{"id":"8e9a2e17-6627-4c16-bbed-fc24b5f24273","arxiv_id":"2412.02804","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The excess of active galactic nuclei in galaxy mergers peaks within 160 million years after coalescence and persists for over a gigayear.","lead":"Using a machine-learning model that estimates how long ago a galaxy merger happened, this paper maps when galaxies switch on their central black holes. It finds the boost in black-hole activity peaks right after the two galaxies merge and can last over a billion years.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MUMMI's T_PM labels may be biased by AGN light: if AGN hosts are preferentially assigned to the 0-0.16 Gyr bin, the claimed peak is not a temporal signal.","rationale":"The reader's CONDITIONAL verdict is appropriate. I agree that MUMMI's T_PM accuracy is the load-bearing premise, and I sharpen it: the relevant failure mode is not random misclassification (which would dilute a real peak) but systematic misclassification correlated with AGN presence, because AGN contribute to the same r-band images used for morphology. The paper provides no test of this. Internal checks (three independent AGN diagnostics, matched control samples, agreement with TNG simulations) strengthen the observational result, but they do not remove the dependence on the unpublished classifier. The highlighted imbalance among T_PM bins and the lack of a published confusion matrix make an AGN-bias test the single most important check. If the injection experiment shows no bias, the central claim stands; if it shows a bias, the peak epoch and the long-lived excess in mid-IR/BLAGN would need reanalysis. I therefore do not change the reader's conditional verdict, but the condition should be made explicit: the T_PM classification must be shown to be invariant to nuclear point-source luminosity.","tokens_in":24988,"tokens_out":7174,"duration_ms":80355,"concrete_test":"Build r-band mock images of TNG post-mergers with known T_PM (same pipeline as Ferreira et al., in prep) and inject a central point source whose brightness spans the AGN range at z~0.1 (e.g., nuclear M_r from -18 to -22). Run MUMMI on the injected and original images, and measure the T_PM-bin confusion matrix as a function of injected nuclear luminosity. If the completeness of the 0-0.16 Gyr bin increases with point-source brightness, or if AGN hosts are systematically shifted to earlier bins, then the observed peak excess is not an unbiased temporal measure and the central claim is not yet supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that AGN excess peaks at 0<T_PM<0.16 Gyr relies entirely on MUMMI's time-bin assignments (Sec. 2.3). These assignments come from an unpublished companion paper that reports only an average 70-80% bin accuracy; no per-bin purity/completeness or dependence on AGN properties is given. MUMMI is trained on TNG mock r-band images that do not include AGN point sources, yet it is applied to real galaxies where luminous AGN can add a central PSF to the same r-band images used for classification. If a bright unresolved nucleus makes a galaxy appear more concentrated and therefore more recently merged, AGN hosts will be preferentially placed in the 0-0.16 Gyr bin, producing a false peak. The highly unbalanced bin sizes (785, 740, 778, 5838) suggest a strong classifier prior toward late times, but without a confusion matrix one cannot exclude the opposite bias in the first bin. Random 70-80% accuracy would dilute a real peak, not create one; the danger is the systematic, AGN-correlated component that is currently untested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper combines a sample of close galaxy pairs and post-merger galaxies identified by the MUMMI machine-vision pipeline in UNIONS, with SDSS DR7 spectroscopy and WISE photometry, to measure how the excess frequency of AGN (relative to a mass- and redshift-matched control sample) evolves across the merger sequence, from wide-separation pairs through coalescence to T_PM = 1.76 Gyr. AGN are selected with three independent diagnostics: narrow emission lines (NLAGN), broad emission lines (BLAGN), and mid-IR W1-W2 colors. The authors report that the AGN excess peaks in the first post-coalescence bin (0 < T_PM < 0.16 Gyr), that the excess is long-lived for mid-IR and BLAGN diagnostics, that more luminous AGN show larger excesses, that the fractional contribution of unobscured (broad-line) AGN rises