{"id":"c4b9c492-160b-4501-834d-7a7dcae7290d","arxiv_id":"2608.08859","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"ORCA dynamically selects the temporal receptive field at inference time using lightweight signal statistics, matching fixed-context anomaly detection baselines without manual tuning of the temporal horizon.","lead":"ORCA is a lightweight anomaly detection system for wearable body sensors that dynamically chooses how much past context to use at each moment, instead of relying on one fixed window size. It matches the best fixed-window baselines while removing the need to hand-tune the window in advance, which matters for edge devices with tight memory and latency budgets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MIMIC-IV regime-stratified AUROC in Table S5 (~0.36–0.44) contradicts overall 0.9977 in Table 1, undermining the claimed 'no performance collapse'.","rationale":"The reader's weakest assumption concerned the unreported thresholds tau_vol and tau_corr in the agentic policy. While that is a real reproducibility gap, the more load-bearing issue is an internal inconsistency in the reported results. The central claim of robust generalization on MIMIC-IV is directly contradicted by the paper's own supplementary Table S5, which reports near-chance AUROC in medium and high volatility regimes. This is not a matter of missing hyperparameters; it is a conflict between two reported empirical results that cannot both be true without an explanation. Since the abstract and main text explicitly highlight 'conservative generalization behavior without performance collapse', the inconsistency strikes at the core of the paper's contribution. In good faith, I do not assume the discrepancy is intentional, but it must be resolved before any verdict beyond conditional acceptance can be justified. The reader's conditional verdict is therefore reinforced, but the condition should be expanded to require reconciliation of Table 1 and Table S5 (and the related WBAN inconsistency between Table S3 and Figure S1).","tokens_in":13940,"tokens_out":5311,"duration_ms":47283,"concrete_test":"Run the ORCA implementation on the same MIMIC-IV test split and compute AUROC both overall and stratified by the same volatility bins used in Table S5. If the stratified values match Table S5 while the overall value does not match Table 1, the claimed overall performance is contradicted by the paper's own regime-level data.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 6.1 and Table 1 report MIMIC-IV AUROC 0.9977 (adaptive k) and describe the result as 'strong detection performance' and 'without performance collapse'. Supplementary Table S5 reports MIMIC-IV AUROC stratified by volatility regimes: medium volatility ORCA = 0.4448, high volatility ORCA = 0.3611; fixed baselines are similar (0.4339–0.4442 and 0.3611–0.3709). The caption states 'Absolute performance is lower due to weakly supervised labels and heterogeneous clinical dynamics'. Even if the overall AUROC is dominated by an unreported low-volatility subset, the claim that ORCA exhibits 'conservative generalization behavior without performance collapse' is false in exactly the nonstationary medium/high volatility regimes the method targets. The two tables cannot both be correct without explicit reconciliation (e.g., mixing proportions, label distribution, or a typo). As written, the empirical support for the MIMIC-IV pillar of the central claim is internally inconsistent: either Table 1 is wrong or Table S5 is wrong. A related inconsistency appears in the WBAN supplementary data: Table S3 lists identical AUROC (0.9991) across all k and ORCA in medium volatility, while Figure S1's narrative claims fixed-k shows a noticeable drop in that regime. These contradictions are unexplained and prevent verification of the claimed adaptive advantage.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ORCA, an anomaly detection framework for multivariate wearable/clinical time series in which a supervisory controller dynamically selects the temporal receptive field (k in {1,3,5}) at inference time from lightweight statistics (variance and autocorrelation). The controller is not trained and adds no parameters; the detection module combines a MinGRU backbone with constant-normalization attention and a relative positional bias truncated to the selected k. The authors evaluate ORCA on a custom WBAN dataset and a MIMIC-IV-derived clinical benchmark, reporting AUROC/AUPRC near 0.99 in Table 1 and claiming that adaptive selection matches the best fixed-context baseline while removing the need to tune the temporal horizon. Supplementary material includes algorithm listings, proofs of gradient preservation and softmax-equivalence conditions, an ablation study, latency/memory analyses, and regime-stratified results.","tokens_in":14227,"tokens_out":4495,"duration_ms":45454,"significance":"If the central claim holds, ORCA is a useful practical contribution: it addresses a real deployment problem for edge WBANs, adds no trainable parameters, and could reduce hyperparameter tuning for temporal context. The paper is also transparent in shipping a public code repository and in