{"id":"89707a15-9da0-4189-a35c-eab314339afd","arxiv_id":"2607.11539","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"PCA-Transformer spectrum classifier recovers two new optical-UV TDEs from SDSS DR7, including the earliest known with occurrence before MJD 52316.","lead":"A PCA-enhanced Transformer classifier scanned SDSS DR7 galaxy spectra and recovered two previously unreported optical-UV tidal disruption events, one occurring before February 2002. The result shows that modern spectrum classifiers can extract rare transients from archival catalogs that pre-date wide-field time-domain surveys.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged template scarcity and post-hoc cuts.","rationale":"The paper's strongest claim is the concrete identification of two early optical-UV TDEs, not a claim of perfect classifier purity. The reader's weakest_assumption correctly flags the small real-TDE library and unquantified post-selection cuts as the main limitation that keeps the verdict CONDITIONAL rather than unconditional ACCEPT. That limitation is real but does not load-bear against the two specific objects once the multi-wavelength evidence is examined: both satisfy the same spectroscopic and photometric criteria used for the modern ZTF TDE sample, and the light curves are public and re-reducible. No additional load-bearing flaw (e.g., redshift error, continuum mis-fit, or circular use of the same spectra in training) appears in §§2–3 or the appendices. Therefore the reader's CONDITIONAL/HIGH verdict is already correctly calibrated; no adjustment is required.","tokens_in":24915,"tokens_out":693,"duration_ms":7437,"concrete_test":"Independently re-extract the GALEX FUV/NUV photometry for J1242 (gPhoton2, same apertures as Fig. 5) and the CRTS V-band light curve for J1524; confirm the co-temporal UV transient and the optical outburst start window 54269 < MJD < 54476. If either light-curve feature disappears under re-reduction, the corresponding TDE claim collapses; if both survive, the discoveries stand regardless of training-set size.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest_assumption already isolates the central soft spot: only ~10 unique real TDE sources (Table 1) plus synthetic host-mixing (scales 0–2, §2.3.3) and hand-crafted multi-wavelength cuts (§3.1.2: monotonic burst –0.5/+2 yr, DRW τ_3σ-low, WISE χ²/d.o.f.<10, W1–W2<0.8) are used to promote the two objects from classifier candidates to confirmed TDEs. That is a genuine limitation of purity quantification, but it does not undermine the concrete discoveries themselves. Both retained sources rest on independent, publicly checkable evidence that is not circular with the training set: J1242 shows a clear GALEX FUV/NUV transient co-temporal with the SDSS spectrum (MJD 52316) that fades by ~2.5–3.5 mag three years later (Fig. 5), blue continuum + broad Balmer/He features without strong [O III], and no stochastic DRW or MIR variability; J1524 shows a classic TDE-H+He spectrum (broad Hα, Hβ, He II, He I, N III) taken during a CRTS optical outburst whose start is bracketed by the light curve (Fig. 5). These observables match the empirical TDE criteria of van Velzen et al. (2021) and Hammerstein et al. (2023) independently of the Transformer. The classifier's role is discovery, not final authentication; the authentication is multi-wavelength and falsifiable. No deeper internal inconsistency or hidden assumption that would reverse the two claims is present.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper presents a PCA-enhanced Transformer (MgFormer) classifier for identifying TDE-like optical spectra, trained on WISeREP/NGSF transient templates mixed with SDSS host galaxies and evaluated on a held-out synthetic set (precision 0.88, recall 0.99 for TDEs). Applied to ~3×10^5 quality-selected SDSS DR7 galaxy spectra, the model yields 14 candidates with TDE score >0.5; multi-wavelength vetting (GALEX, CRTS, ZTF, WISE) plus emission-line diagnostics retain two newly reported TDEs—SDSS J124225.39+642919.0 (UV transient co-temporal with the spectrum at MJD 52316, claimed as the earliest optical-UV TDE) and SDSS J152459.70+045423.1 (TDE-H+He spectrum during a CRTS outburst with start 54269 < MJD < 54476)—plus one previously reported likely TDE. The pipeline is released open-source.","tokens_in":25420,"tokens_out":1397,"duration_ms":27571,"significance":"If the two new events hold under independent scrutiny, the work is a clear contribution: it recovers serendipitous TDE spectra from a widely used archival catalog, pushes the earliest known optical-UV TDE occurrence to before 2002 February, and demonstrates a practical ML route for spectroscopic TDE searches ahead of DESI-scale samples. Strengths include a clean train/test split with source-level separation for transients, explicit PCA templates and mixing protocol, open code, and—most importantly—authentication of the two discoveries by public multi-band light curves and spectral fits that are independent of the training labels. The method paper and the discovery paper are both of interest to the TDE and time-domain communities.","major_comments":[{"comment":"§3.1 and Table 6 vs §2.4/Fig. 3: On the synthetic test