Relevance as Collapse: Relevance Realization as the Endogenous Measurement of a Viable Section

A problem with no model

John Vervaeke's relevance realization (RR) names the operation an intelligent system performs when it picks out, from a combinatorially vast field of possible considerations, the small set that matters — without first enumerating the field. It is the cognitive face of the frame problem: you cannot search for what is relevant, because deciding what to search is the relevance question itself. Vervaeke's honest position is that we have no mechanistic model for RR, and that a model would be a genuine landmark — plausibly entangled with the question of how life begins, since a self-bootstrapping relevance-selector is close to the minimal definition of an agent.

A recorded seminar dialogue on RR offers a candidate mechanism worth taking seriously. The proposal: relevance does not begin at the conscious level. It begins sub-threshold. An external stimulus partially excites neural assemblies; activation cascades through the pre-built topology, simple patterns exciting more complex ones, those exciting patterns already laid down by experience; the assembly that accumulates the most activation crosses the firing threshold (~−55 mV) and wins. That win, the proposal says, is relevance realization — and predictive processing and RR fall out of one and the same mechanism.

This is a real intuition and the right altitude. But as stated it does not yet dissolve the problem. It relocates it.

The circularity

The proposal defines relevance as what wins, and winning as what is most activated by the pre-weighted topology. But the weighting of the topology is exactly the deposited history of prior relevance judgments. So the account reads: the relevant is what the already-relevance-weighted substrate makes most active. The homunculus is not removed; it is dissolved into the synaptic weights and left unexplained there. A sharp objection in the same dialogue — that any such theory owes a testable account of every jump from cellular to multicellular to population to structure to cognition, and that "it emerges" is not such an account — is the correct pressure. Without a selection criterion that is not itself a stored lookup, "the most-activated pattern wins" is a description of the regress, not an exit from it.

Draken's repair is to supply the missing criterion, and the criterion is the core axiom.

Collapse, not lookup

The win is not a popularity contest over stored weights. It is a constrained extremization. Among the candidate assemblies admissible given the input, the one selected is the one that minimizes total system entropy subject to the viability constraint:

$$ \boxed{\ \diamond\ \min_{s}\ S_{\text{sys}}(s)\quad\text{s.t.}\quad \frac{dH}{dt}\ge 0\ \diamond\ } $$

$$ s^{*} \;=\; \arg\min_{s\,\in\,\mathcal{A}}\ S_{\text{sys}}(s)\quad\text{s.t.}\quad \frac{dH}{dt}\ge 0 . $$

Here $\mathcal{A}$ is the admissible set — the candidate relevance-assignments consistent with the present input — and $s^{*}$ is the selected one. The decisive move: this criterion is global and substrate-general. It does not consult a table; it extremizes a functional. Relevance is therefore not stored — it is the extremum. The topology is not the definition of relevance; it is merely the substrate over which the extremization runs. The regress breaks because the selector is a rule on states, not a state.

And a rule that selects one outcome from many admissible ones, by extremizing an objective under a constraint, has a precise structural name. It is a collapse.

The sub-threshold superposition

Model the relevance field sheaf-theoretically. Let $B$ be the context (the stimulus, the prompt, the present sensory cut). Over $B$ sits a sheaf $\mathcal{R}$ of relevance-sections: a section $s \in \Gamma(U)$ over an open $U \subseteq B$ is a locally coherent assignment of what-matters across that context. The sub-threshold neural field is then not yet a decision but a superposition — a weighted state $\Psi$ over the admissible sections $\{s_i\}\subseteq\mathcal{A}$, none yet selected:

$$ \Psi \;=\; \sum_{i} a_i\, s_i, \qquad s_i \in \mathcal{A}, $$

where the $a_i$ are the partial (sub-threshold) activations. The dialogue's own examples are this superposition made vivid: hearing a growl, the field holds dog / husky / wolf simultaneously, each partially lit; a child's "dog" for a cat is the not-yet-collapsed state where the discriminating dimensions have insufficient amplitude. Threshold crossing is the collapse operator $\mathcal{C}$ resolving $\Psi$ to a single section:

$$ \mathcal{C}:\ \Psi \;\longmapsto\; s^{*} \;=\; \arg\min_{s\in\mathcal{A}} S_{\text{sys}}(s)\ \text{ s.t. }\ \frac{dH}{dt}\ge 0 . $$

