Pith. sign in

REVIEW 4 major objections 6 minor 31 references

This paper argues that autoencoders trained only on previously measured ultraperipheral-collision processes can flag rare decays and exotic hadrons—such as J/ψ→4π and a candidate pentaquark—as anomalies, with high purity and without assumin

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Autoencoders trained on simulated known UPC processes flag injected J/psi to 4 pi and pentaquark events with high reported purity in toy ALICE-like data.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A useful toy-level extension of the authors' earlier autoencoder proof-of-concept, but the high-purity and upper-limit claims outrun what the pT=0, no-detector simulation actually supports. the 4 major comments →

arxiv 2508.21728 v1 pith:K5YD2GS3 submitted 2025-08-29 hep-ph hep-ex

AI-powered full-data set search for new physics in ultraperipheral and diffractive events

Simone Ragoni , Brianna Kinkaid , Janet Seger , Christopher Anson , David Tlusty This is my paper
classification hep-ph hep-ex
keywords anomaly detectionautoencodersultraperipheral collisionsdiffractive eventspentaquark searchrare decaysparticle identificationexclusion limits
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper proposes anomaly detection as a model-agnostic search strategy for rare decays and new physics in ultraperipheral and diffractive collisions. Two autoencoder designs are trained exclusively on toy samples of processes that have already been measured—dilepton and dikaon continua plus charmonium decays—then rare channels are injected at rates estimated from known cross sections, branching ratios, and experimental upper limits. Both models flag the injected rare processes as anomalous, selecting J/ψ→4π with high purity and flagging every injected pentaquark with no contamination. If the toy cocktail faithfully mirrors the experimental environment, this would let experiments search for unknown signatures without predefined decay topologies, and turn the absence of anomalous events into upper limits on exotic production.

Core claim

The central claim is that a neural network trained only on known exclusive processes can recognize rare or never-before-seen processes purely by reconstruction error. Two autoencoder architectures are tested: one converts per-track particle-identification information into counts of candidate particles per species, and the other feeds the raw N-sigma PID values directly into the network. An anomaly threshold at mean squared error 0.01, corresponding to the top 4% of the training distribution, flags the with-exotica events. J/ψ→4π events are selected with purity above 90% for the first design and no contamination for the second, at efficiencies of about 11% and above 95%, respectively. All inj

What carries the argument

The central objects are two shallow autoencoders—an encoder and a decoder, each a single dense layer—that compress event information into a low-dimensional latent space and reconstruct it, with mean squared reconstruction error serving as the anomaly score. The load-bearing design choice is the input vector: the first autoencoder counts how many tracks in an event fall within a reference N-sigma of each particle hypothesis, while the second feeds twenty raw N-sigma values plus a track count, using large placeholders for absent tracks and avoiding discrete PID decisions. These inputs are produced by folding decay kinematics (pT ≈ 0, isotropic decays) through a time-projection-chamber energy-l

Load-bearing premise

The toy samples reproduce the real ultraperipheral-collision environment closely enough—all production coherent with pT≈0, isotropic decays, and PID captured by the TPC energy-loss functional form alone—that a model trained on them behaves the same way on real data.

