{"id":"453e59de-0d67-4b6e-8fee-4bb74056e79f","arxiv_id":"2508.06704","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"CISO is a deep learning model that conditions species distribution predictions on incomplete observations of other species, improving performance across plants, birds, and butterflies.","lead":"A new machine learning model, CISO, predicts where species live by combining environmental data with whatever partial observations of other species are available, even when that information is patchy. It reports improved spatial predictions for plants, birds, and butterflies, which could sharpen conservation planning and reveal hidden species interactions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified; central claim unverified because the full text is unreadable and the abstract lacks experimental details.","rationale":"The reader's verdict of UNVERDICTED is appropriate: the supplied full text is unreadable, so the central empirical claim cannot be assessed. The reader's identified weakest assumption was about the practical availability of test-time species observations for deployment; that is a real limitation but it is not the most load-bearing issue for correctness. The more fundamental issue is that the evaluation protocol itself is invisible: no error bars, baselines, split construction, or subset-selection rules are reported in the abstract. Without those details, the claim of improvement on spatially separate test sets could conceal information leakage or baseline mismatch. I do not claim the paper is wrong; I claim it is unverifiable from the provided evidence. A concrete reproduction check with spatial blocking and repeated seeds would settle whether the reported gain is robust. Since no specific technical flaw was identified, the reader's UNVERDICTED verdict should stand unchanged. Credit is due for the stated intention to release datasets, code, and checkpoints; such reproducible artifacts would make the proposed concrete test feasible.","tokens_in":1529,"tokens_out":4043,"duration_ms":55308,"concrete_test":"Download the released code and checkpoints and reproduce the main comparison under a strict spatial-block split (block size comparable to the data grain, e.g., 1 km for SatButterfly) with at least 5 independent seeds. Compute mean and standard deviation of the evaluation metric (AUC or BCE) for CISO, the abiotic-only baseline, and the best alternative method. Accept the central claim provisionally only if CISO's improvement over the abiotic-only baseline exceeds 2 standard deviations and the ordering across seeds is consistent. Also inspect the data-loading code to confirm that conditioning-species observations at test locations are not used for conditioning-subset selection, hyperparameter tuning, or early stopping.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that conditioning on partial species observations improves SDM accuracy on spatially separate test sets—cannot be checked from the supplied material: the full text is a corrupted encoding, so only the abstract is available. The abstract reports no baselines, hyperparameters, data-split construction, conditioning-subset selection rule, error bars, or evaluation protocol. The load-bearing uncertainty is therefore empirical rather than logical: the reported improvement could stem from test-region information (conditioning-species observations at test locations acting as environmental proxies), from an underpowered or mismatched baseline, or from subtle data leakage in the spatial split. This is not a demonstrated flaw, but it is the point on which the entire paper rests, and the available evidence does not expose it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CISO, a deep learning method for species distribution modeling (SDM) that conditions predictions on environmental variables plus a flexible and incomplete set of presence/absence observations of other species. The abstract claims that including this partial biotic information improves predictive performance on spatially separate test sets, that conditioning on a subset of species within the same dataset outperforms alternative methods, and that combining observations from multiple datasets is beneficial. The method is demonstrated on three datasets: sPlotOpen (plants), SatBird (birds), and a new SatButterfly (butterflies) dataset. Only the abstract and title are legible in the supplied manuscript; the full text is a corrupted character encoding, so no methodological or experimental details are accessible.","tokens_in":1613,"tokens_out":2082,"duration_ms":26600,"significance":"If the claims are verified, CISO would address a recognized gap in SDMs by incorporating incomplete biotic interaction data in a flexible deep learning framework, potentially improving predictive accuracy for conservation applications. The cross-dataset combination claim is also novel. However, the significance hinges entirely on empirical validation that is currently unreviewable because the main text is unreadable. The core contribution is a new architecture and training/evaluation protocol whose details and evidence are absent from this submission.","major_comments":[{"comment":"The supplied full text is a corrupted encoding; only the title and abstract are readable. No methods, model architecture, training details, experimental setup, baseline descriptions, results tables, or figures are available. This is a load-bearing deficiency: none of the abstract's central claims can be verified without the complete manuscript. The authors must provide a readable, complete version before any technical assessment can proceed.","section":"Full Text (all