Multi-Messenger Astronomy

Established Multi-messenger astronomy measures one astrophysical event through two or more carriers — photons, neutrinos, gravitational waves, cosmic rays — and treats the coincidence itself as the observable. Each carrier is made by different physics, absorbed by different matter and arrives on a different clock, so a joint detection constrains quantities no single channel reaches: the propagation speed of gravity, the production site of the heaviest elements, the engine inside a burst whose photosphere hides it.

Established The field has three founding events and they are still doing most of the work. SN 1987A delivered roughly two dozen neutrinos across three detectors and then an optical supernova. GW170817 delivered a gravitational wave, a short gamma-ray burst 1.74 ± 0.05 seconds later, and a kilonova in NGC 4993. IceCube-170922A delivered a neutrino near 290 TeV pointing at a flaring blazar. Two are nine years old or older; the third has never reached discovery significance.

Frontier The claim this brief lands is that the binding constraint is statistical and logistical, not instrumental. Gravitational-wave detections are routine — 390 logged since 2015 — and wide-field survey capacity has grown by more than an order of magnitude. The count of binary neutron star mergers with a confirmed electromagnetic counterpart has stayed at one. What fails is the chain between alert and counterpart: sky area, cadence, spectroscopic access, and the arithmetic of chance coincidence in a sky full of unrelated transients.

Established The discipline’s real unit of work is the trials factor. A coincidence is evidence only if you can state how many chances the universe had to produce it by accident: how many alerts were followed, how large the search area was, how many catalogues scanned, how many time windows tried. Every headline association here lives or dies on that number, and the messenger communities do not share a convention for reporting it.

Established A note on sourcing. This brief was commissioned in September 2026 from the Institute’s research base. Reading-list entries without links are cited from the bibliographic record rather than re-fetched, and claims are dated no later than early 2026 unless carried by a linked source.

Established GW170817 is the measured spine of the subject and the numbers are unusually clean. On 17 August 2017 a binary neutron star inspiral was recorded by the two LIGO detectors and Virgo; the Fermi Gamma-ray Burst Monitor independently triggered on GRB 170817A 1.74 ± 0.05 seconds after the merger time. The three-detector network localised the source to roughly 28 square degrees at 90% credibility, tightened to about 16 in the final analysis, at a luminosity distance near 40 megaparsecs. An optical transient was found in NGC 4993 within about 11 hours, ultraviolet emission faded within two days, X-rays appeared at nine days and radio at sixteen. More than seventy collaborations and facilities took part.

Established That single event produced four results that stand on their own. The gamma-ray delay bounds the fractional difference between the speed of gravity and the speed of light to between roughly −3 × 10⁻¹⁵ and +7 × 10⁻¹⁶, removing entire families of modified-gravity models built to imitate dark energy. The kilonova light curve required roughly 0.05 solar masses of neutron-rich ejecta with both lanthanide-poor and lanthanide-rich components, establishing mergers as a rapid-neutron-capture site. Radio interferometry later resolved apparent superluminal motion of 4.1 ± 0.5 c, settling the geometry as a structured relativistic jet seen off-axis. And the event served as a standard siren, giving a Hubble constant of 70.0 (+12.0/−8.0) kilometres per second per megaparsec — an independent distance measurement with an error bar too wide to arbitrate anything.

Established Since then the gravitational-wave side has scaled and the counterpart side has not. The LIGO–Virgo–KAGRA catalogue reached 218 events through GWTC-4.0, which added 128 candidates from the first part of the fourth observing run at astrophysical probability 0.5 or above. GWTC-5.0, released 26 May 2026, reports 390 total detections since 2015, 161 of them between 10 April 2024 and 28 January 2025, with the fourth run accounting for about three-quarters of everything ever found. It contains the sharpest localisation yet, GW240615 at six square degrees, and the loudest signal yet, GW250114 at signal-to-noise 76.9. Confirmed kilonovae in that entire catalogue: one, from 2017.

Established The drought is explainable, which makes it a constraint rather than a mystery. Almost all detections are binary black hole mergers, which are not expected to shine. The second neutron star merger, GW190425, arrived with a 90% localisation near 8,000 square degrees because only one detector was fully operational, and no counterpart was ever identified. Neutron star mergers are rarer and their horizon far closer, so the detectable volume grows slowly with sensitivity while localisation area shrinks only as network geometry allows.

