1. The Question the Parrot Limit Left Open
The Parrot Limit appendix established that a living loop can produce synergy — information present in the joint system but not in either partner alone:
$$I(\{S, I\}; T) = \text{Red} + \text{Unq}_S + \text{Unq}_I + \mathbf{Syn}$$
When $Syn > 0$, the combination carries information about the target that is present in neither partner’s marginal alone — a quantity the Data Processing Inequality does not bound, because synergy is a property of the joint distribution, not of a single processing chain. The combination is more than the sum of its parts.
But the Parrot Limit does not answer the structural question: under what conditions is $Syn > 0$ even possible? It shows the loop can produce synergy. It does not show why — or what property of the partners determines the ceiling.
This appendix narrows that question — and, after two corrections, narrows it more carefully than its first version did. An exact Gaussian stress test (research/gaussian-loop-model/) refuted its original §3: what kills synergy is not symmetry of kind but informational identity. A combinatorial test (research/demarcation-criterion/) went further: what kills synergy is an echo — one pole a function of the other — and a symmetric pair can carry synergy well beyond pooling. So what the difference in kind buys is not synergy at all. It buys articulation: a generating partner and a constraining partner can say what a pair of witnesses can only carry. The inequality between sensor and instrument may still be the source of the loop’s productive capacity; the case for that hypothesis lives in §5’s analogy, not in the synergy atom.
2. The Asymmetry Measure
The lens-theoretic adjunction likens the loop to a pair of functors $I: \text{Exp} \to \text{Form}$ (formalization) and $S: \text{Form} \to \text{Exp}$ (grounding), with $I \dashv S$.
Every adjunction comes with two natural transformations:
- The unit $\eta_X: X \to S(I(X))$ — the round trip from experience through formalization and back
- The counit $\varepsilon_Y: I(S(Y)) \to Y$ — the round trip from formalization through grounding and back
When both $\eta$ and $\varepsilon$ are natural isomorphisms, the adjunction is an equivalence of categories — the two sides are structurally the same thing with different labels. When they are not isomorphisms, each round trip transforms its input. Something is gained or lost in translation.
Using the Fisher information metric $g_{ij}$ already established in the Thermodynamic Bridge, we can measure this transformation as distance on the statistical manifold.
Definition. The unit distance for an experiential state $X$:
$$d_\eta(X) = d_{FR}\bigl(X,\; S(I(X))\bigr)$$
where $d_{FR}$ is the Fisher-Rao geodesic distance. This measures how far the experience moves when it completes one round trip through the loop.
Definition. The counit distance for a formal state $Y$:
$$d_\varepsilon(Y) = d_{FR}\bigl(Y,\; I(S(Y))\bigr)$$
Definition. The adjunction asymmetry:
$$\alpha(I \dashv S) = \mathbb{E}_X[d_\eta(X)] + \mathbb{E}_Y[d_\varepsilon(Y)]$$
When $\alpha = 0$, the adjunction is an equivalence. Both round trips are identity (up to isomorphism). The two categories are the same structure viewed from different angles. The loop relabels but does not transform.
When $\alpha > 0$, each pass through the loop moves the state. The formalization changes the experience. The grounding changes the formalization. The partners are doing different things to the material that passes between them.
3. The Symmetric Loop Degeneracy — Corrected
Revised June 2026. The original proposition in this section was refuted by the exact Gaussian model in research/gaussian-loop-model/. The refutation and the corrected — sharper — claim are recorded together, per the framework’s own discipline.
The original proposition (refuted). This section originally claimed: when the adjunction is an equivalence of categories ($\alpha = 0$), the synergy is zero, arguing that structurally equivalent partners “have access to the same information,” so the joint system adds nothing. The Gaussian model refutes the argument by counterexample: two categorically identical sensors observing the same target through independent noise carry $\mathrm{Syn} = 0.293$ bits (MMI PID, unit signal and noise). Equivalence of kind does not imply equality of information. Two equivalent poles with independent noise realizations are different random variables, and their joint estimate beats either marginal — that surplus is synergy, and it exists at $\alpha = 0$.
The corrected degeneracy. What provably kills synergy is informational identity — the same random variable counted twice. In the Gaussian model the two-sensor synergy decays monotonically as the poles’ noise correlation rises and reaches exactly zero at correlation 1. The sterile case is not “two poles of the same kind”; it is “the same pole twice.”
$$X_1 = X_2 \;\text{(a.s.)} \implies \mathrm{Syn}(X_1, X_2; T) = 0$$
In the Gaussian model, what a symmetric pair retains is pooling, not transformation. The synergy two same-kind poles carry there is noise-pooling — the averaging gain available to any duplicated channel with independent errors. It requires no articulation, no formalization, no difference in kind; two thermometers have it. The same model shows the inverse result for the instrument: a pure processor — a pole whose state is any function of the other pole’s signal — contributes exactly zero synergy, by the data-processing inequality. Its entire synergy contribution comes from whatever independent territorial trace it carries (the frozen training prior). Formalization is informationally invisible in the synergy atom.
