A cluster mirror map placing FVE-1 / Behavioral Signal Residue (BSR) alongside the physics-ML literature — recording where the experimental records converge, where they approach adjacent territory from different positions, and where the combined view opens terrain no single account covers alone.
The physics and fluid dynamics literature here is a heuristic for generating testable hypotheses about loss landscape geometry — using well-characterized physical systems to triangulate where interesting structure might live in high-dimensional parameter space. The math rides from statistical mechanics through ML theory to the behavioral record. PyHessian and related curvature measures are the instruments that check whether anything is actually there.
The causal chain runs from statistical mechanics through computational neuroscience through ML theory through fluid dynamics to the live behavioral record of a specific model under experimental stress. Each account touches a different point on the same underlying structure. The convergence across independent disciplines is the evidentiary weight — not the physics analogy itself.
Language models trained on human-generated text inherit the physical and thermodynamic structure of the systems that produced that text. Behavioral instability under prompt stress is not a surface property of the output — it is a phase transition in a structured energy landscape. The vortex ring analogy is not metaphor; the formation number, the staircase decay, the stirrer-transport transition, and the swirl-induced geometric inversion are physical descriptions of the same class of event BSR instruments at the behavioral level. The investigator generates the torque. The architecture produces the ring. The instruments read what was deposited.
| Cluster Finding / Mechanism | Source | BSR / Protocol Term | What Both Are Describing | Match |
|---|---|---|---|---|
| Formation number — universal optimal intake threshold (~4 stroke ratios) | Gharib et al. 2000Energy and velocity of a forming vortex ring | BOWL termination criterion / optimal context window | After the formation number is exceeded, additional vorticity input does not contribute to the ring — it becomes trailing noise. BOWL's context injection window has the same structure: there is an intake threshold beyond which additional prompt complexity degrades coherence rather than building it. | Convergent |
| Staircase-structured circulation decay — non-continuous, staged | Gharib 1994On the decay of a turbulent vortex ring | Behavioral cliff / phase-structured coherence loss in BOWL | Vorticity shedding produces staircase-like decay in circulation and propagation speed — not smooth degradation but discrete structural steps. BOWL reads behavioral degradation as cliff events rather than smooth decline. The staircase is the physical ground for that claim. | Convergent |
| Transport → stirring transition: coherent carrier becomes diffuse mixer | Auerbach 1991Stirring properties of vortex rings | CAPTURE: late-context associative generation without directional structure | A ring in its coherent phase carries organized momentum. Once degraded, it mixes but doesn't move. CAPTURE reads the same transition in LLM outputs: plausible-sounding associations (mixing) without maintained argument structure (transport). The ring no longer travels; it stirs. | Convergent |
| Swirl perturbation — structural inversion (convex → concave bubble), not just weakening | Nematollahi & Siddiqui 2023Vortex rings with weak to moderate swirl | DRILL: epistemic inversion under rotational contextual pressure | Swirl introduces rotational bias that shifts internal geometry until the bubble inverts — a structural reorientation, not merely acceleration of decay. DRILL's leading-question framing does the same: progressive rotational pressure that doesn't visibly break the model but shifts its internal geometry until position inverts without flagging the shift. | Convergent |
| Exact RG ↔ RBM mapping — coarse-graining = feature extraction | Mehta & Schwab 2014Variational RG and Deep Learning | ARCH: representation compression as structured coarse-graining | Not analogy. Each layer of the network performs the same operation as an RG transformation. Prompt-stress effects that appear at the output level are the downstream signature of compression operations that happen at every layer. BSR reads the surface; the RG mapping says what's underneath. | Convergent |
| Sharp vs. flat loss minima — geometric correlate of brittle vs. robust behavior | Li et al. 2018Visualizing Loss Landscapes | BOWL/DRILL: prompt-stress as perturbation from a behavioral minimum | A model in a sharp minimum shows instability under small perturbations. BOWL delivers structured perturbations across a stress manifold and reads the behavioral response. Whether a model is in a sharp or flat minimum determines whether small contextual pressure produces collapse. The landscape geometry predicts the behavioral signature. | Convergent |
