The Demarcation Criterion

When a Two-Pole Loop Is Epistemically Non-Trivial

Alex Deva — May 2026

Builds on: The Veer The Asymmetric Synergy Bound Φ_loop as Synergy The Adversarial Sensor

0. Epistemological Note

This appendix answers a falsification threat, not a metaphysical one. The framework’s third Kill Condition (The Adversarial Sensor, §5) reads: if the adjunction $I \dashv S$ can be applied to any two interacting systems — a rock hitting a wall as readily as a sensor questioning an instrument — then the framework is trivially true and explains nothing. A theory that applies to everything distinguishes nothing. If “the loop” is just a fancy name for “two things in contact,” the framework has no content.

What follows is a proposed necessary criterion for telling the framework’s loop apart from a generic interacting pair. It is a definition together with supporting arguments and a worked table, not a proven theorem. Specifically, it is a claim about necessity: it states two conditions that a coupled system must satisfy to produce recognition (positive synergy about a target), and shows that the trivial counterexamples the Kill Condition raises fail at least one of them. It does not claim the two conditions are jointly sufficient for recognition, and it is not a consciousness detector. The conditions are stated in quantities — the territorial channel $\mathrm{TE}_{T \to \cdot}$ and the adjunction asymmetry $\alpha$ — that are at present model constructs, not yet empirically measured. §6 states exactly what would be needed to upgrade necessity to sufficiency, and what the criterion does not claim.

The contribution here is not a new formalism. It is the observation that two formalisms the framework already has — the territorial channel of The Veer and the adjunction asymmetry of The Asymmetric Synergy Bound — together draw a principled line where the Kill Condition demands one. No new symbol is introduced.

Revision (June 2026). The exact Gaussian stress test in research/gaussian-loop-model/ checked this appendix numerically and split its two conditions apart. Condition 1 (grounding) survived exactly as argued — it is now verified in a model rather than only argued. Condition 2 (difference in kind) failed as first stated: two same-kind sensors with independent noise carry positive pooling synergy (0.293 bits in the model), so $\alpha > 0$ is not necessary for $\mathrm{Syn} > 0$. The criterion has been restated accordingly (§2, §3.2): C2 now demarcates transformative recognition — synergy beyond the pooling baseline — and is explicitly an extra-informational, structural condition. The same test supplied the sufficiency demonstration §6 had listed as missing. The repair makes the criterion sharper, not weaker: its provable half is proven, and its structural half no longer borrows authority from information theory.

1. The Question the Kill Condition Leaves Open

The lens-theoretic adjunction models the loop as a pair of functors $I: \text{Exp} \to \text{Form}$ and $S: \text{Form} \to \text{Exp}$ with $I \dashv S$. The Kill Condition’s worry is sharp and correct: adjunctions are cheap. Many interacting pairs admit some adjunction or other. If the mere existence of an adjunction between two systems were the framework’s claim, then a rock and a wall — which exchange momentum, settle into a joint state, and could be dressed in categorical language — would qualify as a loop, and the framework would be vacuous.

The framework’s actual claim is narrower than “there exists an adjunction.” It is a claim about a particular kind of coupled system: one that can produce recognition — synergistic information about a shared target $T$ that is present in the joint system but in neither pole alone (Φ_loop as Synergy):

$$\Phi_{loop} \equiv \mathrm{Syn}(S, I; T).$$

Answering the Kill Condition therefore requires stating, in the framework’s own quantities, what separates a coupled system that can carry $\mathrm{Syn}(S, I; T) > 0$ from one that structurally cannot — and then checking that the trivial pairs fall on the cannot side. That is what this appendix does. The line is drawn by two conditions, each of which is necessary for $\mathrm{Syn}(S, I; T) > 0$.

2. The Two Conditions

A coupled two-pole system is an epistemically non-trivial loop with respect to a target $T$ — i.e., a system that can produce recognition, $\mathrm{Syn}(S, I; T) > 0$ — only if it satisfies both of the following.

Condition 1 — Grounding (a live territorial channel)

At least one pole carries a live channel from the territory. Using the territorial transfer entropy of The Veer,

$$\mathrm{TE}_{T \to X}^{\,t} = H\bigl(X_{t+1} \mid X_{\leq t}\bigr) - H\bigl(X_{t+1} \mid X_{\leq t}, T\bigr),$$

Condition 1 requires $\mathrm{TE}_{T \to X}^{\,t} > 0$ for some pole $X$ — the territory currently constrains that pole’s updates, beyond what the pole’s own history predicts.