providing explicit algorithmic pseudocode. However, the empirical support for the main claim is undermined by internal contradictions between the main tables and the supplementary tables, and by missing values for the controller thresholds. The near-saturated aggregate metrics also make it difficult to verify the claimed benefit of adaptation. In its current state, the paper is not yet publishable, but the core idea is defensible and the empirical issues appear addressable with additional reporting and analysis.","major_comments":[{"comment":"The MIMIC-IV results are internally inconsistent. Table 1 reports AUROC 0.9977 for adaptive k and the text says ORCA shows 'conservative generalization behavior without performance collapse,' but Table S5 reports AUROC 0.4448 in the medium-volatility regime and 0.3611 in the high-volatility regime for the same ORCA model. A pooled AUROC of 0.9977 cannot be reconciled with regime AUROCs below 0.45 unless the low-volatility regime overwhelmingly dominates the test set. The paper must report regime sizes, label fractions, and the pooled computation; as written, either Table 1 or Table S5 is wrong. In addition, Table 2 reports full ORCA on MIMIC-IV as 0.9983, which differs from the 0.9977 in Table 1; the discrepancy must be explained as well.","section":"Section 6.1, Table 1 vs. Table S5"},{"comment":"The adaptive policy is under-specified because tau_vol and tau_corr are never reported. These thresholds are the parameters that determine whether k=1, k=5, or k=3 is selected at each window. Without their values or a description of how they were chosen, the method cannot be reproduced, and the claim that ORCA eliminates the need to tune temporal horizons is incomplete, because the thresholds themselves are free parameters that presumably required some selection procedure. Please report the threshold values, the selection method, sensitivity analysis, and whether a single threshold setting works across both datasets and all signal regimes.","section":"Section 4.2 and Supplementary Algorithm 2, Eq. (S5)"},{"comment":"The WBAN regime analysis contradicts itself. Table S3 reports identical AUROC (0.9991) for k=1, k=3, k=5, and ORCA in the medium-volatility regime, while the text around Figure S1 claims that fixed receptive field configurations exhibit a 'noticeable drop' in this same regime. If the fixed baselines do not actually drop, the claimed benefit of agentic adaptation in the medium-volatility regime is not demonstrated. If they do drop, Table S3 is incorrect. The table and the narrative must be reconciled before the results can be interpreted.","section":"Section G.3, Table S3 and Figure S1"},{"comment":"The ablation study is not discriminating: all variants, including random k, achieve AUROC within 0.0001 of each other on both datasets (e.g., 0.9982-0.9983 on MIMIC-IV). The paper's claim that a random policy 'fails to produce systematic gains' is unsupported, because no variant produces a systematic gain; the metrics are saturated. The authors should evaluate on non-saturated subsets such as the medium- and high-volatility regimes from Table S5, or should temper the claim about the agent's contribution to the observed performance.","section":"Section 6.3, Table 2"}],"minor_comments":[{"comment":"The final sentence of Section 5.6 is incomplete and ungrammatical; it should be rewritten to state the implementation settings and threshold-selection details that are deferred to the Supplementary Material.","section":"Section 5.6"},{"comment":"The main text lists Algorithm 1 as the ORCA algorithm but then refers to 'Algorithm 2' in the same section; the numbering between the main text and the Supplementary Material should be unified.","section":"Section 4.1 / Algorithm numbering"},{"comment":"Appendix C describes the controller as using 'instantaneous signal entropy' and interpreting K=5 as 'high stationarity,' which conflicts with Algorithm 2's variance/autocorrelation rule; these descriptions should be aligned.","section":"Supplementary Appendix C"},{"comment":"The MIMIC-IV reference contains the garbled phrase 'lessons et al. Andreo'; the citation should be corrected.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core idea is within scope and potentially useful, but the current manuscript cannot be accepted because the main claims and the supplementary results directly contradict each other. The revision should start from the raw per-regime numbers and rebuild the narrative from them; reporting the threshold values and a sensitivity analysis is essential for reproducibility. Appendix C introduces an LLM-based orchestration layer that is neither evaluated nor aligned with the core method; it should be removed or reconciled with the rest of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know ORCA before reading: the core idea is worth taking seriously. A rule-based controller that picks temporal context length from variance and autocorrelation, letting a MinGRU+ConSmax detector avoid manual k-tuning, is simple, cheap, and interpretable. The paper also does a clean ablation — random, volatility-only, and correlation-only agents all fail to produce systematic gains — which is more than most papers bother with. The distribution of selected k across normal versus anomalous windows (Table 3) is a nice structural sanity check. I would not dismiss the mechanism out of hand.