FMT Prediction Map
The Four-Model Theory of consciousness as a network of claims and predictions, compared with five other theories. Map v1, built from FMT v19, 2026-10.
This map sets out the Four-Model Theory of consciousness (FMT) as a network of claims. Each box is one claim. Each arrow says that the claim it points to holds in virtue of the claim it starts from. The three principles of the theory form the core, in the left column. Every column to the right specifies the one before it, and the last two columns hold the theory's empirical predictions: neuroscientific ones at level 4, cognitive and behavioural ones at level 5. Select a box or an arrow to read it.
The page also carries the maps that Negro, Elbaz, Fischer, Schreiber and Mudrik (2026) drew for four other theories: the Global Neuronal Workspace Theory (GNWT), Integrated Information Theory (IIT), Perceptual Reality Monitoring (PRM) and Recurrent Processing Theory (RPT), and a map of Predictive Processing (PP) that we drew ourselves from its proponents' publications, with the same method. Choose a map with Map, or put a second one next to it with Compare with.
- cognitive
- substrate-general
- brain-specific
- extrapolation (animals, machines)
- core principle
- proposed prediction, not yet in the FMT master
- across levels
- same level
Links between the two maps
In comparison mode, lines between the two maps join claims that say the same thing (green), overlap (amber, dashed) or contradict each other (red). Select a line or a row to read why. These links are our judgement (Matthias Gruber), not the map authors': each was judged from the claim texts alone, every claim of one map against every claim of the other, and each gives its reason. They mark overlap and conflict; they do not say which theory is right. All links of all ten pairs: cross-theory-links.json.
About the format
The format is that of the prediction maps of Negro, Elbaz, Fischer, Schreiber and Mudrik (2026), who mapped the major theories of consciousness as claim networks and measured which claims carry the most weight in each. FMT's principles describe a computational substrate, not the brain, so the map uses the layout they give for substrate-general theories such as Integrated Information Theory: levels 1 and 2 hold mostly substrate-general claims, level 3 brain-specific claims and extrapolations to animals and machines, level 4 neuroscientific predictions and level 5 cognitive predictions. Their maps of GNWT, PRM and RPT use the layout for cognitive theories, in which level 3 holds cognitive predictions, level 4 the neural implementation and level 5 neuroscientific predictions.
How to read it
- Colour gives the claim class: cyan for claims about cognitive functions, green for claims about a physical substrate in general, pink for claims about the brain, yellow for claims about consciousness beyond healthy adult humans. A thick border marks the core. On the FMT map a dashed purple border marks a proposed prediction (E12, E13, E14, E15, F11): added in October 2026 for the temporal echo (C5), the meta-problem (C8) and binding (C9), not yet in the FMT paper. The maps of the other theories keep the colours of the authors' files, except GNWT, whose file has none: there the colours are transcribed from the preprint's Figure 3.
- Solid arrows join claims at different levels; dashed arrows join claims at the same level.
- Select a box to read the full claim and its centrality. On the FMT map the panel also gives the section of the FMT paper the claim comes from, its status (derived, corroborated, untested, proposed and so on), whether other theories make the same prediction, and its references. Select an arrow to read why the one claim follows from the other.
- The measure selector sizes the boxes by one of four centrality measures. They are
computed in this page from the arrows alone, with the definitions Negro et al. use:
- LRC (local reaching centrality): the share of the other claims that depend on this one, directly or through a chain of arrows.
- HC (harmonic centrality): the same count, with each dependent claim weighted by 1/distance.
- BC (betweenness centrality): the share of shortest paths between other claims that pass through this one.
- PR (PageRank, damping 0.85): weight inherited from the claims above. Negro et al. use it to find high-stakes predictions, because a failed prediction with high PageRank reflects on influential claims upstream.
What the measures show
- Reach. After the core, the claim with the widest reach is B1, the architecture of two kinds of model (world and self) on two levels (stored and generated): 77% of the other claims depend on it. Next come B2, the split between stored implicit models and generated explicit ones (46%), and B4, self-referential closure (41%). The architecture claim reaches further than closure does.
- Bridges. B2 has the highest betweenness (0.016). It connects the principles to the cluster of anaesthesia and working-memory predictions and to the cluster about how permeable the explicit models are to the implicit ones. Evidence against the implicit/explicit dissociation (E4, E6, F4) would therefore cut more paths through the theory than evidence against any other single claim.
- Highest PageRank: E3, lucid-dream onset (0.031). E3 predicts that verified onset of lucidity shows phase-transition dynamics that begin in self-model networks. It ranks first because it rests on D7 (REM sleep approaches criticality; lucid onset is a crossing), which has three parents, one of them C12, the criticality signature. A null result on lucid onset refutes the criticality signature (C12 and D7), not the theory. In FMT, near-criticality is a signature that free modelling usually leaves in neural tissue, neither necessary nor sufficient for consciousness, so criticality data cannot disconfirm the core principle A1 either way. PageRank counts parents and cannot represent that distinction.
