IBM-1

↓
the real anatomy crawling16.0 s · 0.920 m · worst joint excursion 11.5°
1,047 bone and muscle surfacesa full measured gait cycle
117 tissue force elementsthree arms — the middle is a negative result
cortex in the loop0.17 mm COM drift under a 5 N push
video prediction, and it does not work−0.25 skill against persistence
cortex in the loop, beside cortex severedupright at 1.019 m · fallen at 0.590 m
one step, with skinfoot airborne at 0.14% body weight
the 22-segment scaffoldwhat the anatomy is posed from
cortex in the loop, beside cortex severed
cortex in the loop, beside cortex severed1.019 m upright · 0.590 m fallen
the acquired atlas
the acquired atlas2,229 surfaces, ~350 MB, ~28 s to load
multisystem physiology
multisystem physiologythe body state the brain reads
cardiopulmonary
cardiopulmonarythe loop that does not stop
skin bioelectricity
skin bioelectricity1,326 patches, 100.00% of the exterior
the materialization ring
the materialization ringevery unmeasured path reads “not measured”
what a materialization contributes
what a materialization contributesand what it does not
asking the brain what it is seeing
asking the brain what it is seeing63.5% top-1, 127× chance

The Implicit Brain Model (IBM-1) is a multimodal foundation model of whole-brain state and dynamics which can be lazily materialized into explicit maps for training and applications. By learning a perspective-agnostic representation of the brain rather than one tied to a particular measurement modality, IBM-1 can train end-to-end across heterogeneous data sources that would otherwise be difficult or impossible to combine.

§1design
fieldsF = { x(q) : q ∈ Ω }

state variables on a support. a component is a physical quantity, never an encoding. supports are not shared: the cortical sheet, the vascular tree and a sensor array are different domains.

anatomya(q) ∈ [0,1]ᵏ

soft membership of a position in named partitions. an atlas is evidence about where things are, not a coordinate system.

topologiesT(i, j)

which state variables may interact, through what. all spatial structure lives here; there is no universal spatial operator.

processesP = (I, O, T, f, θ)

dynamics over a topology. (f, θ) is implementation; implementations compete for a process rather than redefining it.

fig. 1 the mne sample subject as drawn here: pial surface at oct5 coloured by its parcellation, aseg structures as volumes, its outer skin as a ghost, its digitised electrodes.

M = materialize(R, r, B, F, A, T, P)

an explicit model names regions R, a resolution r(q) and a bandwidth B(q); dependency tracing instantiates only what is reachable from the target. the same field can sit at 1 mm across the brain and 50 µm around an electrode at once. the forty entries in the library differ only in R, r and B, and each carries provenance: which implementation ran, which parameters evidence moved, and where a process ran outside its regime.

fig. 2 eeg_forward read off its request: scalp EEG at sixty contacts in, head-volume and contact potentials out. what lights up is what the trace instantiates.
§2research
claimmeasuredoutcome
the shared cortex is load-bearingbypass +324% loss, frozen +180%, association zeroed +180%. on visual retrieval bypass costs 37 of 63 points; on balance, severing drops the body at 2.90 s. load-bearing is not the same as better than a purpose-built baseline — see §2bheld
cross-modal association is learnedoccipito-temporal edges at 4.4× the random-pair baseline — and severing those exact edges costs −0.06%. magnitude is not contributionwithdrawn
band-limited fusion is exactreal EEG moved the posterior 2.03 prior sd in band; out of band, exactly 0. delays as phase ramps agree to 4.4e-16held
the thalamocortical resonance is alpha13.45 Hz, a spindle; held-out +5503 nats/night, p = 7.2e-11prior moved
the slow oscillation sits at the band edge1.000 ± 0.296 Hz across 8 scored N3 nights; τadapt = 0.12 s reproduces itfitted
the paired MEG encoder beats baselineafter the target units were fixed: skill −0.07 to +0.14. the +0.90 to +0.99 reported before measured the target scaleat chance
a positive association kernel is enougheight applications are eight rounds of averaging: rank 1.57 → 1.01, loss diverging. signed fixed itfailed
the neurovascular chain predicts BOLDr = −0.11, permutation p = 0.38failed
Grubb's law holdsslope 0.066 against a declared 0.38failed
MEG resolves finer sources than EEG49.8% vs 63.0% per channel, p = 0.0033failed
the capillary bed carries most of the resistance6–27% of dissipation; 85% of the lengthwrong
tractography is reliable edge by edgerecovery 0.456, FDR 0.725; thresholding does not helpno
§2bstatus

Where the programme actually stands, updated 9 September 2026. Every figure below is measured against an explicit baseline and names the set it was measured on; the two THINGS-EEG2 splits give different orderings and are never quoted interchangeably.