Established On the neutrino side the diffuse astrophysical flux is settled and the sources are not. IceCube established an extraterrestrial flux above atmospheric backgrounds in 2013 and pushed it past five sigma shortly after; the per-flavour spectrum is a power law with index between about 2.4 and 2.9 depending on which event sample is used, and that spread is a real disagreement between cascade-dominated and track-dominated analyses rather than a measurement uncertainty. Frontier Less than a few per cent of that flux has been resolved into identified sources.

Frontier The three best neutrino source claims sit between three and four and a half sigma, and none is a discovery by particle-physics convention. The 2017 alert coincident with the flaring blazar TXS 0506+056 disfavours chance at about 3.0 sigma after trials. An archival search of the same direction found roughly 13 ± 5 excess events in 2014–2015 at about 3.5 sigma — when the blazar was not gamma-ray flaring, which is awkward for the simplest model rather than supporting of it. The Seyfert galaxy NGC 1068 shows about 79 (+22/−20) events at 4.2 sigma with no matching gamma-ray signal, implying a source opaque to TeV photons. An excess along the Galactic plane reached about 4.5 sigma. Four independent near-misses are a pattern, not a detection.

Frontier The most energetic neutrino ever recorded arrived in a detector that was one third built. KM3NeT’s ARCA array registered KM3-230213A on 13 February 2023, published in 2025: a muon of about 120 (+110/−60) PeV implying a parent neutrino near 220 PeV, with a 90% range of roughly 72 to 2,600 PeV. IceCube, with far more exposure over more years, has seen nothing at that energy, and the two results are in tension at roughly 2.5 to 3 sigma. A single event cannot distinguish a fluke from a new population.

Established The negative results carry as much weight as the positives. GRB 221009A in October 2022 was the brightest gamma-ray burst on record; IceCube found no coincident neutrinos, which limits how much burst energy can go into hadronic channels. Repeated optical campaigns on gravitational-wave triggers through the third and fourth runs returned no kilonova. Nulls of this kind are the discipline working correctly, and they are systematically under-published.

Frontier Whether standard sirens can arbitrate the Hubble tension is the largest open question the field owns. The tension is between roughly 73 and roughly 67 kilometres per second per megaparsec from distance-ladder and microwave-background methods, with quoted uncertainties around one per cent. A siren measures luminosity distance directly, with no calibration ladder. GW170817 did that once, to about fifteen per cent. Reaching the two per cent needed to adjudicate requires of order fifty to a hundred events with identified hosts. At one per nine years, that is not a programme; it is a hope.

Frontier Statistical or “dark” sirens are the proposed escape and their systematics are unsettled. Without a counterpart, one marginalises over all catalogued galaxies inside the localisation volume. This has produced constraints near 68, but the answer inherits the incompleteness of the galaxy catalogue at the relevant distance, and nobody has shown the incompleteness correction is unbiased at the precision required.

Frontier The post-merger remnant has never been detected and it is where the nuclear equation of state lives. After two neutron stars touch, the remnant rings at one to four kilohertz for tens of milliseconds — exactly the band where current interferometers are worst, because photon shot noise rises with frequency. Measuring it would pin the pressure of matter above nuclear saturation density, a regime no accelerator reaches. Inspiral tidal deformability from GW170817 already excludes the stiffest equations of state; the remnant signal is the direct measurement.

Speculative A massive black hole binary seen in both gravitational waves and light is the flagship case for space-based interferometry. A merger of 10⁵ to 10⁷ solar masses sits in the millihertz band, is visible for months before coalescence, and could be localised in advance well enough for telescopes to be watching. Whether it emits light at all depends on gas being present, a modelling assumption rather than an observation.

Established Sky localisation is the first bottleneck and it is set by network geometry, not by sensitivity. A source is localised by arrival-time differences between widely separated detectors: two give an annulus, three break it into patches, four shrink the patches. Hence GW170817 on 28 square degrees and GW190425 on 8,000. Raising the sensitivity of an existing detector raises the detection rate and makes the typical localisation worse, because it pulls in more distant events whose signal-to-noise is lower in each instrument.

Established The kilonova cadence problem is a real tension between three requirements that cannot all be relaxed. A kilonova peaks near absolute magnitude −16 and fades roughly one magnitude per day in the optical, faster in the blue. Catching one at 200 megaparsecs means reaching 22nd to 23rd magnitude over hundreds of square degrees and returning within a day. Depth, area and cadence trade against each other on any single telescope; Rubin is the first facility where all three are simultaneously plausible, and even there it means interrupting the main survey.