sections)"},{"comment":"The central claim—including partial biotic information improves predictive performance on spatially separate test sets—is stated without any supporting details: baselines are not named, the evaluation protocol is unspecified, no error bars or statistical significance are reported, and the construction of the 'subset of species' used for conditioning is not described. The claim that 'CISO outperforms alternative methods' is therefore unsupported even at the level of the abstract.","section":"Abstract"},{"comment":"A potential correctness risk is that conditioning on species observations at test locations may indirectly leak information about the target species or serve as environmental proxies. The abstract does not clarify whether the conditioning observations are available at the same test sites, how they are selected, or whether the spatially separate test set still allows such observations. A precise statement of the prediction-time assumptions is essential to assess whether the reported improvements are due to genuine biotic signal or to test-region information.","section":"Abstract / Method (unavailable)"}],"minor_comments":[{"comment":"The abstract states 'CISO enables predictions to be conditioned on a flexible number of species observations alongside environmental variables' but does not specify how the number and identity of conditioning species are chosen at test time. Clarifying this would help readers understand the practical scenario.","section":"Abstract"},{"comment":"The data and code availability statement appears in the corrupted text only as placeholder URLs. Please ensure the final manuscript contains working anonymous links.","section":"Full Text (data availability)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript text is almost entirely corrupted; it appears to be an encoding artifact rather than a deliberate omission. I recommend asking the authors to resubmit a readable PDF before sending it for full technical review. The abstract alone is insufficient to judge soundness; however, the proposed idea is plausible and potentially publishable if the experiments are well designed and the results are accurately reported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The idea behind CISO is genuinely useful: existing joint SDMs force you to have the same set of species everywhere, and they treat interactions as symmetric pairwise effects. CISO instead conditions on whatever subset of species observations you actually have at a site, alongside environmental predictors. That matches messy real survey data, and the new SatButterfly dataset is a concrete contribution beyond the method. The claim that combining multiple datasets helps is plausible.\n\nThe soft spot is the empirical core. From the abstract we learn that CISO improves on spatially separate test sets and beats alternatives, but we see no baselines, no error bars, no details on how the conditioning species were chosen, and no protocol for the spatial split. The supplied full text is corrupted, so I couldn't verify any of this. The stress-test worry is real: if at test time you feed in presence/absence of other species at that same location, those observations can act as proxies for environmental conditions the model didn't see, which would inflate apparent gains without reflecting true biotic interactions. That doesn't make the method flawed, but it means the abstract's claim is underdetermined.\n\nThe other thing to flag is the practical premise: to use CISO you need observations of at least some species at the locations where you want predictions. That's fine for a survey-based workflow, but it limits the 'distribution modeling' use case where you might want to predict into unobserved areas. The abstract doesn't discuss how the model behaves when no species observations are available.\n\nNone of this is fatal. The paper is a serious attempt at a real problem, and the authors are transparent about code and data availability. So I'd send it to review, but the referee should ask to nail down the evaluation: baselines, error bars, sensitivity to the conditioning set, and a check of whether the gains survive when conditioning species are held out at test time. For the reading group: maybe, once we can read the actual full text. I wouldn't cite it yet, but I'd keep an eye on it.","headline":"CISO is a plausible and useful method for incorporating incomplete biotic observations into SDMs, but the empirical claims are unverifiable from what we can read.","tokens_in":2154,"tokens_out":2487,"would_cite":false,"duration_ms":28762,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Conditioning a species distribution model on partial observations of other species improves its predictions, as shown across plant, bird, and butterfly datasets.","keywords":["species distribution modeling","biotic interactions","deep learning","incomplete observations","presence-absence data","conservation planning","cross-taxa prediction","spatial ecology"],"falsifier":"A direct test is to train CISO on a region and then evaluate on a spatially separated test set where the conditioning species are entirely removed, leaving only environmental variables; if the improvement over an environment-only baseline disappears when no other-species observations are supplied, the claimed benefit of partial biotic information is not established.","tokens_in":1382,"feed_emoji":"🦋","tokens_out":4825,"duration_ms":46487,"temperature":0.7,"pith_summary":"The paper introduces CISO, a deep-learning species distribution model that, in addition to environmental variables, takes as input whatever presence/absence observations