What would settle it

Run the same autoencoders on a full detector simulation of the with-exotica cocktail that includes incoherent production, momentum smearing, and tracking inefficiencies: if fewer than all injected pentaquark events pass the 0.01 MSE threshold, or if any non-pentaquark event passes, the flag-all-pentaquarks claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Searches can be run on full data sets without triggering on a predefined decay topology; any event with reconstruction error above threshold becomes a candidate for further study.
  • Rare channels such as J/ψ→4π, normally reachable only through branching-ratio estimates, can be isolated as invariant-mass peaks in the flagged-event sample.
  • The absence of flagged pentaquark events can be converted into an upper limit on σ×BR for Pc(4380) production in ultraperipheral collisions, complementing photoproduction limits.
  • A UPC measurement of the fully-charmed tetraquark T_cccc→4μ would provide the first upper limit in that mass region, since no such measurement currently exists.
  • Because the model is trained only on known processes, the same pipeline transfers to other low-background environments, including future electron-ion collider data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The decisive untested step is transfer from toy to real data: the paper assumes all production is coherent with pT≈0, decays are isotropic, and PID is captured by the TPC energy-loss functional form. A full detector simulation or a run on recorded UPC data including incoherent production, tracking inefficiency, and momentum smearing would test whether the perfect pentaquark selection survives.
  • The second autoencoder's raw-N-sigma input makes no discrete particle-hypothesis decision, so it may generalize to final states beyond the five species and four tracks considered here; a natural extension is to inject channels with electrons or displaced vertices.
  • The two designs sit at opposite ends of a purity-efficiency tradeoff (11% vs above 95% efficiency for J/ψ→4π), suggesting that an ensemble or threshold scan could tune the search to a target process; the paper does not explore that axis.
  • The exclusion limits are computed assuming zero background among flagged events; with real contamination the limits would weaken, and a background-subtracted limit would be a straightforward follow-up.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes two autoencoder architectures for anomaly detection in ultraperipheral and diffractive events at the LHC, using ALICE as the test case. The models are trained on toy samples of exclusive processes that ALICE has already measured, including γγ→μμ, γγ→KK, J/ψ→μμ, J/ψ→p\bar{p}, ψ'→μμ, and ψ'→J/ψππ, with event features consisting of track momenta, energies, and TPC dE/dx PID information. A 'realistic' cocktail is constructed from published ALICE rates, and rare processes are injected: J/ψ→4π and P_c(4380)→J/ψ+p. The authors report that both autoencoders flag the injected rare events with high purity and efficiency, and they derive upper limits on pentaquark and fully-charmed tetraquark production. The central claim is that this technique can enable model-agnostic searches in current and future LHC data.

Significance. If the claims are substantiated, this is a useful proof-of-concept for an alternative to cut-based and supervised searches in low-multiplicity exclusive processes. The strength of the paper is its grounding in concrete ALICE measurements, the explicit modeling of the TPC PID response, and the attempt to estimate event rates from published luminosities and cross sections. The approach is timely and the two autoencoder designs are a sensible starting point. However, the results are currently only demonstrated on a highly idealized toy simulation, and the pentaquark normalization and upper-limit calculations contain inconsistencies. The significance of the paper as a citable search strategy depends on whether the authors can show the method transfers to realistic detector conditions or clearly limit the claims to a toy-level demonstration.

