Pith. sign in

REVIEW 4 major objections 5 minor 60 references

The ATLAS Virtual Research Assistant

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The ATLAS Virtual Research Assistant ranks alerts by real and extragalactic scores and, in production, cut eyeballing workload by 85% while losing under 0.08% of follow-up opportunities.

desk verdict A genuinely useful operations paper: the VRA convincingly cuts eyeballing workload by ~85% while keeping recovery of extragalactic transients high; the headline 'loss' metric is a lower bound, not the whole story. read the letter →

arxiv 2506.09778 v2 pith:RTMOVCER submitted 2025-06-11 astro-ph.IM astro-ph.HEastro-ph.SR

H. F. Stevance , K. W. Smith , S. J. Smartt , S. J. Roberts , N. Erasmus , D. R. Young , A. Clocchiatti This is my paper
classification astro-ph.IMastro-ph.HEastro-ph.SR
keywords skysurveystransientdetectionreal-bogusclassificationgradientboosteddecisiontreesrecallatrankKalerttriageautomatedfollow-upATLAS
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The ATLAS Virtual Research Assistant is a bot that pre-screens the survey's transient alerts so humans only look at the promising ones. The paper argues that a feature-based machine-learning system can rank alerts by how real and how extragalactic they are, rather than trying to classify the transient type, and that this matches how the human eyeballers actually work. In production, the VRA auto-discarded 85% of the daily alert stream and the authors estimate that under 0.08% of follow-up opportunities were lost. If this holds, the same approach could be retrained for other surveys with only a few thousand labeled examples, which matters as LSST and other streams outgrow manual review.

What carries the argument

The load-bearing object is the two-dimensional 'score space' with axes $p_{\rm real}$ and $p_{\rm gal}$, produced by four histogram-based gradient-boosted decision tree classifiers (day-1 and day-N versions of real/bogus and galactic/extragalactic). Alerts are ranked by the VRA score, a normalized geometric distance to the ideal extragalactic point, $$\mathrm{VRAscore}=10\times\frac{\sqrt{$f^{2}$+1}-\sqrt{(1-p_{\rm real})^2+(f\,p_{\rm gal})^2}}{\sqrt{$f^{2}$+1}},$$ with a galactic-axis scaling factor $f=0.5$ that weights the Real axis twice as heavily as the Galactic axis. The evaluation metric is recall at rank $K$ (R@K), with an area-under-curve variant AuRaK benchmarked against the pre-VRA CNN ordering; the auto-garbaging policies apply day-dependent VRA-score thresholds, discarding alerts below 1 on day 1, below 2 on the second visit, and below 3 on later visits.

What would settle it

Re-eyeball every alert auto-garbaged during the production window 2025-04-04 to 2025-06-10, using forced photometry or independent classification, and count how many are genuine extragalactic transients; if that fraction exceeds the paper's reported 0.079% loss of extragalactic events, the central claim would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that alert triage in a messy transient stream can be reduced to two binary questions—'Is it real?' and 'Is it galactic?'—answered by gradient-boosted decision trees trained on human eyeball labels. Each alert gets a Real score and a Galactic score; these are combined into a single 0–10 VRA score defined as a normalized distance to the (real=1, galactic=0) corner of score space, with TNS cross-matches promoted to rank 10 and new light-curve visits updating the scores through separate day-1 and day-N models. On the validation set, ordering by the day-1 models recovers 95% (99%) of the 'Good' extragalactic alerts in the top 15% (25%) of the list, compared with 35% (60%) for the previous CNN real-bogus ordering, and the day-N models reach 95% recall in the top 5% of the list. In production from April to June 2025, the auto-garbaging policies removed 85% of 16,938 alerts from the eyeball list, and the authors place the missed follow-up opportunity at 0.006% of all alerts and 0.079% of extragalactic events.

Load-bearing premise

The human labels that train and validate the models are treated as accurate ground truth, even though the paper notes PM/Garbage confusion, duplicates and AGNs in the Attic, and only partial re-eyeballing; if label errors concentrate in the extragalactic class, the reported recall and loss figures could be optimistic.

