{"id":"25b4435a-650e-4858-8a22-d0a73f363a0a","arxiv_id":"2607.26854","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"A node-invariant gated residual adapter, trained on Colombia and Nicaragua ETSIR-plus-backbone forecasts, improves 8-week dengue MAE in Mexico and Malaysia in 9 of 10 backbone settings using only two target scalars.","lead":"TREA-Net transfers dengue forecasting skill from data-rich countries to data-scarce ones by correcting any neural backbone with a small, shared residual module guided by an environmental SIR model. Health agencies with only 1.5–2 years of local data could get better 8-week case forecasts without retraining large models.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-scoped climatology/generalization caveat.","rationale":"The strongest claim is a multi-backbone, multi-transfer empirical result under explicitly limited target history. Table 1, the 9/10 Wilcoxon result, TiRex best-in-every-setting statement, and ablations (w/o ETSIR, in-domain adapter, per-node 2N) jointly support it within the four-country retrospective design. The load-bearing external condition is indeed that the frozen residual plus climatology prior remains helpful when dynamics or weather differ; the paper already flags climatology and limited prospective scope in Discussion/Limitations. That does not overturn the reported numbers, so the CONDITIONAL verdict (pending code/data and clearer ETSIR citation) needs no further downgrade or upgrade from this pass. Agreement with the reader is full on the weakest assumption and on keeping correctness risk low for the scoped claim.","tokens_in":15632,"tokens_out":461,"duration_ms":9041,"concrete_test":"Re-run the K=104 Mexico TREA-Net–TiRex evaluation replacing training-period weekly climatology with held-out actual temperature/precipitation over the 8-week horizons (or a high-quality weather forecast product); if MAE rises above the zero-shot TiRex baseline or the 29.6% interval-width gain disappears, the operational claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that a single N-invariant gated residual learned from Colombia/Nicaragua backbone–ETSIR triples, plus two global target scalars, improves 9/10 backbone–target settings and yields best MAE with TiRex—is internally supported by Table 1, Wilcoxon tests, and ablations (Table 2). The reader's weakest assumption (climatology weather + frozen correction under source–target mismatch) is real but already correctly scoped as a deployment/generalization limit rather than a failure of the reported tables. No hidden inconsistency in the N-invariant construction, two-scalar adapter, or multi-source reweighting undermines the scoped claim; effect sizes are uneven and one setting slightly degrades, which the paper reports.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes TREA-Net, a backbone-agnostic transfer framework for 8-week-ahead subnational dengue forecasting when the target surveillance system has only K∈{78,104} weeks of history. It feeds normalized backbone forecasts, ETSIR mechanistic projections, and their absolute discrepancy into a shared pointwise gated residual module trained on reweighted Colombia and Nicaragua windows, then adapts to each target with two global scalars (γ, β) while freezing the backbone and correction weights. The N-invariant design is intended to transfer across unequal numbers of administrative units. Empirically, across five backbones and ten backbone–target settings, TREA-Net improves MAE in 9 of 10 comparisons (Wilcoxon), TREA-Net–TiRex attains the lowest MAE on all target datasets (≈7.1% average reduction vs zero-shot TiRex), ablations support the ETSIR feature and multi-source training, and EnbPI intervals maintain empirical coverage while narrowing 8-week width by 29.6% in Mexico.","tokens_in":15885,"tokens_out":1569,"duration_ms":46728,"significance":"If the reported transfer results hold under broader deployment conditions, TREA-Net is a practically useful contribution: a lightweight (~5K-parameter), N-invariant residual adapter that augments both target-trained sequence models and a frozen foundation model (TiRex) without backbone fine-tuning or fixed spatial topology. The multi-country low-data protocol, multi-backbone evaluation, ablations (no ETSIR, in-domain adapter, per-node 2N adapter), Wilcoxon testing, gate-skill analysis, and conformal efficiency results are concrete strengths relative to many hybrid epi-ML papers that remain single-system and end-to-end. The two-scalar target adapter and open evaluation