{"id":"41d16d06-3748-4782-bf0e-406d9eec9635","arxiv_id":"2608.06290","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"The paper proposes the Latent Memory Table, a reusable summary of longitudinal histories, but its key empirical validation is undermined by an unfair baseline comparison.","lead":"A new statistical object, the Latent Memory Table, is proposed to summarize long athlete-monitoring histories so the summary can be saved and reused in later analyses. The paper tests it on soccer monitoring data and simulations, but the main comparison is unfair because the new table can see the target outcome while the comparison baseline cannot.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The P6 incremental-value test is an unfair comparison: the learned latent table is trained on the target wellness channel while the classical baseline excludes it, so the reported R2 gains may reflect outcome autocorrelation rather than reusable memory.","rationale":"The paper's central claim is that a P-validated Latent Memory Table is a reusable statistical object. The strongest empirical support is the SoccerMon Q advantage and the P6 incremental R2. Both rest on the comparison described in Section 7. The learned encoder is trained on the full joint history, including the wellness channel that later serves as prediction target. The classical baseline explicitly excludes the outcome channel. This makes the comparison asymmetrical: current and lagged values of the target are highly autocorrelated, so including them in the representation can raise next-window R2 without adding any memory beyond autocorrelation. The reader's weakest assumption identifies exactly this. I regard it as the most load-bearing concern because P6 is the only property that directly tests reusability in the empirical case study and because Q aggregates P6 into the headline 0.73 vs 0.40. The simulation, which is the more controlled evidence, does not show the learned table beating classical summaries on recovery (Procrustes 0.39 vs 0.46 in Block I), so the empirical Q advantage carries most of the reusability claim. I also note two corroborating weaknesses: P2 is scored 0.62 although the defining within-vs-between inequality fails for same-regime cross-athlete pairs (0.622 vs 0.661), and P5 is scored 0.85 despite the paper admitting only a partial sensitivity grid. These do not replace the main concern but they reinforce that the composite Q is not a reliable oracle. My proposed test directly removes the asymmetry: augment the classical baseline with lagged outcomes and/or train the encoder without the target channel. If incremental R2 disappears, the central claim loses its main empirical support; if it survives, the reader's objection is settled in the paper's favor. Since this confirms rather than changes the reader's REJECT verdict, I set verdict_should_be to UNCHANGED and agreement to agree.","tokens_in":17656,"tokens_out":5884,"duration_ms":66472,"concrete_test":"Recompute Table 7 with two changes: (a) add the lagged target outcome (e.g., previous-window fatigue, sleep quality, readiness, and an EWMA of the target) to the classical baseline; (b) train a second encoder on all channels except the target outcome and use its PCs in the combined model. Keep the same player-grouped splits. If the latent PCs no longer improve held-out R2 over the augmented baseline for more than one target, P6 and the Q advantage are not established; if the gains persist under both changes, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical evidence for reusability (P6) is the incremental held-out R2 in Section 7 / Table 7. The latent PCs are computed from bmi(t), which is estimated by a Transformer trained on all twelve channels, explicitly including the wellness outcomes (fatigue, sleep quality, readiness) that are then used as next-window targets. The classical baseline is described as 'classical load/wellness summaries excluding the outcome channel.' Thus the comparison confounds memory-table value with autoregression: a table containing current fatigue trivially helps predict next-window fatigue. The reported gains (e.g., sleep quality 0.54 vs 0.29; readiness 0.19 combined) are therefore not evidence of incremental longitudinal information beyond a baseline that has access to the same lagged outcome. Because P6 contributes to Q (Reuse score 0.60 in Table 8), the Q=0.73 vs 0.40 advantage is likewise inflated. A secondary internal inconsistency compounds this: P2 is scored 0.62 even though same-athlete cosine (0.622) is below different-athlete/same-regime cosine (0.661), so the defining inequality in Definition 3.5 fails; the simulation also does not favor the learned table (Block I Procrustes 0.39 vs classical 0.46). The paper's admitted limitations (no LSTM/linear baseline robustness; Q as a profile) do not repair these load-bearing issues.