNWT). The similar cases are possible worlds, or models, in which the theory holds; the dissimilar cases are those in which it fails. A good piece of evidence is one that appears where the theory holds and never where it fails. Sustained posterior decoding of a visible stimulus appears in IIT-worlds and also in GNWT-worlds, because GNWT also expects posterior cortex to process visible stimuli. It is a common reason, Dignāga's first type. It can't establish IIT against GNWT however strongly it appears.

The wheel's most useful lesson is that the defect of a common reason doesn't depend on how often it appears. A common reason present in every similar case is no better than one present in some. The modern equivalent is the likelihood ratio: the probability of the evidence if the theory is true, divided by its probability if the theory is false. A result predicted equally by both has a ratio near one, and no number of replications changes that.

The worked case explains a feature of the results that puzzled some commentators: both theories could point to many results "in line with" their predictions, and neither could convert those into a victory. A result in line with your prediction is evidence only if it would be out of line with your rival's. Most of Cogitate's confirmations weren't, because the design couldn't make them so without a conscious versus unconscious contrast.

Why the absences carried so little

If the absences were the decisive results, why didn't they decide? Partly because of the relocations already described. Partly because absence is hard to establish. For the synchrony prediction, the paper reports Bayesian analyses that "supported the null" for posterior iEEG synchrony with Bayes factors from 1.15 to 4.9 in favour of no effect. By common conventions, a Bayes factor between 1 and 3 is weak evidence and one between 3 and 10 is moderate. So the strongest evidence that IIT's synchrony was absent was moderate at best, and some of it was barely more than a coin toss. For the PFC synchrony that GNWT predicted, the factors were 2.62 to 5.32, similar. A null result drawn from 12 electrodes, with factors in that range, is an absence of evidence more than evidence of absence, and the IIT reply exploited exactly that.

That thought experiment cuts against a symmetrical reading of the results, and I think it is right to let it. The GNWT offset result is a well-measured absence of something the theory's leader signed as a prediction. The IIT synchrony result is a poorly measured absence of something the theory's chain predicted. If the question is which failure should have hurt more, the answer on measurement grounds is the GNWT one. If the question is which relocation was more legitimate, the answer is the IIT one, because IIT's reply is about measurement while GNWT's is about what the theory claims. Those two judgements point in the same direction, and the published commentary from the GNWT side doesn't address the first.

Centrality was never fixed

Now the first and, I think, the deeper reason. The Cogitate protocol fixed what each theory predicted, the measurement that would test it and the criteria for passing. It labelled predictions as more or less important. What it didn't fix was a number: how much a failure of each prediction should lower confidence in the theory that made it. And it didn't bind the theorists to say, before the data, which auxiliaries they would give up if a prediction failed.

The consortium's own paper recognises this. Its authors call for "a quantitative framework for systematic theory testing and building", one that integrates evidence "while accounting for prediction centrality, measurement error, and cross-sample consistency". That is a remarkable thing for the authors of a preregistered adversarial test to ask for, and it is an admission that their protocol didn't contain one. Without it, the centrality of each prediction was left to be argued afterwards, and afterwards is when every incentive points towards demotion.

The demotion of offset ignition

Look at how the GNWT offset prediction moved. In the preregistered protocol, the prediction of ignition at both onset and offset was one of GNWT's tested predictions for maintenance, signed off by its proponents. In the summary figure of the Nature paper, the lack of offset ignition is listed among the "key tenets" challenged. In the GNWT commentary, written after the data, offset ignition "is not" a core prediction and is expected "only" when participants attend to offsets. In Dehaene's lecture the word "NOT" is capitalised.

The scientific argument for the demotion is respectable. GNWT links ignition to conscious access, and conscious access to attention. If nobody consciously notices that a task-irrelevant face has disappeared, there is no new content to broadcast, and so no ignition. That is a coherent reading of the theory. The difficulty is only that it wasn't the reading on the table when the prediction was registered. A reader of the protocol would have scored the absence of offset ignition as damaging; a reader of the commentary would score it as irrelevant. The theory didn't change. What changed was the weight given to one of its consequences.

