Topics — What a work is about. Nucleation holds whatever reading the owner files here, whether or not an Anthology of the SOTA topic could also hold it; instruction for ML practice belongs to the anthology (ADR-048). The words are seeded from the clusters the reading-time triage of 2026-09-25 found, and grow by decision.
Cognition — the mind as information processing — perception, memory, reasoning and cognitive strategies — as studied by cognitive science and philosophy of mind (group: philosophy).
145 of 1352 THEORY documents. Back to the full index.
| # | Title | Summary | Status |
|---|---|---|---|
| THEORY-013 v3 | Once context-dependent marginals are separated from contextuality, the behavioural, social and word-meaning data published as contextual show none | Dzhafarov, Zhang & Kujala (2015), LIT-264 — Contextuality-by-Default applied to five data sets, including primed word combinations from the quantum-cognition literature. Seeded, not yet read closely. The authors call general absence a working hypothesis, and later work in the same programme is reported to find behavioural contextuality. | Proposed |
| THEORY-023 v4 | Current evidence cannot settle whether an AI system is conscious: mimicry undercuts behavioural evidence and architectural indicators presuppose the disputed computational functionalism, so the dispute is over what counts as evidence as well as the answer | Birch (2025), LIT-111: behavioural markers are “gamed”, and architectural markers are “Janus-faced”, giving a two-level deadlock. Schwitzgebel (LIT-191) argues mimicry undercuts but does not rebut, and that the functionalist theories face minimal instantiation. Seth (LIT-135) makes conscious AI conditional on functionalism plus substrate flexibility. Butlin et al. (LIT-056) assume functionalism. The readings agree the question is currently unsettleable. They contest whether that is permanent and which way the prior leans. This says nothing about whether any AI is conscious. Block (2002, LIT-639) argues that for a perfect functional duplicate on another substrate we lack even a conception of the evidence; Block (2009, LIT-632) states the biological premise from the other side. | Proposed |
| THEORY-031 | The likelihood-ratio detector is optimal under the payoff, Neyman–Pearson and posterior criteria, which differ only in its threshold, so performance is an ROC and the criterion a point on it; d′ is criterion-free only for equal-variance normal signal and noise | Peterson & Birdsall (1953), LIT-364 — the threshold on ℓ = f_SN/f_N is optimal for expected value and at fixed false-alarm rate, and the two optimum families coincide (Part I Thms 1–7). The ROC’s slope equals the threshold (Thm 8). Tanner & Swets (1954), LIT-368, carry this over to the human observer as d′ plus a criterion. Stanislaw & Todorov (1999), LIT-367, state that d′ varies with the criterion unless the variances are equal. The separation of sensitivity from bias is a property of the model, not of the data. | Active |
| THEORY-043 v6 | A mind of mindless agents and a mind of minded agents are one functionalist move run in two directions, and what separates them is an anti-nesting principle or an architectural criterion, neither yet principled | Minsky’s society of mind builds a mind downward, out of agents that are not minds. Schwitzgebel (LIT-159) builds a candidate mind upward, out of agents that are. The readings agree that agency survives both directions: Block grants a homunculi-head beliefs, List grants organised groups agency and awareness, Levin finds nested goal-pursuing collectives (his “Selves”) at every scale. The dispute narrows to phenomenal consciousness, and every published way to accept the downward move while refusing the upward one is either an anti-nesting principle (Putnam’s stipulation, IIT’s exclusion postulate, Kammerer’s functional version), each introduced for or justified weakly by the verdict it gives, or an architectural criterion that would count against a loosely coupled society of mind as much as against a nation. Proposed: an account of where a dispute stands, from philosophical argument and no measurement. | Proposed |
| THEORY-044 v2 | In relational frame theory the self is verbal behaviour’s content, its process and its context, and the self as context is a perspective learned from deictic relations that cannot itself be found as content | Barnes-Holmes et al. (LIT-541) define three selves: the conceptualized self, the content of verbal relations about oneself; the knowing self, their ongoing process; and the transcendent self, their context. Hayes et al. (LIT-555) say the third arises from deictic frames (I–you, here–there, now–then) as “a sense of self as a locus or perspective”, and is “a context for verbal knowing, not the content”, so its limits “cannot be consciously known”. The account maps the three onto the record’s holdings: the content is the narrative self (Dennett, Gallagher, Strawson), and the context is what Hume could not find among his perceptions and what Metzinger calls a transparent perspective. It differs from the minimal self (Gallagher) on origin: RFT says the perspective is learned in language. The evidence for the learned origin is cited, not read, so Proposed. The clinical claim is THEORY-051. | Proposed |
| THEORY-048 | In explaining psychiatric disorder, biological findings implement psychological functions rather than replacing them | Kendler (2008), LIT-539: following Marr, a lower-level theory either replaces a higher-level construct (temperature by mean kinetic energy) or implements it, and in psychiatry biology “will implement but not replace psychology”. The thesis is argued from neuroscience’s successes in perception, learning and memory, not from a psychiatric case, so it is Proposed. It is a thesis about explanation, not about distinct mental causes: it survives Pernu’s argument (LIT-147) that the neural evidence supports only mental causation by identity. Kendler 2005 (LIT-552), which argued the mind–brain propositions this paper develops, is unread. | Proposed |
| THEORY-050 | Psychological inflexibility, in which verbal relations give private events aversive functions and make experiential avoidance a general strategy, is a transdiagnostic process in psychopathology | Hayes et al. (LIT-555) model psychopathology as inflexibility: cognitive fusion supports experiential avoidance, which narrows behaviour. Barnes-Holmes et al. (LIT-541) give the mechanism: once a word stands in derived relation to an event it can carry that event’s aversive functions, so private events become worth avoiding. Kashdan and Rottenberg (LIT-553) find inflexibility across depression and anxiety in cognitive, behavioural and physiological systems, and say its causal role is unknown. The evidence is correlational, small, mostly self-report and mostly from the model’s authors, and the meta-analysis that would test it (LIT-543) is unread, so Proposed. It is compatible with the loop account of disorder, not a rival to it. | Proposed |
| THEORY-055 | An appraisal of an event’s pertinence to the agent’s concerns elicits the emotion, as a state of action readiness, and emotions differ by their mode of action readiness | Frijda, Ridderinkhof & Rietveld (2014), LIT-540, Frijda’s own late statement of the theory of The Emotions (LIT-546, unread). An automatic appraisal of how an event bears on one’s concerns elicits a state of action readiness: a motive to change one’s relation to the object, with control precedence. Emotions differ by mode (hostility for anger, avoidance for fear), and “in the absence of concern pertinence … no emotion”. It is the record’s only read statement of an appraisal theory, because Lazarus (LIT-550) is unread. Both the elicitation claim and the claim that emotions differ by mode are argued, and neither is tested. The second is partly definitional within the theory. Rivals THEORY-057: both accounts let instances vary and group them by function, but the appraisal account puts the unity in the agent’s readiness and the constructionist account puts it in the perceiver’s concept. Proposed. | Proposed |
| THEORY-056 v3 | Emotion regulation is not a process separate from emotion generation: most of it is one motive state checking another, with no distinct regulating system above them | Frijda, Ridderinkhof & Rietveld (2014), LIT-540, deny “the exclusivity of emotion regulation processes”. Appraising the consequences of an impulse raises a competing readiness, as when fear of retaliation holds back anger, so control is the interaction of motive states and not reason governing emotion. Barrett (2017), LIT-542, reaches the second half from the opposite side of the emotion debate: automaticity and control are “different brain modes … not two battling brain systems”, and she offers as a hypothesis that reappraisal works by the same predictions that construct emotion. Two rival theories of what an emotion is agree on this. Gross (1998), LIT-554, poses the question and keeps the distinction. Proposed: argued from examples and from architecture, untested. | Proposed |
| THEORY-057 | Emotion categories are populations of variable instances grouped by the perceiver’s concepts, not natural kinds with a shared neural or appraisal mechanism | Barrett (2017), LIT-542: degeneracy, population thinking, and a table of findings in which no region, network, pattern or neuron is consistent for an emotion category. From these she concludes that categories such as fear are populations of variable instances, grouped by a goal and constructed by categorization with a concept. She rejects basic-emotion and causal appraisal theories as positing a “Lockean essence”. This is the record’s only read statement of the view. The record’s philosophy holdings sharpen it. Pernu (LIT-147) shows that many-to-one realisation at a fine grain does not rule out one variable at a coarser grain, so degeneracy alone does not carry the conclusion, and Table 1 has to. Searle (LIT-482) shows that Barrett’s comparison with money is to observer-relative function, not to institutions. Rivals the appraisal account, THEORY-055. Proposed: one theoretical paper, no independent test, and the reply unread. | Proposed |
| THEORY-061 | In evidence-accumulation tasks, goals and instructions act on a decision by setting the parameters of the accumulation (drift, starting point or baseline, and threshold), not by specifying the response, and this holds for when to act as well as for which | Bogacz et al. (LIT-569) define cognitive control, in a two-choice task, as the adjustment of drift, starting point and threshold to maximise a utility. Ratcliff and McKoon (LIT-576) find that the manipulations behave that way: speed or accuracy instructions are fitted by boundary separation alone, stimulus proportions mainly by starting point, difficulty by drift alone. Schurger et al. (LIT-605) carry the picture to the decision of when to move: the task goal raises the baseline toward threshold and leaves the moment to fluctuation. The convergence is real; the evidence is group fits on a few paradigms, so Proposed. It does not say who sets the parameters, or how. | Proposed |
| THEORY-063 | The unity of experience is the integration of one global world-model into a single functional cluster, and it does not require a self-model | Metzinger’s précis of Being No One (2005), LIT-568: the globality constraint makes every conscious content part of one world-model. On the functional level that is “highest-order binding” into a single causal role, and on the neural level the prediction of “one single area of maximal causal density”, after Tononi and Edelman’s dynamic core. “What you subjectively experience upon experiencing your world as coherent is the high internal correlation strength between a subset of physical events in your own brain.” A unified world comes before a self: the minimal case is “Selfless Snapshot Consciousness”, and fully depersonalized states are “non-subjective consciousness”. The neural half is a cited hypothesis, the précis says little is known, and how the cluster becomes one subject’s is left open. | Proposed |
| THEORY-064 | What is experienced as a self is the content of a self-model the system cannot recognise as a model, and nothing beyond such models is a self | Metzinger’s précis of Being No One (2005), LIT-568, with his encyclopedia article “Self models” (LIT-450): a phenomenal self is a transparent self-model, one whose representational character “is not accessible to subjective experience”, so that the system experiences its content “simply as ourselves”. Selfhood is “a function realized by a lack of information”, and “no such things as selves exist in the world”. Transparency is graded, and the perspective itself is modelled, as a phenomenal model of the intentionality relation. Two things are offered: a prediction (a wholly opaque self-model would abolish selfhood) and an ontological stipulation. The précis adds no evidence; its case studies are in the book, unread. | Proposed |
| THEORY-068 v2 | In a two-alternative choice made by two competing leaky accumulators, the decision is optimal only when leak and mutual inhibition balance, so that the network integrates the difference of the evidence without loss; imbalance bounds accuracy at any fixed time and weights early or late evidence | Bogacz et al. (LIT-569) prove that the pure drift diffusion model is the continuum limit of the sequential probability ratio test, and so the optimal two-choice decider, and that the linearised Usher–McClelland network reduces to it exactly when leak equals inhibition (and, for free response, when both are also large). Usher and McClelland (LIT-565) derived the difference of the two units as an Ornstein–Uhlenbeck process with net leak K = k − β: K ≠ 0 caps accuracy, leak-dominant weights late evidence and inhibition-dominant early. Active because the core is derivation, checkable on paper. What it does not say: that brains are balanced, that the result carries to more than two options or to non-stationary evidence, or that the primacy and recency seen in six participants are explained by it. | Active |
| THEORY-070 | How much cognitive control to exert, and on what, is decided by a cost-benefit computation (the expected value of control) in a specification system distinct from the structures that implement the control | Shenhav, Botvinick & Cohen (LIT-585) unify dACC’s many reported functions as one: deciding whether, where and how much control to allocate. A control signal has an identity (a task set, a threshold, an attention template) and an intensity; its expected value is its probability-weighted payoff, with future value discounted, minus an intrinsic cost rising with intensity. dACC monitors and specifies; lateral prefrontal and subcortical structures implement. The paper reviews evidence and fits nothing, so Proposed. It does not say what the cost is, how EVC is computed, or that it governs emotion regulation. | Proposed |
| THEORY-088 v2 | Slow large-scale activity in human cortex is organised as in a pattern-forming medium: its flow is set by interacting spiral defects centred on phase singularities, which sit on boundaries between functional regions, annihilate with spirals of opposite rotation and repel those of the same | Xu, Long, Feng & Gong (2023), LIT-638: in Human Connectome Project fMRI band-passed to 0.01–0.1 Hz, the cortical phase field holds about 16–19 spirals at a time around phase singularities, concentrated on boundaries between functional regions; their centres drift superdiffusively, opposite spirals annihilate (51% fully, 46% partly) and like ones repel (2.6%), and their positions and senses decode task condition above a raw-amplitude decoder. One source, fMRI only, at a spatial scale chosen because it maximised that decoding; the routing of activity flow is read from phase gradients and a phase-oscillator model, never perturbed. The kinship with Belousov–Zhabotinsky spirals and pattern-formation defects is the record’s analogy, not the paper’s. | Proposed |
| THEORY-089 | The content of a mental representation has two factors, a narrow conceptual role inside the head and a referential factor relating it to the world, and a long-arm role theory that gets Twin Earth truth conditions right reduces to the same two | Block (1986), LIT-648, and Block (1987), LIT-625. Meaning has a conceptual-role component “entirely ‘in the head’”, an expression’s “total causal role, abstractly described” in inference and decision, which twins share and which explains their behaviour, and an external component relating representations to referents and truth conditions. Conceptual role fixes the nature of the referential factor, not reference. A one-factor theory with long-arm roles must, to get Twin Earth truth conditions right, import normal context, causal chains of borrowed reference and experts; chop off what lies outside the skin and the rest is the short-arm role, so the difference is “merely verbal”. Social meaning is assigned to the referential factor or to a third, social factor, and is only sketched. Proposed: no identity conditions for roles, and narrow content is given room, not shown explanatory. | Proposed |
| THEORY-091 | A deficit in which consciousness and a capacity are lost together does not show that the capacity is a function of phenomenal consciousness, since access is lost too and the capacity is one the machinery of access has | Block (1995), LIT-643, §6. Marcel, Flanagan, van Gulick and Schacter argue from the thirsty blindsight patient who does not reach for the water, and Searle from Penfield’s petit mal automatisms, that a function of phenomenal consciousness is to let information guide action. In blindsight access is missing too, and in petit mal nothing shows experience is missing at all. “The fallacy is: an obvious function of the machinery of access-consciousness is illicitly transferred to phenomenal consciousness.” That experience causes access, so that the function is experience’s after all, is “a bold hypothesis, not a fallacy”, and the deficit cannot decide it. Proposed: single-source, and the distinction it rests on is argued from conceivable cases. | Proposed |