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new unit of analysis for longitudinal data, called the Latent Memory Table (LMT): a table of estimated latent states and uncertainties, produced by a learned memory operator over masked windowed histories, intended to be stored, queried, and reused in the same way as PCA scores or estimated random effects. The method is validated through a six-property system P1–P6 with a composite quality index Q, on a SoccerMon athlete-monitoring case study and on simulations with known memory mechanisms. The paper reports that the learned ensemble table attains Q ≈ 0.73 versus ≈ 0.41 for classical summaries and ≈ 0.39 for lagged PCA, and that the LMT adds incremental held-out predictive value for some wellness targets beyond classical summaries.","tokens_in":17838,"tokens_out":5859,"duration_ms":69829,"significance":"The concept of an exported, validated table of latent memory states is an interesting and potentially useful framing for longitudinal statistics, and the paper includes reproducible code and a clearly stated validation philosophy. If the empirical claims were sound, the proposal could give practitioners a principled alternative to ad hoc embeddings or task-specific features. However, the central evidence for the table's reusability and superiority over classical summaries is undermined by the flawed P6 comparison, an internally inconsistent P2 score, and simulation results in which baseline methods recover the oracle memory better than the learned encoder. The paper's strengths—a formal property system, a composite index, simulation diagnostics, and code release—do not compensate for the fact that the headline quantitative claims currently rest on invalid comparisons.","major_comments":[{"comment":"The incremental-value comparison is not a fair test of the LMT's reusability. The latent PCs are computed from bmi(t), whose encoder was trained on all twelve channels including the wellness outcomes (fatigue, sleep quality, readiness) that then serve as next-window prediction targets. The classical baseline, by contrast, explicitly excludes the outcome channel. Under this design, the reported R2 gains (e.g., sleep quality 0.539 vs 0.286) may simply reflect autocorrelation of the target rather than the incremental longitudinal information in the table. Because the Reuse score (0.60 in Table 8) feeds directly into Q, the Q advantage of 0.734 vs 0.407 is also inflated. A valid comparison would include the outcome channel, or its lagged values, in the classical baseline (e.g., an AR(1) term for each target).","section":"Section 7, Table 7 (P6)"},{"comment":"The personalization property P2 is defined as a strict inequality: within-athlete similarity must dominate cross-athlete similarity after removing adjacency. The reported median cosines are 0.622 for same-athlete pairs and 0.661 for different-athlete/same-regime pairs, so the defining inequality fails. Yet Table 8 assigns the learned table a personalization score of 0.62, approximately equal to the same-athlete cosine. This is internally inconsistent: the paper either needs to revise Definition 3.5 (for example, to require within-athlete similarity to dominate only cross-athlete/different-regime similarity) or correct the score. As written, the paper's own numbers show that the learned table is not P2-valid, and this also affects the composite Q.","section":"Section 6, Definition 3.5, Table 4, Table 8 (P2)"},{"comment":"The simulation results do not support the claim that the learned LMT recovers known memory mechanisms better than baselines. In Block I, the Transformer's Procrustes correlation with the oracle memory is 0.39, while classical and PCA-lagged baselines achieve 0.46 and 0.47, respectively. In Block II, the Transformer again scores lower (0.40) than classical (0.47). The text acknowledges 'partial recovery' but does not directly address the fact that the baselines recover the oracle memory more accurately. Since the paper's central argument is that the LMT deserves the same status as PCA scores or random effects, a simulation in which PCA-lagged windows outperform the learned state is a direct counterexample that must be confronted, not merely reported in a table.","section":"Section 8, Table 11 (Simulation recovery)"},{"comment":"The P1 recoverability property is evaluated using the same Team×Season classification criterion that is used as the auxiliary training objective. Although the evaluation is on held-out athletes, the test essentially verifies whether the auxiliary training succeeded, rather than providing independent evidence that the table retains generalizable longitudinal memory beyond the chosen supervision. This circularity weakens the recoverability claim and should be acknowledged explicitly; for example, the authors could evaluate P1 with a held-out regime rotation or a different auxiliary task not used in training.","section":"Section 4 and Section 6 (P1 recoverability)"}],"minor_comments":[{"comment":"The text describing the three-way cosine comparison states 'Same-athlete similarity remains far above cross-regime similarity', but the more relevant comparison to different-athlete/same-regime similarity (0.622 vs 0.661) is not highlighted; this should be discussed to avoid misleading readers about personalization.","section":"Section 6, Figure 3"},{"comment":"Definition 3.6 says a table is P-valid if it satisfies all six properties, yet the paper later reports Q as an equally weighted average and uses it to rank tables. If P-validity requires all properties to hold, then the learned table, by the paper's own numbers, is not P-valid because P2 fails; the relationship between P-validity and Q should be clarified.","section":"Section 3, Definition 