\n\nWhat is not okay is the empirical reporting. Table 1 reports MIMIC-IV AUROC 0.9977 for adaptive k. Table S5 stratifies the same test set by volatility: medium 0.4448, high 0.3611. The caption says 'absolute performance is lower' due to weak labels, but that does not reconcile a 0.9977 global number with 0.36–0.44 in two of three regimes. Unless the test set is almost entirely low-volatility (not reported), the overall AUROC cannot be 0.998. This is not a footnote; it is the backbone of the 'no performance collapse' claim. There is a second, smaller contradiction: Table S3 shows all WBAN methods at exactly 0.9991 in medium volatility, while Figure S1's narrative says fixed-k 'exhibits a noticeable drop' in that regime. The two tables cannot both be right.\n\nTwo more soft spots, in proportion. First, the thresholds tau_vol and tau_corr are never reported, and the appendix only says the policy 'approximates' an oracle minimizing expected loss. Without those thresholds, the adaptive mechanism is underspecified and the claimed equivalence to fixed best-k cannot be checked. Second, Supplementary Section C introduces a Gemini-2.5-Flash LLM orchestrator that fuses ORCA and TFT outputs for clinical reasoning. That is a completely different architecture from the statistical-probe controller in the main text, it is not evaluated in any results table, and it reads like a second paper stitched in.\n\nIf this crossed my desk, I would send it to reviewers rather than desk-reject — the idea deserves scrutiny. But the action for the authors should be major revision: reconcile Table 1 with Table S5, release the thresholds, and either remove or properly evaluate the LLM section. Without that, I would not rely on any of the quantitative conclusions. It is a useful case study in why supplementary tables must agree with the main text, but not something I would cite or bring to reading group.","headline":"A genuinely useful lightweight adaptive-receptive-field idea is undermined by an unreconciled internal contradiction between Table 1 and Table S5.","tokens_in":14750,"tokens_out":3314,"would_cite":false,"duration_ms":32933,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adaptive temporal context selection matches best hand-tuned anomaly detectors.","keywords":["temporal receptive field","agentic control","anomaly detection","wireless body area networks","physiological time series","inductive bias adaptation","constant-normalization attention"],"falsifier":"Run ORCA, with its controller thresholds fixed as in the paper, on a new public physiological dataset that contains clearly labeled short transients and slow drifts. The central claim fails if the adaptive policy's AUROC or AUPRC falls below the best fixed $k$ on either regime, or if the optimal threshold values change between subjects by more than a small amount.","tokens_in":13735,"feed_emoji":"🩺","tokens_out":11891,"duration_ms":116188,"temperature":0.7,"pith_summary":"This paper tries to establish that the temporal receptive field—the amount of history an anomaly detector considers—should be chosen per input window rather than fixed in advance. It proposes ORCA, a lightweight detector whose supervisory controller reads cheap statistics such as average variance and autocorrelation and picks one of three context lengths at inference time. Across a custom wearable WBAN dataset and the MIMIC-IV clinical benchmark, the adaptive choice matches or slightly beats the best fixed context on AUROC and AUPRC while removing the need for dataset-specific tuning of the temporal horizon. The paper reads this as evidence that state-dependent inductive-bias control is a practical design principle for nonstationary physiological time series under edge constraints.","feed_headline":"Choosing the window on the fly matches tuned anomaly detectors","feed_subtitle":"On wearable and ICU data, ORCA's adaptive context keeps AUROC above 0.99 without horizon tuning.","key_machinery":"The load-bearing object is the agentic controller defined in Supplementary Equation S5. It maps a window to a discrete choice of temporal receptive field $k \\in \\{1,3,5\\}$ using two averaged statistics: mean per-channel variance and mean per-channel sample autocorrelation. High variance selects $k=1$, high autocorrelation selects $k=5$, and the default is $k=3$. That choice sets the support of the relative-position bias matrix $B^{(k)}$ added to ConSmax attention scores, so the detector's effective memory is truncated to the selected neighborhood; MinGRU supplies gated hidden states without recurrent weight matrices in the candidate. No trainable parameters are added by the controller, which is what makes the adaptation agentic in the paper's usage: a policy outside the gradient loop governs a structural inductive bias.","core_discovery":"ORCA claims that the need to pre-tune a temporal horizon can be eliminated by letting a deterministic controller change the attention span online. Concretely, the