major comments (4)
  1. [Secs. 2, 2.1, 2.3] The central claim that the autoencoders 'flag all the injected pentaquark events' with 'no contamination' is made for events generated at pT=0 with isotropic decays and only a Gaussian TPC PID model. Because the input stores raw px, py, pz, E per track, the autoencoders can separate processes by exploiting the exact zero-transverse-momentum and back-to-back kinematics rather than by learning physics that survives in real data. Real ALICE UPC data have finite pT from photon virtuality, nuclear breakup, a small incoherent component, acceptance and trigger losses, tracking inefficiencies, and PID correlations that are not modeled. The manuscript does not provide any validation of transfer to more realistic conditions. Please either add a validation with pT smearing and a detector-response simulation, or restrict the claims to a toy-level demonstration and remove the statements about enablin
  2. [Sec. 2.2, Table 3] The pentaquark normalization is internally inconsistent. The text derives σ=100 nb from the GlueX upper limits BR=4.6% and BR×σ=4.6 nb, but then states that the event counts in Table 3 are 'obtained with the highest of the branching ratio scenarios' (79%). If σ=100 nb and BR=79%, one obtains BR×σ=79 nb, a factor ~17 above the GlueX upper limit. In contrast, the Run 3&4 value of 14 events with L=10 nb^{-1} and an efficiency of about 30% corresponds to BR×σ=4.6 nb, not 79 nb. This makes the expected event counts and the resulting upper-limit interpretation ambiguous. Clarify which normalization is actually used and correct the text accordingly.
  3. [Sec. 2.3, Sec. 3] The pentaquark efficiency and purity claims are statistically unsupported. Table 3 gives 0–1 injected pentaquark events for the Run 2 cocktail, which is the cocktail analyzed in Sec. 2.3. With zero or one events, 'all injected pentaquark events are flagged' is either vacuous or based on a single event, and 'no contamination' cannot be meaningfully established. The Run 3&4 case with 14 events is better but still small. In addition, the MSE threshold of 0.01 is chosen from the same distribution to which it is applied; no validation-set procedure is described. Please report results over a large number of injected events (hundreds or thousands) with statistical uncertainties, and specify how the threshold is fixed before looking at the exotica sample.
  4. [Sec. 3, Fig. 6] The upper-limit curves in Figure 6 are not reproducible as written. The text says the limits are 'simplified' and assume zero background, but it does not give the efficiency as a function of mass, the number of observed events used, or the 90% CL statistical prescription. The pentaquark limit also depends on the unresolved normalization issue in Sec. 2.2. Please provide the full calculation, including the efficiency model and the limit-setting formula, or remove the quantitative limits until the method is validated on more realistic simulations.
minor comments (6)
  1. [Eq. (MSE)] The symbol σ is used for both the cross section in Eq. (2.1) and the number of features in the MSE definition. Use a different symbol, e.g., N_features.
  2. [Sec. 2.2] The branching ratio for J/ψ→4π used in Eq. (2.4) is never quoted. Please give the PDG value and the resulting expected number of events.
  3. [Table 3] The ranges '0 to 1' and '1 to 2' for the pentaquark counts are not explained. Since the expected number is a single number, clarify the rounding and the underlying efficiency values.
  4. [Sec. 2.1] The latent-space dimensionality for each autoencoder is never stated although the text says the two models differ in it. Please provide the exact architecture dimensions.
  5. [Sec. 2.1] The claim that results are 'roughly similar' for N_sigma_ref = 0.5, 1, and 2 is not quantified. Add a small table or figure comparing the three choices.
  6. [General] Minor typographical and notation issues: 'PC(4380)' should be P_c(4380)^+; the title says 'full-data set search' although only toy samples are analyzed, which should be clarified.

Circularity Check

0 steps flagged

No material circularity: simulation results are not derived from their inputs; minor self-citations are non-load-bearing.

full rationale

The paper is an empirical simulation study, not a derivation, and its central claims are not equivalent to its inputs. The training cocktails are built from external ALICE measurements ([9], [29]) and a known TPC response model [24]; exotic samples are generated independently from PDG branching ratios and GlueX upper limits. The autoencoders are trained only on the no-exotica sample; the exotica are never used as training targets or to set the MSE threshold (0.01 is chosen from the no-exotica MSE distribution). Flagging J/psi->4pi and pentaquarks as anomalous is therefore an out-of-sample result of the trained models, not a quantity fitted into them. The authors' prior work ([18], [19]) motivates the technique and the author's thesis [31] supplies a proton efficiency, but neither is load-bearing for the anomaly-detection demonstration; the same test would stand with any reasonable efficiency. The real weakness is transferability (pT ~ 0, isotropic decays, no full detector simulation), which is a validation gap, not circularity. There is also an internal inconsistency in the pentaquark BR scenario (Sec. 2.2 derives sigma=100 nb from BR=4.6% but Table 3 uses 79%), but that is a normalization/correctness issue, not a circular reduction.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The central claim rests on several domain assumptions about the realism of the toy data (coherent-only pT = 0 production, isotropic decays, simplified PID, balanced training) and on fitted normalizations for background processes. No new particles or forces are introduced; the pentaquark and tetraquark are existing hypotheses taken from the literature.