3.6"},{"comment":"The sensitivity analysis covers only the latent dimension d; the paper states in Section 11 that robustness to LSTM/linear baselines remains future work. This is acceptable as a limitation, but the claim of 'stability' (P5) is only partially demonstrated.","section":"Appendix F"},{"comment":"The statement 'Alternative weights are sensitivity analyses' is not accompanied by any reported sensitivity results; either a small table or a reference to an appendix would strengthen this point.","section":"Section 3, Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is broad, and the current evidence does not support it. However, the flaws are fixable in principle: the P6 comparison can be made fair by including lagged outcomes in the baseline, the P2 score can be brought in line with the definition, and the simulation results can be discussed more honestly. I therefore recommend major revision rather than outright rejection, although if the authors are unable to address the P6 and P2 issues, the paper's conclusion would not be credible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere is my read of arXiv:2608.06290. The core idea—export a learned latent state table with uncertainty as a reusable statistical object, validated through a property system—is genuinely worth discussing. The paper is well written, honest about its limitations, and the software architecture that separates encoder estimation from downstream analysis is a good practical contribution.\n\nThe central evidence, however, does not hold up. The P6 incremental-value comparison is unfair. The latent table is trained on all twelve channels, including the wellness outcomes that later serve as targets, while the classical baseline explicitly excludes the outcome channel. The reported R2 gains, especially the sleep-quality jump, are likely autocorrelation artifacts rather than evidence of reusable memory. Because P6 feeds directly into Q, the Q=0.73 versus 0.40 advantage on SoccerMon is inflated.\n\nThere is also an internal inconsistency in P2. Definition 3.5 requires within-athlete similarity to exceed cross-athlete similarity. In Section 6, same-athlete cosine is 0.622, while different-athlete same-regime cosine is 0.661. The defining inequality fails, yet Table 8 scores P2 at 0.62. The high subject-identity classification accuracy may be regime-driven. The score needs a transparent operationalization.\n\nThe simulation does not rescue the claim. Under known memory mechanisms, the Transformer's Procrustes recovery is worse than classical (0.39 vs 0.46 in Block I), and Q is equal. So the empirical case for the table's superiority rests entirely on the flawed SoccerMon comparison.\n\nWhat the paper does well: the negative-control simulation is a good idea, the Procrustes ensemble for uncertainty is a nice touch, and framing classical EWMA as a degenerate memory operator is useful. The paper also avoids overclaiming endpoint prediction.\n\nNet: the LMT is a promising proposal, but the current evidence does not establish it as a reusable unit of analysis. A fair baseline that includes lagged outcomes, a proper P2 test, and a simulation showing at least parity for the learned table would make the case. As it stands, I would not cite it as a validated method, but the proposal deserves a serious referee if the author is willing to fix the baseline.","headline":"Interesting proposal for a reusable latent state table, but the key empirical support rests on an unfair baseline comparison and a P2 score that contradicts the paper's own definition.","tokens_in":18471,"tokens_out":2619,"would_cite":false,"duration_ms":27445,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H25","62M10","62P10","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes the Latent Memory Table as a reusable unit of analysis for longitudinal data, arguing that methods should be compared by the quality of the exported table, not by the encoder that produced it.","keywords":["Latent Memory Table","longitudinal data","unit of analysis","memory operator","property system","quality index","athlete monitoring","multivariate time series"],"falsifier":"Re-run the SoccerMon P6 comparison with a classical baseline that includes the outcome's own recent history (e.g., lagged outcome or an EWMA of it): if latent PCs no longer add held-out R² beyond that baseline, the reusability advantage over classical summaries fails. Conversely, in the simulation framework, add a univariate-memory scenario where the true memory depends on a single channel; if the learned table still scores Q≈0.73 while classical summaries score 0.40, that would suggest Q is inflated by training on the outcome channel.","tokens_in":17275,"feed_emoji":"📊","tokens_out":7684,"duration_ms":79469,"temperature":0.7,"pith_summary":"This paper proposes a new unit of analysis for longitudinal data: the Latent Memory Table, a rectangular export of latent states that summarize each unit's recent windowed history, each with an attached uncertainty estimate. The claim is that this table deserves the same treatment as PCA scores or estimated random effects—stored, queried, and reused across analyses—so methods should be compared by the table's quality, not by the encoder that produced it. To support that claim, the paper defines a