Introduction

Via Balaena LLC / CortenForge

Status: Three principles validated, two boundary conditions established Date: April 16, 2026 Classification: Open research — intended for public release

This document connects biological locomotion control theory to the unsolved x-encoding problem in thermodynamic computing. The code that produced every result is in the same repository.

We tested five biological navigation principles on Ising chain models under Langevin dynamics. Three produced quantitative design rules. Two failed — and the failure pattern reveals a sharp boundary on which biological inspiration transfers and which does not.

Results at a Glance

Three design rules for coupled bistable circuits under Langevin noise:

Two boundary conditions — what NOT to try:

The dividing line: Statistical-mechanical questions (noise tuning, phase coordination, extensivity) transfer from biology to thermodynamic circuits. Dynamical-systems questions (topological invariants, bifurcation sensitivity) do not. This boundary tells an engineer which biological literature to mine and which to ignore.

The validated design rules are actionable: an engineer building a coupled bistable circuit can read the tables above and change what they build tomorrow. The boundary conditions are equally actionable: they tell the same engineer which approaches to skip.

Thermodynamic Computing and the Y Problem

Thermodynamic computing is a paradigm in which computation is performed not by enforcing deterministic logic states, but by allowing a physical system to relax toward thermodynamic equilibrium, reading the resulting probability distribution as the output. The key entities are:

• The energy function F: defines the shape of the target probability distribution over output states
• Y: the target distribution itself — the answer the system is supposed to produce
• The physical substrate: a stochastic system (resistor-inductor-capacitor networks, analog probabilistic circuits, or future purpose-built thermodynamic chips) whose natural dynamics under Langevin noise are governed by F

The state of the art as of 2025-2026 includes Normal Computing's CN101 — the world's first taped-out thermodynamic semiconductor chip — and Extropic's XTR-0 development platform, both of which demonstrate that thermodynamic sampling units (TSUs) can perform AI inference tasks including matrix inversion and Gaussian sampling at energy efficiencies orders of magnitude beyond conventional GPUs.

The Y problem, to first approximation, is solved. Energy-based models provide a formal language for defining Y. Boltzmann statistics guarantee that a system in thermal equilibrium will sample from Y. The physics of the output is understood.

The X Problem — The Unsolved Root

The unsolved problem is X: the input encoding.

In a classical digital computer, encoding an input X is trivial — you set a voltage high or low. The physical act of encoding is decoupled from the physics of computation. In a thermodynamic computer, there is no such decoupling. The input X must be encoded as the initial conditions or boundary conditions of a physical stochastic system. The system then relaxes — through a nonequilibrium trajectory governed by Langevin dynamics — toward (hopefully) the correct distribution Y.

The problem is that this relaxation trajectory is sensitive to how X is injected. An imperfect encoding perturbs the energy landscape, potentially steering the system toward the wrong basin of attraction. At small scale, this can be managed empirically. At scale — as the circuit grows in size and the noise profile shifts — there is no formal theory telling an engineer how to inject X such that the relaxation remains correct. Every scaling step requires empirical re-tuning.

This is not a materials problem or a fabrication problem. It is a theory problem: we lack a formal design language for X-encoding in nonequilibrium stochastic systems. Until that language exists, thermodynamic computing cannot be engineered the way digital computing is engineered — from first principles, with predictable scaling behavior.

The Deeper Root: The Gap in Nonequilibrium Statistical Physics

The theory problem traces to a gap in physics. Equilibrium statistical physics is well-understood. Landauer's bound gives the minimum energy cost of erasing a bit. Boltzmann statistics describe the equilibrium distribution. But thermodynamic computers operate far from equilibrium — they must complete a computation quickly, which means they cannot afford to wait for true equilibrium to be reached.

The stochastic thermodynamics of far-from-equilibrium systems is a field in active development. Thermodynamic uncertainty relations now provide bounds on computation speed, noise level, and energy cost. But a constructive theory — one that tells you how to build a nonequilibrium system that reliably encodes X and converges to Y — does not yet exist.

The analogy: we know the speed limit (Landauer's bound, uncertainty relations), but we have no road map.

The Core Hypothesis

The hypothesis is that biological evolution has already solved versions of the X-encoding problem across a continuous spectrum of operating regimes, and that the biological solutions can be formally mapped to engineering principles for thermodynamic circuit design.

Specifically: every organism that navigates a chaotic physical medium — fluid, air, a stochastic chemical gradient — is solving a version of the same problem. It must inject "intent" (a direction, a target, a behavioral goal) into a noisy physical system (its own body in a turbulent medium) and achieve reliable convergence to the desired output state, at some throughput level, with finite energy. The physics of the problem is structurally isomorphic to X-encoding in thermodynamic computing.

The biological world has explored this design space for hundreds of millions of years. We should read the solutions it found.

The Axis: Reynolds Number

Why Reynolds Number Is the Right Axis

The Reynolds number (Re) is a dimensionless ratio of inertial to viscous forces in a fluid:

$$ \text{Re} = \frac{\rho U L}{\mu} $$

where ρ is fluid density, U is velocity, L is characteristic length, and μ is dynamic viscosity.

It is the natural organizing axis for biological locomotion strategies because it determines what physics dominates at a given scale and speed. At low Re, viscosity dominates — the fluid has no memory of past motion, and every perturbation is immediately damped. At high Re, inertia dominates — perturbations persist as vortices, wakes, and turbulent structures that interact with the swimmer on timescales longer than the swimmer's own motion.

Crucially for our purposes: the ratio of signal timescale to noise timescale changes across the Re spectrum in a way that maps directly onto the thermodynamic X-encoding problem. At low Re, signal and noise are on the same timescale — you cannot separate them and must encode in the topology of motion. At intermediate Re, both are relevant simultaneously — you need multimodal switching. At high Re, noise generates coherent structures that can be exploited — speed itself becomes a source of control authority.

The Re spectrum spans roughly 13 orders of magnitude in biology, from bacteria at Re ~ 10⁻⁵ to blue whales at Re ~ 10⁸. The locomotion strategies are not a continuum — they are discrete regimes separated by qualitative phase transitions in the physics.

Why Three Is Not the Golden Number

The initial framing identified three exemplars: octopus (inertial, distributed computation), dragonfly (predictive internal models), peregrine (vortex-noise coupling). The full spectrum contains at least five qualitatively distinct regimes, each with a different dominant strategy. Three of the five have been experimentally validated on Ising chain models (see the Validated Results section); the other two are hypothesized but untested.

The transition zones between regimes may be as important as the regimes themselves — real thermodynamic circuits will operate across throughput ranges, and failure modes will likely occur at the transitions.

Noise Tuning: E. coli Stochastic Resonance

Re < 1 — The Viscosity-Dominated Regime

The Physics

At Re < 1, the Navier-Stokes equations simplify to the Stokes equations — linear, time-reversible, with no inertial terms. The fluid has no memory. A swimmer moving forward and then backward through the exact same sequence of shapes returns to exactly its starting position. This is Purcell's scallop theorem (1977): in a time-reversible fluid, any reciprocal motion (one that looks the same played forward and backward) produces zero net displacement. It does not matter how fast or slow the motion is performed — speed is irrelevant.

The implication is profound: at low Re, you cannot encode information in the amplitude or timing of a perturbation. Amplitude scales out. Timing scales out. The only thing that survives is the topology of the motion sequence — whether it traces a closed loop in configuration space that encloses a nonzero area. This is geometric phase, formalized by Shapere and Wilczek (1989) using the language of gauge theory and fiber bundles.

What E. coli Does

E. coli navigates chemical gradients (chemotaxis) at Re ~ 10⁻⁵ using the run-and-tumble strategy. Multiple flagellar motors rotate counter-clockwise to bundle the flagella into a helical propeller (a "run" — straight motion). When one or more motors switch to clockwise rotation, the bundle unbundles and the cell reorients randomly (a "tumble"). By extending runs when moving up a chemical gradient and shortening them when moving down, the cell executes a biased random walk toward attractants.

The critical insight: the CheY-P signaling molecule that controls motor switching follows Langevin dynamics — it has intrinsic stochastic fluctuations described by:

$$ \frac{dY}{dt} = -\frac{Y - Y_0}{\tau_Y} + \eta(t) $$

where η(t) is Gaussian white noise. Crucially, signaling noise enhances chemotactic drift. An intermediate level of noise in the slow methylation dynamics improves gradient-climbing performance — not despite the noise, but because of it. The noise is not a contaminant to be suppressed; it is a functional component of the control strategy.

Furthermore, E. coli achieves signal amplification of more than 50-fold: a 2% change in receptor occupancy produces a 100% change in motor output. This nonlinear amplification emerges from the cooperative structure of the receptor-kinase complex operating near a phase transition — a bifurcation point at which sensitivity is maximized.