Editorial extensions

If this is right

  • In production, each day's eyeball list shrinks to roughly one-sixth of its pre-VRA size because 85% of alerts are auto-garbaged, and the remaining list is ranked so human scanners see the most promising extragalactic candidates first.
  • The day-N models use new detection and non-detection information from later visits, so an alert's rank improves as its light curve grows; validation shows 95% recall of Good objects in the top 5% of the ordered list.
  • The same scores drive automated follow-up triggers on the Lesedi telescope, meaning a single learned score stream can both rank for humans and activate robotic observations.
  • The R@K benchmark gives other surveys a direct test: a replacement triage model must beat the AuRaK of the existing ordering before it is worth deploying.
  • Feature-based classifiers trained on a few thousand examples can match or exceed neural-network ordering in this setting, offering interpretability and a route to inject domain knowledge.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 85% auto-garbage fraction was measured during a window that included an anomalous week of image trailing; steady-state workload reduction may be closer to the 80.2% predicted from validation than to 85%.
  • The reported loss rate depends on TNS cross-matching to catch misses, so genuine losses that neither VRA nor TNS flags would go uncounted; a forced-photometry audit of auto-garbaged alerts would test this.
  • Transferring this design to LSST or ZTF will require retraining on each survey's artifacts and cadence, since features like extinction and RA/Dec encode ATLAS-specific tiling and moon-phase effects, but the score-space ranking recipe is survey-agnostic.
  • Extending the score space with the deprioritized 'fast' axis, or adding forced-photometry features on day N, is a natural next step that the authors flag as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents the ATLAS Virtual Research Assistant (VRA), a production system that ranks transient alerts using two Histogram-Based Gradient Boosted Decision Tree classifiers ('Real' and 'Galactic'), with separate day-1 and day-N models that incorporate new light-curve information. Alerts are ranked by a geometric VRA score (Eq. 2), and eyeballing/auto-garbaging policies select which alerts humans inspect. The authors benchmark their ranking with Recall at rank K and AuRaK against the pre-existing Real/Bogus CNN ordering, evaluate policies on a validation set, and report in-production statistics from 2025-04-04 to 2025-06-10, claiming an 85% reduction in eyeballing workload with a <0.08% loss of follow-up opportunity. The system also automatically triggers follow-up observations with the Lesedi telescope.

Significance. If the central claims hold, the VRA is a practical, interpretable, and low-cost automation of transient alert triage that is highly relevant to current and future surveys (e.g., LSST). The paper's strengths include a clear benchmark against the pre-VRA RB-score strategy, explicit release of code and training data (Zenodo), transparent reporting of in-production monitoring including 'potential misses', and a design that uses only a few thousand training samples. These are substantial contributions to astroinformatics practice. The main weakness is that the headline loss metric is conditional on TNS independent discovery, and the training and policy evaluation involve a self-referential loop that the paper acknowledges only in part.

major comments (4)
  1. [Section 4.2 and abstract] The headline figure '<0.08% loss of follow-up opportunity' is computed by cross-matching auto-garbaged alerts to TNS (Sections 2.6 and 4.2). This metric only counts transients that were independently reported to TNS by another survey; any transient that ATLAS alone would have recovered pre-VRA but that no other facility noticed is invisible to the estimate. The paper's own monitoring text supports this concern: 164 'potential misses', 34 below threshold for 15 days, 20 real events, and 1 potentially follow-up-worthy (SN2025hkm) were found only because of the TNS cross-match. The loss rate should therefore be stated as a TNS-conditional lower bound, and the abstract and conclusion should not present it as an overall loss rate without this qualification. A random-sample manual review of auto-garbaged alerts or forced-photometry checks would provide a more complete estimate.
  2. [Section 5.4 and Section 4.1] The auto-garbaging thresholds (<1, <2, <3 on visits 1-3) and the eyeballing threshold (>7) were set or adjusted based on the policy evaluation of Section 4.1; Section 5.4 states that 'the values changed slightly based on the policy evaluation presented in Section 4.1'. Therefore the validation-set performance figures in Section 4.1 (e.g., 80.2% auto-garbaged, 97% of Good objects eyeballed) are in-sample estimates and cannot be read as out-of-sample predictive performance. The in-production statistics in Section 4.2 are a better external test, but the paper should explicitly separate in-sample tuning from out-of-sample validation when interpreting both sets of numbers.
  3. [Section 3.2 and Table 2] The training set includes 1,600 (day-1 training) and 3,571 (day-1 validation) samples labeled 'Auto-garbage' by earlier VRA versions, and this class is used as ground truth for p_real=0. Because the VRA's own outputs become training labels, any systematic errors in auto-garbaging policies are reinforced; the TNS cross-match in Section 4.2 indeed shows that real extragalactic transients can be auto-garbaged. Section 3.1 discusses human-label impurity but does not quantify the fraction of auto-garbage labels that were re-eyeballed, nor does the paper test how results change if auto-garbage samples are removed or re-weighted. This is a load-bearing circularity for the training procedure and should be addressed with a sensitivity analysis.
  4. [Section 4.2 vs Section 6] The abstract and Section 6 state 'no loss of follow-up opportunity' and summarize the result as an unqualified 85% workload reduction. Section 4.2 itself reports that one event (SN2025hkm) 'may have been the object of follow-up' and that the 85% figure is inflated by a hardware/weather week (7,516 alerts, 91.5% auto-garbaged in the week starting 2025-05-23). The conclusions should be qualified to match the body: an 85% reduction over a two-month window that includes a trail-rich week, and a TNS-conditional potential follow-up loss of at most one object by the paper's own accounting.
minor comments (5)
  1. [Section 2.6] Typo: 'prugatory sentinel' should read 'purgatory sentinel'.
  2. [Table 1] The description for 'NON N today' reads 'Number of detections seen today', which should be 'Number of non-detections seen today'.
  3. [Section 6] Spelling inconsistency: the telescope is called 'Lesedi' in Sections 2.4 and 6, but 'Lesidi' appears in Section 6 ('the 1-m Lesidi Telescope'). Please unify.
  4. [Section 3.2.1] Grammar: 'To illustrate this process with show in Figure 6 an example' appears to be missing a verb; consider 'To illustrate this process, we show in Figure 6 an example'.
  5. [Section 4.2] Minor wording: 'transients events miss-labeled' should be 'transient events mislabeled'.