- Next predictions by PageRank (D7, an intermediate claim, sits between them at 0.025): F2, the qualia-structure dissociation (0.028); F1, the content of psychedelic ego dissolution (0.025); E10, the white-box trace (0.025); E2, the gradient between alters in dissociative identity disorder (0.024).
- A low rank that does not mean low stakes. F6, the novel-input probe, tests principle A1 directly, yet ranks 46th of 62 by PageRank. A1 is a root, so it passes on only the base PageRank share, split over its five children.
- Claims with no empirical consequence on the map. C6 (substrate independence) leads only to extrapolations to machines and animals (D8, D9, D11), from which no prediction follows. B3, virtual qualia, reaches a single prediction (F2, through C10). Negro et al. report the same pattern in the map of Integrated Information Theory.
Compare the theories
One row per map. Select a name to open that map next to another.
| Map | Claims | Arrows | Levels | LRC median / 90th | HC median / 90th | BC median / 90th | PR median / 90th | Top 3 by PageRank |
|---|---|---|---|---|---|---|---|---|
| 64 | 101 | 6 | 0.016 / 0.159 | 0.016 / 0.119 | 0.0005 / 0.0070 | 0.014 / 0.022 | E3 0.030; F2 0.027; D7 0.025 | |
| 44 | 58 | 6 | 0.023 / 0.163 | 0.023 / 0.116 | 0.0011 / 0.0055 | 0.019 / 0.030 | F3 0.084; F4 0.054; F5 0.041 | |
| 40 | 53 | 6 | 0.026 / 0.282 | 0.026 / 0.175 | 0.0007 / 0.0118 | 0.023 / 0.037 | E10 0.049; E2 0.045; E8 0.038 | |
| 37 | 53 | 6 | 0.056 / 0.222 | 0.042 / 0.146 | 0.0024 / 0.0090 | 0.021 / 0.044 | F8 0.063; F3 0.052; F7 0.051 | |
| 30 | 40 | 6 | 0.034 / 0.241 | 0.034 / 0.190 | 0.0000 / 0.0086 | 0.029 / 0.046 | F4 0.069; F2 0.055; D1 0.047 | |
| 45 | 60 | 6 | 0.023 / 0.182 | 0.023 / 0.140 | 0.0005 / 0.0042 | 0.019 / 0.037 | D4 0.046; D8 0.044; C2 0.039 |
- Values are computed on this page's copies of the maps with the definitions above. For the four maps of Negro et al. they reproduce 64 of the 66 values printed in the preprint's results to three decimals; the two exceptions are the harmonic centrality of GNWT's B4 and C2 (printed 0.260 and 0.130, the file gives 0.264 and 0.132).
- Median and 90th percentile use the nearest-rank convention. In every map 30 to 50% of the claims have no arrows leaving them, and most others ground only a claim or two, so the medians of LRC, HC and BC are small.
- The maps differ in size and in who drew them. The four maps of Negro et al. were drawn by them and revised with the theories' proponents; the FMT map was drawn by the theory's author following their method, and the PP map by us from PP's published sources, not reviewed by its proponents. PageRank is spread over all claims of a map, so it falls as a map grows; compare ranks within a map rather than PageRank values across maps of different size.
- The authors' files mark no difference between solid and dashed arrows. Here an arrow is drawn dashed when both its claims sit on the same level, the rule the preprint gives.
Quality of the maps
These numbers measure how well a map is articulated and how exposed it is to test, not whether the theory is true.