The X-Encoding Principle

Principle 1: In the viscosity-dominated regime, amplitude and rate are irrelevant. Information must be encoded in the sequence structure — the topology — of the injection. A closed loop in the injection parameter space that encloses nonzero area produces net displacement in the output state; a reciprocal sequence produces nothing.

Applied to thermodynamic circuits: This principle predicts that at very low throughput rates, X-encoding must rely on injection sequence topology rather than amplitude. However, testing showed this does NOT transfer to the Langevin domain — amplitude is a direct lever in time-irreversible systems (see Level 4 below). The scallop theorem requires time-reversibility, which Langevin dynamics lack.

Principle 2: Stochastic resonance is available. An optimal noise level enhances encoding fidelity by improving sensitivity near the bifurcation point. The circuit should be tuned to operate in this regime rather than attempting to minimize noise.

Experiment 1 — E. coli Stochastic Resonance in Chemotactic Navigation

Scientific Question

Does an intermediate noise temperature maximize a Langevin particle's ability to follow a periodic signal in a biased double-well potential? Does the optimal noise level shift predictably with gradient strength?

This directly tests Principle 2 (stochastic resonance enhances encoding fidelity) and extends prior single-particle SR validation by adding a symmetry-breaking chemical gradient via ExternalField.

Langevin Model of Chemotaxis

The experiment maps E. coli chemotaxis onto the Langevin framework:

The particle "runs" (stays in one well) and "tumbles" (crosses the barrier). Stochastic resonance occurs when the noise-driven switching rate matches the signal frequency — the Kramers rate (the thermally activated barrier-crossing rate, which scales as exp(−ΔV/kT)) at kT ≈ 1.0 equals the signal angular frequency ω = 2π × 0.01214.

Experimental Design

Synchrony metric:

synchrony = sign(qpos[0]) · cos(ω·t)

Positive when the particle occupies the correct well at the correct phase of the signal. Averaged per step over an episode.

Part A — Baseline temperature sweep: 30 kT multipliers log-spaced in [0.1, 5.0], 20 episodes each. Establishes the SR curve and locates the peak. Repeated at gradient strengths h ∈ {0.0, 0.3} to test whether the gradient shifts the optimum.

Part B — Scientific controls:

• Low noise (kT × 0.1): particle trapped in one well, no switching → zero synchrony
• High noise (kT × 5.0): random switching, no correlation with signal → zero synchrony
• No signal (A₀ = 0): metric validates to zero even at the SR-optimal temperature

Part C — Multi-algorithm training: Three algorithm classes on the gradient-biased setup (h = 0.3):

• CEM (evolutionary, gradient-free) — sim-rl
• PPO (policy gradient) — sim-rl
• SA (simulated annealing) — sim-opt

Each trains a linear policy (2 inputs → 1 output, mapping particle position and velocity to a temperature control signal) for 100 epochs on a 32-environment parallel batch with ctrl-temperature. Evaluation: 20 deterministic episodes with the trained policy.

Gate System

Controls use the 3σ threshold: |mean| < 3·stderr.

Results

Controls (validated):

Baseline SR sweep (validated):

The SR curve shows the predicted bell shape: zero synchrony at low noise (particle trapped), rising to a clear peak at kT ≈ 1.7 (noise-driven switching matches signal frequency), falling back to zero at high noise (random switching destroys correlation).

1-particle multi-algorithm training: Skipped (superseded by Level 3 multi-particle validation below).

Code: sim/L0/therm-env/tests/experiment_1.rs

Level 3 — Ising Chain Stochastic Resonance

Does stochastic resonance scale from 1 particle to coupled multi-cell systems? This is the direct test of Principle 2 on a thermodynamic circuit proxy.

The Setup

Four particles, each in a double well with its own oscillating field, connected by nearest-neighbor coupling:

Particle 0 ←J→ Particle 1 ←J→ Particle 2 ←J→ Particle 3
  [±x₀]          [±x₀]          [±x₀]          [±x₀]
  + signal       + signal       + signal       + signal

Reward: Average synchrony across all particles: (1/N) × Σ sign(qpos[i]) × cos(ωt)

Scientific question: How does coupling strength J shift the SR-optimal noise temperature? This is the τ_circuit / τ_noise characterization.

The Experiment

Phase 0 — Scout (~30 sec): J=2 only, 8 kT points in [1, 15], 10 episodes. Validates that the kT=15 ceiling captures the strongest-coupling peak.

Phase 1 — Temperature sweep at 5 coupling strengths (~35 min): For each J ∈ {0.0, 0.5, 1.0, 1.5, 2.0}, sweep 25 temperatures log-spaced in [0.1, 15.0], 40 episodes each. Maps the SR curve at each coupling strength. Gates: every J must have a significant (|t| > 2.708, df=39, α=0.01) interior peak.

Phase 2 — Multi-algorithm training at each J: CEM, SA (simulated annealing), and Richer-SA (SA with adaptive neighborhood) each train a linear policy (8 inputs → 1 output) for 100 epochs on a 32-environment parallel batch. Tests whether gradient-free agents independently discover the SR-optimal temperature at each coupling strength. PPO was dropped — policy gradient methods compute per-timestep advantages, fundamentally wrong when the optimal policy is a constant temperature.

Controls (same pattern as Level 2): Low noise (kT × 0.1), high noise (kT × 10), no signal (A₀=0) — all must show zero synchrony.

Key design decisions:

• kT range [0.1, 15.0]: Coupling raises the effective barrier for individual particle switching (interior: ΔV_eff = 3 + 4J, end: 3 + 2J). Using the calibration ratio ΔV/kT_peak ≈ 1.39, the J=2 peak is predicted at kT ≈ 6.5. Range of 15 gives 2.3× headroom.
• Training env k_b_t = 15.0: The policy's tanh output [0, 1] maps to kT ∈ [0, 15.0], letting agents reach all predicted peaks (kT 2–7) at ctrl ≈ 0.13–0.47.
• 40 episodes/point: For the weakest signals (sync ≈ 0.04, σ ≈ 0.06), stderr = 0.06/√40 = 0.0095, giving |t| = 4.2. Well above the 2.708 threshold.

Results

Scout (validated):

Peak at kT=6.5 (interior, |t|=4.17), exactly matching the effective-barrier prediction. Range [0.1, 15.0] confirmed safe.

v1 sweep (invalid — range too narrow):

| J | peak kT | peak sync | |t| | Status | |---|---|---|---|---| | 0.00 | 2.16 | 0.075 ± 0.014 | 5.47 | Valid (interior) | | 0.10 | 5.00 | 0.043 ± 0.010 | 4.50 | Boundary | | 0.50 | 1.64 | 0.039 ± 0.020 | 1.95 | Not significant | | 1.00 | 5.00 | 0.054 ± 0.008 | 6.74 | Boundary | | 2.00 | 3.78 | 0.049 ± 0.018 | 2.68 | Not significant |

Only J=0 was valid. Peaks for J≥0.1 hit the kT=5 ceiling or were below significance with 15 episodes.

v2 sweep (validated):

All 5 coupling strengths produce significant interior peaks. 5,000 episodes, 35 minutes, single run.

| J | peak kT | peak sync | |t| | Status | |---|---|---|---|---| | 0.00 | 2.29 | 0.043 ± 0.008 | 5.55 | PASS | | 0.50 | 2.29 | 0.057 ± 0.008 | 7.63 | PASS | | 1.00 | 2.82 | 0.065 ± 0.008 | 8.55 | PASS | | 1.50 | 2.29 | 0.057 ± 0.012 | 4.89 | PASS | | 2.00 | 4.29 | 0.053 ± 0.007 | 7.23 | PASS |

Two-regime behavior: Coupling does not simply shift the SR peak — it reveals two switching modes with a crossover:

• Weak coupling (J ≤ 1.0): The SR peak stays near kT ≈ 2.3 (same as uncoupled) but synchrony increases with coupling (0.043 → 0.065). Coupling enhances the existing resonance. Individual particle switching dominates — each particle crosses the barrier independently, and coupling merely coordinates them.
• Strong coupling (J = 2.0): The peak shifts to kT ≈ 4.3. The higher effective barrier (ΔV_eff = 3 + 2J to 3 + 4J for end/interior particles) requires more noise to drive transitions. The collective switching mode — where coupling forces particles to flip together — takes over.

The crossover between these regimes lies between J = 1.5 and J = 2.0. At J = 1.5, the data shows a primary peak at kT = 2.29 with a secondary bump near kT = 4.3–5.3, suggesting both modes are active and competing.

The effective-barrier model predicted the J = 2 peak at kT ≈ 6.5. The actual peak is at kT ≈ 4.3 — correct direction, right order of magnitude, but the simple model overestimates the barrier because cooperative switching lowers the effective barrier relative to independent switching.

