{"id":"02716475-4ae8-4b12-8d05-06f616b13eb6","arxiv_id":"2604.01088","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"In a deterministic cellular automaton, the information needed to predict the final pattern is not accessible at the start and is instead constructed over time by emergent vortex and loop structures that reveal outcomes only near the end.","lead":"In a fully deterministic cellular world where every cell follows fixed local rules, the final fate is fixed from the first instant, yet no machine-learning model could predict it from the initial state. This paper shows the predictive information is genuinely built later, as vortices and winding loops self-organize and make some endings legible only at the end.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The strongest evidence for late-time predictability is retrospective: figures align trajectories to the final event and no out-of-sample forward classifier is tested, so the 'dynamically constructed' claim lacks a demonstrated temporal protocol.","rationale":"The reader's weakest assumption concerned completeness of the t=0 ML negative. That is a real concern and remains valid. But the single most load-bearing point for the central claim is the temporal comparison: the paper needs to show that fate becomes predictable at some intermediate time without using knowledge of the final event. The current evidence for late predictability is mostly retrospective. Fig. 4e aligns runs by the final event; Fig. 5 uses backward-time windows anchored to the final configuration. The only forward-looking visualization, Fig. 4f, is descriptive and lacks an actual prediction evaluation. Therefore the gap between determinism and predictability is not measured with one consistent yardstick. This does not invalidate the interesting topological mechanism or the extensive ML negative at t=0. It does mean the headline 'predictability is dynamically constructed' is not yet fully demonstrated; it remains a plausible but conditional conclusion. I therefore recommend keeping the reader's CONDITIONAL verdict rather than moving to ACCEPT or REJECT. The proposed prospective benchmark is a single, decisive check that would either support the claim or show that the late-time 'predictive signatures' are artifacts of end-aligned analysis.","tokens_in":48021,"tokens_out":7795,"duration_ms":100697,"concrete_test":"Run a prospective benchmark on held-out runs: at fixed observation times t (e.g., every 100 steps from t=0 to the maximum run length), train a simple classifier (logistic regression or gradient-boosted trees) on features computed from data up to t only—raw lattice snapshot, vortex counts, NCL-string count, sliding-window mean and slope—to predict the final fate class. Report balanced accuracy versus t on a frozen test set. If accuracy stays near chance until the final ~20 steps and then rises, the late-legibility claim survives. If accuracy rises only when features are aligned using knowledge of the final event (as in Figs. 4e and 5), the paper's late-time 'prediction' is retrospective and the dynamic-construction claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires showing that fate is unreadable at t=0 and becomes readable along the trajectory. The t=0 half is already provisional (Supp. Note 2.8 concedes untested architectures). But the more load-bearing problem is the late half. Fig. 4e and Figs. 5b–5e align trajectories by the 'final event' or 'final configuration' (backward time τ), so the observer effectively knows when the run ends. Fig. 4f uses a sliding window ending at t and does not require final-event knowledge, but the 'prediction' is only a colored scatter in (mean NCL count, slope); no classifier is trained on data available at time t, no balanced accuracy is reported as a function of time-to-event, and no held-out runs are used. The statement that 'an observer, without knowing how many timesteps remain, can use the forward-time dynamics of NCL strings to predict' is therefore not actually demonstrated. Similarly, the trajectory clustering at τ=20 uses the last 20 steps before a known final configuration. This is retrospective description, not forward prediction. The claimed temporal asymmetry between determinism and practical predictability is thus not measured with a consistent predictive protocol: chance-level ML at t=0 is compared with visually separated, end-aligned trajectories at late times.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a deterministic generalized cellular automaton of secrete-and-sense cells on a triangular lattice, whose trajectories end in one of three macroscopic fates (static configurations, rectilinear waves, or spiral waves). The authors report three main findings: (i) a battery of machine-learning classifiers, trained on up to 850,000 runs, cannot predict the binary static-versus-dynamic fate from the initial 14×14 configuration above chance; (ii) recoding the four cell states as a discrete phase field reveals topological objects—charged vortices, strings, and non-contractible loops (NCLs)—whose late-time dynamics differ among the three fate classes; and (iii) trajectory-similarity analysis of the total vortex-core-size time series shows that distinct histories collapse onto a small number of terminal profiles only in the final ~20 timesteps. The paper concludes that predictability is not a property of the initial condition but is dynamically constructed along the trajectory through self-organizing topological modes.","tokens_in":48184,"tokens_out":3285,"duration_ms":41998,"significance":"If the central claim were established with a consistent predictive protocol, the paper would make a significant conceptual contribution: a deterministic, non-chaotic