after coalescence, and that the starburst excess peaks contemporaneously with the AGN excess.","tokens_in":25212,"tokens_out":5428,"duration_ms":54963,"significance":"If the T_PM assignments are reliable, this is the first observational measurement of the post-coalescence time evolution of AGN triggering, a result that directly tests simulation predictions and bears on feedback timescales. The study has several genuine strengths: the AGN diagnostics are independent of the machine-learning labels (so there is no circularity in the measured excess), the control matching is carefully constructed and demonstrated, the sample is large (8,141 post-mergers with quality cuts), and the paper makes clear, falsifiable predictions, such as the peak excess at 0-0.16 Gyr and the persistence of mid-IR/BLAGN excesses to 1.76 Gyr. The agreement with Illustris-TNG and disagreement with EAGLE, if correct, would be an important constraint on subgrid models of black hole fueling. However, the central claim depends entirely on the accuracy of MUMMI's time-post-merger bins, which is currently only summarized as a global 70-80% figure from an unpublished companion paper, so the observational result is not yet as secure as the abstract implies.","major_comments":[{"comment":"The central claim that the AGN excess peaks at 0 < T_PM < 0.16 Gyr rests entirely on MUMMI's T_PM bin assignments, but the only accuracy information provided is a global '70-80 per cent' statement from an unpublished companion paper. No per-bin confusion matrix, purity/completeness values, or dependence of classification accuracy on AGN properties is given. The training images are TNG mock r-band images that do not include AGN point sources, so a bright unresolved nucleus in a real galaxy could plausibly make the remnant appear more concentrated and shift it into the shortest T_PM bin, creating the observed peak even if the true temporal trend is flat. This concern is compounded by the highly unbalanced bin populations (785, 740, 778, 5838 in the four T_PM bins), which suggest a strong classifier prior. The authors should provide the confusion matrix and per-bin accuracy from Ferreira et al. (in prep), and test for AGN-correlated biases by injecting point sources into the TNG mock images or by comparing T_PM assignments with an independent morphological or stellar-population age indicator.","section":"2.3"},{"comment":"The quoted uncertainties on the AGN excess are binomial errors on the fractions, and therefore propagate neither the 70-80% T_PM classification uncertainty nor the scatter in the control matching (which is iterative and reuses control galaxies across samples). The headline comparison (excess 3.1 +/- 0.2 vs 2.0 +/- 0.2 in the first two bins) is thus presented as more significant than the data actually support. The authors should propagate the T_PM confusion matrix (e.g., by Monte Carlo resampling of bin assignments) and the control-matching scatter (e.g., by bootstrap resampling of the control pools) into all excess values and significance claims.","section":"3.1, Tables 1-4"},{"comment":"The interpretation of the post-merger decline and the comparison with EAGLE presumes that the four T_PM bins have comparable purity and completeness. The fourth bin (0.96 < T_PM < 1.76 Gyr) contains 5,838 galaxies versus ~750 in each of the first three bins, and the bin lengths are deliberately unequal. If the longer bin has lower completeness or a different contamination rate (both plausible given the stated 70-80% average accuracy and the increasing difficulty of time prediction at late stages), the apparent decline in excess beyond ~1 Gyr - which is used to claim that the NLAGN excess ends by ~1 Gyr and that EAGLE's delayed ~300 Myr peak is ruled out - could be an artifact. Per-bin accuracy and completeness figures, plus a test of how the derived excesses change if the last bin is subdivided or if classification errors are simulated, are necessary to support the temporal-decay claims.","section":"2.4 and 4.1"},{"comment":"The conclusion that there is a deficit of broad-line AGN in the pre-merger phase is based on excess values that the text itself describes as 'marginal,' and several bins appear consistent with unity within their error bars. Since this deficit is used to motivate the dust-covering-fraction narrative in the abstract and Section 5, the paper should either quote the relevant significances (or show the deficit is significant after including