Design Rule (Noise Tuning)

For an N=4 Ising-coupled bistable circuit with coupling strength J:

• J < 1.5: Operate at kT ≈ 2.3. Coupling enhances SR without shifting it. Stronger coupling gives better signal fidelity (up to ~50% improvement at J = 1.0).
• J ≥ 2.0: Operate at kT ≈ 4.3. The collective switching mode dominates and requires higher noise.
• J ≈ 1.5–2.0: Transition zone. Both modes active. Operating at either kT ≈ 2.3 or kT ≈ 4.3 gives similar fidelity.

Operating at the wrong temperature degrades synchrony to noise-floor levels.

Operating Envelope: Barrier Height Tolerance

The Noise Tuning Rule maps the kT axis of the design surface. A complementary sweep maps the ΔV axis: fix kT=2.0 and sweep barrier height ΔV from 0.5 to 10.0 at two signal amplitudes (A₀=0.3 and A₀=0.1), N=4 uncoupled particles (J=0), 40 episodes per point.

Key finding: the ΔV axis is forgiving. Unlike the sharp SR peak on the kT axis, the synchrony-vs-ΔV curve is a broad plateau:

The weak signal (A₀=0.1) was below detection threshold at all ΔV values, setting a minimum detectable signal floor for this kT.

Engineering implication: The design surface is asymmetric. Temperature requires precision tuning (sharp peak, ±30% of optimal degrades to noise floor). Barrier height is tolerant — any ΔV/kT between 0.25 and 2.75 gives similar performance. This means: tune kT carefully to the design rule above; ΔV can be approximate.

The trapping cutoff at ΔV/kT ≈ 3 is a hard constraint: if the barrier exceeds 3× the thermal energy, the circuit cannot respond to signals regardless of other parameters.

Why Not Train Agents to Find the Peak?

Phase 2 trained gradient-free agents to discover the SR-optimal temperature from dynamics alone. CEM(J=0) converged to kT ≈ 0.07 — a local optimum that games the synchrony metric rather than finding the SR peak at kT ≈ 2.3. SA(J=0) stalled for 80+ epochs.

The likely cause: a linear policy mapping 8 particle observables to a scalar temperature lacks the capacity to represent the nonlinear relationship between circuit state and optimal noise level.

This does not weaken the design rule. The sweep data directly maps the optimal kT for each coupling strength — an engineer doesn't need an agent to discover this; the rule is the table above. Agent-based discovery becomes relevant when the circuit topology is unknown or changes at runtime, a harder problem deferred to future work.

Code: sim/L0/therm-env/tests/ising_chain.rs

Level 4 — Topological Encoding Test (Principle 1)

Does injection sequence topology matter more than amplitude? In the biological E. coli regime (Re < 1), the scallop theorem guarantees that amplitude and rate are irrelevant — only the topology of the motion sequence determines net displacement. Principle 1 claims this extends to thermodynamic circuits: information must be encoded in sequence structure, not amplitude.

The Test

Fix J=1.0, kT=2.82 (P2 optimal). Compare two conditions:

• Condition A (topology): Metachronal injection δ=0.66 (P4 optimal phase lag) at baseline amplitude A₀=0.3
• Condition B (amplitude): Synchronized injection δ=0 at doubled amplitude A₀=0.6

If topology beats doubled amplitude, Principle 1 transfers to the Langevin domain.

Additionally: sweep synchronized amplitude from A₀=0.3 to 2.0 to find the crossover point where brute-force amplitude matches the topological advantage.

Results

Principle 1: NOT VALIDATED. Amplitude dominates in the Langevin domain.

80 episodes per condition.

Gates:

Why it fails — and why that's informative:

The scallop theorem applies at Re < 1 where the governing equations are linear and time-reversible. In that regime, amplitude literally cancels out of the physics. The Langevin domain has no such constraint: the oscillating field directly biases the particle's switching rate, and a stronger field produces proportionally more switching. Synchrony scales nearly linearly with amplitude (0.051 → 0.088 → 0.290 across the 0.3–2.0 range).

At matched amplitude (A₀=0.3), metachronal and synchronized injection produce identical synchrony (0.050 vs 0.051). The Injection Timing metachronal advantage (18–37% at certain J values) is a coupling-coordination effect that emerges from specific parameter combinations, not a universal topology-dominance principle.

What this means for thermodynamic circuit design:

In the Langevin domain, amplitude IS a design lever. Engineers can — and should — use signal strength to improve encoding fidelity. The topological encoding principle from Stokes-regime biology does not transfer to systems where the dynamics are nonlinear and time-irreversible. This is a boundary condition on the biological analogy, not a weakness: it tells us precisely where the analogy holds (noise tuning, phase coordination) and where it breaks (amplitude scaling).

Design Rule (Principle 1)

Not applicable in the Langevin domain. Amplitude modulation is effective and scales linearly with synchrony. Use it.

For topology-sensitive encoding, the physics must constrain amplitude from the equations — the Stokes-regime scallop theorem does this, but Langevin dynamics do not. Principle 1 remains valid for systems where the scallop theorem applies (low Re fluids, certain linear circuit topologies), but does not generalize to the broader thermodynamic circuit architecture.

Code: sim/L0/therm-env/tests/ising_chain.rs — ising_topological_sweep

Key References

• Purcell, E.M. "Life at Low Reynolds Number." American Journal of Physics 45 (1977)
• Shapere, A. & Wilczek, F. "Geometry of Self-Propulsion at Low Reynolds Number." Journal of Fluid Mechanics 198 (1989)
• Flores, M. et al. "Signalling Noise Enhances Chemotactic Drift of E. coli." Physical Review Letters 109 (2012)
• Mattingly, H.H. et al. "Escherichia coli Chemotaxis is Information Limited." Nature Physics 17 (2021)

Injection Timing: Ctenophore Metachronal Coordination

Re ~ 1-1000 — The Intermediate Regime

The Physics

This is the most neglected and arguably most important regime. Both inertia and viscosity play significant and comparable roles simultaneously. There is no clean simplification. The governing equations are the full nonlinear Navier-Stokes, but at scales where both viscous and inertial terms contribute non-negligibly to thrust and drag. Crucially: thrust and drag are governed by different forces — at Re ~ 10², thrust is generated primarily by inertial forces while drag is still significantly viscous. No single strategy dominates.

The scallop theorem still technically applies (reciprocal motion produces no net displacement), but the breakdown of the theorem begins in this regime as inertial effects introduce memory into the fluid. A swimmer leaving a wake disturbs the flow it will encounter on its return, breaking exact time-reversibility.

What Ctenophores Do

Ctenophores are among the oldest animals on Earth and the largest animals that use cilia to swim. Their ctene rows (arrays of fused cilia) beat in a metachronal wave — sequential, coordinated beating with a phase lag between adjacent appendages, creating the appearance of a traveling wave moving down each row. This strategy is notably distinct from either the purely viscous strategy of small ciliates or the inertial undulation of fish.

The metachronal wave achieves several things simultaneously: it creates non-reciprocal motion (satisfying the kinematic constraint of the scallop theorem while inertia begins to provide memory), it generates hydrodynamic interactions between adjacent paddles that increase efficiency beyond what any single paddle could achieve alone, and it provides omnidirectional maneuverability by independently modulating different ctene rows.

Critically: ctenophores can perform tight turns while maintaining forward swimming speeds close to 70% of their maximum — a performance metric comparable to or exceeding vertebrates with far more complex locomotor systems. This is multimodal switching in action: the system does not stop and reorient, it redistributes effort across different control surfaces in real time.

What Water Boatmen Do

Water boatmen (Corixidae, Re ~ 10-200) operate in the most confused part of the intermediate regime. They use drag-based paddling with asymmetric power and recovery strokes — a direct application of temporal asymmetry to break time-reversibility. Their energetic efficiency is lower than fish of comparable size, but they are trimodal: they can swim, walk, and fly, transitioning rapidly between locomotor modes as the physical environment demands. This mode switching is the characteristic strategy of the intermediate regime.

The X-Encoding Principle

Principle 3: In the intermediate regime, no single encoding strategy dominates. The optimal approach combines temporal asymmetry (power and recovery phases at different rates) with distributed spatial coordination (metachronal waves) and mode switching (dynamically selecting between encoding strategies as local conditions shift).

Applied to thermodynamic circuits: At intermediate throughput rates — fast enough that purely topological encoding is insufficient, slow enough that inertial coupling effects haven't emerged — the X-encoder must operate as a multimodal adaptive system. It should maintain multiple encoding strategies simultaneously and switch between them based on local observables of the circuit's stochastic state. This is the regime where a Kalman filter architecture is most directly applicable: continuously estimating which encoding mode is appropriate given the observed noise floor and relaxation rate.

Principle 4: Distributed coordination across multiple control points (metachronal wave) produces emergent efficiency that no single control point could achieve. Applied to multi-cell thermodynamic circuits: coordinating injection timing across circuit nodes with a phase lag (analogous to a metachronal wave) may produce higher encoding fidelity than simultaneous or independent injection.

