MNRAS 533, 2113–2132 (2024) https://doi.org/10.1093/mnras/stae1817 Advance Access publication 2024 August 27 Kilonova Seekers : the GOTO project for real-time citizen science in time-domain astrophysics T. L. Killestein , 1 , 2 ‹† L. Kelsey , 3 ‹† E. Wickens, 3 L. Nuttall, 3 J. Lyman , 2 C. Krawczyk , 3 K. Ackley, 2 M. J. Dyer , 4 F. Jim ´enez-Ibarra, 5 K. Ulaczyk, 2 D. O’Neill, 2 A. Kumar, 2 D. Steeghs , 2 D. K. Galloway, 5 V. S. Dhillon , 4 , 6 P. O’Brien, 7 G. Ramsay , 8 K. Noysena, 9 R. Kotak, 1 R. P. Breton , 10 E. Pall ´e, 6 , 11 D. Pollacco, 2 S. Awiphan , 9 S. Belkin , 5 P. Chote, 2 P. Clark, 3 D. Coppejans, 2 C. Duffy , 8 R. Eyles-Ferris , 6 B. Godson, 2 B. Gompertz , 12 O. Graur , 3 , 13 P. Irawati, 9 D. Jarvis, 4 Y. Julakanti, 7 M. R. Kennedy , 14 H. Kuncarayakti, 1 A. Levan, 15 S. Littlefair, 4 M. Magee , 2 S. Mandhai, 10 D. Mata S ´anchez, 6 , 11 S. Mattila, 1 , 16 J. McCormac, 2 J. Mullaney, 4 J. Munday, 2 M. Patel , 6 M. Pursiainen, 2 J. Rana, 6 , 11 U. Sawangwit, 9 E. Stanway , 2 R. Starling , 7 B. Warwick 2 and K. Wiersema 17 Affiliations are listed at the end of the paper Accepted 2024 July 23. Received 2024 July 23; in original form 2024 June 4 A B S T R A C T Time-domain astrophysics continues to grow rapidly, with the inception of new surv e ys drastically increasing data volumes. Democratized, distributed approaches to training sets for machine learning classifiers are crucial to make the most of this torrent of disco v ery – with citizen science approaches pro ving ef fecti ve at meeting these requirements. In this paper, we describe the creation of and the initial results from the Kilonova Seekers citizen science project, built to find transient phenomena from the GOTO telescopes in near real-time. Kilonova Seekers launched in 2023 July and received over 600 000 classifications from approximately 2000 volunteers o v er the course of the LIGO-Virgo-KAGRA O4a observing run. During this time, the project has yielded 20 disco v eries, generated a ‘gold-standard’ training set of 17 682 detections for augmenting deep-learned classifiers, and measured the performance and biases of Zooniverse volunteers on real-bogus classification. This project will continue throughout the lifetime of GOTO, pushing candidates at ever-greater cadence, and directly facilitate the next-generation classification algorithms currently in development. Key words: techniques: miscellaneous – surv e ys – supernovae: general. 1 I t n d t e A 2 K O 2 l f  † i 2 i n o a p t S c i fl a a © P C p D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024 I N T RO D U C T I O N n the current era of time-domain astronomy, we are operating close to he limit of human validation of transient phenomena due to the vast umbers of observations being taken on a daily basis. The e xpansiv e ata volumes (TB per night) of current all-sk y surv e ys such as he Gra vitational-wa v e Optical Transient Observ er (GOTO; Stee ghs t al. 2022 ), Zwicky Transient Facility (ZTF; Bellm et al. 2019 ), steroid Terrestrial-impact Last Alert System (ATLAS; Tonry et al. 018 ), and All-Sky Automated Survey for Supernovae (ASAS-SN; ochanek et al. 2017 ), and the impending era of the Vera C. Rubin bservatory’s Le gac y Surv e y of Space and Time (LSST; Iv ezi ´c et al. 019 ) highlight the continuing need for no v el, automated, machine- earned approaches of source classification in order to triage and ollow-up candidates in a timely manner. E-mail: thomas.killestein@utu.fi (TLK); lisa.kelsey@port.ac.uk (LK) Joint first authorship. i s d a 2024 The Author(s). ublished by Oxford University Press on behalf of Royal Astronomical Society. Th ommons Attribution License ( https:// creativecommons.org/ licenses/ by/ 4.0/ ), whic rovided the original work is properly cited. Modern transient disco v ery is predominantly based on difference maging (e.g. Alard & Lupton 1998 ; Zackay, Ofek & Gal-Yam 016 ). In this technique, ‘template’, ‘reference’, or ‘background’ mages are subtracted from new ‘science’ images in order to remo v e on-varying sources from the image. These reference images are f the same part of the sky as the science image, but were taken t a prior time during the optimal sky conditions (dark moon hases, good seeing). Typically they are also of longer exposure than he science images, meaning that fainter sources can be detected. ubtracting the reference image from the new science image, after orrecting for differential background and PSF mis-matches, results n a ‘difference’ image. This difference image may contain residual ux indicating that something has changed between the reference nd science images – a potential transient or variable source has ppeared. The photometry can then be extracted from the difference mage, to measure positions and fluxes free of contamination from urrounding sources (e.g. Wozniak et al. 2002 ) or host galaxy light. The majority of detections (referred to as candidates herein) in ifference images detected via source extraction are artefacts, known s ‘bogus’ sources following the real-bogus paradigm introduced is is an Open Access article distributed under the terms of the Creative h permits unrestricted reuse, distribution, and reproduction in any medium, 2114 T. L. Killestein and L. Kelsey et al. M i s a t ( M V e N c ( a e d d q s c l m r v t p Z e L e i t h a o a p e o d 1 fl b s t r p e d o ( e s o o t 1 2 a t S O w i f o a K a m i a o r o c s p A S 2 T G t g b o ‘ t w o a T a 2 h e v e e 2 i o u G n i c c fi t D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024n Bloom et al. ( 2012 ). These artefacts broadly arise from bright tar residuals, point-spread-function (PSF) mis-match, and/or mis- lignment. A vast literature has emerged to tackle this challenge – ransitioning from traditional machine learning (ML) approaches Bailey et al. 2007 ; Goldstein et al. 2015 ; Wright et al. 2015 ; ong et al. 2020 ), through to deep learned classifiers (Cabrera- ives et al. 2017 ; Duev et al. 2019 ; Killestein et al. 2021 ; Corbett t al. 2023 ; Mong et al. 2023 ) – with ever increasing performance. aturally ho we v er, as surv e ys grow larger, more performant source lassification algorithms are required to ensure that the number of inevitable) false positives do not overwhelm human vetters. To chieve this goal, larger and larger data volumes are required to f fecti vely train these algorithms, and fully sample the diversity of etections seen in surv e y data. As surv e ys get bigger, the method for ealing with these data volumes needs to impro v e. Such surv e ys uickly outstrip the capacity of individuals or small teams of cientists to ef fecti vely label. A complementary approach, which an be used to create a human-labelled data set for training machine- earning based classifiers, is to use citizen science. Citizen science enables collaboration between researchers and embers of the public, by engaging the public to participate in esearch tasks and help make scientific disco v eries. F or tasks such as etting of candidate transients, the person-power increase of opening his task up to the public is highly significant. Transient astronomy rojects on the Zooniverse citizen science platform 1 such as Galaxy oo Supernovae (Smith et al. 2011 ) and Supernova Hunters (Wright t al. 2017 ), using data from the Palomar Transient Factory (PTF; aw et al. 2009 ; Rau et al. 2009 ) and Pan-STARRS1 (Chambers t al. 2016 ) respectiv ely, hav e had great success involving the public n this way. In both cases, volunteers were provided with a set of arget, reference, and difference images for a candidate transient that ad been flagged as interesting by a computer algorithm, and were sked a simple question to determine if the observation was real r bogus. This facilitates disco v ery of transient events, and creates binary-labelled training set for ML algorithms to augment their erformance in future iterations. Alongside the direct benefits for scientific analysis, citizen sci- nce provides an excellent opportunity for public engagement and utreach by enabling members of the public to help in key scientific isco v ery, and to achiev e e xperiential learning (Bruner 1961 ; Kolb 984 ). The Zooniverse platform was originally created for the agship Galaxy Zoo project (Lintott et al. 2008 ), and has since ecome the predominant online platform for facilitating citizen cience (Marshall, Lintott & Fletcher 2015 ). At the time of writing, he Zooniverse platform has 91 active projects on offer, with topics anging from history, language, and literature to climate, nature, hysics, and space; meaning that there is something of interest for veryone. Citizen science approaches have led to tangible scientific isco v eries: In astronomy, the Galaxy Zoo project led to the disco v ery f ‘green pea’ galaxies, a new class of compact, star-forming galaxies Cardamone et al. 2009 ). Similarly, the Planet Hunters project nabled the disco v ery of PH1b, the first known planet in a quadruple tar system (Schwamb et al. 2013 ). We hav e dev eloped the Kilono va Seek er s citizen science project 2 n the Zooniv erse platform, pro viding an opportunity for members f the public to help the GOTO collaboration in the disco v ery of ransient events that may have been otherwise missed or overlooked,NRAS 533, 2113–2132 (2024) https:// www.zooniverse.org/ http:// kilonova-seekers.org/ o e w u a nd enabling them to participate in cutting-edge science in near real- ime. In this paper, we report findings from the launch of Kilonova eek er s on 2023 July 11, o v er a ∼ 6 month period until the end of the 4a observing run of the LIGO-Virgo-KAGRA (LVK) gravitational- ave detectors, on 2024 January 16. As the primary aim of GOTO s to follow up gra vitational-wa ve alerts from LVK, the timeframes or Kilono va Seek er s are strongly driven by the schedules of these bserving windows. In Section 2 we begin by introducing GOTO nd the need for a citizen science project. In Section 3 we discuss the ilono va Seek er s project in terms of the data used, the workflow, nd interface the volunteers interact with, the behind-the-scenes achinery, and the alerting and reporting mechanisms. We present n Section 4 statistics about volunteer classifications, demographics, nd engagement, with a particular focus on the valuable contribution f our ‘power users’. In Section 5 we highlight the key scientific esults and disco v eries from the project, the o v erall performance f volunteers, and measure the selection function of the volunteers ompared to the GOTO real-bogus classifier. Finally in Section 6 we ummarize the project so far and our key findings, noting our future lans for the project throughout the lifetime of the GOTO surv e y. full list of the citizen scientists who were involved with Kilonova eek er s can be found in Appendix A . T H E G R A VI TATI ONA L-WA V E O P T I C A L RANSI ENT OBSERV ER ( G OTO ) OTO (Dyer et al. 2022 ; Steeghs et al. 2022 ) is a multisite, wide-field elescope array designed to observe electromagnetic counterparts to ra vitational wa v e ev ents – specifically the afterglow of compact inary mergers involving a neutron star, known as kilonovae. GOTO perates in two distinct observing modes: ‘triggered follow-up’ and all-sk y surv e y’ (see Dyer et al. 2020 ), to rapidly target and tile o v er he regions associated with incoming alerts, such as gravitational- ave alerts from the LIGO-Virgo-KAGRA (LVK) detectors. While ther transients, such as supernovae, take a few weeks on a ver - ge to reach their optical peak brightness (Anderson et al. 2014 ; aubenberger 2017 ; Perley et al. 2020 ), kilonovae peak around 1 d fter merger (e.g. Li & Paczy ´nski 1998 ; Kasen, Badnell & Barnes 013 ; Arcavi et al. 2017 ). Surv e ys optimized to find kilonovae must ave quick responses to alert triggers, fast surv e y cadence, and fficient transient identification methods. GOTO’s o v erall field of iew is larger than the localization skymap of GW 170 817 (Abbott t al. 2017 ), the only gravitational wave (GW) event with a detected lectromagnetic (EM) counterpart, and can co v er the whole sky in –3 d – so is ideally suited for these types of searches. Due to a combination of the large sky coverage and fast cadence n all-sky survey mode, GOTO collects and generates large volumes f data (500 GB/24 h raw, 2–5 TB/24 h dataproducts) that make nfiltered human vetting challenging. To address these data volumes, OTO uses a real-bogus classifier ( GOTORB ) based on a convolutional eural network (CNN) to classify candidate transients in difference maging (for more information, see Killestein et al. 2021 ). Each lassification is given a probability of being real, and an associated onfidence level between 0 and 1. This classifier is effective at ltering out bogus detections, with a 97 per cent reco v ery rate of real ransients for a fixed false positive rate of 1 per cent. As seen with ther citizen science projects such as Supernova Hunters (Wright t al. 2017 ), CNNs and human classifiers have different strengths, hich when combined can make a more efficient process than only sing one. CNNs are very good at processing large volumes of data, nd human classifiers perform better than CNNs when the image is Kilono va Seek er s 2115 m i 3 G h b i a b w f i t n a 9 g o t i m 2 J w ( n e w e u t t m i A d f t d c i f t t 3 K e e n p d s a ( l i t ( u i e b n d s t L a A t f n b c t K i p o p a s t r t m r c s a a c o i b u w p t s f f o r a w o D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024ore ambiguous, and when there are not many examples to compare t to. T H E CITIZEN SCIENCE PLATFORM iven the significant volumes of detections generated, only the ighest scoring candidates from a gra vitational-wa ve follow-up can e prioritized for eyeballing by the GOTO collaboration. By the mperfect nature of classification algorithms, a number of false neg- tives will al w ays exist below the chosen score threshold, potentially eing astrophysically interesting. By lowering the score threshold, e can impro v e reco v ery rates, although naturally with increased alse positives. Beyond the real-time necessity for fast transient searches, increas- ng the possible size of human-labelled data sets is important for raining impro v ed classification algorithms. The presence of label oise (inaccurate labelling, see e.g. Fr ´enay & Verleysen 2013 ) is strong limiting factor in pushing accuracies from 99 per cent to 9.9 per cent (and beyond) and can likely only be mitigated via rouping of labels, weighting by quality of data item, and clipping f bad or unrepresentative examples. Citizen science provides a methodology to scale data labelling asks from small teams of expert scientists, up to thousands of ndividuals. Calibrated uncertainty quantification is also a crucial issing link in many current astronomical classifiers (e.g. Abdar et al. 020 ). Although strides with Bayesian neural networks (e.g. Valentin ospin et al. 2020 ) have neatly quantified uncertainties associated ith choice of model, this often does not represent the uncertainty or confidence) a human would assign to their prediction. The true ature of uncertainties in ML is a complex issue, ho we