The figure shows my reading of the shift, and I want to be clear about its status. The weights before and after are mine, not anyone's published numbers. Nobody published weights, which is the point. I've placed them to reflect the language used: "key tenet" in the joint paper, "not" core in the commentary, and for IIT, "may be wrong with respect to the frequency range" for the synchrony prediction, which moves it from a direct test of the theory towards a test of a bridge assumption.

Why centrality matters more than predictions

Here is the general point. A theory that makes three predictions, each of which fails, is in trouble only if those predictions were central. If they were peripheral, their failure is consistent with a revision at the periphery. Every serious theory has a core and a protective belt, in the language Imre Lakatos used for scientific research programmes: the core is what you won't give up; the belt is the auxiliaries you adjust to keep the core in contact with the data. Lakatos thought a programme was progressive if its belt adjustments predicted new facts, and degenerating if they only accommodated old ones.

An adversarial test can only damage a core. If the predictions it tests sit in the belt, the most it can do is force an adjustment. So the critical design question for any adversarial collaboration is how to get predictions that sit as close to each theory's core as the measurement allows, and the critical fairness question is how to stop either side redrawing the line between core and belt after the results. Cogitate addressed the first question with great care. It left the second to good faith.

The story matters for fairness. A reader could take the commentary's demotion of offset ignition as goalpost-moving. The refractory-period experiment shows the GNWT side treating its revised auxiliary as a hypothesis that needs evidence. The same is true of the IIT side's call for denser sampling of early visual cortex. My complaint is with the protocol, not with the people.

A twist from the reanalysis

There is a further turn. In mid-2026 a group including Marta Garrido posted a reanalysis of the Cogitate MEG data from 100 participants using dynamic causal modelling, a method that estimates directed influences between regions instead of correlations. They "found support for both GNWT predictions of prefrontal ignition at stimulus onset and offset; noting that the latter was not previously observed in the original Cogitate analyses." They also tested IIT's sustained posterior integration, but a validation comparing face and object networks "did not yield strong evidence for face selectivity", which they say warrants "a conservative interpretation of the IIT findings in particular". It is a preprint and hasn't been peer reviewed.

If it holds up, the result is awkward for everyone. The GNWT side demoted offset ignition after it wasn't found; a reanalysis now finds it. Should it be promoted back? If centrality is set after the data, the honest answer is that it can't be either promoted or demoted on the basis of the data, because the data are what is being scored. This is the clearest illustration I know of why weights have to come first. An earlier paper from the same group, Bandara, Rowe and Garrido, modelled a different EEG dataset in which faces were sometimes seen and sometimes not, and found that "both theories could explain the data, with a very slight preference for frontal theories". Two careful reanalyses, two verdicts close to a tie.

The marking scheme analogy

Think of an examination in which the marking scheme lists the points a good answer must make, but doesn't say how many marks each point is worth. Two examiners read the same script. One thinks the third point is the heart of the question; the other thinks it is a detail. They agree on everything the script contains and disagree on the grade. No amount of rereading the script will resolve their disagreement, because it isn't about the script. Cogitate's predictions were a marking scheme without marks.

What a decisive test would need

Put the two reasons together and the shape of a better test appears. It needs predictions whose failure would count against a theory's core, by the theorists' own advance judgement, and it needs evidence that is present in one theory's worlds and absent in the other's. The first is a matter of commitments; the second is a matter of design. The likelihood ratio is the bridge between them.

For each result, ask two questions: how probable was this result if theory A is true, with its stated auxiliaries? And how probable was it if theory B is true, with its stated auxiliaries? The ratio of those two numbers is how much the result should shift the balance between them. A common reason has a ratio near one. Dignāga's valid types have high ratios: present where the theory holds, absent where it fails. If the theorists write down these probabilities in advance, the posterior verdict is a calculation, and nobody can relocate the failure without admitting that their advance probabilities were wrong.

