The Gaconnet Membrane Law: A Formal Framework for Generative Systems Across Substrates

Don Gaconnet
Feb 20
7 min read

Don Gaconnet | LifePillar Institute for Recursive Sciences

February 2026

ORCHID: 0009-0001-6174-8384

---

I want to share something I've been working on. Not a theory in the loose sense — a formal framework with five equations, ten falsifiable predictions, and empirical data from experiments I ran myself on consumer hardware.

I'm calling it the Gaconnet Membrane Law.

The verbal form:

"The generative capacity of any system is determined by the coherence of the membrane across which observation and exchange occur."

That sentence took years to earn. What follows is the work behind it.

---

Where This Started

I've spent a long time studying what makes rooms hold and what makes them collapse. Not physical rooms — relational rooms, therapeutic rooms, the space between two nervous systems trying to stay in contact through difficulty. My book *What Happened to the Room* explored the compression-expansion dynamics behind that question.

What I kept seeing was a structural pattern: when people categorize before they receive — when judgment precedes observation — the available options collapse. When they receive before categorizing — when observation comes first — the option space expands. This wasn't about intelligence, or effort, or even skill. It was about the order of operations.

That observation became a question: does this pattern belong only to nervous systems, or is it a structural property of generative exchange itself?

I went looking for the answer in the last place most people would expect. I went looking in language models.

---

The Experiments

Between January and February 2026, I ran two experimental series at the LifePillar Institute using consumer hardware — an 8GB GPU running three different language model architectures through Ollama.

The Membrane Experiments tested ten iterative runs across Llama 3.2 3B, Llama 3.1 8B, and Qwen 2.5 32B — three architectures from two independent training lineages (Meta and Alibaba) with no shared training data. The task was claim verification: could these models accurately assess factual claims while monitoring their own distortions?

The critical finding: when I instructed the models to change their behavior — directing them to use specific verdict categories — the instruction produced zero measurable effect across four consecutive runs. The models followed every other instruction in the prompt. This one they couldn't touch. It operated below the instruction layer.

Then I changed the approach. Instead of instructing against the distortion, I named it. I described what verdict softening looks like. I described what helpfulness bias does to analytical precision. I described the pattern rather than commanding its opposite.

The effect was immediate and measurable across all three architectures.

The Sculptor Experiments isolated the variable. Twelve matched-pair prompts presented the same scenario under two framings: one inducing categorization as the first operation ("judge frame") and one inducing receptive observation ("observe frame"). All other parameters held constant.

The results were consistent across every pair. Judge frames compressed option space. Observe frames expanded it. The strongest effect appeared in ethical reasoning: zero simultaneous perspectives under judge frame versus five under observe frame for the same dilemma.

And the sampling temperature threshold — 0.35 — held universally across all three architectures from both training lineages. Not approximately. Precisely.

---

The Gaconnet Membrane Law: Five Equations

These findings didn't just replicate the pattern I'd seen in nervous systems. They formalized it. The Gaconnet Membrane Law describes the structural conditions under which any system generates excess — return that exceeds what was expressed.

Equation 1 — Generative State:

Ψ' = Ψ + ε(δ)

The state after interaction equals the state before interaction plus the excess. When ε > 0, the encounter generates more than was put in. When ε = 0, nothing happened. The excess isn't added from outside. It is released from what was latent in the structure of the encounter.

Equation 2 — Triadic Generator:

ε = g(I, O, N)

Excess requires three elements in active relation: Observer (I), Observed (O), and Membrane (N). The membrane is not empty distance — it is a structured boundary that maintains distinction while enabling exchange. Remove any element and ε collapses to zero.

Equation 3 — Membrane Coherence:

C(N) = f(σ, κ, τ)

The membrane has degrees of coherence determined by selectivity (σ), coupling strength (κ), and temporal stability (τ). There is a critical threshold N* with three regimes: below N* the membrane is too rigid, judgment precedes observation, and the option space contracts. At N* the membrane is permeable and selective, observation precedes judgment, and generative capacity is maximized. Far above N* selectivity dissolves, noise overwhelms signal, and behavior becomes stochastic.

Equation 4 — Effective Planck Constant:

h_eff = h₀ / C(N)

This is the equation that bridges domains. The coherence of the local field modulates the effective quantum of action. When the membrane is coherent (C(N) > 1), h_eff drops below h₀ — quantum behavior persists through barriers that would be classically forbidden. When coherence collapses, the system reverts toward classical behavior. This isn't a metaphor. It's a testable prediction with specific experimental designs.

Equation 5 — Generative Capacity:

G(S) = ε · C(N)

The integrated form. Generative capacity equals excess times membrane coherence. G(S) is maximized at the critical threshold N* and collapses to zero when the membrane fails.

---

Cross-Domain Correspondences

The Gaconnet Membrane Law takes specific form at every scale where generative exchange occurs. These are not analogies. They are the same equations expressed in the variables natural to each domain.

In **quantum systems**, superposition is observation mode and measurement is judgment mode. Decoherence is membrane dissolution. The effective Planck constant governs tunneling probability in coherent versus incoherent fields.