| Map | Predictions | Reach | Fruitfulness | Exposure | Robustness | Scope | Extrapolations (excluded) | |||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| share of theoretical claims | idle claims | predictions per theoretical claim, mean / median | per core claim | mean / median | predictions resting on the core | ≥2 parents | ≥2 disjoint paths from the core | substrate-general | cognitive / brain / extrapolation | |||
| FMT | 26 | 97% (33 of 34) | C6 | 5.3 / 2.0 | 25.3 | 20% / 22% | 100% | 38% | 35% | 34% | 23% / 36% / 7% | D8, D9, D10, D11 |
| GNWT | 22 | 100% (21 of 21) | none | 5.1 / 4.0 | 22.0 | 23% / 21% | 100% | 45% | 36% | 0% | 52% / 45% / 2% | D11 |
| IIT | 18 | 95% (20 of 21) | B3 | 5.0 / 3.0 | 18.0 | 28% / 31% | 100% | 39% | 33% | 33% | 5% / 59% / 3% | F3 |
| PRM | 19 | 100% (17 of 17) | none | 5.0 / 3.0 | 19.0 | 26% / 24% | 100% | 47% | 42% | 0% | 65% / 32% / 3% | D1 |
| RPT | 15 | 86% (12 of 14) | C3, C4 | 4.0 / 3.0 | 15.0 | 27% / 21% | 100% | 27% | 13% | 3% | 60% / 33% / 3% | D2 |
| PP | 17 | 85% (22 of 26) | B6, B9, C3, E3 | 2.9 / 1.0 | 8.3 | 17% / 19% | 100% | 35% | 29% | 10% | 56% / 31% / 4% | D9, D10 |
- Which claims count as predictions
- The level of a claim decides it, as in Negro et al. (preprint p. 7): levels 4 and 5 in the layout for substrate-general theories (FMT, IIT), levels 3 and 5 in the layout for cognitive theories (GNWT, PRM, RPT). Every other claim, the core included, is a theoretical claim. Arrows point from a claim to the claim that holds in virtue of it, so the claims above a prediction are the ones a failed prediction puts at risk. One rule overrides the levels in every map: a claim of the extrapolation class (alone or in a two-class box) counts as neither theoretical claim nor prediction and is listed in the last column as one of the extrapolations (excluded), because extrapolations to animals or machines are neither part of the theory's derivation base nor currently testable predictions in either layout. Without the rule the layouts would treat them unequally: FMT's extrapolations sit at level 3, which its layout counts as theory, while each other map's level-3 extrapolation would count as a prediction. Arrows still run through an excluded claim when the measures follow a chain. Scope still counts the extrapolation class.
- Reach
- The share of theoretical claims from which a chain of arrows reaches at least one prediction; the others are listed as idle claims. A claim with no route to a prediction cannot be tested through the map; these are the claims Negro et al. call "not particularly productive ... in terms of empirical consequences" (pp. 18–19).
- Fruitfulness
- The number of distinct predictions reachable from a theoretical claim, as mean / median over all theoretical claims, and averaged over the core claims (FMT has three, the other maps one). Kuhn counted fruitfulness, a theory's yield of new consequences, among the criteria for choosing between theories; Negro et al. read their maps for "which claims yield more downstream claims" (p. 27).
- Exposure
- For each prediction, the share of the map's theoretical claims among the claims it holds in virtue of, directly or through a chain, given as mean / median over the predictions; and the share of predictions that rest on a core claim. Popper made exposure to refutation the mark of an empirical theory. Lakatos described how auxiliary claims absorb a refutation and shield the core (preprint p. 3); exposure counts how much of the theory stands behind each test.
- Robustness
- The share of predictions with two or more parent claims, and the share reached from the core along two or more paths that share no arrow (counted as a maximum flow). Negro et al. note whether a claim is "supported by a single or multiple 'parent' claims" (p. 27). A prediction derived along two independent routes does not hang on a single auxiliary claim, so a failed test is harder to blame on that one claim alone.
- Scope
- The share of substrate-general claims, and of cognitive / brain-specific / extrapolation claims, among the claims outside the core, in the four classes of Negro et al. (p. 7). A claim drawn in two classes (their "hybrid claims", p. 24) counts half in each; the core is left out because the authors' files draw it grey, without a class. Kuhn's scope is the range of phenomena a theory speaks to; the substrate-general share says how much of a theory is stated without assuming a brain.
Experimental record in the ConTraSt database
ConTraSt (Yaron, Melloni, Pitts & Mudrik, 2022, Nature Human Behaviour 6, 593–604, https://doi.org/10.1038/s41562-021-01284-5) collects experiments whose authors interpreted their findings in light of a theory of consciousness, and records whether they read the findings as supporting or challenging it. ConTraSt groups the theories into four families. Counts of experiments per family, read from the database's public interface at https://contrastdb.tau.ac.il/, retrieved 2026-10-19 (the 2022 paper covered 412 experiments; the database has grown since):
| ConTraSt family | Map on this page | Supporting | Challenging |
|---|---|---|---|
| Global Workspace | GNWT | 244 | 57 |
| Integrated Information | IIT | 129 | 10 |
| Higher Order | PRM (family only) | 14 | 11 |
| First Order & Predictive Processing | RPT (family only) | 145 | 24 |
| none | FMT | not yet tested in ConTraSt | |
ConTraSt does not list PRM or RPT separately: PRM belongs to the higher-order family (Negro et al., preprint p. 5), RPT to the first-order family, so their rows count every theory in the family. An experiment interpreted in light of several theories counts once for each. The counts record how authors read their own results, most of them post hoc rather than as tests of critical predictions (Yaron et al., 2022), and they are not weighted by how central the tested prediction is to the theory. Data: ConTraSt, licensed CC BY-SA 4.0; the counts are quoted here, no data file is copied.
All claims
Method and credit
Method: Negro, N., Elbaz, E., Fischer, S., Schreiber, M., & Mudrik, L. (2026). Prediction maps as a gateway to better understand and test scientific theories. PsyArXiv. https://doi.org/10.31234/osf.io/kqr87_v1 (licence CC BY 4.0).
