{"id":"8ec81b6a-0a26-424c-a02b-f61a6c419cfe","arxiv_id":"2607.21689","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A neural-flow emulator of the CECA femtoscopic source, fit to 49 ALICE pp correlation functions, reduces claimed uncertainties on antinuclei coalescence parameters B2 and B3 to a few percent and ~10% respectively.","lead":"This paper trains a neural network to imitate a model of particle emission and fits it to 49 ALICE proton-proton correlation measurements, then uses the result to predict how often cosmic-ray collisions would create antideuterons and antihelium-3. It matters because dark-matter searches using cosmic-ray antinuclei need a precise Standard-Model background prediction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Extrapolation from LHC to cosmic-ray energies rests on untested energy-independence of the source; existing 7 TeV data can test it.","rationale":"The reader's weakest assumption identified the extrapolation from TeV to GeV as the key risk, and I concur. The concrete test using 7 TeV ALICE data is feasible with existing data and directly probes whether the multiplicity scaling is energy-independent. If the scaling fails, the entire cosmic-ray application is undermined; if it holds, the central claim is substantially supported. The concerns about the abstract's overstatement of uncertainty (few percent vs 25% at low pT) and the missing p-p→p-n systematic are secondary: they affect the interpretation of the uncertainty bands but do not invalidate the pipeline. The reader's CONDITIONAL verdict is appropriate, with the condition that the extrapolation be validated or its uncertainty inflated. This stress-test does not change that verdict, hence UNCHANGED. I agree with the reader's weakest_assumption, so agreement_with_reader is 'agree'.","tokens_in":13618,"tokens_out":7473,"duration_ms":74795,"concrete_test":"Test the energy-independence assumption using existing ALICE p–p femtoscopy data at √s=7 TeV (Ref. [38]). Apply the same CECA+NF surrogate fitting pipeline to these data, extracting (rd, h, τ) for the same multiplicity classes (e.g., ⟨dNch/dη⟩≈2.55, 4–5, and 6–8) used in the 13 TeV analysis. If the resulting eigenmode parameters (η, ξ) as a function of ⟨dNch/dη⟩^{1/3} do not fall on the 13 TeV linear scaling within uncertainties, the source is energy-dependent at fixed multiplicity, falsifying the extrapolation. If they do overlap, the extrapolation to √s≈24 GeV gains credibility, though a residual extrapolation above 13 TeV would remain untested.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that B2/B3 uncertainties shrink to few-percent/ten-percent levels depends on the assertion that the emission source scales with multiplicity but not collision energy, enabling extrapolation from √s=13 TeV ALICE data to the √s≈24 GeV cosmic-ray regime. This assertion is anchored by a linear fit in the (η, ξ) eigenmode space using only the three MB multiplicity classes (Supplemental Material A), with the HM class excluded as 'distinct phase-space'. The extrapolation target ⟨dNch/dη⟩≈2.2 lies below the lowest anchor (2.55), so it is an extrapolation of a three-point linear model, with no direct data in that regime. The only validation is the NA61/SHINE B2 band, which is explicitly preliminary and constructed by fitting Lévy-Tsallis functions to proton and deuteron spectra, then propagating via bootstrap. If the source actually depends on collision energy at fixed multiplicity (e.g., due to changes in resonance composition, radial flow, or hadronization dynamics), the extrapolated source bias would not be captured by the quoted uncertainties, and the claimed order-of-magnitude uncertainty reduction would be unsupported for the cosmic-ray regime. The paper itself acknowledges 'no direct data' in this regime, and the mT-extrapolation systematic—truncating the source to the lowest measured mT bin—does not address energy dependence. This is the most load-bearing concern because a systematic bias in the source would directly shift B2/B3, not merely broaden their uncertainty bands.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This Letter trains a conditional normalizing-flow surrogate for the CECA femtoscopic source model, fits it to 49 ALICE p–p correlation functions at √s = 13 and 13.6 TeV, and then extrapolates the source in multiplicity and transverse mass to the low-multiplicity regime relevant for cosmic-ray antinuclei production. Using the ToMCCA coalescence afterburner, the authors compute B2 and B3 and claim that uncertainties shrink from factors of 10–100 and ~1000 to the few-percent and ten-percent level, respectively, in the measured kinematic range. The paper includes closure tests on withheld CECA configurations, a reproduction of the published Run-2 CECA fit, and a comparison with a preliminary NA61/SHINE B2 