Frontier The scarce resource is not imaging but spectroscopy. A wide-field imager can produce a hundred candidates in one error region in one night. Deciding which is a kilonova requires a spectrum from an eight-metre-class telescope, scheduled months ahead and heavily oversubscribed. In 2017 the community suspended normal operations on multiple large telescopes for one event. That is not repeatable at several alerts per week, and the institutional fix — standing target-of-opportunity allocations on large spectrographs — exists only patchily.

Established Alert latency is now good and alert purity is the limiting quantity. Preliminary gravitational-wave alerts are public within roughly a minute, with pre-merger early warning possible for the closest binaries, and IceCube dispatches its best track alerts on a comparable timescale. The cost is that a meaningful fraction of automated alerts are later retracted, and every retraction consumes follow-up time that cannot be recovered.

Established Data triage at survey scale is a filtering problem of a different order. A modern wide-field survey generates of order a million alerts a night; Rubin is designed for around ten million, at a target of sixty seconds from shutter close to distribution and on the order of twenty terabytes per night. No human inspects that stream. It is reduced by independent brokers running machine-learning classifiers whose selection function — what they systematically discard — is uncharacterised for rare transient classes, precisely the class kilonovae belong to.

Frontier There is no shared significance convention across the messengers, and this is a real defect. Gravitational-wave astronomy reports a false-alarm rate in inverse years plus a per-event astrophysical probability; neutrino astronomy reports a post-trials p-value converted to sigma; optical transient astronomy often reports neither. Combine a three-sigma neutrino coincidence with a two-sigma optical flare and the joint statement is frequently not a well-defined probability of anything.

Established Interferometer sensitivity depends on two materials problems that have resisted a decade of work. Mirror coating thermal noise limits the mid-band, where neutron star binaries live, and the amorphous tantala-silica coatings in use have improved only incrementally; the world’s supply of large-aperture low-loss coatings runs through very few facilities. Quantum noise is now addressed by injecting frequency-dependent squeezed light, which works and buys tens of per cent rather than a factor.

Frontier Kilonova interpretation depends on atomic data that mostly does not exist as measurements. The opacity of lanthanide-bearing ejecta is computed from theoretical line lists of millions of transitions, because laboratory spectroscopy of most lanthanide and actinide ions in the relevant ionisation states has not been done. Strontium was identifiable in the GW170817 spectrum because strontium is light enough to have good laboratory data. This is the least glamorous dependency in the subject and arguably the tightest: ejecta masses quoted to two significant figures rest on opacities with factor-level uncertainty.

Established Everything downstream depends on numerical relativity waveform banks. Parameter estimation compares data against template families calibrated to supercomputer simulations. Systematic error in those templates propagates directly into masses, spins and tidal deformabilities, and for the loudest events waveform systematics are already comparable to statistical errors.

Established Neutrino astronomy depends on the optical properties of a cubic kilometre of natural ice. IceCube reconstructs direction and energy from photon arrival times in glacial ice whose scattering and absorption vary with depth. Angular resolution for tracks is around half a degree at high energy and much worse for cascades, and source identification is limited by that resolution as much as by statistics. Frontier Dark-siren cosmology depends similarly on an all-sky galaxy catalogue complete to the relevant depth — unglamorous survey work with no single owner, whose incompleteness is modelled rather than measured.

Frontier The single result that would most change this brief is a second binary neutron star merger localised tightly enough that a kilonova is found and a host redshift measured. It would convert the counterpart rate from a quantity estimated off one detection into a measured number, test whether AT2017gfo was typical, give a second independent standard siren, and show that the alert-to-spectrum chain works under routine conditions rather than the emergency mobilisation of 2017.

Frontier The window is the fifth observing run, and the date has moved repeatedly. The fourth ran from May 2023 through the middle of this decade with successive extensions; the fifth is planned for the late 2020s at improved sensitivity, with the localisation benefit depending on how many detectors run simultaneously. A four-detector network localises far better than a three-detector network with one instrument down, and the difference between those states is budgetary rather than technical.