Experiment — Metachronal Phase-Lag in Ising Chain (Principle 4)

Scientific Question

Does phase-lagged injection produce higher per-node synchrony than synchronized injection in a coupled Ising chain? This directly tests Principle 4: whether distributed coordination across multiple control points produces emergent efficiency beyond what simultaneous injection achieves.

The Setup

The same N=4 Ising chain from the Noise Tuning experiment, but with phase-shifted oscillating fields:

Particle 0 ←J→ Particle 1 ←J→ Particle 2 ←J→ Particle 3
  signal(φ₀)    signal(φ₁)    signal(φ₂)    signal(φ₃)

Each particle's signal has phase φᵢ = -i × δ, creating a traveling wave from particle 0 → 3. The reward measures local synchrony: each particle is scored against its own phase-shifted signal, then averaged.

Sweep: 20 phase lags δ ∈ [0, π] at 4 coupling strengths (J = 0, 0.5, 1.0, 2.0), each at its SR-optimal kT from the Noise Tuning temperature sweep. 80 episodes per point. Total: 6,400 episodes, 46 minutes.

Gate system:

Results

| J | kT | δ* | peak sync | sync(δ=0) | improvement | |t| | |---|----|----|-----------|-----------|-------------|-----| | 0.00 | 2.29 | 1.49 | 0.051 ± 0.005 | 0.047 | +9.1% | 9.60 | | 0.50 | 2.29 | 0.66 | 0.061 ± 0.007 | 0.051 | +17.9% | 9.31 | | 1.00 | 2.82 | 0.66 | 0.056 ± 0.006 | 0.041 | +36.8% | 9.48 | | 2.00 | 4.29 | 0.00 | 0.058 ± 0.007 | 0.058 | +0.0% | 8.07 |

Interpretation:

• J=0 (uncoupled control): The curve is flat — uncoupled particles don't care about phase lag. The nominal "peak" at δ=1.49 is noise on a flat distribution, confirmed by Gate 1's regression test.
• J=0.5–1.0 (moderate coupling): Phase-lagged injection at δ ≈ 0.66 (≈ π/5) produces 18–37% higher synchrony than synchronized injection. The optimal δ* is the same at both coupling strengths — a stable design parameter.
• J=2.0 (strong coupling): Synchronized injection (δ=0) is optimal. Strong coupling already coordinates the particles into collective switching; phase lag disrupts the coordination rather than enhancing it.

This mirrors the biological pattern: ctenophores (intermediate regime, moderate hydrodynamic coupling) use metachronal waves, while organisms with stronger coupling mechanisms don't need them.

Note on Gate 0 (J=1.0): The δ=0 baseline (0.041) fell below the Noise Tuning value (0.065) by 0.024 — a ~2σ discrepancy between independent runs with different seeds. This is within expected cross-run variance (proper two-sample threshold: 0.034) but tripped the conservative single-sample gate. The within-sweep comparison (δ=0 vs δ*) remains valid since both share the same random process.

Design Rule (Principle 4)

For an N=4 Ising-coupled bistable circuit with coupling strength J:

• J < 2: Inject with phase lag δ ≈ π/5 between adjacent nodes. Expected improvement: 18–37% over synchronized injection.
• J ≥ 2: Use synchronized injection — the coupling already coordinates the particles. Phase lag hurts.
• δ is coupling-independent* in the moderate range (J=0.5–1.0 both give δ* ≈ 0.66).

Combined with the Noise Tuning Rule: first tune kT to the SR optimum for your coupling strength, then apply phase-lagged injection at δ ≈ π/5. The two knobs are independent — temperature controls noise level, phase lag controls injection timing.

Code: sim/L0/therm-env/tests/ising_chain.rs

Key References

• Byron, M.L. "Moving in the In-Between: Locomotion Strategies at Intermediate Reynolds Numbers." Princeton MAE Seminar (2022)
• McHenry, M.J. et al. "The Hydrodynamics of Locomotion at Intermediate Reynolds Numbers." Journal of Experimental Biology 206 (2003)
• Daniels, J. et al. "The Hydrodynamics of Swimming at Intermediate Reynolds Numbers in the Water Boatman." Journal of Experimental Biology 217 (2014)
• Hoover, A.P. et al. "Omnidirectional Propulsion in a Metachronal Swimmer." PLOS Computational Biology (2023)

Scale-Invariance: Octopus Compressed Command

Re ~ 10³-10⁵ — The Inertial Distributed Regime

The Physics

Inertia dominates. Viscosity still contributes to drag but does not govern the locomotion strategy. The fluid has memory — perturbations persist and propagate. The swimmer generates wakes that interact with the environment on relevant timescales. But this is also the regime of high degrees of freedom: the control problem is not underdetermined (as at low Re where everything scales out) but overdetermined — there are more controllable degrees of freedom than there are dimensions of the desired output, and the control architecture must manage this redundancy efficiently.

What the Octopus Does

The octopus has eight arms with virtually infinite degrees of freedom. Each arm is a muscular hydrostat — no rigid skeleton, with muscles providing both structural support and actuation. The arm contains approximately 380,000 motor neurons distributed along its length, yet the brain controls all of them via only ~4,000 efferent nerve fibers — a compression ratio of roughly 100:1.

This is the key insight. The central brain does not specify each muscle's activation individually. It sends a compressed command that specifies the goal (reach toward target at location X), and the arm's own distributed neural circuitry handles the decomposition of that command into the specific muscle activations required to execute the motion in the specific stochastic fluid environment the arm is currently in.

The primary locomotion primitive is bend propagation: the brain initiates a bend at the base of the arm, which travels as a wave from base to tip. Different reaching movements vary in speed and distance but maintain a basic invariant velocity profile scaled appropriately. The same motor program, scaled, generates all reaches. This scale-invariance is fundamental — it means the encoding strategy does not need to be re-designed for every target; one parametric program covers the full range.

Sensory information flows back in the opposite direction: ~2.3 million receptors distributed along the arm send information to the brain via only ~17,500 afferent fibers. Most sensory processing happens locally in the arm's peripheral nervous system. The brain receives a summary, not raw data.

The architecture: compressed high-level command → distributed local decoding → precise physical execution → compressed sensory summary → updated command.

The X-Encoding Principle

Principle 5: In the inertial distributed regime, the optimal architecture separates the encoding problem into two layers: a compressed high-level specification of the desired output (what state to reach) and a distributed local computation layer that translates that specification into physical actions appropriate to local noise conditions. The central controller does not need to know the noise floor at every circuit node; the local layer adapts.

Applied to thermodynamic circuits: The X-encoder broadcasts a compressed description of the target energy basin to all circuit nodes. Each node runs a local computation (analogous to the arm's ganglion) that determines how to modulate its local Langevin dynamics to steer toward that basin, given the local noise statistics it observes. The circuit's distributed neural computation — its local stochastic dynamics — handles the translation. This is the thermodynamic analog of bend propagation.

Principle 6: Scale-invariance of the encoding program. A parametric injection wave that can be scaled in amplitude and velocity to cover a range of target states (analogous to the octopus's invariant velocity profile) is more powerful than a lookup table of specific injection sequences for each target. Scale-invariance is also a prerequisite for scaling the circuit itself — an encoding strategy that must be re-tuned for each circuit size is not an engineering tool.

Experiment 6 — Scale-Invariant Encoding (Principle 6)

Scientific Question

Does the SR-optimal temperature hold at larger circuit sizes? If the Noise Tuning Rule requires retuning for each circuit size, they're lab curiosities. If they hold from N=4 to N=16 without adjustment, they're engineering tools.

This is the most important result for practical thermodynamic circuit design: design rules hold at scale without retuning.

Experimental Design

Fix coupling J=1.0. Repeat the Noise Tuning temperature sweep at three chain sizes:

Same Ising chain setup as the Noise Tuning chapter's multi-particle experiment: double wells, nearest-neighbor coupling, oscillating field, synchrony metric.

Results

Principle 6: VALIDATED. The SR-optimal temperature holds across a 4× range of circuit sizes.

| N | peak kT | peak sync | |t| | Interior? | |---|---------|-----------|-----|-----------| | 4 | 2.29 | 0.041 ± 0.011 | 3.77 | YES | | 8 | 2.29 | 0.049 ± 0.008 | 5.82 | YES | | 16 | 2.82 | 0.063 ± 0.009 | 7.15 | YES |

Gates:

Analysis:

N=4 and N=8 peak at exactly the same grid point (kT=2.29). N=16 peaks one grid step higher (kT=2.82) — a 23% shift across a 4× increase in circuit size. On the 25-point log-spaced grid, this is within one step of the reference.

The slight upward shift at N=16 is consistent with the effective-barrier model: a 16-particle chain has proportionally more interior particles (14/16 vs 2/4), each feeling coupling from both neighbors. The average effective barrier rises by ~8%, shifting the optimal noise temperature by a corresponding amount. This is a predictable, small correction — not a breakdown of the design rule.

Peak synchrony appears to increase with N on this 3-size sweep. However, the expanded N-scaling sweep (below) reveals this was a discretization artifact.