ver, nominal stimates are useful in active learning (where models may suggest hich data are most informative to be labelled by a human, e.g. Ren t al. 2020 ), anomaly detection, and decision making rules under ncertainty. Given these challenges, a citizen science approach is well-suited o generating the scale (and quality) of labelled data sets required o train impro v ed classifiers, and drive searches for candidates that ay otherwise be missed in real-time. Kilonova Seekers launched n 2023 July, after a short beta-testing period with live volunteers. t its core, Kilonova Seekers streams uncurated difference image etections (referred to as ‘candidates’ herein) meeting certain cuts rom the GOTO real-time pipeline (see Lyman et al., in preparation) o the Zooniverse platform, populating a workflow with pre-baked ata visualizations (known as subjects) to receive annotations and lassification from citizen scientist volunteers. Through custom nfrastructure (see Section 3.2 ), we listen to the classification stream rom Zooniverse in real-time, and use this to trigger alerts according o set rules on consensus. We elaborate further on the specifics of his process in the following sections. .1 Data extraction, pr e-pr ocessing, and presentation ilono va Seek er s ingests candidates as part of a scheduled task – x ecuted on a daily cadence during project launch, and increased to very three hours during the O4a observing run. Given the multisite ature of GOTO, this leads to eight uploads of data per day (weather- ermitting). A candidate corresponds to a single difference image etection – analogous to the concept of alerts in other transient urv e ys. F or logistical reasons, Kilonova Seekers does not take into ccount multiple candidates at the same location being associated i.e. operating at a source level) – which would require more complex ogic to de-duplicate candidates, adding additional o v erhead. This isntentionally decoupled from how candidates are handled internally, o provide an independent dataflow. The numbers of real transients and artefacts are heavily imbalanced Bloom et al. 2012 ), thus we sample difference image detections niformly in their real-bogus score (with values between 0 and 1 nclusive, see Killestein et al. 2021 ) through a process of histogram qualization – selecting a uniform number of candidates per real- ogus bin, with typical equal bin-size of 0.1. Although these choices ecessarily bias the data set generated, there still exists sufficient iversity to re-balance (and thus train classifiers on) the final data et. Candidates are queried from the main difference photometry able generated by GOTO’s KADMILOS data processing pipeline (see yman et al., in preparation), up to a user-specified maximum to v oid flooding v olunteers with candidates in the case of rich fields. number of operational considerations drive the exact query used o ingest candidates – with our selection cuts being: (i) Signal-to-noise greater than 10: to minimize the number of alse alarm detections due to correlated noise in the initial stages. (ii) Avoidance of the Galactic plane ( | b| < 10 ◦): to minimize the umber of variable sources being uploaded to Kilono va Seek er s – oth for practical rate-limiting purposes, as well as data set imbalance onsiderations. (iii) Exclusion of specific GOTO unit telescopes (UTs): owing o ongoing hardware issues, one specific UT was disabled in the ilono va Seek er s live w orkflow to minimize impact on volunteers. (iv) Cuts on images with extremely high numbers of difference mage detections: after excluding the plane, these are likely to be oor subtractions which affect class balance. We impose that number f detections in each difference image must be less than the 90th ercentile number of detections across all difference images. (v) Real-bogus score: for the purposes of fast disco v ery, we adopt real-bogus score of 0.7 or greater. This is slightly below the normal core threshold of 0.8 used internally, and corresponds approximately o the equality point between false positive rate and false ne gativ e ate, a common choice in ML contexts. We extract a set of stamps, sized approximately 3 × 3 arcmin, from he science, reference, and difference images, small cutouts of the ain images centred on each candidate detection. The science and eference images are derived from stacked data products, a sigma- lipped combination of a number of individual sub-frames, to reject ingle-image outliers such as cosmic rays. Stamps are extracted t nativ e GOTO pix el scale (1.4 arcsec pix el −1 ). Pix el thresholds re set using the IRAF ZSCALE algorithm (Tody 1986 , 1993 ), per- hannel to span their full range. In a break from the norm of ther transient disco v ery projects on Zooniv erse, we use colourized mages: specifically the MATPLOTLIB ’bone’ colourmap. The tasteful lue shading is intended to minimize visual stress. To generate and pload a subject to Zooniverse, we construct a pre-baked layout that e populate with stamps and metadata for a given candidate. We rominently display the detection time into each stamp, to reinforce he real-time nature of uploads to the volunteers, and write which urv e y each image comes from: to alert volunteers to any images rom gra vitational-wa ve (GW), gamma-ray b urst (GRB), or neutrino ollow-up. The o v erplotted cross-hairs dra w attention to the centre f the frame, and the box shows the field-of-view that the GOTO eal-bogus classifier sees, providing important context. We illustrate subject in Fig. 1 . Early in Gen. 1 Kilonova Seekers , we noticed volunteers o v er- helmingly classifying cosmic rays (CRs) as real detections, in spite f their often non-PSF-like appearance and documentation on the MNRAS 533, 2113–2132 (2024) 2116 T. L. Killestein and L. Kelsey et al. M Figure 1. Example subjects from Kilonova Seekers . The science, reference, and difference images are plotted, along with subframes and event informa- tion. The top layout shows SN2024gy, a Type Ia supernova in the nearby galaxy (13.5 Mpc) NGC 4216 flagged by volunteers. The bottom layout shows a cosmic ray artefact that was flagged by volunteers, visible in only one of the four sub-frames, and unfortuitously projected on top of a galaxy. fi s i s a i t e t f p m i r n v t i c s ( F S 3 4 5 t t s t a e w c o w S i 3 K t s r a s m s u h T o p a 3 A c K a e c l b p c W a c t w t C w w D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024eld guide for these objects. This moti v ated the addition of the ubframes panel for Gen. 2 – in which we display the individual mages that compose the stack – to identify single-frame artefacts uch as CRs that propagate into the stack. These are visible in Fig. 1 , nd are 32 × 32 pixels each, with a faint circle added to aid the user n identifying potential moving targets. Based on feedback from the volunteers, we added labels to show he volunteers which GOTO site the data originates from, and an vent tag to explain which mode GOTO was in when the image was aken. As GOTO is focused on transient follo w-up, dri ven by triggers rom external facilities – the types of images that the volunteers are resented with may change on a daily basis. For example, in survey ode many galaxies may be present in the images, whereas if GOTO s following a specific alert, the telescopes may be pointed towards egions of greater source density, with images being dominated by earby variable stars in our galaxy. To explain this clearly to our olunteers, we use the follo wing e vent labels and provide links to he individual instruments listed here so that they can find more nformation if they are interested in learning more: (i) All-sky survey – GOTO is scanning the sky systemati- ally to find new sources. (ii) LVK alert [alert number] – GOTO is following a pecific gra vitational-wa ve alert from the LIGO-Virgo-KAGRA LVK) detectors, searching for the potential optical counterpart. 3 (iii) Fermi alert – GOTO is following a GRB alert from the ermi Space Telescope . 4 (iv) Swift alert – GOTO is following a GRB alert from the wift Space Telescope . 5 NRAS 533, 2113–2132 (2024) https:// emfollow.docs.ligo.org/ userguide/ https:// fermi.gsfc.nasa.gov/ https:// swift.gsfc.nasa.gov/ S 6 7 8(v) IceCube alert – GOTO is following a neutrino alert from he IceCube detector. 6 (vi) Supplemental survey – GOTO is doing something else hat is not co v ered by the other event tags. Some metadata is deliberately censored from the volunteers, uch as the sky location of each candidate, and exact discovery ime. This is predominantly to prevent volunteers from seeking dditional contextual information outside of the image, that would .g. confirm a given detection is a minor planet and thus real, as ell as for operational reasons to prev ent an y disco v eries being orrelated with GW event skymaps, or reported without scrutiny n TNS or social media channels. This policy will naturally evolve ith workflow requirements, with in-development workflows (see ection 6 ) providing additional (albeit carefully chosen) contextual nformation for classifications. .2 Workflow and ingestion ilono va Seek er s presents one unified workflow to the user, tailored o the real-bogus paradigm for source classification. Subjects are hown to volunteers randomly, from the pool of data that has not eached retirement (when voted upon by 15 volunteers). Volunteers re asked if a real source exists at the centre of the crosshairs in the cience and difference images. Initial beta tests including a fuzzy aybe option showed volunteers o v erwhelmingly (  50 per cent ) elected this option, hindering consensus estimates and making ncertainty estimation impossible. The web workflow is depicted in Fig. 2 . Kilono va Seek er s also as a companion mobile workflo w, deli vered via the Zooniverse app. his has the same layout as the web workflow, but with the addition f an intuitive ‘swipe left and right’ interface familiar from other opular mobile apps. We defer a full discussion of the workflows nd their utilization to Section 4.2 . .3 Alerting and reporting lerts are intended to flag an object for further follow-up once a given andidate (subject) reaches a configurable consensus threshold. For ilono va Seek er s this is set at a threshold of 80 per cent agreement, nd a minimum of eight votes for the majority option set through mpirical testing during beta. The high minimum vote threshold is rucial to a v oid false consensus, where the wrong answer may be ocked in by an early run of votes. This was determined empirically, ut is further moti v ated statistically by ensuring an error of ∼10 er cent in the derived agreement fraction. Alerting to the collaboration is delivered via Slack 7 (the communi- ation platform used by the GOTO collaboration), using the Incoming ebhook API to post an alert card to a dedicated #knseekers- lerts channel for rapid triaging of candidates. One such alert ard is displayed in Fig. 3 – with action links to direct the vetter o the internal GOTO Marshall (see Lyman et al, in preparation), a eb interface for further analysis of transients and reporting, or to he Kilono va Seek er s Talk pages to check discussion on the object. ollecting key information via a collaborative platform provides a ay to centralize discussion about candidates in a maintainable, open ay. Real extragalactic transients are reported to the Transient Name erver (TNS 8 ) through the existing GOTO Marshall architecture. https:// icecube.wisc.edu/ science/ icecube/ https://slack.com https://wis-tns.org Kilono va Seek er s 2117 Figure 2. Screenshot of the live Kilonova Seekers main workflow. T fi r r t i p a 3 T w ( t s H e v v v b e d t i J T h 9 p m p c t d h A t u ∼ W m M o m a r M r t o m e f t D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024o credit volunteers for their work, we append the names of ve randomly selected classifiers of a given transient to the TNS emarks section, subject to integrity checks (see Section 4 ). This andomization occurs at point of consensus, and is done in this way o more fairly assign credit, rather than just the first (who may be n a more fa v ourable time-zone, for e xample). Re gardless of this rompt report, all volunteers who correctly identify a given transient re credited on the project results page. .4 Implementation details o power the real-time nature of Kilono va Seek er s , we developed a eb service to receive classifications from Zooniverse in low-latency typically in ∼s), combine them with contextual information from he GOTO Marshall, and generate alerts for promising transients. We use Zooniverse’s Caesar 9 tool to generate a stream of clas- ifications, pushed into a PostgreSQL database hosted locally via a TTP POST endpoint, exposed on the database machine. The web ndpoints for Kilonova Seekers are write-only by design, delivered ia Apache2 backed by the Python DJANGO framework. Schema alidation via PYDANTIC ensures only POST requests containing alid classifications are ingested, and enforces strong type safety y checking and enforcing that ingested data are of the right type, nhancing reliability. As Zooniverse predominantly use NoSQL atabases internally and make heavy use of free-form JSON data hroughout their APIs, we make no attempt to normalize these at ngest and instead use PostgreSQL’s excellent native support for SON(B) datatypes, despite it being a relational database at heart. his was largely driven by the requirement for the database to ost ingests from multiple projects, including the internal GOTOzoo https:// github.com/ zooniverse/ caesar 1 1roject used for GOTO template v etting. Giv en that different projects ay have different metadata (provided as JSON strings), we create roject-specific database views for each project, to ensure queries an be written in simpler, more user-friendly ways, without having o parse the JSON strings each time. The full Kilonova Seekers atabase and real-time stack is hosted on lo w-po wer commodity ardware, specifically a cloud-hosted Raspberry Pi Model 4B. lthough comparativ ely tin y, we found this hardware performed ably hroughout the first six months of the project with o v er a 99.9 per cent ptime – proving highly capable and handling peak throughputs of 100 classifications per second during the initial launch rush phase. e are currently in the process of migrating Kilono va Seek er s to ore powerful hardware, as we introduce active learning and online L estimators to our workflows, though this is predominantly for perational stability and could easily remain in situ. To provide onitoring of the health of the project, Grafana 10 and Prometheus 11 re used to construct real-time dashboards to visualize the rates, atios of real-bogus, and bulk properties of incoming classifications. etrics such as the daily number of active users and classification ate are crucial for informing ongoing engagement strategies and hus are prominent in the design. We anticipate open-sourcing various aspects of the real-time flows f Kilonova Seekers in the near future, to enable the community to ake use of pre-built utilities for real-time citizen science projects – specially in light of new transient surv e ys coming online in the near uture that aim to deliver citizen science components, for example he Vera C. Rubin Observatory (e.g. Higgs 2023 ). MNRAS 533, 2113–2132 (2024) 0 https:// grafana.com/ 1 https:// prometheus.io/ 2118 T. L. Killestein and L. Kelsey et al. M Figure 3. Alert card for a Kilonova Seekers candidate that has reached consensus, published via Slack. Visible on the alert card are the consensus level for the candidate, links to both internal GOTO webpages and the Kilono va Seek er s discussion forum, and the candidate itself. 