Transfinite Anatomy Theory

Medical AI increasingly builds hierarchical structure into its models. Hierarchical Markov Blankets describe biological organization as nested conditional-independence structures. The free-energy principle treats boundaries as inference interfaces. Digital-twin programs reconstruct physiology scale by scale. They disagree on much, but share a thesis: disease is a phenomenon over biological hierarchies, not merely a pattern in observation space.

The vocabulary used to describe the boundaries that populate those hierarchies is narrower. In practice a boundary is treated either as a conditional-independence separator or as a sufficient-statistic interface — in both cases, a statistical factorization device.

The blood–brain barrier admits glucose through stereospecific carrier-mediated transport while excluding plasma proteins of comparable molecular weight, and actively effluxes xenobiotics. The immune system distinguishes self from non-self through a receptor repertoire shaped by exposure history.

What makes these boundaries healthy is not that they separate. It is what they choose to let through.

Separation is necessary; it is not sufficient. Two boundaries with identical separation properties can hold opposite selectivity profiles — one transmitting what the organism needs, the other transmitting what harms it. Conditional independence cannot tell them apart. For diagnosis and intervention, that distinction is the whole point.

Take a hierarchical unit with external state X, internal state Y, boundary state B, and a task variable T encoding the role the unit plays in the larger hierarchy. A selective boundary maximizes I(B; T) subject to I(B; T^⊥) ≤ β — with conditional independence entering as a standing constraint rather than as the objective.

Among all boundaries that separate, a selective one prefers those that keep what matters and discard what does not.

Under trivial task structure and unbounded β, the formulation reduces exactly to the Hierarchical Markov Blanket condition. TAT does not invalidate HMB-based models. It locates them — as the selectivity-neutral region of boundary space, a region genuinely occupied by interfaces whose role really is exhausted by separation. Any analysis valid for HMB stays valid there.

Closing: task-relevant transmission is lost.

Leaking: task-irrelevant information exceeds β while separation holds.

Structural breach: separation itself fails, as influence bypasses the regulated channel.

Used in its compositional sense — trans- a finite bound. Any unit may be resolved into a finer-scale sub-hierarchy, with practical depth bounded by measurement resolution and modeling purpose rather than by a ceiling internal to the theory.

This is not fractality. What recurs down the hierarchy is the partition pattern, not the dynamical content: cardiac mechanics and molecular kinetics obey different equations, with different state spaces and time constants.

A sudden loss of function at one node, a slow drift of setpoints across many, a breached selectivity profile, a collapse of regulatory complexity — these are mathematically distinct phenomena. A framework that collapses them into one anomaly score discards information that is clinically essential.

TAT proposes a vocabulary of six primitives, each a distinct causal-graph alteration paired with an operational signature observable in multivariate physiological time series.

The vocabulary is deliberately neither mutually exclusive nor exhaustive; most real diseases instantiate several primitives concurrently or in sequence. Chronic heart failure typically runs P2 → P5 → P6. Sepsis often opens with P4, broadcasts into P1, and terminates in P3 or P6.

Cross-specialty commensurability, first: a cardiologist's compensated heart failure (P2+P5) and an endocrinologist's insulin resistance (P2) become structurally comparable. And an intervention class implied by the active primitive: a P4 breach calls for restoring interface selectivity, a P5 shift for setpoint correction.

The paper develops boundary breach in detail with the BBB as the boundary of study, yielding three predictions that go beyond biomarker-threshold framing.

Q1 — Leakage carries differential information. A boundary that has lost selectivity should leak preferentially in the task-irrelevant direction. In early or mild disruption, the cross-barrier information increase should be carried disproportionately by T^⊥-aligned features — peripheral cytokines, immune cells — rather than regulated metabolic substrates. A barrier that simply “opens” predicts symmetric increase; TAT predicts asymmetry.

Q2 — Selectivity precedes overt biomarker change. The ratio I(B;T)/I(B;T^⊥) can fall before Q_Alb crosses clinical threshold. Multivariate selectivity estimators should therefore detect dysfunction earlier than univariate permeability biomarkers in gradual cases.

Q3 — Recovery is hysteretic. After an intervention normalizes a univariate permeability biomarker, the selectivity ratio should recover with a lag and along a different path — a hysteresis loop in the (biomarker, selectivity) plane. Testable in relapsing–remitting courses such as multiple sclerosis.

The same structure transfers, with substitutions, to host–pathogen breach in sepsis, intestinal barrier dysfunction, autoimmune self/non-self breach, and tissue barriers in metastasis.

The equations coupling levels are not specified. The formalization presumes a task variable T whose choice is itself a clinical commitment. The primitives are anchored at organ-system scale.

Only four kinds of boundary information are formalized here — separation, sufficiency, task-relevant transmission, task-irrelevant suppression — out of a dozen or more that a biological boundary plausibly carries. Channel capacity, state-dependent permeability, hysteresis, identity. Cataloguing the rest is the framework's principal open problem.

We do not claim the six primitives are complete; we expect extension. What we commit to is a form — every primitive is a distinct causal-graph alteration paired with an operational signature.

TAT is falsified not by a disease the current six miss, which merely extends the vocabulary, but by a clinically important dysfunction that cannot be cast in this form at all.

Empirical validation, formalization of accommodation dynamics across levels, and integration with neighbouring frameworks remain open.