{"id":"58be0173-0c85-441f-bef2-fc1c721608bc","arxiv_id":"2607.10439","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A port-Hamiltonian GNN trained on EEG phasors matches the cortex's avalanche-branching ratio (σ≈1) but misses its 1/f spectrum and long-range correlations.","lead":"This paper fits a port-Hamiltonian model of brain dynamics to EEG recordings, with a graph neural network learning how brain regions exchange and dissipate energy. It claims the model reproduces near-critical avalanche behavior of real cortex, while openly failing two other statistical signatures of brain activity.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The single passing invariant—avalanche branching σ≈1—is not established as a port-Hamiltonian prediction: it depends on an uncalibrated noise temperature T=0.2 and the learned R(x), with no null model or sensitivity analysis.","rationale":"The reader's weakest assumption is exactly the load-bearing soft spot: the branching success is a possible fitted consequence of T and R(x), not a clean port-Hamiltonian prediction. I agree. The paper deserves credit for honestly reporting the spectral-slope and DFA failures, for a leakage-free by-subject split, and for a concrete upgrade path. But the central positive evidence—near-critical branching—would be decisive only if it were robust and structure-dependent. As written, no sensitivity analysis and no ablation are provided, so the σ≈1 result is compatible with a null explanation: any sufficiently noisy stochastic rollout with a tuned noise scale can produce branching ratios near 1 under threshold-based avalanche estimation. The proposed concrete test would settle whether the port-Hamiltonian conservative coupling (J) and the learned dissipation (R) are actually responsible for the branching match, or whether the single free parameter T is responsible. Because this concern directly undermines the paper's strongest claim, and because the reader already judged the overall evidence insufficient, the verdict should remain REJECT; no adjustment is needed.","tokens_in":17638,"tokens_out":6505,"duration_ms":76879,"concrete_test":"On the same fitted model, with the same threshold/bin avalanche estimator, (i) sweep T over {0.05, 0.1, 0.2, 0.5, 1.0} and report σ; and (ii) at T=0.2, repeat the free run with the learned skew-symmetric J(x) set to zero (pure dissipation+noise) and also with R(x) replaced by the identity. If σ remains within ~0.05 of 1 under any of these conditions, the branching match is a fluctuation-level/estimator artifact rather than evidence for the port-Hamiltonian conservative structure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's only positive validation result is the free-running stochastic model's branching parameter σ=1.00 versus the measured 0.94 (Section V A). This is not shown to be a consequence of the port-Hamiltonian/metriplectic structure. The fluctuation-dissipation noise in Eq. 11 has covariance 2T R(x), where R(x) is the learned dissipation MLP and T=0.2 is a scalar chosen without a priori justification or sensitivity analysis. Since T and R(x) enter multiplicatively, the effective noise scale is freely adjustable through the un-pinned magnitude of R; the training loss does not constrain that scale. Moreover, the threshold-crossing avalanche estimator is applied to a phase/frequency state whose precise relation to the recorded EEG signal is not specified, and many non-critical stochastic processes can yield descendant-to-ancestor ratios near 1 under suitable thresholding. The paper reports no null model (e.g., J(x) set to zero, or a generic diffusion with the same covariance) and no T sweep. The noiseless rollout produces no avalanches by construction, so the chosen fluctuation level is doing the work; whether it encodes near-critical cortical dynamics or simply a tuned noise amplitude is untested. Thus the sole passing rung does not yet support the central claim that the port-Hamiltonian structure captures a model-independent criticality invariant.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a port-Hamiltonian/GNN model of cortical dynamics. The state is an alpha-band phase/frequency phasor vector; the Hamiltonian is decomposed into five band sub-energies; the skew-symmetric coupling J(x) is gated by an empirical alpha-band phase-locking value (PLV); dissipation is a diagonal positive-semidefinite R(x) learned with a softplus network; and the training loss combines kinematic reconstruction with passivity, energy-balance, PLV, and PAC regularizers. The model is trained on ~1.1M phasor samples from the PhysioNet EEGMMIDB with three held-out subjects, reaching a held-out kinematic MSE of 1.30e-4. The free-running stochastic model is then scored against model-independent invariants: avalanche branching is reported as sigma = 1.00 versus a measured 0.94 (pass), while the 1/f spectral slope (1.96 vs 1.18) and DFA exponent (1.68 vs 0.68) fail. The paper frames the branching result as evidence for the port-Hamiltonian/metriplectic structure and outlines an extensive upgrade path toward closed-loop neuromodulation.","tokens_in":17985,"tokens_out":4806,"duration_ms":56176,"significance":"The paper has real strengths: it reports a leakage-free by-subject split, it states its validation ladder explicitly, it honestly reports the failed spectral