{"id":"87f2bfe2-7ec7-4055-ac6c-dde7ce37a387","arxiv_id":"2607.28510","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Uncertainty-aware Probabilistic Random Forest on 11 photometric light-curve features classifies TDEs without host data more stably than XGBoost and recovers 14 new candidates from ZTF.","lead":"A host-agnostic Probabilistic Random Forest classifies tidal disruption events from ZTF light curves by treating photometric features as distributions rather than point estimates. This reduces overconfident errors on faint, noisy alerts expected from the Rubin Observatory and recovers new TDE candidates from archival data.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The comparative claim that PRF is essential for low-S/N Rubin data rests on a single-source synthetic degradation test whose noise model and feature pipeline may not generalize.","rationale":"The reader correctly flags aggressive uncertainty-aware SMOTE (4.3.1) as a weak link for generalization, and the paper itself calls linear feature-space interpolation “not ideal.” That concern is real but secondary: SMOTE affects both classifiers equally inside the LOOCV, so it does not uniquely underwrite the PRF-versus-XGBoost stability/FP-rejection contrast that is the paper’s strongest empirical result. The more load-bearing step for the headline claim (“uncertainty-aware classifiers are essential for the Rubin era”) is the single-source magnitude-degradation test in §6.2/Fig. 10. That test is carefully executed and directionally informative, yet it remains a single morphological template under ZTF noise statistics. Expanding it to multiple real TDEs and an LSST-like noise/cadence model is the concrete check that would either solidify or qualify the forward-looking claim. Until then the verdict stays CONDITIONAL, with the primary condition shifted from SMOTE realism alone to the robustness of the low-S/N degradation result. The ZTF-internal comparisons, candidate list, and public data release are solid and do not require a harsher verdict.","tokens_in":30412,"tokens_out":667,"duration_ms":14007,"concrete_test":"Repeat the §6.2 degradation protocol on at least the 10 post-cutoff TDEs of Table 4 (or a stratified subset of the 59 training TDEs), injecting both the original ZTF noise model and a thinned, LSST-like cadence+depth noise model; if the median PRF–XGB probability gap at the faint end shrinks below ~0.2 or loses monotonicity for a majority of sources, the Rubin-essential claim is not supported by the present evidence.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper’s central forward-looking claim (abstract + §6.2 + §7) is that uncertainty-aware PRF is essential for the faint, low-S/N Rubin regime because it degrades gracefully while XGBoost collapses. That claim is supported almost entirely by one controlled experiment (Fig. 10 / §6.2) that takes a single high-confidence TDE (TDE2021axu), scales its GP mean flux, and re-injects noise drawn from an empirical flux-error window of the same ZTF light curve. The resulting feature posteriors are then fed to both classifiers. Because the noise model, cadence, and baseline are still pure ZTF, and because only one morphological template is used, the experiment does not yet demonstrate that the same graceful-vs-abrupt behaviour holds for the broader TDE population, for Rubin-like sparse sampling, or for the aggressive SMOTE-trained decision boundaries that were learned on bright ZTF features. If the PRF advantage disappears under multi-source or cadence-mismatched degradation, the “essential for Rubin” conclusion weakens while the more modest ZTF comparative results remain intact.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper develops a host-agnostic photometric classifier for tidal disruption events (TDEs) in the ZTF nuclear-transient stream. Eleven light-curve features (rise/decay timescales, blackbody temperature evolution, pre-transient and plateau fractional rms, plateau strength/SNR, etc.) are extracted via a multi-band MCMC fit that extends the van Velzen et al. (2021) model with a late-time plateau. Classification is performed with a Probabilistic Random Forest (PRF) that treats features as distributions, benchmarked against deterministic XGBoost under a repeated stratified Leave-One-Out Cross-Validation (K=59 TDEs, 250 seeds) with uncertainty-aware SMOTE. The authors report complementary regimes (XGBoost higher recall at high precision; PRF fewer false positives at high recall and more stable probabilities for ambiguous sources), recover 10/12 post-2023 spectroscopic TDEs, and identify 11 new archival candidates plus 3 possible label contaminants. A single-source synthetic flux-degradation test is used to argue that PRF degrades more gracefully than XGBoost and is therefore essential for the low-S/N Rubin regime.","tokens_in":30697,"tokens_out":1850,"duration_ms":44495,"significance":"Host-agnostic, purely