band.","tokens_in":13907,"tokens_out":4778,"duration_ms":47612,"significance":"The methodological core is genuinely useful: replacing expensive CECA simulations with a differentiable surrogate, validating it on held-out configurations, and using femtoscopic data to constrain the coalescence source is a step forward for antinuclei predictions. If the extrapolation assumptions are correct, the impact on cosmic-ray antinuclei searches and dark-matter interpretations is substantial. However, the headline uncertainty reduction depends on an energy-scaling hypothesis that is not yet empirically established, and the current external benchmark is preliminary. The paper is therefore significant but not yet fully convincing as a basis for quantitative cosmic-ray flux predictions.","major_comments":[{"comment":"The central extrapolation from LHC pp collisions (√s=13/13.6 TeV, ⟨dNch/dη⟩≈2.55 and higher) to the cosmic-ray regime (√s≈24 GeV, ⟨dNch/dη⟩≲5) rests on the assertion that the source scales with multiplicity but not collision energy. This is supported only by a three-point linear fit in (η,ξ) eigenmode space using the three MB classes, with the extrapolation target 2.2 below the lowest anchor (2.55). The NA61/SHINE comparison in Fig. 2 is a useful but weak check: the B2 band is preliminary, reconstructed by Lévy-Tsallis fits to spectra with bootstrap propagation. I ask for a quantitative test of energy independence using existing ALICE pp femtoscopic data at √s=7 TeV [38] at matched multiplicity, or an explicit uncertainty penalty for unvalidated energy extrapolation. Without this, the few-percent uncertainty claim is not supported in the cosmic-ray regime.","section":"Main text, 'Crucially...' and Appendix A"},{"comment":"The statement that the NF surrogate simultaneously describes all 49 correlation functions with χ²/NDF = 1.39 is not interpretable as written because NDF is not reported. Since the fit quality is the main evidence that CECA captures the source geometry, please report NDF, per-class χ², and the treatment of bin-to-bin correlations. If NDF is large the quoted value is uninformative; if small it changes the significance. This is a reporting gap in a load-bearing result.","section":"Full-dataset fit, χ²/NDF = 1.39"},{"comment":"The systematic for pT/A < 0.4 GeV/c is estimated by 'truncating the source to the lowest measured mT range.' This is an arbitrary proxy for extrapolation and is not a validated model of the mT dependence below the measured range. It does not cover the possibility that the source shape changes at low mT (e.g. due to resonance composition or radial flow), nor does it address energy dependence. Since the cosmic-ray flux is dominated by low kinetic energies, this truncation procedure is load-bearing for the claimed uncertainty reduction below 0.4 GeV/c. Please justify the procedure against a physics-motivated alternative or use a more conservative systematic.","section":"mT-extrapolation systematic; Fig. 2"}],"minor_comments":[{"comment":"The Fig. 2 legend shows 'ToMCCA-B3 ×360' while the caption says the B3 prediction is scaled by a factor of 30 for visual clarity. Please harmonize the stated scaling factor.","section":"Fig. 2"},{"comment":"There is an unbalanced parenthesis in Eq. (A1): lnL_M^i,s((r_d,h,τ|m_T^(i)) has an extra opening parenthesis. Please correct.","section":"Appendix A, Eq. (A1)"},{"comment":"'SciPi.optimize.curve fit' should be 'scipy.optimize.curve_fit'.","section":"Appendix D"},{"comment":"The abstract's 'few percent and ten-percent level' should be qualified by pT range: Fig. 2 shows the source uncertainty reaching up to 23% below 0.4 GeV/c. The abstract currently overstates the uniformity of the uncertainty reduction.","section":"Abstract and Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically impressive, but the energy extrapolation is the key risk. I would like to see a 7-TeV femtoscopic test or an equivalent quantitative check before accepting the cosmic-ray claim. The missing NDF and the arbitrary mT-truncation systematic are also addressable with additional analysis. This is not a rejection: the framework and closures are well executed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper does something genuinely useful: it replaces expensive CECA simulations with a normalizing-flow surrogate, fits 49 ALICE p–p correlation functions at once, and uses an eigenmode decomposition to push source parameters toward low multiplicity. That pipeline is a real step forward for femtoscopy-to-coalescence calculations. The surrogate passes held-out validation, the Run-2 closure reproduces published source parameters, and the ALICE B2 point is consistent within 0.84σ. The work is careful and honestly flags many limitations.\n\nThe soft spots are exactly where the headline claim lives. The abstract promises B2 uncertainty at the few-percent level and B3 at ten percent, but Fig. 2 shows total B2 uncertainty reaching ~25% at low pT, plus a ~15% wavefunction systematic. That mismatch needs to be resolved, either in the numbers or in the claim.