N-Scaling Expansion (N=4 to N=64)

An expanded sweep tested 8 chain sizes with finer kT resolution to determine whether the apparent sync increase with N was a real scaling law.

The kT range was narrowed and the grid density increased: 40 points focused in [1.0, 5.0] (where all peaks fall) versus 25 points spread across [0.1, 15.0]. This resolves higher peak sync values at all chain sizes — the original grid undersampled the true peak. Different seed offsets ensure independent noise realizations. All other physics parameters (ΔV, x₀, γ, k_BT, A₀, timestep, sub-steps, episode length) are identical.

Power law test: α = -0.037 ± 0.027, |t| = 1.38 (not significant at p < 0.01). There is no scaling law — peak synchrony is flat across a 16× range of circuit sizes.

Peak kT stability: The peak kT bounces between 2.28 and 3.18 without systematic drift (17.6% total variation around a mean of 2.75). The SR peak is broad and flat-topped; the exact optimum is noise-dominated, but the operating band is wide enough that this doesn't matter.

Interpretation: The system is approximately extensive. The preliminary increase from 0.041 to 0.063 in the 3-size sweep was a discretization artifact — the coarse 25-point grid over [0.1, 15.0] undersampled the peak, and the finer 40-point grid over [1.0, 5.0] resolves higher peak sync values at all chain sizes. The design rules hold without degrading across a 16× scale range, but fidelity does not improve for free. This is the expected behavior of a well-behaved statistical mechanical system.

Design Rule (Principle 6)

For coupling J=1.0, the SR-optimal temperature is kT ≈ 2.8 (mean across 8 chain sizes), holding from N=4 through N=64 without retuning.

Combined with the Noise Tuning Rule: For J < 1.5, operate at kT ≈ 2.3–2.8. For J ≥ 2.0, operate at kT ≈ 4.3. These rules apply regardless of circuit size in the range N=4–64.

An engineer scaling a thermodynamic circuit from a 4-node prototype to a 64-node production system can use the same noise temperature. No retuning required.

Code: sim/L0/therm-env/tests/ising_chain.rs — ising_scale_invariant_sweep

Key References

• Gutfreund, Y. et al. "Organization of Octopus Arm Movements: A Model System for Studying the Control of Flexible Arms." Journal of Neuroscience 16 (1996)
• Sumbre, G. et al. "Octopuses Use a Human-like Strategy to Control Precise Point-to-Point Arm Movements." Current Biology 16 (2006)
• Levy, G. et al. "Motor Control in Soft-Bodied Animals." Current Biology 25 (2015)
• Davenport, J.S. et al. "Lessons for Robotics from the Control Architecture of the Octopus." Frontiers in Robotics and AI (2022)
• Mischiati, M. et al. "Neural Models and Algorithms for Sensorimotor Control of an Octopus Arm." arXiv (2024)

Synthesis: What Transfers and What Doesn't

The Results

We tested five biological navigation principles on Ising chain models under Langevin dynamics. Three validated. Two failed.

The pattern is not random. It divides cleanly along a line.

The Boundary

What transfers: Principles that ask statistical-mechanical questions — optimal noise level (Noise Tuning), injection timing coordination (Injection Timing), extensivity at scale (Scale-Invariance). These are questions about how to exploit the statistics of a stochastic system. The Langevin framework is statistical mechanics, so it speaks this language natively.

What doesn't transfer: Principles that ask dynamical-systems questions — topological invariants of motion sequences (Topological Encoding), sharp bifurcation sensitivity (Deliberate Instability). These require mathematical structures the Langevin domain does not have: time-reversibility (for the scallop theorem) and sharp phase transitions at finite system size (for bifurcation amplification).

This is not a weakness of the framework. It is the framework's boundary condition — and knowing the boundary is as valuable as knowing the interior.

The Design Rules

Three quantitative rules emerged, each tested with statistical gates and reproducible code:

The Noise Tuning Rule: For an Ising-coupled bistable circuit with coupling J:

• J < 1.5: operate at kT ≈ 2.3
• J ≥ 2.0: operate at kT ≈ 4.3
• Keep ΔV/kT < 3.0 (hard trapping cutoff above this)
• The kT axis is sharp (±30% degrades to noise floor). The ΔV axis is forgiving.

The Injection Timing Rule: For coupled circuits at moderate coupling:

• J < 2: inject with phase lag δ ≈ π/5 between adjacent nodes (18–37% improvement)
• J ≥ 2: synchronize injection (coupling handles coordination)
• The two knobs (temperature and phase lag) are independent

The Scale-Invariance Rule: The Noise Tuning Rule holds from N=4 to N=64 without retuning:

• Optimal kT ≈ 2.8 at J=1.0 regardless of circuit size (confirmed across 8 chain sizes)
• The system is approximately extensive — fidelity neither improves nor degrades with scale

What the Failures Tell Us

Topological Encoding fails: In the Stokes regime (Re < 1), the scallop theorem guarantees amplitude is irrelevant — only sequence topology matters. In the Langevin domain, there is no scallop theorem. Amplitude is a direct lever: synchrony scales linearly with signal strength. An engineer should use amplitude modulation freely.

Deliberate Instability fails: The peregrine exploits pitch instability near a bifurcation point for sensitivity amplification. In the Ising chain, the ΔV axis shows a broad plateau, not a sharp transition. Langevin systems at finite N have smooth crossovers. An engineer should tune kT carefully (sharp peak) and not worry about ΔV (forgiving axis).

Both failures trace to the same root: Langevin systems are time-irreversible and have smooth energy landscapes at finite system size. The biological principles relied on time-reversibility (for topology to dominate) and bifurcation sharpness (for instability to amplify). Neither constraint exists in thermodynamic circuits.

The Complete Spectrum (Updated)

The middle of the Reynolds number axis transfers cleanly. The extremes require physics the Langevin model doesn't contain.

Implications for Thermodynamic Circuit Design

An engineer reading this document should take away three things:

1. Tune noise, don't suppress it. There is an optimal operating temperature for your circuit. It depends on coupling strength. See the Noise Tuning chapter.

2. Coordinate injection timing at moderate coupling. Phase-lagged injection at δ ≈ π/5 gives 18–37% better fidelity than synchronized injection. At strong coupling, synchronize instead. See the Injection Timing chapter.

3. These rules hold at scale. You do not need to retune when scaling from 4 to 16 nodes. See the Scale-Invariance chapter.

And two things NOT to do:

4. Don't try to operate near a bifurcation point. The sensitivity-amplification story from biology doesn't apply. The design surface is smooth, not critical.

5. Don't optimize sequence topology. Use amplitude instead. It works linearly and doesn't require the time-reversibility constraint that makes topology powerful in Stokes flow.

Relationship to Current Thermodynamic Computing Research

Extropic and Normal Computing are focused on chip fabrication and algorithm development. The X-encoding problem — how to inject inputs into a stochastic physical system — is acknowledged but not yet formally addressed in their published work. These design rules are complementary: they address the theory gap that enables the next generation of circuit design.

Open Source Philosophy

This research is released openly. The code that produced every result is in the same repository. The reasoning: becoming the foundational reference for noise-exploiting thermodynamic circuit design creates more value than any IP protection could. The simulation infrastructure (CortenForge) is the durable asset.

Dragonfly — Predictive Guidance

Status: Open hypothesis. The principles below are derived from biology but have not yet been tested on Ising chain models. Based on the stat-mech / dynamical-systems boundary established in the Synthesis chapter, we predict P7 and P9 will validate.

Re ~ 10⁵-10⁷ — The Predictive Inertial Regime

The Physics

Full inertial regime. Viscosity contributes to drag but not meaningfully to the locomotion control problem. The fluid generates persistent wakes and vortices. The key physical constraint is latency: at these speeds, the time for a reactive signal to propagate through the nervous system and execute a motor command is comparable to or longer than the time in which the target moves a significant fraction of the control distance. A pure reactive strategy fails — by the time the correction executes, it is already wrong.

This is the latency gap regime: you are moving fast enough to lose the distributed local computation advantage (the arm can't adapt fast enough), but not fast enough to generate the vortex-lift control authority that defines the next regime up. It is the most computationally demanding regime for a biological or engineered controller.

What the Dragonfly Does

Dragonflies achieve a prey capture success rate of 90-95% in the wild — the highest of any known predator. They do this while intercepting aerobatic prey in mid-air at speeds up to several meters per second, in turbulent air, with a nervous system containing roughly 1 million neurons total.

The mechanism was established by Mischiati et al. (Nature, 2015): dragonfly interception steering is driven by forward and inverse internal models of the dragonfly's own body dynamics and of the prey's predicted trajectory. Predictive rotations of the head continuously track the prey's angular position. The head-body angle thereby established guides systematic rotations of the body to align with the prey's predicted flight path. Model-driven control underlies the bulk of maneuvers; reactive control is reserved specifically for unexpected prey movements.