photometric TDE selection is a genuine and timely need for LSST/Rubin, where host WISE/AGN indicators and spectroscopic follow-up will be unavailable for most faint nuclear flares. The work’s concrete strengths include: (i) explicit propagation of feature posteriors into a PRF rather than point estimates; (ii) a carefully documented multi-class LOOCV with 250 repetitions, threshold scenarios, confusion matrices, and SHAP importances; (iii) an independent post-cutoff TDE test set; (iv) a controlled signal-degradation experiment; and (v) public release of forced-photometry light curves and per-source class probabilities on Zenodo. If the stability and false-positive advantages of uncertainty-aware classification hold under more realistic Rubin-like sampling, the methodological message is important for the broader transient-classification community. The archival candidate list is a useful byproduct even if purity remains unquantified on the unlabeled pool.","major_comments":[{"comment":"§6.2 and Fig. 10 are the primary quantitative support for the abstract’s and §7’s claim that uncertainty-aware PRF classifiers are “essential for the Rubin era.” The experiment uses a single high-confidence template (TDE2021axu), scales its GP mean flux, and re-injects noise drawn from that same ZTF light curve’s empirical flux–error relation. Cadence, baseline coverage, and morphological diversity therefore remain pure ZTF. A single-template, same-survey noise model cannot establish that the graceful-vs-abrupt behaviour generalises across the TDE population, under Rubin-like sparse sampling, or under the SMOTE-trained boundaries learned on bright ZTF features. Either expand the degradation test to a multi-source ensemble and/or cadence-mismatched noise, or substantially qualify the Rubin-forward claim so that it is presented as a motivated hypothesis rather than a demonstrated necessity","section":"§6.2, Fig. 10; Abstract; §7"},{"comment":"§4.3.1 applies aggressive uncertainty-aware SMOTE that oversamples the 59 real TDEs (and SNe) up to the AGN count inside every LOOCV training fold. The paper itself notes that linear interpolation in feature-and-uncertainty space “is not ideal” and defers physically motivated injections to future work. Because the reported decision boundaries, SHAP rankings, and threshold metrics are learned on this heavily synthetic minority population, the generalisation claim to real faint/low-S/N events rests on an untested assumption. At minimum, the manuscript should quantify sensitivity to the oversampling factor (e.g., metrics vs. milder SMOTE ratios or no SMOTE with class weights) and state clearly which performance numbers are most affected. Without that, the LOOCV metrics cannot be read as reliable estimates of behaviour on real Rubin-like light curves.","section":"§4.3.1"},{"comment":"The PRF–XGBoost comparison confounds two distinct design choices: (i) native uncertainty propagation vs. point estimates, and (ii) bagging vs. sequential gradient boosting. §4.4 and §6.1 attribute stability and false-positive rejection largely to uncertainty handling, yet the bagging architecture alone (independent trees, averaged predictions) already reduces variance under LOOCV perturbations. A standard (non-probabilistic) Random Forest baseline, or an ablation that feeds PRF median features without uncertainties, is needed to isolate how much of the reported advantage is truly due to treating features as distributions. Without that control, the central methodological claim—that uncertainty awareness is what makes the classifier robust—is only partially supported.","section":"§4.1–4.4; §6.1"},{"comment":"§5.1 explicitly states that reported probabilities are uncalibrated classifier outputs and “should not be interpreted as posterior probabilities,” yet the high-precision / balanced / high-recall scenarios (§4.3.3, Table 3, Fig. 4–5) and the archival search (§5.2) apply absolute probability thresholds (e.g. ≈0.85) as if they were meaningful on an absolute scale. Thresholds chosen on uncalibrated scores are fold- and architecture-dependent; the complementary-regime narrative and the candidate cuts therefore need either a calibration step (Platt/isotonic on held-out folds) or a clear statement that all thresholds are relative operating points only, with purity/completeness re-estimated under that caveat when the ensemble is applied to unlabeled data.","section":"§4.3.3; §5.1–5.2; Table 3"}],"minor_comments":[{"comment":"Table 3 reports asymmetric uncertainties with occasional identical upper/lower signs (e.g. PRF high-recall Recall 0.966+0.017+0.000). Clarify whether these are 16th/84th percentiles and correct any typographical sign errors.","section":"Table 3"},{"comment":"Figure 3 violin plots are dense; labeling only a subset of TDE IDs or providing an interactive/supplementary version would improve