The simulation lets you play with this. Set the prior odds between the theories, then set, for each of three results, how strongly the theorists expected it under each theory. Watch how the posterior moves. Three points stand out. Results that both theories expected barely move the needle, however many there are. A single result that one theory confidently expected and the other confidently forbade can move it a long way. And if the probabilities are set after the result, any posterior can be reached.

Designing for uncommon evidence

The second requirement is evidence that discriminates. The GNWT commentary named one: a contrast between conscious and unconscious processing of the same stimulus. Such designs already exist, using masking, inattentional blindness or stimuli at the threshold of visibility. They allow a researcher to ask whether a posterior signal tracks the stimulus or tracks the experience. If it tracks the experience, IIT gains; if it tracks the stimulus regardless of experience, the posterior decoding becomes what GNWT said it might be, a record of processing that precedes access.

There is a reason Cogitate chose visible stimuli instead, and it is a good one. Threshold designs make consciousness depend on attention and report, which entangles the variable of interest with the processes GNWT says are its basis. Critics of report-based paradigms, including researchers close to IIT, have argued that such designs are biased towards frontal theories, because what they measure is partly the machinery of report. So the choice of stimuli was itself a contested auxiliary. A design that both sides accept as discriminating may need several paradigms at once: visible stimuli with long durations, which favour IIT's maintenance predictions, and threshold stimuli without report, which favour GNWT's access predictions, with each side's likelihoods registered for each.

The coverage problem in numbers

One more design point, specific to intracranial work. The iEEG data came from patients whose electrodes were placed for clinical reasons. That gave 655 PFC electrodes for the onset-offset analysis and 12 in V1/V2 for IIT's synchrony analysis. In a fair test, statistical power should be comparable for each theory's decisive predictions, and with clinical placement it can't be. The honest response is to state, in the preregistration, the minimum coverage at which a null result will count as a failure, and to treat any analysis below that minimum as inconclusive by agreement. That rule would have converted IIT's synchrony null from a contested failure into an agreed non-test, which is what the evidence supports, and spared everyone an argument.

The consortium's call also names cross-sample consistency, and the decoding figure shows why. Each method reported something slightly different about prefrontal content: present but brief in iEEG, weaker in fMRI, and for orientation present only in MEG, where leakage from posterior sources could explain it. A framework that weights predictions should also say in advance how disagreements between methods will be resolved: whether a finding must appear in all three, in two of three, or in the method best placed to see it. Without that rule, a theory can lean on whichever method favours it, and its rival can lean on the others. That is one more place where a choice made afterwards can quietly decide a verdict.

The same structure in machines

The question this paper asks is not unique to consciousness science. It turns up wherever rival mechanistic stories compete to explain the same system. In machine learning interpretability, researchers propose that a network computes some function by a particular circuit, and test the proposal by intervening on parts of the network. A proposal can make a prediction about what happens when a component is removed, and the prediction can fail, and the proposal's defenders can say that the network has redundant backup components, or that the intervention moved activations outside their normal range. These are the same two relocations: to an auxiliary about how the mechanism shows up, and to the measurement. The field has the advantage that it can measure everything, all units, at every time step, with any intervention it likes. It still has the problem of centrality: which predictions of a proposed circuit are core, and which depend on choices about how it is measured.

The machine case shows that perfect coverage doesn't remove the problem. If Cogitate had had electrodes in every cortical column, the IIT reply about sparse coverage would have been unavailable, but the reply about frequency bands would remain, and so would the GNWT reply about attention to offsets. Coverage fixes the measurement joint. Only advance commitment fixes the theoretical joints.