Circularity Check

3 steps flagged · score 4.0 of 10

No fatal circularity: the workload reduction is a measured production outcome, but the loss metric is definitionally TNS-conditional, and the training/policy loop contains self-referential elements.

  1. self definitional [Section 2.6 (Monitoring) and Section 4.2 (In production performance)]
    "Potential misses are defined as alerts which would have not met the VRA rank threshold but whose rank was raised to 10 by a cross-match to TNS. ... as the current potential loss rate is only 0.006% of all the alerts entering the eyeball list, and 0.079% of all the Extra Galactic events."

    The loss-of-follow-up metric counts only transients that appear in TNS: a 'potential miss' is, by definition, an alert whose VRA score was low but whose rank was rescued by a TNS cross-match. Any extragalactic transient that the VRA auto-garbaged and that no other survey reported to TNS is invisible to the metric. The reported <0.08% loss is therefore not an estimate of total lost follow-up opportunity but a measure of TNS-visible loss, so the headline loss figure is partly an artifact of the definition rather than an independent measurement of all missed science.

  2. fitted input called prediction [Section 4.1.2 and Section 5.4 (Choosing the policies)]
    "Based on our validation set, we can expect that over the course of a week 80.2% of the incoming alerts will be auto-garbaged, 3.7% will be left in purgatory and 16% will be sent to human scanners for eyeballing. ... In future iterations of the VRA the garbaging logic remained the same but the values changed slightly based on the policy evaluation presented in Section 4.1."

    The auto-garbaging thresholds were adjusted using the Section 4.1 policy evaluation on the validation set, and then the same evaluation is quoted as an expected production outcome ('80.2% ... will be auto-garbaged'). That expectation is an in-sample consequence of thresholds fitted to that same validation set, not an out-of-sample prediction. The later production measurement of 85% is genuinely independent, so this step is a presentation weakness rather than the main result.

1 more flagged steps
  1. other [Section 3.2 (Data and Features), Table 2]
    "This is the first dataset that is impacted by the VRA, which is reflected in the large fraction of data that are labelled as 'Auto-garbage'. ... Auto-garbage [preal=NaN;p_gal=0] 1,600 3,571"

    A substantial part of the training and validation labels are 'Auto-garbage' outputs produced by earlier versions of the VRA itself. The classifiers are therefore being optimized, in part, to reproduce the VRA's own previous auto-garbaging decisions, creating a feedback loop in the training data. This does not force the headline result because the production loss claim is checked against TNS, but it does mean the model's agreement with 'Auto-garbage' labels is not fully independent evidence that those labels are correct.

full rationale

The central derivation is not circular in the strongest sense. The VRA classifiers are trained on human eyeballing labels (with acknowledged impurities), benchmarked against the existing CNN RB score via R@K and AuRaK, and the headline 85% workload reduction is a direct measurement on 16,938 production alerts between 2025-04-04 and 2025-06-10, not a value recovered from the training set. The TNS cross-match is an external source of ground truth, and the paper manually reviews the 34 sub-threshold potential misses, which is genuine validation effort. However, three steps contain partial self-reference or in-sample fitting. First, the loss metric counts only TNS-registered transients: a 'potential miss' is defined as an alert whose VRA score was low but whose rank was raised by TNS cross-match, so any ATLAS-only miss not reported to TNS is invisible; the <0.08% figure is therefore a bound on TNS-visible loss by construction. Second, the auto-garbaging policy thresholds were adjusted using the Section 4.1 policy evaluation, and the same evaluation is then quoted as the expected 80.2% auto-garbage fraction, so that expectation is in-sample rather than a genuine prediction; the later production number is independent. Third, a portion of the training labels are 'Auto-garbage', i.e., outputs of earlier VRA versions, so the models are partly trained to reproduce the VRA's own previous decisions; this is a feedback loop rather than a derivation, and it is mitigated by the TNS check. The self-citations to the Technical Manual and data release are for implementation details, not for load-bearing uniqueness or correctness claims. Overall the central claim has substantial independent content, so the circularity score is 4, reflecting the TNS-conditional loss definition and the in-sample policy estimate rather than a collapse of the derivation.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claims rest on one fitted ranking metric (VRA score), two hand-tuned scalers, four policy thresholds, and a set of labeling and auditing assumptions. The score-space concept is an internal ranking device, not a physically invented entity. The most consequential assumption is that human eyeball labels are pure enough to serve as ground truth and that TNS catches all missed transients; both are acknowledged as imperfect in the text.