readability. Cross-reference to Table A1 is helpful but easy to miss.","section":"Fig. 3"},{"comment":"Eq. (5) normalises excess variance by F_peak rather than by mean host flux; a one-sentence justification relative to the more common f_var definition would help non-specialist readers.","section":"§3.3, Eq. (5)"},{"comment":"The plateau sigmoid timescale is fixed at τ_plat = 3.264 days (§3.1) to reach 99% of F_plat near t_peak+30 d. State whether results are sensitive to this choice.","section":"§3.1"},{"comment":"Several in-text citations and TNS report formatting are inconsistent (e.g. mixed “TNS Classification Report” styles). A pass for uniform reference formatting would help.","section":"References; Table 4–5 notes"},{"comment":"Data availability (§8) is commendable. Adding a short README pointer to which catalog columns map to the 11 features and the PRF/XGB probability columns would lower the barrier for reuse.","section":"§8"},{"comment":"In §2.2, the zero-point cut is written as |log10 ZP / med(ZP)| < 0.4; confirm the exact expression and units so the cut is reproducible.","section":"§2.2"}],"recommendation":"major_revision","confidential_remarks":"The core ZTF host-agnostic pipeline and the PRF-vs-XGBoost stability comparison are publishable after revision; the main risk is over-claiming readiness for Rubin on the basis of one single-source degradation test and aggressive SMOTE. If the authors tone the forward-looking language and add the requested ablations/controls, this is a solid MNRAS contribution. Scope fit is appropriate for MNRAS (transients / surveys / ML methods)."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a careful, usable comparison of uncertainty-aware PRF against XGBoost on a uniform 11-feature ZTF nuclear set, plus a concrete candidate list and public light curves. The Rubin-forward claim is the softest part, not the ZTF results.\n\nWhat is new is not host-agnostic TDE ML itself (Gomez, Stein, Bhardwaj, Lanza et al. already exist), but treating the extracted features as distributions via Reis et al. PRF, then measuring stability under repeated stratified LOOCV (250 seeds) against a strong deterministic baseline. They show complementary regimes cleanly: XGBoost higher recall at high precision; PRF rejects more FPs in high-recall and stays unimodal on ambiguous sources. The post-mid-2023 TDE holdout (10/12 above ~0.85), SHAP importances, and the ensemble search that yields 11 unknowns plus 3 plausible label flips are real deliverables. Data release on Zenodo is a plus.\n\nSoft spots, in proportion. Aggressive uncertainty-aware SMOTE up to the AGN count is the weakest training assumption; they admit linear feature-space interpolation is not ideal. Probabilities are uncalibrated. The new candidates lack spectra, so purity is inferred from FPR, not confirmed. The stress-test on Fig. 10 is fair: the graceful-vs-collapse story for faint Rubin data is almost entirely one high-confidence TDE (2021axu) with ZTF cadence and a ZTF-derived noise model. That supports a qualitative point about uncertainty propagation; it does not yet prove the architecture is essential under multi-template, sparse LSST sampling. The more modest ZTF comparative claim still stands without that leap.\n\nMath and pipeline look standard (van Velzen-style multi-band model + plateau, two-stage MCMC, emcee). Citations are appropriate; no circular label games. For people building photometric nuclear classifiers or planning LSST TDE filters, this is worth reading. I would send it to referees. Engage if you care about uncertainty-aware transient ML; treat the Rubin essentiality language as aspirational until multi-source degradation and better minority synthesis are shown.","headline":"Solid ZTF methods paper: PRF vs XGBoost on host-agnostic TDE features is carefully done; the “essential for Rubin” claim rests too heavily on one synthetic degradation test.","tokens_in":31363,"tokens_out":560,"would_cite":true,"duration_ms":17591,"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":"TDEs can be found from light curves alone if classifiers treat features as uncertain distributions rather than fixed numbers.","keywords":["tidal disruption events","photometric classification","probabilistic random forest","host-agnostic","measurement uncertainty","ZTF","Rubin Observatory","nuclear transients"],"falsifier":"Apply the identical PRF and XGBoost pipelines, without re-tuning, to a large set of simulated Rubin-like nuclear light curves whose true labels are known; if the PRF no longer shows higher stability and lower false-positive rate at faint magnitudes, the central claim fails.","tokens_in":31249,"feed_emoji":"🌌","tokens_out":864,"duration_ms":15624,"temperature":0.7,"pith_summary":"Tidal disruption events are rare flares that mark a star being torn apart by a supermassive black hole. Upcoming surveys will find far more of them than can ever receive