free parameters (6)
  • NKK continuum normalization = 34 events (1.1 to 1.4 GeV/c2)
    Inferred from statistical uncertainties in ALICE K+K- cross-section data [29], used to normalize the exponential continuum distribution in the cocktail.
  • gamma gamma to mu mu continuum exponential slope = not quoted
    Obtained by fitting the invariant mass distributions in [9] to an exponential; used to determine the number of continuum mu-pair events in the cocktail.
  • MSE threshold = 0.01 (top 4%)
    Hand-set selection threshold for flagging anomalies; described as corresponding to the top 4% of the MSE distribution without stating which cocktail was used to define it.
  • N_sigma_ref for autoencoder 1 = 0.5
    Reference N-sigma value used to decide whether a track is a candidate for a given particle species; the paper notes results are similar for 0.5, 1, or 2.
  • Pc(4380) cross section = 100 nb
    Derived from GlueX upper limits BR = 4.6% and BR times sigma = 4.6 nb; used to set the pentaquark injection rate.
  • Pc(4380) detection efficiency = 30.4% (12.16% times 2.5)
    Product of proton TPC efficiency and J/psi efficiency from [30], [31], with a factor 2.5 for additional J/psi decay channels; treated as point values without uncertainties.
axioms (6)
  • domain assumption Coherent UPC production dominates and pT is set to 0 for all processes.
    Sec. 2 states 'we focus on coherent production. We therefore assume that pT ~ 0 GeV/c for each resonance and process under consideration.' This excludes incoherent contributions and pT smearing present in real data.
  • domain assumption Decays are isotropic in the parent or gamma-gamma rest frame.
    Sec. 2: 'All processes have been generated with isotropic distributions in the centre-of-mass frame of the original particle, or in the gamma-gamma frame.' This ignores spin correlations and acceptance effects.
  • domain assumption The TPC dE/dx functional form from [24] adequately models experimental PID and N-sigma values.
    Sec. 2: 'We have implemented the functional form of the ALICE TPC PID response.' No comparison to full detector simulation or real data is provided.
  • domain assumption The number of continuum events can be inferred from statistical uncertainties of published cross sections.
    Sec. 2.2, Eqs. 2.1 to 2.3 assume the statistical uncertainty comes entirely from NKK, giving NKK = 34; if the acceptance-efficiency uncertainty is not negligible, the cocktail composition changes.
  • domain assumption Reconstruction error (MSE) is a valid anomaly score with a threshold at the top 4%.
    Sec. 2.3: events with MSE above 0.01 are selected as exotica; this assumes anomalies are separable from known events by reconstruction error alone.
  • domain assumption Balanced training samples (30,000 events per process) transfer to the highly unbalanced real cocktail.
    Sec. 2.1 trains on 30,000 events per process while Table 3 has abundance ratios up to roughly 3000 to 1; no reweighting or prior correction is discussed.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of AI-powered full-data set search for new physics in ultraperipheral and diffractive events." pith.science (2026). https://pith.science/paper/K5YD2GS3

@misc{pith2026250821728,
  author       = {Pith},
  title        = {Pith review of: AI-powered full-data set search for new physics in ultraperipheral and diffractive events},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K5YD2GS3}},
  note         = {Machine review of arXiv:2508.21728}
}
Share X Bluesky LinkedIn Reddit HN
abstract

We present possible strategies for anomaly detection of rare particle decays and exotic hadrons, such as pentaquarks, in low-background environments such as those characteristic of diffractive events and ultraperipheral \pp, \pA, or \AAcoll collisions at the CERN Large Hadron Collider (LHC). Our models are trained with toy samples representing the UPC processes measured until now by the ALICE Collaboration. When samples containing rare processes such as $\jpsi\rightarrow4\pi$ and pentaquark production, where the number of injected pentaquark events is estimated based on current experimentally available upper limits, and those for $\jpsi\rightarrow4\pi$ are estimated through the branching ratio of the decay channel, are analyzed, the rare processes are flagged as anomalous by the models. This approach demonstrates the applicability of such a technique for searches for new physics in the current and future data sets at collider experiments with high purity, while also allowing for the measurement of upper limits for the production of exotica.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 12 canonical work pages · 4 internal anchors

  1. [2]