\n\nThe larger issue is the extrapolation to cosmic-ray energies. The paper assumes the source scales with multiplicity, not collision energy. That is plausible but tested only by a linear fit to three multiplicity classes, with the target ⟨dNch/dη⟩≈2.2 actually below the lowest anchor. The NA61/SHINE band used for validation is preliminary and constructed indirectly via Lévy-Tsallis fits and bootstrapping. If the source has any energy dependence at fixed multiplicity—resonance composition, radial flow, hadronization dynamics—the claimed uncertainty reduction in the cosmic-ray regime is unsupported. The paper itself admits there are no direct data there, and the mT-extrapolation systematic (truncating to the lowest mT bin) doesn't address energy dependence. A systematic for the p–p→p–n source substitution is missing, and the χ²/NDF=1.39 has no NDF given.\n\nNone of this is fatal. The framework is modular, the surrogate is fast, and the pipeline can be updated when new data arrive. But the paper overstates what is currently demonstrated. I'd send it to a serious referee, not desk-reject. The referee should ask for: a reconciliation of abstract and figure numbers, publication of the linear fit parameters and an explicit validation plan for energy independence, a systematic on the p–n substitution, and code/data for the surrogate. With those, it becomes a valuable reference for anyone estimating antinuclei backgrounds for AMS-02 and GAPS. As is, I would not cite the central numbers in my own work.\n\nMy advice: engage with it, but treat the cosmic-ray predictions as conditional on an assumption that still needs testing.","headline":"Ambitious and well-engineered pipeline, but the headline uncertainty reduction is only partly earned; the cosmic-ray extrapolation sits on an untested energy-independence assumption.","tokens_in":14504,"tokens_out":5146,"would_cite":false,"duration_ms":44169,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that a normalizing-flow surrogate of the proton–proton emission source, fit to 49 LHC correlation functions, reduces antinuclei coalescence uncertainties from factors of 10–1000 to a few-to-ten percent.","keywords":["cosmic-ray antinuclei","coalescence","femtoscopy","source function","normalizing flows","CECA source model","B2/B3 coalescence parameters","dark matter indirect detection"],"falsifier":"A direct femtoscopic measurement of p–p correlations at sqrt(s) ~ 24 GeV with multiplicity <dNch/deta> ~ 2–5, compared with the multiplicity-extrapolated LHC source at matched mT; if the inferred source radius and shape disagree by more than the quoted uncertainty, the scaling premise is falsified. A published deuteron spectrum from the ongoing SPS fixed-target program whose B2(pT) deviates from the surrogate-based band by more than the combined uncertainty would also falsify the coalescence prediction.","tokens_in":13411,"feed_emoji":"🌌","tokens_out":9967,"duration_ms":84828,"temperature":0.7,"pith_summary":"This paper argues that the dominant uncertainty in cosmic-ray antinuclei production—the shape of the proton–neutron emission source—can be removed by combining femtoscopy with machine learning. The authors train a normalizing-flow surrogate of the CECA source model on 49 proton–proton correlation functions from LHC collisions, spanning several multiplicity classes and transverse-mass bins, and fit all 49 simultaneously. Using the established scaling of the source with event multiplicity and pair transverse mass, they extrapolate from TeV-scale collisions to the low-multiplicity GeV regime where cosmic rays produce antinuclei. Propagated through a Wigner-function coalescence model, this data-constrained source reduces the uncertainty on the deuteron coalescence parameter B2 from a factor of 10–100 to a few percent (plus a ~15% wavefunction systematic) and on the three-nucleon parameter B3 from about a factor of 1000 to ~10% in the measured kinematic range. If correct, this makes predictions of the secondary antinuclei background for dark-matter searches quantitatively controlled.","feed_headline":"Machine-learned source model cuts antinuclei yield uncertainty to ~10%","feed_subtitle":"A data-driven surrogate of particle emission turns a 10-to-1000x dark-matter background uncertainty into ~10%.","key_machinery":"The load-bearing object is the conditional normalizing-flow surrogate of CECA (Common Emission in CATS), an effective femtoscopic source model parametrized by three parameters: core radius rd, shape h, and lifetime tau. The flow maps a Gaussian latent distribution through invertible transformations to the resonance-deformed, non-Gaussian source distribution S(r* | rd,h,tau; k*, mT), and is trained on 1.84 x 10^8 samples