spectra, and many will be too faint for host-galaxy tricks that current classifiers rely on. This paper shows that eleven photometric light-curve features—rise and decay times, temperature evolution, and pre-flare variability—are enough to separate TDEs from supernovae and AGN without any host information. The key is a Probabilistic Random Forest that treats every feature as a distribution rather than a single number, so measurement noise is carried into the decision. Against a standard deterministic booster the probabilistic model is more stable on ambiguous cases and rejects substantially more false positives when completeness is prioritized. Applied to archival ZTF data it recovers eleven new candidates and flags three previously mislabeled sources. The result is a host-independent pipeline whose conservative behavior is designed for the low-signal regime that will dominate the next decade of transient astronomy.","feed_headline":"Light curves alone can flag TDEs if noise is treated as data","feed_subtitle":"A probabilistic forest that carries feature errors beats rigid classifiers on faint, ambiguous flares","key_machinery":"The Probabilistic Random Forest (PRF): a bagging ensemble that routes each object through every tree split according to the cumulative distribution of its measured features, so measurement variance is propagated into both the predicted class probabilities and the feature-level contributions.","core_discovery":"TDEs can be reliably identified from photometric light-curve features alone in a host-agnostic framework. A Probabilistic Random Forest that ingests feature uncertainties as distributions yields higher stability for ambiguous candidates and rejects far more false positives in the high-recall regime than a deterministic XGBoost baseline, establishing that uncertainty-aware classifiers are required to avoid brittle, overconfident errors at low signal-to-noise.","pith_inferences":["The same uncertainty-propagation idea should transfer to other rare nuclear classes (changing-look AGN, partial TDEs) whose light curves are equally sparse.","Once physically motivated light-curve injections replace linear SMOTE, the same eleven features could support early-time rather than full-light-curve classification.","A gradient-boosting variant that natively accepts feature distributions would combine XGBoost’s recall with PRF’s noise handling."],"forward_implications":["Host-galaxy cuts and WISE AGN indicators become optional rather than required for photometric TDE selection.","In high-completeness searches the PRF can cut roughly a third more contaminants than a deterministic booster at the same recall.","Eleven new archival candidates and three reclassified labeled sources become concrete spectroscopic follow-up targets.","Faint-end classifiers for the Rubin alert stream should propagate feature uncertainties rather than treat medians as exact.","PRF and XGBoost occupy complementary regimes and can be combined as an orthogonal filter."],"fun_headline_variants":["Light curves alone flag TDEs when noise is treated as data","Host-agnostic PRF classifies TDEs from photometry with uncertainties","Probabilistic forest beats rigid models on faint ambiguous flares","TDEs identified without hosts by feeding feature errors to PRF","Uncertainty-aware PRF cuts false TDE positives at low signal-to-noise"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That aggressively oversampling the handful of real TDEs by linear interpolation in feature-and-uncertainty space produces synthetic examples realistic enough for the learned boundaries to hold on genuinely faint, sparsely sampled light curves.","fun_headline_variants_meta":{"raw":{"variants":["Light curves alone flag TDEs when noise is treated as data","Host-agnostic PRF classifies TDEs from photometry with uncertainties","Probabilistic forest beats rigid models on faint ambiguous flares","TDEs identified without hosts by feeding feature errors to PRF","Uncertainty-aware PRF cuts false TDE positives at low signal-to-noise"]},"model":"grok-4.5","effort":"low","cost_usd":0.004279,"raw_usage":{"total_tokens":1357,"prompt_tokens":857,"num_sources_used":0,"completion_tokens":75,"cost_in_usd_ticks":42788000,"prompt_tokens_details":{"text_tokens":857,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":425,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":857,"tokens_out":75,"duration_ms":8238,"temperature":1.0,"reasoning_tokens":425,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T05:04:19.032523+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Apply the identical PRF and XGBoost pipelines, without re-tuning, to a large set of simulated Rubin-like nuclear light curves whose true labels are known; if the PRF no longer shows higher stability and lower false-positive rate at faint magnitudes, the central claim fails.","supporting_citations":[],"review_version":1}