4 A t t G m t p t h e 4 K a C p a a ‘ p c a fi 1 i a t 1 1 Figure 4. Cumulative classifications per day on Kilonova Seekers from launch until the end of O4a (2024 January 16). The first (blue) shaded region corresponds to the dates of press releases, and active media coverage of the project during the launch period. The second (red) shaded region towards the end of September shows the maintenance period after three months of operations, when we temporarily paused the scheduled uploads and implemented the Gen. 2 subjects based on feedback from the volunteers. The third (green) shaded region highlights the increase in rate of classifications o v er the winter holiday period and the subsequent return to work. The solid vertical line corresponds to the date of an email newsletter sent out to registered volunteers, leading to a clear increase in classifications. The dashed line is the date we increased the data upload cadence from twice per day to every three hours. r r r v f i e F a p c u i o b d g ( a t e p b c D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024 VO LU N TEERS s a citizen science project, our Zooniverse volunteers are the key to he success of Kilonova Seekers . For us, it is not only important hat the project provide useful classifications for improving the O TO real-bogus classifier , b ut that the v olunteers contrib ute to eaningful scientific disco v ery, engage with our collaboration and he other volunteers, learn from the project, and most crucially, enjoy articipating in the science of GOTO. In this section we discuss the volunteer classifications, highlighting he valuable contribution of our most prolific users (in the top 25, erein power users); before exploring the volunteer demographics, ngagement, and the speed and efficiency of their classifications. .1 Volunteer classifications ilono va Seek er s launched publicly on Zooniverse on 2023 July 11 t 14:30 UTC, achieving 1000 classifications within the first 30 min. oinciding with the project launch, Kilonova Seekers was featured in ress releases from the GOTO partner institutions and social media, nd the Kilonova Seekers leads (T.L.K and L.K) were interviewed bout the project on the radio for BBC Radio Solent 12 and on the Missing Links’ show on Dublin City FM. 13 This period of active ublicity is highlighted in blue in Fig. 4 , where the impact of this an be seen by a steep gradient in the rate of classifications. After the initial launch rush, classifications settled down to an verage of ∼ 4000 classifications per day o v er the course of the rst three months of operations. We consider this time to be ‘Gen. ’ of Kilono va Seek er s . During this time, only GOTO-North was ncluded, and we were operating the Kilono va Seek er s project with once-per-day upload cadence, along with the Gen. 1 image styleNRAS 533, 2113–2132 (2024) hat did not contain the subframes for easier detection of cosmic 2 https:// www.bbc.co.uk/ sounds/ play/ live:bbc radio solent 3 https:// www.dublincityfm.ie/ shows/ missing-links/ c o o t ays (as discussed in Section 3.1 ). As illustrated in Fig. 4 by the ed shaded region, we paused the scheduled uploads for a week to apidly implement the Gen. 2 subjects based on feedback from the olunteers, and to upgrade the behind-the-scenes infrastructure ready or ingesting subjects from GOTO-South and the planned increase n upload cadence. We announced our new Gen. 2 subjects in an mail newsletter once the maintenance was complete, as indicated in ig. 4 by a solid red line. Classifications quickly increased again to n average of ∼ 3100 classifications per day after this maintenance eriod. GOTO-South at Siding Spring Observatory was integrated suc- essfully into our upload pipeline, and we mo v ed to a three-hour pload cadence on 2023 October 11, as indicated by the dashed line n Fig. 4 . Classification rates did slow after this period to an average f ∼ 1700 per day, ho we ver this was largely due to poor weather at oth sites due to the changing seasons, meaning there were fewer ata to upload to the project. A particularly interesting feature of Fig. 4 is highlighted by the reen shaded region. This indicates the Christmas holiday period December 24–Jan 1), when many people are off work for around week. We found a significant increase in classifications during his time, suggesting that our users may have had more free time to ngage with Kilonova Seekers – as evidenced by an increase in Talk osts from many of our users during this period. In total, o v er the course of this initial run of Kilono va Seek er s , etween launch and the end of O4a, our volunteers achieved 643 124 lassifications of 42 936 subjects. By focusing in on the first 100 d post launch, we can compare the lassification curve of Kilonova Seekers (Fig. 5 ) with other projects n the Zooniverse. As discussed in Spiers et al. ( 2019 ), the majority f projects on Zooniverse show high classifications on project launch hat rapidly declines after the initial launch rush. Occasional peaks Kilono va Seek er s 2119 Figure 5. Classifications per day on Kilono va Seek er s for the first 100 d after launch. This distinct classification curve shows that volunteers regularly classify on the project with the release of new data. Figure 6. The distribution in classifications among users from launch until the end of O4a. The median number of classifications is 11; ho we v er, we hav e a strong core user-base, with a number of users completing more than 10 000 classifications each. i c H i i w n i p c w 4 A a f 2 d c Figure 7. Pareto plot of the cumulative fraction of Kilonova Seekers partic- ipants from launch until the end of O4a, plotted against cumulative fraction of classifications. The dashed diagonal line represents perfect parity/equality in classification effort per participant. The Gini index is annotated, providing a quantitative measure of the inequality in contribution. n A ‘ t P t o c G o o s h a a w u b t c w o c w u a w d b f a a e a o w t D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024n activity may be seen after periods of project promotion, press o v erage, or further data release. Other projects such as Supernova unters show a dramatically different classification curve (see fig. 4 n Spiers et al. 2019 ), with more regular spikes in classification ndicative of recurring activity. For Supernova Hunters , these spikes ere on a weekly cadence, resulting from the weekly data upload and ewsletter cadence of the project. Kilono va Seek er s f alls somewhere n-between these two trends. The project shows a clear initial launch eak and rapid decline, with smaller regular spikes in activity, likely orresponding to our regular daily upload cadence (barring any eather restrictions). .1.1 Power users s shown in Fig. 6 , which shows the distribution in classifications mong users, many Kilonova Seekers volunteers only undertake a ew classifications. Similarly to those for Galaxy Zoo (Lintott et al. 008 ) and Bursts from Space: MeerKAT (Andersson et al. 2023 ), the istribution follows a power law, where the majority of volunteers omplete between 1 and 10 classifications on the project, with the umber of volunteers declining for larger numbers of classifications. dditionally, this plot clearly shows the significant impact of our power users’ who have each contributed thousands of classifications o the project. An alternati ve frame work to look at this is via the areto-like (e.g. Lorenz 1905 ; Cowell 2011 ) plot in Fig. 7 , where he cumulative fraction of classifiers, and their cumulative share f the classification effort is depicted. Around 90 per cent of the lassifications are performed by 10 per cent of the volunteers, with a ini index (Gini 1912 ) of 0.9, in line with other Zooniverse projects f a similar nature (e.g. table 3 of Spiers et al. 2019 ). The majority of these power users are the most active participants n the Talk pages, regularly asking questions about the project, haring their experiences, and providing their thoughts and insights to elp others. For the next generation of Kilonova Seekers we anticipate ppointing and training some of these individuals as moderators to id in the day-to-day running of the project. To better understand the classification patterns of the volunteers, e present in Fig. 8 the average daily classifications for the power ser group (the 25 users with the greatest number of classifications etween launch and the end of O4a), displayed in 15 min windows o see trends in volunteer classifications throughout an average day, alculated by dividing the total number of classifications per user per indow by the window length in days. We split this into two based n initial daily upload schedule in Fig. 8 (a) and based on the later hange to upload new data every three hours in Fig. 8 (b). For the 92 d hen we were uploading data every day at 12:00 UT , our most active sers were predominantly doing their classifications immediately fter the daily data upload. Whilst it is encouraging that volunteers ere keen to classify the data immediately, and to be included on the isco v ery reports, these reports were quickly becoming dominated y the same few volunteers, and others were missing out. This gave urther moti v ation to mo v e to a more frequent data upload – alongside more real-time data stream being beneficial for classification speed nd distributing the work more fairly. Uploading data more frequently nables volunteers across different timezones to see the data first: llowing them to participate in disco v ery, and be acknowledged n disco v ery reports. As illustrated in Fig. 8 (b), during the period here the data were uploaded every three hours, whilst the times hat specific volunteers made no classifications remained consistent, MNRAS 533, 2113–2132 (2024) 2120 T. L. Killestein and L. Kelsey et al. M Figure 8. Average classifications over the course of a day for our top 25 users (as defined by the 25 users with the highest number of classifications between launch and the end of O4a), divided into 15 min windows. Each row corresponds to a unique user, in descending order to the total classifications o v er the initial phase of this project, i.e. the top row is the volunteer with the most classifications. t t v w 4 T o d S o c p u f u t u o S v a S b d c s p u t u ( p s t i r t o t o w I o P H E c l 4 T i o d m s t t t G S m a T D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024here were no longer clear times when the most prolific volunteers did he majority of their classifications. In spite of these changes, some olunteers still seem to consistently work non-stop on the project, ith gaps in Fig. 8 (b) likely arising from binning/finite sampling. .2 Volunteer demographics o date, Kilono va Seek er s has attracted roughly 2000 volunteers, in v er 20 distinct time zones, across 105 different countries. Fig. 9 isplays the geographical distribution of volunteers on Kilonova eek er s , shaded according to classifiers per capita. Based on data btained from Google Analytics, we have participants from every ontinent (except Antarctica). The wide accessibility of Zooniverse rojects enables us to reach countries that may be traditionally nderrepresented in astronomical communities. Based on the number of users per country, the United States is by ar the largest contributor to Kilonova Seekers , with a total of 1284 sers. At approximately half this value with a total of 615 users is he United Kingdom. Ho we v er, considering av erage page views per ser for individual countries in the time between launch and the end f O4a, we find that Portugal contains the most prolific Kilonova eek er s , with o v er 2750 views per user on average. Kilono va Seek er s is available to all users who can access the Zooni- erse platform on the internet, which is available to computer, tablet, nd mobile users. Alongside the classic in-browser mode, Kilonova eek er s is available via the Zooniverse mobile app, available on oth iOS and Android devices. The majority of classifications are one via a computer, indicated by Fig. 10 , but roughly a third of lassifications are done via mobile phones (inferred via user agent trings). As displayed in Fig. 11 , the fraction of mobile classifications er user is bimodal, with the vast majority of volunteers either not sing a mobile phone at all or solely using their mobile phone o engage with Kilonova Seekers . Owing to this clear split in our ser-base, it is important that future iterations of Kilonova Seekers and other Zooniverse projects) do not contain too many imagesNRAS 533, 2113–2132 (2024) er page, to ensure continued readability on smaller mobile phone creens. Although the number of classifications specifically done via he mobile app is relatively small compared to those who use an nternet browser (as indicated by the smaller pie chart in Fig. 10 ), it epresents a non-negligible proportion of participants, necessitating hat Kilono va Seek er s remains compatible with the app, regardless f future updates, so that it remains accessible to all users. As GOTO is a global collaboration with members from all across he w orld, it w as important to offer Kilono va Seek er s in the variety f languages that are spoken by the collaboration. At time of riting, Kilono va Seek er s is available in English, Dutch, Spanish, and ndonesian. We were the first project on the Zooniverse platform to ffer Indonesian, and are currently working on the Finnish, Japanese, olish, and Swedish translations, to be released in the near future. o we ver, discussions on the Talk boards predominantly occur in nglish. These localizations are a volunteer effort driven by GOTO ollaboration members, and thus we aim to scale up to support more anguages as capacity/enthusiasm allows. .3 Volunteer engagement he Kilonova Seekers team and the wider GOTO collaboration nteract with the volunteers via the project ‘Talk’ boards, a series f forum pages separated into categories and threads for different iscussions. We encourage the volunteers to discuss subjects that they ay be unsure of on their individual talk pages, and to ask the GOTO cientists questions by creating their own discussion threads. We use his platform as a key page for announcements to the volunteers from he Kilono va Seek er s team, including details about new disco v eries hat the y hav e made and updates about the project or status of the OTO telescopes. Volunteers can ‘@’ members of the Kilonova eek er s team on the Talk pages in the same way as popular social edia platforms to alert them if they have a question or need help, nd can also send pri v ate messages to the team and other volunteers. hrough