Established The second decisive test needs no new hardware and no new funding, because it is already running. A core-collapse supernova in the Milky Way would deliver a neutrino burst minutes to hours before optical brightening, and the inter-detector coincidence network built for this purpose would issue a pointing alert in that interval. The expected rate of roughly one to three per century makes it the purest natural experiment in observational astrophysics and entirely unschedulable. SN 1987A gave about two dozen neutrinos from 50 kiloparsecs; a Galactic event at ten kiloparsecs would give thousands, enough to resolve the explosion mechanism directly.

Frontier The third test is one the field could run deliberately and mostly has not: a pre-registered follow-up protocol with a stated trials factor. Fixing in advance which alerts are followed, which catalogues searched and which time windows count would turn recurring three-sigma coincidences into a properly normalised population statement. Pre-registration reduces the chance of a spectacular single result, which is what the incentive structure rewards.

Frontier A fourth test settles the ultra-high-energy neutrino tension by arithmetic. If the flux implied by the single 220 PeV KM3NeT event is real, a completed cubic-kilometre array in the Mediterranean plus continued exposure elsewhere should produce further events within years. If none appears, the original event was a fluctuation.

Established The detection hardware sits at the edge of what mechanical engineering permits, and the specifications are public. A LIGO detector uses four-kilometre evacuated arms inside a vacuum volume of about 10,000 cubic metres, among the largest on Earth, and resolves a change in mirror separation of roughly one ten-thousandth the width of a proton.

Established Third-generation designs do not refine that architecture, they restart it. The European proposal is a triangular underground observatory with ten-kilometre arms, pairing cryogenic low-frequency interferometers with room-temperature high-frequency ones; it remains in site selection between Sardinia, the Euregio Meuse-Rhine and Lusatia, with groundwork reported as of August 2026. The American proposal is a surface L with arms of tens of kilometres. Both are civil-engineering projects whose cost is dominated by tunnelling and vacuum, not optics.

Established The survey side is a data-handling machine with a telescope attached. Rubin pairs an 8.4-metre primary with a 3.2-gigapixel camera covering about 9.6 square degrees per pointing, with an alert pipeline designed to publish within about a minute. Cost was about $680 million; first light 23 June 2025, full survey operations 30 June 2026. The aperture and cost scaling governing the whole ground fleet is treated in Mega-Telescopes.

Established The coordination layer is ordinary software engineering doing load-bearing scientific work. Alerts move as structured messages over publish-subscribe transport, through the successor to the gamma-ray coordinates network and through community brokers, against a shared transient naming registry. The most consequential decisions of the past decade — who may publish an alert, how fast, under what retraction policy — were made here.

Established The follow-up capacity this subject depends on is owned by the large-aperture fleet. Mega-Telescopes carries the cost-scaling result that explains the short spectrograph queue: construction cost rises as roughly the 2.7 power of aperture, so the eight-metre class will not be casually duplicated for target-of-opportunity work.

Established The detection-versus-emission asymmetry belongs to a sibling brief and bounds this subject. Gravitational Wave Engineering records 390 confirmed detections against zero devices ever built, and the reason: nothing couples to the gravitational channel at laboratory scale. Multi-messenger astronomy is therefore permanently observational, never experimental in the accelerator sense.

Established Propagation tests across messengers are a shared frontier with fundamental physics. Quantum Gravity carries the constraint programme built on exactly these coincidences — photon time-of-flight from bursts, neutrino flavour ratios over cosmological baselines, gravitational-wave dispersion — and reports that every such test has returned a tightening null. Black Hole Physics Applications carries the compact-object census and the two imaged horizons.

Frontier Two adjacent subjects share the plumbing rather than the physics. Planetary Defense depends on the same wide-field stream for near-Earth object discovery, and Interstellar Archaeology runs commensally on the same telescopes. The broker and classifier layer is an applied case of the pattern in Artificial Scientists: automated selection works where a cheap exact verifier exists downstream, and a kilonova spectrum is exactly such a verifier.

Established The most important institutional change of the past decade was making gravitational-wave alerts public by default. Through the second observing run, alerts went to partners under memoranda of understanding with embargo terms; GW170817 was handled that way. From the third run onward, candidate alerts have been published openly and immediately, converting follow-up from a membership privilege into an open competition and materially widening participation.

Established Authorship at this scale has broken the conventional model and nobody has replaced it. The GW170817 multi-messenger paper carried several thousand authors across more than seventy collaborations. There is no accepted way to assign individual credit inside such a list, which affects hiring in a field where early-career participants do most of the observing.