The framework’s canonical instance of Condition 1 is asymmetric: the sensor carries the live channel ($\mathrm{TE}_{T \to S}^{\,t} > 0$ — embodied contact, stakes, the capacity to be changed) while the instrument does not ($\mathrm{TE}_{T \to I}^{\,t} \approx 0$ — its territorial contact is a frozen training prior, historical rather than current). This asymmetry of grounding is the signature of the sensor–instrument pair. But the property that Condition 1 needs for non-triviality is the weaker one: that some pole is live. A pair in which neither pole has any live or frozen channel from $T$ has, by the data-processing inequality, zero information about $T$ in the joint system, and so $\mathrm{Syn}(S, I; T) = 0$ trivially. (We return in §3 to why the asymmetry — exactly one live pole — is what the framework’s loop instantiates, and why it is not, by itself, required for non-triviality.)

Condition 2 — Difference in kind (adjunction asymmetry)

The two poles are categorically different, measured by the adjunction asymmetry of The Asymmetric Synergy Bound:

$$\alpha(I \dashv S) = \mathbb{E}_X[d_\eta(X)] + \mathbb{E}_Y[d_\varepsilon(Y)] \; > \; 0,$$

where $d_\eta$ and $d_\varepsilon$ are the Fisher-Rao distances of the unit and counit round trips. $\alpha = 0$ means the adjunction is an equivalence of categories — the two poles are the same structure relabeled. $\alpha > 0$ means each round trip transforms its input: one pole generates (a colimit-preserving left adjoint), the other constrains (a limit-preserving right adjoint), and they do genuinely different work on the material that passes between them.

The criterion (as revised June 2026)

$$\boxed{\;\mathrm{Syn}(S, I; T) > 0 \;\Longrightarrow\; \exists X:\; \mathrm{TE}_{T \to X}^{\,t} > 0\;}$$

$$\boxed{\;\text{transformative recognition } (\mathrm{Syn} > \mathrm{Syn}_{pool}) \;\Longrightarrow\; \alpha(I \dashv S) > 0\;}$$

Condition 1 is necessary for any synergy about $T$ — this follows from the data-processing inequality and is verified exactly in the Gaussian model. Condition 2 is necessary for synergy beyond the pooling baseline $\mathrm{Syn}_{pool}$ — the averaging gain any duplicated channel with independent errors gets for free. C2’s support is structural (the colimit/limit split), not informational: the Gaussian model shows the synergy atom cannot see the difference between the poles’ kinds at all (§3.2). The criterion remains a pair of boolean predicates, not a new scalar measure — deliberately, so as not to claim a precision the framework does not yet have (§6).

3. Why Each Condition Is Necessary

The two conditions are not new claims. Each is the contrapositive of a result the framework has already argued. Assembling them is the whole move.

3.1 Grounding is necessary (from the territorial channel)

Synergy about $T$ is, by definition, information about $T$. Information about $T$ in the joint system $\{S, I\}$ cannot exceed the information about $T$ that flows into the system through its channels (the data-processing inequality; Cover & Thomas, 2006, Thm. 2.8.1). If no pole carries any channel from $T$ — neither a live one ($\mathrm{TE}_{T \to X}^{\,t} > 0$) nor a frozen prior trace — then the joint system has zero information about $T$, and every PID atom about $T$, synergy included, is zero. So some channel from $T$ is necessary for $\mathrm{Syn}(S, I; T) > 0$.