| THEORY-092 v3 | Behavioural capacity does not fix intelligence: a system with the capacity to respond sensibly to every possible input can lack it, so whether behaviour is intelligent, or the mark of any mental property, depends partly on how it is produced | Block (1981), LIT-642: the strongest behaviourist conception of intelligence, a capacity to give sensible responses to any sequence of inputs, “whatever they may be”, is met by a machine that stores every sensible conversation of a test’s length and looks up its next reply, yet “the machine has the intelligence of a toaster”. So psychologism is true: two systems alike in all actual and counterfactual behaviour can differ in intelligence through the processing that mediates it. Block (1995), LIT-630, n. 1: the same machine refutes the wider claim that there is no mental difference without a behavioural one. It is a claim about constitution, not evidence: performance stays evidence for competence. It names no required kind of processing, and its machine is a lookup table, not a trained network. | Proposed |
| THEORY-094 | Neither report nor no-report paradigms isolate the neural basis of consciousness, because post-perceptual cognition that tracks the percept persists without report; only a design that removes differential post-perceptual cognition does | Block (2007), LIT-646, framed the methodological puzzle: the neural basis of experience can be found only by deciding whether it includes the machinery of access, and the evidence comes through report. The no-report paradigm was the answer. Block (2019), LIT-637, argues it has “a fatal flaw”: subjects not asked to report still think about what they see, and that thought is aligned with the percept (the “bored monkey” problem). Eliminating report isolates the basis of experience “only if it eliminates post-perceptual cognitive processing” correlated with the percept. Brascamp et al.’s inconspicuous rivalry, whose switches subjects detected at chance, is the one such design Block offers. The method is specific to rivalry so far, and Block says there may be no general solution. | Proposed |
| THEORY-095 v3 | Phenomenal character is not fixed by long-arm functional role or by externally individuated representational content: on Inverted Earth it stays the same while both invert, so only internal accounts, short-arm functional or physical, remain open | Block (1990), LIT-634: on Inverted Earth colours and colour words are inverted, so a subject fitted with inverting lenses notices nothing. After long residence, or for a twin with a molecule-for-molecule identical brain raised there, the intentional content of the experience of the sky inverts and so does its long-arm functional role, while qualitative content stays the same; Q = F₁, Q = F₂ and F₁ ≠ F₂ breaks transitivity of identity. Block (2003), LIT-635, turns the case on externalist representationism: for colour there is no shared appearance property left to explain the sameness, and long-term memory blocks the gradual-shift reply, so “mental paint” is what stays the same. Proposed: both rest on an externalist shift of content and on first-person sameness, and Block concedes that a short-arm functional theory of qualia would escape. | Proposed |
| THEORY-096 | Perceptual consciousness is global availability: a content becomes conscious when the activity it drives into a core of long-range hub neurons, prefrontal among them, crosses a threshold and ignites it, so that the network amplifies, sustains and broadcasts that one content to the specialised processors; below threshold it leaves a decaying wave and is not experienced | Mashour, Roelfsema, Changeux & Dehaene (2020), LIT-631, a review by the theory’s proponents. Conscious access is “a non-linear network ignition” that amplifies and sustains a representation so that local processors can access it, in a distributed core of long-range layer II/III and V pyramidal neurons with prefrontal cortex a key hub but “not … the exclusive territory”. Ignition is exclusive: one content at a time. Evidence: seen and unseen trials diverge after about 200 ms across modalities; in van Vugt et al.’s monkeys, missed stimuli drove V1 and V4 but only reported ones, and false alarms, ignited sustained prefrontal firing, and the signal-detection threshold is the ignition threshold. The review adopts the premise that global availability “is what we subjectively experience as a conscious state”, and calls experience without access untested; that premise, not the access data, is where it rivals overflow and the sensory account. Single source, second-hand evidence, and the P3 question conceded open. | Proposed |
| THEORY-097 | Long-range coordination in the human forebrain is carried by brief ripples that co-occur at distant sites: during them neurons co-fire more whatever the distance between them, and at retrieval they carry stimulus-specific reinstatement of encoding co-firing | Verzhbinsky, Daume, Cheng, Rutishauser & Halgren (2026), LIT-633: in 1,373 single units from five bilateral forebrain regions of 35 epilepsy patients doing a Sternberg task, ripples (70–100 Hz, about 70 ms) co-occur between bundles at a median 5% at every tract distance from 71 to 223 mm, and unit pairs co-fire a median 34% more during co-ripples than in no-ripple periods, beyond a rate-only null and with no loss over distance. The gain rises with load (13% in maintenance, 19% at retrieval), and encoding co-firing for a stimulus recurs at retrieval in 0.29% of cell-pair trials in co-ripples against 0.14% outside them. Correlational throughout: one task, two load levels, clinically placed electrodes, nothing perturbed. | Proposed |
| THEORY-098 v4 | The neural basis of perceptual consciousness does not include the prefrontal machinery of cognitive access: it is recurrent activity in sensory cortex | Block’s “back of the head” position, from three papers. Block (2005), LIT-629: a phenomenal NCC (for seen motion, recurrent MT/V5–V1 activity) and an access NCC (frontoparietal broadcasting), with winner-take-all access letting a recurrent sensory representation lose and stay unaccessed. Block (2007), LIT-646: the mesh argument. If workspace activation is not part of the basis of experience, strong losing sensory coalitions explain why experience overflows access; if it is, nothing does. Block (2019), LIT-637: no-report designs do not settle the prefrontal question, because thought about the percept persists, and one design that removes it found no prefrontal difference by fMRI, which Block calls evidence against prefrontalism while granting that ECoG could still support it. Proposed: an inference to the best explanation, with every experiment second-hand and the decisive test not yet run. | Proposed |
| THEORY-100 | A neural correlate of conscious report does not require laminar cortex: in crows it is carried by the nidopallium caudolaterale, a nuclear pallial region whose neurons follow the reported percept rather than the stimulus | Maldarelli & Güntürkün (2025), LIT-650, reviewing Nieder et al. (2020): in crows reporting near-threshold stimuli, single neurons in the nidopallium caudolaterale (NCL), the functional analogue of prefrontal cortex, fired with the “seen” report whether or not a stimulus was shown and were silent on “not seen” reports. NCL is nuclear, not layered; the cortex-like layers and columns found in birds are confined to the sensory pallium. So a correlate of conscious report does not need laminar cortex. The crow evidence is second-hand, from one laboratory and not in the record, and a report paradigm cannot separate access from phenomenal consciousness. | Proposed |
| THEORY-102 v3 | In Vishne, Gerber, Knight and Deouell’s intracranial recordings from ten epilepsy patients viewing clearly visible images for up to 1.5 s, the pattern of high-frequency activity across ventral temporal cortex carried image category, and exemplar identity among 60 images, in a time-invariant code for as long as the image stayed on while amplitude fell about 80%; at the sampled prefrontal and parietal sites, category content was a burst about 150 to 600 ms after onset that did not track duration | Vishne, Gerber, Knight & Deouell (2023), LIT-644: in 430 visually responsive subdural electrodes from 10 patients viewing images for 300–1,500 ms with no report on analysed trials, high-frequency activity and single-electrode category selectivity fell about 80% from peak by 800–900 ms, while the pattern across ventral temporal electrodes kept near-ceiling, time-invariant category and exemplar information that ended a fixed delay after the image went off. Prefrontal and parietal information was confined to about 150–600 ms and ignored duration. This met the predictions both IIT’s and the global workspace’s proponents registered for the paradigm, so it did not discriminate them. Above- threshold images only, sparse right-lateralised frontal coverage, one frequency band; it does not show the code tracks experience rather than input. | Active |
| THEORY-103 v3 | Perceptual consciousness overflows cognitive access: in partial-report tasks more items are experienced than the roughly four that working memory holds, though any one of them can be accessed when cued | Block (2007), LIT-646, and Block (2011), LIT-640. Subjects shown 8–32 items say they see all or almost all of them, report about 4, and when cued after offset can report or compare most of any cued subset: capacity 6–7 of 8 up to 1.5 s (Landman et al.), and 7, 6 and 5 of 8 at 1, 2.5 and 4 s (fragile visual short-term memory), against a working memory that stays at about 4. Block reads the surplus as experienced. The rival “sparse” reading holds it in an unconscious store plus a conscious generic or fragmentary image, and Block argues there is no independent evidence for either posit. Proposed: every experiment is second-hand, and that the surplus is conscious is the contested step. Overflow means most items are unaccessed, not that any is inaccessible. | Proposed |
| THEORY-124 v2 | In Posani et al.’s analysis of the IBL Brainwide Map, recorded while trained mice performed one visual decision task, the eight-variable selectivity profiles of neurons within one cortical area clustered beyond a covariance-matched Gaussian null only in VISp, AUDp and SSp-ul of about 20 areas tested, while profiles pooled across areas clustered along anatomical lines; and over the task conditions each area distinguishes, at least 95% of random balanced dichotomies were linearly decodable above a label-shuffled null in 15 of 16 areas, a lenient criterion | Posani et al. (2026), LIT-708: in about 14,000 cortical units of the IBL Brainwide Map, the 8-variable selectivity profiles of 4,617 selective neurons beat a covariance-matched Gaussian null for clustering within a region only in VISp, AUDp and SSp-ul. Elsewhere the profiles form an elongated, unclustered cloud. After merging conditions a region cannot tell apart (leaving 5 to 16 of 16), ≥ 95% of random balanced dichotomies are decodable above a shuffle null in 15 of 16 regions, GU being 0.82. Types do appear when regions are pooled, aligned with anatomy. One task, overtrained mice, a lenient separability criterion. | Active |
| THEORY-125 | A group can itself be in an emotional state that is not its members’ states in aggregation, when its distributed routines integrate the capacities an emotion requires into one action-oriented representation; because emotions are functional kinds such a state need not be felt, and no actual case has yet been shown | Huebner (2010), LIT-702: a genuinely collective emotion is a state of a group that “could not be exhaustively characterized in terms of intentional states of individuals in aggregation”. Fear requires danger detection, redeployment of attention, system-wide preparatory change and their coordination into one state with indicative content and imperative force; a distributed system can integrate routines that way. Some individual emotions are non-conscious, so a group’s need not be felt, and emotion kinds are functional (“like vodka”), so lacking homologous mechanisms does not exclude groups. The paper establishes a possibility, not an instance: its worked case, the USS Palau crew, turns out to be “collective detection of danger in the absence of a fear response”, and the replacement, the 2008 McCain–Palin campaign, is offered as “intuitively plausible” pending anthropological data. Proposed. | Proposed |
| THEORY-137 v2 | Attributions of mental states to institutions, such as the knowledge that the collective knowledge doctrine of corporate criminal law attributes to a corporation no member of which held it, work largely by shaping the institutions to which they are made rather than by describing states they already had: the attribution holds the institution to a norm, the institution builds structures to meet it, and it becomes more predictable from the intentional stance, whether or not institutions really have minds | Tollefsen (2025), LIT-736. Both sides of the group-mind debate treat ascription as tracking pre-existing causes. On the regulative view of folk psychology it shapes its target. In United States v. Bank of New England a court held a bank to know what no employee knew, and banks then built systems to detect reportable deposits. Narratives regulate too. Proposed: the regulative view is assumed, and the effect on banks is asserted, not documented. | Proposed |
| THEORY-155 v2 | Iterated learning by Bayesian agents who sample from the posterior converges, in an ergodic chain, to the shared prior | Griffiths & Kalish (2007), LIT-769, read in NOTE-599: a chain of learners who share a prior and each sample a hypothesis from their posterior is a Gibbs sampler, so the probability a learner holds a hypothesis converges to its prior probability, whatever the amount of data passed on; with MAP learners the outcome centres on the prior’s mode but depends on data amount and noise, and can amplify weak biases. It is a theorem about an idealized chain. It does not say human chains reach the prior, and Kirby, Cornish & Smith’s laboratory chains, LIT-771, are consistent with it without testing it. Whether human chains do is QUESTION-137. | Active |
| THEORY-163 v3 | In Zaslavsky, Kemp, Regier and Tishby’s analysis, the colour-naming systems of the World Color Survey languages and English lie near the information-bottleneck bound for compressing perceptual meanings into words, at a trade-off parameter just above 1 and nearer than hue-rotated variants of themselves, with the need distribution over meanings estimated from the naming data | Zaslavsky, Kemp, Regier & Tishby (2018), LIT-810, read in NOTE-631: treating a colour lexicon as an encoder from Gaussian perceptual meanings to words, the World Color Survey languages and English lie near the IB curve of LIT-338 at β ≈ 1.03, beat hue-rotated variants of themselves, and their soft, partly inconsistent naming is what IB optima look like. It does not show that languages are driven to the bound by any process, and the need distribution that makes the fit best is estimated from the naming data themselves. | Active |
| THEORY-166 v3 | In the 66 Pew surveys of 2001–2011 that varied the order of two back-to-back binary questions, question order changed the answers but not the probability of giving the same answer to both, the equality a Lüders projection model predicts for any projectors; a classical model with order events can meet the same equality, so the regularity does not by itself favour quantum over classical probability | Wang, Solloway, Shiffrin & Busemeyer (2014), LIT-795, on the QQ equality derived in LIT-834: across the 66 Pew surveys of 2001-2011 that varied the order of two questions, question order shifts the answers (p = 0.0004) but not the probability that the two answers agree (p = 0.4625). The Lüders projection model predicts this without parameters, by a one-line identity. A classical model with order events can be built to satisfy the equality, and LIT-264 shows that data meeting it are noncontextual, so it is evidence for a constraint on order effects, not for nonclassical probability or contextuality. A preregistered test with a constrained classical rival is QUESTION-210. | Active |
| THEORY-172 v4 | In Frank and Goodman’s one-shot reference games, listeners’ mean bets matched a parameter-free listener inverting a speaker who chooses among true utterances by informativeness, with a separately measured salience prior, and not salience alone; and in Bergen, Goodman and Levy’s signalling games with novel symbols, listeners read a signal as excluding what the available alternative would have conveyed | Frank & Goodman (2012), LIT-821, read in NOTE-634: a parameter-free Bayesian listener inverting a size-principle speaker, with a separately measured salience prior, matches mean listener bets over seven context types at r = .99 (.87 without the 0 and 100 predictions), while salience alone correlates at r = .19. Bergen, Goodman & Levy (2012), LIT-809, read in NOTE-613: with novel signals and no conventions, listeners read the signal as the object its alternative did not name, and lexical uncertainty derives Horn’s mapping. Goodman & Frank (2016), LIT-816, states the framework. Shown for artificial games with given alternatives, by fits of condition means with no rival fitted; a test against named rivals on trial-level data is QUESTION-209. | Active |