T_PM and matching uncertainties) or explicitly state that the pre-coalescence obscuration signal is only suggestive. As written, the strength of the claim in the abstract exceeds the statistical support in the figure.","section":"3.3 and Figure 6"}],"minor_comments":[{"comment":"In the first row of Table 3, the post-merger AGN fraction is reported as '0.13 +/- 0.1'; the error should presumably be approximately 0.01, based on the binomial error for 104/785.","section":"Table 3"},{"comment":"In the first row of Table 4, the control BLAGN fraction is listed as '0.017 +/- 0.007'; for 435/25,905 the binomial error is approximately 0.0008, so the quoted error appears to be off by an order of magnitude.","section":"Table 4"},{"comment":"The phrase 'Figure of Donley et al. 2012' should be 'Figure 1 of Donley et al. 2012' for clarity.","section":"2.5"},{"comment":"In the introduction, 'Omori et el. 2023' should be 'Omori et al. 2023'.","section":"1"},{"comment":"There is a typo in 'occured' in the sentence describing the first snapshot after coalescence; it should be 'occurred'.","section":"2.3"},{"comment":"The summary states that 'radiation pressure from the AGN redistributes this dust' as if it were a direct measurement; since only the BLAGN fraction is measured, consider softening this to 'is consistent with redistribution of nuclear dust by AGN feedback.'","section":"5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript leans very heavily on Ferreira et al. (in prep) for the central T_PM reliability claim. The editor may wish to ensure that the companion paper is either published or that the per-bin validation metrics are included in the present paper before final acceptance. The current global 70-80% accuracy statement is insufficient for the strong temporal conclusions drawn here. The paper is otherwise well structured and the scientific question is important; the requested validation appears feasible within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the genuinely new thing: this is the first observational census of AGN excess binned by time since coalescence. That alone makes it worth a hard look. The peak at 0 < T_PM < 0.16 Gyr is reproduced across three independent AGN diagnostics, with matched controls and binomial errors that make the factor-of-three excess in the first bin hard to dismiss. The luminosity dependence and the obscuration narrative are nice add-ons, and the contemporaneous starburst peak is a useful datapoint.\n\nThe soft spots are real but not fatal. The load-bearing piece is MUMMI's T_PM assignments, which are only described in an unpublished companion paper. The paper quotes 70–80% bin accuracy but gives no per-bin confusion matrix and no test of whether misclassification correlates with AGN properties. The stress-test worry that AGN light makes hosts look more concentrated and pushes them into the first bin is plausible; I can't rule it out from this paper. Random 70–80% accuracy would dilute a real peak, not create one, but the systematic component is untested. The authors should either include a per-bin purity table in an appendix or show the trend survives in a sample restricted to low-AGN-contribution galaxies. The quoted errors also omit classification and matching scatter; that's a minor complaint since the binomial errors are subdominant to the systematic question. The pre-merger BLAGN deficit is called marginal in the text but appears as a result in the abstract; the abstract oversells that one. And the pair sample and post-merger sample come from different survey footprints, so the joint sequence is not perfectly homogeneous; the authors acknowledge the matching but don't discuss this.\n\nThe core result is probably right, but it's not yet settled. This deserves a serious referee. I'd send it to review, conditional on the authors opening up the MUMMI systematics and softening the BLAGN-deficit language. It's a strong paper that needs one more round of robustness work.","headline":"First time-resolved post-merger AGN census, plausible but the classifier systematics need to be opened up before I'd bet on the peak bin.","tokens_in":25817,"tokens_out":2233,"would_cite":true,"duration_ms":24044,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Galaxy mergers trigger AGN most strongly in the first 160 million years after coalescence, with elevated rates persisting past 1.76 Gyr for mid-infrared and broad-line selected AGN.","keywords":["galaxy