{"id":"134752d3-97c1-486a-a465-8da867ec11ca","arxiv_id":"2607.29434","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"In a driven active-matter swarm, the driver-induced change in entropy production and heat flow — not absolute dissipation — predicts reservoir-computing performance across dynamical regimes.","lead":"A simulation study of an active swarm reservoir shows that entropy-production measures — especially the change caused by the input signal — track how well the system predicts a chaotic Lorenz-63 time series. The authors offer thermodynamic metrics as a screening tool that could identify good physical computers without task-by-task evaluation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (9)'s sign-selected passive-bath heat is not thermodynamically derived; the paper's own admission that total EP is negative when driving is weak makes η_bath and CV(Q) results load-bearing on an unvalidated decomposition.","rationale":"The single most load-bearing concern is indeed the heat-flow identification in Eq. (9), as the reader identified. The paper's headline contribution is that entropy-production-based observables — especially the relative difference η_bath and the robustness measure CV(Q) — track reservoir performance. Both of these depend on Q, which is defined by selecting only the decelerating part of the non-conservative force. This choice is not derived from a microscopic bath model; it is a physically motivated postulate that guarantees Q > 0. The paper candidly admits in Sec. V that the total entropy production in Eq. (2) becomes negative in the undriven regime and for weak driving, and appeals to the neglected constituent entropy Ṡ_cons to restore the second law. That admission is not a minor caveat: it means the decomposition into system and bath entropy is incomplete, and the quantity η_bath is not established as a thermodynamic heat flow. If an alternative, thermodynamically consistent definition of heat (e.g., from a Langevin thermostat) changes the location of the maximum in η_bath or the sign of the CV(Q)-performance correlation, the central association would reduce to the more trivial statement that stronger input coupling improves performance. This is exactly the kind of unvalidated decomposition that can be tested by recomputation. Other secondary concerns — the post-hoc construction of W_d(1−c_d), the exclusion of the overdamped regime, and autocorrelation across parameter grids — are real but do not strike at the core claim as directly. The reader's CONDITIONAL verdict is appropriate: the paper is transparent and the issue is addressable by adding a consistent bath coupling, but until then the thermodynamic interpretation of the headline results remains conditional. I therefore see no reason to change the verdict, and I agree with the reader's identification of the weakest assumption.","tokens_in":23909,"tokens_out":9436,"duration_ms":103154,"concrete_test":"Replace Eq. (9) with a thermodynamically consistent heat-flow definition: add a Langevin thermostat (friction γ, temperature T) to the non-conservative forces and compute the bath heat as the work done by the thermal force, Q_bath = ∫ Σ (γ p_i/m_i + ξ_i(t))·(p_i/m_i) dt, or, minimally, recompute all results with the Θ-term removed, Q_all = -∫ Σ q̇_i·B_i dt. Repeat the speed-controller scan (Sec. IV) and the driver-force scan, recompute η_bath, CV(Q), and the correlations in Tables III/IV. If the near-critical peak in η_bath and the strong negative CV(Q)-performance correlation do not survive under either alternative, Eq. (9)'s sign-selected passive-bath postulate is the decisive unvalidated assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that driver-induced entropy-production observables track reservoir performance — rests on the identification of the bath heat flow in Eq. (9). There, only the component of the non-Hamiltonian forces opposing the particle velocity (q̇_i · B_i < 0) counts as heat delivered to a passive bath. This sign selection makes Q̇ strictly positive by construction and yields the distinctive near-critical peak in η_bath (Fig. 5) and the CV(Q) robustness claim (Fig. 4). But the paper itself states in Sec. V that Eq. (2), the total entropy production, 'is positive only where the driver dominates the system's dissipative response, turning negative when driving is weak or absent (small K_d, small r_d, undriven case, overdamped regime).' The rescue via neglected constituent entropy Ṡ_cons is an added assumption, not a derivation. If Eq. (9) does not correspond to the physical heat exchanged with the bath, then η_bath and CV(Q) are not thermodynamic measures; the association with performance could simply reflect that stronger driving produces more deceleration events, i.e., 'input work correlates with performance.' Since the abstract explicitly advertises 'dissipation coincide with peak performance' and 'heat flow provide complementary diagnostics,' this decomposition is the load-bearing element. The reader's weakest_assumption correctly pinpoints this.