are available for a subset of other species at the same locations. The central claim is that this incomplete biotic information improves predictions of where the remaining species occur, compared with models that rely only on abiotic conditions. The authors test this across plant, bird, and butterfly datasets and report consistently better performance on spatially separated test sets, as well as further gains when observations from multiple datasets are combined. If the claim holds, SDMs can make productive use of the patchy, opportunistic species observations that field surveys and citizen science typically produce.","feed_headline":"Patchy species data improves distribution forecasts","feed_subtitle":"A deep-learning model uses a few observed species to predict the rest, and works across plants, birds, and butterflies.","key_machinery":"The key machinery is the conditioning mechanism: a deep-learning model that accepts an arbitrary number and combination of species observations, alongside environmental covariates, and produces a prediction for the target species. The flexibility to handle incomplete sets is what distinguishes CISO from pairwise co-occurrence methods, which require aligned species-by-site matrices. The mechanism allows the model to encode whatever partial observations exist at a location, including observations from different taxa, and use them to modulate the environmental signal.","core_discovery":"CISO's core discovery is that a model conditioned on a flexible, incomplete set of species observations can outperform both environment-only SDMs and methods that require complete pairwise co-occurrence data. The paper demonstrates this by conditioning predictions for a target species on observations of a subset of other species, showing that the remaining species are predicted more accurately. It also shows that observations from different taxonomic datasets can be combined to improve performance, indicating that the model captures cross-taxon associations rather than just within-dataset correlations. The authors position CISO as a practical tool that relaxes the symmetry and completeness a","pith_inferences":["A direct extension not tested in the paper: when predicting in a new region or future climate, the conditioning species may be absent or compositionally different; the reported gains could degrade unless the model is trained to handle shifted species sets. This could be tested by training on one region and conditioning on observations from a novel region.","The implicit claim that partial observations carry information about the target species suggests a sampling-design implication: the marginal value of an additional species observation could be quantified, allowing survey effort to be allocated to the species that most reduce prediction uncertainty.","If the model truly captures cross-taxa interactions, it might be used as a data-driven screen for potential biotic interactions, generating candidate species pairs for field or experimental validation, though the paper does not directly validate the ecological causality of these associations."],"forward_implications":["SDMs can be applied to data where only a few species are recorded (e.g., citizen-science checklists) without discarding incomplete records, widening the usable data pool for ecological modeling.","Conservation planners can use observations of common or easily detected species to improve predictions for rare or elusive species at the same sites, which is directly relevant to monitoring programs.","The method provides a way to combine monitoring data from different taxonomic groups, potentially improving predictions across ecosystem boundaries where single-taxon data are sparse.","If the model's learned associations are interpretable, it can generate hypotheses about which species interactions matter most for distribution, guiding targeted ecological field studies."],"supporting_citations":[],"fun_headline_variants":["Patchy sightings sharpen species range predictions","Flexible species input improves distribution forecasts","CISO uses sparse species data to predict ranges","Incomplete observations, better SDM forecasts","Cross-taxon data boosts species distribution accuracy"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The method's practical advantage depends on having, at the time of prediction, presence/absence records for at least some other species at the same locations where one wants to predict; without such observations the model has no biotic input and reduces to an environment-only SDM.","fun_headline_variants_meta":{"raw":{"variants":["Patchy sightings sharpen species range predictions","Flexible species input improves distribution forecasts","CISO uses sparse species data to predict ranges","Incomplete observations, better SDM forecasts","Cross-taxon data boosts species distribution accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000167,"raw_usage":{"total_tokens":1107,"prompt_tokens":773,"completion_tokens":334,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":269}},"tokens_in":517,"tokens_out":334,"duration_ms":4621,"temperature":1.0,"reasoning_tokens":269,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:34:39.325079+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to train CISO on a region and then evaluate on a spatially separated test set where the conditioning species are entirely removed, leaving only environmental variables; if the improvement over an environment-only baseline disappears when no other-species observations are supplied, the claimed benefit of partial biotic information is not established.","supporting_citations":[],"review_version":1}