many-body system in which practical predictability has a genuine temporal onset, carried by emergent topological degrees of freedom. The strengths are substantial. The machine-learning analysis is exemplary in design: a frozen test set of 100,000 runs, learning curves spanning three orders of magnitude in training data, threshold calibration on validation data, six model classes including spatially structured CNNs, and a model-agnostic mutual-information control. The late-time analyses are also internally consistent and replicated across 19–20 independent trajectory sets. The topological mechanism—pairwise vortex annihilation enforced by toroidal charge neutrality, with a proof in Supplementary Note 5—is a concrete, falsifiable structural finding. However, the paper's strongest interpretive claim is currently ahead of the evidence, because the late-time 'predictability' is demonstrated retrospectively, not with a forward prediction protocol, and the t=0 negative result carries an explicit caveat in Supplementary Note 2","major_comments":[{"comment":"The statement that 'an observer, without knowing how many timesteps remain, can use the forward-time dynamics of NCL strings to predict' (text near Fig. 4f) is not demonstrated. Fig. 4e aligns all runs by the final event (last vortex pair perishes or spiral forms), so the observer in this plot knows when each run ends. Fig. 4f is a colored scatter in (mean NCL count, slope) with no classifier, no held-out runs, and no reported balanced accuracy or ROC-AUC as a function of time-to-event. The opacity encoding (1–30 vs 31–60 timesteps before the final event) again uses future information. To support the forward-time claim, the authors should train a temporal classifier on features available at time t (e.g., sliding windows ending at t, without any final-event alignment) and evaluate its out-of-sample accuracy on held-out runs as a function of remaining time. This is a load-bearing gap becau","section":"Results, 'Non-contractible loops predict final pattern types' (Fig. 4e-f)"},{"comment":"The trajectory-similarity analysis is also retrospective. Every backward-time window [1, τ] is anchored at the final configuration, so the analysis requires knowing when the run ends. The collapse to 23 connected components by τ=20 and to 977 exact terminal profiles in the pooled set is a descriptive statement about trajectory geometry, not a demonstration that a predictor operating online—without knowledge of the terminal time—can classify fate. In addition, no quantitative cluster-purity or classification metric is reported for the components in Fig. 5c; the statement that clusters are 'segregated by final pattern type' is based on visual inspection. A forward metric (e.g., nearest-neighbor classification using only data up to time t, or a time-indexed balanced accuracy) is needed to establish that predictability is dynamically constructed rather than merely retrospectively described.","section":"Results, 'Fate-dependent microstructures emerge only in the final timesteps' (Fig. 5b-e)"},{"comment":"The t=0 half of the central claim is explicitly provisional. Supplementary Note 2.8 concedes that permutation-invariant architectures and graph neural networks were not tested and that 'we cannot exclude the possibility that alternative representations or substantially different model classes could uncover predictive structure not detected here.' The Discussion nevertheless states that 'the degrees of freedom that carry predictive signatures are not present in any accessible representation initially (at t = 0).' This overstates the evidence: the ML and NMI results support the weaker claim that no predictive signal was found within the tested model classes, representations, and data scales. If any representation-aware learner extracts above-chance signal from the initial configuration, the conclusion degrades from 'predictive structure is absent at t=0' to 'predictive structure exists at","section":"Supplementary Note 2.8; Discussion"}],"minor_comments":[{"comment":"There are repeated typographical artifacts, e.g., 'Y et' at the start of the Introduction and in a later paragraph. These should be corrected.","section":"Throughout"},{"comment":"The definition of 'final event' in panel (e) is given in the caption but should also appear in the main text where the panel is first referenced. As written, the reader must infer that the alignment uses information not available to an online observer.","section":"Fig. 4 caption"},{"comment":"The caption states that the color bar at right indicates final pattern type, but the mapping from colors to the three fate classes is not given in the caption or figure. Please add a legend.","section":"Fig. 5b caption"},{"comment":"The padding of shorter trajectories beyond their actual duration is described, but it would be helpful to state explicitly that all analyses are restricted to times within the actual duration for each trajectory, so padding cannot create spurious matches.","section":"Methods, 'Unique-trajectory analysis'"},{"comment":"The sentence 'predictability is constructed late and asymmetrically' is repeated almost verbatim in the following paragraph. Consider consolidating to avoid redundancy.","section":"Discussion"}],"recommendation":"major_revision","confidential_remarks":"The paper has real strengths, particularly the careful ML negative result and the topological mechanism. The main weakness is methodological asymmetry: a rigorous chance-level benchmark at t=0 is compared with retrospective, end-aligned descriptions at late times. This can likely be fixed within the manuscript's scope by adding a forward, out-of-sample temporal classifier and by qualifying the t=0 claim in line with Supplementary Note 2.8. I do not see a load-bearing error that would require rejection, but the central claim as currently worded is not yet supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things. First, the paper's t=0 negative result is unusually well done: six model classes, learning curves up to 850k samples, frozen test set, threshold calibration, shuffle controls, and a model-agnostic NMI check all converge on chance-level performance. Second, the late-time mechanism — NCL strings threading vortex pairs — is genuinely new and the asymmetry between annihilation-bound and spiral-bound runs is crisp. But the phrase 'predictability is dynamically constructed' is stated more strongly than the evidence supports, because most of the late-time analysis is retrospective.