Where this sits on RTM's ladder

This series uses a five-layer map, the Recursive Tantric Model (RTM), to locate questions about mind. L1, spanda, is the layer of dynamics, the activity that can be measured. L2, prakāśa, is manifestation, the fact of experience. L3, vimarśa, is reflexive awareness, the capacity of experience to grasp itself. L4, pratyabhijñā, is recognition. L5, cit, is consciousness as ground. The terms come from Kashmir Śaiva philosophy and are used here as labels for kinds of question, not as claims about metaphysics.

Cogitate measured L1. Both theories make claims about the relation between L1 and L2: GNWT says L2 contents are those broadcast across a particular L1 network; IIT says L2 is identical to a cause-effect structure that L1 activity reflects. The bridges from L1 measurement to L2 claims are exactly the auxiliaries this paper has been tracing. On RTM's map, an adversarial test is a test of bridges between layers, and its results can only be as decisive as the bridges are fixed. That is the ladder version of my thesis.

The strongest case against my answer

The strongest objection is that I am asking for a precision that consciousness science doesn't yet have, and that would distort it if forced. Theories of consciousness are young. Their links to measurement are uncertain in ways their founders can't fully specify. Asking Dehaene or Tononi to write down a probability for offset ignition or posterior gamma synchrony is asking them to make up a number, and made-up numbers dressed as likelihoods would give the verdict a false authority. On this view, the post hoc arguments about centrality aren't a failure of the protocol. They are the science working: a test reveals which bridges were weak, the theorists learn, and the next test is better. Cogitate did exactly what an adversarial collaboration should do, and my complaint confuses learning with losing.

There is a second objection, from the other direction. Perhaps the results did decide, and the theorists simply wouldn't accept it. On this view, the absence of offset ignition across 655 electrodes is a clear failure of a prediction GNWT's leader signed, and the absence of sustained synchrony, together with the failure of the "maximum integrated information" prediction, is a clear failure for IIT. Calling these inconclusive lets both theories off. A tougher reading would say both theories were damaged, and that the right response is to lower confidence in both, as the joint paper's language of "substantially challenging key tenets of both theories" suggests.

I take both objections seriously. To the first, my reply is that the numbers don't have to be precise to be useful. A theorist who writes "about 0.6" for offset ignition has made a commitment that "not core" later contradicts; the value of the number is that it is checkable, not that it is exact. And the alternative to made-up numbers isn't no numbers: it is numbers made up afterwards, implicitly, in the choice of what to call core. To the second, my reply is that I agree both theories should lose some confidence, and the worked likelihood case shows how much. What I deny is that the results falsified either decisively, which is the question I set out to answer. A decisive falsification needs a central prediction, a discriminating result and adequate power together, and no result in Cogitate had all three.

A counterexample and a failure boundary

A counterexample to my thesis would be a well-designed adversarial test, with centrality weights and likelihoods registered in advance, that still produced a dispute in which each side relocated failures. If that happened, my diagnosis would be wrong: the problem would lie deeper than the protocol, perhaps in the theories' lack of determinate content. I don't know of any adversarial test of theories of consciousness that has registered numeric weights. Until one does, the counterexample is hypothetical.

The failure boundary is this. My thesis applies to theories that make claims about measurable brain activity through stated bridges. It doesn't apply to a theory whose core claim implies nothing measurable even with bridges. If IIT's core, as some critics argue, is unfalsifiable because integrated information can't be computed for brains, then no weighting scheme will make it testable, and the problem with Cogitate's IIT predictions would be that they never touched the core at all. I think that view is too strong, because IIT's core does imply that posterior cortex is the main substrate in humans, which is measurable. But if it is right, my proposal fixes GNWT's side of the test and leaves IIT's untouched.

A negative control

Here is a check on my own reasoning. If the problem were simply that theorists defend their theories, one would expect both sides to relocate every failure. They didn't. The IIT statement accepts that the synchrony prediction failed as tested and offers reasons; it doesn't call it peripheral. The GNWT commentary accepts that PFC failed to represent some content dimensions well. And in the decoding results, where the theories' predictions were fixed and both were partly met, neither side disputed centrality at all. The relocations cluster where the protocol left room: unweighted predictions with contested bridges. That is what my thesis predicts, and it is evidence that the cause is structural.