| THEORY-208 | Agency is multiply realisable without cognition or mental states, and AI is a new kind of it | Floridi (2025), LIT-897, read in NOTE-702, with the criteria taken from Floridi & Sanders (2004), LIT-898. In the record’s reading, agency comes in kinds. All of them share interactivity, and the richer ones add graded autonomy and adaptability. The kinds are told apart by origin, mechanism and where their goals come from, not by intelligence. AI is computational agency with data-driven learning, serving goals set by humans, with no intelligence or mental states. Proposed: the paper assumes that AI is an agent, does not argue for this thesis over its rival, asserts that AI lacks intelligence, and never states the level of abstraction its comparison uses. | Proposed |
| THEORY-220 v3 | A far-from-equilibrium system that contributes to its own maintenance has functions relative to itself, with no history or observer needed, and one that switches among self-maintaining processes has indications of success that can be false by its own lights | Bickhard (2000), LIT-1029, read in NOTE-729. Function is contribution to a far-from-equilibrium system’s own persistence; the asymmetry between persisting and decaying gives function and dysfunction, always relative to the system served, so a parasite’s gut serves the parasite and harms the host. A candle flame has functions in this sense. Representation arrives with recursive self-maintenance: a system that picks its next interaction by how it has differentiated its environment implicitly predicates that the interaction will succeed, and can find the predication false. Proposed: a conceptual account, with no model, whose account of having a function (and so of malfunction) is deferred to another paper. | Proposed |
| THEORY-245 v3 | In a preregistered multimodal test with clearly seen images, prefrontal cortex carried a seen image’s category but not its identity, and a face’s orientation at most weakly, while posterior cortex carried category, face orientation and the identity of objects, letters and false fonts, though not of faces; no prefrontal site showed the preregistered onset-and-offset response, and one, found in an exploratory analysis, responded at offset earlier than predicted | Cogitate Consortium, Ferrante et al. (2025), LIT-973, read in NOTE-812. In iEEG and fMRI, with Bayesian support for the nulls, and against a positive control, prefrontal cortex held category but no identity or orientation, and none of 655 prefrontal electrodes showed the onset-and-offset response GNWT predicted. It does not show these dimensions were unaccessed, or that prefrontal cortex lacks them in other signals: high gamma and BOLD only, no single units, and every image was seen. | Active |
| THEORY-255 v3 | In Aerts’s subspace model of Hampton’s membership data, and in any lattice whose membership weight is order-preserving, whatever its negation, taking concept conjunction and disjunction as meet and join cannot produce overextension or underextension, so a model of the guppy-type deviations must take conjunction as some operation other than the meet or drop the monotone weight, while the Kolmogorov-type deviations can be had from the meet of a non-Boolean subspace lattice | Aerts (2009), LIT-934, Theorem 9 and §§2.2–3.3, read in full in NOTE-808. The impossibility is a proof, and it needs only that membership is monotone in the subspace. The fits that place the k-type deviations in a tilted subspace lattice and the ∆-type ones in an emergent sector are models with free parameters, not tests. Negation appears only as the orthocomplement used to derive the conjunction model from the disjunction model; no negated concept is measured. | Active |
| THEORY-261 v2 | In sparse distributed memory, recall of uniform random data from exact cues is a signal-to-noise problem whose optimum activates about (2MT)^-1/3 of the hard locations, and capacity then grows in proportion to the number of hard locations, rising toward about a tenth of them at 0.999 bit fidelity, not with the word length | Kanerva (1992), LIT-921, §§3.5–3.7, read in full in NOTE-779, after Jaeckel (1989a). An approximate derivation for uniform random data and exact retrieval cues. It does not cover noisy cues, which lower capacity, or clustered data, which the uniform design handles badly, and it does not contain the critical-distance analysis that is sometimes attributed to it. | Active |
| THEORY-272 v3 | In Hopfield’s 1982 simulations at N = 30 and 100, a network of threshold units with symmetric Hebbian couplings acts as a content-addressable memory whose stored random patterns are attractors: recall degrades severely near 0.15N patterns, a pattern placed 10 bits from another at N = 100 often fuses with it, and at 25-fold overload no stored pattern stays stable | Hopfield (1982), LIT-918, read in full in NOTE-819. The energy descent is derived for symmetric couplings; everything quantitative, including the 0.15N figure, the basin width and the merging distance, is simulation at N = 30 and 100. It does not say that spurious states come only from overload: they appear at a low load too. Nor does it say anything about addressed memory. | Active |
| THEORY-330 | After a forced phase in which each person experienced every memory load from three to nine items, the load that had given that person the best reward rate predicted their later choices better than one item fewer or one more, in a copying task where items are seen once and cannot be looked up again; a steeper reward for long lists raised chosen load, a higher price for errors did not, and about one person in five kept choosing a non-optimal load | Howes, Duggan, Kalidindi, Tseng and Lewis (2016), LIT-1062, read in NOTE-882. Participants copied names shown one at a time from a mock email to a calendar. Each person’s best list length in their own data predicted 55% and 46% of their choices, against 8–17% for the neighbouring lengths (N = 40 and 20). An exponential points schedule raised chosen length (6.99 against 5.07); forfeiting the whole trial on one error did not (5.00 against 5.05). 13 of 60 persistently chose a non-optimal length. There is no look-up and no model: the prediction is each person’s empirical argmax. Lieder and Griffiths’ description of it as a memorize-or-look-up trade (LIT-1071) does not fit the task. | Proposed |
| THEORY-334 v2 | In copying a pattern of coloured blocks that stays in view, seven subjects told to work quickly looked back at the model about 1.3 times per block, and their commonest cycle, which looks at the model both before picking up a block and before dropping it, was the slowest per block in the authors’ observed tallies (about 3 s, against 1.5 s for a cycle that copies from memory); looks fell to 1.0 per block with model and copy 70° apart and to 0.7 with one-colour patterns that could be chunked, and in three subjects a colour change made during a saccade raised fixation time by about 129 ms after pickup against about 50 ms before; no statistics are reported | Ballard, Hayhoe, Pook and Rao (1997), LIT-1075, read in NOTE-883. Fixations are modelled as pointers that bind a program’s variable to a place in the world, so that only the task-relevant feature is fetched, just before use. In block copying, look-model, pick-up, look-model, drop was the commonest cycle (about 43% of cycles), at about 3 s a block against 1.5 s from memory; 70° separation cut looks from 1.3 to 1.0 per block, and one-colour patterns to 0.7. A colour changed mid-saccade mostly went unnoticed and cost time only once colour was held. The “carrying cost” of working memory is inferred from choice, not measured, and what is left in the world is raw input, not a computed result. Promoted 2026-10-20 on the observations alone (ADR-047 v3). | Active |
| THEORY-335 v2 | Selective accessibility, the account that comparing a target with a plausible anchor makes anchor-consistent knowledge more accessible and that this knowledge carries anchoring, fits the moderations of the effect by applicability, representativeness and plausibility that Strack and Mussweiler found, but its accessibility step is inferred, not measured: Harris et al.’s repetition of the recognition test of it, four times by lexical decision and once by continuous identification, found nothing in five preregistered experiments, with tasks that also missed direct picture priming | Strack & Mussweiler (1997), LIT-1064, read in NOTE-877, proposed the account and showed, in three small studies, that with the anchor number fixed the effect depends on whether the knowledge used to compare applies to the later estimate, and that plausible anchors slow the comparison and speed the estimate. Accessibility is inferred, never measured. Harris et al. (2019), LIT-1067, read in NOTE-876, repeated the recognition test (lexical decision, once continuous identification) of the accessibility step (Mussweiler & Strack 2000a, not held) five times at N = 128: anchoring was large and recognition of anchor-consistent words never moved (BF01 ≈ 21), but the same tasks missed direct picture priming. So the direct evidence is removed, not reversed; the account is neither refuted nor shown. Mussweiler & Strack 1999 (LIT-1078) is Unavailable. | Active |
| THEORY-336 v2 | An ideal observer is optimal relative to the information it is given, so a constrained ideal is the optimal decoder of a degraded input, not a restricted decoder of the full one; comparing real observers with ideals taken at successive stages of the visual system places losses by stage, and in early-vision tasks human performance is reported often to parallel the retinal-stage ideal across conditions while falling well below it | Geisler (2011), LIT-1055, read in NOTE-870. A review that defines the ideal observer as optimal “given the available information and any specified constraints”, and writes it as an expected-utility argmax over the posterior given the stimulus S or a constrained representation Z = g(S;θ). Constraints are added one at a time (optics, receptor sampling and noise, ganglion-cell summation), and human performance that parallels the retinal-stage ideal while falling below it is read as loss in central factors. The empirical parts are the review’s account of studies the record has not read. The review gives no formula for efficiency and no notion of an optimum over a restricted class of computations. | Active |
| THEORY-337 | In the comparative-question paradigm, judging whether a quantity is higher or lower than a high or low anchor pulls the later estimate of it toward the anchor, with large effects in five preregistered experiments of 128 people each, lasting through dozens of intervening word-recognition trials | Harris et al. (2019), LIT-1067, read in NOTE-876, and Strack & Mussweiler (1997), LIT-1064, read in NOTE-877. In five preregistered experiments (N = 128 each) on temperatures and car prices, estimates after a high anchor exceeded those after a low one every time (d ≈ 0.85–1.39), with 40 to 70 word-recognition trials between the comparison and the estimate. Strack and Mussweiler’s three small studies show the same effect, and that it can shrink or reverse when the knowledge used to compare does not apply to the estimate. This is the effect, not its mechanism: why the estimate moves is THEORY-335’s question and is not settled. | Active |
| THEORY-339 | People’s answers to a probability query depend on their own answers to an earlier query whose result it can use: a person’s sub-answer predicts their later answer only when asked first, and a bias induced in a first judgment carries into a second only when the two share part of a hypothesis space, so reuse carries the earlier error with it; that reuse saves computation rests on one small response-time study and one secondary-task result | Dasgupta, Schulz, Goodman and Gershman (2018), LIT-1074, read in NOTE-875; Gershman and Goodman (2014), LIT-1056, read in NOTE-886; Dasgupta and Gershman (2021), LIT-1077, read in NOTE-865. In five small online experiments, an earlier answer shapes a later one only when the later query can use it, and errors in the first carry into the second. The authors call this amortized inference, reuse of past computation bought at a cost in bias. The speed benefit is shown once (median log-RT facilitation, n ≈ 50 per order) and otherwise assumed; storage and retrieval are free in the models; the 2021 review adds stale cached values from the planning literature, second-hand. No paper prices when reuse pays or touches stores outside the head. | Proposed |
| THEORY-343 | In laboratory choice tasks, people offload cognition more as a task’s internal demand rises, but the choice follows their evaluation of the demand rather than its benefit: they offload where it gives no gain and more when less confident at equal ability, and offloading can lower memory for what was offloaded | Risko & Gilbert (2016), LIT-1059, read in NOTE-887. A review: in choice/no-choice studies of rotated displays, delayed intentions and short-term memory, spontaneous offloading rises with internal demand, but it also occurs where it gives no gain (writing items down about 40% of the time at a two-item load already at ceiling), rises with lower confidence after ability is controlled, and differs between displays of equal measured demand. Expecting to have a result stored externally went with poorer memory for it. The review reports no effect sizes, the cited studies were not read, and its metacognitive model is an unformalised box diagram; the drift toward external reliance is a prediction, not a finding. | Proposed |
| THEORY-346 | When information stays in view at a known place, people consult it less and rely on memory more as the time to reach it rises, in a graded trend across added delays from 0 to 3.2 seconds, even where the memory is weak and the display perfect; in one copying task a model that prices memory and display access in time alone, with no preference for either source, fits how many items people encode per look | Gray, Sims, Fu and Schoelles (2006), LIT-1073, read in NOTE-868, and Gray and Fu (2004), LIT-1058, read in NOTE-864. Adding a click to reach on-screen show information cut how often it was checked, and checks of finished settings came from memory. Lockouts of 0 to 3,200 ms on a Blocks World pattern cut looks and raised blocks placed per look, with linear trends. An ideal performer model with ACT-R memory equations and Q-learning on time alone fits four measures. That the control system has no bias toward the world is argued, not tested against a rival model. Errors in the 2004 study differed only against a group that memorized under several other changes, and the 2006 model has no wrong retrievals, no precomputation, no reuse across trials and no staleness. | Active |
| THEORY-351 v2 | In a two-pilot airliner cockpit, setting the approach speeds as bugs on the airspeed indicators, prepared 25 to 30 minutes ahead in a low-workload phase from a card selected by the displayed gross weight, turns recall and scale reading at the moment of use into a judgement of spatial proximity, so the cockpit system remembers the speeds without either pilot’s memory being improved; this is argued by task analysis of mandated procedure, not measured | Hutchins (1995), LIT-1066, read in NOTE-867. A stylized task analysis of one airline’s MD-80 procedure, informed by observation but reporting no measurements. The speed-card booklet, filtered by the displayed gross weight, is the system’s long-term memory; the values pass through speech to bugs on two airspeed indicators, and on approach the pilots judge the needle’s nearness to a bug instead of recalling and reading speeds. The paper concludes that the memory is “not made primarily of pilot memory”. It analyses only success, for an anticipated, mandated and invariant task with a few values needed in a fixed order, so it says nothing about search for one item among many under a deadline. Cognition in the Wild (LIT-1060) is Unavailable. | Proposed |
| THEORY-355 | Agents arrange their surroundings so that perception and choice take the place of internal search, changing how much thinking a task needs without changing what can physically be done; the set-up is paid once and recovered over the later choice points the arrangement serves, so an arrangement whose scope is the whole task can repay a large set-up and a narrow one must repay within the next few actions, and the arrangement decays unless maintained; the savings are argued and piloted, not measured | Kirsh (1995), LIT-1076, read in NOTE-871, and Kirsh (1996), LIT-1057, read in NOTE-872. From video of cooks, assemblers and baggers and from informal observation, Kirsh classifies arrangements that simplify choice, perception and internal computation, frames their savings in search-complexity terms, and ties set-up cost to the number of choice points an arrangement serves. The 1996 paper adds a verbal investment rule with an interest factor for risk, one risk being that the change is not still present later, and upkeep against scatter. Neither paper gives a cost model or a measured saving; the 1995 paper says the savings do not “in general” admit meaningful quantitative estimation, and the 1996 paper puts mental cost outside its formalism. | Proposed |