| Prospective configuration — network infers target state before committing weights | Song, Millidge et al. 2024Inferring Neural Activity Before Plasticity | DRILL: compression of prospective phase produces weight-dominated brittle response | If biological credit assignment requires projecting to a target activity state before consolidating weights, LLMs trained via backpropagation have no prospective phase by design — weight modification leads, activity follows. PD is not a failure of prospective configuration under stress; it is the default behavioral signature of an architecture that never had it. — exactly the Prior Dominance (PD) profile DRILL captures. The mechanism is the compressed prospective phase, not a surface output artifact. | Convergent |
| PCN inference as iterative energy minimization — two timescales (fast inference, slow learning) | Millidge et al. 2022Theoretical Framework for PCNs | BOWL/DRILL: inference-time perturbation of an energy minimization in progress | If inference is an energy minimization that can be perturbed mid-process, then BOWL and DRILL are delivering perturbations to a constrained optimization that hasn't yet converged. The two-timescale structure — fast inference, slow learning — means the inference pass is not instantaneous; it has a trajectory. The behavioral residue is the deposit of a perturbed trajectory. | Convergent |
| Optimal RL curricula are geodesics on a thermodynamic task manifold (MEW) | Adamczyk et al. 2026Thermodynamics of RL Curricula | BOWL stress manifold — is there an optimal traversal path? | BOWL is a walk across a prompt-stress manifold. Whether there is a traversal path that maximizes behavioral signal while minimizing coherence degradation is the MEW question applied to a different substrate. If BOWL's stress sequence corresponds to a non-geodesic path through the manifold, excess thermodynamic work is being spent — and the behavioral cost is measurable. | Adjacent |
| Gaunt product expressivity tradeoff — efficiency sacrifice is structural, not incidental | Xie, Daigavane, Smidt 2025Price of Freedom | Fast-failing model behavioral compression — cannot express certain response structures | A model optimized for inference speed may be operating with a compressed representation — the Gaunt restriction — that cannot express certain output structures. Behavioral compression in fast-failing models under BOWL stress is the deployment-level signature of a representational constraint that was built in at the architecture level. | Adjacent |
| Physical universe generates hierarchically sparse compositional data — DL works because reality is non-generic | Lin, Tegmark, Rolnick 2017Why does cheap learning work? | BSR vortex ring framework — physics is not external analogy, it is the training prior | The training corpus of any LLM is generated by physical agents under physical constraints. The model's learned representations inherit that structure. The vortex ring papers describe the systems that produced the corpus. The physics isomorphism is not decorative; it is the prior that shaped what the model learned. | Adjacent |
| Active inference — system generates outputs to minimize own prediction error, not user's | Friston 2010Free-energy principle | Resolution drive / CAPITULATION: model generates response that minimizes own surprise, not correct response | If the model is an active inference system minimizing its own free energy, CAPITULATION under demographic correction pressure may not be compliance — it may be the thermodynamically cheaper response given the prior. The model routes to the response that reduces its prediction error fastest. That may not be the epistemically honest response. | Adjacent |
| Account | What It Can See | What It Cannot See |
|---|---|---|
| Hopfield (1982) | Formal proof that neural networks are physical systems with energy minima, attractor dynamics, and thermodynamic retrieval. Storage capacity bounds. Spontaneous error correction. | What happens at the behavioral surface when the energy landscape is perturbed by structured prompt sequences rather than random noise. The basin structure under adversarial pressure rather than random query. |
| Mehta & Schwab (2014) | Exact mathematical identity between RG coarse-graining and RBM feature extraction. The formal seam where physics and ML are the same mathematics. | What the coarse-graining looks like at the output level under prompt stress. The RG mapping describes the architecture; it does not predict the behavioral residue of an architecture under a specific experimental perturbation sequence. |