and DFA rungs, and it makes code and data availability commitments. If the branching match were robust and attributable to the metriplectic structure, it would be a noteworthy result. However, the sole passing invariant is not established as a consequence of the advertised physics: the implemented model is not the non-equilibrium steady-state system described in the theory, the PLV-based connectome validation is circular, and the branching result depends on an uncalibrated noise temperature. The significance is therefore substantially below what the abstract claims.","major_comments":[{"comment":"The implemented and validated model is not the metriplectic NESS system described in the theory. Equation (8) introduces the metabolic port b_met(x) and the irreversible bracket M∇S, and Eq. (10) defines a steady-state power balance P_met = P_diss. But the training loss (19b) explicitly penalizes positive energy increase via L_p = E[ReLU(H_dot)], and Eq. (19c) enforces the port-Hamiltonian power balance without any metabolic port. The stochastic free-run (Eq. 11) also omits b_met. Thus the validated dynamics are a passive stochastic system, not the NESS system with a metabolic supply; the branching result cannot be credited to the metriplectic non-equilibrium structure.","section":"II E, III F, V A"},{"comment":"The branching result sigma = 1.00 versus 0.94 is the only passing rung, but it is not shown to be a prediction of the port-Hamiltonian/metriplectic structure. The noise covariance is sigma sigma^T = 2 T R(x), with T = 0.2 chosen without a priori justification or sensitivity analysis. Since R(x) is a learned MLP whose overall scale is not pinned by the loss, T and R(x) enter multiplicatively and the effective noise amplitude is effectively free. The deterministic noiseless rollout produces no avalanches by construction, so the fluctuation level is doing the work. The paper provides no T-sweep, no null model with J=0, and no comparison with a generic diffusion process having the same covariance. A non-critical stochastic process can produce descendant/ancestor ratios near 1 under suitable thresholding.","section":"V A, Eq. (11)"},{"comment":"The connectome validation is circular. The learned coupling J(x) is multiplied by the measured alpha PLV matrix in Eq. (16), and the PLV coherence prior L_PLV in Eq. (19d) explicitly anchors |J| to the same measured PLV. Figure 10b then reports a positive association between |J| and the empirical alpha PLV as evidence of anchoring. That association is enforced by the loss and the multiplicative gate, not discovered independently. It therefore cannot serve as evidence that the model 'recovers' synchrony structure.","section":"III C, Eq. (16), IV D, Fig. 10"},{"comment":"The avalanche estimator is not clearly defined on the model output. The model state is x = [phi, omega] (Eq. 5), i.e., phase and angular frequency variables, not an EEG voltage or amplitude. Section III H describes thresholding a 'multichannel signal' at 2.5 standard deviations to define avalanches, but Section V A does not specify what observable of the free-run is thresholded, how it relates to the recorded EEG, or how the same estimator is applied to both. Without this specification, the reported sigma = 1.00 is not interpretable as a cortical avalanche branching parameter.","section":"III H, V A"}],"minor_comments":[{"comment":"The text states the model's aperiodic exponent is beta = 1.96, while the Fig. 9 caption gives beta ~ 2.1 for the same model spectrum. Please reconcile the two numbers.","section":"V A vs Fig. 9"},{"comment":"Eq. (3) is introduced in Section II A, but Section II E refers to 'the Stuart-Landau oscillator of Section II E'. The cross-reference should be corrected.","section":"II E"},{"comment":"The notation in Eq. (19d), f|J| - PLV(alpha) with tilde over PLV, is not defined precisely. Specify the normalization and the Frobenius-norm indexing.","section":"III F"},{"comment":"Equation (11) drops the metabolic port b_met that appears in Eq. (8), without comment. Either include it consistently or state explicitly that the implemented free-run neglects the metabolic port and justify that simplification.","section":"II E, V A"},{"comment":"The abstract claims the model is a 'physically principled, structure-preserving substrate for closed-loop neuromodulation', but the paper itself notes in Section V I that only about 5% of the state space is directly reachable with the three anatomical ports and that closed-loop control results are deferred. Softening the abstract claim would better match the evidence.","section":"Abstract"}],"recommendation":"reject","confidential_remarks":"I recommend rejection. The only positive validation result, avalanche branching, is not tied to the port-Hamiltonian/metriplectic structure: the implemented model is not the NESS system described in the theory, the PLV validation is circular, and the branching value depends on an uncalibrated noise temperature with no null model or sensitivity analysis. These are load-bearing problems with the central claim rather than presentation issues, and they would require substantial new experiments and re-analysis to address."