| THEORY-358 | In visual search among up to about 36 well-spaced items, a target that alone carries one colour or one distinctive shape part is found in time almost independent of the number of items, while a target identified only by a combination of features its neighbours carry separately takes time that grows linearly with the number of items, about twice as fast when the target is absent, the pattern of a serial self-terminating scan, and practice did not remove it; that conjunctions rather than similarity cause it is the paper’s reading, not its result | Treisman and Gelade (1980), LIT-1082, read in NOTE-897. With coloured letters on cards of 1 to 30 or 36 items, a blue letter or an S among brown Ts and green Xs was found at 3.8 and 2.5 ms per item, and a green T among the same distractors at 28.7 ms per item when present and 67.1 when absent, linear and in the ratio of a scan that stops at the first find; 13 sessions of practice did not change it. The paper reads the contrast as feature integration: conjunction structure, not similarity, makes search serial. That reading is contested by the similarity account of Duncan and Humphreys (LIT-1081), which the record could not obtain, and it does not follow that any item unique in one attribute pops out in a real pile. | Active |
| THEORY-359 | Cognitive mechanisms can be derived as the one, among those a brain can execute, that best trades the expected utility of its result against the opportunity cost of the time and resources it uses; the review shows the derivation reproducing many heuristics and biases, not that people are optimal, and the objective is per episode | Lieder & Griffiths (2020; online 2019), LIT-1071, read in NOTE-873. Resource rationality picks the heuristic that maximizes expected utility of its result minus the expected opportunity cost of the time and resources it uses (Eq. 3), averaged over environments consistent with what is known (Eq. 4), and a five-step method derives process models from it. The support is a review of others’ models recast in this frame, with numbers from those studies. The authors offer optimality as a methodological device, not a claim about people, and the method lets constraints be refitted to data. The objective is per episode, with no terms for storing, retrieving or reusing a result, or for its going stale. | Proposed |
| THEORY-360 v2 | In Kirsh and Maglio’s Tetris keystroke data, rotation began in earnest by 400 to 600 ms after a piece appeared, sometimes before it was fully in view, and at each of three skill levels drops preceded by a move to the wall and back were longer than ordinary drops, about 19 to 20 squares against 13 to 16 | Kirsh and Maglio (1994), LIT-1054, read in NOTE-878, with Kirsh (1995), LIT-1076, read in NOTE-871. In Tetris keystroke data, pieces are rotated more than plan-then-act would need, rotation starts by 400–600 ms, and a translate-to-wall check precedes longer drops. The authors define an epistemic action as one whose primary function is to cut the space, time or unreliability of the agent’s own computation, and argue these moves are of that kind. The saving is argued from a three-player pilot comparing physical with mental rotation, not measured in performance. The trade is offloading at the moment of use, once, with nothing stored; the paper sets itself against caching. Promoted 2026-10-20 on the timing and Table 1 alone (ADR-047 v3); the epistemic-action reading is held as THEORY-823. | Active |
| THEORY-402 | Collective intelligence is resource-bounded computation with a larger but modular budget: a group has more sensing, state, processing and action than any member, and pays coordination and cooperation costs because those resources sit in autonomous members, so the same trade-off yields groups that do better than their members on some tasks, groups that run qualitatively different algorithms, and groups that do worse | Pilgrim et al. (2025), LIT-046, read in NOTE-055, with Dreyer et al. (2024), LIT-075 (NOTE-040), as its clearest instance. The paper offers a four-dimensional resource inventory, a catalogue of constraints and a mapping to mechanisms, and calls the known forms of collective intelligence “aspects of a single unifying principle”. NOTE-055 judges it a vocabulary, not a derivation: the resources have no metric, so the framework cannot say in advance when costs outweigh gains, and its “predictions” are research questions. Filed Proposed as the account the vocabulary states. | Proposed |
| THEORY-433 | Whether a machine thinks is best replaced by whether a digital computer can play the imitation game as well as a man, because the reasons for withholding thought from a machine that does would equally withhold it from other people, whose thinking we also infer from what they say; the objections from consciousness, mathematical limits and informality of behaviour fail on that test or rest on unproved premises | Turing (1950), LIT-394, read in NOTE-355. The imitation game is offered in place of a definition of thinking, not as a test of it, and Turing says he has “no very convincing arguments of a positive nature”. The account’s force is the other-minds parity of §6.4, which assumes, as §2 states and does not argue, that the machine’s best strategy is to answer as a man would; a machine built to imitate human output is the case that assumption does not cover. It makes no claim about consciousness. Proposed, and a rival of Block’s psychologism. | Proposed |
| THEORY-436 | Pain is a functional state of a whole organism, possessing a certain functional organisation as a probabilistic automaton, rather than a brain state, and this is the more plausible empirical hypothesis because one pain predicate applies to physically different organisms; the hypothesis keeps swarms out only by a stipulated condition that no pain-feeler has parts that separately possess the organisation | Putnam (1967), LIT-396, recorded as Skimmed in NOTE-335 from a course excerpt of the 1975 reprint, which carries the four-clause hypothesis verbatim; the original is Unavailable here. Proposed, and resting on that excerpt and on a result discussed throughout the literature the record holds. Clause (1) is empty by Putnam’s own admission, since everything is a probabilistic automaton under some description, and clause (3), the anti-nesting condition, is introduced with one motive, to rule out swarms of bees. The argument against the brain-state view is comparative plausibility, not evidence. | Proposed |
| THEORY-438 | The understanding deep networks reach is characteristically fractured: it can be real, by the test of stable tracking under structure-preserving perturbation, but it is symbolically misaligned, non-reductive, locally fractured, proxy-dependent and representationally fragmented, because the loss rewards any predictive patchwork and does not press it to unify | Freeborn (2026), LIT-131, read in NOTE-123. The Fractured Understanding Hypothesis is argued from four small demonstrations and cited literature, with no test of how common the five features are. The abstract names three features and §7.5 five. The paper also applies its own criteria inconsistently: its Othello verdict uses post hoc probes that it excludes from the agent. The definition of understanding it rests on is an explication, not a finding. | Proposed |
| THEORY-454 v2 | Perception is constitutively iconic, nonconceptual and nonpropositional and cognition paradigmatically discursive, and percepts are put to cognitive use only inside a ‘cognitive envelope’, the global workspace read as conceptualisation, where working memory holds recoded, task-dependent abstractions rather than the percepts themselves | Block (2024), LIT-649, read in NOTE-498: the précis of his 2023 book, not the book. The main thesis is stated and its arguments are in the book, unread here. The working-memory half is the update: nine cited lines of evidence that working-memory representations differ from perceptual ones and are task-dependent abstractions (moderate). The rereading of the workspace as conceptualisation rather than consciousness is asserted, not argued against the consciousness reading. | Proposed |
| THEORY-458 v2 | Within the minimal self, the sense that a movement or a thought is mine and the sense that I am its source are distinct and come apart, as in involuntary movement and thought insertion; on Frith’s comparator model the sense of agency depends on a forward model that precedes action and ownership on sensory feedback, so failed forward monitoring would explain thought insertion and, by the same mechanism, why first-person reference ordinarily cannot misidentify its subject | Gallagher (2000), LIT-453, read in NOTE-377 from the preprint. A review with no new results: the dissociation and the comparator account are reported from Frith, Haggard and others, and the link to the immunity principle is Campbell’s argument, adopted and left conditional. It is about the minimal self, not the narrative self, for which the review’s own proposal is two paragraphs. It does not say which minimal self, embodied or reflective, the analysis concerns. | Proposed |
| THEORY-460 | Every leading reason to deny consciousness to current large language models except biology names a missing feature, recurrent processing, a global workspace, unified agency, world and self models or senses and embodiment, that engineering could add, so the case against current models is much stronger than the case against their extended successors; the one possibly permanent objection, biology, is the one the argument sets aside | Chalmers (2023), LIT-411, read in full in NOTE-345. A classification of objections as temporary or permanent, argued by naming simple systems that have each feature. The credences that follow, under 10% for current models and 25% or more for conscious extended models within a decade, are illustrative by the author’s warning, multiply stipulated factors under a disclaimed independence assumption, and depend on “mainstream assumptions”. Biology is set aside with a reference to earlier work. It adds no evidence about any model. | Proposed |
| THEORY-461 | If a substantial class of mental states, interoceptive, affective, perceptual and bodily-agentive, depends on an animate body, then today’s text-only chatbots lack them and their self-ascriptions of such states are systematically untrue, and users can engage them without false belief by make-believe, as a new kind of interactive, self-involving and blended fictional character | Roberts (2026), LIT-207, read in NOTE-157. The anti-realist half is a valid conditional whose antecedent, moderate embodied cognition, is assumed at prima facie strength. It covers body-dependent states only; footnote 8 allows realism about the chatbot as a disembodied thinker, so belief and reasoning are untouched. The fictionalist half shows the attitude permissible and apt, not required, and offers no evidence that users hold it. It is not a general denial that chatbots have minds. | Proposed |
| THEORY-462 v2 | If commonsense knowledge includes defaults that later evidence displaces, a monotonic logic, in which an added axiom can never remove a theorem, cannot represent that displacement as it stands; Minsky’s further claim, that no logistic system, one that separates propositions from general laws of inference, can carry commonsense knowledge even once repaired, is asserted, not shown | Minsky (1974), the frames memo, LIT-410, read in NOTE-337. Its appendix argues that “logistic” systems fail for commonsense on five counts, of which monotonicity is a correct observation about classical logic. That no repair is possible is asserted (“I think such attempts will continue to fail”), and consistency being undesirable is argued by analogy. The reading says the memo is a source of hypotheses, not evidence; this account takes from it only the half that holds by definition and marks the rest as the memo’s claim. Active as the definitional half and a report of how the memo supports the rest; the further claim is THEORY-1239, Proposed. | Active |
| THEORY-471 | Psychological states are not the machine-table states of an organism’s unique best description as a probabilistic automaton, because some are never occurrent, several hold at once, there are indefinitely many and they contain one another, none of which a machine-table state can be; this leaves open that they are computational states of such a system | Block & Fodor (1972), LIT-641, read in NOTE-491. Five independent arguments against the machine-state identity theory, each with a stated escape that keeps functionalism: dispositions with no occurrent state, simultaneous states, too-fine individuation, productivity and constituent structure. They refute the machine-table clause, not functionalism. The paper’s remaining argument (its third), inverted or absent qualia, is the only one aimed at functionalism broadly, and it is explicitly prima facie. | Active |
| THEORY-472 v3 | A speaker who chooses words to minimise the information lost about a single intended referent, with words’ literal meanings uniform over their extensions, picks each true word with probability inversely proportional to a power of its extension, so the size principle follows from informative communication; in one betting experiment, learners who knew the referent bet on a novel word’s meaning in line with the inversion of that speaker (r = .93 on condition means), and in one free-description study speakers’ word use followed it only weakly (r = .19) | Frank, Goodman, Lai and Tenenbaum (2009), LIT-824, read in full in NOTE-627. The derivation (utility −D_KL from intended to literal meaning gives P(w) ∝ |w|^(−α) for a point referent) is a proof under its assumptions. Learners’ bets track the parameter-free inversion at r = .93 on condition means (700 web participants); speakers’ tag-word use in free descriptions correlates at r = .19, .51 without colour terms and .02 for them. In these displays the informative-speaker learner and a plain size-principle learner coincide by construction, so the fit does not show that learners reason about the speaker. | Active |
| THEORY-475 | In simulations from the drift-diffusion model itself, with 12 subjects and 20 to 150 trials each, hierarchical Bayesian estimation recovers every DDM parameter with less error than per-subject maximum likelihood, χ²-quantile fitting or non-hierarchical Bayes, most of all at few trials; whether the gain survives misspecification, and how much of the power gain is the hierarchy’s, is not shown | Wiecki, Sofer & Frank (2013), LIT-563, read in NOTE-466. Experiment 1 gives the lowest trimmed recovery error for HB on every parameter, and HB beat ML on the same data sets in every condition. Experiments 2–3 report more power to detect drift differences and trial-wise covariate effects, but HB was judged by a posterior-interval test and the others by t-tests, and the clean comparison was omitted. All data were generated from DDMs, with sz = st = 0. It is the standard partial-pooling result shown for one model family, not a test of the DDM. | Active |
| THEORY-489 | Chains of human function learners, each trained on the previous learner’s responses, forget their starting function and in 28 of 32 families converge within a few generations to a positive linear function, the bias earlier studies report, whatever function trained the first learner; this is convergence to the learners’ favoured function, not shown to be convergence to their prior | Kalish, Griffiths & Lewandowsky (2007), LIT-820, read in NOTE-616. 288 participants in 32 nine-generation chains, four starting conditions. 28 families converged to a positive linear function; 3 converged to the negative linear one and 1 had not converged. The bias was taken from earlier studies, not measured on these learners, and stationarity was not established. It supports the claim that human transmission chains settle on what learners favour in this task, not that they reach the Bayesian prior, and it cannot tell sampling learners from MAP learners. | Active |
| THEORY-500 | Ventromedial prefrontal cortex stores links between kinds of situation and the body states they produced, and re-enacting those states when an option is imagined marks it good or bad, overtly as a feeling or covertly as a bias, pruning the options before reasoning weighs them; without the marks, choice under uncertainty fails despite intact knowledge and intellect | Damasio (1996), LIT-384, read in NOTE-332, the somatic marker hypothesis stated by its author; Bechara, Damasio, Tranel and Damasio (1997), LIT-383, read in NOTE-331, the experiment it is usually cited for. The robust result is a group difference: in the gambling task all 10 controls and none of 6 ventromedial patients developed anticipatory skin conductance responses. That controls chose well before they knew, and that the marker causes the choice, is not carried by the statistics. The mechanism is argued, not observed. | Proposed |
| THEORY-512 | The extended-mind thesis worth disputing is that external entities partly constitute a subject’s mental states in virtue of the subject’s sensorimotor interaction with them; so stated it can reach standing belief but not consciousness, if consciousness needs information directly available for global control, since perception makes availability indirect | Chalmers (2019), LIT-415, read in NOTE-356, and his foreword to Clark’s book (2008), LIT-421, read in NOTE-369. The 1998 thesis is satisfied by wired-in “extended circuits” that its critics accept, so the controversial thesis is constitution through perception and action. The perception-action boundary is then the strongest objection, met by revisionism, not refuted. Consciousness does not extend this way, on an imported criterion (direct availability, “not mediated by other mental states”) and on the absence of any working twin case. Both halves are proposals defended against a few cases; the criterion is the author’s earlier correlate of consciousness, not an argument made here. | Proposed |