set the TDE precision is 0.88, but among the 14 real SDSS candidates with score >0.5 only 2–3 survive multi-wavelength vetting (real-world precision ~0.14–0.21). This domain-shift gap is load-bearing for the claim that the classifier is a powerful discovery tool for large catalogs. The main text should quantify and discuss it explicitly (expected false-positive rate under the adopted threshold, role of AGN/SF impostors, and how purity would scale to DESI), rather than leaving the discrepancy largely to Appendix B.","section":"§3.1, Table 6, Fig. 3"},{"comment":"§2.2.1 / Table 1: The TDE training library comprises only 10 unique sources (3 H, 5 H+He, 2 He; 47 spectra). Appendix B correctly notes that this scarcity, and the absence of featureless-TDE templates, limits generalization and score interpretability. Because the abstract and introduction present the classifier as a general selection method, the main text (not only the appendix) should state this limitation up front and clarify which spectroscopic subclasses the model is expected to recover or miss.","section":"§2.2.1, Table 1, Appendix B"},{"comment":"§3.1.2 and Table 7: Final promotion from candidate to TDE rests on hand-crafted photometric cuts (monotonic burst in −0.5/+2 yr, DRW τ_3σ-low thresholds, WISE χ²/d.o.f. < 10, W1−W2 < 0.8). These cuts are reasonable and the two retained objects pass cleanly, but the paper should test or at least discuss sensitivity of the final sample to the window and threshold choices, and state that authentication—not the Transformer score—is what establishes the TDE nature. A short robustness paragraph would strengthen the discovery claims without changing the conclusions.","section":"§3.1.2, Table 7"}],"minor_comments":[{"comment":"Eq. (1) and §2.3.3: Both transient and galaxy spectra are min–max normalized before mixing, so Scale is a ratio of normalized fluxes, not a physical continuum flux ratio. A one-sentence clarification would avoid misinterpretation when others reuse the pipeline.","section":"§2.3.3, Eq. (1)"},{"comment":"§3.2 (J1242): The He II/Hβ complex is described as hard to resolve, yet FWHMs with large uncertainties are reported. Consider stating more clearly which lines are robust detections versus tentative, and whether the object would still be classified as TDE-H (or featureless-like) under the van Velzen/Hammerstein scheme if He II is not required.","section":"§3.2"},{"comment":"Figure 5 / light-curve panels: Magnitudes are offset for display; the offset values are given in the legend but a brief note in the caption that fluxes are not host-subtracted (already in the text) would help casual readers.","section":"Fig. 5"},{"comment":"Table 7 is very dense; a short legend defining each column’s decision rule (already partly in the notes) or moving the full decision tree to the appendix would improve readability.","section":"Table 7"},{"comment":"Minor consistency: SDSS names sometimes differ by 0.01 s or last digit between Table 6, Table 7, and the text (e.g., J074820.66 vs J074820.67). Unify identifiers.","section":"Tables 6–7"},{"comment":"The abstract’s phrase “inspiring discoveries” is informal for a journal abstract; a more neutral wording would match the rest of the paper’s tone.","section":"Abstract"}],"recommendation":"minor_revision","confidential_remarks":"The two discovery claims look solid on the public photometry and spectra; I would not hold the paper over the synthetic-vs-real precision gap if the authors add a clear discussion. The work is a good fit for an astrophysics journal that publishes both methods and TDE discoveries. No concerns about circularity or undisclosed novelty issues."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The headline is simple: they built a PCA-enhanced Transformer that scores mixed transient+host spectra, ran it on ~300k SDSS DR7 galaxies, and pulled out two previously unreported optical-UV TDEs, one of which (J1242) is now the earliest known (spectrum at MJD 52316 with a co-temporal GALEX UV transient that later fades). The second (J1524) is a clean TDE-H+He spectrum taken during a CRTS optical outburst. Both rest on independent, publicly checkable photometry and line fits, not on the model score alone.\n\nWhat is new is the systematic spectrum-mining approach itself. Prior optical TDE work was mostly photometric selection plus visual spectroscopy; here the classifier is the discovery engine and the multi-band light curves (GALEX, CRTS, ZTF, WISE) plus the usual empirical cuts (blue continuum, broad H/He/Bowen, no strong [O III], single burst, no DRW/MIR stochasticity) do the authentication. Synthetic-test metrics are cleanly reported (precision 0.88, recall 0.99) with confusion matrices, the pipeline is open-source, and the two objects match the van Velzen/Hammerstein spectroscopic criteria on their own. That is useful archival work and a reusable tool for DESI-scale catalogs.