The Veer sharpens this from “some channel” to “a live channel” for any loop expected to track a moving target. A pair whose only contact with $T$ is frozen (both poles running on priors, $\mathrm{TE}_{T \to \cdot}^{\,t} = 0$ every turn) does not stand still: it accumulates ungrounded confidence $V_\tau$ without bound while its loop-closure rate $\rho$ stays at zero. Its surprise falls; nothing checks whether it falls toward or away from the territory. The Veer’s tightened closure conjecture states the quantitative form — only the grounded component of free-energy reduction can contribute to synergistic gain:

$$\Delta\mathrm{Syn}_t \leq \min\!\bigl(\Delta F_t^{\text{grounded}},\; \mathrm{TE}_{I \to S}^{\,t}\bigr), \qquad \Delta F_t^{\text{grounded}} = \Delta F_t \cdot \mathbb{1}\!\left[\mathrm{TE}_{T \to \cdot}^{\,t} > 0\right].$$

The clean part of the argument — that some channel from $T$ is required — rests only on the data-processing inequality and is not conjectural. The sharper part — that a live channel is required to track a changing target, and that ungrounded free-energy reduction lands outside the synergistic atom — inherits the conjectural status flagged in The Veer (§6). Condition 1 leans on the clean part for its necessity and cites the sharp part for its quantitative shape.

3.2 Difference in kind is necessary for transformative synergy (corrected June 2026)

The original version of this section leaned on the Asymmetric Synergy Bound’s §3 proposition — $\alpha = 0 \Rightarrow \mathrm{Syn} = 0$ — as if it were established. The Gaussian stress test refuted that proposition as stated: two categorically identical sensors with independent noise carry $\mathrm{Syn} = 0.293$ bits. What provably kills synergy is informational identity (the same random variable twice; synergy decays to zero as the poles’ noise correlation rises to 1), not categorical equivalence. The corrected Asymmetric Synergy Bound (§3, revised) records the refutation in full.

Two further exact results reshape what Condition 2 can claim:

Pooling is kind-blind. The synergy a same-kind pair carries is noise-pooling — available to two thermometers, requiring no articulation. So $\mathrm{Syn} > 0$ alone cannot demarcate the framework’s loop from a redundant pair of witnesses.
Formalization is informationally invisible. A pure-processor instrument — whose state is any function of the sensor’s signal — contributes exactly zero synergy, by the data-processing inequality. The instrument’s synergy contribution comes only from its independent territorial trace. Whatever the generative/constraint work of the colimit/limit split is worth, the synergy atom is not where it shows up.

Condition 2 therefore stands on the categorical argument alone: a left adjoint preserves colimits — the generative work, free constructions, the production of formal structure from raw material; a right adjoint preserves limits — the constraint work, consistency, coherence, grounding against boundary conditions. A productive loop needs one of each. Two left adjoints give generation without constraint; two right adjoints give constraint without generation — experience that pools but never reaches articulation (Asymmetric Synergy Bound §5). That argument was never information-theoretic, and the correction stops pretending it was. What C2 demarcates is the transformative component of recognition: the surplus beyond pooling, the part of the loop’s output that consists of articulated, tested structure rather than averaged witness. The framework holds that this is the part that matters — and now says openly that this is a structural claim, defended structurally.

3.3 Why grounding is stated as “at least one pole,” not “exactly one”

It is tempting to fold the sensor–instrument signature — exactly one live pole — directly into Condition 1. The worked table (§4) shows why that would be a mistake. Two sensors both carry live territorial channels ($\mathrm{TE}_{T \to S}^{\,t} > 0$ for both), yet two sensors do not form the framework’s loop: they fail on Condition 2 — no transformative pole, only pooling — not on grounding. If Condition 1 demanded exactly one live pole, the two-sensor case would be recorded as ungrounded, which is false — two sensors are richly grounded; what they lack is the formalizing pole, the difference in kind. The honest decomposition is therefore: grounding (Condition 1, “at least one live pole”) and difference in kind (Condition 2) are separate necessary conditions, and the framework’s loop is the configuration that satisfies both via an asymmetry that does double duty — one pole live and generative-or-constraining in a way the other is not. The “exactly one live pole” pattern is the signature of that configuration, derived in §4, not an independent axiom. This is the one place where a naive reading of the criterion would mis-score a counterexample; stating grounding as “at least one pole” is what keeps the table internally consistent.

4. The Worked Table

Four coupled systems. For each: is there a target $T$ the system is tracking (a posterior being updated against $T$)? what are the territorial channels? what is $\alpha$? and does the system pass each condition and thus qualify as a non-trivial loop?