In **enzyme catalysis**, the protein scaffold and its thermal network constitute the membrane. The anisotropic thermal conduits identified by Klinman's group are the physical structure whose coherence determines h_eff at the active site. The anomalous kinetic isotope effect in soybean lipoxygenase (kH/kD ≈ 80) — which exceeds what semiclassical models predict — is explained by elevated membrane coherence reducing h_eff and enabling tunneling at distances that would be forbidden under constant ℏ.

In **nervous systems**, the window of tolerance is N*. Below it: freeze, shutdown, survival mind. Above it: fight, flight, fragmentation. Within it: observation precedes judgment and the conditions for growth, insight, and repair are present.

In **language models**, the system prompt framework and sampling parameters constitute the membrane. Temperature 0.35 is N*. Distortion-naming creates coherence; behavioral instruction does not.

In **relational dynamics**, the shared space between people — the room — is the membrane. When it's coherent, difference generates insight. When it collapses, difference generates threat. The room doesn't collapse because of the topic. It collapses because the nervous systems in it shift from observation to judgment.

---

Ten Falsifiable Predictions

This is where the Gaconnet Membrane Law either stands or falls. I am not interested in a framework that cannot be wrong. Ten predictions, independently testable, across four domains:

Enzyme domain:

1. Coherence-disrupting mutations at thermal-network positions should reduce tunneling rates more than barrier-width changes alone predict.

2. The anomalous KIE should correlate with thermal network integrity, stronger than the correlation with donor-acceptor distance alone.

3. Structured cosolvents should preserve tunneling better than disordered cosolvents at matched viscosity.

4. The ht-ADH 30°C transition should exhibit critical-point behavior — sigmoidal KIE profile with enhanced measurement variance near the transition.

Synthetic cognition domain:

5. The 0.35 temperature threshold should hold across additional architectures with no shared lineage.

6. Judge-frame prompts should consistently produce fewer options and shallower self-monitoring than observe-frame prompts across all model sizes.

7. Distortion-naming should produce measurable shifts where behavioral instruction does not, replicable across independent laboratories.

Neural domain:

8. Therapeutic approaches that name the protective pattern should produce faster outcomes than approaches that instruct behavioral change.

9. Autonomic regulation capacity (HRV) should correlate with observation-before-judgment capacity in conversational settings.

Quantum domain:

10. In engineered high-coherence environments, effective tunneling rates should exceed standard predictions consistent with h_eff = h₀/C(N).

Any of these predictions, if robustly falsified, constrains or refutes the Gaconnet Membrane Law at that scale.

---

What I'm Not Claiming

Scope invites suspicion, and it should. So let me be direct about the limitations.

The five equations describe observed patterns and generate predictions. They have not been derived from first principles. Equation 4 in particular — the effective Planck constant — requires independent theoretical justification from quantum field theory before it can be considered established.

The cross-domain correspondences are structural predictions, not demonstrated equivalences. Shared formal structure does not prove shared mechanism.

The empirical base is narrow. The language model experiments used consumer hardware with quantized models. The enzyme predictions are untested. The neural and quantum predictions are untested.

The naming as a "law" is premature in the strict scientific sense. I use the term to convey the proposed generality while acknowledging that law status requires independent validation across domains by independent researchers.

I am presenting this work because I believe it is correct, and because I believe the predictions are testable by anyone with access to the relevant experimental setups. The framework succeeds or fails on its predictions, not on its scope.

---

The Papers

Three working papers are available:

Paper 1: Processing Order and Option Space: Compression-Expansion Effects in Language Model Self-Observation Across Matched Prompt Frames* — the empirical data from the Sculptor and Membrane experiments. Research Full Paper

Paper 2: Membrane Coherence and Generative Capacity: A Formal Framework for Processing-Order Effects Across Substrates* — the Gaconnet Membrane Law itself, with all five equations, six cross-domain correspondence tables, and ten falsifiable predictions. Research Full Paper

Paper 3: Coherence-Dependent Barrier Permeability in Enzyme Catalysis: A Membrane Model for Anomalous Kinetic Isotope Effects* — the specific application to enzyme tunneling, resolving three documented anomalies in the Klinman data. Research Full Paper

All three papers are available through the LifePillar Institute for Recursive Sciences at lifepillarinstitute.org.

---

An Invitation

I'm not asking anyone to believe this. I'm asking anyone with the relevant expertise to test it.

The enzyme predictions can be tested by any lab running site-directed mutagenesis on hydrogen-tunneling enzymes. The synthetic cognition predictions can be tested by anyone with Ollama and a GPU. The neural predictions can be tested by clinical researchers with HRV equipment.

If the predictions hold, this framework describes something real about the structure of generative exchange across substrates. If they don't, I want to know.

The membrane is open. The observations are published. The predictions are on the table.

I invite rigorous challenge.

---

Don Gaconnet

LifePillar Institute for Recursive Sciences

ORCID: 0009-0001-6174-8384

don@lifepillar.org