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript analyzes a deterministic active-swarm reservoir studied previously in the reservoir-computing literature and asks whether entropy-production observables can serve as screening metrics for computational performance. It introduces a Monte-Carlo estimator for the Gibbs entropy rate Ṡ_sys based on the generalized Liouville equation (Eqs. 24–31), proposes a heat flow Q̇ into a passive bath via sign-selected non-Hamiltonian work (Eq. 9), and performs two parameter scans: the speed-controller scan (K_sc, s) and the driver-force scan (K_d, r_d). The results compare driven vs undriven entropy, driver work, and coefficients of variation to Lorenz-63 prediction performance, with additional reproducibility checks on four other attractor drivers (Appendix B). The central claim is that the driver-induced contrast in entropy production, rather than absolute dissipation, marks the near-critically damped band as the best computational regime, making EP-derived quantities candidate screening metrics for physical reservoirs.","tokens_in":24146,"tokens_out":6227,"duration_ms":74953,"significance":"If the interpretation holds, the paper would provide practical thermodynamic diagnostics for physical reservoir computers and a general, parameter-free estimator for phase-space contraction in deterministic ODE systems. Strengths include: the generalized-Liouville estimator derivation is clean and machine-checkable; data are openly available; the correlation analysis is repeated across five driver signals; and the paper is unusually explicit about its own limitations. The main weakness is that the bath heat flow in Eq. (9) is an unvalidated Heaviside projection that makes Q̇ positive by construction, and the paper itself admits in Sec. V that the total entropy balance in Eq. (2) is negative when driving is weak or absent. Because η_bath, CV(Q), and the headline 'dissipation coincides with peak performance' rely on identifying this Q̇ with physical heat, the current thermodynamic language overstates what is established. With validation or a careful re-scoping of Q̇ as a deceleration-work diagnostic, the empirical associations may still be useful; as written, a load-bearing assumption remains unproven.","major_comments":[{"comment":"The heat flow Q̇ is defined by a Heaviside projection keeping only non-Hamiltonian force components that oppose a particle's velocity, so Q̇ ≥ 0 by construction and the bath never feeds back. The authors concede in Sec. V that Eq. (2) 'is positive only where the driver dominates the system's dissipative response, turning negative when driving is weak or absent (small K_d, small r_d, undriven case, overdamped regime)', and the rescue via neglected constituent entropy Ṡ_cons is an added assumption rather than a derivation. Since η_bath (Eq. 40), CV(Q) (Fig. 4), and the central claim that 'dissipation coincides with peak performance' all depend on Q̇ being a physical heat flow, this decomposition is load-bearing. Please either validate Eq. (9) in a thermodynamically consistent coupling (e.g., a Langevin limit or Nose–Hoover-style thermostat) or remove the thermodynamic interpretation and de","section":"Eq. (9), Sec. V"},{"comment":"The empirical core of the paper is the Pearson/Spearman correlations between performance and thermodynamic metrics, but no sample sizes, confidence intervals, or p-values are reported. Several entries are weak, e.g., W_d has Pearson r = 0.0301 in Table IV while Spearman ρ = 0.7001, and the text describes correlations as 'strong' without uncertainty quantification. Given a two-dimensional scan, a post-hoc capper filter, and multiple metrics, these correlations need N (number of parameter configurations retained), bootstrap intervals or p-values, and some awareness of multiple comparisons. Without this, the relative ranking of metrics in the tables cannot be assessed.","section":"Tabs. III–IV, Sec. IV D"},{"comment":"W_d is the work performed by the input channel on the reservoir, so a correlation between W_d and task performance is partly built in: a reservoir whose drive does no work cannot respond to the input. The paper partly acknowledges this by showing W_d alone fails in the driver-force scan and by introducing the heuristic W_d(1−c_d). To support the stronger conclusion that entropy-based metrics add predictive value beyond input power, the authors should provide a control (e.g., random or power-matched drive) or a partial-correlation analysis controlling for W_d. Otherwise the claim that EP, rather than input-coupling strength, tracks performance remains underdetermined.","section":"Eq. (41), Sec. IV D"},{"comment":"The quantity called the 'system entropy production' Ṡ_sys is the time derivative of the Gibbs entropy and is negative for every undriven and driven configuration shown in Fig. 2. A negative rate may be a legitimate open-system entropy change, but it is not an entropy-production rate. The paper's own second-law bookkeeping is left unresolved, and the title/abstract advertise 'entropy production' for an object that is negative. Please either rename the observable (e.g., phase-space contraction rate) or provide a full bookkeeping in which Ṡ_sys + Ṡ_env ≥ 0. The current ambiguity affects the interpretation of η_sys in Eq. (39), whose denominator is negative.","section":"Sec. II B, Eqs. (4)–(6), Fig. 2"}],"minor_comments":[{"comment":"The symbol D appears in Eqs. (32a)–(32b) without an explicit definition; if it is the spatial dimension, state this before the equation. The expression 'D Ka' in Eq. (32a) also appears garbled and should be typeset as a product or dimension factor.","section":"Eq. (32)"},{"comment":"The figure plots |η_sys| and |η_bath|, but the text discusses signs and directions. Please define the sign convention for η_sys and η_bath and explain why absolute values are used, especially because S_sys is negative over most of the parameter space.","section":"Fig. 5, Eqs. (39)–(40)"},{"comment":"The replacement of the tanh wrapper by the force capper and the global force rescaling by 20 shifts the K_sc axis relative to prior work. The text says this 'slightly shifts' the axis, but for reproducibility it would help to give an explicit mapping or a table of equivalent parameters.","section":"Sec. III, Eq. (15)"},{"comment":"The robustness tables in Appendix B report correlations without the same N and uncertainty information as the main tables. Please either include the same statistical details or note explicitly that the appendix is only a qualitative consistency check.