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a real core idea: combining port-Hamiltonian structure, GNN surrogates, and a phase-locking-gated connectome to model EEG dynamics, then validating against a ladder of model-independent invariants. It also reports its failures openly — the spectral slope and DFA exponents don't match, and it says so plainly. That honesty is genuine and refreshing. The by-subject split and seed robustness are also handled cleanly, and the upgrade path is concrete and falsifiable.\n\nBut the central claims outrun the evidence. The metriplectic/NESS theory in Section II E is never actually implemented: the metabolic port b_met and entropy functional S(x) don't appear in the training loss or in the free-running simulations. Instead, training uses a passivity regularizer that penalizes energy increase, which is the opposite of the advertised non-equilibrium steady state. So the model that is fit and validated is not the model the abstract describes.\n\nThe one positive validation result, avalanche branching σ≈1.00 vs. measured 0.94, also doesn't yet carry the weight put on it. It comes from a stochastic free-run with a single arousal temperature T=0.2 chosen without a priori justification, and the noise covariance is 2T R(x) where R(x) is a learned MLP. T and R(x) enter multiplicatively, so the effective noise level is freely adjustable through the un-pinned magnitude of R. No sensitivity analysis or null model is reported. A diffusion with the same covariance but no port-Hamiltonian structure might give the same branching under the right threshold. So the branching match is not established as a consequence of the port-Hamiltonian form.\n\nThe connectome validation is circular. The learned coupling J(x) is gated by the measured alpha PLV (Eq. 16), and the training loss L_PLV anchors |J| to that same PLV (Eq. 19d). The paper then shows a correlation between learned coupling and the same PLV as evidence of anchoring. That's not validation; it's a description of the training target.\n\nThese are load-bearing gaps because the abstract claims a 'physically principled, structure-preserving substrate' and stability guarantees for neuromodulation. The implemented model is a fitted neural ODE with a passivity regularizer and a tuned noise term. That can still be a useful tool, but it's not yet what the title and abstract promise.\n\nWho gets value from this? Researchers working on physics-constrained deep learning for neural time series, especially those interested in validation ladders and honest benchmarking. It deserves a serious referee — the framework is worth developing — but it needs major revisions: either implement the metriplectic structure or drop those claims, add a null model and T-sweep for branching, and replace the circular PLV validation with something held out.\n\nMy recommendation: send it to peer review, but make clear to the authors that the current implementation doesn't support the advertised theory and the branching result needs much stronger evidence.","headline":"Interesting framework with honest failure reporting, but the implemented model doesn't match the advertised theory and the one success may be a tuning artifact.","tokens_in":18528,"tokens_out":1551,"would_cite":false,"duration_ms":20195,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A port-Hamiltonian GNN surrogate of cortex, fitted to scalp EEG phasors, reproduces near-critical avalanche branching (σ=1.00 against observed 0.94) but fails the 1/f and DFA rungs; the paper presents the gap as a concrete upgrade path.","keywords":["port-Hamiltonian systems","metriplectic dynamics","graph neural networks","cortical criticality","avalanche branching","EEG phasors","non-equilibrium steady states","neuromodulation"],"falsifier":"Re-run the stochastic free-run with T swept from, say, 0 to 1 (or learned from data) and check whether σ stays near 1 across a broad range; also replace the learned R(x) with scalar dissipation and re-score. If branching collapses or moves with T, the invariant is an artifact of tuning rather than an emergent property.","tokens_in":17400,"feed_emoji":"🧠","tokens_out":8186,"duration_ms":83277,"temperature":0.7,"pith_summary":"This paper tries to establish that a physics-constrained model of cortical dynamics—phases and frequencies of EEG oscillations coupled through a port-Hamiltonian structure with a learned dissipation graph network—can be fitted to human recordings and then judged by invariants it was never trained to match. The free-running model reproduces the cortex's near-critical avalanche branching (σ=1.00 against a measured 0.94), and the paper takes this as evidence that the dissipation–fluctuation structure is doing real physical work. The same free-run fails two other model-independent tests: the aperiodic 1/f spectral slope (β=1.96 vs 1.18) and long-range temporal correlations (DFA α=1.68 vs 0.68), a gap the paper states plainly and attributes to specific, named upgrades: excitatory–inhibitory criticality control and volume-conduction-robust connectivity. A sympathetic reader would care because a model that clears one nontrivial invariant while being explicit about the rungs it misses offers a testable route from black-box EEG classification toward mechanism, energy accounting, and stimulation with stability guarantees.","feed_headline":"Physics-constrained cortex model hits one criticality test, misses two","feed_subtitle":"A GNN port-Hamiltonian EEG twin reproduces the branching