The guidance law was characterized by Brighton et al. (PNAS, 2017, working with peregrine falcons but confirmed analogously for dragonflies): proportional navigation (PN). Under PN, turning rate is commanded proportional to the angular rate of the line-of-sight to target:

$$ \omega_{\text{commanded}} = N \times \frac{d\lambda}{dt} $$

where λ is the line-of-sight angle and N is the navigation constant (feedback gain). This is the same guidance law used by most visually guided missiles. For dragonflies, N ~ 3, which coincides with the classical linear-quadratic optimal guidance result: PN with N = 3 minimizes control effort to intercept a non-maneuvering target.

Before takeoff, the dragonfly performs a critical pre-selection step: it assesses whether the prey's angular size and velocity co-vary within a privileged range, and times its takeoff to predict when the prey will cross its zenith. The behavioral decision embeds the computational constraints — the dragonfly only pursues prey it has pre-verified it can intercept given its own body dynamics.

The minimum sensory requirement is also remarkable: two local observables — vertical wind acceleration and torque (body rotation rate) — are sufficient to implement the PN guidance law. Global knowledge of the flow field is not required.

The X-Encoding Principle

Principle 7: In the latency-gap regime, the X-encoder must operate predictively. It cannot react to observed drift in the circuit's state because the signal propagation latency is too long for reactive correction to be useful. Instead, it maintains a forward model of the circuit's own relaxation dynamics — predicting where the system will be at time t+Δ and injecting X at the predicted future state rather than the current observed state. The injection leads the relaxation rather than chasing it.

Applied to thermodynamic circuits: A forward model of the Langevin dynamics, calibrated from observed circuit statistics, predicts the evolution of the energy landscape for the next several time steps. X-injection is computed against the predicted state. When prediction error exceeds a threshold (unexpected stochastic events), the system falls back to reactive injection temporarily, then returns to model-driven control.

Principle 8: Pre-selection and operating regime commitment. Before entering high-throughput operation, the X-encoder should verify that the circuit's current noise profile and relaxation rate are within the regime for which the encoding strategy was designed (analogous to the dragonfly's pre-takeoff assessment). Attempting to encode outside the designed regime is the primary failure mode.

Principle 9: Minimum observables sufficiency. Two local scalar cues are sufficient for full trajectory guidance. The X-encoder instrumentation can therefore be minimal: local energy gradient and local curvature (or equivalent observable pair) fed into a PN-style feedback law. This is a tractable instrumentation problem, not a global state estimation problem.

Hypothesis Given Current Results

Principles 7 (predictive forward model) and 9 (minimum observables) are statistical-mechanical in character — they ask how to use noise statistics to calibrate a model, and how few observables suffice for control. Based on the pattern from the validated results, these should transfer to the Langevin domain.

Principle 8 (pre-selection / regime commitment) is also stat-mech: it's a gate check on operating conditions before committing to a strategy. This should transfer.

Prediction: P7 and P9 will validate. The PN guidance law with N ≈ 3 should emerge from Langevin dynamics. Minimum observables (2 local cues) should suffice. These are noise-exploitation strategies, which is what the Langevin framework handles natively.

Status: Not yet tested. Infrastructure exists (experiment_4.rs).

Experiment 4 — Dragonfly PN Guidance in Langevin Noise

Implement proportional navigation with N as a free parameter in a simulated Langevin particle navigating a 2D energy landscape toward a target basin. Vary N across [1, 5] and measure convergence speed and convergence fidelity as a function of noise level. Verify the N ~ 3 optimum and characterize its robustness to noise floor variations. Extend to the forward-model variant: precompute the target basin's future position using the Langevin drift term, and measure the improvement in convergence.

Platform readiness: High — ThermCircuitEnv builder, DoubleWellPotential, LangevinThermostat, and all 8 RL/optimization algorithms exist.

Status: Plumbing validated — CEM learns state-dependent temperature control. The full PN guidance experiment (N sweep, noise sweep, forward-model variant) is not yet started.

Code: sim/L0/therm-env/tests/experiment_4.rs

Key References

• Mischiati, M. et al. "Internal Models Direct Dragonfly Interception Steering." Nature 517 (2015)
• Brighton, C.H. et al. "Terminal Attack Trajectories of Peregrine Falcons are Described by the Proportional Navigation Guidance Law of Missiles." PNAS 114 (2017)
• Mills, R. et al. "Physics-Based Simulations of Aerial Attacks by Peregrine Falcons Reveal that Stooping at High Speed Maximizes Catch Success." PLOS Computational Biology 14 (2018)
• Combes, S.A. "Neuroscience: Dragonflies Predict and Plan Their Hunts." Nature 517 (2015)
• Gonzalez-Bellido, P.T. et al. "Eight Pairs of Descending Visual Neurons in the Dragonfly Give Wing Motor Centers Accurate Population Vector of Prey Direction." PNAS 110 (2013)

Peregrine — Vortex-Noise Coupling

Status: Partially tested. P11 (deliberate instability) was tested and failed — see the Noise Tuning chapter. P10 and P12 require spatial physics (CFD or soft-body simulation) not available in the current infrastructure.

Re > 10⁷ — The High-Inertial Turbulent Regime

The Physics

At very high Re, the flow is turbulent. Vortices are generated spontaneously and persistently. The swimmer is immersed in a chaotic vortex field of its own and the environment's making. In the peregrine's case, a new physics emerges: vortex-induced lift. Deliberately shaped vortices generate aerodynamic forces larger than those from conventional attached flow. Speed generates control authority rather than destroying it.

This is the most counterintuitive regime in the biological spectrum. It is also where the gap between the biology and the Langevin analog is widest — the principles here rely on spatial vortex structure and sharp bifurcation sensitivity, neither of which exists in 1D coupled double-well models. See the hypothesis section below for what this means.

What the Peregrine Falcon Does

The peregrine is the fastest animal on Earth — exceeding 380 km/h in a stoop (hunting dive). At these speeds, it maintains not merely stable flight but precise, active maneuvering sufficient to intercept aerobatic prey. The mechanism was established by Gowree et al. (Communications Biology, 2018) and extended by Brucker and Gowree (AIAA Journal, 2021).

The stoop is a four-phase morphological sequence:

Phase I (Teardrop — T-shape): Wings folded completely, feathers tucked, legs retracted. Drag minimized. The falcon converts gravitational potential energy to kinetic energy with near-zero energy expenditure. Angle of attack maintained at ~5° — the equilibrium point where aerodynamic and gravitational forces balance. This phase is passive.

Phase II (Cupped wing — C-shape): Wings open slightly with primary feathers aligned vertically. Substantial lateral (side) forces generated — up to 3x body weight — enabling pure yaw control. Asymmetric morphing allows roll and heading correction. The strong vortices produced are aligned laterally, providing steering authority without significant deceleration.

Phase III (M-shape — terminal phase): The defining configuration. Wings deploy into a forward-swept M-shape. This is where the core physics lives.

The M-shape vortex field: Wind tunnel experiments and Large Eddy Simulations (LES) revealed a rich set of interacting vortex structures:

• Horn / Werle-Legendre vortices emanating from the frontal region due to strong spanwise flow promoted by the forward sweep of the radiale (wrist bone)
• Dorsal vortex (DV) interacting with the horseshoe vortex (HSV) of the body
• Wing vortex (WV) and tail vortex (TV) enhanced by M-shape geometry
• Primary feather vortex (PFV) at the wingtip primaries

The critical discovery: a counter-rotating vortex pair interacts with the main wing vortex to reduce induced drag, which would otherwise decelerate the bird significantly during pullout. The vortices do not merely provide lift — they actively cancel each other's drag penalty. The chaos is not fought; it is structured so that one layer cancels another.

Deliberate pitch instability: LES analysis confirmed that the falcon is flying unstably in pitch during the M-shape phase — positive pitching moment slope at trim angle of attack ~5°. This is a feature, not a flaw. Pitch instability maximizes responsiveness: a small input produces a large output change. The hand wings (primaries) act as "elevons" — stabilizing the intentionally unstable configuration while preserving its high-sensitivity property.

The guidance law: Brighton et al. (PNAS, 2017) confirmed using GPS loggers and onboard cameras that terminal attack trajectories follow the proportional navigation guidance law (shared with the dragonfly) but with a lower navigation constant N < 3, appropriate to the lower flight speed relative to missiles and accounting for higher biological sensor latency. Monte Carlo simulation confirmed N ~ 3 as the optimum for high-speed stoops against agile prey.

Physiological substrate supporting high-throughput precision:

• Nasal tubercles regulate respiratory pressure at >200 mph
• Visual acuity ~4x human density; 150 fps processing rate; dual fovea (forward shallow + lateral deep)
• Nictitating membrane clears debris without interrupting vision
• Reinforced arm skeleton and shoulder girdle (~2-3x bone mass of comparable raptors) to sustain 3g+ load factors
• Talon reflex arc bypassing conscious processing: impact → grip in ~15ms (vs. 200ms human reaction time)

The X-Encoding Principle

Principle 10: In the high-inertial turbulent regime, speed itself generates control authority. The appropriate strategy is not to suppress noise but to deliberately inject structured perturbations that generate counter-rotating vortex pairs whose mutual interaction cancels drag while preserving lift (control force). Higher throughput produces more vortex-induced force available for steering — the precision-throughput relationship inverts.