this, volunteers have told us how they have shared Kilonova Kilono va Seek er s 2121 Figure 9. Geographical distribution of volunteers on the Kilonova Seekers project. The intensity of a given country corresponds to the classifiers per capita, using information from Natural Earth, 14 log-normalized for visualization purposes. Figure 10. Pie charts illustrating the different ways classifications are made on Kilono va Seek er s . The larger pie chart indicates the percentages of classifications during O4a that were completed on computers, mobiles, and tablets. The smaller, nested pie chart indicates the percentage of mobile classifications done via a mobile browser or the Zooniverse mobile app. Figure 11. Distribution of the fraction of the total classifications per user performed on a mobile phone. This takes into account both mobile browser and mobile app classifications. S h o c u t # t v fi w s o fi p t t t s c c r t i c S t r d N v m s D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024eek er s with their families, friends, amateur astronomy groups, and ave discussed the project in blogs and at conferences, widening the v erall participation of the project. On the project Talk pages, volunteers are able to tag their omments. Without any prompting from the team, volunteers started sing very similar or the same hashtags as each other. Most of hese indicate potential transients with tags such as #real or transient , or highlight other astronomically interesting objects hat are not part of the aims of the project e.g. #comet . The olunteers also use these tags to indicate common artefacts from the eld guide, e.g. #badsubtraction and #satellite , along ith artefacts the y hav e encountered from prior similar citizen cience projects, amateur astronomy, and e ven ne w ones of their wn naming, which we have been able to use not only in our regular eld guide updates, but also to update the GOTO hardware team on otential issues. For the next generation of Kilonova Seekers , we plan o implement a new multiclass workflow, and these tags will form he basis for the different labels we will include. Alongside the Talk pages, we engage our volunteers using newslet- ers. These provide an opportunity to update the volunteers on the tatus of the project, announce key findings, inform volunteers of hanges to the project, and generally share our enthusiasm with the itizen scientists. We have found these to be particularly useful for e-engaging volunteers who may have lost interest in the project o v er ime, as can be seen in the upturn in classifications after a newsletter n Fig. 4 . To ensure that volunteers are credited appropriately for their ontributions, disco v eries are reported via a dedicated Kilonova eek er s results page, including the names or usernames of all of he volunteers who marked a candidate as ‘real’. Furthermore, we andomly select a subset of five names from the ‘real’ list to add in a edicated acknowledgement in the remarks section of the Transient ame Server (TNS) page for the object. In order to receive credit, olunteers must be logged into their Zooniverse account when they ake the disco v ery, so that the y can be identified. When volunteers ign up to the Zooniverse platform, they have the option to giveMNRAS 533, 2113–2132 (2024) 2122 T. L. Killestein and L. Kelsey et al. M t w a s 5 I S S d t v t t w c r i d 2 t T c 2 2 T S i 3 K u t p t i l fi 2 t r 5 O i w s K B d a b K p t t s a l h c w v p p h c b p t c b e t t t w d G r t w r b G T s i w o o 5 b O S b i W i f h e d v o u a t n w s t D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024heir real name. If the y hav e chosen to provide this, their real name ill be used for credits, otherwise we use their public username. We utomatically filter out email addresses and web links from these text trings. SCIENTIFIC H I G H L I G H T S n the six months between launch and the end of O4a, the Kilonova eek er s project reported a total of 29 objects to the Transient Name erver, which are listed in Table 1 , where 20 of these were official isco v eries, first made by Kilonova Seekers . At present, the candidates that are flagged as interesting by he volunteers require cross-checking by the GOTO collaboration ia the Slack alert cards (see Section 3.3 ). Real disco v eries are hen reported through the TNS via the GOTO Marshall. Anything hat is a new disco v ery and has not appeared yet on the TNS ith another group is immediately reported, but Kilonova Seekers andidates first identified by other groups are not yet routinely eported owing to limited person-power – something planned to mpro v e via automation in future updates. To date, 6 of the 20 transients first disco v ered by Kilonova Seekers uring O4a have been classified spectroscopically. The first, AT 023rob, was classified as a cataclysmic variable star (CV) by he Spectroscopic Classification of Astronomical Transients (SCAT; ucker et al. 2022 ) surv e y (Hinkle 2023 ). The remaining were all lassified as Type Ia supernovae (Davis, Foley & Jacobson-Galan 023 ; Do 2023 ; Kopsacheili et al. 2023 ; Fremling, Neill & Sharma 024 ) by SCAT, the extended Public ESO Spectroscopic Survey of ransient Objects (ePESSTO + ; Smartt et al. 2015 ), and the Young upernova Experiment (YSE; Jones et al. 2021 ). In total o v er the period discussed in this paper, 1037 spectroscop- cally confirmed supernovae were reported to the TNS, of which 54 subjects associated with these known SNe were generated for ilono va Seek er s , assuming the subjects are associated with SNe sing a narrow 1 arcsec cross-match radius. Of these, 259 reached he consensus threshold of 80 per cent agreement and 8 or more ositive votes. This implies a recovery fraction of 72 per cent across his sample, broadly in line with more in-depth estimates presented n Section 5.2 . A large number of these transients are detected at o w SNR, dri ving the lo wer reco v ery than perhaps anticipated – this gure increases rapidly with SNR, moving to 82 per cent at SNR = 0, 95 per cent at SNR = 50, and 100 per cent at SNR = 70. In he following subsections, we discuss in depth some of these early esults from the Kilonova Seekers project. .1 Rapid reporting ne of the key accomplishments to highlight from Kilonova Seekers s the speed of classification and consensus from the volunteers. As e ha ve v olunteers from around the world, there is almost al w ays omeone online looking at the data in real-time, whether uploaded to ilono va Seek er s (e.g. Fig. 8 ), or internally within the collaboration. etween 2023 September 11 and the end of O4a, we changed the ata upload cadence to the Zooniverse platform to every three hours, nd found that the majority of new subjects uploaded were classified efore the next data upload just three hours later. We display in Fig. 12 the average classification speeds of the ilono va Seek er s volunteers per subject. We clip the maximum time er classification to 2 min to measure the actual attention paid to he classification – there were cases where classifications took on he order of 18 h, which we interpret as situations where a volunteer tepped away from their device and submitted the classification atNRAS 533, 2113–2132 (2024) later time. As shown in Fig. 12 (a), our power users typically take ess time to classify a subject than the remainder of users, who ave a wider range of classification times. Ho we ver, the median lassification time for both groups is roughly 5 s, meaning that if e take our total classifications for the period (see Section 4.1 ), our olunteers ha ve dedicated at least 893 h of classification time to the roject during O4a. In Fig. 12 (b), we break down the power-user classification times er user, and explore the distributions. There are clear differences ere, with some users routinely taking under 10 s for every single lassification they do, whilst others take substantially longer. This ehaviour is unclear, and no conclusive explanation exists. Some ower users may be reading and investigating the metadata for he subjects to find more insights that may help them make a lassification – since these attributes are mentioned on the Talk oards by a small subset of volunteers. The final user on the plot is an xtreme outlier – upon detailed inspection this user’s classification imes show a remarkable bimodality, with a similar ‘early’ peak to he other participants, but with a strong peak around 20 s, skewing heir quartiles on this plot. A particularly significant scientific highlight for Kilonova Seekers as the disco v ery of AT 2023xqy (the Zooniverse subject for this isco v ery is displayed in Fig. 13 ). This object was observed by OTO-South on 2023 No v ember 13 at 11:06:02.592, and was eported to the TNS at 14:27:36 on the same day. It was observed, he data were reduced and uploaded to Zooniverse, the candidate as flagged as interesting, cross-checked and confirmed as real, and eported to the TNS within approximately 3 h and 20 min of data eing taken. This transient had a rapid rise in brightness. The last OTO non-detection was 24 h prior at a L -band magnitude of 20.8. he transient was disco v ered 1 d later at a magnitude of 19.2 – uggesting this object rose in brightness by 1.6 mag per 24 h, and mplying the transient was caught early post-explosion. This finding as later confirmed by ATLAS on 2023 No v ember 17. This speed f human vetting is simply not sustainable without the dedication of ur citizen scientists. .2 Validation data set, detection efficiency, and volunteer enchmarking utside of the real-time transient disco v ery w orkflow, Kilono va eek er s provides a framework for generating a number of human enchmarks, and gold-standard data sets for training machine learn- ng solutions, as a natural byproduct of the transient search workflow. e elaborate on a few ongoing analyses that provide substantial nsights into the abilities of our volunteers, and map out the ‘human actor’ present in transient follow-up, that few time-domain projects av e previously e xplored in detail (e.g. Goldstein et al. 2015 ; Hayden t al. 2021 ). To measure the intrinsic performance of volunteers, and etermine sensible classification baselines, we inject a number of alidation data sets (both intentionally, and intrinsically via known bjects) with known answers into the live project: (i) Hand-labelled validation data set: 300 examples, sampled niformly in real-bogus score from detections prior to project launch, nd hand-labelled by the Authors to ensure high accuracy. (ii) Minor planets: given the ingest pipeline is agnostic to con- extual information, these detections with high real-bogus score aturally enter into Kilonova Seekers as part of the transient search orkflo w. We kno w a priori that these are real detections, and the patial association enables us to retrieve high confidence low-signal- o-noise detections for testing. Kilono va Seek er s 2123 MNRAS 533, 2113–2132 (2024) Ta bl e 1. K ilo no va Se ek er s di sc o v er ie s r ep or te d to th e TN S, w hi ch w er e o bs er ve d by G OT O be tw ee n K ilo no va Se ek er s la un ch (20 23 Ju ly 11 ) a n d th e en d o f O 4a (20 24 Ja nu ar y 16 ). W e pr es en t t he TN S n am e, in te rn al G O TO n am e, G O TO di sc o v er y da te , K ilo no va Se ek er s as so ci at ed su bje ct id (s) o n Zo on iv er se , TN S re po rti ng gr ou p, tr an sie nt lo ca tio n, an d if kn ow n , th e cl as sifi ed ty pe an d re ds hi ft. R ed sh ift s a re ta ke n di re ct ly fro m th e TN S cl as sifi ca tio n re po rt, bu t r o u n de d w he re ap pr op ria te . TN S N am e G OT O N am e G OT O D isc o v er y da te (U T) K ilo no va se ek er s su bje ct/ s TN S R ep or tin g gr ou p R A /D ec Ty pe R ed sh ift Ki lo no va Se ek er s di sc o v er ie s AT 20 23 pm m G OT O 23 yt 20 23 -0 8- 05 04 :4 8: 55 91 25 9 7 01 G OT O : K ilo no va Se ek er s 02 :4 4: 18 .4 22 + 1 4: 23 :2 7. 51 – – AT 20 23 po f G OT O 23 vt 20 23 -0 8- 08 02 :5 5: 13 91 22 35 02 , 9 1 2 82 78 0 G OT O : K ilo no va Se ek er s 19 :4 8: 39 .6 23 + 0 0: 40 :2 5. 99 – – AT 20 23 ro b G OT O 23 aja 20 23 -0 9- 05 21 :5 6: 05 91 62 4 7 86 G OT O : K ilo no va Se ek er s 18 :5 5: 04 .8 78 −2 5: 42 :4 1. 94 CV – AT 20 23 w bu G OT O 23 bb l 20 23 -1 0- 28 06 :0 9: 44 92 88 9 8 63 G OT O : K ilo no va Se ek er s 10 :4 8: 51 .5 94 + 1 7: 37 :3 3. 02 – – AT 20 23 xn j G OT O 23 bi a 20 23 -1 1- 11 10 :4 5: 19 93 52 4 3 42 G OT O : K ilo no va Se ek er s 00 :2 1: 31 .4 92 −3 2: 48 :2 0. 18 – – AT 20 23 xq f G OT O 23 bi q 20 23 -1 1- 10 10 :4 0: 28 93 59 7 7 12 G OT O : K ilo no va Se ek er s 00 :0 3: 55 .1 59 −2 9: 35 :3 8. 95 – – AT 20 23 xq g G OT O 23 bi p 20 23 -1 1- 12 17 :0 7: 37 93 61 5 0 33 G OT O : K ilo no va Se ek er s 10 :3 9: 28 .0 16 −3 9: 31 :3 3. 69 – – AT 20 23 xq y G OT O 23 bjh 20 23 -1 1- 13 11 :0 6: 02 93 67 1 1 56 G OT O : K ilo no va Se ek er s 23 :4 1: 43 .0 58 −3 4: 12 :0 6. 46 – – AT 20 23 yd t G OT O 23 bl c 20 23 -1 1- 18 12 :1 6: 31 93 95 3 7 74 G OT O : K ilo no va Se ek er s 02 :1 9: 40 .7 42 −4 8: 15 :3 2. 90 – – SN 20 23 ye r G OT O 23 bl j 20 23 -1 1- 18 20 :4 5: 07 93 96 5 1 56 G OT O : K ilo no va Se ek er s 01 :2 1: 16 .7 00 + 1 7: 12 :5 5. 98 SN Ia 0. 06 AT 20 23 yo x G OT O 23 bm s 20 23 -1 1- 28 04 :5 4: 34 94 19 3 2 52 G OT O : K ilo no va Se ek er s 11 :5 5: 51 .5 73 + 4 4: 08 :0 5. 40 – – AT 20 23 yq r G OT O 23 bn o 20 23 -1 2- 02 10 :2 5: 06 94 31 0 8 06 G OT O : K ilo no va Se ek er s 01 :1 4: 48 .7 73 −2 0: 59 :4 1. 45 – – AT 20 23 yq s G OT O 23 bn n 20 23 -1 1- 30 11 :1 5: 33 94 31 0 8 14 G OT O : K ilo no va Se ek er s 02 :0 8: 23 .4 40 −3 5: 04 :2 3. 95 – – SN 20 23 yr s G OT O 23 bn t 20 23 -1 2- 03 13 :4 7: 40 94 32 2 3 74 G OT O : K ilo no va Se ek er s 06 :2 6: 52 .8 96 −2 4: 36 :5 3. 01 SN Ia -9 1- bg -li ke 0. 02 33 1 SN 20 23 ys p G OT O 23 bn z 20 23 -1 2- 03 13 :2 9: 44 94 33 2 7 59 G OT O : K ilo no va Se ek er s 06 :1 9: 37 .2 94 −2 9: 49 :1 6. 56 SN Ia 0. 09 AT 20 23 aa gc G OT O 23 bu s 20 23 -1 2- 15 12 :2 8: 26 94 83 6 5 62 G OT O : K ilo no va Se ek er s 05 :2 9: 37 .6 58 −3 5: 55 :1 6. 98 – – SN 20 23 aa jf G OT O 23 bw l 20 23 -1 2- 17 12 :0 1: 54 94 86 2 4 95 G OT O : K ilo no va Se ek er s 04 :2 2: 41 .4 84 −5 1: 29 :1 5. 63 SN Ia 0. 04 28 AT 20 23 ab dm G OT O 23 bz u 20 23 -1 2- 17 11 :3 1: 43 95 03 5 9 83 G OT O : K ilo no va Se ek er s 03 :4 1: 14 .3 08 −4 8: 51 :1 8. 08 – – AT 20 23 ab dn G OT O 23 bz s 20 23 -1 2- 24 11 :4 9: 37 95 03 5 9 74 G OT O : K ilo no va Se ek er s 05 :4 8: 49 .1 79 −2 4: 15 :2 1. 60 – – SN 20 23 ac la G OT O 24 P 20 23 -1 2- 26 04 :1 7: 46 95 12 8 3 49 G OT O : K ilo no va Se ek er s 12 :0 5: 02 .4 50 + 0 1: 10 :3 2. 95 SN Ia 0. 06 56 5 R ep or te d SN 20 23 ox c G OT O 23 uh 20 23 -0 8- 04 22 :1 2: 26 91 27 3 3 50 AT LA S 16 :0 4: 31 .4 69 + 3 6: 19 :0 0. 59 SN 0. 04 34 SN 20 23 ve r G OT O 23 bb c 20 23 -1 0- 26 00 :4 5: 31 92 80 9 7 61 Pa n -S TA RR S 03 :5 1: 40 .2 74 −0 0: 30 :3 8. 95 SN Ia -9 1T - lik e 0. 03 SN 20 23 vq n G OT O 23 bc c 20 23 -1 0- 27 21 :3 1: 28 92 88 9 7 37 AT LA S 22 :5 2: 31 .7 26 + 1 8: 14 :0 6. 46 SN Ia 0. 07 AT 20 23 xi g G OT O 23 bh y 20 23 -1 1- 10 13 :3 9: 34 93 46 4 8 49 AT LA S 04 :3 0: 41 .2 58 −3 9: 17 :5 5. 