free parameters (6)
  • f (VRA score axis scaler) = 0.5
    Eq. 2 weights the galactic axis; tested 0.4-1.0 all giving AuRaK 0.9518, chosen for interpretability (Sections 2.4, 5.4).
  • f (galactic flag scaler) = 0.9
    Sets distance to (1,1) in score space; chosen by visual inspection to cover galactic distribution without encroaching on bogus (Sections 2.4, 5.4).
  • galactic flag distance threshold = <0.4
    Conservative cutoff chosen by visual inspection, not systematic search (Section 5.4).
  • extragalactic eyeballing threshold = >7
    VRA score threshold for human review; value chosen without grid search (Section 2.4, 5.4).
  • auto-garbaging thresholds = VRA<1 day1; <2 visit2; <3 mean visit3+
    Derived from in-production score distributions and adjusted by policy evaluation on validation data (Sections 2.5, 5.4).
  • dayN cutoff = 15 days
    Models only trained up to day 15; purgatory sentinel flags older alerts (Sections 2.6, 3.2.3).
assumptions (4)
  • domain assumption Human eyeball classification is a valid ground truth for real/bogus and galactic/extragalactic status
    All training labels and R@K evaluations use these categories; the paper acknowledges mislabeling (Section 3.1).
  • domain assumption TNS cross-matching of the garbage list is an adequate external audit of missed follow-up opportunities
    The <0.08% loss claim is measured by matching VRA-garbaged alerts to TNS (Section 4.2); transients never reported to TNS would be invisible.
  • domain assumption The upstream cut (CNN RB score > 0.2) and other pre-VRA filtering are fixed
    VRA is trained and deployed only on alerts that pass earlier pipeline steps (Sections 2.1, 3.2).
  • ad hoc to paper The VRA score distance formula (Eq. 2) orders alerts by scientific priority
    The ranking is a designed heuristic with a scalar f; it is not derived from a utility model (Sections 2.4, 5.4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of The ATLAS Virtual Research Assistant." pith.science (2026). https://pith.science/paper/RTMOVCER

@misc{pith2026250609778,
  author       = {Pith},
  title        = {Pith review of: The ATLAS Virtual Research Assistant},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RTMOVCER}},
  note         = {Machine review of arXiv:2506.09778}
}
read the original abstract

We present the Virtual Research Assistant (VRA) of the ATLAS sky survey which performs preliminary eyeballing on our clean transient data stream. The VRA uses Histogram Based Gradient Boosted Decision Tree Classifiers trained on real data to score incoming alerts on two axes: "Real" and "Galactic". The alerts are then ranked using a geometric distance such that the most "Real" and "Extra-galactic" receive high scores; the scores are updated when new light curve data is obtained on subsequent visits. To assess the quality of the training we use the Recall at rank K, which is more informative to our science goal than general metrics such as accuracy or F1-Scores. We also establish benchmarks for our metric based on the pre-VRA eyeballing strategy, to ensure our models provide notable improvements before being added to the ATLAS pipeline. Finally, policies are defined on the ranked list to select the most promising alerts for humans to eyeball and to automatically remove the bogus alerts. In production the VRA method has resulted in a reduction in eyeballing workload by 85% with a loss of follow-up opportunity <0.08%. It also allows us to automatically trigger follow-up observations with the Lesedi telescope, paving the way to automated methods that will be required in the era of LSST. inally, this is a demonstration that feature-based methods remain extremely relevant in our field, being trainable on only a few thousand samples and highly interpretable; they also offer a direct way to inject expertise into models through feature engineering

Figures

Figures reproduced from arXiv: 2506.09778 by the authors.