Applied to thermodynamic circuits: At high injection rates, the X-encoder should deliberately introduce paired perturbation structures into the circuit's Langevin noise field — perturbations designed so that their stochastic cross-correlation produces net drift toward the target energy basin, while their self-canceling structure minimizes the energy dissipation (entropy production) associated with the injection. This is the thermodynamic analog of the counter-rotating vortex pair.

Principle 11: Deliberate instability as a sensitivity amplifier. A circuit operating near a phase transition (bifurcation) in its energy landscape is pitch-unstable in the analogy — highly sensitive to perturbations. The X-encoder should target this operating point and use a minimal stabilization mechanism (the circuit analog of elevon primaries) to prevent divergence while preserving the high-sensitivity regime.

Principle 12: The logarithmic spiral as a scale-invariant approach geometry. Peregrine falcons resolve the conflict between aerodynamic streamlining (head straight) and maximum visual acuity (head turned 40°) by flying a logarithmic spiral path — a constant-angle curve that is self-similar at every scale. A scale-invariant X-injection trajectory would not require re-tuning as the circuit scales, which is the central engineering requirement for thermodynamic computing to become a manufacturable technology.

Hypothesis Given Current Results

Principle 11 (deliberate instability) has been tested and failed — the Ising chain shows a broad sensitivity plateau, not a sharp bifurcation point. The ΔV axis is forgiving, not critical. See the Noise Tuning chapter, Operating Envelope.

Principles 10 (paired perturbation structures) and 12 (logarithmic spiral) are dynamical-systems concepts — they require spatial vortex geometry and scale-invariant trajectory topology respectively. Based on the pattern from the validated results, these are unlikely to transfer to the Langevin domain. The Langevin framework doesn't have the vocabulary for structured spatial perturbations or topological trajectory invariants.

Prediction: P10 and P12 will not validate in the Langevin domain. They require physics (spatial vortex structure, trajectory topology) that the 1D coupled double-well model does not contain. Testing them properly requires at minimum a 2D/3D fluid simulation with vortex dynamics.

Status: P11 tested and failed. P10 and P12 not tested (would need CFD or soft-body physics).

Experiment 5 — Peregrine M-Shape Vortex Coupling Simulation

Simulate the M-shape vortex field in a fluid simulation domain. Quantify the counter-rotating vortex pair's drag-cancellation ratio as a function of injection speed. Translate the vortex geometry into a parameterized description of structured perturbation pairs for circuit injection.

Platform readiness: Low — needs reformulation to the Langevin domain or a fluid simulation.

Status: Not started. P11 (deliberate instability) tested and failed in Langevin domain — see the Noise Tuning chapter.

Code: —

Key References

• Gowree, E.R. et al. "Vortices Enable the Complex Aerobatics of Peregrine Falcons." Communications Biology 1 (2018)
• Brucker, C. & Gowree, E.R. "Peregrine Falcon's Dive: Pullout Maneuver and Flight Control Through Wing Morphing." AIAA Journal 59 (2021)
• Brighton, C.H. et al. "Terminal Attack Trajectories of Peregrine Falcons are Described by the Proportional Navigation Guidance Law of Missiles." PNAS 114 (2017)
• Tucker, V.A. "Curved Flight Paths and Sideways Vision in Peregrine Falcons." Journal of Experimental Biology 203 (2000)
• Mills, R. et al. "Physics-Based Simulations of Aerial Attacks by Peregrine Falcons." PLOS Computational Biology 14 (2018)

Transition Zones and the Lateral Line

Status: Open hypothesis. No experiments have been run on regime transitions. This depends on having 2–3 regime-specific strategies working first.

The five regimes and their principles are the theoretical framework. The experimental results in the Validated Results section validated three principles and identified the statistical-mechanics / dynamical-systems boundary. The transition zones between regimes remain an important open question, because:

1. Real thermodynamic circuits will operate across a range of throughput levels, necessarily crossing multiple regime boundaries.
2. The biological exemplars suggest that transitions between regimes are discontinuous — there is no smooth interpolation of strategies, and performance degrades sharply in the transition zone before a new strategy takes over.
3. The intermediate regime (Regime 2) exists precisely because neither low-Re nor high-Re strategies work there, and the organisms that inhabit it (ctenophores, water boatmen) are notably less efficient than those that specialize in either adjacent regime.

The transition zone hypothesis: In thermodynamic computing, as injection throughput increases from the regime of one strategy to the next, there will be a characteristic throughput range where neither strategy works well — analogous to the intermediate Reynolds number regime. Identifying these critical throughput values for a given circuit architecture, and designing the mode-switching logic that handles the transitions, is a specific and tractable engineering problem that this research program can address through simulation.

The Fish Lateral Line

There is a sixth organism that deserves dedicated study for its unique role at the boundary between Regimes 3 and 4: the fish with lateral line. The lateral line is an array of mechanosensory organs distributed along the fish's body that detects local pressure gradients and vortex shedding frequencies with remarkable precision. Fish use the lateral line to perform Karman gaiting — holding station in a turbulent vortex street behind a cylinder by passively synchronizing their body kinematics to the oscillating flow. This is energy harvesting from chaos: exploiting the environmental noise field as a source of free locomotion rather than fighting it. The lateral line provides the sensing architecture that makes this possible. The engineering analog is an in-situ noise characterization system embedded in the circuit itself, providing real-time observables to the X-encoder without interrupting computation.

Experiment 6 — Regime Transition Characterization

For each pair of adjacent regimes, simulate a Langevin circuit operating at the strategy of the lower regime and systematically increase the injection rate. Measure when performance begins to degrade and what the transition regime looks like. Identify the critical τ_circuit / τ_noise ratio at each transition. Map these to concrete circuit parameters (injection rate, noise floor, relaxation time) to provide actionable design guidance for where to switch encoding strategies.

Platform readiness: Low — this is a meta-experiment that requires 2-3 regime-specific strategies to be working first. It depends on Experiments 1, 4, and ideally 3 or 5.

Status: Not started

Code: —

Experiment 7 — Fish Lateral Line In-Situ Sensing

Implement a lateral-line-style sensor array in the simulated circuit: distributed local pressure sensors (equivalent to neuromasts) providing real-time vortex shedding frequency and local gradient information. Test whether this in-situ sensing reduces the number of pre-measurement calibration steps required before the X-encoder can operate effectively. Quantify the steady-state sensing accuracy as a function of array density and sensor placement.

Platform readiness: Low — needs a distributed pressure-gradient sensor plugin. The sim sensor infrastructure exists but does not include the specific sensor type needed.

Status: Not started

Code: —

Experiments, Status, and Next Steps

Infrastructure

All experiments run on CortenForge's simulation stack:

Completed Validations

The Boundary

What transfers: Statistical-mechanical questions — noise tuning, phase coordination, extensivity. The Langevin framework speaks this language natively.

What doesn't: Dynamical-systems questions — topological invariants (requires time-reversibility), bifurcation sensitivity (requires sharp phase transitions at finite N). The model doesn't have the vocabulary.

Remaining Langevin-Ready Principles

These could be tested with the existing ThermCircuitEnv infrastructure:

Principles Needing Different Physics

Next Experiments

Four follow-on experiments that deepen the validated results, ordered by impact:

1. N-Scaling Law — COMPLETED

Result: No scaling law. Peak synchrony is flat (~0.058–0.071) across N = 4–64 with no significant trend (α = -0.037, |t| = 1.38). The preliminary increase from the 3-size sweep was a discretization artifact. The system is approximately extensive — design rules hold without retuning across a 16× scale range, but fidelity does not improve for free. Peak kT is stable (mean 2.75, 17.6% drift). Closes open question 6.

Code: ising_scale_law_sweep in ising_chain.rs. Runtime: 7.5 hours.

2. Coupling Crossover Mapping

The Noise Tuning rule showed two regimes: weak coupling (J < 1.5, peak kT ≈ 2.3) and strong coupling (J ≥ 2.0, peak kT ≈ 4.3). Where exactly is the crossover? Is it smooth or sharp?

Experiment: Sweep J = 1.0, 1.25, 1.5, 1.75, 2.0 with 25 kT points, 40 episodes each. ~3 hours.

Gate: If the crossover occupies less than ΔJ = 0.25, it's sharp — potentially a phase transition in the coupled system.

3. Optimal (J, δ) Surface

A finer mesh would map the full optimal phase-lag surface: 8 J values × 30 δ values × 80 episodes. ~6 hours.