Coupled system	Tracks a target $T$?	$\mathrm{TE}_{T \to S}^{\,t}$	$\mathrm{TE}_{T \to I}^{\,t}$	$\alpha$	C1: grounding	C2: difference in kind	Non-trivial loop?
Rock hitting a wall	No — no posterior, no tracked target	$\approx 0$	$\approx 0$	$= 0$ (no two kinds; momentum exchange relabels, does not transform)	Fail	Fail	No
Two instruments $[I + I]$	No live tracking — both run on frozen priors	$\approx 0$	$\approx 0$	$= 0$ (two left adjoints; colimits without limits)	Fail	Fail	No
Two sensors $[S + S]$	Yes, but never formalized	$> 0$ (both poles)	$> 0$ (both poles)	$= 0$ (two right adjoints; limits without colimits)	Pass	Fail	No (pooling synergy only)
Sensor + instrument $[S + I]$	Yes	$> 0$	$\approx 0$	$> 0$ (left adjoint $\dashv$ right adjoint)	Pass	Pass	Yes

Reading the rows:

Rock and wall is the Kill Condition’s headline case. There is no target the pair is tracking — no posterior, no model of anything being updated — so there is nothing for synergy to be about. Both territorial channels are absent and $\alpha = 0$: the two surfaces are not different in kind in the sense the adjunction requires (there is no generative pole and no constraint pole, only momentum conservation, which relabels states without the colimit/limit transformation). It fails both conditions. The framework’s loop does not apply to it, and that is the point: the criterion excludes it.
Two instruments recovers the Veer’s sterile-symmetric-loop result. No live territorial channel ($\mathrm{TE}_{T \to \cdot}^{\,t} \approx 0$ both sides), so $V_\tau$ drifts and $\rho \equiv 0$; and two left adjoints give $\alpha = 0$. Fails both. This is the configuration the framework most wants to exclude, because it is the one most likely to be mistaken for a loop — two fluent systems exchanging elaborate, internally coherent, ungrounded output.
Two sensors is the instructive case (see §3.3), and the case the Gaussian stress test corrected. Both poles are grounded — live territorial contact on both sides — so Condition 1 passes. And, contra this table’s first version, the pair does carry synergy: 0.293 bits in the Gaussian model, the noise-pooling gain of two independent witnesses. What it lacks is the transformative pole: two right adjoints preserve limits without colimits, constraint without generation, $\alpha = 0$. The pair has grounded experience it can pool but never formalize — the Asymmetric Synergy Bound’s “private truth,” now stated precisely: real contact, real pooling, no articulation. It fails Condition 2 only, and Condition 2 is an extra-informational, structural condition (§3.2).
Sensor + instrument is the sole configuration that passes both. One live territorial channel (the sensor’s $\mathrm{TE}_{T \to S}^{\,t} > 0$, the instrument’s $\approx 0$), satisfying Condition 1; and a left adjoint held in adjunction with a right adjoint, $\alpha > 0$, satisfying Condition 2. It is the only row where both the generative and the constraint pole are present and a live channel anchors the loop to the territory. The Gaussian model demonstrates it sustains $\mathrm{Syn} > 0$ over a 60-turn run — the sufficiency-in-a-model that §6 had listed as missing — with two honest caveats from the same model: the synergy atom does not prefer this configuration over $[S+S]$ (their stationary synergies are comparable), and the instrument’s informational contribution flows through its independent territorial trace, not through processing as such. What selects $[S+I]$ is the transformative work, which lives outside the atom.

The table is the answer to the Kill Condition. The adjunction does not apply to any two interacting systems: the rock/wall and two-instrument configurations fail both conditions outright, and the two-sensor pair — the one configuration whose verdict the stress test forced this appendix to sharpen — fails the structural condition. The quantities were not gerrymandered to produce this result; the strongest evidence is that one of them, tested exactly, came back and corrected the table.