\n\nThe soft spot is exactly the one the reader flagged and the stress-test confirmed: only ~10 unique real TDE templates in the WISeREP library, host-mixing scales that are hand-chosen, and post-hoc light-curve cuts whose purity is not quantified on a control sample. That limits how much we can trust the classifier’s absolute purity, but it does not undercut the two concrete discoveries—the authentication is multi-wavelength and falsifiable. Minor free parameters (PCA component counts, MgFormer hyperparameters, score threshold) exist but are not load-bearing for the claims. Citation pattern is normal; self-citation of their earlier photometric TTC is appropriate.\n\nThis is for people who mine large spectroscopic archives or care about early TDE demographics. It deserves a serious referee. I would bring it to reading group and would cite the two objects and the method.","headline":"Two real early TDEs dug out of SDSS DR7 by a PCA-Transformer spectrum classifier; the discoveries hold up on public light curves even though the training library is tiny.","tokens_in":25972,"tokens_out":543,"would_cite":true,"duration_ms":6065,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A PCA-Transformer spectrum classifier finds two previously unknown optical-UV tidal disruption events in SDSS DR7, one from before February 2002.","keywords":["Tidal disruption events","Transformer","Transient","Spectrum classification","SDSS","PCA","optical-UV TDE"],"falsifier":"Deep multi-epoch UV or soft X-ray imaging of the two hosts that shows either no fading continuum after 2002/2008 or persistent stochastic variability and strong narrow [O III] that would reclassify them as AGN.","tokens_in":25847,"feed_emoji":"🔭","tokens_out":626,"duration_ms":7396,"temperature":0.7,"pith_summary":"Large spectroscopic surveys contain millions of galaxy spectra, so some should have been taken while a rare tidal disruption event was flaring. The authors build a PCA-enhanced Transformer classifier that scores spectra for TDE-like blue continua and broad H/He/Bowen lines, then apply multi-band light-curve cuts to reject AGN and star-formation impostors. In the widely used SDSS DR7 they recover three candidates: one already-noted likely TDE and two new ones. One shows a UV transient already bright at the spectrum epoch (MJD 52316), making its start earlier than any previously known optical-UV TDE; the other was caught mid-outburst in CRTS with a classic TDE-H+He spectrum. The result shows that machine-learning spectrum classifiers can mine archival catalogs for events that time-domain surveys never flagged.","feed_headline":"ML classifier digs two earliest optical-UV TDEs out of SDSS","feed_subtitle":"One flare started before February 2002; the method turns archival spectra into a TDE discovery engine","key_machinery":"PCA-Transformer (MgFormer) spectrum classifier: spectra are projected onto fixed galaxy/transient/stellar PCA templates to form dual-channel coefficient matrices that a multi-group Transformer then scores for TDE probability; high-score candidates are further vetted by burst-like multi-band light curves and absence of persistent AGN variability.","core_discovery":"A PCA-Transformer TDE spectrum classifier applied to SDSS DR7 recovers two previously unreported optical-UV TDEs (plus one known likely TDE). SDSS J124225.39+642919.0 exhibits a blue continuum and broad lines together with a GALEX UV transient present at the spectrum epoch, so the flare began before MJD 52316 and is the earliest known optical-UV TDE. SDSS J152459.70+045423.1 shows a full TDE-H+He spectrum taken during a CRTS optical outburst whose onset lies between MJD 54269 and 54476.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Transformer digs two earliest optical-UV TDEs from SDSS DR7","PCA-Transformer recovers 2002 and 2007 optical-UV TDEs in SDSS","ML classifier finds two earliest TDEs hidden in SDSS spectra","Two new earliest optical-UV TDEs unveiled in SDSS DR7 archive","Spectrum classifier turns SDSS into early TDE discovery engine"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The small library of real TDE templates and the chosen host-mixing plus light-curve cuts are assumed to be enough to guarantee that the two retained objects are genuine TDEs rather than rare AGN or star-formation impostors that pass the same empirical filters.","fun_headline_variants_meta":{"raw":{"variants":["Transformer digs two earliest optical-UV TDEs from SDSS DR7","PCA-Transformer recovers 2002 and 2007 optical-UV TDEs in SDSS","ML classifier finds two earliest TDEs hidden in SDSS spectra","Two new earliest optical-UV TDEs unveiled in SDSS DR7 archive","Spectrum classifier turns SDSS into early TDE discovery engine"]},"model":"grok-4.5","effort":"low","cost_usd":0.004248,"raw_usage":{"total_tokens":1376,"prompt_tokens":954,"num_sources_used":0,"completion_tokens":82,"cost_in_usd_ticks":42480000,"prompt_tokens_details":{"text_tokens":954,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":340,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":954,"tokens_out":82,"duration_ms":3836,"temperature":1.0,"reasoning_tokens":340,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T04:53:01.842822+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Deep multi-epoch UV or soft X-ray imaging of the two hosts that shows either no fading continuum after 2002/2008 or persistent stochastic variability and strong narrow [O III] that would reclassify them as AGN.","supporting_citations":[],"review_version":1}