Figure 1
Figure 1. encapsulates the problem we will address in this paper, the real, extragalactic transient sources are still only a few percent (5.5% over this period) of the objects that a human on duty will scan through man￾ually. Further automation of the eyeballing process is [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Histogram of the time delay between an alert entering the eyeball list and a human making a classification between 27th March 2024 and 13th August 2024. 2.2. Scope and benchmark [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Left: Distribution of the Real/Bogus score for the human-labeled ”garbage” and ”good” alerts over the period 27th March - 16th August 2024. Right: Recall at rank K for the data set ordered by real/bogus score. When ordering by real/bogus score the eyeballers would have to, on average, eyeball the top 35% of the list to recover 95% of the good objects (amounting to 5.3% of the list). To recover 99% of the good object… view at source ↗
Figures from the paper (21 more)
Figure 4
Figure 4. Figure 4: A toy representation of our Transient agnostic score space, defined by three dimensions: Real (x), Galactic (y), Fast (z). The latter relates to the timescale on which the transient lightcurve evolves - it is not used in this iteration of the VRA and is a more subjecti…
Figure 5
Figure 5. Figure 5: Alert type distribution in our full data sets spanning 27th March 2024 to 22nd January 2025. Some of these alerts were re-eyeballed during development as their human labels were discrepant with their location in Score Space. A few Active Galactic Nuclei (AGNs) were fou…
Figure 6
Figure 6. Figure 6: Visualisation of the lightcurve history used to create the VRA features. The star markers show detections (any filter) and the triangles show the 5-sigma limiting magnitude in the ATLAS frame of each non-detections. The time axis is given with respect to the time of fi…
Figure 7
Figure 7. Figure 7: Score space and R@K of our validation data set for the day 1 models, colour coded according to their human-given labels. The score space is that obtained with the models trained excluding the pruned features (see [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Permutation importance of the day 1 model fea￾tures. implementation; the results for both our Real and Galac￾tic scoring models are presented in [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Score space and R@K of our validation data set for the day N models, colour coded according to their human-given labels. The score space is that obtained with the models trained excluding the pruned features. The R@K plot shows both the R@K curves obtained for the day …
Figure 10
Figure 10. Figure 10: Permutation importance for the day N models of the bogus detections. We can visualize these poli￾cies against our model predictions in [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Visualisation of the extragalactic eyeballing policies (blue line) and galactic candidate policies (yellow line) compared to the distribution of our alert samples. All VRA ranks from 1 to 9 are also shown (grey solid lines). log10 std ra min5d, log10 std dec min5d and…
Figure 12
Figure 12. Figure 12: Transient web server page for AT2024lwd which is mistakenly labeled as garbage by the VRA day 1 models during policy evaluation [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Long term history of AT2024lwd (AT￾LAS ID=1225835351364240100) as recorded by the VRA feature calculator [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: Position of the key transients (lime) listed in [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: VRA outcome for the “Good”, “Galactic” and “Bogus” alerts (which include “Garbage”, “PM” and “Auto￾Garbage”) after four visits. The eyeballed outcome encom￾passes both alerts which are eyeballed as extragalactic and galactic candidates. inspected and many were mislabe…
Figure 17
Figure 17. Figure 17: Alert type distribution over the period 2025-04- 04 to 2025-06-10. See the label description in Section 3.1. 164 “potential misses” - alerts which did not meet the VRA score threshold initially but whose VRA score was raised to 10 when a cross-match to TNS was detecte…
Figure 18
Figure 18. Figure 18: Real Bogus score distribution splits by alert types. The Auto-garbage alerts are plotted in grey over each plot. Note that we logged the x-axis for better visualization. The features given to the scoring algorithms are not logged. Also note that the rb pix feature dis…
Figure 19
Figure 19. Figure 19: Right Ascension (RA) and Declination (Dec) distribution of our data split by types: “Good” (blue), “Galactic” (yellow), “PM” (orange), “Garbage” (red), “Auto-garbage” (grey). A.1.3. Additional on-sky localization features There are three other features related to the …
Figure 20
Figure 20. Figure 20: Additional positional feature distributions. From left to right: The separation between the alert and the most nearby cross-matching catalog source; the standard deviation of the RA and Dec localizations of all detections recorded. All these features were logged (base…
Figure 21
Figure 21. Figure 21: Distribution of the redshift measurements across our alert types. We separate the spectroscopic (z) and photometric redshifts (photo z). We plot separately the labels given by human scanners and show the Auto-garbage label distribution in grey over-top. the other dist…
Figure 22
Figure 22. Figure 22: Distribution of the E(B − V ) feature for our different alert types. We plot separately the labels given by human scanners and show the Auto-garbage label distribution in grey over-top. A.2. Day 1 Light curve features A.2.1. Long term history For a description of how …
Figure 23
Figure 23. Figure 23: Long term history (from -100 days with respect to first alert) features, from left to right: The mean and standard deviation of the number of non detections between each detection, and the standard deviation of the magnitude values of these detections. We plot separat…
Figure 24
Figure 24. Figure 24: Short term history features. We plot separately the labels given by human scanners and show the Auto-garbage label distribution in grey over-top. A.3. Day 1 features of AT2024lwd We only show the features that are most out of distribution for a “Good” alert [PITH_FUL…
Figure 25
Figure 25. Figure 25: The five day 1 features of AT2024lwd whose values are anomalous for a “Good” object, shown on top of the distributions for these features split by alert types. The grey distribution is the “Auto-Garbage” which is superimposed over the human-vetted labels [PITH_FULL_I…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 14 canonical work pages

  1. [1]

    Alwosheel, A., van Cranenburgh, S., & Chorus, C. G. 2018, Journal of Choice Modelling, 28, 167, doi: https://doi.org/10.1016/j.jocm.2018.07.002

    doi:10.1016/j.jocm.2018.07.002 2018

  2. [2]

    C., Kulkarni, S

    Bellm, E. C., Kulkarni, S. R., Graham, M. J., et al. 2019, PASP, 131, 018002, doi: 10.1088/1538-3873/aaecbe

    doi:10.1088/1538-3873/aaecbe 2019

  3. [3]

    2013, in ECML PKDD Workshop: Languages for Data Mining and Machine Learning, 108–122

    Buitinck, L., Louppe, G., Blondel, M., et al. 2013, in ECML PKDD Workshop: Languages for Data Mining and Machine Learning, 108–122

    work page 2013

  4. [4]

    B., Kulkarni, S

    Cenko, S. B., Kulkarni, S. R., Horesh, A., et al. 2013, ApJ, 769, 130, doi: 10.1088/0004-637X/769/2/130

    doi:10.1088/0004-637x/769/2/130 2013

  5. [5]

    C., Magnier, E

    Chambers, K. C., Magnier, E. A., Metcalfe, N., et al. 2016, ArXiv e-prints. https://arxiv.org/abs/1612.05560 11 It also makes the modelsbiasedtowards the fast recovery of extra- galactic transients, at the expense of galactic transients, but that is a design choice that is known and reported. 12 and since astronomers are not primarily AI professionals wea...