from 2300 CECA configurations. It reduces a single source evaluation by over 100x, making a simultaneous fit of all 49 correlation functions possible. The extrapolation is carried by the empirical mT–multiplicity scaling of the source—the claim that in small systems the emiss","core_discovery":"Working in the coalescence picture of light-nucleus formation, the paper shows that the femtoscopic source function S(r*) for proton–proton pairs—extracted from 49 correlation functions via the Koonin–Pratt equation with a high-precision strong interaction—can be represented by a fast, differentiable normalizing-flow surrogate of the CECA source model. Because the source scales with multiplicity and transverse mass rather than with collision energy, the authors fit LHC data at sqrt(s) = 13 and 13.6 TeV across four multiplicity classes and seven mT bins, then extrapolate along a linear eigenmode of the fit to low multiplicity (<dNch/deta> <~ 5), the regime relevant for cosmic-ray interactions","pith_inferences":["The paper explicitly sets aside proton–Lambda correlations to avoid interaction-model systematics; with improved hyperon potentials, feeding those channels through the same surrogate would constrain the source with a second isospin sector and test the universality of the extracted geometry.","Because the surrogate is differentiable and the full chain is modular, one could invert the direction of inference: instead of predicting B2 from femtoscopic data, fit cosmic-ray antinuclei measurements (once they exist) to the source parameters, making the coalescence calculation part of a Bayesian analysis of future observations.","The ~15% wavefunction systematic for B2 is at present irreducible in the paper's approach, but a matched pair of chiral potentials for the two- and three-nucleon sectors would bring the B2 and B3 wavefunction systematics under a single consistent framework."],"forward_implications":["Secondary antideuteron and antihelium flux predictions for dark-matter indirect searches become reproducible at the few-to-ten percent level rather than spanning factors of 10–1000.","The source functions replace pion-pair proxies in coalescence calculations, removing a known bias from long-lived resonance contributions that overestimates radii by up to a factor of three.","The framework accepts new femtoscopic measurements from the energy range of interest (sqrt(s) ~ 20–30 GeV) as direct constraints, without retraining the surrogate.","The A=3 predictions place antihelium-3/antitriton coalescence on a controlled footing for the first time, and the same pipeline can be extended to heavier antinuclei.","The differentiable surrogate sets up an end-to-end inference route from correlation functions through coalescence to galactic propagation."],"fun_headline_variants":["Neural flow shrinks antinuclei yield odds from 1000x to ~10%","AI surrogate pins down antinuclei rates for dark matter searches","Femtoscopic ML model cuts cosmic-ray antinuclei uncertainty","Dark matter signal sharpened: antinuclei production uncertainty down to <10%","From factor-1000 to 10%: neural surrogate fixes antinuclei yields"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole extrapolation stands or falls on the scaling premise that the proton–proton emission source is fixed by event multiplicity and pair transverse mass rather than by collision energy—a regime with no direct data, where the mT extrapolation systematic is estimated by truncating to the lowest measured bin rather than by a validated model.","fun_headline_variants_meta":{"raw":{"variants":["Neural flow shrinks antinuclei yield odds from 1000x to ~10%","AI surrogate pins down antinuclei rates for dark matter searches","Femtoscopic ML model cuts cosmic-ray antinuclei uncertainty","Dark matter signal sharpened: antinuclei production uncertainty down to <10%","From factor-1000 to 10%: neural surrogate fixes antinuclei yields"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000777,"raw_usage":{"total_tokens":3259,"prompt_tokens":714,"completion_tokens":2545,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":2438}},"tokens_in":458,"tokens_out":2545,"duration_ms":16506,"temperature":1.0,"reasoning_tokens":2438,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T07:21:37.119390+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct femtoscopic measurement of p–p correlations at sqrt(s) ~ 24 GeV with multiplicity <dNch/deta> ~ 2–5, compared with the multiplicity-extrapolated LHC source at matched mT; if the inferred source radius and shape disagree by more than the quoted uncertainty, the scaling premise is falsified. A published deuteron spectrum from the ongoing SPS fixed-target program whose B2(pT) deviates from the surrogate-based band by more than the combined uncertainty would also falsify the coalescence prediction.","supporting_citations":[],"review_version":1}