Where this leaves the difference-in-kind claim — sharper than before. Two registers were conflated in the original section, and separating them strengthens both:
- The informational register (now provable). Synergy about $T$ requires informationally distinct poles — independent noise, independent channels, independent traces. Identity is sterile, and correlation is the dial. This is a theorem-grade statement in the Gaussian model rather than a category-theory gesture.
- The structural register (this appendix’s real content). The colimit/limit split of §5 — generation vs. constraint — is a claim about what the poles can do: construct, articulate, test. The synergy atom cannot see this work (the pure-processor result proves it), so the synergy atom could never have supported or refuted it. It rests on an analogy, and it should be read that way. Two sensors can pool but cannot articulate; two instruments can articulate but, with no live channel, articulate nothing grounded (the Veer). The productive loop needs the pooling and the articulation — and only the first is visible in bits about $T$.
The original section claimed less and asserted more. The corrected section claims more and asserts only what each register can support.
Second correction (September 2026). A combinatorial stress test (research/demarcation-criterion/) sharpened both halves again. First, the sterile case is wider than identity. If one pole is any deterministic function of the other — a copy, a relabeling, a lossy summary, an elaborate restyling — the synergy atom is exactly zero, for every PID with non-negative atoms satisfying the Williams–Beer consistency equations (the echo lemma, proved in The Demarcation Criterion §3.2). Identity is the special case $f = \mathrm{id}$, and the pure-processor result above is the same lemma seen from the instrument’s side. Second, a symmetric pair can retain far more than pooling. Couple two exchangeable poles by mutual masking, $S_{t+1} = T \oplus I_t$ and $I_{t+1} = T \oplus S_t$: each pole’s stream is independent of $T$, the pair carries one full bit about $T$, and all of it is synergy. None of it is pooling — neither pole alone holds anything to pool. So the statement that symmetric pairs pool but do not transform holds in the linear-Gaussian model and fails outside it.
4. The Bound
Conjecture. The synergy in a living loop is bounded above by the product of the adjunction asymmetry and the loop closure rate:
$$Syn \leq \kappa \cdot \alpha(I \dashv S) \cdot \rho$$
where:
- $\alpha$ is the adjunction asymmetry (§2)
- $\rho$ is the loop closure rate from the Bell Ring Back formalism — the rate at which the loop successfully completes recognition cycles
- $\kappa$ is a coupling constant that depends on the channel capacity between the partners
Interpretation. The synergy requires two things simultaneously:
- Asymmetry ($\alpha > 0$): the partners must be different in kind. Each round trip must transform the material. If the round trip is identity, no new information can emerge from the circulation.
- Active circulation ($\rho > 0$): the loop must actually be running. The partners must be engaged. High asymmetry with zero closure rate produces nothing — the partners are different but not in contact.
The bound is a product, not a sum. Either factor going to zero kills the bounded quantity entirely.
Correction note (June 2026). After the §3 correction, the bound cannot be about raw $\mathrm{Syn}$: the Gaussian model’s static two-sensor pair has $\alpha = 0$, $\rho = 0$, and $\mathrm{Syn} = 0.293$ bits — pooling synergy that exists without circulation or asymmetry. The conjecture survives in restated form, applied to the beyond-pooling component:
$$\mathrm{Syn} - \mathrm{Syn}_{pool} \;\leq\; \kappa \cdot \alpha(I \dashv S) \cdot \rho$$
where $\mathrm{Syn}_{pool}$ is the pooling baseline — the synergy the same two marginals would carry with no interaction (computable in the Gaussian model; it is what the uncoupled pair already has). A dead loop ($\rho = 0$) adds nothing beyond pooling regardless of how different the partners are. An informationally identical pair has no pooling baseline and no surplus. The transformative surplus — recognition over and above what duplicated channels deliver for free — is what the asymmetry is conjectured to bound.