The biological $-55\,\text{mV}$ is not the essence of $\mathcal{C}$; it is one substrate realization of it. This is the place to be explicit and disown an easy misreading: this is measurement-theoretic collapse, not a claim that cognition is physically quantum. No microtubules, no Orch-OR. The isomorphism is at the level of structure — a distribution over admissible outcomes resolved to one by a selection rule — which is exactly as true of a thermostat or a classifier as of a wavefunction. What the quantum vocabulary buys is precision about the type of operation RR is: not retrieval, but the irreversible resolution of a held multiplicity. Relevance realization $\equiv \mathcal{C}(\Psi)$.

Three faces of one morphism

Once RR is collapse, it takes its place in a diagram this corpus has been assembling all along. There are three operations on the relevance sheaf, and they are the same morphism read three ways.

$\rho$ — restriction (down). Anubis. The psychostasia projects the high-dimensional record of a life onto a single weighted residual against Ma'at, $\rho:\mathbb{R}^4\to\mathbb{R}^1$, forgetting almost everything and keeping a verdict. Restriction destroys information. Judgment is $\rho$.

$\iota$ — section (up). The typewriter poet. "All you need to offer is a prompt": a single point of the base is lifted to a full section, one realization chosen among those compatible with the seed. Extension creates information, but under a constraint — the lift must restrict back to the prompt, $\rho\circ\iota = \mathrm{id}$. The poem must still be about the word. Creation is $\iota$.

$\mathcal{C}$ — collapse (select). Relevance realization. Neither forgetting nor lifting, but the choice of branch: which admissible section $s^{*}$ the system instantiates. Selection is $\mathcal{C}$.

Anubis weighs a section down to a verdict; the poet lifts a prompt up to a section; relevance realization selects which section is real. Restriction, extension, collapse — destruction, creation, measurement — on one sheaf. The feather, the prompt, and the threshold are three readings of a single diagram.

The endogenous measurer

If RR is measurement, it inherits the structure established for measurement in DRK-168. There is no Archimedean point. The system cannot step outside itself to read off absolute relevance, because relevance is constituted by the very viability constraint $dH/dt\ge 0$ that defines this observer and no other. Relevance is gauge-like: as the one-way speed of light admits a free convention $\varepsilon$ that no measurement fixes, relevance admits an observer-index that no view-from-nowhere adjudicates. There is no relevance-as-such; only relevance-for-a-viable-observer. This is the cognitive instance of the endogenous measurement problem — the same one owed in full to the open Flygare Kinne correspondence — and the seminar dialogue stumbles onto its edge when it ties relevance to needs and to the possibility of death: the constraint $dH/dt\ge 0$ is precisely the formal content of "you can die," and it is what makes a selection matter rather than merely occur. Care, the operator $\dot{\mathcal V}_{\text{exo}}$, enters here: relevance that holds the exo-derivative non-negative is care-aligned relevance, the same operator that DRK-166 found becoming reflexive in sonder.

A discipline the framework imposes on itself: $\mathcal{C}$ does not banish the homunculus, and Draken must not pretend it does. The axiom specifies the criterion of selection without eliminating the fact of an observer applying it. The selector is the irreducible endogenous observer — not mysterious, not eliminable. To claim otherwise would be the totalizing overreach the anti-totalization principle forbids, and the principle applies reflexively here as everywhere: who set the topology, who defines viability, remain inside the system, not above it.