framing are well matched to health agencies with short surveillance histories. Significance is applied rather than foundational: gains are modest and geography-dependent, but the design is portable and the empirical package is stronger than typical residual-hybrid forecasting notes.","major_comments":[{"comment":"Table 1 and the abstract claim improvement in 9 of 10 settings, but effect sizes are highly heterogeneous: Mexico gains are often material (e.g., PatchTST K=78: 89.5→74.4; TiRex K=104: 60.1→51.2), whereas several Malaysia deltas are near noise (e.g., N-HiTS K=104: 29.0→29.1 degradation; TCN K=104: 40.0→39.9). Wilcoxon significance on paired seeds does not establish operational value. The Results/Discussion should foreground effect-size heterogeneity (not only win count), report paired percentage reductions with uncertainty, and qualify the portable early-warning claim by when the frozen correction helps versus leaves the backbone essentially unchanged.","section":"Results and Comparisons; Table 1"},{"comment":"Methodology (ETSIR test-time covariates) replaces future temperature/precipitation over the H=8 horizon by training-period week-specific climatology. This is load-bearing for both the ETSIR feature and the learned gate/Δ mapping, yet there is no sensitivity experiment (oracle future weather vs climatology vs persistence weather; or years with anomalous climate). Without that, it remains unclear how much of the transferred residual is robust epidemiological structure versus source-trained compensation for a systematically biased climate prior. A compact ablation or error breakdown on high-anomaly windows would substantially strengthen the central transfer claim.","section":"Methodology: ETSIR Mechanistic Prior; Discussion: Limitations"},{"comment":"Related Work cites transfer/hybrid methods (CALI-Net, EINNs, EGDL, EARTH, PETSA-style adapters) but Experiments compare mainly backbone-only, ARIMA, ETSIR alone, and ETSIR-PINN. For the claim of a superior transferable residual adapter under unequal N and short K, at least one competitive transfer or parameter-efficient adaptation baseline (e.g., target-only residual on (yB,yE), FiLM-style adapter, or simple source-trained residual without multi-source reweighting beyond the existing ablation) should be reported under the same K protocol. Table 2’s in-domain adapter helps but does not fully close this gap.","section":"Related Work; Experiments; Table 2"}],"minor_comments":[{"comment":"Figure 3 is informative, but aggregate metrics (MAE by horizon, coverage by horizon, and interval width already partly in Figure 6) should be cross-referenced in the main caption so readers do not over-generalize from four selected units.","section":"Figure 3; Uncertainty Quantification"},{"comment":"Notation: byE / ŷE and MinMax tildes are clear, but the manuscript sometimes switches between incidence I_t and log(1+I_t) scales when discussing MAE on raw counts; state explicitly at each table that metrics are always on inverse-transformed case counts.","section":"Methodology; Table 1"},{"comment":"Wang et al. 2025 ETSIR citation is incomplete (“as cited in the main paper” / “et al.”). Provide a full bibliographic entry or preprint identifier for reproducibility of the mechanistic prior.","section":"References"},{"comment":"Multi-source reweighting (w_g = D_min/|D_g|) is sensible; briefly state whether windows are sampled with replacement or loss-reweighted in the implementation, and whether backbone-specific source models share identical hyperparameters across Colombia and Nicaragua.","section":"Multi-Source Training; Supplementary settings"},{"comment":"Abstract states conformal prediction “maintains empirical coverage” while the appendix reports ~87% (Mexico) and ~81% (Malaysia) vs 90% nominal under distribution shift. Align abstract wording with the undercoverage already documented.","section":"Abstract; Empirical Coverage under Distribution Shift"},{"comment":"Typos/style: “Rodr íguez” spacing artifacts; “Aedes aegypti” should be consistently italicized; arXiv footer date “29 Jul 2026” looks like a metadata glitch and should be corrected if unintentional.","section":"Throughout"}],"recommendation":"minor_revision","confidential_remarks":"Solid applied ML-for-epi paper with a clean transfer protocol and honest ablations; not a conceptual breakthrough, but above the usual hybrid-forecasting bar. I would not reject for lack of theory. Main risk is over-reading the 9/10 headline given small Malaysia deltas and climatology dependence—minor revision should force clearer effect-size reporting and one weather-sensitivity check. Scope fits cs.LG / applied forecasting venues; if the journal expects strong baselines against prior epi transfer methods, that is the main editorial pressure point."