six-property validation system with a composite quality index Q, shows in simulation that Q discriminates genuine multivariate memory from negative controls and reacts to window misspecification, and reports a soccer-monitoring case study in which the learned table attains Q≈0.73 versus about 0.40 for classical and lagged principal-component baselines. If the claim is right, one validated table can feed many downstream statistical operators without retraining, making the exported table the scientific product rather than the model.","feed_headline":"Latent Memory Table becomes the reusable unit of longitudinal analysis","feed_subtitle":"The learned table scores 0.73 versus 0.40 for classical summaries on athlete monitoring data.","key_machinery":"The central object is the Latent Memory Table T, collected from a memory operator Mθ : HW → Rd that maps a masked, windowed history Hi,t to a d-dimensional latent memory state mi(t), together with per-row ensemble covariance bΣi(t) obtained from a Procrustes-aligned replicate ensemble. The table carries the scientific content, and its validity is assessed by a property system P = {P1, ..., P6} aggregated into the composite quality Q(T) = (1/6) Σ Sj(T). Because latent coordinates are rotation non-identifiable, comparisons use geometry after orthogonal Procrustes alignment rather than raw coordinates, which is why the export is a table of states with uncertainty instead of a named-axis embedding.","core_discovery":"The central claim is that the Latent Memory Table T = {(bmi(t), bΣi(t)) : i = 1,...,N, t ∈ Ti}, formed by a memory operator Mθ mapping each masked windowed history to a finite-dimensional state together with its ensemble covariance, is a reusable unit of analysis on the same footing as principal-component scores or estimated random effects. Methods should be compared by the quality Q of the table, which is an equally weighted aggregate of six property scores measuring recoverability, personalization, temporal coherence, interpretability, stability, and reusability, rather than by the encoder or by endpoint accuracy. Classical fixed summaries such as exponentially weighted moving averages are shown to be degenerate univariate special cases of the same operator class. Empirically, the constructed table attains Q≈0.73 versus about 0.40 for classical and lagged-PCA baselines on the SoccerMon case study, with incremental held-out value for some wellness targets, while simulations with known memory mechanisms show that Q, not regime classification accuracy alone, separates genuine memory from negative controls and responds to window misspecification.","pith_inferences":["Editorial inference: the reported reusability gains rest on comparing latent states trained on all twelve channels, including wellness outcomes, against classical summaries that exclude the outcome channel; a baseline that included lagged outcomes might erase the incremental R², so the reusability advantage should be re-tested with outcome-aware baselines.","Editorial inference: if the Latent Memory Table becomes a standard export, the same P/Q ladder could be applied to clinical or device-monitoring panels, and the framework predicts that Q, not endpoint accuracy, is the quantity that transfers across settings.","Editorial inference: the simulation finding that Q peaks near the true memory horizon suggests window length is an empirical property of the data, so practitioners could use Q-sensitivity grids to choose windows instead of importing domain conventions.","Editorial inference: the Procrustes ensemble provides per-row uncertainty, and a natural next step the paper does not take is to propagate that uncertainty into downstream models via measurement-error or Bayesian mixed models on table rows."],"forward_implications":["Any sequence encoder—Transformer, LSTM, linear map, or PCA of lags—is only an estimator of the memory operator, so comparisons should be made by the quality of the resulting table, not by the architecture.","Classical EWMA-style workload summaries and acute:chronic ratios are degenerate cases of the same operator class, so the framework subsumes standard sports-science dashboards rather than competing with them.","A P-validated table licenses a family of statistical operators—paths, anomaly scores, neighbourhoods, partitions, and covariate use in mixed, survival, or functional models—without retraining the encoder.","Regime classification accuracy is not sufficient to certify a memory table: in the negative-control simulation, structure stays high while interpretability and state recovery collapse, catching a failure that accuracy alone would miss.","On SoccerMon, the learned table's quality Q≈0.73 substantially exceeds classical summaries (≈0.41) and lagged PCA (≈0.39), with incremental held-out R² gains for fatigue and readiness and a latent-only advantage for sleep quality."],"supporting_citations":[{"why":"Supplies the SoccerMon dataset analyzed in the empirical case study, providing the paper's main empirical domain.","marker":"Midoglu et al., 2024"},{"why":"Defines the training–injury paradox and the classical workload summaries used as baseline in the P6 comparisons.","marker":"Gabbett, 2016"},{"why":"Establishes the acute:chronic workload ratio as a classical scalar memory summary against which