    ATLAS collaboration, Observation of the γγ →τ τProcess in Pb+Pb Collisions and Constraints on the τ -Lepton Anomalous Magnetic Moment with the ATLAS Detector , Phys. Rev. Lett. 131 (2023) 151802 [ 2204.13478]

    Pith/arXiv arXiv 2023

  2. [3]

    CMS collaboration, Probing Small Bjorken-x Nuclear Gluonic Structure via Coherent J/ ψ Photoproduction in Ultraperipheral Pb-Pb Collisions at √sNN = 5 .02 Te V, Phys. Rev. Lett. 131 (2023) 262301 [ 2303.16984]

    Pith/arXiv arXiv 2023

  3. [5]

    STAR collaboration, Exclusive J/ ψ, ψ(2s), and e+e − pair production in Au+Au ultraperipheral collisions at the BNL Relativistic Heavy Ion Collider , Phys. Rev. C 110 (2024) 014911 [ 2311.13632]

    Pith/arXiv arXiv 2024

  4. [6]

    ALICE collaboration, Coherent photoproduction of ρ0 vector mesons in ultra-peripheral Pb–Pb collisions at √sNN = 5.02 TeV , JHEP 06 (2020) 035 [ 2002.10897]

    Pith/arXiv arXiv 2020

  5. [7]

    STAR collaboration, ρ0 photoproduction in ultraperipheral relativistic heavy ion collisions at√sN N= 200 GeV , Phys. Rev. C 77 (2008) 034910 [ 0712.3320]

    Pith/arXiv arXiv 2008

  6. [8]

    ALICE collaboration, Energy dependence of coherent photonuclear production of J/ ψ mesons in ultra-peripheral Pb-Pb collisions at √sNN = 5.02 TeV , JHEP 10 (2023) 119 [2305.19060]

    Pith/arXiv arXiv 2023

  7. [9]

    ALICE collaboration, Coherent J/ψ and ψ′ photoproduction at midrapidity in ultra-peripheral Pb-Pb collisions at √sNN = 5.02 TeV , Eur. Phys. J. C 81 (2021) 712 [2101.04577]

    Pith/arXiv arXiv 2021

  8. [10]

    LHCb collaboration, Study of exclusive photoproduction of charmonium in ultra-peripheral lead-lead collisions, JHEP 06 (2023) 146 [ 2206.08221]

    Pith/arXiv arXiv 2023

  9. [11]

    ALICE collaboration, Exclusive four pion photoproduction in ultraperipheral Pb-Pb collisions at √sNN = 5.02 TeV, 2404.07542. – 14 –

    arXiv

  10. [12]

    STAR collaboration, Photoproduction at the Relativistic Heavy Ion Collider with STAR , Nucl. Phys. A 830 (2009) 507C [ 0907.2351]

    work page internal anchor Pith review Pith/arXiv arXiv 2009

  11. [13]

    Particle Data Groupcollaboration, Review of particle physics , Phys. Rev. D 110 (2024) 030001

    work page 2024

  12. [14]

    GlueX collaboration, First Measurement of Near-Threshold J/ ψ Exclusive Photoproduction off the Proton , Phys. Rev. Lett. 123 (2019) 072001 [ 1905.10811]

    Pith/arXiv arXiv 2019

  13. [15]

    Gon¸ calves and B.D

    V.P. Gon¸ calves and B.D. Moreira,Fully - heavy tetraquark production by γγ interactions in hadronic collisions at the LHC , Phys. Lett. B 816 (2021) 136249 [ 2101.03798]

    Pith/arXiv arXiv 2021

  14. [16]

    LHCb collaboration, Observation of a narrow pentaquark state, Pc(4312)+, and of two-peak structure of the Pc(4450)+, Phys. Rev. Lett. 122 (2019) 222001 [ 1904.03947]

    Pith/arXiv arXiv 2019

  15. [17]

    LHCb collaboration, Observation of charmonium pairs produced exclusively in pp collisions, J. Phys. G 41 (2014) 115002 [ 1407.5973]

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  16. [18]

    Machine learning opportunities for online and offline tagging of photo-induced and diffractive events in continuous readout experiments