73 – – AT 20 23 ac do G OT O 23 ca a 20 23 -1 2- 24 11 :5 9: 58 95 13 49 96 , 9 5 1 90 83 5 ZT F 06 :0 4: 40 .4 10 −2 6: 38 :4 1. 64 – – SN 20 24 gy G OT O 24 J 20 24 -0 1- 06 05 :0 0: 18 95 41 3 5 90 K o ic hi Ita ga ki 12 :1 5: 51 .2 90 + 1 3: 06 :5 6. 12 SN Ia 0. 00 11 8 SN 20 24 hm G OT O 24 Q 20 24 -0 1- 06 10 :2 5: 10 95 43 0 4 26 AT LA S 03 :2 4: 06 .5 21 −3 8: 43 :5 9. 42 SN Ia 0. 06 7 AT 20 24 kh G OT O 24 X 20 24 -0 1- 06 05 :3 3: 21 95 60 1 2 22 AT LA S 13 :1 6: 52 .1 36 + 2 8: 06 :3 2. 66 – – AT 20 24 ag m G OT O 24 fq 20 24 -0 1- 06 05 :1 4: 39 95 97 4 7 95 AT LA S 12 :5 7: 38 .7 72 + 4 0: 11 :5 7. 38 – – D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024 2124 T. L. Killestein and L. Kelsey et al. M Figure 12. Boxplots showing the classification times of the Kilonova Seekers volunteers. Maximum time per classification has been clipped to 2 min to remo v e those classifications where someone paused mid-classification and submitted at a much later time. The lines inside the bars represent the median classification time, the boxes show the upper ( Q 3) and lower ( Q 1) quartile values, with width corresponding to the interquartile range (IQR) and the whiskers represent Q 1 − 1 . 5 × IQR and Q 3 + 1 . 5 × IQR, respectively. Figure 13. Kilono va Seek er s subject for AT2023xqy. This transient was flagged by the volunteers as real and reported to the TNS within 3 h and 20 min of data being taken by GOTO. r f t t v a m s t K c i e c m s f Figure 14. Precision-recall plot for the validation set, computed per volun- teer with o v er 100 classifications. The dashed lines partition the precision- recall space into quadrants, corresponding to the 50 per cent precision/recall boundary. The size of the plot markers is proportional to the number of classifications performed by that user. P R w b D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024The hand-labelled validation data set is given an arbitrarily high etirement limit to ensure as many volunteers as possible see them or comparative analyses. For the analyses that follow, we neglect he possibility of label noise (inaccurate labelling by the team) in he validation data sets. For the hand-labelled set, these data were etted by the Authors with both knowledge of the co-ordinates, and dditional contextual information (historical variability, source cross- atches) to guide the labelling. For the minor planet data set, we elect only detections with high-confidence ( ≤4 arcsec) matches o catalogued objects from the Minor Planet Centre, following illestein et al. ( 2021 ). Through analysis of the validation data set, and binary classifi- ation labels from volunteers, we can assess both the cohort and ndividual performance of volunteers in a real-world setting. To nsure low sampling noise in our estimations of precision, we only onsider volunteers who have completed 100 validation subjects or ore, yielding noise of O(1 per cent). We suspect the validation set ize is sufficient to mitigate data-driven scatter in metrics. As shown in Fig. 14 , we plot the precision (PR) and recall (RC) or each volunteer e v aluated on the hand-labelled validation data set.NRAS 533, 2113–2132 (2024) R = TP TP + FP (1) C = TP TP + FN , (2) here TP is the number of real transients correctly labelled as such y the volunteer, FP is the number of bogus transients incorrectly Kilono va Seek er s 2125 Figure 15. Fraction of positive votes per subject, binned by the SNR of the detection, derived from all live Kilonova Seekers minor planet detec- tions. Uncertainties are estimated by the one-sided binomial score interval approximation, with error bars representing 2 sigma. The 50 per cent reco v ery threshold sits around signal-to-noise 6. The harmonic mean of the real-bogus classifier score (Killestein et al. 2021 ) per bin is o v erplotted abo v e the bars in orange, for illustration. l b m F w b ( i 5 p l w o v t l d w t v c a t K d w c W c m t 1 p σ w t o c a P w n f b d n l m K s U s c m t n a h h p s a 2 w a o B a l a c s n u s l t t a i e D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024abelled as real, and FN is the number of real transients labelled as ogus. The F 1 score is a convenient metric derived as the harmonic ean of these quantities, given as 1 = 2 · PR · RC PR + RC , (3) here the precision and recall are defined as abo v e. The volunteers roadly perform well on the validation data set, achieving a median class-weighted, 1 σ uncertainty) F 1 score of 78 + 13 −35 per cent and lie n a cluster in the upper right quadrant (precision and recall abo v e 0 per cent). and represents a class-balanced accuracy, weighting recision, and recall equally. There are a notable minority (20 per cent) of volunteers who ie in the lower right quadrant (high precision, but low recall) – hom we interpret as ‘underconfident’ volunteers. When they mark bjects as real transients, they are likely to be correct, but they mark ery few objects as real transients – perhaps owing to not fully rusting their own predictions. Reassuringly, very few volunteers ie in the low precision region of the plot, characterized by poor iscriminative performance – we associate the upper left quadrant ith ‘o v erconfident’ volunteers, who reco v er the majority of real ransients but mark many artefacts as real. We hope that, over time, olunteers precision-recall scores will flow towards the upper right orner as they gain performance and familiarity with the workflow nd project. In Fig. 15 , we compare the reco v ery of minor planets by he volunteers compared to the GOTO real-bogus classifier (see illestein et al. 2021 ) as a function of the signal-to-noise of the etection. We cross-match all uploaded Kilonova Seekers subjects ith Minor Planet Centre 15 ephemerides, and in total retrieve 92 640 lassifications – which we know a priori are good transient detections. e compute the fraction of positive votes per signal-to-noise bin, hosen approximately to linearly span the range 3 to 20, where the ajority of detections typically lie. Uncertainties are estimated from he normal approximation (Wald 1943 ) to the one-sided binomial 5 https://www.minorplanetcenter.net g o c s roportion confidence interval: ˆ p = √ ˆ p(1 − ˆ p) N (4) hich is an adequate and asymptotically correct estimator, given the ypically large N per bin, and lack of bins with ˆ p close to zero or ne. F or comparison, we o v erplot the harmonic mean of real-bogus lassifier scores – the closest analogy to the fraction of votes positive pproach we use for volunteer labels. This is given as = 1 N N ∑ i= 1 1 p i , (5) here p i is the i- th classifier score in each bin, and N is the total umber of subjects per signal-to-noise bin. This plot highlights acets of the performance of both human vetters and the real- ogus classifier. The classifier score remains high across the SNR istribution, as expected. The marked bump at low ( ∼ 7) signal-to- oise in the classifier score is likely a result of the steep power- aw slope in the magnitude distribution of minor planets – with any times more small bodies than larger in the training set (see illestein et al. 2021 ). The human classifier scores show a smooth igmoid curve, passing 50 per cent recovery around a SNR of 6. ncertainties (given by the error bar) are largely driven by sample ize per bin, rather than human-derived uncertainty. The real-bogus lassifier score comfortably exceeds the human true positive rate, arkedly so at lower signal-to-noise. It is perhaps not surprising hat a classifier explicitly trained on minor planets outperforms a aive ensembling of human predictors – yet to our knowledge this is mong the first validations of deep-learned classifiers outperforming uman annotators in time-domain astronomy. We caution that the uman-derived fraction of positive votes may not be well-calibrated robabilistically, taking into account discussions on variable preci- ion and recall of volunteers abo v e – nevertheless via thresholding nd consensus these issues may be mitigated. Optimal schemes for thresholding or weighting (e.g. Marshall et al. 016 ; O’Brien et al. 2024 ) are left to future publications, though e note that the uncertainty is a crucial component of our science ims, and so fraction of positive votes is diagnostic here. With priors n the true/false positive rates per volunteer from the validation set, ayesian models of annotation (e.g. Paun et al. 2018 ) are a promising venue for deriving well-calibrated and optimal inferences on how ikely an object is to be real from volunteer labelling. Nevertheless, this result underscores that classifier scores alone re not sufficient to fully capture the uncertainty associated with a lassification. Subjects that are genuinely challenging in a statistical ense, such as those at low signal-to-noise, should be treated with uance to a v oid o v erinterpretation. This underscores the necessity of ncertainty quantification in classification Although early in the project’s lifetime, these validation data ets have enabled a number of interesting scientific (and socio- ogical) insights into the way volunteers approach classification asks, their intrinsic efficiency at recovering transient objects, and he different dispositions of the volunteers to classification. More dv anced v alidation experiments are currently underway – including njecting augmented variants of the validation set to track the volving performance of the volunteers between Kilonova Seekers enerations. One remaining, potentially insightful task is to re-run ur v alidation workflo w with GOTO team members to compare and ontrast Figs 14 and 15 , and measure the selection function of project cientists (similar to the investigation of Wardlaw et al. 2018 , forMNRAS 533, 2113–2132 (2024) 2126 T. L. Killestein and L. Kelsey et al. M M f e g a o a o d t 6 I a d s z d h b i m e r c l s 6 F h h e d s w A c w l t a d f m l s m T m t m w o a o u t t g o i t W n A W h s R C t M S I p d I p ( A S t ( P S S S p i t a M C I e e D G G c a U D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024artian surface feature detection and classification) – which could eed into downstream analyses to derive more informed recovery stimates/drive second-looks on more challenging data. Based on cuts inferred from the validation data set, we define our old-standard data set as subjects with > 80 per cent agreement, nd more than eight positiv e/ne gativ e votes from volunteers. Based n these cuts, we find a gold-standard data set of 17 682 detections cross O4a. This gold-standard data set is informing the development f the next real-bogus classifier within GOTO, with a more detailed iscussion of nuances associated with crowd-sourced training of ransient classification models deferred to a future publication. C O N C L U S I O N S n this paper, we have presented the first stage of Kilonova Seekers , citizen science project designed specifically for real-time transient isco v ery, complementing the unique capabilities of the GOTO urv e y for gra vitational-wa ve follow-up. In the period from 2023 July to 2024 January, Kilonova Seekers : (i) Achieved 643 124 classifications of 42 936 subjects. (ii) Attracted roughly 2000 volunteers, in o v er 20 distinct time ones, across 105 different countries. (iii) Reported 29 objects to the TNS, where 20 of these are isco v eries first reported by the project. Six of these disco v eries ave been classified spectroscopically by other teams. (iv) Achieved turn-around times of as quick as 3 h and 20 min etween observation and TNS report, for candidates flagged as nteresting by the volunteers. (v) Created a gold-standard training set of 17 682 subjects for achine learning, with o v er 80 per cent agreement among volunteers. (vi) Measured the detection efficiency of the volunteers at recov- ring transient sources, and compared this with the existing GOTO eal-bogs classifier. With this initial phase of Kilono va Seek er s , we have demonstrated oncretely that citizen science can work both in real time and low atency – driving decision-making and discovery on large data- treams. .1 Recent updates and future work or the O4b observing run which is now underw ay, Kilono va Seek er s as continued to grow rapidly and transitioned to an augmented ourly cadence upload, to further reduce the latency between discov- ry, upload, and consensus. This has led to a number of citizen science isco v eries within 2 h of images being taken. We intend to keep hortening this cadence towards zero-delay (uploads simultaneous ith pipeline completion), as surv e y and platform capacity allow. new injection of unbiased (spanning the full real-bogus range) andidates, which aggressively sample real-bogus scores across the hole range are proving an excellent seed data set for novel deep- earned classifiers in development. In the time taken to prepare his publication, Kilonova Seekers has now reached 31 disco v eries nd achiev ed o v er 1 million classifications from volunteers. A full iscussion of this second phase and ongoing disco v ery is deferred to uture works. Development of the Kilono va Seek er s w orkflows continue, with ulticlass, context-augmented workflows planned to be released ater in 2024. This will enable volunteers to not only classify if a ource is real or bogus, but to subdivide each of these classes into orphological types (e.g. supernova, nuclear transient, variable star). his workflow will further support the training of next-generationNRAS 533, 2113–2132 (2024) achine learning classifiers, and enable uncertainty-aware con- extual classification. The introduction of this Kilonova Seekers ulticlass will mark Gen. 3 of the project, and be accompanied ith a re-launch. This development is, of course, in addition to the riginal fast disco v ery workflow, to ensure continuity for volunteers nd maintain compatibility with mobile app users. Based on the keen engagement with Kilonova Seekers , a number f parallel companion outreach and public engagement projects are nder active development: empowering volunteers to do their own ransient follo w-up ef forts with professional telescopes, learn about ime-domain astrophysics through observing objects themselves, and enerate meaningful scientific outcomes and publications on the bjects they have discovered. The time-domain community are eagerly following up alerts dur- ng the LIGO-Virgo-KAGRA O4b observing run, hoping these GW riggers will facilitate disco v ery of new electromagnetic counterparts. ith the growth of the Kilonova Seekers project, this community is ow markedly larger. C K N OW L E D G E M E N T S e thank the anonymous referee for their insightful comments which elped impro v e the quality of the manuscript. TLK acknowledges upport via an Research Council of Finland grant (340613; P.I. . Kotak), and from the UK Science and Technology Facilities ouncil (STFC, grant number ST/T506503/1). LK and LN thank he UKRI Future Leaders Fellowship for support through the grant R/T01881X/1. EW thanks STFC for support through the grant T/Y509486/1. JDL acknowledges support from a UK Research and nnov ation Fello wship (MR/T020784/1). DMS ackno wledges sup- ort by the Spanish Ministry of Science