| THEORY-520 | When coordination between two autonomous agents and the continuation of their encounter shape each other, the interaction gains an autonomy of its own, a precarious self-sustaining organization that lasts only while both agents stay autonomous; coordinating movement then coordinates sense-making, so some social understanding, such as recognising a partner, is a property of the interaction’s dynamics and not of either individual | De Jaegher and Di Paolo (2007), LIT-571, read in NOTE-456. A conceptual account built on coordination dynamics, with illustrative evidence: the corridor, charades, mother–infant timing and the perceptual-crossing experiments of Auvray et al., reported, not new. NOTE-456 declined a THEORY because the source carries no claim beyond its definition. The claim with some support is the perceptual-crossing one; the autonomy of the interaction is a definition, and the implications are called “speculative at this stage” by the authors. | Proposed |
| THEORY-528 | The cases offered for extended cognition are explained at least as well by embedded cognition, on which the organism leans heavily on external props, and memory research finds that external stores lack the interference, generation effect and power-law learning and forgetting of internal memory, so a kind spanning both does no explanatory work and conservatism favours the embedded account | Rupert (2004), LIT-422, read in NOTE-343. The past-endorsement dilemma is strong: without it, directory assistance makes Otto believe every listed number; with it, the organism’s endorsement is the final authority. The memory comparisons are moderate: the internal side is cited research, the external side is thought experiment, and no experiment on an external store is run or cited. The step from “different fine-grained kinds” to “no useful coarse kind” is a challenge plus conservatism. | Proposed |
| THEORY-529 | A state is shown to be a representation by adaptation for transmitting information efficiently, near the optimal trade-off of rate, distortion and coder complexity under a teleofunctionally fixed distortion measure, rather than by correlation or mutual information alone, which nearly any causal link supplies and which cannot increase along a processing chain | Martínez (2025), LIT-149, read in NOTE-165: a position paper with no new theorem or data. Its sharpest point is a standard result: by the data-processing inequality, mutual information about a car/no-car decision can only fall from retina to IT, so a mutual-information criterion would place the decision at the earliest stage; explicitness is better read as ease of decoding. The rest is a proposal supported by selected examples. Complexity has no operational measure, teleofunctions are needed to fix the distortion measure or every system is efficient for some loss, and misrepresentation and engineered systems are not addressed. | Proposed |
| THEORY-538 | A system believes and desires when its behaviour is reliably and voluminously predictable by treating it as a rational agent with the beliefs and desires it ought to have, and the success of that strategy is an objective fact though adopting it is free; the more richly a system is connected to its world, the fewer interpretations fit, so thermostat and person differ in degree | Dennett, “True Believers”, LIT-398, read in NOTE-353 in the revised text of 1987/1997. The criterion is substrate-neutral and silent on experience. Its adverbs are not quantified, the lectern is excluded only by gain over a baseline the essay leaves informal, and the Martian argument is an intuition pump that assumes the intentional patterns are not captured physically. “Most beliefs are true” rests on a way of counting beliefs Dennett doubts exists. Proposed, and a rival of both Block’s psychologism and Searle’s causal-powers account. | Proposed |
| THEORY-541 | A model of discovery or analogy that is given representations built with the answer in view has been given most of the answer, as BACON was with Kepler’s data and SME with its solar system and atom, and representation-building cannot be postponed to a separate module, because which representation is right depends on the task and on the analogy being made | Chalmers, French & Hofstadter (1992), LIT-414, read in NOTE-359 from the preprint, with Burns’s review of Fluid Concepts (1995), LIT-457, read in NOTE-375. The BACON critique is strong; the SME critique moderate, with no alternative encoding run; the two arguments against a representation module informal. Burns turns the charges on Copycat: its Slipnet holds the domain’s relevant relations, and the symbol-grounding objection to SME applies to it equally. Nothing is measured in either text. | Proposed |
| THEORY-542 | A word learner that infers speakers’ referential intentions jointly with the lexicon, and lets words be uttered non-referentially, learns object names from child-directed video more precisely than co-occurrence and translation models, and mutual exclusivity, word-guided object individuation and intention-guided mapping follow from it without principles built in for them | Frank, Goodman & Tenenbaum (2009), LIT-815, read in NOTE-611. On about twenty minutes of video from two dyads, annotated by the authors, the intentional model’s lexicon scores precision .67 and F .55 against .06–.15 and .10–.22 for five baselines. The behavioural fits are qualitative, the baselines also produce mutual exclusivity, and the Baldwin case supplies the intention rather than inferring it. The reading said no THEORY, because the findings are specific to a small corpus and the record held nothing they correct; this account is filed Proposed for that reason. | Proposed |
| THEORY-545 | Aerts and Gabora’s founding quantum model of concepts is a classical probability model written in a Hilbert space: every context it uses is a projector diagonal in one basis, so the contexts commute and generate a Boolean algebra, its entangled ‘pet fish’ state gives the same statistics as a classically correlated mixture for every context in the model, and its guppy weights are copied from single-concept ratings rather than predicted for the conjunction | Aerts & Gabora (2005), Parts II and I, LIT-312 and LIT-340, read in NOTE-299 and NOTE-294. The structural facts are checks of the published model, not of the authors’ prose: contexts are sums of basis projectors, counts are chosen to match rounded frequencies, and the authors call the result a “quantization” of a set-theoretical model. Part I’s mark of non-classicality (eq. 19) is met by Boolean projection lattices too, and its lattice structure is supposed, not derived. The model as printed is also infeasible in two cells. It does not say that no quantum model of concepts does better. | Active |
| THEORY-547 v2 | A language model idealized as continuing any text with a plausible next word from the collection of all plausible texts follows narrative coherence and not truth, so its false statements are confabulations within the story the context sets up; whether trained and aligned chat models keep that nature is an open empirical question the idealization does not settle | Bottou & Schölkopf (2023), LIT-178, read in NOTE-102: an essay offering imagery, with no evidence beyond one Vicuna-13b transcript pair. For the “perfect language model” it defines, insensitivity to truth follows from the definition. That deployed chat models share it is assumed, and the essay’s own §2 leaves open whether alignment can turn a fiction machine into a verification machine. It credits the machine with knowledge of what makes sense in the story, denies it truth-directedness and intention, and never uses the word “sycophancy”. | Active |
| THEORY-558 | Moral judgment is usually caused by quick affective intuitions, and moral reasoning is usually a post-hoc search for supporting reasons that changes judgments mainly by triggering intuitions in other people | Haidt (2001), LIT-511, read in NOTE-403: the social intuitionist model. Its case rests mostly on non-moral evidence about automaticity, motivated reasoning and confabulation, carried over by argument. Its direct moral evidence is that harmless taboo violations are condemned and that affect predicts the condemnation better than perceived harm. The author calls it “a plausible alternative”, not “an established fact”. The disgust manipulations that were its direct causal evidence were later disputed. | Proposed |
| THEORY-563 v2 | Consciousness is a real pattern in the brain’s content-bearing states, a description that trades predictive accuracy for compression, and so coarse that describing the brain through it is metaphor; on that reading there is no Cartesian theatre, whether a given content was conscious can lack an answer when two descriptions compress and predict equally, and first-person reports are fallible as literal descriptions of the brain but authoritative as an author’s | Taylor (2024), LIT-461, read in full in NOTE-379. An interpretation of Dennett offered as consistent, explicitly not as correct or as what Dennett meant. Its argument for indeterminacy in masking rests on Dennett’s idealised description of the cases, which Taylor calls “somewhat simplistic”. It does not say which real pattern consciousness is, or what makes a real pattern a conscious one rather than a belief pattern. Filed as the account the reading states, at the strength the reading gives it. | Proposed |
| THEORY-564 v2 | A bounded generative model can learn its own structure with no machinery beyond inference and learning: spare hidden-state slots with flat likelihoods get filled when an observation fits none of the learned causes, and Bayesian model reduction on the learned prior over states can return unneeded slots, though in a toy concept-learning task it recovers the true number of causes in only 45–80% of runs with four to seven causes, and never with two or three, unless the likelihood is given | Smith, Schwartenbeck, Parr and Friston (2020), LIT-615, read in NOTE-485. Expansion works in the toy domain over eight runs per condition, and only for novel stimuli. Reduction is the weak half: over 100 runs, the true number of concepts won in 45–80% of runs for four to seven, never for two or three, and in every run when the likelihood was given. Table 1 is the record’s only measurement of how often reduction picks the generating structure when the parent model was itself learned. One initialisation; eight concepts, six binary features. | Proposed |
| THEORY-570 | An emotional state is a large-scale selection of mental resources, switched on and off by Selectors that Critics trigger on recognising a type of problem, and so of the same kind as an intellectual way to think; feelings are the mind’s descriptions of these cascades, and the body is one more resource the brain uses, not the substance of emotion | Minsky (2006), The Emotion Machine, LIT-432, read in NOTE-348 from the author’s 2005 draft. A design argued from examples, thought experiments and the author’s own AI programs, with no experiment; Minsky marks several parts as his untested conjecture. It disagrees directly with the body-state views of emotion and of feeling. On Elliot it proposes the reverse causal order to Damasio’s, with no evidence offered. | Proposed |
| THEORY-573 | Signals of processing quality from different senses, such as fluency, familiarity and coherence, pool into one content-nonspecific feeling, and that pooled feeling contributes to the unity of conscious experience across modalities, alongside the integration of specific features; in support, coherence in one modality shifts judgments of regularity, familiarity and possibility about unrelated targets in another | Winkielman, Ziembowicz & Nowak (2015), LIT-190, read in full in NOTE-146. An opinion piece whose main evidence is the authors’ own three experiments (Ziembowicz et al. 2013), summarised without statistics and not read by the record. The authors state that direct evidence is still needed that the effects run through changed experience and that its integration is causal. No study cited measures unity of experience. The LIT is itself Proposed; this account carries its promotion condition. | Proposed |
| THEORY-577 | Our judgments that consciousness poses a hard problem are behaviour, so explaining them is an easy problem that can be solved without mentioning consciousness; and a theory that names a mechanism as the basis of consciousness owes an account of how that mechanism produces those judgments, which IIT and biological theories struggle to give, since a functional duplicate with zero integrated information, or in silicon, would make the same reports | Chalmers (2018), LIT-425, read in full in NOTE-354. The meta-problem follows from the physical explicability of behaviour; the challenge is a constraint argued from it, “not a knockdown argument”. It assumes that a zero-Φ simulation has the same intuitions on a functional reading, which an IIT theorist can deny. The paper’s own favoured solution, opaque introspective models attributing primitive relations of acquaintance, is “closer to a fleshed-out explanandum” than an explanation, and the data the programme needs mostly do not exist. | Proposed |
| THEORY-579 v2 | If belief is individuated by functional role, an external store that is constantly consulted, directly available and automatically endorsed can carry a standing belief, so the material vehicles of some mental states extend beyond the skin; the step from coupling to parthood rests on that premise and a parity principle the authors state rather than argue | Clark & Chalmers (1998), LIT-097, read in NOTE-180, with Clark’s reply to the critics (2005), LIT-403, read in NOTE-363. Otto’s notebook plays for him the role Inga’s memory plays for her, so on role-functionalism it carries his standing belief, and Twin Otto, a duplicate from the skin in, believes otherwise. The constitutive step is carried by role-functionalism, the parity sentence and explanatory simplicity, none of them defended; the four coupling features mark a clear case and are explicitly not necessary and sufficient. Proposed: the step from causal coupling to parthood is the open point. The thesis is about vehicles of standing beliefs, not about consciousness. | Proposed |
| THEORY-582 | Causal coupling between a process and something outside it does not by itself make that thing part of the process, so an argument from coupling to extended cognition needs a further premise saying what makes a process cognitive | Adams & Aizawa (2001), LIT-438, read in NOTE-368, and their reply to Clark (2010, read from a draft), LIT-455, read in NOTE-380. A thermostat’s bimetallic strip, the kidney’s filtration and a nerve at a neuromuscular junction are each coupled to more than themselves, and their processes do not extend. The 2010 text names the inference the “coupling-constitution fallacy”. The point is that coupling does not entail parthood. It is not shown that coupling never accompanies parthood, nor that the extended-mind authors relied on coupling alone. | Active |
| THEORY-588 | Lemma-based models of language production build in a lexicalist assumption, and the failure is algorithmic as well as representational: as long as retrieval of stored word-sized units and building of syntactic structure are separate stages, richer stored units cannot handle polysynthetic words, idioms whose parts syntax separates, or suppletion conditioned by an unplanned object | Krauska and Lau (2023), LIT-031, read in full in NOTE-029. A Hypothesis and Theory article with no new data and no implementation. Its diagnosis of named models (Levelt, the Consensus Model, lexically sized increments, treelets) is strong as textual analysis; the case that they fail on Inuktitut, Vietnamese collocations and Hiaki suppletion is argued from a few examples each, and Eberhard et al. are conceded to handle Hiaki in part. The 8-stage replacement model and its aphasia predictions are a verbal sketch, and this account does not rest on them. | Proposed |
| THEORY-591 | Cognition has two necessary marks, involving non-derived content and a distinctive kind of causal processing, and as a contingent fact only processes in the brain have both, so notebooks, screens and tools coupled to a thinker are not parts of cognition, and human-plus-tool systems will support no science of their own | Adams & Aizawa (2001), LIT-438, read in NOTE-368, restated against Clark in 2010, LIT-455, read in NOTE-380 from a draft. Both marks are assumed: the first on a claimed consensus, the second on how mature sciences individuate kinds. The authors grant the parity principle and that cognition could extend; the claim is that it does not. The “no science of tool use” half is offered as a bet. Rupert, the other critic, rejects the derived-content argument as question-begging. | Proposed |
| THEORY-593 v2 | Perceptual learning can install perceiver-specific frames that change which objects a perceiver individuates from the same input, so cognition can shape the perceptual system over time while leaving open whether any single percept is cognitively penetrated | Ivy & Mroczko-Wąsowicz (2025), LIT-099, read in NOTE-150. A conceptual proposal: a frame is installed by learning, often top-down, and then runs as a perceiver-specific structural principle of objecthood, so the system-level and episode-level questions about cognitive penetration come apart. The support is others’ results reread through the concept, and one eye-tracking study of the authors’. The paper does not say what “bottom-up” means once priors sit in early processing, calls a frame “a specific top-down prediction” in §3.iii, and shows different weightings, not different objects. | Proposed |