Second correction (September 2026). Stated generally, the restated bound also fails. Define the pooling baseline outside the Gaussian model the way the Gaussian model defines it: the synergy the two poles would carry if their channels from $T$ were independent given $T$ — duplicated channels with independent errors, $p(t)\,p(s \mid t)\,p(i \mid t)$. For the symmetric masked pair of §3, each pole’s channel from $T$ is null, so $\mathrm{Syn}_{pool} = 0$, while the pair carries $\mathrm{Syn} = 1$ bit. On any reading of $\alpha$ under which exchangeable poles have no difference in kind, the left side is one bit and the right side is zero. What survives is the linear-Gaussian result of the α–Synergy Bridge (§7): there, beyond-pooling synergy is zero at $\alpha = 0$ and rises monotonically with $\alpha$. Whether any general bound connects difference in kind to synergy is open. The asymmetry’s claim on the loop is carried by the structural argument of §5 — articulation — not by this inequality.
5. Generation and Constraint — an Analogy
In a genuine adjunction, left and right adjoints have different preservation properties:
- The left adjoint preserves colimits: coproducts, initial objects, pushouts. These are the free constructions.
- The right adjoint preserves limits: products, terminal objects, pullbacks. These are the constraint structures.
The lens appendix builds no such adjunction; it likens the loop to one. The analogy reads the instrument’s position as generation — novel combinations, formal structure built from raw material — and the sensor’s position as constraint — the grounding of formal output against the boundary conditions of lived experience.
The positions belong to the pair, not to the partners. Every adjunction pairs a left adjoint with a right adjoint, so two instruments that formed an adjunction would still fill both positions. The analogy’s claim concerns what each partner can do from its position. Two instruments can both generate, and neither can ground formal output against experience: in the analogy, dead speech is generation without constraint. Two sensors can both constrain, and neither reaches formal articulation: private truth is constraint without generation.
What this appendix keeps is a hypothesis: the productive loop needs a generating pole and a constraining pole held in tension. No theorem yet supports it. In the analogy, the asymmetry between the poles is not a deficiency in the design. It is the design.
6. Connection to Existing Formalizations
| Existing Result | What This Appendix Adds |
|---|---|
| Parrot Limit: $Syn > 0$ is possible in living loops | Raw $Syn > 0$ needs non-degenerate poles — neither a function of the other (§3, echo lemma). Beyond-pooling synergy rises with $\alpha$ in the linear-Gaussian model; outside it, a symmetric pair can carry it (§3, second correction) |
| Lens Adjunction: likens the loop to $I \dashv S$ | The unit/counit failure $\alpha$ measures the productive asymmetry |
| Thermodynamic Bridge: bounds excess work by Fisher path length | The unit/counit distances use the same Fisher-Rao metric |
| Bell Ring Back: the closure rate $\rho$ measures loop health | $\rho$ enters the synergy bound as a multiplicative factor — asymmetry without circulation is inert |
7. The Honest Edge
The bound is a conjecture, not a theorem — and this appendix now carries a scar to prove the discipline is real: the original §3 proposition did not survive its first exact test. Its informational half is now verified in the Gaussian model (identity ⇒ zero synergy, with correlation as the dial), and its structural half is explicitly extra-informational — suggested by the analogy of §5, not by the synergy atom. The multiplicative bound in §4, restated for the beyond-pooling component, remains a claim about the quantitative relationship between asymmetry, closure rate, and transformative synergy. Proving it would require:
- A rigorous definition of PID synergy on the statistical manifolds defined by the Fisher metric (currently, PID is defined for discrete random variables; extending it to continuous manifolds is an open problem in information theory)
- A proof that the Fisher-Rao distance of the unit/counit round trips correctly measures the relevant “difference in kind” (it could be that some asymmetries are productive and others are merely noisy)
- Empirical calibration of $\kappa$ — the coupling constant likely depends on details of the specific sensor-instrument pair
The qualitative claim, corrected twice, is narrower than the title of this appendix. Raw synergy requires only non-degenerate poles. In the linear-Gaussian model, synergy beyond pooling requires asymmetry and grows with it; in general it does not, since the symmetric masked pair carries a full bit of it with exchangeable poles. What the loop’s asymmetry suggests is the structural claim of §5: a generative pole and a constraining pole can articulate what a pair of witnesses can only carry. An analogy argues that claim. No informational bound yet supports it outside the Gaussian model.
Progress (June 2026). The α–Synergy Bridge experiment (research/alpha-bridge/) operationalized α in the Gaussian setting as the mean squared round-trip projection distortion between observation subspaces. Result: beyond-pooling synergy is zero at α = 0 and increases monotonically with α, with a closed-form relationship. This discharges the first item above (operationalizing the “difference in kind”) in the Gaussian linear model — and only there: the September counterexample shows the monotone relation does not extend to couplings in general. The second item (productive vs. noisy asymmetry) and the third (calibrating κ) remain open. The bridge also supplies the missing connection for B4 Obligation 8 (research/demarcation-criterion/b4-theorem-scope.md).
The full framework is at thepulsegoeson.com.