Spike-constrained free energy control mechanisms in biological neural circuits will synchronize with the topological diversity of timescales in recurrent networks, producing measurable improvements in robustness to perturbations in both in silico cortical models and LEO satellite fault-tolerant computing architectures.

In recurrent spiking neural networks (SNNs) subject to a free-energy-minimizing control objective with hard spike-count/metabolic constraints, the resulting controller will (a) preferentially route compensatory activity through nodes/edges whose intrinsic timescales match the topological diversity (heterogeneity of eigenvalue/timescale spectra) of the network's connectivity graph, and (b) this timescale-topology alignment will produce a measurable increase in robustness — defined as maintained task/state-estimation performance under node lesion, edge removal, or noise injection — relative to (i) unconstrained controllers, (ii) constrained controllers with randomized/shuffled timescale-topology pairing, and (iii) topology-blind constant-gain controllers. The same control-topology alignment principle, instantiated as a resource-constrained fault-tolerant scheduler with heterogeneous "recovery timescales" per compute node, will produce a quantitatively analogous robustness gain (≥15% improvement in mean-time-to-recovery or task-completion-under-fault, effect size Cohen's d≥0.5) in a simulated LEO satellite compute fabric under randomized single- and dual-node fault injection.

Minimum viable test (Stage 0, ~10 days): Single recurrent SNN (500 units, Watts-Strogatz small-world topology with heterogeneous synaptic time constants τ∈[5,200]ms), free-energy control objective (variational free energy proxy: prediction error + spike-cost regularizer), compare against 2 baselines (unconstrained, shuffled-τ) under random node-silencing (10%, 20%, 30%) over 1,000 trials. If no significant separation (Stage 0 disproof criteria met), abort before scaling to satellite simulation.

Full protocol: 3 cortical topology classes (small-world, scale-free, geometric random) × 3 controller conditions × 5 perturbation types (node lesion, edge removal, weight noise, delay jitter, input dropout) × 3 perturbation magnitudes = 135 conditions, 200 trials each (27,000 simulation runs). Satellite analog: discrete-event simulation of 50–200 node LEO compute mesh, same 3×5×3 factorial design mapped to fault-injection scenarios (node failure, link failure, radiation-induced bit flips, latency spikes), using real fault statistics from published CubeSat/LEO reliability data where available.

# Stage 0: minimal viable test
for topology in [small_world]:
    G = generate_topology(topology, n=500, seed=range(10))
    for tau_regime in [aligned, shuffled, homogeneous]:
        tau = assign_timescales(G, regime=tau_regime)
        net = RecurrentSNN(G, tau)
        controller = FreeEnergyController(net, lambda_cost=sweep([0.01,0.1,1.0]))
        controller.train(task=noisy_lorenz_estimation, epochs=until_convergence)
        for pert_frac in [0.1, 0.2, 0.3]:
            for trial in range(200):
                net_p = perturb(net, method='node_lesion', frac=pert_frac)
                perf = evaluate(controller, net_p, task)
                alignment_idx = spearman(tau, graph_centrality(G))
                log(topology, tau_regime, pert_frac, perf, alignment_idx)

mixed_effects_anova(data, dv='perf', factors=['topology','tau_regime','pert_frac'])
mediation_analysis(iv='tau_regime', mediator='alignment_idx', dv='perf')

# Stage 1: satellite port (only if Stage 0 passes success criteria)
sat_graph = load_LEO_topology(Hypatia, n_nodes=[50,200])
recovery_latency = assign_timescales(sat_graph, regime=[aligned, shuffled])
scheduler = FreeEnergyScheduler(sat_graph, recovery_latency, budget_constraint)
inject_faults(sat_graph, mode=['single_node','dual_node','regional_radiation'])
compare_MTTR(scheduler_conditions)