What would change my mind

What I don't know

I don't know how either theory's leaders would weight their predictions if asked now; I have read their published words and inferred, and the weights in my figure are my reconstruction. I don't know whether the MEG reanalysis that found offset ignition will survive review, or whether dynamic causal modelling of MEG can detect prefrontal feedback reliably enough to settle the question. I don't know whether Dehaene's refractory-period result has been published in full. I don't know whether a conscious versus unconscious contrast acceptable to IIT's proponents can be built, given their argument that report and attention contaminate such contrasts. And I don't know whether centrality weights would survive contact with a surprising result, or whether theorists would then dispute the weights themselves, which is the counterexample I can't yet rule out.

A preregistration of commitments

Here is the instrument this paper proposes, offered for the next adversarial collaboration and usable by anyone reading one. It adds four lines to a standard preregistration. Each line is a commitment that can be checked after the data.

The sorter lets you try the distinction between relocations that a protocol should allow and those it should flag.

Claim Ledger

Dated predictions

Glossary

Adversarial collaboration

A study designed jointly by proponents of rival theories, who agree in advance on predictions and their interpretation.

Anaikāntika (अनैकान्तिक)

Inconclusive: a reason that does not lead to a single conclusion.

Asādhāraṇa (असाधारण)

Uncommon: a reason present in neither similar nor dissimilar cases.

Auxiliary hypothesis

An assumption that links a theory’s core to a measurable prediction.

Bayes factor

The ratio of the probability of the data under one hypothesis to that under another.

Global neuronal workspace theory

The theory that contents become conscious when broadcast through a prefrontal-parietal network after ignition.

Hetu (हेतु)

The reason or evidence in an Indian inference.

Hetucakra (हेतुचक्र)

Dignāga’s wheel of reasons: nine types from presence in similar and dissimilar cases.

Ignition

In GNWT, a sudden self-sustaining burst of activity that broadcasts a content.

Integrated information theory

The theory that consciousness is the cause-effect structure of a maximally integrated system.

Likelihood ratio

How much more probable a result is under one hypothesis than another.

Sādhāraṇa (साधारण)

Common: a reason present in both similar and dissimilar cases.

Sapakṣa (सपक्ष)

Similar cases, in which the property to be proved is known to hold.

Spanda (स्पन्द)

Vibration or dynamism; in RTM, the layer of measurable activity, L1.

Trairūpya (त्रैरूप्य)

The three conditions a valid reason must meet.

Vipakṣa (विपक्ष)

Dissimilar cases, in which the property is known to be absent.

Viruddha (विरुद्ध)

Contradictory: a reason found only where the property is absent.

Umpire's call, revisited

Raghu wasn't satisfied. "If both batsmen stayed in," he said, "then the machine wasn't good enough." Meera said the machine was fine. "The problem was that nobody agreed beforehand how close was close enough. So afterwards, each side said the ball was missing." Raghu thought about it. "Then next time," he said, "make them sign the margin."

That is my answer. The posterior synchrony and prefrontal offset results failed to falsify either theory decisively for three connected reasons. The predictions were fixed but their centrality wasn't, so each side could argue after the data that the failed prediction was peripheral or rested on a bridge it didn't own: GNWT by calling offset ignition non-core, IIT by locating the synchrony failure in the frequency band, region selection and the 12 electrodes in early visual cortex. Many of the confirming results were common reasons in Dignāga's sense, present whether either theory was true, because all stimuli were clearly visible. And the decisive absences were measured with very different power, so that the better-measured one was the easier one to demote. None of this is a failure of honesty. It is a gap in the design of adversarial tests, and the consortium's own call for a framework that accounts for prediction centrality points straight at it.