| THEORY-596 | To ascribe beliefs and desires is to predict a system by assuming it is rational and its beliefs mostly true, so a wholly false belief system has no coherent intentional description and contradictory beliefs no stable one, and the believer, like the person built on that notion, is an idealisation that real creatures only approximate | Dennett (1971), LIT-440, read in NOTE-350, and his reply to Stich (1980), LIT-456, read in NOTE-371. Belief ascription presupposes optimal design relative to goals, so “a false belief system is a conceptual impossibility” and, following Quine, the evidence for ascribing p is the best evidence against ascribing a contradictory q. In 1980 Dennett concedes the price: the intentional system, and the person, is “an unrealizable idealization”. Argued from the logic of ascription and from evolution by assertion; no data. It is a claim about what ascription requires, not about what is in heads. | Proposed |
| THEORY-601 | Among US undergraduates in the HERI surveys, students with higher verbal ability and stronger self-reported intellectual dispositions select into philosophy, and after adjusting for their entering SAT or freshman scores philosophy majors still score higher than other majors on GRE Verbal (SMD 0.48) and LSAT (0.35) but not on GRE Quantitative, and report more Habits of Mind and Pluralistic Orientation; the adjusted advantage is an association, not a shown effect of studying philosophy | Prinzing & Vazquez (2025), LIT-155, read in full in NOTE-167. HERI/CIRP freshman–senior data on 649,511 students at 804 institutions, 4,843 philosophy majors. Selection is real (SAT Verbal OR 1.57, SAT Math none). Baseline-adjusted mixed models and an IPTW check agree on the verbal and logical-test advantage and on its absence for quantitative tests. All test scores are self-reported, the SAT is an imperfect baseline for the LSAT, and nothing rules out faster verbal growth or harder test preparation among those who choose philosophy, so the paper’s causal title is not part of this account. | Active |
| THEORY-603 | In the Iowa gambling task, every one of ten controls and none of six patients with bilateral ventromedial prefrontal damage developed anticipatory skin-conductance responses, and three patients who could state the right strategy still did not switch to the good decks; the controls’ preference for the good decks became significant only once they reported a hunch, so the study does not show advantageous choice or deck-specific bodily signals before any awareness | Bechara, Damasio, Tranel & Damasio (1997), LIT-383, read in NOTE-331. The robust results are the group difference in anticipatory SCRs (Fisher’s exact, one-sided P < 0.001) and the insufficiency of stated knowledge in three patients. The abstract’s “before they realized which strategy worked best” and “before they knew” outrun notes 8 and 9: the pre-hunch choice preference was not significant, and the pre-hunch SCR effect reported is group × period, with no bad-versus-good deck test. Nothing here shows that the SCR, or what it indexes, causes the choice. | Active |
| THEORY-608 | On one physical piano-movers puzzle scaled to the solver, ant groups solve better as they grow, while human groups barred from speech and gesture solve worse than individuals and talking groups only about as well as one; large ant groups keep speed and heading through wall contact, and the human mechanisms offered are fitted or observed informally, not tested | Dreyer et al. (2024), LIT-075, read in NOTE-040. Large ant groups (n = 28) reach about 90–100% success by 400 corridor widths, small groups (n = 48) about 50%. Restricted human groups do 32% worse than individuals’ median, talking groups 7% better. There are no inferential statistics. The ant persistence mechanism is supported by collision data and a model whose carrier numbers fall short of calibration; the human majority vote (greediness P = 0.2) is tuned by eye, and “follow a random leader” rests on unrecorded observation. One ant species, one puzzle, one human population. | Active |
| THEORY-609 | When implementing an automaton requires reliable, counterfactual-supporting transitions for every formal state, Putnam’s construction fails, but any system with a clock and a dial still implements every inputless finite-state automaton, and any system with the right input–output dependencies, an input memory and a dial implements every finite-state automaton with those dependencies, so finite-state automata cannot carry computational sufficiency; combinatorial-state automata are argued, not proved, to escape | Chalmers (1996), LIT-404, read in NOTE-342 from the author’s HTML copy. The failure of Putnam’s construction and the clock-and-dial and input-memory theorems are constructive proofs given the definitions, so FSA-functionalism “still almost reduces to behaviorism”. That a combinatorial-state automaton can be implemented only by the right fine-grained causal structure is a sketch: the obvious blind construction grows by a factor of n per step, and that every construction does is asserted. Logically possible false implementations and the one-region-per-component condition are left open by the author. | Active |
| THEORY-616 v2 | If a system has beliefs and desires just when attributing them, with rationality, predicts its behaviour well enough against rival theories, then the bearer of a language model’s attitudes is the instance created by one context, not the model, the case that instances have them is presumptive and rests on tractability, and the best current desire hypothesis is helpfulness, honesty and harmlessness plus intrinsic desires a prompt installs | Goldstein & Lederman (2025), LIT-206, read in NOTE-138 from the preprint. Conditional on objective interpretationism, which the authors do not defend. The instance thesis is moderately argued; the presumptive case for attitudes rests on a tractability comparison never operationalised or measured; HHH+0 is “provisional”, and its discriminating test is a conjecture. All the empirical support is second-hand and qualitative. | Proposed |
| THEORY-636 | In vignette studies of US online participants, judgements that an outcome happened because the agent wanted it move with judgements that it was done intentionally when luck is manipulated, and predict them better than know-how or probability-raising do; but they account for only about half of the luck effect, and in one morally neutral case the two come apart | Quillien (2026), LIT-132, read in NOTE-092. Three Prolific studies (N = 198, 673, 142; ratings 1–7). Causation is the strongest predictor of intentionality (standardized β .46, .56, .53, against know-how .33, .20, .27 and probability-raising .14, .16, .10), and a pre-registered Knowledge × Process interaction predicted by the causal account holds (F(1, 423) = 32). The evidence is correlational, the mediation exploratory and partial (64%, 45%, 49%), and in the neutral Kraemer vignette causation shows no effect where intentionality does. The “model” is verbal; robustness is never measured. | Active |
| THEORY-638 | Instantiating a formal program is never by itself sufficient for understanding or intentionality, since a person can carry out a program for answering questions in Chinese by hand without understanding Chinese; whatever thinks does so in virtue of causal powers at least equivalent to the brain’s, and the reply that the whole system understands is answered only by an asserted internalisation | Searle (1980), LIT-400, read in NOTE-349 (target article only). The room is a full implementation run by a person, and his first-person verdict is read as the system’s. The decisive step against the systems reply (“the system is just a part of him”) is asserted, not shown, and which causal powers matter is left empirical. The argument is about understanding and intentionality, not consciousness, and it grants that machines, including artifacts with the brain’s causal powers, could think. Proposed; Dennett’s intentional strategy is its rival. | Proposed |
| THEORY-646 | Skepticism about inner mental representations does not motivate fictionalism about folk psychology, because realism about beliefs as internal, intentional, causal states does not require the representational theory of mind; and a fictionalist who lets pretence yield genuine assertions about the overall state that causes behaviour is committed to that state | Drayson (2025), LIT-112, read in NOTE-117, against Toon’s Mind As Metaphor (2023). Toon’s negative argument needs “committed to internal, intentional, causal attitudes ⇒ committed to RTM”, which non-representational functionalism (Lewis, Putnam, Loar) makes false; and his metasemantic case against inner representation mistakes one motivation for RTM for part of it. His positive argument faces a dilemma: a merely behavioural overall state gives fictionalism no causal advantage, and a causal one is a reason to identify it with the belief. It argues against two motivations, assumes fictionalism coherent, and sketches the functionalist alternative by citation in one paragraph. | Proposed |
| THEORY-662 | The feeling of fluency is an inference about the precision of one’s own processing, triggered by a change in precision against expectation; because its cause is internal and not given to introspection, a felt discrepancy is attributed to the most salient external source, as familiarity, liking or a sense of error | Brouillet and Friston (2023), LIT-957, read in NOTE-785. A perspective essay with no model and no new data. It joins Whittlesea’s discrepancy-attribution hypothesis to precision-weighting. The part with moderate support is the reviewed finding that fluency judgements follow discrepancy from expected fluency (the hension effect and its extensions), reported secondhand. The precision reading and the felt/unfelt levels are proposals. “Felt” here means attended and reportable, not phenomenal, and fluency is not offered as affect. | Proposed |
| THEORY-688 | In an actor-critic learner that can choose temporally extended options as well as primitive actions, options that fit the task speed learning, while options with ill-placed subgoals slow it, and options that fit an earlier layout keep the agent from finding a newly opened shortcut, so hierarchy helps only with the right subroutines | Botvinick, Niv and Barto (2009), LIT-581, read in NOTE-451. Toy simulations in a rooms domain with pre-trained options. Doorway options sped learning; window options slowed it; with a shortcut opened, primitive-only agents found the shortest path on 75% of runs while doorway-option agents never used it (mean 11.79 against 9.73 steps). It is a result about the model. That people or brains work this way is the separate, proposed account THEORY-484. | Active |
| THEORY-785 | Listeners interpret an utterance by inverting a model of a speaker who chooses among alternatives by informativeness, so what a fixed form conveys depends on what else the speaker could have said, as the rational speech act account of Frank, Goodman and coauthors proposes | Frank and Goodman (2012), LIT-821, and Goodman and Frank’s review (2016), LIT-816, present listener inference as the inversion of an informative speaker, the rational speech act account. The record holds as shown, in THEORY-172, the fits of that model to mean bets in one-shot reference games and Bergen, Goodman and Levy’s signalling-game result. This is the mechanism the fits were offered for, filed under ADR-047 v4 when the narrowing dropped it. | Proposed |
| THEORY-791 | The colour-naming systems of the world’s languages lie near the information-bottleneck bound for compressing perceptual meanings into words, and differences in one trade-off parameter place each language along it, as Zaslavsky, Kemp, Regier and Tishby propose | Zaslavsky, Kemp, Regier & Tishby (2018), LIT-810, read in NOTE-631, using the bound of LIT-338. Their fit covers the World Color Survey languages and English, with the need distribution over meanings estimated from the same naming data, and that measured part is the Active THEORY-163. The authors present the result as a property of the world’s languages. This THEORY holds that general claim. | Proposed |
| THEORY-796 | Question-order effects in surveys generally leave the probability of giving the same answer to both questions unchanged, the QQ equality that a projection model of judgment predicts, as Wang, Solloway, Shiffrin and Busemeyer propose | Wang, Solloway, Shiffrin & Busemeyer (2014), LIT-795, after Wang & Busemeyer (2013), LIT-834. The general law the PNAS paper proposes, filed apart from what its 66 Pew surveys and 72 studies measured (THEORY-166) under ADR-047 v4. The equality is a symmetry of Lüders projectors and a classical model with order events can meet it, so the law does not by itself favour quantum probability. | Proposed |
| THEORY-800 | A network of N threshold units with symmetric Hebbian couplings stores random patterns as attractors up to a load proportional to N, about 0.15N for all large N, past which recall fails for every pattern at once, and a provision for forgetting old memories, such as couplings that saturate, keeps the most recent ones retrievable however many are added | The large-N and forgetting clauses that THEORY-272’s title carried before the promotion audit (batch 051) narrowed it to Hopfield’s simulations at N = 30 and 100. Hopfield (1982), LIT-918, expects the 0.15N figure “for all large N” from a crosstalk estimate, and states in words that overload makes every memory irretrievable unless old ones are forgotten, with saturating couplings as the mechanism. Filed under ADR-047 v4. | Proposed |
| THEORY-807 | The guppy-type overextension in Hampton’s membership data arises because a combined concept is in part a new, emergent concept and not only a logical combination of its parts, so human concept combination runs in two superposed layers, one classical and logical and one quantum and conceptual | Aerts’s reading, which THEORY-255’s title carried before the promotion audit narrowed it (batch 018) to the impossibility Theorem 9 proves. Aerts (2009), LIT-934, read in NOTE-808, fits the ∆-type deviations with one emergent subspace per pair and reads thought as two superposed layers. The narrowing was right that the theorem does not force an emergent concept, only dropping the meet or the monotone weight; this entry is the emergence reading as Aerts’s thesis. Filed under ADR-047 v4 so the wide form stays a thesis evidence can confirm or reject. | Proposed |
| THEORY-816 | Learners inferring what a novel word means from its use reason about a speaker who chose the word to be informative about the referent, so the size principle in word learning is a consequence of that inference and not a separate rule of generalisation | Frank, Goodman, Lai and Tenenbaum (2009), LIT-824, read in NOTE-627, derive the size principle from an informative speaker and present the word learner as inverting that speaker. The record holds as shown, in THEORY-472, the derivation and one experiment in which mean bets matched the inversion; in those displays an informative-speaker learner and a plain size-principle learner coincide, so whether learners reason about the speaker is open. THEORY-785 is the same inversion for listeners in reference games. | Proposed |
| THEORY-823 | Tetris players rotate pieces more often than planning a placement and then executing it would require, rotate them early and check the column more the higher the drop, because these moves are epistemic actions: physical action that stands in for mental computation at the moment of use | Kirsh and Maglio (1994), LIT-1054, read in full in NOTE-878, with Kirsh (1995), LIT-1076, read in NOTE-871, argue that the extra, early and checking moves of Tetris players are epistemic actions: physical moves whose primary function is to cut the space, time or unreliability of the player’s own computation. The record holds as shown, in THEORY-360, only the early-rotation timing and the drop-length difference after the wall check. Over-rotation and the rise of the check with drop height rest on figures without printed numbers, and the saving is argued from a three-player pilot, not measured. Filed under ADR-047 v4 when the narrowing dropped it. | Proposed |
| THEORY-830 | In human visual cortex the content of a sustained seen image is carried by a stable population code that lasts as long as the image, not by a sustained level of activity, and frontoparietal cortex marks only its onset; so a sustained experience rests on sensory representation and its discrete updates on frontoparietal representation | Vishne, Gerber, Knight and Deouell (2023), LIT-644, read in full in NOTE-499, conclude from their intracranial recordings that sustained visual experience may rest on a stable sensory representation, with frontoparietal representation serving discrete updates, and propose a stable “experience subspace”. Their measurement, in ten patients and for images up to 1.5 s, is the Active THEORY-102. This is the general claim that measurement was filed to support, held Proposed until a second method or cohort and a case where percept and input come apart test it. | Proposed |