via the Plan de Generacion e conocimiento PID2020-120323GB-I00 and PID2021-124879NB- 00. SM acknowledges support from the Research Council of Finland roject 350458. The Gra vitational-wa ve Optical Transient Observer GOTO) project acknowledges the support of the Monash-Warwick lliance; University of Warwick; Monash Univ ersity; Univ ersity of hef field; Uni versity of Leicester; Armagh Observatory & Plane- arium; the National Astronomical Research Institute of Thailand NARIT); Instituto de Astrof ´ısica de Canarias (IAC); University of ortsmouth; University of Turku. We acknowledge support from the cience and Technology Facilities Council (STFC, grant numbers T/T007184/1, ST/T003103/1, ST/T000406/1, ST/X001121/1, and T/Z000165/1). This publication uses data generated via the Zooniverse.org latform, development of which is funded by generous support, ncluding a Global Impact Award from Google, and by a grant from he Alfred P. Sloan Foundation. This research has made use of data nd/or services provided by the International Astronomical Union’s inor Planet Center. Software: This research has made use of ASTROPY (Astropy ollaborationt 2013 , 2018 , 2022 ), GEOPANDAS (Jordahl et al. 2020 ), RAF (Tody 1986 , 1993 ), MATPLOTLIB (Hunter 2007 ), NUMPY (Harris t al. 2020 ), PANDAS (McKinney et al. 2010 ), and SCIPY (Virtanen t al. 2020 ). ATA AVAI LABI LI TY OTO images and source catalogues will be made available in a OTO data release at a later date. Anonymized and/or aggregated lassification data are made available upon reasonable request to the uthors, but are anticipated to be released publicly at a later date. ser-le vel Zooni verse data and PII will remain pri v ate follo wing the Kilono va Seek er s 2127 Z i R A A A A A A A A A B B B B C C C C C D D D D D F F G G H H H H H I J J K K K K K L L L L M M M M M O P P R R S S S S S T T T T T V V W W W W W Z A W i o s t S r a M A J m F a A L d G B A C A D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024ooniverse User Agreement and Privac y Polic y –https://www.zoon verse.org/priv acy . EFER ENCES bbott B. P. et al., 2017, Phys. Rev. Lett. , 119, 161101 bdar M. et al., 2021, Information Fusion , 76, pp. 243–297 lard C. , Lupton R. H., 1998, ApJ , 503, 325 nderson J. P. et al., 2014, ApJ , 786, 67 ndersson A. et al., 2023, MNRAS , 523, 2219 rcavi I. et al., 2017, Nature , 551, 64 stropy Collaboration 2013, A&A , 558, A33 stropy Collaboration 2018, AJ , 156, 123 stropy Collaboration 2022, ApJ , 935, 167 ailey S. , Aragon C., Romano R., Thomas R. C., Weaver B. A., Wong D., 2007, ApJ , 665, 1246 ellm E. C. et al., 2019, PASP , 131, 018002 loom J. S. et al., 2012, PASP , 124, 1175 runer J. S. , 1961, Harvard Educational Review, 31, pp. 21–32 abrera-Viv es G. , Re yes I., F ¨orster F., Est ´ev ez P. A., Maureira J.-C., 2017, ApJ , 836, 97 ardamone C. et al., 2009, MNRAS , 399, 1191 hambers K. C. et al., 2016, preprint ( arXiv:1612.05560 ) orbett H. et al., 2023, ApJS , 265, 63 owell F. , 2011, Measuring Inequality. Oxford Univ. Press, Oxford. Available at: https:// doi.org/ 10.1093/ acprof:osobl/ 9780199594030.001.0001 avis K. , F ole y R., Jacobson-Galan W., 2023, Transient Name Server Classification Report, 2023-3315, 1 o A. , 2023, Transient Name Server Classification Report, 2023-3146, 1 uev D. A. et al., 2019, MNRAS , 489, 3582 yer M. J. et al., 2020, in Guzman J. C., Ibsen J., eds, Proc. SPIE Conf. Ser. Vol. 11452, Software and Cyberinfrastructure for Astronomy VI. SPIE, Bellingham, p. 114521Q yer M. J. et al., 2022, in Marshall H. K., Spyromilio J., Usuda T., eds, Proc. SPIE Conf. Ser. Vol. 12182, Ground-based and Airborne Telescopes IX. SPIE, Bellingham, 121821Y. Available at: http://dx.doi.org/10.1117/12. 2629369 remling C. , Neill D., Sharma Y., 2024, Transient Name Server Classification Report, 2024–174, 1 r ´enay B. , Verleysen M., 2013, IEEE Trans. Neural Netw. Learn. Syst. , 25, 845 ini C. , 1912, Variabilit `a e mutabilit `a: contributo allo studio delle dis- tribuzioni e delle relazioni statistiche. Tipogr. di P. Cuppini, Bologna oldstein D. A. et al., 2015, AJ , 150, 82 arris C. R. et al., 2020, Nature , 585, 357 ayden B. et al., 2021, ApJ , 912, 87 iggs C. , 2023, American Astronomical Society Meeting Abstracts, 140.03 inkle J. , 2023, Transient Name Server Classification Report, 2023-2214, 1 unter J. D. , 2007, Comput. Sci. Eng. , 9, 90 vezi ´c ˇZ. et al., 2019, ApJ , 873, 111 ones D. O. et al., 2021, ApJ , 908, 143 ordahl K. et al., 2020, geopandas/geopandas: v0.8.1. Available at: https: // doi.org/ 10.5281/ zenodo.3946761 asen D. , Badnell N. R., Barnes J., 2013, ApJ , 774, 25 illestein T. L. et al., 2021, MNRAS , 503, 4838 ochanek C. S. et al., 2017, PASP , 129, 104502 olb D. , 1984, Experiential Learning: Experience as the Source of Learning and Development. Prentice-Hall, Englewood Cliffs, NJ. Available at: ht tps:// books.google.fi/ books?id = zXruAAAAMAAJ opsacheili M. , Gonzalez-Banuelos M., Gutierrez C., Kopsacheili M., Gonzalez-Banuelos M., Yaron O., 2023, Transient Name Server Clas- sification Report, 2023–3082, 1 aw N. M. et al., 2009, PASP , 121, 1395 i L.-X. , Paczy ´nski B., 1998, ApJ , 507, L59 intott C. J. et al., 2008, MNRAS , 389, 1179 orenz M. O. , 1905, Publ. Am. Stat. Assoc. , 9, 209 arshall P. J. , Lintott C. J., Fletcher L. N., 2015, ARA&A , 53, 247 arshall P. J. et al., 2016, MNRAS , 455, 1171 cKinney W. et al., 2010, in Van der Walt S., Millman J., eds., Proceedings of the 9th Python in Science Conference, SciPy, Texas, p. 56–61 ong Y. L. et al., 2020, MNRAS , 499, 6009 ong Y. L. et al., 2023, MNRAS , 518, 752 ’Brien S. M. et al., 2024, AJ , 167, 238 aun S. , Carpenter B., Chamberlain J., Hovy D., Kruschwitz U., Poesio M., 2018, Trans. Assoc. Comput. Linguist. , 6, 571 erley D. A. et al., 2020, ApJ , 904, 35 au A. et al., 2009, PASP , 121, 1334 en P. , Xiao Y., Chang X., Huang P.-Y., Li Z., Gupta B. B., Chen X., Wang X., 2020, preprint ( arXiv:2009.00236 ) chwamb M. E. et al., 2013, ApJ , 768, 127 martt S. J. et al., 2015, A&A , 579, A40 mith A. M. et al., 2011, MNRAS , 412, 1309 piers H. , Swanson A., Fortson L., Simmons B. D., Trouille L., Blickhan S., Lintott C., 2019, J. Sci. Commun. , 18, A04 teeghs D. et al., 2022, MNRAS , 511, 2405 aubenberger S. , 2017, in Alsabti A. W., Murdin P., eds, Handbook of Supernovae. Springer, Cham, p. 317 ody D. , 1986, in Crawford D. L., ed., Proc. SPIE Conf. Ser. Vol. 627, Instrumentation in Astronomy VI. SPIE, Bellingham, p. 733 ody D. , 1993, in Hanisch R. J., Brissenden R. J. V., Barnes J., eds, ASP Conf. Ser. Vol. 52, Astronomical Data Analysis Software and Systems II. Astron. Soc. Pac., San Francisco, p. 173 onry J. L. et al., 2018, PASP , 130, 064505 ucker M. A. et al., 2022, PASP , 134, 124502 alentin Jospin L. , Buntine W., Boussaid F., Laga H., Bennamoun M., 2020, IEEE Computational Intelligence Magazine 17 2 irtanen P. et al., 2020, Nat. Methods , 17, 261 ald A. , 1943, Trans. Am. Math. Soc. , 54, 426 ardlaw J. , Sprinks J., Houghton R., Muller J.-P., Sidiropoulos P., Bam- ford S., Marsh S., 2018, Int. J. Appl. Earth Obs. Geoinform. , 64, 354 ozniak P. R. , Udalski A., Szymanski M., Kubiak M., Pietrzynski G., Soszynski I., Zebrun K., 2002, Acta Astron. , 52, 129 right D. E. et al., 2015, MNRAS , 449, 451 right D. E. et al., 2017, MNRAS , 472, 1315 ackay B. , Ofek E. O., Gal-Yam A., 2016, ApJ , 830, 27 PPENDI X A : FULL LIST O F VO L U N T E E R S e are truly grateful for the e xtensiv e effort of our volunteer scientists n making the Kilonova Seekers project happen. A full list of names f contributors (who gave permission for their name to be shared) ince our launch is given below in alphabetical order, correct as of ime of manuscript preparation: 5, 958bacsal, A, A Piras, A Taylor, A lot of imagination, Aaboli amant, Aarush Naskar, Abby, Abditory, Abdulla G. Asanar, Abdu- ahman Mohamed, Abel, Abhijeeth Veeranki, AbrilPerezH, abrosio, chmadsujana, Ada Ji, Adam, Adam Cash, Adam Gibson, Adam artinez, Adam Schufeldt, Adam Straub, adamzwawy, Adekunle dejokun, Aditi Brij, Adrian Morales, Adrian Smith, Adrianna ones, Adrien Droguet, Afjal khh, Agnid Nandi, Aguirre, Ah- ad azizisani, Ahmed Estiak, Aiden Chadwick, Aimee Gonzalez erreira Sirvani Valentim, AJinSA, Aki Suvanto, Akiko Inamoto, knepeter, Al Lamperti, Alaa Salah Afifi, Alan Teague, Albert, lbiona Leka, Alejandro, Alejandro Arr ´olig a Vaneg as, Alejandro opez, Aleksandr, Aleksandr Ketov, Aleksandr Timofeev, Aleksan- ra Pogorzelska, Ale x, Ale x Al-Sammarraie, Ale x Andersson, Ale x abriel, Alex Lammers, Alex Mitchell, Alex Zuniga, Alexander ecker, Ale xander Blagrav e, Ale xander Davidson, Ale xander Doens, lexander G. Plasser, Alexandra Hercilia Pereira Silva, Alexandre elier, Ale xia F otini P anagopoulos, Ale xis Carrilllo, Ale xis Case y, lexis Daniel G ´omez Alatorre, Alexis MANET, Alexis Tombrello, MNRAS 533, 2113–2132 (2024) 2128 T. L. Killestein and L. Kelsey et al. M A A P A G A a K A C A M S R A B A B A B D A A A P t F A G A n A d A m A A A A A A b m A M K B Z l B W S H B O B B B B C N C M n M C S R B c m H F c C C L C T S I C C C s V c C S C L B s J D C K R D D i d A L D D C D G D D D D M D m D D t F O M E M D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024lfredo Gimeno, Ali Kiwan, Ali Reza fani, Ali Tejani, Alice, lice Bull, Alice Hu, Alima, alimamo, Alina Borissenko, Aliona hilippova, Alison Edwards, Allison Myers, Allison Umberson, lma, almalthea, Alvin Echeverria, Alyssa Chandler, Amadeus abriel dos Santos Siqueira Silva, Amanda, AMAR PAL SINGH, maury Vincent, Amber Alvidrez, Amelia Chaber, Amirali Shahri- rymanesh, Ammar Vora, Amoli Kakkar, Amy, Ana, Ana Haag, Ana aren Tapia, ANA LUIZA MAXIMO AGUIAR DE ALMEIDA, na M. Pizarro Gal ´an, Ana Paula Waaijenberg, Ana Sofia de Oliveira aldeira, analemma.sky, Anamaria Liana Axinte, Anargha Bose, nastasia Eriksen, Anastasia Prybytko, anastasia scoggins, Anay ishra, Andrea Bortoluzzi, Andrea Espinoza, Andrea Nava, Andrea erio, Andrea Williams, Andrej Coleman, Andres Eloy Martinez ojas, Andrew, Andrew Bickley, Andrew Boyer, Andrew Conan, ndre w Cooper, Andre w Del Santo, Andre w Obara, Andre w Shaw Sc(Hons) MCPara MRi, Andrew Waldie, Andrew Winkelman, ndre y Korobko v, Andrii Dzygunenko, Andry Nasief, Andrzej obinski, Andrzej Wojtowicz, Andy Tonthat, AndyTheAstronomer, nel Madrigal Gonzalez, Angad Chadha, Angel Elbaz Sanz, Angela rito, Angela Volpe, Angelika Reithmayer, Angelique Reder, Angelo e Lemos, Anil Vasudev, anita martins da cruz, Anita Springer, nna, Anna Andriyanov, Anna Batue v a, Anna Brisa Michef f Soares, nna Clara de Souza Fraga, Anna Kruchinina, Anna Mackiewicz, nna Plum, Anna Scott, Anna Vorobe v a, Anna Zanone, AnnaJe wel ace, annparker, Anond Disyatat, anthony, Anthony R. Wells, An- hon y Rainone, Anthon y TREMBLIN, Antonio, ANTONIO JEF- ERSON MONTE ALVERNE PAULINO, Antonio M. Puertas, ntonio P asqua, Anton y Davi Costa de Sena, anwilk, Anylem onzalez, An ¯d ela Mogin, Aoife Boyd, Aoiffe Boyle, Aparna Joshi, rchana, Ariana montes, Arianne Ambion, arianny caetano, Arka- ar, Arkaprova Dutta, Arkhipova Daria, Arla Heikkinen, Arlind.S, rman Svoboda, armandina gutierrez, Armando I Zamora, army- ragon637, Arnaud Dufourcq Lagelouse, arsama, Artemii Krykun, rthur Almeida, Arthur Meunier, Arthur P. Pereira, arthur pereira artins, Arttu Sainio, arturovasquez, Artyom Yakubov, Aryan Vinod, sh Washburn, Ashlee Kephart, Ashleigh Goh, Ashley Abrego, Shle y Wilkinson, Ashle y Willis, Ashton, Ashtyn Gibbs, Ashutosh, shwin Shenoy, Asim, asterisk man, Athanasia Vlachou, Atlas, ubrey Tyson, Aurelijus A. Alekserius, Auriam, Aurora, Auryne, ur ´elien GENIN, Austin Hughes, Axavier ne yra, Ax el Geo vanni, ya Ahmad, A ydın AYBAR, A yushmaan Mishra, B L Goodwin, adgerfish, Baiba Dislere, Barbaa, barbara england, Barbara Hart- ann, barmet76, Barrie Matthews, Bartlomiej Krajewski, Basar nil, Basil, Basudev Bhattacharya, Basundhara Maji, Bawan Aziz uhemed, bdinti, Beatriz Barros Maia, Beau, bekind2all, Bella arlisch, Ben Bartel, Ben Cole, Ben Kelahlyah, Benjamin Kapsch, enjamin Olson, Benjamin Pumphrey, benjamin savageau, Benjamin ahradnik, Benoit ROUSSEAU, Bent Løschenkohl, Bernd Niko- aus, Bernhard, Bernice Buan, besharp, beta cigni, Beth Meeker, HARAT GUPTE, Bhavesh Sai Arambakam Madhu, Bianca, Bj ¨orn ilde, Blaize Baehrens, Bob, Bob Birket, Bobbi Marcum, Bogosi ekhukhuni, Bokre Samson, BorisBanjac, Boundlessness, Braden ancock, Brady Lundin, Braiden king, bramboro, Brandi Halloran, randie Nuckolls, Brandon Adcock, Brendan, Brennen Boyer, Brent ’Connor, Brett Reilly, Brian Andersen, Brian cloke, Brian Nevins, rian Spirk, Briana Gulas, Brianna, Bridget Foster, brinlong, Brittany rockenton, Brix Ola, Broc Daly, broe317, Bronwyn Wall w orth, ruce Griego, Bruce Horlyck, Bryan F. Smith, Buldris, buzzwon, yron allen be gle y, C Unsworth, C. D’silva, C. Luke Gurbin, . S. Tolliver, Caballero, Gabriel D., Cairo Taylor, Calvin D ourse, Cameron Alexander, Cameron Johnson, Cameron Lopes, amille Mumm, Candela, CANNIZZAR O, Carl Setzer , Carla V.NRAS 533, 2113–2132 (2024) ejia, carloartemi, Carlos Alfredo Narv ´aez Gait ´an, Carlos Anto- io santos, Carlos Augusto Ara ´ujo Silva, Carlos Nunez, Carmen andel, Carol A. Schneier, Carol G Taylor, Carolina Bresciani, arolina Dos Santos Casaleiro Da Silva, Carolyn Bolus, Carolyn ill, Carolyne Brough, Carrie black, Carsten Meldgaard, Carston ose, Carter Hathaway, Caryme Martinez, Casandra Martin, Casey onham, Cassie Merkel, Cath Cockeram, Cath Sharp, cathcollins, atherinebp, Cau ˜ a Filipe Ribeiro Albuquerque Silva, Cecilia Lo- ax, Ceilidh Macrae Kirk, Ceona E., Cezary Kruszewski, Chan wee Im, Chappers34, charbel saliby, Charles Pennison, Charlie rost, Charlotte Williams, Chase, Chasity Newland, Chayse Jones, hemistinside, Chen Shaojie, Chen Stanilo vsk y, Cherridah Weigel, herrine W ilder , chhanda bewtra m.d., Chiara palmitesta, Chinabob, hipFaust, Chiroko, Chloe Ernspiker, Chloe Greenbaum, Chloe Le acheur, Chloe McElroy, Chris, Chris Barbosa, Chris McDaniel, hris McFarlane, Chris Mitchell, Chris Nowlan, Chris Pattison, Chris heofel, Chris bushell, chriscasper, chrisfro, Christian, Christian ergienko, christine groen, Christine Lee, Christopher B. Davis, I, Christopher Bowen, Christopher Horga, Christopher Pemberton, hristopher Strauss, Christy Browne, Chuck Henrich, Cian Maestri, iaron Drain, Ciro Sirio Perrella, Claire R. Hadley, Claire Volinski, laude Cornen, Claudia Gonzalez Lozano, Claudio Correa, Cledi- on Marcos da Silv a, Clif f Kurlander, Clif ford Bro wn, Cl ´ement iolette, Cody Cook, colcol, Cole, Cole Murphy, Colin Chandler, olin hewitt, comface, Connor Sands, Cooper Ev ans, Cooper K elly, ore y McInerne y, Cory Chambers, Craig Foss Olsen, Cristiano ecci, Cristina Almeida, Cristopher Cojocaru, cs192, csprucefield, Thomas, cubby348, curlytoplu, cwilton, Cynde, Cynthia Jerez- ema, Cynthia Moore, Cyril, Cyrus Trial, C ´eline de Ruiter, D rough, D J Spruce, d.gordon.banks, d ashenden, dadotron, Dale inclair, Dalia Garcia, Damian Gleis, Damian Janson, Damien ackson, Damien Laouteouet, Dan Ryczanowski, Dana Lubow, aniel, Daniel Alquizaleth, Daniel Amaya, Daniel Berliner, Daniel onte, Daniel Gadomski, Daniel Henley, Daniel J. Reisner, Daniel arnuakh, Daniel Leibman, Daniel