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing worth knowing is that this is a practical transfer recipe that mostly works as advertised. They train one shared, node-count-invariant gated residual on backbone forecast, ETSIR projection, and their discrepancy from Colombia and Nicaragua, freeze everything, then fit two global scalars on 78 or 104 weeks of Mexico/Malaysia data. Across five backbones and ten settings it beats the paired backbone in 9/10, with Wilcoxon support, and TREA-Net–TiRex is best everywhere (~7% average MAE cut vs zero-shot TiRex).\n\nWhat is actually new is the packaging, not the parts. ETSIR priors, residual adapters, FiLM-style scale/shift, channel-independent weights, and foundation-model zero-shot are all known. The useful move is making the correction backbone-agnostic and N-invariant so you can move across different numbers of admin units without graph surgery or backbone fine-tuning, plus multi-source reweighting so Colombia does not dominate. Ablations are honest: drop ETSIR and error rises; in-domain-only correction is weaker; 2N node-specific adapters do not help. They also ship EnbPI intervals and show a real 29.6% width cut at h=8 in Mexico at similar coverage. Protocol hygiene is good—held-out after week K, training-only ETSIR/MinMax/climatology, no test leakage.\n\nSoft spots are real but scoped. Gains are uneven (stronger in Mexico, small in Malaysia; one tiny degradation). Future weather is training-period weekly climatology, so anomalous climate can mislead the prior. Evaluation is retrospective on two targets and one disease; the ETSIR citation is thin (\"Wang et al. 2025\" as cited). Code is promised anonymous, not yet a public artifact in the text. None of that breaks the tables.\n\nThis is for people building portable epi early-warning under short surveillance histories, and for foundation-model users who want a cheap mechanistic residual without retraining. Math is elementary; data and design look solid for the claim as stated. I would send it to referees. Engage if you care about low-data health forecasting; skip if you only want new theory.","headline":"Solid applied transfer setup for low-data dengue forecasting: N-invariant residual on backbone+ETSIR, two-scalar target adapt, real multi-backbone gains—not a new theory, but clean and usable.","tokens_in":16511,"tokens_out":549,"would_cite":true,"duration_ms":11121,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A tiny transferable correction, guided by a climate-aware epidemic model, improves 8-week dengue forecasts when a new country has only 1.5–2 years of data.","keywords":["dengue forecasting","transfer learning","ETSIR","residual adaptation","time-series foundation models","conformal prediction","limited surveillance data","N-invariant models"],"falsifier":"On a held-out target country with short history, check whether TREA-Net raises mean absolute error versus its frozen backbone (and loses Wilcoxon significance) when source–target transmission dynamics or anomalous climate diverge, especially if real weather forecasts replace training-period climatology and still fail to help.","tokens_in":16477,"feed_emoji":"🦟","tokens_out":968,"duration_ms":21165,"temperature":0.7,"pith_summary":"Health agencies need multi-week dengue forecasts to plan mosquito control and hospital capacity, but new surveillance systems often have too little history to train reliable neural models. TREA-Net keeps any pretrained forecasting backbone frozen and adds a lightweight residual correction that blends the backbone’s forecast with an Environmental Time-Series SIR (ETSIR) projection of climate-sensitive transmission. The correction is trained on long dengue records from Colombia and Nicaragua, is invariant to the number of regions, and is adapted to a new country by fitting only two global scalars on 78 or 104 weeks of local data. Across five backbones and ten transfer settings into Mexico and Malaysia, this improves the backbone in nine settings with statistically significant gains; with the TiRex foundation model it posts the lowest mean absolute error on every target set, and conformal intervals stay calibrated while narrowing at longer leads in Mexico. The practical claim is that mechanistic guidance can be ported as a small residual