mergers","active galactic nuclei","post-merger galaxies","time post-merger","AGN triggering","machine vision classification","nuclear obscuration","starburst-AGN connection"],"falsifier":"Re-date the same post-merger galaxies with an independent method, such as fitting tidal features or stellar populations, and recompute the AGN excess in the same four time bins; if the peak moves out of the 0 to 0.16 Gyr bin, or the long-lived mid-infrared and broad-line excess disappears, the original bin assignments were creating the temporal signal.","tokens_in":24760,"feed_emoji":"🔭","tokens_out":9530,"duration_ms":93222,"temperature":0.7,"pith_summary":"This paper asks when, along the sequence of a galaxy merger, a supermassive black hole is most likely to be actively accreting. The authors combine a sample of pre-coalescence galaxy pairs with post-merger remnants whose time since coalescence is estimated by a machine-vision pipeline, and compare AGN fractions in each stage with matched non-merger controls. They find that the excess of AGN peaks immediately after coalescence, in the bin 0–0.16 Gyr, regardless of whether AGN are selected by narrow emission lines, mid-infrared colours, or broad emission lines. The excess remains statistically significant out to the longest bin, 0.96–1.76 Gyr, for mid-infrared and broad-line selected AGN. This matters because it places the first direct observational timing constraint on when mergers feed black holes and connects that timing to nuclear dust clearing and starbursts.","feed_headline":"Black hole feeding peaks within 160 Myr of galaxy merger","feed_subtitle":"Three independent AGN diagnostics agree, and excess activity lasts past 1.76 Gyr.","key_machinery":"The load-bearing tool is MUMMI, a machine-vision ensemble of neural networks trained on simulated galaxies that both identifies post-merger remnants and predicts the time since coalescence in four bins: 0–0.16, 0.16–0.48, 0.48–0.96, and 0.96–1.76 Gyr. This converts the post-merger population into a timeline, which is what allows the paper to ask when AGN triggering peaks and how long it lasts. The three AGN diagnostics probe distinct physical regions—narrow lines from ionized gas on roughly kiloparsec scales, mid-infrared colours from the dusty torus, and broad lines from the innermost accretion region—and their different responses to merger stage carry the obscuration argument. The AGN excess, defined as the ratio of AGN fractions in mergers versus matched controls, is the common metric on which all trends are drawn.","core_discovery":"The paper's central discovery is that the statistical enhancement of AGN in merging galaxies is largest right after the two galaxies coalesce, at $0 < T_{\\rm PM} < 0.16$ Gyr, and then declines as the remnant ages. The result holds for all three AGN diagnostics, and the pattern differs by diagnostic in a way the authors interpret as evolution of nuclear obscuration: narrow-line and mid-IR selected AGN show an excess already in the pair phase, while broad-line AGN show a mild deficit before coalescence that becomes an excess afterward, implying the broad-line region is hidden by inflated dust covering fractions during the encounter and partially cleared after it. The excess is largest for the most luminous and bolometrically dominant AGN, mirroring simulations; merger-hosted narrow-line AGN are about 2.5 times more luminous than their secular counterparts. The paper also shows that the statistical peak of starbursts is contemporaneous with the AGN peak, with any lag between the two processes constrained to less than roughly 150 Myr.","pith_inferences":["Editorial inference: because the first time bin is only 0.16 Gyr wide, the true peak of accretion could be even sharper and closer to coalescence than the bin-averaged excess shows; simulations or samples with finer time resolution in the first few hundred megayears would reveal it.","Editorial inference: the long-lived mid-infrared excess may be partly a byproduct of merger-driven quenching rather than continued black hole growth; splitting post-mergers by specific star formation rate would test whether less star-forming remnants are more likely to show a mid-infrared AGN signature.","Editorial inference: the covering-fraction picture predicts that X-ray column densities toward AGN in close pairs should be systematically higher than toward AGN in post-mergers; a dedicated X-ray census of these systems would directly test the blowout