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"I am not recommending rejection because the generalized-Liouville estimator is sound, the data are available, and the associations may well survive a re-scoped interpretation. The decision hinges on the load-bearing status of Eq. (9): either the authors validate the heat-flow decomposition against a thermodynamically consistent model, or they clearly re-label Q̇ and η_bath as 'deceleration work' diagnostics and soften the thermodynamic claims in the abstract and title. I would also ask for basic uncertainty quantification on the correlation tables before the empirical claims can be evaluated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a genuine empirical contribution, not a repackaging. The Liouville-based phase-space-contraction estimator (Eqs. 26-30) is clean, parameter-free, and more general than the swarm model; the driven/undriven contrast (η_sys, η_bath) is a new diagnostic; and they test it across five driver signals with unusually honest reporting. If you work in physical reservoir computing, this is worth your time.\n\nThe main finding is the negative result plus the positive one: absolute EP magnitude is not the predictor; the driver-induced change in EP, and driver work, track the performance landscape. The near-critical band already known to be optimal from the same group's prior work shows the sharpest driven/undriven contrast. So the paper does not discover a new regime; it provides a thermodynamic signature for a known one. That is a valid step, but worth flagging because the abstract's 'coincide' phrasing can overstate.\n\nThe soft spot is real and load-bearing. Eq. (9) defines bath heat using only decelerating non-Hamiltonian force components, which makes Q_dot positive by construction. The paper admits in Sec. V that the total entropy production is negative where driving is weak, and the rescue via neglected constituent entropy is an assumption, not a derivation. If the passive-bath split is wrong, η_bath and CV(Q) are not thermodynamic measures; the association could just be 'more driving produces more deceleration events,' i.e., input work correlates with performance. That does not sink the paper—the empirical association survives replication across signals—but it means the headline 'dissipation' claim is weaker than it looks. The post-hoc composite W_d(1-c_d) and the excluded overdamped regime are secondary issues; the core estimator and the contrast metrics are solid.\n\nVerdict: deserves a serious referee. The authors should be pushed to either implement a thermodynamically consistent bath coupling or reframe η_bath as a phenomenological response measure rather than a heat flow. With that revision, this could be a useful screening heuristic. I would bring it to a reading group and would cite it for the Liouville estimator and the contrast diagnostic.","headline":"A serious, transparent empirical study: driven/undriven entropy contrast tracks reservoir performance, but the headline heat metric rests on an unvalidated passive-bath split.","tokens_in":24748,"tokens_out":2005,"would_cite":true,"duration_ms":21602,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Entropy production tracks reservoir computing performance—but only the driver-induced part, not the absolute value.","keywords":["entropy production","reservoir computing","active matter","swarm reservoir","phase-space contraction","heat flow","Lorenz-63 prediction","near-critical damping"],"falsifier":"Replace the passive-bath heat-flow rule with a thermodynamically consistent bath coupling, e.g., a Langevin thermostat with an Einstein-relation diffusion constant, or explicitly model the constituents' internal entropy production ˙Scons. If the η_bath peak in the near-critically damped band disappears or its correlation with prediction performance drops below significance across both parameter scans, the paper's central claim is falsified. Alternatively, an experiment with synthetic active particles and controlled drag could measure whether the driven-undriven heat-flow contrast actually trac","tokens_in":23647,"feed_emoji":"🐝","tokens_out":2251,"duration_ms":29738,"temperature":0.7,"pith_summary":"This paper asks whether entropy production (EP) can serve as a physical indicator of computing performance in a driven active-matter reservoir. Simulating a swarm of particles that must predict a Lorenz-63 chaotic signal, the authors compute two separate entropy channels: the system's internal EP from phase-space contraction and the bath EP from heat flow. They find that performance does not track the absolute magnitude of EP, but rather the driver-induced change in the thermodynamic response—specifically the ratios η_sys and η_bath between driven and undriven system entropy and heat. The near-critically damped dynamical regime, already known to be the best-performing reservoir, is exactly where this driver-induced dissipation is sharpest and where the entropy metrics are most robust to initial conditions. A sympathetic reader would take this as evidence that EP-derived quantities, especially relative driven-undriven contrasts and driver work, offer a task-agnostic screening metric for physical reservoir