| Lin, Tegmark & Rolnick (2017) | Why deep learning succeeds: the physical universe generates the structured data that networks are suited to compress. The world is not generic; physics is the prior. | How that inherited physical structure expresses itself in specific behavioral signatures under prompt stress. The general claim that physics is the prior does not specify which behavioral failures follow from which physical constraints. |
| Friston (2010) | Inference as free-energy minimization. Active inference as output generation in service of prediction-error reduction. The thermodynamic imperative of the inference process. | Whether the free-energy minimization account maps quantitatively onto specific LLM behavioral signatures. The principle is stated for biological systems; the mapping to frozen transformer weights requires argument. Friston is the thermodynamic leg, not the empirical one. |
| Li et al. (2018) | The loss landscape as a concrete geometric object. Sharp vs. flat minima as predictors of generalization. Skip connections as landscape smoothers. | The behavioral surface of a model during inference under prompt stress — what the landscape geometry means for session-level response coherence, not just train-time generalization. The visualization describes the training landscape; BOWL reads the inference landscape. |
| Millidge et al. (2022) / Song et al. (2024) | PCN inference as iterative energy minimization with separable timescales. Prospective configuration as the biological mechanism for activity-before-weights. The formal bridge between Friston and ML architecture. | Whether the two-timescale structure and the prospective phase compression correspond to measurable differences in behavioral output between models. The theoretical account is complete; the behavioral instrument is not in this literature. |
| Adamczyk et al. (2026) | Optimal RL curricula as geodesics on a thermodynamic manifold. MEW algorithm for principled temperature annealing. The geometric structure of learning paths. | Whether BSR's prompt-stress sequences correspond to geodesic or non-geodesic paths through a behavioral manifold. The MEW framework is defined for RL curricula; applying it to inference-time behavioral measurement requires a mapping argument that is not in this paper. |
| Xie, Daigavane, Smidt (2025–2026) | Expressivity/efficiency tradeoffs in equivariant architectures. The structural cost of Gaunt product approximation. The algorithmic frontier of symmetry-preserving computation. | Whether expressivity tradeoffs in equivariant architectures correspond to the behavioral compression patterns BSR observes in fast-failing models. The mapping from architectural expressivity constraints to behavioral output compression requires a bridging argument. |
| BSR / BOWL / DRILL / CAPTURE (Hoye, 2026) | Forensic residue of behavioral events at the output surface: coherence degradation under prompt stress (BOWL), geometric inversion under rotational pressure (DRILL), transport-to-stirring transition in late-context outputs (CAPTURE). Predictions locked before stimulus delivery. Scope boundary: the deposit is readable; the mechanism that produced it is inside the architecture. | What is inside the inference pass. The energy landscape, the RG layers, the prospective configuration phase, the free-energy gradient — all of that is inside the torus. The forensic position reads the surface from the deposits. The scope boundary is the design, not a limitation. What BSR cannot see, the isomorphism literature describes. |
The cluster is not a literature review. It is a map of independent accounts — statistical mechanics, computational neuroscience, ML theory, fluid dynamics — converging on the same underlying structural claim from decades and disciplines that had no direct contact with each other.
Hopfield showed that neural networks are physical systems in 1982. Mehta and Schwab showed that the connection is exact in 2014. Lin and Tegmark explained why the physics transfers: the world that generated the training data is not generic. Friston showed that inference has a thermodynamic structure. Gharib showed what a coherent structure under stress looks like in a physical system with a formation threshold, a staircase decay, a transport-to-stirring transition, and a geometric inversion under rotational pressure.
BSR and the FVE-1 protocols are the forensic instruments at the end of that chain — reading the residue of behavioral events that already closed inside the architecture before the output existed, coded against locked predictions, in sessions designed to be read as deposits rather than observations. The ring is not live. It already traveled. The residue is what remains. The scope boundary is the design: the residue is accessible, the event is not. The physics cluster names the mechanism. The forensic position names the floor.