    arXiv 2016

  6. [6]

    A., Modjaz, M., et al

    Crawford, A., Pritchard, T. A., Modjaz, M., et al. 2025, arXiv e-prints, arXiv:2503.03735, doi: 10.48550/arXiv.2503.03735 Dom ´ ınguez S´ anchez, H., Huertas-Company, M., Bernardi, M., et al. 2019, MNRAS, 484, 93, doi: 10.1093/mnras/sty3497

    arXiv doi:10.48550/arxiv.2503.03735 2025

  7. [7]

    J., Ackley, K., Jim´ enez-Ibarra, F., et al

    Dyer, M. J., Ackley, K., Jim´ enez-Ibarra, F., et al. 2024, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 13094, Ground-based and Airborne Telescopes X, ed. H. K. Marshall, J. Spyromilio, & T. Usuda, 130941X, doi: 10.1117/12.3018305

    doi:10.1117/12.3018305 2024

  8. [8]

    A., Piascik, A

    Erasmus, N., Steele, I. A., Piascik, A. S., et al. 2024a, Journal of Astronomical Telescopes, Instruments, and Systems, 10, 025005, doi: 10.1117/1.JATIS.10.2.025005 30

    doi:10.1117/1.jatis.10.2.025005

Show all 60 references
  1. [9]

    B., van Gend, C

    Erasmus, N., Potter, S. B., van Gend, C. H. D. R., et al. 2024b, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 13096, Ground-based and Airborne Instrumentation for Astronomy X, ed. J. J. Bryant, K. Motohara, & J. R. D. Vernet, 130968K, do...

    doi

  2. [10]

    Friedman, J. H. 2001, The Annals of Statistics, 29, 1189 , doi: 10.1214/aos/1013203451

    2001

  3. [11]

    I., & Aleo, P

    Gagliano, A., Contardo, G., Foreman-Mackey, D., Malz, A. I., & Aleo, P. D. 2023, ApJ, 954, 6, doi: 10.3847/1538-4357/ace326

    2023 doi

  4. [12]

    2019, ARA&A, 57, 305, doi: 10.1146/annurev-astro-081817-051819

    Gal-Yam, A. 2019, ARA&A, 57, 305, doi: 10.1146/annurev-astro-081817-051819

    2019 doi

  5. [13]

    2021, in American Astronomical Society Meeting Abstracts, Vol

    Gal-Yam, A. 2021, in American Astronomical Society Meeting Abstracts, Vol. 237, American Astronomical Society Meeting Abstracts, 423.05

    2021

  6. [14]

    2021, ARA&A, 59, 21, doi: 10.1146/annurev-astro-111720-030029

    Gezari, S. 2021, ARA&A, 59, 21, doi: 10.1146/annurev-astro-111720-030029

    2021 doi

  7. [15]

    H., Rhodes, L., Srivastav, S., et al

    Gillanders, J. H., Rhodes, L., Srivastav, S., et al. 2024, ApJL, 969, L14, doi: 10.3847/2041-8213/ad55cd

    2024 doi

  8. [16]

    2018, The Journal of Open Source Software, 3, 695, doi: 10.21105/joss.00695

    Green, G. 2018, The Journal of Open Source Software, 3, 695, doi: 10.21105/joss.00695

    2018 doi

  9. [17]

    J., Bloemen, S., Vreeswijk, P

    Groot, P. J., Bloemen, S., Vreeswijk, P. M., et al. 2024, PASP, 136, 115003, doi: 10.1088/1538-3873/ad8b6a

    2024 doi

  10. [18]

    2025, arXiv e-prints, arXiv:2502.18558, doi: 10.48550/arXiv.2502.18558

    Gupta, R., & Muthukrishna, D. 2025, arXiv e-prints, arXiv:2502.18558, doi: 10.48550/arXiv.2502.18558

    2025 doi arXiv

  11. [19]

    R., Millman, K

    Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357, doi: 10.1038/s41586-020-2649-2

    2020 doi

  12. [20]

    N., Denneau, L., Tonry, J

    Heinze, A. N., Denneau, L., Tonry, J. L., et al. 2021, PSJ, 2, 12, doi: 10.3847/PSJ/abd325

    2021 doi

  13. [21]

    Hunter, J. D. 2007, Computing in Science & Engineering, 9, 90, doi: 10.1109/MCSE.2007.55

    2007 doi

  14. [22]

    2023, arXiv e-prints, arXiv:2310.12528, doi: 10.48550/arXiv.2310.12528

    Huppenkothen, D., Ntampaka, M., Ho, M., et al. 2023, arXiv e-prints, arXiv:2310.12528, doi: 10.48550/arXiv.2310.12528

    2023 doi arXiv

  15. [23]

    E., et al

    Kaiser, N., Aussel, H., Burke, B. E., et al. 2002, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 4836, Survey and Other Telescope Technologies and Discoveries, ed. J. A. Tyson & S. Wolff, 154–164, doi: 10.1117/12.457365

    2002 doi

  16. [24]