The practical step is the commitments sheet. When the next adversarial collaboration publishes its protocol, read it for four things: weights on each prediction, likelihoods under each theory, named owners for each bridge, and a power floor for nulls. If those are missing, expect another umpire's call, and read any later claim that a failed prediction "was never core" as a revision, to be weighed on its new evidence. If they are present, the experiment can do what Cogitate set out to do and come close to doing: make a theory of consciousness lose.

Sources

Primary research

• Cogitate Consortium et al. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature 642, 133–142. pmc.ncbi.nlm.nih.gov/articles/PMC12137136/
• Cogitate Consortium (2023). An adversarial collaboration protocol for testing contrasting predictions of global neuronal workspace and integrated information theory. PLOS ONE. journals.plos.org/plosone/article?id=10.1371/journal.pone.0268577
• Melloni, L., Mudrik, L., Pitts, M. and Koch, C. (2021). Making the hard problem of consciousness easier. Science 372, 911–912. pubmed.ncbi.nlm.nih.gov/34045342/
• Cogitate Consortium. An adversarial collaboration to critically evaluate theories of consciousness (preprint). people.socsci.tau.ac.il/mu/mudriklab/files/2024/04/Cogitate_et_al_prep
• Boly, M., Koch, C. and Tononi, G. IIT proponents’ statement on the Cogitate results (supplement). www.iit.wiki/validation/arc-cogitate-supplement
• Naccache, L., Sergent, C., Dehaene, S. and colleagues (2025). Commentary on the Cogitate results. Neuroscience of Consciousness, niaf037. academic.oup.com/nc/article/2025/1/niaf037/8280147
• Dehaene, S. (2026). Collège de France course, lecture of 27 February 2026 (slides). www.college-de-france.fr/sites/default/files/media/document/2026-02/de
• Testing the spatiotemporal predictions of global neuronal workspace theory and integrated information theory (2026). MEG reanalysis with dynamic causal modelling, preprint. sciety.org/articles/activity/10.64898/2026.07.16.738188
• Bandara, K. H., Rowe, E. G. and Garrido, M. I. (2025). Computational modelling shows evidence in support of both sensory and frontal theories of consciousness. Neuroscience of Consciousness, niaf033. academic.oup.com/nc/article/2025/1/niaf033/8262378

Primary texts and translations

• Dignāga’s hetucakra, overview with the nine types. en.wikipedia.org/wiki/Hetucakra
• Iwata, T. Study of the hetucakra, Austrian Academy of Sciences.

Secondary or leads

• Dignāga, Encyclopaedia Britannica. www.britannica.com/biography/Dignaga
• Dharmakīrti, Stanford Encyclopedia of Philosophy. plato.stanford.edu/archives/fall2018/entries/dharmakiirti/
• Urbain-Jean-Joseph Le Verrier, Encyclopaedia Britannica. www.britannica.com/biography/Urbain-Jean-Joseph-Le-Verrier
• The hunt for Vulcan, NOVA (PBS). www.pbs.org/wgbh/nova/article/hunt-for-vulcan/
• Seven important dates in Mercury history, Encyclopaedia Britannica. www.britannica.com/list/7-important-dates-in-mercury-history
• Siegfried, T. Trusting gravity isn’t always the best astronomy policy, Science News (quoting Lahav and Massimi). www.sciencenews.org/blog/context/trust-gravity-isnt-always-best-astron
• Lahav, O. and Massimi, M. (2014). Dark energy, paradigm shifts, and the role of evidence. arXiv:1405.1350. arxiv.org/abs/1405.1350
• A 25-year-old bet about consciousness has finally been settled, Scientific American. www.scientificamerican.com/article/a-25-year-old-bet-about-consciousne
• Prominent consciousness theory is slammed as bogus science, Scientific American. www.scientificamerican.com/article/prominent-consciousness-theory-is-s