controller computes the average signal variance and the average autocorrelation over the current window and sets the receptive field to $k=1$ under high volatility, $k=5$ under high autocorrelation, and $k=3$ otherwise; this $k$ truncates the relative-position bias in a constant-normalization attention head built on a minimal gated recurrent backbone. On ORCA's custom WBAN dataset the adaptive policy reaches AUROC 0.9993 (fixed $k=3$: 0.9994) and on MIMIC-IV 0.9977 (fixed $k=3$: 0.9974). The paper also reports that during anomalous events the controller shifts from short contexts to $k=5$ in 73% of windows, which it presents as evidence that the adaptation tracks physiological regime rather than acting randomly.","pith_inferences":["Inference: the threshold values $\\tau_{\\mathrm{vol}}$ and $\\tau_{\\mathrm{corr}}$ are never reported, so a reimplementation cannot reproduce the controller from the text alone; this omission leaves open whether the constants are dataset-specific.","Inference: since the WBAN metrics are saturated above 0.999, the strongest case for adaptive control in this paper is removing tuning effort and stabilizing latency, not raising the accuracy ceiling; a harder benchmark with subtler contextual anomalies would be needed to expose detection gains.","Inference: the controller's preference for $k=5$ during anomalous windows could be turned into an unsupervised anomaly-confidence score, giving a low-cost way to rank or localize events without labels.","Inference: the same variance-and-autocorrelation probes could be attached to forecasting or classification heads, since the only requirement is a mechanism that accepts a variable-length context."],"forward_implications":["Anomaly detectors on edge devices can drop the dataset-specific search over temporal horizons and instead adjust their memory online at no extra parameter cost.","The controller's behavior is interpretable: it lengthens the context as an anomalous window approaches and shortens it afterward, so the selected $k$ can act as an event-aligned diagnostic signal.","In medium-volatility regimes where no single fixed context is clearly right, adaptive selection preserves AUROC while fixed-$k$ baselines degrade.","Because the adaptation lives in a truncated relative-position bias, the same controller idea transfers to any attention-based sequence model that accepts a variable-length neighborhood."],"supporting_citations":[{"why":"Defines the full-attention transformer whose quadratic cost ORCA is designed to avoid and serves as the architectural baseline.","marker":"Vaswani et al. (2017)"},{"why":"Supplies the relative position encoding that ORCA truncates with the selected receptive field.","marker":"Shaw et al. (2018)"},{"why":"Provides ConSmax, the constant-normalization attention operator that replaces softmax in ORCA.","marker":"Liu et al. (2024)"},{"why":"Motivates the minimal recurrent backbone by questioning whether full self-attention is necessary.","marker":"Feng et al. (2024)"},{"why":"Defines the Temporal Fusion Transformer used as the decoupled post-hoc interpretability and forecasting module.","marker":"Lim et al. (2021)"},{"why":"Supplies the MIMIC-IV clinical benchmark used for out-of-distribution evaluation.","marker":"Johnson et al. (2023; 2024)"}],"fun_headline_variants":["Adaptive window matches tuned anomaly detectors","ORCA adapts its context, no tuning needed","No horizon tuning: adaptive window matches baselines","Anomaly detection that selects its own context window"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one fixed pair of decision thresholds for variance and autocorrelation correctly separates signal regimes across every sensor channel and patient; the paper never reports the threshold values or shows that they transfer.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive window matches tuned anomaly detectors","ORCA adapts its context, no tuning needed","No horizon tuning: adaptive window matches baselines","Anomaly detection that selects its own context window"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001234,"raw_usage":{"total_tokens":5089,"prompt_tokens":984,"completion_tokens":4105,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":4046}},"tokens_in":600,"tokens_out":4105,"duration_ms":33873,"temperature":1.0,"reasoning_tokens":4046,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:20:45.401596+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ORCA, with its controller thresholds fixed as in the paper, on a new public physiological dataset that contains clearly labeled short transients and slow drifts. The central claim fails if the adaptive policy's AUROC or AUPRC falls below the best fixed $k$ on either regime, or if the optimal threshold values change between subjects by more than a small amount.","supporting_citations":[{"cited_title":"Were rnns all we needed? arXiv preprint, 2024","cited_arxiv_id":null,"evidence_quote":"Motivates the minimal recurrent backbone by questioning whether full self-attention is necessary."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MIMIC-IV clinical benchmark used for out-of-distribution evaluation."}],"review_version":1}