what runsmeasuredagainst
one implicit kernel, publishedfused from 34 checkpoints, Procrustes-aligned, weighted by measured transfer. recovers 88.7% of a task-specific kernel on the task that kernel was trained for0% for random
16 materializations, one 3.84M kerneltrained simultaneously to step 32,000, consolidated every 500 steps, replica drift cosine 0.998–1.000, none collapsed. visual_eeg 30.4% top-1, optic_nerve 22.8%, ten per-subject terms mean 15.4%. two of the sixteen — video and audio_visual — sit at negative skill against persistence and are reported as failures, not omitted0.5% chance
image → cortex → EEG retrieval63.5% top-1 on the designated test set, 200 images, median rank 1127× chance
a declared optic nerve into occipital cortexthree retinal populations at 2.5 / 4.2 / 8.3 ms, resolved. displacing the port to a random region of equal size collapses retrieval to 0.44%8.69% in place
the substrate resolution has an interior optimumcontrolled sweep, resolution the only variable: 128 → 19.4%, 512 → 23.0%, 2048 → 23.0%, 8192 → 21.7%both tails worse
the sheet delivers between disjoint regionsimage → occipital → sheet → a precentral-only readout (ports verified disjoint) reaches 48.4× chance on real THINGS-EEG2, against a severed sheet at 1.00× — severed has across-image variance of exactly zero and effective rank 1.0. a positive control that permutes only the readout columns, which cannot change the answer, reads 48.6×severed = chance
cortical dynamics hold a body upright250 exchanges, five seconds, 5 N pelvis push: 0.17 mm COM displacement. severed, the body is on the floor at 2.90 s at 499 mmheld

And what does not work, stated with the same weight:

claimwhat the control saidoutcome
the cortex beats a dynamics-free encodermatched selection on the designated test set: control 66.5% against the cortex model's 63.5%. indistinguishable, and the direction is gonewithdrawn
the kernel's learned content carries perception between regionsa kernel with its site rows permuted reads 50.8× against the trained kernel's 48.4× — indistinguishable, with amplitude, effective rank, train loss and train top-1 all preserved. transport is a property of the graph and the weight statistics, not of what was learnedwithdrawn
the kernel's learned content drives movementa kernel with its site rows permuted recovers the same push. the dynamics are load-bearing; what they learned from vision and audio is notwithdrawn
next-frame video training shapes cortical wiringpairing each frame with a random frame still transfers 83–86% of the 91.5%. only ~5 points depend on predictionwithdrawn
sharing one kernel helps the terms that share itone subject trained solo for the same own-steps reaches 17.4% against the shared run's 15.4%neutral
video continuation predicts motion25% worse than emitting the previous frame unchanged. the clips looked sharp because at horizon 8 frame t+8 resembles frame twithdrawn
the body walksthree independent CMA searches each produce one genuine step — swing foot unloaded below 2% of body weight, 3.5–6.4 cm clearance, 4.6–10.3 cm advance, landing on measured contact — and then fall. the pelvis travels 0.2–5.8 mm while the foot advances 46–103 mm: the body steps in place, so the landing leg brakes a centre of mass that never moved, and the COM collapses 18–20 cmone step
the cortex can route sensation to motor cortex0.03–0.07% of the driven signal reaches the disjoint motor region, and the ratio scales linearly with association gain rather than compoundingstructural

Fifty-one claims have been withdrawn. Not one was a modelling error. Every one was a quantity computed correctly and then compared against the wrong thing — the wrong population, the wrong units, the wrong split, the wrong point in time, or no baseline at all. The worst flattered itself by 280× while never beating a zero baseline; the largest was an image–EEG pairing that was 99.94% wrong and produced three hours of confident negative results about a corpus that was fine. The ledger keeps every one, with the check that caught it.

§3data
§4releases
forward models
  • av_meg_predictor(WIP)
  • av_predictor(WIP)
  • meg_encoder(WIP)
  • visual_evoked(WIP)
  • eeg_forward(WIP)
  • meg_forward(WIP)
  • bold_forward(WIP)
  • fnirs_forward(WIP)
  • ecog_forward(WIP)
decoding
  • meg_to_text(WIP)
  • speech_envelope(WIP)
  • eeg_to_image(WIP)
  • inner_speech(WIP)
stimulation
  • tms_response(WIP)
  • tes_response(WIP)
  • tfus_response(WIP)
  • dbs_response(WIP)
  • tms_eeg_tep(WIP)
state and disorder
  • sleep_dynamics(WIP)
  • anesthesia(WIP)
  • seizure_propagation(WIP)
  • pharmaco(WIP)
the embodied loop
  • invasive_bci(WIP)
  • handwriting_bci(WIP)
  • motor_mapping(WIP)
surrogates
  • macro_surrogate(WIP)
  • connectome_gnn(WIP)
  • encoding_model(WIP)
  • resting_state_fc(WIP)