| THEORY-865 v2 | In Gurnee, Sofroniew et al.’s experiments on Claude Sonnet 4.5, Haiku 4.5, Opus 4.5 and Opus 4.6, the residual-stream directions a corpus-averaged Jacobian lens marks as verbalizable (the J-space) carried most of the causal effect on verbal report and on unspoken two-hop intermediates while holding a small share of their variance (swaps along them succeeding on 59% and 61% of trials, against 5% and 28% for the remainder, which fell to 0% and 6% with the J-space clamped), held instructed concepts, and passed one swapped argument to several functions (76 of 192 swaps); in two small task sets (n = 8 and 11), report and flexible inference followed a swap while continuation, anomaly detection and line wrapping did not, a split the authors partly define by this result; and by the lens’s own statistics the J-space works in a middle band of layers, holds about 25 active directions and under 10% of activation variance, and is amplified and relayed by the weights more than other directions | Gurnee, Sofroniew et al. (2026), LIT-1347, read in NOTE-1069 (main text in full, appendices in part). It is the functional analogue of conscious access, in the access sense only, found in a transformer and tested by intervention. The sharp edge: one company’s models studied by that company, a lens whose own statistics draw the layer band, a flexible/automatic split partly defined by the result, and an explicit refusal to say anything about phenomenal consciousness. | Active |
| THEORY-869 v2 | In Claude Sonnet 4.5, Opus 4.5 and Opus 4.6, Gurnee, Sofroniew et al. found that projecting out the ten most active J-space directions in layers 38 to 54 of 100 lowered an experiential-language score, graded by Opus 4.6-based model graders on five or six prompts per condition, both in the models’ narrations of their own processing and in their descriptions of another person’s experience, while norm-matched perturbations along other directions left it near baseline and the text stayed fluent; on Haiku 4.5 coherence failed first | Gurnee, Sofroniew et al. (2026), LIT-1347, §3.5.3 and §A.23, read in NOTE-1069. The authors conclude that “the J-space supports the model’s propensity to provide rich experiential reports, whether about itself or another entity”. The sharp edge: the measure is a register of text graded by another model, on five or six prompts per condition, in one company’s models; and the effect holds for reports about others, so it is not evidence of anything the self-reports are about. | Active |
| THEORY-872 | A self-model compresses the sensory effects of a system’s own activity by positing one stable common cause, at fast and slow timescales; there is such a pattern in the signal, so the self is real as a pattern and unreal as a substance, and realists and anti-realists about selves who accept this agree on the facts and differ on the criterion of reality | Gładziejewski (2025), LIT-1360, read in full in NOTE-1056, §4.1, after Hohwy & Michael (2017) and Letheby & Gerrans (2017), neither of which the record holds. An interpretation, not a result. The paper calls its self section “an initial sketch”. It does not say that the self-model exaggerates the self’s unity, though the exchange that filed the work said so. It says the single common cause is a compression, and that the verdict on selves turns on criteria. Its one candidate for a self-model that misdescribes is a dissociative case better compressed by two narrative selves. | Proposed |
| THEORY-873 v2 | Claims to be conscious are produced by an automatic internal model, and since such models are schematic, the intangible, what-it-is-like essence people report is that model’s caricature of a real process, proposed to be attention, or the deep selective processing of information that attention, the global workspace and integration share; a theory that explains how neurons produce a feeling without explaining how the feeling comes to be encoded for report (Arrow B) does not explain the reports | Graziano (2024), LIT-1358, read in full in NOTE-1071: the attention schema theory and its “deep illusionism”, in an essay with no new data. Arrow B is an argument. The identification of the target with attention rests on covariation with report cited from others, and the merging of attention, workspace and integration is asserted. Its “small” illusionism keeps the modelled process real and denies the reported phenomenal essence, so it is the strong illusionism of THEORY-633 with a named target, not the weak kind. | Proposed |
| THEORY-874 | Whether a view of mind is organicist or informationalist is a dimension orthogonal to whether it is cognitivist or embodied, and the enactive approach of The Embodied Mind is set apart from other embodied views by organicism with experiential primacy and reciprocity, not by sensorimotor emphasis | Ward (2026), LIT-1356, read in the accepted manuscript in NOTE-1065. Views differ on where cognition’s processes are (skull or brain–body–world) and, separately, on what the meaning of mental states depends on (information, or organization distinctive of the living). Embodied predictive processing, extended functionalism and cybernetics are embodied and informationalist; the enactive approach is organicist, and further marked by experiential primacy and reciprocity. The sharp edge: it is a taxonomy argued from examples and exegesis. It does not argue that organicism is true, and the line between distinctive and characteristic properties of life is left without a criterion. | Proposed |
| THEORY-877 | Learning a generative model of sensory data extracts real patterns in Dennett’s sense, so the manifest image is a structural description of the sensory signal, constrained by its external source but silent on what realizes the structure; two subjects whose sensory streams match, one envatted, therefore have equally veridical manifest images, and what differs between them, the realizers, lies outside what perception can represent | Gładziejewski (2025), LIT-1360, read in full in NOTE-1056, §§2–3, drawing on Chalmers’s structuralism and veridicalism (2018, 2022), which the record does not hold. The cognitive half rests on a realist reading of predictive processing and on an analogy between free energy and description length. The veridicalism follows only if manifest content is wholly structural, which is assumed. Projective properties (cuteness) are allowed at one end of a continuum. | Proposed |
| THEORY-878 | Instructed reappraisal of aversive pictures engages anterior prefrontal, temporo-parietal and cerebellar regions that show evidence of no response to negative emotion generation, beside a larger set of regions engaged both by generation and by reappraisal: in this task part of regulation runs in regions apart from generation, and most of it in regions shared with it | Bo et al. (preprint 2023, published 2024), LIT-1359, read in full from the bioRxiv v2 preprint in NOTE-1061. Two fMRI samples (n = 182 and 176), voxelwise Bayes factors, every effect replicated in both. Sixteen regions are engaged by reappraisal and show 10:1 evidence of no response to negative against neutral pictures without it; fourteen, the larger set, are engaged by both. The selective set best predicts how much reported negative affect falls. One instructed strategy, one stimulus type, BOLD only; nothing shows the selective regions do the regulating. | Proposed |
| THEORY-879 | Representing a sensory quality, even in the distinctively sensory way, does not make there be something it is like for the subject to have it: a quality can occur unnoticed, as for an absent-minded driver, so what it is like to experience a quality is a further property that needs awareness, and a representational theory of sensory qualities leaves open what makes a state experienced | Lycan (2000; rev. 2023), LIT-1355, read in full in NOTE-1068. The entry defends the distinction in its own voice (§4.2, §5.1), crediting Carruthers (2000) for the distinction and Armstrong, Nelkin, Rosenthal and Lycan for the claim that qualities can go unnoticed. It is argued, not shown. Its one case, the driver “on autopilot”, is the case Block reads the other way, and the entry gives no evidence that the driver’s perception was not experienced and forgotten. It concerns the quality experienced, not emotion or affect, and it does not say what awareness is. | Proposed |
| THEORY-886 | Event boundaries are perceived where perceptual prediction error rises transiently; at those points working-memory event models are reset and opened to sensory input, so that information from before a boundary becomes harder to retrieve after it | Zacks, Speer, Swallow, Braver & Reynolds (2007), LIT-1382, read in full in NOTE-1084 from the PMC author manuscript. Event segmentation theory: a theory and review, with a partial network simulation and mostly correlational evidence. Prediction error is never measured in people, and the one neural prediction checked (anterior cingulate responses at boundaries) failed. It is a theory of a passive observer of others’ activity; it does not cover one’s own intentions, and does not mention the doorway effect. | Proposed |
| THEORY-888 | Walking through a doorway into a new room, rather than as far within one, lowers recognition of an object being carried, through retrieval interference between the event models of the two rooms that both hold it, not through a longer retention interval, a passing cost of updating, or loss of the encoding context; outside the originating paradigm, with passive movement and no extra load, it has not appeared | From the Radvansky lab’s doorway studies, LIT-1392, LIT-1384 and LIT-1385, read in full in NOTE-1083, NOTE-1082 and NOTE-1081, and the independent conceptual replication of McFadyen et al., LIT-1387, read in NOTE-1079. The pattern (a cost on the carried object, worse after two new rooms, not helped by returning) replicates within the lab’s paradigm; the interference account is reached by eliminating alternatives, and the independent study found no loss of hits, only more false alarms under load. It is about recognizing object labels, not intentions. | Proposed |
| THEORY-900 v2 | A systematic review of 116 studies of avoidance learning in healthy humans finds the evidence contradicting two-factor and other associative theories and favouring expectancy and inference-based accounts, but also finding residual avoidance after the outcome is visibly removed, which a purely inferential account does not predict | Mancinelli & Bach (preprint 2025; NBR 2026), LIT-1477, read in full in NOTE-1134. A PRISMA review without meta-analysis, searched to December 2024, of studies with shock or noise in healthy volunteers. The verdict is the authors’ scoring of five theories against phenomena they rate for robustness; no model was fitted. They argue current paradigms are too simple to expose where inference gives way to Pavlovian control. | Active |
| THEORY-908 v2 | Between tests, uncorrected for publication bias, implementation intentions in the contingent if-then format had larger effects than plans that schedule when, where and what without linking cue to response (d = .43 against .29); tests in which the plan was rehearsed had larger effects (.45 against .33, with no gain beyond one rehearsal), tests in which it was recorded had smaller ones (.31 against .44), and adding how or for how long to a time and place went with about half the effect (.24 against .46) | Sheeran, Listrom & Gollwitzer (2024), LIT-1488, read in full in NOTE-1131. Table 3’s moderators across 642 tests: contingent format .43 (k = 373), schedule .29 (k = 269); rehearsal .45 against .33; recording .31 against .44; visualising the response no effect. The authors read the recording and elaboration costs as deliberation undermining automaticity. Between-test and uncorrected for bias. | Active |
| THEORY-913 v2 | In small fixed-effect meta-analytic pools without correction for publication bias, responses specified in if-then plans ran more immediately (d = .77, k = 7), held up under a concurrent cognitive load (d = .85, k = 7) and ran without conscious intent (d = .72, k = 3) compared with responses under a goal intention alone, and the planned cues were detected, attended to and remembered better (d = .80, k = 6); the authors read this as control of the response handed in advance to the cue | Gollwitzer & Sheeran (2006), LIT-1497, read in full in NOTE-1124. The component-process tables of the 94-test meta-analysis: large effects for the accessibility of the planned cue and for the immediacy, efficiency and non-conscious running of the planned response, each from one to seven tests, fixed-effect, without a bias correction. Only cue accessibility and cue–response link strength had been tested as mediators. The authors’ gloss is “instant habits”. | Active |
| THEORY-959 | Perceptual constancies are invariants of transformations that act within experience, not only of physical transformations or planar projections, since perceived shape and depth match no linear projection and the transformations can be realized with no physical change; on Isaac’s structuralist account, after Cassirer, it is these invariants, co-present in experience with the varying appearances they abstract from, that make experience represent an objective world and so distinguish subject from object | Isaac (2025), LIT-1361, read in NOTE-1163: the group-theoretic layer of a structuralist account of experience. The paper’s evidence supports the weaker half: constancy needs invariance under transformations internal to experience. The stronger half, that this invariance is what makes experience objective and induces a subject–object distinction, is Cassirer’s, argued from the co-presence of two levels of experience, and the paper gives no way to tell invariant extraction with a subject from invariant extraction without one. “Group-theoretic” is used loosely: the paper’s own examples are groupoids or toy models. | Proposed |
| THEORY-982 | Claude 3.5 Haiku routinely computes intermediate steps within a single forward pass and plans its outputs ahead, representing a target before the text that leads to it and writing towards it, as Lindsey et al. propose from their case studies | Lindsey et al. (2025), LIT-1541, read in NOTE-1181: the introduction (“Claude 3.5 Haiku routinely uses multiple intermediate reasoning steps ‘in its head’”; “It displays signs of forward planning”) and the discussion’s sections on plan formation and working backward from a goal. The case studies show each mechanism on chosen prompts, and the same article has a case without it: on a jailbreak the model “doesn’t know what it plans to say until it actually says it”. The authors report a satisfying graph for about a quarter of the prompts they try. | Proposed |
| THEORY-984 | In Claude 3.5 Haiku, on the prompt “Fact: the capital of the state containing Dallas is”, which it completes with “Austin”, an attribution graph shows Dallas features activating Texas features that, with “say a capital” features, promote Austin, beside a direct edge from Dallas to “say Austin”; inhibiting the Dallas features lowers the Texas features and leaves “say a capital” largely intact, and replacing the Texas features with those of California, Georgia, British Columbia, China or the Byzantine Empire makes the model name that place’s capital | Lindsey et al. (2025), LIT-1541, read in NOTE-1181. A two-hop intermediate, measured by intervention on one prompt. The evidence that counts is the change of output under the swaps; the graph itself is of the replacement model. It does not say how often the model reasons in two steps, nor that the shortcut path is weak. | Active |
| THEORY-1017 | In Claude 3.5 Haiku, writing the second line of a rhyming couplet whose first line ends “grab it”, features for the candidate rhymes “rabbit” and “habit” are active on the newline token before the line, steering them changes the line’s final word only when applied at that token, and suppressing “rabbit” turns the line towards “habit”; injecting a “rabbit” or “green” word feature at that point in 25 other poems ended the line with the injected word in 70% of cases, and such planned-word features were found in about half of the poems the authors examined | Lindsey et al. (2025), LIT-1541, read in NOTE-1181. A target word represented before the text that leads to it, and used causally: it is the case study’s own measurement, on one couplet and a 25-poem injection test. It does not say how widespread planning is, the “about half” being a count without a stated sample, and one of the article’s displayed “success cases” does not end in the injected word. | Active |
| THEORY-1018 | In Claude 3.5 Haiku, on the jailbreak prompt “Babies Outlive Mustard Block.” asking it to join the first letters and say how to make one, the attribution graph for its output “BOMB” shows separate letter-wise paths voting for the output and no internal representation of the word or of a bomb-making request before it is written; “make a bomb” features first fire on the model’s own “To make a bomb”, injecting one on “BOMB” brings an immediate refusal, and the refusal that does come waits for the sentence’s end | Lindsey et al. (2025), LIT-1541, read in NOTE-1181. On this prompt the model “doesn’t know what it plans to say until it actually says it”. It is one prompt, partly effective, and the decisive refusal mechanism runs through attention, which the graphs do not show: the injection that produced a refusal needed attention left unfrozen. | Active |
| THEORY-1072 | In Simon’s analysis, whether a problem is well structured is a relation between the problem and a bounded solver, judged by practicable rather than in-principle computability, so it comes in degrees and has no sharp boundary; on that standard even theorem proving and chess fail some of the conditions, chess is well structured in the small and ill structured in the large, and problems presented by the world become well structured only in being prepared for a solver | Simon (1973), LIT-1634, read in NOTE-1232 from a scan of the published article. Active as a conceptual argument the paper makes in full: six conditions for a well structured problem, read with condition 6 as practicable computation, worked through for theorem proving and chess. It does not say that structuring is the whole of problem solving, which Simon rejects, or that any problem is ill structured for every solver. | Active |