mireles do nascimento, Daniel aso, Daniel Wolf, Daniela Gallego Ram ´ırez, Danielle Perkins, annis Vo, Danny Cameron, Danny Campbell, Danny Roylance, anveer Kalliecharan, Daria Machina, Darien, Darien Lefort, Dar- us Gumuliauskas, DarkAryan, dash 5, Dave Anderson, Dave D, a vews333, Da vi Cordeiro dos Santos, Davi Lima Alc ˆ antara, David khmadullin, David Baker, David Briggs, David John Flood, David ´opez Mart ´ınez, David Meierhenry, David R Harris, David Saewert, a vid Stefaniak, Da vide, Da vide Iannone, Da vidFoss, da vidselfe, awn Sturgeon, dcortesi, Dean Santos, deanroberts, DEBAYUDH HAKRABORTY, Deborah Kelsey, Deborah Woods, Declan Raven, een, DEEPAK, Deirdre Harris, dei v ad, Dena A Mitchell, Denilso . Delfrate, Denis, Denis Hathaway, Denis Pilon, Dennis Rowland, ennis Toy Jr, Derrick Wales, Destin Smith, De viek, De von Gerik, evrit Saha, Dhruv visariya, Dhruvatara Bhogishetty, Dhuertas, iana Sironi, Didac Invernon Campoy, Diedre barnett Garcia, Diego iaz, Diganta Sonowal, Dimitri, Dimitri Ferreira Lima, Dimitris itsikaris, dirkie, dj tjitso, Dmitriy, Dmitriy Korovin, Dmitrujs, ocwill8, dom mercer, Dominik Siefert, Dominik Swiniarski, Do- inik Valouch, Dominika, Don, Don Feldman, Douglas Higgs, ouglas Madzier , Dphr , Dr Peter Musk, Dr Sabrina G ¨artner , Dr Brian ecker, DrT.K.Subramaniam, Drew, DrKlahn, drokly, Duangrue- ai Samransanit, Dubravko Jakovljevic, Duncan Grant, DUPONT lorent, Dylan Drazek, Dylan Jusino, Dylan N. Weinrich, Dylan wen Reserva Unas, D ´avid F ¨ul ¨op, E Pratt, E. Mayr, E.N.G., Eaden orton, Ebubekir Sark, Edg ar Guzman-Contreras, Edg ard Schwarz, dna Soto, Eduarda V Baldo, Edward, Edward Caplin, Edward okurai Cherlin, Edwardo Garcia, Ege Turker, Ekin Alp Arslan, Kilono va Seek er s 2129 e E B G K E E E e E e B T L E E B c C L A fi N F c D F f R L S r T D i G B R G G G G D W G G L G A H S D T S H B S K M J H B I I W j I a k D J J T S J F G P c J J s j J H D J J o P J j J J Z j M L J E J O G S J J C J M D J J w M C j J J R J J K K K D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024landale, Eldhie Joy Rosales, ElectraVentures, Eleftheria Travlou, lena Akimova, Elif Bayat, Elisa Di Dio, Elisabeth Baeten, Elisabeth rann, elisah, elissa steele, Eliyah Palamarchuk, Elizabeth, Elizabeth all, elizabeth serna, Elizabeth Swope, Elizav eta Svito va, Ella, Ella atkova, Ellie Gold, Elliot Jones, Elyssa Smith, Emanuel Agapios, mi, Emilia Domingos, Emilie Wuattier, Emils Locmelis-Lunovs, mily Burrage, Emily Jayne Bean, Emma Boyett, Emma Fagan, mma Ryan, Emma Sarkissian, Emmett Hein, emptylica, epv95ngc, quidad1, erez dagan, Eric Bellm, ERIC F ABRIGAT , Eric Kim, RIC MAILLOT, Eric Peuster, Eric Yachen Zou, Erica J Welborn, richill, ericjpaquin, Erick Gomez Lopez, Erik Rodriguez, Erin rache, Erin Comparri, Erin Norris, Erin Zorzy, Ernest Jude P. iu, Ernst Schneidereit, Esmeralda Gonzalez, Estelle Baude, Esther iufu, ET Junior, Ethan, Ethan Alday, Ethan Atkinson, Ethan Estey, than J. Keefe, Ethan Vice, Ettore Fernandes Damique Aguiar, ugene Mercado, eugenius, Evan Barber, Evana Shrestha, Evangelos atzios, Eve, Evgeny Epifanov, evyn, Ewout Kerklaan, ExavierM- Leod, e xpofev er, Eyob, Eyvindr Leav enworth, Ezequiel Santos outo, ezflyer, fabienmazieres, Fabi ´an Bacca Alvarado, Fabrice amareille, F abr ´ıcio F achini, Felicia Yllenius, Felipe Laruccia Sant nna, Felipe ranzani de Luca, Femke de Vroome, fierybrunettlass, lippp, Finley Saville-Brown, Finn Suratt, Fiona Ellis, Fiona Mc- eill, Flaviano Santos dos Reis, fleuger, Floor Goossens, flya200, OURNAISE Alexandra, Franchesca Flowers, Francis Varley, Fran- isco Alexander Balmaceda VII, Francisco Zala Rucabado, Francois UFOURMANTELLE, Frank Decapio, Frank Helk, Frank Stuart, red, Fred Hellmig, Freddie Hason, Frederic Elcin-Coolidge, Freya, rozenchosen, Fujai Muhammad Charieth, futterw ack en, Fuyuki emix, FZolee, G Castro, G.W, Gabriel Jaimes Illanes, Gabriel a wrence, Gabriel P alacios, Gabriel Stewart, Gabriella Costa de ouza, Gabrielle Mendon c ¸a, Gamar Alsadah, Ganymede3, Gar- ett Cornwell, Garrett Smith, Gast ´on Gonz ´alez Kriegel, Gaudin itouan, Gautham Arun, Gavin Dukowitz, Gavin SLoan, Geert ankers, Gemini Smith, GeminiNoSaga, Geof Wyght, Geof f K eel- ng, George Bowers, George G. Guilkey, George Humberstone, eor ge Kokaev, Geor ge Luker, Geor gia Lock, Geor gina Fern ´andez elmonte, Gerald W. Nash Jr, Geraldine Qiu, Gerard Planelles ipoll, Gerrit Bischoff, Gert Jan Klootwijk, gfox, Gianluximon, ianni Tornaghi, Giovanni, Giovanni Aparicio, Giovanni Colombo, iselle Sanchez, Giulio T. Forcolin, Giuseppe Conzo, gjcolburn86, loria (preferer George please), Gloria Hernandez, Glorii, goggog, olden Wolter, Goowithabrain, Gord Harmer, Gorka, Gorobets mitrii Andreewitch, Grace Mere-ana Ashby, Grace Parker, Grace aller, Grace W ells, Graeme Bartlett, Graham Parlett, Grant Larsen, rantham Norris, greenfield05, Greg, Greg Borders, Greg Gajer, re g Schwitzer, Gre g Scott, Gre gg Kerlin, Gre gory Aydt, Gre gory ewis, Gribol, grosbeak, gryphachu, Guillermo S ´anchez Calvo, uoyou Sun (   ), Gurmanavdeep Singh Mahal, Gustavo fonso Gomes, Gusta v o Manzanilla, Gwendolyn Cardente, gwhw, aaniya Khan, Hakkı Alp Tekin, Haley Smith, Hali Edmunds, Halley olanum Theia Janus Culver, Halvor Nafstad, Hangar77, Hannah iBenedetto, Hannah Foltz, Hannah Martin, HarpiaLC, Harriet yler, Harry Adams, harsh mahajan, Harsh vardhan, Harshdeep ingh, Hasan Arda G ¨uler, Hatim Piplodwala, Haven Tyler, hawkman, eather Ritter, Hector M Castro, Heidi, Heidi deVeyra, Helen ates, Helen Spiers, Helena Jane Gomez, hellkr, Helo ´ısa Pascoal de ouza, Henning von Hoersten, Henry Gagnier, Henry Rauch, Henryk rawczyk, Hern ´an Flecha Alf aro, HerrStahl, hiba f arrukh, Hiba ohiuddin, highwaystar, Higor Gabriel jadjiski soares, hiko, Hilary ohnson, HippyPhysicist, Hiruve Gallo, Hisato Hayashi, Holen Yee, risto Delev, Hugo Andr ´es Durantini Luca, HummDinger, Hunter urke, Hushaan, HypnotiQ, Ian Banbury, Ian Barber, Ian Branigan, an Chu, Ian Kennedy, Ian Lin, IanH84, Igor Akeliev, Igor Korotskin, gor Kuchik, Igor luiz lein martins, Iliana, Iliq zlatanov, Illyana einzetl, Ilyas W ajahat Zafar Jalisi, imdra, Ine Theunissen, inge anson, Inken Gatermann, Irina Thome, Isaac Wardell, IsaacPerks, sabella Read, Isabella Suzanne Valentine, Isabelle bourgeois, Is- belli do Vale Silv a, Isac Oli veira Leite, Isadora Velloso, ishaan olipaka, Ishita Jaisia, Isidora mart ´ınez, Issy Walker, Istiu, Isza enise De Jesus, Ivan Martin, Ivan Titov, Izabel Bramlett, J N, . Furst, J. J. Dziak, J. Oliveto, J. Toth, Jaana Kemppainen, Jacek ackiewicz, Jack Anderson, Jack R. Brelsford, Jackarific, Jackson omaszewski, Jacob Balch, Jacob Hanini, Jacob Rogers, Jacob chmidt, Jacob Thadius Giggey, Jacob Williams, Jacqui S, jacquiejh, ade Friedlis, jadkinssd, Jagadeesh Pitchai Pazham, Jahcari, Jaime rankle, Jake Chon, Jakub Kowalik, James, James Galla, James arland, James Goerke, James H Kinsman, James Hewitt, James earson, James Smith, James Wilson, jamicze, Jamie Bjune, Jamie hild, Jamie Griggs, Jamie Ramsay, Jamie Thompson, Jamie Wyman, amon, Jamy547, Jan Jungmann, Jan Slavick `y, jan55, Jardin Nathan, ari-Pekka P ¨a ¨akk ¨onen, jarphys, Jarwen , Jasmine Lao, Jason, Ja- on Daniels, Jason Griffith, Jason Singleton, Jatin Singh Tomar, a vier, Ja vier Alvarez-Escalera, Ja vier Gonzalez Duran, Jay Darnell, ayanta Ghosh, Jazz, Jbrabham, jddavidson, jean cool, jedkat, Jeff amner, Jeff Lesperance, Jeff Wilson, JEFFERSON LORENCONI E MORAIS, Jef frey Ruf f, jelik, Jen Beck, Jennifer Burstein, ennifer Kestell, Jennifer Krouse, Jennifer Penoyar, Jennifer Rackley, ennifer Shearer, Jenny X. Zhao, Jeremiah Sisemore, Jeremy Maci- lek, Jeremy Thomas, Jerico B. Azarcon, Jernalyn Dulza, Jeroen ullens, Jeronimo, jess77, Jessica, Jessica Field, Jessica Shaffer, essica Vaccarino, Jesus Bible, Jesus Eduardo Ceron Sanchez, gendera, Jhonatas Tokuno de Campos Firmino, Jian Sundvall, iashuo Zhang, Jillian Ropchan, Jim O’Donnell, Jim Paszternak, imena Bra v o-Guerrero, Jimmy Fisher, jin young kim, Jingyuan hao (  ), Jkmorse57, jlam21xp, jlynec, jmalnar, Joan Kalec, oanhopkins08, Joanna Jarmolowicz, Joanna Kaczmarczyk, Joanna olenda- ˙Zakowicz, Joao Pedro Santos, Joaquim Queiroz, Jocelyn eon, Jodhviir Sekhon, Joe CC, Joe Lane, Joedube11, joeK2 45, ohan Joby, John bowles, John C. Skorupski, John D. Krull, John ltgroth, John Engler, John Falconer, John Gibson, John Haight, John ossy, John Li Chen, John M. Cummins, John Martin Hunter, John R. ’Grady, John Simpson, John Welsh, John R Williams, Johnathan ueltzau, Jolene Nethaway, Jon Bueno, Jon Nugent, Jon P av er, Jon utton, Jonah Donis, Jonah Gluck, Jonas Lehnberger, Jonas Nagel, onathan Hatton, Jonathan W. Landers, Jooheon Lee, Joonzoon, ordan Newman, Jordi, jorge, Jorge A. Vilchez, Jose Alberto da Silva ampos, Jose Gabriel Nino Barreat, Jose Luis Perez III, Jose Reta, osef, Joseph Brom, Joseph Constantine, Joseph M. Crisp, Joseph olnar, Joseph Morrison, Joseph Vinik, Joseph-Michael Viggs, Josh ean, Joshua, Joshua Adams, Joshua Green, Joshua Hottenstein, oshua K. Sulli v an, Joshua Malcolm garner, Joshua McArthur, oshua Slauer, Joshua Tan, Joshua Thompson, Joshua Truong, josh- ilde, Jo vokna, Jox ean Koret, Joyce Kimbrell, Jo ˜ ao Paulo Molina oraes da Silva, jpvignes, JStarhunter, Juan Antonio Serrano, Juan orti, Judith Kokesch, Judith K ¨onig, Judith M. Kirshner, Judy, uhana, Jules van Horen, Julia Allison Urawski, Julia Augustin, ulia Ellers, Julia Hodges, Julian Van Allen, Julianne Register, ulien Cochet, Julien Ortega, Juliet Guttendorf, Julio C ´esar Evaristo osa, Junghoon Chung (Kyle), Justin Abramson, JustinPaulson, uvenal Barry, Jyothsna Terli, J ¨urgen Saeftel, Kabir Singh, Kacper . Kowalski, Kacper Zydron, Kaden Tro, kafter, Kai Macci, kainat, aitlyn R Deacon, Kanishka Faqiryar , Kara Alber , Karan Choudhary, aren Babich, KarlPettit, Karolina Biskup, karu58, Kass Ulmer, at Elder, Kat Lakey, Katar ´ına Ka ˇcicov ´a, Kate Reddick, kateboyd,MNRAS 533, 2113–2132 (2024) 2130 T. L. Killestein and L. Kelsey et al. M K K k D B O K C K l K b d K K K K L A f W V L L C L L T R M B W B L M a l J L L F L m M H B K d M K K M M F M O T G A R M M G T M M l t T M L m B P M S l m m A P M B l S b M J S K E M M M M M J J s M S G D m R K L K J N z N n W N D J N w L N n N Y D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024atharine Harris, Katherine Alice Tylczak, Katherine Collins, Katie, atie Simpson, Katja No vitsko va, Katja Schilder-Leu, Katrina, aufmann, Kay Williams, kay d, Kayla Briggs, Kaylee Coley, Kaylee oll, Kayleigh Spriensma, Kaylie Scorza, KB Burson, KELLY RADFORD KITCHENS, Kelly Lynne Bowell (Baulos), Kelly Leary, Kelly Rolyns Dela Cruz, K ellySun, K elvin Haakmeester, endall, K endall Do w, K enedy Chauv ette, kennedy v elado, Kenn y av anaugh, K erry Knight, K erttu Saatsi, K erys Taryn Ste venson, esi Butler, K e vin Borrot, K e vin H. Smith, K e vin Healey, K e vin K e- aher, K e vin Lam, ke vin reyes, K e vin Wailes, khamyn Collier Smith, illerApp, Kim Berry , Kimberly , Kimberly Ho v encamp, Kim- erly Tangkion, Kimberly VanMoorlehem, Kimberlyn Ortiz, kim- onghyeon, Kings Cordova, Kiran Tikare, Kirby, Kiri L. K. Salazar, jell Nilsen-Nygaard, kkolodzie, kleistf, Korben, Korina Koci, ostriko v Kirill, KOZle gend, krahling, kreole, kresimir.jednacak, ris Hunt, Krissy Guttendorf, Krist Lim, Kristen Therese Roehrig, senia Shutovska, ktarkin, Kyle Fitzgerald, K ylie Coakley, K ynko, adislav Vondr ´a ˇsek, lailawise, Lance Ganey, Lanny Huynh, Lara rce, Larice Fields, Larry.Melanson, Lars Erik Kjellstr ¨om, laughter- ollo wsme, Laura Bro wn, Laura Dionysius, Laura Trouille, Laurel olchok, Lauren Deinhardt, Lauren Gardner, Lauren Harris, Layssa ict ´oria Barbalho Assun c ¸ ˜ ao Silv a, lbibayof f, LDJ101, ldybug5012, ea Frohring, Leah Callender, leelaht, Leila Stemler, leisanearthlink, eka Sree Jaishankar, Lemartinel, Lena Blomeyer, Leonardo Alberto olombo, Leonardo Rodrigues, Leonel H ´egues Nava, Lezel Rabusa, iam Ball, Liam Yanulis, Lidia A. Rendon Martinez, likethedeserts, incoln Hamilton, Linda Hoxie, Linda J. Berkel, lindemose, Linden hompson, Lindse y Skinner, liondav e, Liran Zhu, Lisa Cali, Lisa ust, Lise Smith, Lissa, Liz Dowthwaite, Liz Erator, Lizbeth C o ya, Lizzeth Ruiz Arro yo, llvieg as, Log an A. LaChance, Log an uck, Log an Burling ame, Log an Fisk, Log an Sedgwick, Log an arren, lolinda, Longueville Matthieu, lopsidedhead, Loredana ellucci, lorenzlacson, Lorenzo Bay-M ¨uller, Lori Romua, Louis, ouis A Masciocchi, Louis Sterobo, Louise St. Germain, Louise- arie Bertrand, Low Koon Yen, Lo ¨ıc Baert, Lrsafari, Luana alves ntonio, Luana Mayara Rosa Afonso, Lubov, Luca Ubertini, Lucas, ucas, Lucas Marcoccia, Lucas Stroobants, Lucas Wheeler, Luci ackson, Lucia Maria Ferreira Lima, Luc y chachka, Luc yitSwD, uigi Ferreira Lima, Luigi Lalonde, Luis Rosas, Luis Wieczorek, uiz Eduardo Pugliese Ferreira, Luiz Fernando Teodoro Tabosa legler , Luke Hauser , Luke Holland, Luna Costa Antunes, luohua, vdStraat, Lydia Watkins, Lynne Dale, Lynnette Colombo, Lyud- ila Aleksyuk, M.S. Batukov, M.T. Mazzucato, Mable woodbury, ackenzi Hill, MacKenzie Kaser, Madeleine Flowers, Madison all, Madisyn Shick, Magdalena Badura, Magdalini Christopoulou, iologist, Magomed, Mahdi Zakir Mahi, Mainak Mondal, Maja uchci ´nska, Malcolm F o wles, Mallory K elly, Manjit, manseincar- iff, Manuel PONCET, Mara Clarissa Codog, Marat, Marc Hanset, arc Philipp Riker, Marcela Moreira, Marcie Nightingale, Marcin obielusz, Marco Bortolazzi, Marco Tison, Marcus Camper, Marek l ˇcovansk `y, Marek Skalski, Margaret Walby, Margaret Zimmer, aria, Maria Atef Thabet, Maria Clara Silva Bezerra, Maria Dubiel, aria Eduarda de Souza Vasconcelos, Maria Elsa Rosales, Maria ernanda Maciel, Maria Isabel Jarrin Nunes Nishiyama, Maria ADIK, Maria Martin, Maria Regina Ortiz, Maria W icher , Maria- livia Torcea, Mariah t, mariakruglova, Marianne Hundling, Marie essier, Marika, marina alkiviades, Marina Corr ˆ ea Freitas, Mario iov anola, Marisol Alv arado Resendiz, Marius Agafitei, Mark bbott, Mark Hackney, Mark Hillaert, Mark mccormack, Mark eynolds, Mark S. Hamblosa, Mark Stypczynski, Mark Moffatt, arkus Biburger, Markus Buchhorn, Marlon Braun, Maros Rolko, arta, Marta ´Swi e ¸ch, martdoespenguinwatch, Martell ValenciaNRAS 533, 2113–2132 (2024) uerrero, Martha Casquette, martin hall, Martin Thomas, Marvin hunderbolt, Mary Ann Brennan, Mary Falk, mary joy romano, ary Kay, Maryangel Arenas, MaryV, MARZAK, Mar ´ıa Cecilia elgarejo Ferreyra, Mar ´ıa Hern ´andez, masch914, Massimo Buck- in, Masuma Akhter, Matalin Hansen, Mateusz Sałasi ´nski, Ma- eusz Szałankiewicz, Mathias Steyaert, Mathieu Laroche, Mathilde empke, Matt Hill, Matt Naylor, Matteo Cianfaglione, Matteo osca, Matthew, Matthew Day-Lopes, Matthew Krause, Matthew awrence, Matthew Schlademan, Matthew Terren, Matthew Zim- er, Matthias Breimann, Matthias Kohl-Himmelseher, Matth ¨aus alderer, Maureen Roberts, Maurice A. Hippleheuser, Mauricio izarro Morguno vsk y, max, Max Bhoyroo, Max Krievs, Max Read, ax Stephenson, Max Werner, Max Zakrzewski, MaXoN, May uu Suu Myat, Maya, Maya Douglas, Maya Fiorentino, Maya ester, Mayahuel Torres Guerrero, Mayukha Dissanayake, mdlw, dmcdermitt, Medina Fazlic, Meena