interpretation.","Editorial inference: if this timing is universal, merger-AGN surveys at higher redshift that find no excess may be washing out the signal by averaging over a wide range of merger stages; applying time-bin classification to those samples would sharpen the comparison."],"forward_implications":["The peak in AGN excess at $0 < T_{\\rm PM} < 0.16$ Gyr appears with every AGN diagnostic, so the finding does not depend on a single selection method.","Mid-infrared and broad-line AGN remain more common than in controls out to the longest time bin, meaning merger-enhanced nuclear activity outlasts the star-formation enhancement, which fades by about 1 Gyr.","The excess is stronger for more luminous and bolometrically dominant AGN, and merger-hosted narrow-line AGN are about 2.5 times more luminous than secular ones, implying mergers preferentially produce powerful accretion episodes.","The pre-merger deficit and post-merger excess of broad-line AGN implies the nuclear dust covering fraction rises during the encounter and falls after coalescence, consistent with AGN feedback clearing some of the obscuring material.","Starburst and AGN excesses peak in the same 0.16 Gyr interval, so any statistical delay between the two triggers is less than about 150 Myr."],"supporting_citations":[{"why":"Presents the MUMMI machine-vision pipeline and the merger classifications applied to the imaging sample.","marker":"Ferreira et al. 2024"},{"why":"Supplies the time-post-merger bin predictions and the 70–80 percent bin-assignment accuracy that define the four temporal stages.","marker":"Ferreira et al. in prep"},{"why":"Provides a cosmological simulation benchmark that merger-enhanced AGN accretion peaks at coalescence and persists about 2 Gyr.","marker":"Byrne-Mamahit et al. 2024"},{"why":"Simulation showing reduced line-of-sight column density roughly 40 Myr after a merger, used to interpret the broad-line deficit and post-coalescence excess.","marker":"Blecha et al. 2018"},{"why":"Gives the conversion between pair separation and time before merging, placing pairs on the same timeline as post-mergers.","marker":"Patton et al. 2024"},{"why":"Supplies the stringent emission-line demarcation used to classify narrow-line AGN.","marker":"Kewley et al. 2001"},{"why":"Provides the broad-line AGN catalog used as one of the three AGN diagnostics.","marker":"Liu et al. 2019"}],"fun_headline_variants":["AGN triggering peaks immediately after galaxy coalescence","Three independent diagnostics agree: AGN peak at merger","Post-merger AGN excess lasts over 1.76 Gyr","AGN and starburst peaks coincide at galaxy merger","Luminous AGN show strongest excess after merger"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The timeline assumes the classifier puts each post-merger galaxy into the right time-since-coalescence bin, but it does so correctly only 70 to 80 percent of the time, so if the mistakes are tied to AGN properties the apparent peak right after coalescence could be an artifact.","fun_headline_variants_meta":{"raw":{"variants":["AGN triggering peaks immediately after galaxy coalescence","Three independent diagnostics agree: AGN peak at merger","Post-merger AGN excess lasts over 1.76 Gyr","AGN and starburst peaks coincide at galaxy merger","Luminous AGN show strongest excess after merger"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001167,"raw_usage":{"total_tokens":4936,"prompt_tokens":1162,"completion_tokens":3774,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":778,"completion_tokens_details":{"reasoning_tokens":3695}},"tokens_in":778,"tokens_out":3774,"duration_ms":29082,"temperature":1.0,"reasoning_tokens":3695,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:05:38.653037+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-date the same post-merger galaxies with an independent method, such as fitting tidal features or stellar populations, and recompute the AGN excess in the same four time bins; if the peak moves out of the 0 to 0.16 Gyr bin, or the long-lived mid-infrared and broad-line excess disappears, the original bin assignments were creating the temporal signal.","supporting_citations":[{"cited_title":"J., Dopita, M","cited_arxiv_id":null,"evidence_quote":"Supplies the stringent emission-line demarcation used to classify narrow-line AGN."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the broad-line AGN catalog used as one of the three AGN diagnostics."}],"review_version":1}