\n\nWhat is actually good: the vortex/NCL observables are a real addition to the coarse-graining literature. The pairwise annihilation constraint (charge conservation on the torus) is cleanly demonstrated and provable, and the Brownian particle model recapitulating termination times (rho = 0.95) is a nice reduction. The trajectory-collapse result — 38,000 distinct histories to 977 terminal profiles in the last 20 steps — is reproducible across 19 independent sets and makes the late-time-legibility point concrete. Code and data are up on GitHub; methods are detailed.\n\nWhere the soft spots are, in proportion: (1) The t=0 half is inherently provisional; Supp Note 2.8 concedes permutation-invariant and graph architectures were not tested. That is honest, but it means the strong claim 'not present in any accessible representation' should be 'not detectable by the model classes and representations we tried.' (2) The stress-test note is right: Fig 4e aligns trajectories to the final event, and Fig 4f provides a colored scatter, not a trained forward classifier with balanced accuracy as a function of time-to-event. The claim that an observer 'without knowing how many timesteps remain' can predict is therefore not actually demonstrated. A simple classifier on (mean NCL count, slope) at time t would fix this; without it, the assertion overreaches. The trajectory clustering of Fig 5 is likewise useful evidence of late-time convergence, but it is retrospective description. (3) The Brownian model fits sigma and r_threshold from the same data it matches — disclosed, but worth remembering when evaluating the correlation.\n\nNone of this sinks the core finding. The temporal asymmetry — chance at t=0, legible late — is real and backed by independent observables. The paper just needs to reframe the late-time results as 'predictive signatures become detectable' rather than 'an observer can predict.'\n\nWho it's for: statistical physicists and anyone working on computational irreducibility, cellular automata, or topological defects in discrete systems. A serious referee should engage with it; the fixes are mostly reanalysis and wording. I would accept it for peer review and would cite it, but I'd push for the forward-classifier test.","headline":"Solid computational study with a genuine topological mechanism; the late-time predictability evidence is real but partly retrospective, so the 'forward prediction' claim needs softening.","tokens_in":48849,"tokens_out":1805,"would_cite":true,"duration_ms":23136,"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":"In a fully deterministic cellular automaton, the final pattern is fixed from the start yet unpredictable from it, because the topological structures that make the outcome legible are constructed only as the dynamics unfold.","keywords":["deterministic predictability","cellular automata","self-organization","topological defects","vortices","non-contractible loops","machine learning","phase field"],"falsifier":"Retrain a graph neural network or other permutation-equivariant classifier on the same 850,000 training examples, preserving the frozen test set; if balanced accuracy on static-versus-dynamic classification exceeds chance using only the initial 14x14 configuration, the key claim—that predictive structure is not present in any accessible representation at t=0—is falsified.","tokens_in":47707,"feed_emoji":"🌀","tokens_out":8234,"duration_ms":80253,"temperature":0.7,"pith_summary":"This paper tries to establish that in deterministic, non-chaotic many-body systems, the future can be fixed yet practically unpredictable, because the structures that make the future legible—topological defects such as vortices and loops—do not exist at the start and are built during the dynamics. Using a generalized cellular automaton of secreting and sensing cells, the authors show that six families of machine-learning models, trained on up to 850,000 runs, cannot tell from the initial 14x14 configuration whether a run will end static or dynamic. By recoding cell states as a discrete phase field, they uncover charged vortices connected by strings that can wrap the torus as non-contractible loops; the loss of these loops announces vortex annihilation and hence static or rectilinear-wave outcomes, while spiral waves form abruptly with no advance signature. The load-bearing conclusion is that predictability is dynamically constructed and arrives late and asymmetrically, closing differently for annihilation-bound versus spiral-bound trajectories. If right, this reframes self-organization as the physical construction of predictive information, not the unfolding of a pre-encoded outcome.","feed_headline":"Predictability is built by vortex strings late in deterministic runs","feed_subtitle":"Machine learning can't guess final patterns from the start; only late-stage vortex dynamics reveal them.","key_machinery":"The key machinery is the recoding of the four cell states as phase vectors (0, π/2, π, 3π/2) on a toroidal lattice, turning the dynamics into a discrete phase field. This reveals three classes of vortices—+1, −1, and 0—whose total topological charge is conserved even though no update rule states it. Each vortex pair is threaded by same-state strings; a string that winds around the torus is a non-contractible loop (NCL). NCL strings carry the predictive signal: a sliding-window decline in their count announces imminent annihilation (static or rectilinear fate), while a near-one count accompanies spiral formation without warning. A supporting tool is an effective Brownian-particle model of ran","core_discovery":"Central claim: a deterministic system's macroscopic fate can be fixed by the initial state and still be unreadable from it, because the predictive degrees of freedom are collective topological structures that do not exist at t=0. Recoding the four cell states as phase vectors reveals charged vortices joined by strings of like-state cells, some of which form non-contractible loops around the toroidal lattice. The dynamics of these loops carry the result: as the last vortex pair approaches annihilation the loops vanish, making static and rectilinear fates readable in forward time; spiral-wave fates keep their loops until the wave forms abruptly. Since six machine-learning models and a mutual-i","pith_inferences":["The authors' caveat implies a testable falsifier: if a permutation-equivariant or graph neural network trained on the same data beat chance from the initial configuration, the conclusion would weaken to 'predictive structure exists but is hard to find' rather than 'predictive structure is absent at t=0.'","The Brownian-vortex-gas analogy suggests annihilation times in larger lattices should follow first-passage statistics of diffusing annihilating particles; pattern-forming biological tissues could be searched for analogous late-time topological signatures as early-warning signals.","If the mechanism generalizes, the common modeling assumption that macroscopic outcome is encoded in initial conditions may need to be replaced by a distinction between formal determination and physically accessible determination, with the latter depending on when topological structures coalesce.","For machine learning on physical data, the paper implies that static snapshot classification can be fundamentally limited even with perfect labels; adding temporal or topological features may help, but the usefulness of any added feature will itself be time-dependent."],"forward_implications":["Determinism alone does not guarantee practical predictability: a complete set of update rules can coexist with the absence of any tractable way to read the outcome from the current state.","The negative machine-learning result is not a data or capacity failure inside the tested classes: more data, deeper models, and spatial inductive biases all saturate at chance, and feature-wise mutual information collapses to the shuffled baseline.","Vortex annihilation, and therefore static and rectilinear-wave outcomes, is predictable in forward time from the progressive loss of non-contractible loop strings; spiral-wave formation is not foreshadowed and remains heterogeneous down to the final timestep.","Globally distinct histories collapse to a few hundred exact terminal profiles in the final ~20 timesteps, a 97.4% reduction in trajectory diversity, showing the late-time dynamics are far more constrained than the full evolution.","The same minimal ingredients—cyclic internal states, finite-range interactions, and topological constraints—should make late-constructed predictability a general class property across excitable media, coupled oscillators, and lattice models with discrete rotational symmetry."],"fun_headline_variants":["Fate becomes readable only when vortex strings self-organize","Predictability emerges late—via topological strings","Vortex strings decide which futures become legible","Hidden in the start: topology decides future legibility","Vortex loops make unpredictable futures legible late"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper's claim that no predictive structure exists in the initial configuration rests on the assumption that the six tested model classes and the per-cell mutual information cover all practically accessible ways of reading the lattice; the authors' own Supplementary Note 2.8 concedes that permutation-invariant and graph architectures were not tested, so if any such representation-aware learner extracted above-chance signal from the initial state, the strong 'absent at t=0'","fun_headline_variants_meta":{"raw":{"variants":["Fate becomes readable only when vortex strings self-organize","Predictability emerges late—via topological strings","Vortex strings decide which futures become legible","Hidden in the start: topology decides future legibility","Vortex loops make unpredictable futures legible late"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000962,"raw_usage":{"total_tokens":3908,"prompt_tokens":694,"completion_tokens":3214,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":3141}},"tokens_in":438,"tokens_out":3214,"duration_ms":20839,"temperature":1.0,"reasoning_tokens":3141,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T05:35:31.996535+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain a graph neural network or other permutation-equivariant classifier on the same 850,000 training examples, preserving the frozen test set; if balanced accuracy on static-versus-dynamic classification exceeds chance using only the initial 14x14 configuration, the key claim—that predictive structure is not present in any accessible representation at t=0—is falsified.","supporting_citations":[],"review_version":1}