    2017, in Advances in Neural Information Processing Systems, ed

    Ke, G., Meng, Q., Finley, T., et al. 2017, in Advances in Neural Information Processing Systems, ed. I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, & R. Garnett, Vol. 30 (Curran

    2017

  17. [25]

    https://proceedings.neurips.cc/paper files/paper/2017/ file/6449f44a102fde848669bdd9eb6b76fa-Paper.pdf

    Associates, Inc.). https://proceedings.neurips.cc/paper files/paper/2017/ file/6449f44a102fde848669bdd9eb6b76fa-Paper.pdf

    2017

  18. [26]

    L., Lyman, J., Steeghs, D., et al

    Killestein, T. L., Lyman, J., Steeghs, D., et al. 2021, MNRAS, 503, 4838, doi: 10.1093/mnras/stab633

    2021 doi

  19. [27]

    L., Kelsey, L., Wickens, E., et al

    Killestein, T. L., Kelsey, L., Wickens, E., et al. 2024, MNRAS, 533, 2113, doi: 10.1093/mnras/stae1817

    2024 doi

  20. [28]

    M., Kulkarni, S

    Law, N. M., Kulkarni, S. R., Dekany, R. G., et al. 2009, PASP, 121, 1395, doi: 10.1086/648598

    2009 doi

  21. [29]

    2023, Universe, 10, 10, doi: 10.3390/universe10010010

    Liang, R., Liu, Z., Lei, L., & Zhao, W. 2023, Universe, 10, 10, doi: 10.3390/universe10010010

    2023 doi

  22. [30]

    2023, arXiv e-prints, arXiv:2305.02997, doi: 10.48550/arXiv.2305.02997 M¨ oller, A., Peloton, J., Ishida, E

    McElfresh, D., Khandagale, S., Valverde, J., et al. 2023, arXiv e-prints, arXiv:2305.02997, doi: 10.48550/arXiv.2305.02997 M¨ oller, A., Peloton, J., Ishida, E. E. O., et al. 2021, MNRAS, 501, 3272, doi: 10.1093/mnras/staa3602

    2023 doi arXiv

  23. [31]

    2025, arXiv e-prints, arXiv:2502.20479, doi: 10.48550/arXiv.2502.20479

    Moreno-Cartagena, D., Protopapas, P., Cabrera-Vives, G., et al. 2025, arXiv e-prints, arXiv:2502.20479, doi: 10.48550/arXiv.2502.20479

    2025 doi

  24. [32]

    S., Biswas, R., & Hloˇ zek, R

    Muthukrishna, D., Narayan, G., Mandel, K. S., Biswas, R., & Hloˇ zek, R. 2019, PASP, 131, 118002, doi: 10.1088/1538-3873/ab1609 ONS-UK. 2022, Office of National Statistics, https: //www.ons.gov.uk/peoplepopulationandcommunity/ personalandhouseholdfinances/incomeandwealth/ bull...

    2019 doi

  25. [33]

    2019, A&A, 627, A21, doi: 10.1051/0004-6361/201834473

    Pasquet, J., Pasquet, J., Chaumont, M., & Fouchez, D. 2019, A&A, 627, A21, doi: 10.1051/0004-6361/201834473

    2019 doi

  26. [34]

    2011, Journal of Machine Learning Research, 12, 2825

    Pedregosa, F., Varoquaux, G., Gramfort, A., et al. 2011, Journal of Machine Learning Research, 12, 2825

    2011

  27. [35]

    A., Fremling, C., Sollerman, J., et al

    Perley, D. A., Fremling, C., Sollerman, J., et al. 2020, ApJ, 904, 35, doi: 10.3847/1538-4357/abbd98

    2020 doi

  28. [36]

    A., Ho, A

    Perley, D. A., Ho, A. Y. Q., Fausnaugh, M., et al. 2025, MNRAS, 537, 2362, doi: 10.1093/mnras/staf125

    2025 doi

  29. [37]

    B., Erasmus, N., van Gend, C

    Potter, S. B., Erasmus, N., van Gend, C. H. D. R., et al. 2024, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 13098, Observatory Operations: Strategies, Processes, and Systems X, ed. C. R. Benn, A. Chrysostomou, & L. J. Storrie-Lombardi, ...

    2024 doi

  30. [38]

    J., Maguire, K., Smartt, S

    Prentice, S. J., Maguire, K., Smartt, S. J., et al. 2018, ApJL, 865, L3, doi: 10.3847/2041-8213/aadd90 Radhakrishnan Santhakumari, K. K., Battaini, F., Di

    2018 doi

  31. [39]

    2024, arXiv e-prints, arXiv:2407.17288, doi: 10.48550/arXiv.2407.17288

    Filippo, S., et al. 2024, arXiv e-prints, arXiv:2407.17288, doi: 10.48550/arXiv.2407.17288

    2024 doi arXiv

  32. [40]

    A., Jegou Du Laz, T., et al

    Rehemtulla, N., Miller, A. A., Jegou Du Laz, T., et al. 2024, ApJ, 972, 7, doi: 10.3847/1538-4357/ad5666

    2024 doi

  33. [41]