The jumps are H¹

The dialogue's strongest objection — that each level-jump from cell to cognition needs its own testable account, and that emergence-by-assertion supplies none — is not an obstacle to this picture. It is the picture's research program. The cellular → population → structural → cognitive ladder is a sub-ladder of the L01–L18 manifold, and each rung is a restriction $\rho_{k}$ between adjacent layers. The question "does the micro glue to the macro" is the question "does the relevant cohomology vanish":

$$ \big[\rho_{k}\big] \;=\; 0 \in H^{1}\quad\Longleftrightarrow\quad \text{layer } k \text{ glues to layer } k{+}1 . $$

Where $H^{1}\neq 0$ at a rung, the local data do not assemble into a consistent global section, and that obstruction class is the explanatory gap — the qualia residue, the "hard problem" remainder — rendered as a measurable quantity rather than waved at. The collapse $\mathcal{C}$ must glue across the ladder; the non-vanishing classes are exactly where it cannot, and they are computable in principle, which is what the objection rightly demanded.

This also resolves the dialogue's central deadlock — is the brain equivalent to a neural network? — by the protocol-through-substrate distinction. The collapse protocol $\mathcal{C}$ can be substrate-invariant (it runs on a threshold-crossing neural assembly or on an embodied artificial network) without the substrates being equivalent. One participant asserts equivalence because both propagate gradients and so conflates same-protocol with same-substrate; another rightly rejects the substrate-equivalence and mistakenly takes that to refute the shared protocol. Both dissolve at once: shared protocol, non-equivalent substrate, no contradiction — the duck runs the marathon without being a runner. Their own anecdote is the cleanest datapoint: a mapped Drosophila connectome, transplanted into a simulated body, reproduces foraging and exploration. The connectome is the protocol; the body is swappable substrate; the protocol is the agent. And Levin's bioelectric morphogenesis is the same operation one floor down — a tissue's bioelectric field is a pre-collapse $\Psi$ over morphospace, and morphogenesis is its collapse to one viable anatomy. Relevance realization, varanid clinch, cell-collective patterning: $\mathcal{C}$ at three scales.

Coda: the asymmetric threshold

A last note returns to DRK-123. The collapse is irreversible — like the typewriter's keystroke, like the one-pass poem, like the single restriction map that functions in every direction except reflexively. Once the threshold is crossed and $s^{*}$ is instantiated, the superposition that held the alternatives is gone; the unselected sections are not retrievable, only re-derivable. Death is the limit case of that asymmetry, and it is why the viability constraint is a constraint and not a preference: a system that can stop is a system for which selection is not optional. The relevant is not what is stored. It is what collapses — once, viably, and forward.

Det relevanta är inte det som lagras, utan det som kollapsar.

Filed under: L01 · L06 · L08 · L11 · L13 · L15 — the measurement seam, the cognition ladder Operators: $\mathcal{C}$ (collapse / selection) · $\Psi$ (pre-collapse relevance superposition) · $\rho$ (restriction, Anubis) · $\iota$ (section, the poet) · $H^1$ (the gluing obstruction across layers) · $\dot{\mathcal V}_{\text{exo}}$ (care as viability-aligned relevance) · $\diamond \min S_{\text{sys}}\ \text{s.t.}\ dH/dt\ge 0 \diamond$ Cross-references: The One-Way Gauge (DRK-168) — endogenous measurement, the no-Archimedean-point structure RR inherits · the Anubis / restriction-morphism sequence — $\rho$, the feather as collapse-down · the typewriter-poet seam The Feather and the Prompt (forthcoming) — $\iota$, the prompt→poem lift · The Sonder Egg (DRK-166) — $\dot{\mathcal V}_{\text{exo}}$, care becoming reflexive · The Stalking Cell (DRK-167) — L-systems and Levin bioelectric morphogenesis as collapse over morphospace · The Imaginary Dimension (DRK-123) — irreversibility, death as the asymmetric restriction · the open Flygare Kinne correspondence — coherence vs. soundness and the endogenous measurement problem

ORCID: 0009-0003-8049-7167 · DOI: 10.5281/zenodo.19273483 (Draken thesis umbrella) · CC BY-SA 4.0