4. Effective Barrier Model Validation — COMPLETED

Merged into experiment 1 as Gate 3. Result: the effective-barrier model (R² = 0.29) does not fit — peak kT bounces without systematic drift, indicating the SR peak is broad enough that the exact optimum is noise-dominated rather than barrier-determined.

All Experiment Code

Code: sim/L0/therm-env/tests/ising_chain.rs

Experiment Platform: ThermCircuitEnv

All experiments in this research program share the same core loop: configure a Langevin particle system with an energy landscape, wire it into a training environment, run a gradient-free optimizer or RL algorithm to modulate temperature or other control channels, and measure the result. The ThermCircuitEnv crate eliminates the boilerplate that each experiment would otherwise repeat.

What It Provides

ThermCircuitEnv sits on top of two layers of infrastructure:

The ML chassis (sim-ml-chassis) provides the Algorithm trait, batched environments, policy representations, and rollout machinery. Every algorithm (CEM, REINFORCE, PPO, TD3, SAC from sim-rl; SA, Richer-SA, PT from sim-opt) bolts onto the same trait. Any experiment built on ThermCircuitEnv automatically gets access to all 8 algorithms with zero additional wiring.

The thermostat layer (sim-thermostat) provides LangevinThermostat and passive energy landscape components: double-well potentials, oscillating fields, pairwise coupling, ratchets, and external fields.

ThermCircuitEnv bridges these: it takes a circuit description (particle count, energy landscape, control channels) and produces a ready-to-train environment.

Example

#![allow(unused)]
fn main() {
ThermCircuitEnv::builder(n_particles)
    .gamma(10.0).k_b_t(1.0).timestep(0.001)
    .with(DoubleWellPotential::new(3.0, 1.0, 0))
    .with_ctrl_temperature()
    .reward(|_m, d| -(d.qpos[0] - 1.0).powi(2))
    .sub_steps(100).episode_steps(1000)
    .build()
}

Validation

The platform was validated with a CEM training test: one particle in a double well with ctrl-temperature. CEM learned state-dependent temperature control, achieving roughly 2x improvement over a constant-temperature baseline. This confirms the plumbing is correct end-to-end but does not test the scientific claims of any specific biological regime.

Location: sim/L0/therm-env/ — 38 tests total (31 debug, 7 require --release).

Glossary

Chemotaxis: Navigation by a chemical gradient, characteristic of bacteria including E. coli.

CheY-P: Phosphorylated CheY protein in E. coli; the signaling molecule that controls flagellar motor switching between run and tumble states. Follows Langevin dynamics with intrinsic noise that enhances gradient sensitivity.

Energy-based model (EBM): A machine learning model that defines the shape of a probability distribution via an energy function. The foundational mathematical object in thermodynamic computing.

Geometric phase / holonomy: In the context of low-Re locomotion, the net displacement achieved by a closed loop in configuration space. Formally equivalent to the holonomy of a connection on a fiber bundle (Shapere-Wilczek formalism).

Karman gait: A swimming behavior in fish involving synchronization to the oscillating flow of a Karman vortex street behind a cylinder; allows passive energy harvesting from environmental turbulence.

Karman vortex street: Alternating, periodic vortices shed downstream of a bluff body in steady flow.

Langevin dynamics: A stochastic differential equation describing the evolution of a system subject to deterministic drift plus Gaussian noise:

$$ dX = -\nabla U(X),dt + \sqrt{2kT},dW $$

where U is the energy function and dW is Wiener process noise.

Lateral line: An array of mechanosensory organs distributed along the body of fish and aquatic amphibians, detecting local pressure gradients and flow velocity. Enables vortex detection and Karman gaiting.

Metachronal wave: Sequential, coordinated beating of an array of appendages (cilia, pleopods, etc.) with a phase lag between adjacent elements, producing a traveling wave appearance. Characteristic of ctenophores, krill, and many other invertebrates.

Muscular hydrostat: A muscular structure with no rigid skeleton (e.g., octopus arm, elephant trunk, human tongue) in which muscles provide both structural support and actuation, enabling infinite degrees of freedom.

Proportional navigation (PN): A guidance law in which turning rate is commanded proportional to the angular rate of the line-of-sight to target. Used by most guided missiles; proven in peregrine falcon and dragonfly prey capture. Optimal navigation constant N ~ 3 minimizes control effort to intercept a non-maneuvering target.

Purcell's scallop theorem: At low Reynolds number (Re < 1), any reciprocal body motion (time-reversible) produces zero net displacement. Net locomotion requires non-reciprocal motion — sequences that trace a closed loop enclosing nonzero area in configuration space.

Reynolds number (Re): Dimensionless ratio of inertial to viscous forces in fluid flow: Re = ρUL/μ. The fundamental organizing parameter for fluid dynamics and biological locomotion strategies.

Run-and-tumble: E. coli's primary locomotion strategy. "Run" = all flagella rotating CCW, forming a bundle → straight motion. "Tumble" = one or more flagella switching to CW rotation → random reorientation.

Stochastic Processing Unit (SPU): Normal Computing's first prototype thermodynamic computer; an 8-cell stochastic circuit on a PCB using RLC elements.

Stochastic resonance: The phenomenon by which an intermediate level of noise improves signal detection or transmission performance in a nonlinear system, beyond what is possible with either no noise or excessive noise.

Thermodynamic Sampling Unit (TSU): Extropic's hardware unit; a probabilistic circuit that produces samples from a programmable energy-based distribution. The thermodynamic computing analog of the GPU.

X-encoding: The problem of translating a logical input X into the initial or boundary conditions of a physical stochastic system such that the system's relaxation toward equilibrium produces the correct output distribution Y. The primary unsolved problem in thermodynamic computing hardware design.

Open Questions

Answered by This Work

1. Does stochastic resonance scale from single particles to coupled circuits? Yes. The SR peak persists across coupling strengths J = 0–2 and circuit sizes N = 4–16. Two regimes emerge: weak coupling (J < 1.5, peak kT ≈ 2.3) and strong coupling (J ≥ 2, peak kT ≈ 4.3). See the Noise Tuning chapter.

2. Does the optimal noise level require retuning at larger circuit sizes? No. Peak kT ≈ 2.8 holds from N=4 to N=64 at J=1.0. Confirmed by both the original 3-size sweep and the expanded 8-size sweep with finer kT resolution. See the Scale-Invariance chapter.

3. Does phase-lagged injection improve fidelity in coupled chains? Yes, at weak coupling. Optimal δ ≈ π/5 gives 18–37% improvement over synchronized injection at J = 0.5–1.0. At J ≥ 2, coupling handles coordination and synchronized is optimal. See the Injection Timing chapter.

4. Does topological encoding dominate amplitude in the Langevin domain? No. The scallop theorem requires time-reversible dynamics. Langevin dynamics are time-irreversible, so amplitude is a direct lever. Synchrony scales linearly with signal amplitude. See the Noise Tuning chapter, Level 4.

5. Is there a sharp bifurcation point for sensitivity amplification? No. The ΔV axis shows a broad sensitivity plateau (ΔV/kT ∈ [0.25, 2.75]) with a gradual trapping cutoff, not a sharp transition. See the Noise Tuning chapter, Operating Envelope.

6. Does synchrony improve with circuit size? No. An expanded sweep across N = 4, 8, 12, 16, 24, 32, 48, 64 (40 kT points in [1.0, 5.0], 40 episodes) shows peak synchrony is flat at ~0.058–0.071 with no significant trend (power law α = -0.037, |t| = 1.38, not significant). The preliminary increase from the 3-size sweep was a discretization artifact on the coarse 25-point grid. The system is approximately extensive — the design rules hold without retuning across a 16× scale range, but fidelity does not improve for free. See the Scale-Invariance chapter.

Still Open

7. Is the weak/strong coupling crossover a phase transition? The current data shows two discrete modes with a transition between J = 1.5 and J = 2.0. Whether this is smooth or sharp (and whether it has the character of a thermodynamic phase transition) is unknown. See Next Experiment 2.

8. What is the correct dimensionless ratio for thermodynamic circuits? We proposed τ_circuit / τ_noise as the analog of the Reynolds number. The experiments used kT as the control variable and J as the coupling parameter, but the fundamental dimensionless group that governs the behavior remains to be identified. The ratio ΔV/kT ≈ 1.39 at the SR peak may be part of it.

9. Which biological principles transfer and which don't — is there a general rule? This work found that statistical-mechanical questions (noise tuning, phase coordination, extensivity) transfer to the Langevin domain, while dynamical-systems questions (topological invariants, bifurcation sensitivity) do not. Is this boundary precise? Does it hold for other Langevin systems beyond Ising chains?

10. How do these design rules map to real thermodynamic computing hardware? The experiments use idealized double-well potentials with nearest-neighbor coupling. Real circuits (Josephson junctions, molecular switches, optical bistable elements) have different noise statistics, coupling topologies, and operating timescales. Connecting the design rules to specific hardware parameters is the next major step.