    2024, Nature Methods, 21, doi: doi:10.1038/s41592-023-02150-0

    Reinke, A., & et al. 2024, Nature Methods, 21, doi: doi:10.1038/s41592-023-02150-0

    2024 doi

  34. [42]

    J., Finkbeiner, D

    Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, ApJ, 500, 525, doi: 10.1086/305772 31

    1998 doi

  35. [43]

    J., Prieto, J

    Shappee, B. J., Prieto, J. L., Grupe, D., et al. 2014, ApJ, 788, 48, doi: 10.1088/0004-637X/788/1/48

    2014 doi

  36. [44]

    W., et al

    Sheng, X., Nicholl, M., Smith, K. W., et al. 2024, MNRAS, 531, 2474, doi: 10.1093/mnras/stae1253

    2024 doi

  37. [45]

    J., Valenti, S., Fraser, M., et al

    Smartt, S. J., Valenti, S., Fraser, M., et al. 2015, A&A, 579, A40, doi: 10.1051/0004-6361/201425237

    2015 doi

  38. [46]

    J., Clark, P., Smith, K

    Smartt, S. J., Clark, P., Smith, K. W., et al. 2018, The Astronomer’s Telegram, 11727, 1

    2018

  39. [47]

    W., Smartt, S

    Smith, K. W., Smartt, S. J., Young, D. R., et al. 2020, PASP, 132, 085002, doi: 10.1088/1538-3873/ab936e

    2020 doi

  40. [48]

    J., et al

    Stalder, B., Tonry, J., Smartt, S. J., et al. 2017, ApJ, 850, 149, doi: 10.3847/1538-4357/aa95c1

    2017 doi

  41. [49]

    K., Ackley, K., et al

    Steeghs, D., Galloway, D. K., Ackley, K., et al. 2022, MNRAS, 511, 2405, doi: 10.1093/mnras/stac013

    2022 doi

  42. [50]

    A., Smith, R

    Steele, I. A., Smith, R. J., Rees, P. C., et al. 2004, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 5489, Ground-based Telescopes, ed. J. M. Oschmann, Jr., 679–692, doi: 10.1117/12.551456

    2004 doi

  43. [51]

    2025a, ATLAS VRA Technical Manual, Zenodo, doi: 10.5281/zenodo.14944208 —

    Stevance, H. 2025a, ATLAS VRA Technical Manual, Zenodo, doi: 10.5281/zenodo.14944208 —. 2025b, ATLAS VRA v1 - Training Data and Code, Zenodo, doi: 10.5281/zenodo.15195392

    doi

  44. [52]

    2025, HeloiseS/atlasvras: VRA 1.1, v1.1, Zenodo, doi: 10.5281/zenodo.14363396

    Stevance, H., & Smith, K. 2025, HeloiseS/atlasvras: VRA 1.1, v1.1, Zenodo, doi: 10.5281/zenodo.14363396

    2025 doi

  45. [53]

    F., Leland, J., & Smith, K

    Stevance, H. F., Leland, J., & Smith, K. W. 2025, arXiv e-prints, arXiv:2506.06403. https://arxiv.org/abs/2506.06403 The PLAsTiCC team, Allam, Jr., T., Bahmanyar, A., et al. 2018, arXiv e-prints, arXiv:1810.00001, doi: 10.48550/arXiv.1810.00001

    2025 arXiv

  46. [54]

    L., Denneau, L., Heinze, A

    Tonry, J. L., Denneau, L., Heinze, A. N., et al. 2018, PASP, 130, 064505, doi: 10.1088/1538-3873/aabadf

    2018 doi

  47. [55]

    2011, A&A, 536, A105, doi: 10.1051/0004-6361/201117752 Wes McKinney

    Vernet, J., Dekker, H., D’Odorico, S., et al. 2011, A&A, 536, A105, doi: 10.1051/0004-6361/201117752 Wes McKinney. 2010, in Proceedings of the 9th Python in Science Conference, ed. St´ efan van der Walt & Jarrod Millman, 56 – 61, doi: 10.25080/Majora-92bf1922-00a

    2011 doi

  48. [56]

    Stevance, H. F. 2024, RAS Techniques and Instruments, 3, 385, doi: 10.1093/rasti/rzae027

    2024 doi

  49. [57]

    D., Francis, G

    Williams, R. D., Francis, G. P., Lawrence, A., et al. 2024a, RAS Techniques and Instruments, 3, 362, doi: 10.1093/rasti/rzae024 —. 2024b, RAS Techniques and Instruments, 3, 362, doi: 10.1093/rasti/rzae024

    doi

  50. [58]

    L., O’Connor, J

    Worters, H. L., O’Connor, J. E., Carter, D. B., et al. 2016, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 9908, Ground-based and Airborne Instrumentation for Astronomy VI, ed. C. J

    2016

  51. [59]

    Simard, & H

    Evans, L. Simard, & H. Takami, 99083Y, doi: 10.1117/12.2231636

    doi

  52. [60]

    Young, D. R. 2023, Sherlock. Contextual classification of astronomical transient sources, doi: 10.5281/zenodo.8038057

    2023 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.