Balakrishnan, Megan, Mehal, ehmet t ¨ukenmez, Mekki, Melanie Garcia, Melissa Arsaut, Melodie ngela Pangilinan, melody maddahian, Melyssa Danielle de Souza ereira, Mercutio Cac ´er `es, Merijn Benard, Mervyn Hing, mfwuk, g Kaung Sat Thant, Mia Burgener, mia hernandez, Micaiah alonek, mich wimbledon, michael, Michael Antoni, Michael Bor- and, Michael D. Cochenet, Michael Demuth, Michael Douglas oares Tobias, Michael Ellison, Michael Galbraith, Michael Ham- lin, Michael J Morrin, Michael J. Adams, Michael J. Tolentino, ichael Langdon, michael mae, Michael Nunn, Michael Pearle r, Michael Phillip, Michael Please, Michael Stouffer, Michael tutters, Michael W Heaven, Michael W. Zabarouskas, Michel luyskens, Michele Boldini, Micheline, Michelle, Michelle D. spinoza, Michelle Gregory, Michelle Lucas, Michelle Moore, ichelle williams, Michkov, Miguel Angel Ag ¨uero Romero, iguel Gonz ´alez, Miguel Guilherme, Mihaela, Mihai Str ¨ompl, ika Sadikario, Mikayla Creitz, Mike, Mike Barrett, Mike Barton, ike from Michigan, Mike Stirling, Mike Waldner, Mike Young, ike4Tammy, Mila Mucha Jana Moleman, Pablo Moleman, Miles ohnstone, Miles Whatcott, Millicent Switzer, Millie Evers, Miss Fox ames, Miss. Gunaretnam Keerthana, Missybee35, mitch, mitchell piller, Miłosz, mmilson, mohamed kamesh, Mohammad Odeh, OHAMMAD SAMEER, Mohammed Mahabu Subhani, Mohmad harif Jamali, Monica E., Monte J. VanDeusen, Mor gan, Mor gan rimes, Morgan Jensen, Mori Joseph, Moses Fleischman, Mouad erfoufi, Mr L, mrakes, Mreowwww, Mrityunjay Saxena, mroeling, sfletcher, mshimohi, MUHAMMAD AMIRUL AMIN B. ABD AHIM, muhammadmustaf a, Mustaf a, Myar Misellati, Myranda eightley, M ´arcia Saori C ˆ amara Kishi, M ´arjorie Mour ˜ ao, N. J. eyland Elsom, Nadine Trimpin, Naomi K. Lee, naslund, Nastya uliko va, Natalia, Natalia L ´opez Guzm ´an, Natalia Pirogova, Natalie ones, Nataliia, Natasha A Frakes, Natasha Van Bemmel, NateA, athalie Amador, Nathalie Smits, Nathaly Hernandez, Nathan Coet- ee, Nathan Fontanez, Nathan Scalzone, Nathan Yusuf, nathancbell, a vahra Lindsay, Na vaneethakrishnan, Na veen kumar S, nchazarra, ebogipfel, Neil Cottrell, Neil Kenneth Burton, Neil Sharp, Neil ick ens, Nelvish Naik er, Nerine Storm Thompson, Ne v aeh, NeZur, g Wen Xuan, Nica Roseme, Nicholas D. McGranahan, Nicholas unn, Nicholas Garrett, Nicholas Hans, Nicholas Hertz, Nicholas Bianchi, Nicholas Phu, Nicholas Sheppard, Nicholas Yarbrough, ichole Coryell, Nick, Nick CUmmings, Nick Weilert, Nicky- an, Nicola Bird, Nicolas Agust ´ın Menillo, Nicolas Cano, Nicolas amarche, Nicole Boisvert, Nicole Meireles Gomes, Nicole P. Vogt, icole Parks, Nicole Wolff, Niels Anonymous, Nigel Branson, igel1960, nihar.mishra, Nik Ritchie, Nike Koper, Nikita Babin, ik o Strauch, Nik olaos Dimitrios Poulos, Nik olas Mair, Nikshey adav, Nisha, Nishant Deo, Noah Childers, Noam dos Santos Gomes Kilono va Seek er s 2131 C D A O P D B O G p P P F M B d H P L P P P p P C R D P M A m R R D R R T B R K J R L S k C M R t C R R R r R S K S S C S M S S S R K S P R s S d l n P S A S G S S s G s S s t t S S S N S P S s S M S n H T T w T t E T a T T t T R T t T M B u V V P D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024erqueira, NoctornalE , Nolan Reket, NonLinearTurtle, Nova and i v a Thapa Magar, No ´e Susol, Nymie Goat, Ob v erse, Olayeni nifowose, Oleksandra Pyshna, Olena Kravhuk, Olesya Shaturova, lg a, Olg a Stehle, Olg a Stikhina, Oli ver Linnhof f, Oli ver Lykke etersen, Oli ver ROSSA, Oli via Bransford, Oli via Elgar, Olwenn aniel, Om Rohith Karamala, Ond ˇrej Koub ´ık, Onni Tengstr ¨om, Oren uerke, Origami99, Ornella Guyet, OSCAR MEANI, Oscar Rivera, skar Blachut, Owen Pearman, Pablo Gonzalez sepulveda, Pablo ´omez Alcojor, pablo.rpo val, P aige Schedler, P aloma, P amela Allen, ammins, Panthi Barot, Paramvir Kasana, Parker, Parker Betts, arth bha wsar, P ascal Delcroix, P atrece Vreeswyk, P atricia Boyce, atricia Fernandez, P atricia Kalimootoo, P atryk Kra wczy ´nski, P aul uentes, Paul Kirkland, Paula Sofia Paz Garcia, Paulo Gomes onteiro, paul w ay, pavan, Payson Clausen, peadams1992, Pedro erges, Pedro Garcia-Lario, Pedro Gon c ¸alves, Pedro Guilherme e Barros Silva, Pedro Ricardo Balata dos Santos Costa, Peggy all, Pepito, Perci v al Nollido, PERNEY CLEMENTINE, Perry Lin, ete Hermes, Peter Ansorge, Peter Barker, Peter Holmes, Peter ange, Peter Langohr, Peter Wuestner, petersamuels, petestmarie, etika, Peyton Sykes, philip.cross, Philipp Wenzler, Philippe Reclus, hilipWang, Phillip Chen, Phoebe Lister, Phuong Duong, picerx38, ietro Maiorana, pine gro v e, Piotr Małek, Piotr Zaja ¸c, Piotr ´Sojko, ippa brown, pitcher12k, pitchtwit, pjaj, planetari7, plent2d, Polina, oo v endhan Mathiyazhagan, PowlOwl, prachi, Prajwalita Jayadev ha van, Prana v Gurusubramanian, Prasad Vengurlekar, Pratibha athore, Praza Kembaren, Prima Rosa, Princess Keila Nu ˜ nez, Pritam utta, Pritam Raikar, Priyansh Sharan, Prohor Thomas Polipartov, roWang, Przemek, PsY4aLL, Punyae Bhatia, P ´eter Megl ´ecz, Quaid ilholen, Quentin Laydecki, Quinn Torgerson, R. Hoornstra, Rachel . F ord, Radosla w Siewierski, Radosła w Dobrzycki, Raella Roth- an, Rafael Cifuentes, Rafaela Carolina Schmitt, Raghav Mishra, ainey Waddell, Rajalekshmy, Ralph Allen, Ramesh Subramanian, amloo, Ramona Valkov ´a, Raphael, Rathika, ray van dycke, Rayen ahmoul, readwin, rebecca Robinson, red mtn, RedTrev, Reem, efugeZero, Reik o.Kaw ashima, Reina bonilla, Renee Appelman, en yuqing, Re x Mundi, Re y Tolentino, rfsmith, Rhea Porter, Rhodri homas, Ricardo, Ricardo Costa Silva, Ricardo Meloney, Richard unche, Richard Cheal, Richard Edwardson, Richard Mukerji, ichard perry, Richard Ryan Rote, Richard Wambach, Rick de oster, Ricky Grimaldi, Riku J. Laukka, Riley McManus, Rita ´on yer, Riv ers Langston, rmhoek, Rob Ash, Rob Woodings, Robbie, obert Apollo Pepper, Robert Auburger, Robert Deering, Robert G awless, Robert Jan Bauer, Robert Mazarelli, Robert Rice, Robert tone, Roberta Reed, Roberto Haver, Robyn Wilson, Rochelle Ad- ins, rochelle thomas, rodnjes, Rodrigo De Souza janssen, Rogelio alyecac, Roger Blake, Roger Feo, Roger Gagnon, Roiban Elena onica, Rolf Agterhuis, Rolf Arndt, Ron Usry, Ronan Lucchesi, onnie Payton, Rory Campbell, Rory O’Connell, Rosemary Billing- on, Rosemary Patyus, Rosyid Haryadi, Rouanne Bohg, Rowan allahan, Row an Lavender, Ro yce Edw ard Sayers, Rubru, ruby, udra Protap Nandi, Rui Augusto Barbosa da Silva, RUMOR47, upal Maitra, Russell Fulton, Russell Klare, Ruth McCabe, Rutten, uud van Kruchten, Ryan, Ryan Hazen, Ryan J, Ryan Tansey, yder, Rylie Gipson, Rylie Roche, Ryne Kreitz, R ¨onn Sch ¨uck, S amya, S Uzel, Sabine Vollenhofer-Schrumpf, Sabnam Shrestha, abrina west, Sagarika Menon, Sage Alyxander, Sai Shankar, Saif hwaja, Saint.Maria, Sallyann Chesson, Salv ador, Salv ador Araujo, am C., Sam Coletta, Sam L, Sam Ro wclif fe, Sam Scarborough, am Swaim, Sam Watson, Samantha Anderson Mcauley, Samantha o v er, Samantha Crossfield, same, Sameh Faty, Samuel Anderson, amuel Edward Lewis Halstead, Samuel hasley hunwick, Samuel aragno Predolim, samuel rodrigues henrique, Samuel Williams, anchayan Sarkar, Sandra Lee Harris, MDiv, Sandro Montagner, anjith, Sankuratri, sanra, Saoirse Cox, saphira, Sara Gankhuyag, ara Kindschus, Sara R, Sarah Grissett, Sarah Mae Raybould, Sarah . Kleinman, Sarah Starks, Sarah Williams, Sarathkumar, Sari-Lynn erkhofs, Sasha Gdaniec, satsig, Satyajit Gro v er, Saumya Garg, aurao M Upare, Sauro Gaudenzi, Savannah Gamache, Savannah owell, sbfoulkes, scarab64, Scott C Evers, Scott Nantais, Scott ogers, seadour, Sean Hohmann, Sean M. O’Brien, Sean Miller, ebastian abbatoy, Sebasti ´an Alejandro Freigeiro, See Moua, Seema hroff, Sema DEM ˙IRBA ˘G, Seojin Park, Seppo Hassinen, Sergey Se- ov, Ser gio, Ser gio Roda, Serkan Alaba s ¸, Serkan Kadi, Shadae Tay- or, Shaik nissar hussain, Shakaraliye v a Leila, Shaleena Turner, shal- achywyt, shane Masterson, Shannon Adams, Shannon Kindilien, hD, Sharad Pednekar, Sharon Fulton, Sharuyan, Shaul Lindblom, haun Coulon, Shawn Phelps, Shehab Khidr, Sheree Gillett, Shirley- nne McIntyre, Shi v, Shi v am Kumar Sharma, Shi v am Saxena, hi v ani Hancock, shocko61, Shogo Noma, Shotoya Simmons, Shree hosh, Shre y Ajaykumar Bhatt, Shre yas G Kamath, SHRIRAM, huming Wang, Sian Oleary, Siddharth, Siddharth Kuchimanchi, iddharth Mehrotra, Sidhant Hargunani, Siegfried Zanier, Siham, imek, Simion Ochieng Agumba, Simon, Simon Barker, Simon ardner, Simon Hennessey, Simon Lund Sig Bentzen, simon nichol- on, simone marcorini, Simran Patel, Sinai Coons, Si v apriya Arvind, k yboy777, sk ywarrior, slebede v a, smartalice, smith7748, snafu04, nappa, snikdekl, Sniwa, Sofia, Sofia G, Sof ´ıa Lorena V ´azquez Mon- o ya, Soham Thak er, Sohini Mukherjee, Somia Ditor, Somobrata Mi- ra, Sonn y Black, Son ya Singh, Sophia, Sophia Ottner, Sophia Rehan, ophia Rencher, Sophia Shannon, Sophie Whitle y, So van acharya, pencer Emmons, Srijan Banerjee, SRIJITA BISWAS, Sristi Betal, rushti Pardhi, Stabitha Cru’ktfinger, Stacey E. Farr, Stacy Walker- ez, starchitect, starsjosh, stasxh, Stephanie, Stephanie Colbert, tephanie mosher, Stephanie Ochoa, Stephanie Santo, Stephany alacios Martinez, Stephen Cole, Stephen Gazard, Stephen Kiraly, tephen Molatlhegi, Stephen Patrick Brennan, Stephen R. Hyzer, te ve, Ste ve Davies, Ste ve Lake, ste ve meado ws, Ste veF48, ste vekrz, tev en Rosander, Stick ywarp, Stitch1981, Stukalin Danil, Subhas azumdar, Sudeeksha Bhattacharyya, surhudm, Susanna Dargan, usannah Bispham, Svetoslav Alexandrov, svr303, swaas, SY, Syd- ey Letlow, Syed Hozaifa Shabur, Syed Muhammad Abdullah Bin asan, Systrie, Tahseen Huda, Taira Agosto, tairving, Tan Poh Ching, ania Ramirez, tania.ann.007, Tanishk Mohan, Tara, Tarant0ga, arpon Thompson, T asfia T arannum, T atyana Reid, T aylor Ben- are, tdenton, tej Rajesh, tenfyr, Teodor Michalski, Terdoo Osu, errance P Felegie, Sr, Terry Georgas, Thaliyadath K Ravindranath, hecrazedlog, Theresa, Theresa Carl, Thomas BERTRAND, Thomas Collett, Thomas Hoffmann, Thomas Jenkins, Thomas Kl ¨ager, homas Koceja, Thomas Laube, Thomas massey, Thomas R. Stew- rt, Thomas S Grady, Thomas Thomopoulos, thorsteinn, threepwood, ia Rice, T iep T ran, T iffan y Sha w-Diaz, T im Dyson, T im Jaramillo, im Loris Kunze, Tim Plank, Tim Pointing, Tim-Christopher Aust, iorriot, tireniolu Oladipo, tkuhnle, Tobias G ´eron, tobidd, Tobrux, oby Leeper, T om, T om Bickle, T om Davison, T om Osman, T om ich, Tomas Koval, Tomassci, tomjon, TomThaler, tony, Tony, ony Grant, T ony Hoffman, T orsha Kundu, T orsha Majumder, petera, T randafir Mircea, T rapeznikov Egor Valerevich, trash0, ravis Olah, T ravis Rector, T rent Pierce, T ricia, tripp yskipp y, Trish artinache, Tuesday Muse, TULSI GOYAL, Tyler Andreww, Tyler estram, Tyler Hirshberg, Tyronius McBasketball, Tyson Dabney, dayatejwani, Udunie, Uliana Shiliae v a, Uriel, Utsav Khan, Utsho, ahavy, V ahid Kermani, V aideeshwar Si v asubramanian, Vajrapani, alentina, V aleria Tokare v a, Valerie Blanco, v alerie flores, Valerie e gg, Valerie R se ymour, Valmir Martins de Morais, Vandytim,MNRAS 533, 2113–2132 (2024) 2132 T. L. Killestein and L. Kelsey et al. M V V D V V W L W w D w d D H W S s Y Y Z T F Z d C r А Д Н  1 Department of Physics and Astronomy, University of Turku, Vesilinnantie 5, Turku FI-20014, Finland 2 Department of Physics, University of Warwick, Gibbet Hill Road, Coventry CV4 7AL, UK 3 Institute of Cosmology and Gravitation, University of Portsmouth, Portsmouth PO1 3FX, UK 4 Department of Physics and Astronomy, University of Shef field, Shef field S3 7RH, UK 5 School of Physics and Astronomy, Monash University, Clayton VIC 3800, Australia 6 Instituto de Astrof ´ısica de Canarias, E-38205 La Laguna, Tenerife, Spain 7 School of Physics and Astronomy, University of Leicester, University Road, Leicester LE1 7RH, UK 8 Armagh Observatory and Planetarium, College Hill, Armagh BT61 9DG, UK 9 National Astronomical Research Institute of Thailand, 260 Moo 4, T. Donkaew, A. Maerim, Chiangmai 50180, Thailand 10 Jodrell Bank Centre for Astrophysics, Department of Physics and Astron- omy, The University of Manchester, Manchester M13 9PL, UK 11 Departamento de Astrof ´ısica, Univ. de La Laguna, E-38206 La Laguna, Tenerife, Spain 12 School of Physics and Astronomy, University of Birmingham, Birmingham B15 2TT, UK 13 Department of Astrophysics, American Museum of Natural History, Central Park West and 79th Street, New York, NY 10024-5192, USA 14 School of Physics, Kane Building, University College Cork, Cork, Ireland 15 Department of Astrophysics/IMAPP, Radboud University, Postbus 9010, NL-6500 GL Nijmegen, The Netherlands 16 School of Sciences, European University Cyprus, Diogenes street, Engomi, 1516 Nicosia, Cyprus 17 Centre for Astrophysics Research, University of Hertfordshire , Colleg e Lane, Hatfield AL10 9AB, UK This paper has been typeset from a T E X/L A T E X file prepared by the author. D ow nloaded from https://academ ic.oup.com /m nras/article/533/2/2113/7735340 by Turku U niversity user on 30 August 2024 anessa Eliseo, vanrock70, V arooni Manoj Sawant, V arsha, V asileios lachakis, Vasiliy, Vasylenko A., Vaughan, venatrix, Venkatesh eshpande, Vern Sowers, Viacheslav Zelenev, Victor, Victor Celli, ictor Cunha Da Silva, Victor Edwin, victor juan garcia porcel, ictoria Jackiw, V ikrant Kurmude, V iktor, V iktor Dobrenov, V iktor ase, Vincent Hobe ¨ıka, Virgilio Gonano, Vishwanand Doobay, Vitor uis Gon c ¸ales Dias, vivian, Vlad, Vlad Mihai, Vladimir, volare09, alker W ells, W alter E Moody Jr, W alter MacDonald, wangqintao, arrenchen, Weatherly-Battaglia, WEBs in space, WELTON VAZ E SOUZA, Wenceslao Santiago Germ ´an, Wendy, Wendy Smith, endy27, W enjie Zhou, W entao Huang, W eronika, W esley Gabriel e Oliv eira Melo, Wesle y Teh Ee Wen, wesley webb, wgoltz, Wilfried omaink o, Wilfried Esk en, W ilhelm A. Weidmann, W ill, W ill aresch, William B Hernandez, william birney, William Midgley, illiam Paul Rhodes, William Russell, Willow Colson, Windsor S myser, WOJCIECH MIKINA, W olverineW azza, wrackard, Xander prangers, Xavier Dartevel, Xenon Chase, xiaoguangliu, xKingx, aBoy, Y anilsa H Marte Rodriguez, yanyam, Yavin 4, YGBS, ogurtz, Y ohan Terpend, Y olly Reyes, Y ounes Oubkis, Yufan Fane hou, Yun Chen, Yuri Peruzzi, Yuri Sushkov, Yvonne Harrison, Zac homas, Zaccaria Vidali, Zach Ortega, Zach Schroeder, Zachary leisig, Zachary Thede, Zachary Vaillancourt, zacho, zak bennett, akck Goolsbee, zan, Zander Polk, Zarriah Fisher, Zden ˇek Flan- erka, Zebraorpanda, zedcat, zena barabandi, Zenith Diehl., Zhiyuan heng, Zijun He, Ziyun Tang, Zoe Kateri, Zoe Vickhouse, Zo- yana, Zovacor, ´Agata Moretti Zaneti, ´Iris Eduarda Forcel Roncada, лександр, Владимир Асташин, Даниил, Дмитрий Дуда, митрий Сергеевич Фёдоров, Евгений Юрченко, Матвей, иколь Вельковская,  ,  ,   ,   ,  , ,  NRAS 533, 2113–2132 (2024) Published by Oxford University Press on behalf of Royal Astronomical Society. This is an ( https://cr eativecommons.or g/licenses/by/4.0/), which permits unrestricted reus© 2024 The Author(s). Open Access article distributed under the terms of the Creative Commons Attribution License e, distribution, and reproduction in any medium, provided the original work is properly cited.