it never trained on; the failed 1/f and DFA rungs are the concrete next targets.","key_machinery":"The load-bearing object is the metriplectic/port-Hamiltonian cortical twin: canonical phasor coordinates x=[φ,ω], a Hamiltonian decomposed into five band sub-energies, a skew-symmetric connectome J(x) gated by a phase-locking prior, state-dependent dissipation R(x), and a single 'arousal temperature' T that scales the fluctuation-dissipation noise. Its role is to guarantee structure: conservative routing in one bracket, entropy production in the other, a steady-state power balance instead of decay to silence, and stability guarantees under stimulation. The GNN surrogate makes the state-dependent operators learnable at 64 channels while preserving the physical shape constraints by constructio","core_discovery":"The central claim, on the author's own terms, is that a 'Cortical GNN-pHNN'—a graph-neural-network surrogate embedded in a port-Hamiltonian/metriplectic structure over band-limited neural phasors—can be fitted to large-scale scalp EEG from multiple subjects and then evaluated as a physical model rather than a classifier. The Hamiltonian is stratified into five interpretable sub-energies (delta through gamma); the coupling matrix J(x) is kept skew-symmetric by construction and gated by measured phase-locking; dissipation R(x) is state-dependent and positive-semidefinite; and a fluctuation-dissipation noise channel with a single arousal temperature T turns the system into a non-equilibrium ste","pith_inferences":["Editorial inference: the branching result is currently overdetermined by the unexamined choice T=0.2 and the learned dissipation R(x); until a temperature sweep is reported, the port-Hamiltonian structure cannot be separated from fitted noise.","Editorial inference: because the metabolic port is absent from the implemented free-run, the validated model is closer to a stochastic dissipative oscillator than to the advertised metriplectic non-equilibrium steady state; the non-equilibrium claim is prospective.","Editorial inference: a cheap control experiment—train a black-box recurrent surrogate on the same phasor objective and score the same invariants—would show whether the branching pass comes from structure preservation or merely from stochastic dissipation.","Editorial inference: if the proposed criticality-control upgrade closes the spectral and DFA gaps, the same validation ladder could be applied to pathological regimes, interpreting seizure as dissipation failure and anaesthesia as metabolic collapse."],"forward_implications":["If the branching match is not a temperature artifact, a physics-constrained EEG model can reproduce a model-independent invariant it never trained on, lending support to cortical criticality as a genuine dynamical signature.","The two reported failures set quantitative targets: an excitatory–inhibitory criticality control and a volume-conduction-robust connectome should shallow the 1/f slope and pull the DFA exponent below 1.","The metriplectic reformulation replaces relaxation-to-silence with a non-equilibrium steady state, so resting cortex can sustain oscillation indefinitely—matching the waking brain.","Fluctuation-dissipation noise gives trial-to-trial variability and a single arousal axis, linking discrete EEG conditions to locations on a continuous temperature scale.","The neuroanatomical stimulation ports plus energy-shaping control provide a stability-guaranteed route to closed-loop neuromodulation, contingent on TMS-EEG perturbational validation."],"fun_headline_variants":["GNN port-Hamiltonian EEG twin: branching matches, 1/f and DFA don't","Port-Hamiltonian EEG twin: one invariant matched, two missed","Cortex model: branching fits, 1/f and DFA don't","Metriplectic cortex twin reproduces branching, not spectral slope","Physics-coded cortex model: criticality hit, spectral and correlation misses"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The avalanche-branching match rests on the unexamined choice of arousal temperature T=0.2 and the learned dissipation matrix; if that choice is doing the work, the port-Hamiltonian structure adds little beyond a stochastic oscillator.","fun_headline_variants_meta":{"raw":{"variants":["GNN port-Hamiltonian EEG twin: branching matches, 1/f and DFA don't","Port-Hamiltonian EEG twin: one invariant matched, two missed","Cortex model: branching fits, 1/f and DFA don't","Metriplectic cortex twin reproduces branching, not spectral slope","Physics-coded cortex model: criticality hit, spectral and correlation misses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001481,"raw_usage":{"total_tokens":5846,"prompt_tokens":860,"completion_tokens":4986,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":4885}},"tokens_in":604,"tokens_out":4986,"duration_ms":35621,"temperature":1.0,"reasoning_tokens":4885,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T07:13:33.002154+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the stochastic free-run with T swept from, say, 0 to 1 (or learned from data) and check whether σ stays near 1 across a broad range; also replace the learned R(x) with scalar dissipation and re-score. If branching collapses or moves with T, the invariant is an artifact of tuning rather than an emergent property.","supporting_citations":[],"review_version":2}