adapter rather than rebuilt inside each new surveillance system.","feed_headline":"Tiny residual fix lifts dengue forecasts with 1.5 years of data","feed_subtitle":"A climate-SIR prior plus two local scalars improves frozen neural and foundation backbones across countries.","key_machinery":"N-invariant gated residual correction: for each horizon–region pair it takes the normalized backbone forecast, ETSIR forecast, and their absolute gap, then outputs a gate and residual so the corrected forecast is backbone plus gate times residual; only two global target scalars (scale and shift) are fit at deployment.","core_discovery":"Epidemiological knowledge learned from data-rich dengue surveillance can be transferred to structurally different, data-scarce systems by freezing a neural forecasting backbone and applying an N-invariant gated residual correction driven by ETSIR projections, with only two target-specific scalars for local adaptation. In transfers from Colombia and Nicaragua to Mexico and Malaysia under 78 or 104 weeks of target history, this improves the paired backbone in 9 of 10 settings and yields the best mean absolute errors when paired with zero-shot TiRex.","pith_inferences":["If the two-scalar adapter is enough here, other public-health transfer settings may benefit more from freezing large models and learning tiny residual bridges than from full fine-tuning under short histories.","Replacing climatology with operational weather forecasts is a direct next stress test: gains should grow when anomalies drive outbreaks and shrink if the residual overfits source climate regimes.","The gate’s spatial link to ETSIR skill in Malaysia hints that monitoring gate activation could flag regions where the mechanistic prior should be down-weighted in operations."],"forward_implications":["Agencies with ~1.5–2 years of weekly dengue data can improve multi-week forecasts without retraining large backbones or matching the source country’s number of regions.","Zero-shot foundation-model forecasts can be sharpened with a ~5K-parameter residual module and two local scalars estimated in under a minute on CPU.","Calibrated conformal intervals can stay near baseline coverage while becoming substantially narrower at 8-week lead when the point forecast is stabilized (as reported for Mexico).","The same backbone-agnostic adapter can be re-evaluated as an open early-warning add-on for other climate-sensitive pathogens once local ETSIR and short target histories exist."],"fun_headline_variants":["Two scalars adapt frozen backbones for dengue forecasts in new countries","ETSIR residual transfer beats backbones in 9 of 10 scarce-data settings","Gated ETSIR correction lifts TiRex to lowest MAE on Mexico and Malaysia","N-invariant residual adapts Colombia knowledge to 78-week target surveillance","Climate-SIR prior plus two globals cuts 8-week dengue error across borders"],"cache_read_input_tokens":128,"weakest_assumption_plain":"A single residual map learned from Colombia and Nicaragua, plus two country-wide scalars and weather replaced by historical weekly averages over the forecast horizon, is assumed to capture enough local epidemiology to help rather than hurt in a new place.","fun_headline_variants_meta":{"raw":{"variants":["Two scalars adapt frozen backbones for dengue forecasts in new countries","ETSIR residual transfer beats backbones in 9 of 10 scarce-data settings","Gated ETSIR correction lifts TiRex to lowest MAE on Mexico and Malaysia","N-invariant residual adapts Colombia knowledge to 78-week target surveillance","Climate-SIR prior plus two globals cuts 8-week dengue error across borders"]},"model":"grok-4.5","effort":"low","cost_usd":0.004086,"raw_usage":{"total_tokens":1290,"prompt_tokens":859,"num_sources_used":0,"completion_tokens":84,"cost_in_usd_ticks":40864000,"prompt_tokens_details":{"text_tokens":859,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":347,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":859,"tokens_out":84,"duration_ms":6974,"temperature":1.0,"reasoning_tokens":347,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T19:08:09.359403+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a held-out target country with short history, check whether TREA-Net raises mean absolute error versus its frozen backbone (and loses Wilcoxon significance) when source–target transmission dynamics or anomalous climate diverge, especially if real weather forecasts replace training-period climatology and still fail to help.","supporting_citations":[],"review_version":1}