CALLIOPE · BCOM

Most whole-brain models (WBMs) in current use assume a single canonical column — a uniform local microcircuit tiled over the cortical mantle, coupled by a structural connectivity matrix. That assumption is convenient but wrong: the human cortex is a differentiated patchwork of cytoarchitectonic types, from the three- to four-layered allocortex to the hyper- granular koniocortex, each with distinct cellular density, myelination, laminar organization, and characteristic connectivity. A next-generation WBM should encode this heterogeneity as part of its meta-architecture, not abstract it away. In this working paper we propose CORTYPEX — an open, fsaverage-based atlas of cortical types with four complementary products: (i) a topological map of the concentric and island organization of cortical differentia- tion grades together with their subcortical coprocessor loops; (ii) a reference table linking each type to its expected computational role and EEG/SEEG spectral signature; (iii) a short-range adjacency matrix extracted from the topographical (“political”) boundaries of the types; and (iv) long-range connectivity rules derived from dMRI tractography as probability density func- tions over type pairs. We then cast CORTYPEX as a substrate for the Kolmogorov-Theory (KT) agent: cortical rings realize the Modeling Engine, allocortical–basal-ganglia loops realize the Objective Function, and agranular–cerebellar loops realize the Planning Engine, with a dedicated audit of laminar and subcortical connectivity labels. The goal is a biologically disciplined meta-architecture that groups like the BCOM Archeon circuits-to-agent pipeline and the WBM Platform for Depression can plug into directly. CORTYPEX will be released as an open repository accompanying this paper.

CORTYPEX proposes a biologically honest map of cortical diversity as the foundation for the next generation of whole-brain models — replacing the fiction of a uniform cortex with a typed, subcortically-wired substrate.

Every current whole-brain model (WBM) makes the same quiet lie: it tiles the same generic microcircuit across every cortical region and connects the tiles with a structural wiring matrix. The math is tractable, but the biology is wrong. The human cortex runs a steep gradient from the primitive three-layer allocortex (hippocampus, olfactory cortex) all the way out to the hyper-granular koniocortex of V1, A1, and S1 — seven distinguishable types with different cell densities, myelination, laminar organization, dominant rhythms, and connectivity rules. Treating primary visual cortex and subgenual cingulate as the same "chip" isn't a harmless simplification; it erases the structure that makes cortex computationally interesting in the first place.

CORTYPEX is a four-part atlas designed to fix this. It provides: (1) a topological map of how the seven types are arranged — concentrically from the allocortical core outward, with koniocortical and unimodal-association "islands" embedded in a mid-grade matrix, obeying a strict rule that non-adjacent grades cannot touch; (2) a reference table linking each type to its expected computational role, myelination signature, and EEG/SEEG spectral fingerprint — priors for parameterizing type-specific neural mass models; (3) a short-range adjacency matrix encoding which parcels physically border which, derived from the fsaverage cortical surface; and (4) long-range connectivity rules from diffusion MRI tractography, expressed as probability distributions over connection strength for each pair of cortical types. A notable empirical regularity the paper highlights: older-to-older (allocortical) tracts are more curved, newer-to-newer (isocortical) tracts are straighter — a developmental fingerprint readable in the tractogram.

The paper then maps this anatomy onto the Kolmogorov-Theory (KT) agent framework, which divides cognition into a Modeling Engine (ME), Objective Function (OF), and Planning Engine (PE). The mapping is clean: the sensory hierarchy from koniocortex up through eulaminate rings is the ME, compressing inputs into progressively abstract representations via predictive-coding feedforward/feedback loops; the allocortex plus basal-ganglia-thalamo-cortical loops are the OF, implementing value gating and action selection through disinhibitory dynamics; and the dysgranular/agranular motor cortex plus cerebellar loops are the PE, proposing and refining actions via efference-copy forward models. The paper also does careful anatomical housekeeping — correcting a prior diagram that mislabeled basal-ganglia-to-thalamus edges as "valence assessment" (it's selection by disinhibition) and incorrectly drew direct cortex-to-cerebellum connections (the real route goes cortex → pons → cerebellum → motor thalamus → cortex).

The source is honest that CORTYPEX is a proposal and a substrate, not a finished model. The fsaverage atlas, adjacency matrix, type-pair connectivity distributions, and LaNMM parameter priors are promised as open deliverables; the figures in the paper are placeholders. The immediate downstream targets are BCOM's Archeon circuits-to-agent pipeline and a WBM platform for depression, where the OF-dysfunction framing maps directly onto limbic and basal-ganglia circuitry. The validation benchmark — showing that typed priors outperform uniform-column priors on a biomarker-matching task — is listed as a next deliverable, not a result in hand.