Frontier The forward programme is funding-limited at every node, and the risks are correlated rather than independent. Third-generation interferometers, a next-generation neutrino array, a space mission and continued survey operations compete in several cases for the same national budgets. Established The concentration risk is concrete: a two-site interferometer network that loses one site stops being able to localise. Budget proposals in the mid-2020s contemplating reduction of one observatory to a single site were, whatever their fate, a demonstration that localisation sits on a political variable rather than a technical one. Scientific Funding Models carries the general finding that funding design changes what scientists attempt.

Established Standards work has been done and it works. Alerts share a machine-readable schema lineage, transient names are issued by a single registry so independent discoveries can be reconciled, and open archives publish calibrated strain and neutrino event samples after defined proprietary periods — the reason a coincidence found by one group can be checked by another.

Frontier What is missing is a body that can commit follow-up resources. No allocation committee is empowered to interrupt multiple independent observatories on a shared trigger; each facility decides individually, in real time, under local policy. For the one event where it mattered the community self-organised. Whether that scales to several alerts per week has never been tested, because the alerts have not yet been worth it.

Established Follow-up capability is distributed very unevenly, and the distribution is geographic. Rapid response requires telescopes at the right longitude at the right local time. Networks in the southern hemisphere and in Asia are structurally necessary for full sky coverage, and the institutions hosting them have historically had weaker claims on the resulting publications than the northern consortia that received the alert.

Frontier Satellite constellations degrade exactly the survey mode this field needs. Low-orbit constellations leave trails in wide-field long-exposure imaging; masking works at the cost of discarding pixels, and the affected fraction rises with constellation size. No regulatory instrument has jurisdiction over the night sky as a scientific commons.

Established Public alerts create a public retraction problem. Openly published candidates reach the press before vetting, so retractions read as failures rather than as the system working. The alternative — delaying until human review — costs precisely the hours in which a kilonova is bright.

Speculative No dual-use concern in this subject survives examination. Claims that gravitational-wave or neutrino infrastructure has covert military application recur; the coupling physics makes them arithmetic failures rather than policy questions, and the sibling brief on gravitational-wave engineering documents an official study that dismissed one such claim on those grounds.

Established The field has already changed a fact about the origin of matter. Before 2017 the astrophysical site of rapid neutron capture — the process making gold, platinum, uranium and roughly half the elements heavier than iron — was a textbook open question. One event moved it to a measured, if incomplete, answer. That is a rare civilizational output: a change in the provenance of ordinary material objects.

Established These instruments give access to states of matter no laboratory will reproduce. Matter above nuclear saturation density, fields of 10¹⁵ gauss and bulk relativistic outflows are not manufacturable, so the observatories are the only experimental access route — an argument for treating them as infrastructure of the same class as accelerators rather than as telescopes.

Speculative A Galactic supernova would be a civil as well as a scientific event. A publicly distributed neutrino-triggered warning of hours would put a naked-eye astronomical event on a published schedule for the first time in four centuries. Who issues that notice, with what confidence statement attached, has been discussed and not settled.

Handwave Claims that this infrastructure is a step toward practical control of the gravitational or neutrino channels do not survive the numbers. The argument that detection capability implies emission capability works by assertion.

These horizons track when the counterpart rate changes, not the detection rate — the detection rate is already solved.

• 10 yr: Frontier A four- or five-detector network at improved sensitivity, plus a wide-field survey with a standing interrupt allocation, should produce the second and third kilonovae and a handful of well-localised neutron star mergers. Expect the siren Hubble constant at the five-to-ten per cent level, narrowing the tension without arbitrating it, and one to three more neutrino sources between four and five sigma with continued argument about trials accounting.
• 25 yr: Speculative Third-generation interferometers, if built, detect essentially every neutron star merger in the observable volume, converting the problem from finding counterparts to choosing which to follow. Millihertz observation is operating and a first joint millihertz-plus-optical detection is plausible but not assured. Expanded neutrino arrays resolve a substantial fraction of the diffuse flux into sources.
• 50 yr: Speculative Standard sirens function as a primary cosmological probe with systematics independent of the ladder. The nuclear equation of state is measured from post-merger signals rather than inferred from inspirals. A Galactic supernova has occurred with probability roughly one in two, settling the explosion mechanism observationally.
• 100 / 250+ yr: Handwave Lunar-surface or solar-orbit interferometer arrays, detection of the cosmic neutrino background, gravitational-wave cosmology reaching before recombination: coherent extrapolations with no costed programme and no demonstrated enabling technology. Nothing constrains them except that none violates a conservation law.