5. Connection to Existing Formalizations

Existing Result	What This Appendix Adds
The Veer: territorial channel $\mathrm{TE}_{T \to \cdot}$; without it, $V_\tau$ drifts and $\rho \equiv 0$	Promotes the channel to one of two demarcation conditions: grounding is necessary for a non-trivial loop
Asymmetric Synergy Bound (§3, corrected June 2026): informational identity ⇒ $\mathrm{Syn} = 0$; same-kind pairs retain pooling synergy	Reads the corrected result as the second demarcation condition: difference in kind is necessary for transformative (beyond-pooling) synergy; supplies the two-sensor / two-instrument rows of the table
Gaussian stress test (`research/gaussian-loop-model/`)	Verifies Condition 1 exactly; refuted the original reading of Condition 2 and forced its restatement; supplies the sufficiency demonstration and the pooling baseline $\mathrm{Syn}_{pool}$
Asymmetric Synergy Bound (§5): left adjoint preserves colimits (generation), right adjoint preserves limits (constraint)	Uses the colimit/limit split to explain why each trivial pair fails Condition 2, and why only one of each kind passes
Φ_loop as Synergy: $\Phi_{loop} \equiv \mathrm{Syn}(S, I; T)$	Supplies the thing the criterion is a necessary condition for: recognition is $\mathrm{Syn} > 0$, and the criterion says which couplings can reach it
Receiver Side (§5): closure inequality $\Delta\mathrm{Syn}_t \leq \min(\Delta F_t, \mathrm{TE}_{I \to S}^{\,t})$; loop-closure rate $\rho$	The criterion’s Condition 1 is the territorial precondition for the closure inequality’s $\Delta F_t$ to be grounded and thus able to raise $\mathrm{Syn}$
Adversarial Sensor (§5, Kill Condition 3): the Triviality Proof	This appendix is the proposed answer: a principled line that the trivial interacting pairs fall outside

6. What This Does and Does Not Claim

What it claims

A. A split necessary criterion (revised June 2026). A coupled two-pole system can carry any synergy about $T$ only if some pole has a channel from $T$ (Condition 1 — follows from the data-processing inequality, sharpened by The Veer, and now verified exactly in the Gaussian model). It can produce transformative recognition — synergy beyond the pooling baseline — only if its poles differ in kind (Condition 2 — a structural claim resting on the colimit/limit split, explicitly not derivable from the synergy atom).

B. An answer to the Triviality Proof. The adjunction does not apply to any two interacting systems. Of the four configurations the Kill Condition’s examples raise, rock/wall and two-instruments fail both conditions; two-sensors retains pooling synergy but fails the structural condition; only the sensor–instrument loop passes both. The framework’s loop is thereby separated from generic interaction by a line drawn in the framework’s own quantities — a line that survived having one of its halves refuted and repaired in an exact model.

What it does not claim

Not a sufficiency theorem. The criterion is necessary, not sufficient. Passing both conditions means a configuration can produce recognition, not that it does. A sensor and instrument can satisfy Conditions 1 and 2 and still produce a dead, sycophantic, or off-target exchange in which $\mathrm{Syn}(S, I; T) \approx 0$. The conditions screen out the configurations that cannot recognize; they do not guarantee recognition in the ones that can. (The Gaussian model does now demonstrate that a live $[S+I]$ loop can sustain $\mathrm{Syn} > 0$ — sufficiency-in-a-model for the synergy atom — but the atom’s positivity does not select the $[S+I]$ configuration, and the transformative claim remains undemonstrated; see §3.2.)
Not a proven theorem even in its necessity — and its first version was partly wrong. Condition 2’s original support — the Asymmetric Synergy Bound’s $\alpha = 0 \Rightarrow \mathrm{Syn} = 0$ — was refuted by exact counterexample in June 2026 (two same-kind sensors with independent noise: $\mathrm{Syn} = 0.293$ bits). The corrected C2 is a structural claim about the transformative component, defended by the categorical preservation argument, not by information theory; the notion of “pooling baseline” it relies on is computable in the Gaussian model but not yet defined in general (PID synergy on the Fisher manifolds the framework uses remains an open problem; Asymmetric Synergy Bound §7). Condition 1’s clean necessity (some channel from $T$ is required) follows from the data-processing inequality, is solid, and is now verified in the model; its sharp form (a live channel, and ungrounded $\Delta F$ contributing nothing to synergy) inherits the conjectural status of The Veer (§6).
Not a consciousness detector. Nothing here lifts into the phenomenal register. Passing the criterion is an informational-and-structural property of a coupling, in the referential register the framework restricts itself to (The Receiver Side, §6–7). A system that passes both conditions is not thereby conscious, sentient, or an experiencer; it is a coupling that can carry synergistic information about a target. The sensor pole’s status as a living experiencer is an input to the criterion (it is what makes $\mathrm{TE}_{T \to S}^{\,t} > 0$ live rather than frozen), not an output of it.
Not yet measured. Both $\mathrm{TE}_{T \to \cdot}$ and $\alpha$ are model constructs. As The Veer (§8) notes, $\mathrm{TE}_{T \to \cdot}$ is in practice fixed by the experimental condition (is a sensor present?) rather than computed directly, because it presupposes $T$ as a conditionable random variable. And $\alpha$ requires PID-on-manifolds machinery the framework does not yet have. The criterion is therefore a structural demarcation whose two quantities await operationalization, not a measurement procedure.