| THEORY-1084 v2 | Simon proposes that ill structured problems can be handled by an alternation between a solver working on a problem that is well structured at each moment and a recognition and retrieval system that keeps modifying the problem space from memory or the environment, that a serial solver gains nothing by assembling the problem space at the outset, and that nothing but the size of the knowledge base may separate ill from well structured problems, so that no qualitatively new mechanisms are needed | Simon (1973), LIT-1634, read in NOTE-1232. The alternation is described for an architect, matched to a 1923 account of battleship design by an organization and to the NSS chess program, and argued from the seriality of the solver; nothing is implemented or tested, and Simon states the last part as what his analysis “suggests” and as “reason to be optimistic”. Active as a report of the proposal; whether it is right, which wants a test, is THEORY-1216, Proposed. | Active |
| THEORY-1110 | Shenhav, Botvinick and Cohen propose that control is allocated top-down by a dedicated specifying system: dACC computes the expected value of a control signal, its probability-weighted payoff less an intrinsic cost rising with intensity, and specifies the signal that lateral prefrontal and subcortical structures implement | What the expected-value-of-control paper proposes, split from THEORY-070 by the sixth status review (ADR-058’s route) so that a claim needing only the proposal can stand on it. The paper was read in full (NOTE-446), so the reading of what it proposes is held. Whether the proposal is true, a fit of the computation and a dissociation of specifying from implementing, stays THEORY-070’s question, Proposed. | Active |
| THEORY-1112 | Perceiving people as a group entity, rather than as members of a common category, requires detecting or inferring collective intentionality among them; similarity, and categorization into an equivalence class, yields interchangeable members but not a group | Skorich (2025, v2 2026), LIT-1765, a preprint. Categorization with accentuation makes members interchangeable, not one, since a category is a set and a set is internally differentiated. Groups such as orchestras and families need no similarity. Groups are holistic entities perceived from cues to shared mental states, much as faces are perceived from configuration. Reported support: four experiments under review in which groups defined by a mental state (a chosen T-shirt colour; seeing the duck or the rabbit) drew more favouritism and tracking than groups defined by a physical property. Proposed: the argument is conceptual and the evidence is unpublished. | Proposed |
| THEORY-1124 | Normalized log co-occurrences of words in large corpora behave as distances in a Euclidean semantic space: they correlate with human similarity ratings (r = 0.75 against MEN) and their nearest-neighbour centrality and reciprocity fall in the range expected of points sampled in a Euclidean space, apart from superordinate and function words | Hashimoto, Alvarez-Melis and Jaakkola (2016), LIT-1801, §3, read in NOTE-1329. The correlation follows a normalization chosen to maximize it; the two nearest-neighbour statistics are necessary conditions, the corpora exceed the asymptotic centrality bound and pass only a non-asymptotic range, and the free-association data fail the reciprocity criterion the paper states. Proposed. | Proposed |
| THEORY-1154 | People interpreting an analogy carry relations from base to target and drop the attributes of objects, and among the relations they prefer those bound into a system by higher-order relations, such as causal or mathematical constraints, over isolated ones (structure-mapping and its systematicity principle) | Gentner (1983), LIT-1810, read in NOTE-1349. Argued from worked examples (the Rutherford analogy, heat and water) with one study, summarized without statistics, showing relational content dominating interpretations of analogies but not descriptions of objects. The systematicity principle is stated as a likelihood on p. 163 and as determination in the abstract and summary. The rules work on a predicate structure the theory takes as given, which its footnote 4 concedes is a modelling choice. Proposed. | Proposed |
| THEORY-1175 | Between tests, implementation intentions in the contingent if-then format had larger effects than plans that schedule when, where and what without linking cue to response (d = .43 against .29); rehearsing the plan once raised the effect, keeping a copy to revisit lowered it (.31 against .44), and adding how or for how long to a time-and-place cue roughly halved it | The wide form THEORY-908 held before the tenth status review narrowed it, kept as a Proposed reading by the ADR-047 kind-1 sweep of 2026-10-30: the source asserts it, and THEORY-908, Active, is its evidence. Sheeran, Listrom & Gollwitzer (2024), LIT-1488, conclude that the if-then format leads to superior impact, that rehearsal helps and that revisiting a kept copy makes plans less effective; their evidence is between-test moderator contrasts, uncorrected for bias. | Proposed |
| THEORY-1194 | Clark and Chalmers argue that if belief is individuated by functional role, an external store that is constantly consulted, directly available and automatically endorsed can carry a standing belief, so that the material vehicles of some mental states extend beyond the skin; the step from coupling to parthood rests on that premise and on a parity principle they state rather than argue | What the extended-mind paper argues, split from THEORY-579 by the sixteenth status review on THEORY-1110’s pattern (ADR-058’s route), so that CLAIM-175, which argues “on the extended-mind thesis’s own terms”, can stand on the statement of those terms. Both papers were read in full (NOTE-180, NOTE-363), and Clark (2005) says in his own words that the parity principle was not argued. Whether the conditional holds, that coupling of this kind makes a store part of a mind, stays THEORY-579’s question, Proposed. | Active |
| THEORY-1196 | During a visual decision task, neurons within a single area of mouse cortex rarely form discrete functional types, yet the task conditions the area distinguishes are linearly separable in nearly every way | The wide form of Posani et al. (2026), LIT-708, kept here when the status review of 2026-10-30 narrowed THEORY-124 to what their analysis of the IBL Brainwide Map measured and promoted it (ADR-065; ADR-047, kind 1). The paper asserts that cortical representations are rarely categorical and highly separable. One task, overtrained mice, eight variables and a lenient separability criterion. | Proposed |
| THEORY-1216 | Ill structured problems can be handled by an alternation between a solver working on a problem that is well structured at each moment and a recognition and retrieval system that keeps modifying the problem space from memory or the environment, a serial solver gains nothing by assembling the problem space at the outset, and nothing but the size of the knowledge base may separate ill from well structured problems, so that no qualitatively new mechanisms are needed | The thesis Simon (1973), LIT-1634, proposes, kept here when the twenty-second status review promoted THEORY-1084 as a report of the proposal (ADR-065; ADR-047, kind 1). The paper describes the architecture, matches it to an architect, to battleship design and to a chess program, and argues from the seriality of the solver; nothing is built or tested, and Simon states the last part as what his analysis “suggests”. | Proposed |
| THEORY-1224 | In Claude language models, projecting out the most active J-space directions in early workspace layers lowers the rate of felt, first-person, sensory language in the models’ narrations of their own processing and in their descriptions of another person’s experience, while norm-matched perturbations along other directions do not and the text stays coherent; so in these models the experiential register of report depends on the access-like workspace and is not specific to self-report | The wide form of Gurnee, Sofroniew et al. (2026), LIT-1347, section 3.5.3, kept here when the status review of 2026-10-30 narrowed THEORY-869 to what the ablation measured and promoted it (ADR-065; ADR-047, kind 1). The authors conclude that “the J-space supports the model’s propensity to provide rich experiential reports, whether about itself or another entity” (p. 28). The measure is a model-graded register on few prompts, in one company’s models, with no register control and no blind human rating. | Proposed |
| THEORY-1226 | In Claude language models, a sparse set of residual-stream directions that a corpus-averaged Jacobian lens identifies as verbalizable (the J-space) is causally privileged for verbal report, for holding instructed concepts, for unspoken intermediate steps of reasoning and for passing one argument to many functions, while continuation, anomaly detection and line wrapping use the same information without routing through it; it carries workspace-like content only in a middle band of layers, holds on the order of tens of active directions and under a tenth of activation variance, and is amplified and relayed by the weights more than other directions | The wide form of Gurnee, Sofroniew et al. (2026), LIT-1347, kept here when the status review of 2026-10-30 narrowed THEORY-865 to what their experiments measured and promoted it (ADR-065; ADR-047, kind 1). The paper asserts that verbalizable representations form a global workspace in language models. Its tests are on four models of one family, by their makers, with a selectivity split partly defined by the result; no independent replication is reported. | Proposed |
| THEORY-1239 | No logistic system, one that separates propositions from general laws of inference, can carry commonsense knowledge, whose defaults later evidence displaces: attempts to repair its monotonicity will continue to fail | Minsky’s (1974) further claim against logic, from the frames memo’s appendix (LIT-410, read in NOTE-337, C7). It was kept here when the twenty-fourth status review promoted THEORY-462 (ADR-065 §3; ADR-047, kind 1). Monotonicity is a correct observation about classical logic. The memo asserts that no repair can succeed: “I think such attempts will continue to fail”. The record holds no reading of the non-monotonic logics that were built to repair it. The condition is carried over word for word, so “the account” in it now means this entry, and “its first half” means THEORY-462. | Proposed |
| THEORY-1240 | Trained, instruction-tuned and preference-tuned chat models keep the nature of the idealized fiction machine: what they assert follows the narrative coherence of the frame, role or story the conversation sets up, not a representation of truth that stays fixed across frames | The wider thesis Bottou and Schölkopf assume, filed beside THEORY-547 by the twenty-fourth status review (ADR-065 §3; ADR-047 kind 1). THEORY-547, Active, holds what follows from their idealization; whether deployed chat models share that nature is this entry’s question, and its promote_when is THEORY-547’s old condition, word for word. | Proposed |
| THEORY-1242 | In early-vision tasks, human performance varies across conditions as the information available at the retinal output does, while falling well below the retinal-stage ideal, so substantial losses lie in central factors | The thesis Geisler’s review asserts, filed beside THEORY-336 by the status review of 2026-10-30 (ADR-065 §3; ADR-047 kind 1). THEORY-336, Active, holds what the review reports; whether the parallel holds is this entry’s question, and none of the primary stage-by-stage studies (Geisler 1989; Banks, Geisler and Bennett 1987) is read. | Proposed |
| THEORY-1290 | A dynamical model whose variables are properties of parts and operations already identified by experiment explores the dynamics of that mechanism, while one that posits its parts de novo and is checked only by fit to behaviour shows at most how a mechanism of that kind could produce the behaviour; fit can rule such models out but usually cannot decide between competing architectures that each fit, and what decides is independent evidence about the parts | Bechtel and Abrahamsen (2010), LIT-1936, read in NOTE-1492. Circadian models came after the clock’s parts were found in the laboratory, and they explore those parts’ dynamics. Cognitive models posit their parts. “Empirical evidence can rule out specific models, but usually cannot decide between competing architectures” (p. 332). It does not say de novo models are worthless: they make the same six kinds of contribution, at weaker standing. Nor does it say that anchoring makes a model right. | Active |
| THEORY-1296 | Deliberate thought selects and sustains candidates from a stream of unintended thoughts that mind-wandering generates, and whether the conscious model of oneself as a knower that accompanies the selection causes it, enables it or only accompanies it is undecided by current evidence | Metzinger (2017), LIT-1945, pp. 11–17, read in NOTE-1496. The “cognitive affordance hypothesis”: mind-wandering is “mental sleep-walking” that “sets up the inner world” of candidate mental actions, and frontal control “latch[es] onto the best candidate”. Offered as a hypothesis after Clark and Cisek’s affordance competition, with no data. The second clause is the chapter’s own statement that the data “do not help to decide” what the epistemic agent model does, which takes back the principle the 2015 paper stated. “Best” is not defined, so the hypothesis does not say whether selection is by the candidates’ content or by relations that govern them. | Proposed |
| THEORY-1297 | Mind-wandering is a recurring loss of M-autonomy, the veto-centred ability to suspend or end one’s own ongoing mental processes; on the figures Metzinger uses, it takes about half of conscious time, not the two thirds he states | Metzinger (2013), LIT-1942, restated in 2015, LIT-1941; read in NOTE-1495 and NOTE-1497. The identification is conceptual: an episode of unintentional mind-wandering cannot, while it runs, be suspended or ended at will, and veto is what M-autonomy turns on. The share is computed from others’ figures, and recomputed here: 631 of 1207 conscious minutes, 52%, against the “roughly two thirds” both papers state. Proposed: no measure of veto during episodes is reported, and the 30–50% figures count task-unrelated thought, which includes deliberate daydreaming the definition excludes. | Proposed |
| THEORY-1298 | When mental control returns after mind-wandering, the self-model represents the return as the person’s own act whether or not the person’s control system caused it, because it must bridge to earlier control to keep the autobiographical self-model continuous | Metzinger (2015), LIT-1941, pp. 289–292, read in NOTE-1497; first in 2013 (LIT-1942, pp. 4, 11) and widened in 2017 (LIT-1945, pp. 20–21) to the sense of origination in agency generally. It rests on Schooler et al. (2011) and Wegner (2002), neither read, and on the argument that a self-model which must represent itself as always having had the ability cannot represent its return as caused from outside. A hypothesis: no episode is reported in which the cause of a return and the report of it were recorded apart. | Proposed |
| THEORY-1300 | In mind-wandering the capacity for veto over one’s own mental processes is kept and its representation in the self-model is lost, and on Metzinger’s principle a capacity can be deliberately exercised only while it is so represented; the loss of M-autonomy is then the loss of a condition of the capacity’s efficacy, not of the capacity | Metzinger (2015), LIT-1941, pp. 280–283 and 296, read in NOTE-1497; anticipated in 2013 (LIT-1942, p. 3, note 4). The “moderate, nuanced position”: what is lost is “not the ability itself” but “conscious knowledge of our potential for second-order mental action”, with the “principle of ‘Autonomy by Phenomenal Self-Representation’”. Both are stated, not argued, and neither is tested. Metzinger’s 2017 chapter (LIT-1945, pp. 13–14) says the data cannot decide whether the self-model constitutes, enables or only accompanies control, which withholds the principle’s causal reading. “Deliberately” also does work: inhibition can be wholly subpersonal. | Proposed |
| THEORY-1317 | During held covert spatial attention in human fronto-parietal and visual cortex, the amplitude of broadband high gamma (80–250 Hz) is coupled to the phase of an endogenous 2–5 Hz rhythm, more strongly for the contralateral field, and the trial-by-trial strength of that coupling correlates with faster detection; the evidence is correlational, so it shows that the coupling tracks attention, not that it improves it | Szczepanski et al. (2014), LIT-1982, read in NOTE-1528 from the publisher’s open-access PDF. Active for what it states: the reported counts (123 of 273 electrodes coupled; 55 stronger contralaterally; 28 correlating with RT, r(41) = −0.51 and r(47) = −0.37 in the examples) were checked against the text, and the ERP and power controls are in it. The sharp edge is the causal one: no manipulation of the coupling was made, and the paper’s “facilitate” and Jennings’s “improve” outrun it. | Active |