• Depends on This subject waits on results owned elsewhere on the map. Follow-up capacity is bounded by the aperture cost scaling in Mega-Telescopes, which is why spectroscopic confirmation is the scarce step rather than imaging. The detection side’s ceiling is the coupling asymmetry in Gravitational Wave Engineering. The propagation-limit science these coincidences feed is owned by Quantum Gravity. No physics result blocks the observational programme itself.
• Requires (not on this map) A second qualified supplier of large-aperture low-loss optical coatings, because mid-band sensitivity is coating-thermal-noise limited and that supply chain is effectively single-sourced. Laboratory-measured opacities for lanthanide and actinide ions, because every quoted kilonova ejecta mass rests on theoretical line lists with factor-level uncertainty. Standing interrupt allocations on eight-metre-class spectrographs, because the 2017 response was an emergency mobilisation that does not scale to weekly alerts. Construction money for at least one third-generation ground interferometer, because localisation past the current network needs new facilities rather than upgrades. And a common convention for reporting trials factors across the three communities, because joint significance statements are currently not well-defined probabilities.
• Enables A cosmological distance measurement immune to distance-ladder systematics; a direct measurement of matter above nuclear saturation density; identification of the sources of the highest-energy particles known; and hours of warning for a Galactic supernova.
• Adjacent Black Hole Physics Applications for the compact-object census, Planetary Defense and Interstellar Archaeology for shared survey and alert infrastructure, Deep Space Communications for the data-transport constraint on space-based nodes, and Artificial Scientists for the automated-triage layer.

Frontier Claim: GW170817 proved neutron star mergers made the gold. It proved mergers make heavy elements, which is a smaller statement. Whether they make most of them over cosmic history depends on the merger rate, the delay-time distribution and the per-event yield, and competing sites remain viable. The abundance record in metal-poor stars currently favours more than one channel.

Frontier Claim: TXS 0506+056 was the first identified source of astrophysical neutrinos. The 2017 coincidence reaches about 3.0 sigma after trials and the archival 2014–2015 excess about 3.5 sigma, both below the five-sigma threshold particle physics uses for discovery. Worse for the simple picture, the archival flare occurred when the blazar was not gamma-ray bright, so the two pieces of evidence do not fit one steady-state model. The correct description is a strong candidate, not an identification.

Speculative Claim: the 220 PeV KM3NeT event reveals a new ultra-high-energy population or new physics. It is one event. Poisson uncertainty on a single count is of order the count, the implied flux is in roughly 2.5-to-3-sigma tension with non-detection by a longer-running instrument, and the honest reading is that either the event is an upward fluctuation or the other instrument has had a downward one. Only more exposure decides.

Established Claim: more gravitational-wave detectors improve localisation proportionally. The improvement is non-linear and geometric: two detectors to three collapses a degenerate annulus into small patches, a qualitative change; three to four is incremental. The operational corollary is that a three-detector network with one instrument offline degrades by orders of magnitude, not by a third.

Established Claim: since detections are routine, counterparts should be too. Almost all detections are binary black hole mergers, for which no emission is expected on any standard model; the one claimed exception, an active-galactic-nucleus flare proposed as the counterpart of a high-mass merger in 2020, is contested on chance-coincidence grounds. The relevant denominator is neutron star mergers, a handful, not the catalogue of 390.

Frontier Claim: five sigma is the standard of evidence in this field. It is the standard in particle physics and, through false-alarm-rate thresholds, effectively the standard gravitational-wave collaborations apply. It is not the operating standard for source associations in neutrino or transient astronomy, where three sigma post-trials is routinely reported as evidence. The prior odds genuinely differ, so this is not simple dishonesty — but the inconsistency is real and the trials factor often goes unreported.

Handwave Claim: the reported coincidences are mostly the look-elsewhere effect and the field is fooling itself. The sharpest external criticism, and half right: trials accounting for individual association claims has been inconsistent and several claims weakened on re-examination. But the diffuse neutrino flux, the gravitational-wave catalogue and the GW170817 coincidence are each independently above five sigma by their own communities’ conventions, and a gamma-ray delay of 1.74 seconds against a merger localised to one galaxy at 40 megaparsecs is not a look-elsewhere artefact by any accounting.