The authors built an app for drawing and analysing prediction maps:
https://xk29y8.csb.app. The file fmt-prediction-map.platform.json below loads in that
app through its Load Map button. This FMT map was drawn by the theory's author following their
method; Negro et al. did not build or review it, nor this page.
Maps of the other theories
- Global Neuronal Workspace Theory (GNWT): map by Niccolò Negro, Eden Elbaz, Shai Fischer, Maor Schreiber and Liad Mudrik (2026), file JSON MAPS/GNW_network_2026-03-12.json in the OSF project https://osf.io/9xug8/, retrieved 2026-10-19. Licensed under CC BY 4.0. Modified for display: claim IDs and levels read from the label prefixes, the prefix removed from the claim text, boxes re-laid out in one column per level, same-level arrows drawn dashed, centrality recomputed. The claims and arrows are unchanged; the original file is reproduced unchanged under "map". Class colours added: the file leaves 43 of its 44 claims uncoloured (#ffffff). Each claim's class (cognitive, brain-specific, both, or extrapolation) was transcribed from the preprint's Figure 3, where every box is coloured, and is listed under "added_classes" (table: gnwt-classes.md); the page draws those classes in the shades of the authors' IIT and RPT files. The colours under "map" are the file's own, unchanged. Class table:
gnwt-classes.md. Download:negro2026-gnwt.json(the authors' file unchanged, with its provenance). - Integrated Information Theory (IIT): map by Niccolò Negro, Eden Elbaz, Shai Fischer, Maor Schreiber and Liad Mudrik (2026), file JSON MAPS/IIT_network_2026-03-12.json in the OSF project https://osf.io/9xug8/, retrieved 2026-10-19. Licensed under CC BY 4.0. Modified for display: claim IDs and levels read from the label prefixes, the prefix removed from the claim text, boxes re-laid out in one column per level, same-level arrows drawn dashed, centrality recomputed. The claims and arrows are unchanged; the original file is reproduced unchanged under "map". Colours as in the file. Download:
negro2026-iit.json(the authors' file unchanged, with its provenance). - Perceptual Reality Monitoring (PRM): map by Niccolò Negro, Eden Elbaz, Shai Fischer, Maor Schreiber and Liad Mudrik (2026), file JSON MAPS/PRM_network_2026-03-12.json in the OSF project https://osf.io/9xug8/, retrieved 2026-10-19. Licensed under CC BY 4.0. Modified for display: claim IDs and levels read from the label prefixes, the prefix removed from the claim text, boxes re-laid out in one column per level, same-level arrows drawn dashed, centrality recomputed. The claims and arrows are unchanged; the original file is reproduced unchanged under "map". Colours as in the file. Download:
negro2026-prm.json(the authors' file unchanged, with its provenance). - Recurrent Processing Theory (RPT): map by Niccolò Negro, Eden Elbaz, Shai Fischer, Maor Schreiber and Liad Mudrik (2026), file JSON MAPS/RPT_network_2026-03-12.json in the OSF project https://osf.io/9xug8/, retrieved 2026-10-19. Licensed under CC BY 4.0. Modified for display: claim IDs and levels read from the label prefixes, the prefix removed from the claim text, boxes re-laid out in one column per level, same-level arrows drawn dashed, centrality recomputed. The claims and arrows are unchanged; the original file is reproduced unchanged under "map". Colours as in the file. Download:
negro2026-rpt.json(the authors' file unchanged, with its provenance). - Predictive Processing (PP): Drawn by Matthias Gruber (aIware) from the published sources of PP's proponents, following the method of Negro et al. (2026). Not drawn or reviewed by PP's proponents or by Negro et al. Treated here as a theory of conscious experience. Every claim cites its sources, listed under the claim. Download:
pp-prediction-map.json.
The maps are provided as is, without warranties, as stated in the licence. Their use here does not imply that the authors endorse this page or the FMT map.
Downloads
fmt-prediction-map.json: the map with every field (level, class, status, uniqueness, FMT section, references, arrow justifications).fmt-prediction-map.platform.json: the same map in the format of the Negro et al. app. That format has no fields for level, status or references, so it carries the claims, colours and arrows.- The four maps of Negro et al. and our map of Predictive Processing: see their entries above.
cross-theory-links.json: the links between claims of different maps, each with its relation and reason (our judgement).
The theory
Gruber, M. (2026). The Four-Model Theory of Consciousness. Zenodo. https://doi.org/10.5281/zenodo.18669891 (concept DOI, resolves to the latest version). An introduction in short articles: the FMT wiki at https://fmt.matthiasgruber.com/, with a page on this map at https://fmt.matthiasgruber.com/predictions/prediction-map/.