the learned table is compared.","marker":"Hulin et al., 2016"},{"why":"Provides the EWMA formulation of acute:chronic ratios, the prototype degenerate memory operator.","marker":"Murray et al., 2017"},{"why":"Documents mathematical coupling in classical workload ratios, motivating a multivariate learned memory over fixed scalars.","marker":"Lolli et al., 2019"},{"why":"Supplies the masked-sequence encoder lineage for irregular multivariate time series that the memory-operator estimator extends.","marker":"Che et al. 2018"},{"why":"Provides the generalized Procrustes analysis used to align replicate state tables across random seeds, the basis of per-row uncertainty.","marker":"Gower, 1975"},{"why":"Supplies the GAMM framework used to test P4 interpretability via partial smooths of PC coordinates on domain constructs.","marker":"Wood, 2017"},{"why":"Names the joint-model framework that rows of the table are intended to enter as time-varying covariates under P6.","marker":"Rizopoulos, 2012"},{"why":"Provides the joint longitudinal–survival modelling background that motivates reusability of the exported table.","marker":"Tsiatis and Davidian, 2004"}],"fun_headline_variants":["Latent Memory Table: a reusable unit for athlete longitudinal data","New memory table beats classical summaries in athlete study","Athlete data: Latent Memory Table scores 0.73 vs 0.40","Focus on the table, not the model: athlete memory states","Why Latent Memory Tables are the new reusable statistics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reusability claim depends on a fair comparison: latent states are built from all channels including the wellness outcome being predicted, while the classical baseline excludes that channel, so the reported gains could partly reflect outcome autocorrelation rather than the value of the table.","fun_headline_variants_meta":{"raw":{"variants":["Latent Memory Table: a reusable unit for athlete longitudinal data","New memory table beats classical summaries in athlete study","Athlete data: Latent Memory Table scores 0.73 vs 0.40","Focus on the table, not the model: athlete memory states","Why Latent Memory Tables are the new reusable statistics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00039,"raw_usage":{"total_tokens":2105,"prompt_tokens":1046,"completion_tokens":1059,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":662,"completion_tokens_details":{"reasoning_tokens":972}},"tokens_in":662,"tokens_out":1059,"duration_ms":11849,"temperature":1.0,"reasoning_tokens":972,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:21:54.264055+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the SoccerMon P6 comparison with a classical baseline that includes the outcome's own recent history (e.g., lagged outcome or an EWMA of it): if latent PCs no longer add held-out R² beyond that baseline, the reusability advantage over classical summaries fails. Conversely, in the simulation framework, add a univariate-memory scenario where the true memory depends on a single channel; if the learned table still scores Q≈0.73 while classical summaries score 0.40, that would suggest Q is inflated by training on the outcome channel.","supporting_citations":[{"cited_title":"A., Johansen, D., Riegler, M., Halvorsen, P.\\ and Hicks, S.\\ (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the SoccerMon dataset analyzed in the empirical case study, providing the paper's main empirical domain."},{"cited_title":"J.\\ (2016)","cited_arxiv_id":null,"evidence_quote":"Defines the training–injury paradox and the classical workload summaries used as baseline in the P6 comparisons."},{"cited_title":"T., Gabbett, T","cited_arxiv_id":null,"evidence_quote":"Establishes the acute:chronic workload ratio as a classical scalar memory summary against which the learned table is compared."},{"cited_title":"B., Gabbett, T","cited_arxiv_id":null,"evidence_quote":"Provides the EWMA formulation of acute:chronic ratios, the prototype degenerate memory operator."},{"cited_title":"M., Hawkins, R., Kelly, D","cited_arxiv_id":null,"evidence_quote":"Documents mathematical coupling in classical workload ratios, motivating a multivariate learned memory over fixed scalars."},{"cited_title":"Recurrent neural networks for multivariate time series with missing values","cited_arxiv_id":null,"evidence_quote":"Supplies the masked-sequence encoder lineage for irregular multivariate time series that the memory-operator estimator extends."},{"cited_title":"C.\\ (1975)","cited_arxiv_id":null,"evidence_quote":"Provides the generalized Procrustes analysis used to align replicate state tables across random seeds, the basis of per-row uncertainty."},{"cited_title":"N.\\ (2017)","cited_arxiv_id":null,"evidence_quote":"Supplies the GAMM framework used to test P4 interpretability via partial smooths of PC coordinates on domain constructs."},{"cited_title":"Joint Models for Longitudinal and Time-to-Event Data","cited_arxiv_id":null,"evidence_quote":"Names the joint-model framework that rows of the table are intended to enter as time-varying covariates under P6."},{"cited_title":"A.\\ and Davidian, M.\\ (2004)","cited_arxiv_id":null,"evidence_quote":"Provides the joint longitudinal–survival modelling background that motivates reusability of the exported table."}],"review_version":1}