    S. Ragoni, J. Seger, C. Anson and D. Tlusty, Machine learning opportunities for online and offline tagging of photo-induced and diffractive events in continuous readout experiments , 2410.06983

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [19]

    Zero-bias new particle searches using autoencoders in UPCs and diffractive events

    S. Ragoni, J. Seger and C. Anson, Zero-bias new particle searches using autoencoders in UPCs and diffractive events , 2411.00903

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [20]

    ALICE collaboration, Measurement of the Lifetime and Λ Separation Energy of 3 H Λ, Phys. Rev. Lett. 131 (2023) 102302 [ 2209.07360]

    Pith/arXiv arXiv 2023

  19. [21]

    Hinton and R.R

    G.E. Hinton and R.R. Salakhutdinov, Reducing the Dimensionality of Data with Neural Networks, Science 313 (2006) 1127647

    work page 2006

  20. [22]

    Klein, J

    S.R. Klein, J. Nystrand, J. Seger, Y. Gorbunov and J. Butterworth, STARlight: A Monte Carlo simulation program for ultra-peripheral collisions of relativistic ions , Comput. Phys. Commun. 212 (2017) 258 [ 1607.03838]

    Pith/arXiv arXiv 2017

  21. [23]

    ALICE collaboration, Production of light-flavor hadrons in pp collisions at√s = 7 and √s = 13 TeV, Eur. Phys. J. C 81 (2021) 256 [ 2005.11120]

    Pith/arXiv arXiv 2021

  22. [24]

    ALICE collaboration, Performance of the ALICE Experiment at the CERN LHC , Int. J. Mod. Phys. A 29 (2014) 1430044 [ 1402.4476]

    Pith/arXiv arXiv 2014

  23. [25]

    Farina, Y

    M. Farina, Y. Nakai and D. Shih, Searching for New Physics with Deep Autoencoders , Phys. Rev. D 101 (2020) 075021 [ 1808.08992]

    Pith/arXiv arXiv 2020

  24. [26]

    Chollet et al., Keras, GitHub, 2015, https://github.com/fchollet/keras

    F. Chollet et al., Keras, GitHub, 2015, https://github.com/fchollet/keras

    work page 2015

  25. [27]

    Dertat, Applied deep learning - part 3: Autoencoders , October, 2017, https://towardsdatascience.com/applied-deep-learning-part-3-autoencoders-1c083af4d798

    A. Dertat, Applied deep learning - part 3: Autoencoders , October, 2017, https://towardsdatascience.com/applied-deep-learning-part-3-autoencoders-1c083af4d798

    work page 2017

  26. [28]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel et al., Scikit-learn: Machine Learning in Python , Journal of Machine Learning Research 12 (2011) 2825

    work page 2011

  27. [29]

    ALICE collaboration, Photoproduction of K+K- Pairs in Ultraperipheral Collisions , Phys. Rev. Lett. 132 (2024) 222303 [ 2311.11792]

    Pith/arXiv arXiv 2024

  28. [30]

    ALICE collaboration, Inclusive J /ψ production at midrapidity in pp collisions at √s = 13 TeV, Eur. Phys. J. C 81 (2021) 1121 [ 2108.01906]. – 15 –

    Pith/arXiv arXiv 2021

  29. [31]

    S. Ragoni, Hadron spectra measurement in Xe–Xe collisions at √sNN = 5.44 TeV with the ALICE experiment at the LHC , laurea magistrale thesis, Alma Mater Studiorum – Universit` a di Bologna, Bologna, Italy, September, 2018

    work page 2018

  30. [32]

    Cowan, Statistical Data Analysis , Clarendon Press International Series on Particle Physics, Oxford University Press, Oxford, UK (1998)

    G. Cowan, Statistical Data Analysis , Clarendon Press International Series on Particle Physics, Oxford University Press, Oxford, UK (1998)

    work page 1998

  31. [33]

    ONNX, ONNX, ONNX Community, 2021, https://onnx.ai. – 16 –

    work page 2021

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.