substrates.","feed_headline":"Entropy contrast, not entropy, predicts reservoir performance","feed_subtitle":"Driver-induced dissipation peaks exactly where a swarm predicts chaos best, offering a training-free diagnostic.","key_machinery":"The argument runs on two entropy estimators plus a work measure. First, a Monte-Carlo estimator based on the generalized Liouville equation computes the system EP rate as the phase-space contraction rate, i.e., the sum of Lyapunov exponents, using only the deterministic flow field. Second, the bath EP is computed from a heat-flow identification in which only non-conservative forces opposing a particle's velocity—those with ˙qi·Bi < 0—count as heat delivered to a passive bath. The central comparative quantities are the relative differences η_sys and η_bath (Eqs. 39–40), the driver work W_d, and the coefficient of variation CV of the entropy metrics. The sharp contrast between driven and undri","core_discovery":"For a driven active-swarm reservoir performing Lorenz-63 prediction, computational performance is not set by the absolute entropy production rate. Instead, the predictive signature is the driver-induced change in the thermodynamic response: the relative difference in system entropy (η_sys) and in transferred heat (η_bath) between driven and undriven conditions peaks in the near-critically damped band, coinciding with peak prediction performance. In the undriven system this band is actually a minimum of transferred heat; under driving it becomes a maximum, indicating that the near-critical regime is distinguished by sensitivity to the driver, not by intrinsic dissipation. Driver work W_d also","pith_inferences":["If this holds, entropy-based screening could be applied to other physical reservoir candidates—photonic, mechanical, or chemical—where a clean Hamiltonian/non-Hamiltonian split exists, potentially replacing brute-force task sweeps.","The near-critical damping band may be an operational realization of 'edge of chaos': the entropic contrast between driven and undriven response could serve as a measurable edge-of-chaos detector for physical substrates.","A testable extension is to use η_bath or driver work as an online control signal, dynamically tuning K_sc or the driver coupling to keep the reservoir near the peak-entropy-response regime during operation.","The passive-bath assumption means the bathentropy estimate is only valid where the driver dominates dissipation; in weakly driven regimes the neglected internal constituent entropy likely becomes relevant, so a full entropy balance would be needed to sharpen the screening metric."],"forward_implications":["Entropy ratios and driver work could flag good computational regimes in physical reservoirs without running the full training and prediction task.","The near-critically damped band is shown to be special thermodynamically only under driving, not intrinsically—so thermodynamic diagnostics must compare driven and undriven states, not just measure dissipation.","Robustness to initial conditions, quantified by low CV of system entropy and heat flow, is itself a predictor of high reservoir performance.","The generalized-Liouville system-entropy estimator applies to any system of first-order ODEs, independent of the swarm model, making phase-space contraction a broadly usable diagnostic.","The same qualitative associations hold for four additional attractor-based driver signals, suggesting the effect is not specific to Lorenz-63 inputs."],"fun_headline_variants":["Driver-induced entropy gap flags best reservoir performance","Entropy contrast, not rate, marks computing sweet spot","Swarm computing peaks where entropy shift is sharpest","Predictive power tied to entropy change under driving","Reservoir performance peaks at maximal entropy response"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The heat-flow identification in Eq. (9) assumes only the decelerating part of the non-conservative forces transfers entropy to a passive bath that never couples back; if accelerating forces also exchange entropy with the environment, or the bath feeds back, then the computed bath entropy rate is not the true physical entropy flow and the central association may reduce to 'input work correlates with performance.'","fun_headline_variants_meta":{"raw":{"variants":["Driver-induced entropy gap flags best reservoir performance","Entropy contrast, not rate, marks computing sweet spot","Swarm computing peaks where entropy shift is sharpest","Predictive power tied to entropy change under driving","Reservoir performance peaks at maximal entropy response"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1171,"prompt_tokens":772,"completion_tokens":399,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":336}},"tokens_in":516,"tokens_out":399,"duration_ms":4449,"temperature":1.0,"reasoning_tokens":336,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T07:08:49.804412+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the passive-bath heat-flow rule with a thermodynamically consistent bath coupling, e.g., a Langevin thermostat with an Einstein-relation diffusion constant, or explicitly model the constituents' internal entropy production ˙Scons. If the η_bath peak in the near-critically damped band disappears or its correlation with prediction performance drops below significance across both parameter scans, the paper's central claim is falsified. Alternatively, an experiment with synthetic active particles and controlled drag could measure whether the driven-undriven heat-flow contrast actually trac","supporting_citations":[],"review_version":1}