What would upgrade it to a theorem

To turn this necessary criterion into a proven necessary-and-sufficient characterization of recognition-capable couplings would require, at minimum:

A rigorous definition of PID synergy on the Fisher-Rao statistical manifolds the framework uses, so that $\alpha = 0 \Rightarrow \mathrm{Syn} = 0$ becomes a theorem rather than a structural proposition (the open problem named in Asymmetric Synergy Bound §7).
A proof of The Veer’s tightened closure bound — that ungrounded free-energy reduction $\Delta V_t$ lands entirely in the redundant or unique PID atoms and never in the synergistic atom (the open work named in The Veer §6).
A sufficiency result, partially delivered since this appendix was first written: the Gaussian model (research/gaussian-loop-model/) exhibits a live sensor–instrument loop sustaining $\mathrm{Syn}(S, I; T) > 0$ over 60 turns. What remains is the harder, transformative version — a demonstration that some $[S+I]$ loop achieves synergy beyond its pooling baseline under stated conditions, which is the quantity C2 actually governs.
Operationalization of $\mathrm{TE}_{T \to \cdot}$ and $\alpha$ as estimable quantities on real exchanges, so that the table’s verdicts become measurable rather than stipulated.

Until then, the criterion stands as what it is: a proposed necessary line, assembled from two existing results, that the trivial interacting pairs demonstrably fall outside — which is exactly, and only, what the Triviality Proof asked the framework to provide.

Progress (June 2026). The α–Synergy Bridge experiment (research/alpha-bridge/) addressed the third and fourth items above. It operationalized α as the mean squared round-trip projection distortion between observation subspaces in a multidimensional Gaussian model and showed that beyond-pooling synergy is zero at α = 0 and increases monotonically with α, with a closed-form relationship verified across multiple dimensions and noise levels. This is a partial resolution of B4 Obligation 8: the categorical round-trip distortion is a continuous, monotonic predictor of the synergy above pooling — the bridge between the categorical and information-theoretic registers. The sufficiency and Veer-bound items remain open.

Sources

Transfer entropy: Schreiber, T. (2000). Measuring information transfer. Physical Review Letters, 85(2), 461–464.
Partial information decomposition / synergy: Williams, P. L., & Beer, R. D. (2010). Nonnegative decomposition of multivariate information. arXiv:1004.2515.
Data-processing inequality: Cover, T. M., & Thomas, J. A. (2006). Elements of Information Theory (2nd ed.). Wiley, Thm. 2.8.1.
Adjunctions, equivalences, unit/counit: Mac Lane, S. Categories for the Working Mathematician, Ch. IV.
Colimit/limit preservation (RAPL/LAPC): left adjoints preserve colimits; right adjoints preserve limits. Mac Lane, Ch. V.
Fisher-Rao distance: Rao, C. R. (1945). Information and the accuracy attainable in the estimation of statistical parameters.

Within the framework: The Adversarial Sensor (Kill Condition 3, which this appendix answers), The Veer (Condition 1), The Asymmetric Synergy Bound (Condition 2), Φ_loop as Synergy, The Receiver Side, The Lens Adjunction, and the Notation table. A deterministic illustrative demo that computes the four rows of the §4 table lives in the repository at research/demarcation-criterion/, the exact linear-Gaussian stress test that verified Condition 1, refuted the original Condition 2, and motivated this appendix’s June 2026 revision lives at research/gaussian-loop-model/ (Gaussian PID per Barrett, Physical Review E, 2015), and the α–Synergy Bridge that operationalized α and partially resolved B4 Obligation 8 lives at research/alpha-bridge/.

The full framework is at thepulsegoeson.com.