{"id":"eacfe108-25ab-48b4-a6b4-2568a2fb3437","arxiv_id":"2507.10738","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An information-field-theory based model reconstructs inclined air shower radio events, recovering energy fluence well while electromagnetic energy and Xmax remain biased.","lead":"Cosmic ray radio detectors see noisy signals, and this paper builds a Bayesian reconstruction that infers the electric field, energy, and shower maximum all at once using information field theory. It is a generalist-relevant step because current radio reconstruction is noise-limited and struggles to attach reliable uncertainties to its estimates.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The likelihood noise width is set to the RMS of the measured data, which includes the signal itself; this misspecification can explain the miscalibrated pulls and biases, so the central inference claim is not yet supported.","rationale":"The reader's weakest_assumption identifies the Gaussian likelihood as the key concern. My stress-test sharpens this to a specific and testable modeling error: using the RMS of the signal-containing data as the noise standard deviation. This is more than a generic 'likelihood misspecification' because it is explicitly stated in Section 2.4 and because it has a direct, mechanistic link to the observed pull miscalibration: high-SNR traces have large RMS, inflating the error bars and producing overconfident or underconfident posteriors depending on how the inflated variance interacts with the forward model's correlations. The paper's own results show that the uncertainty estimates are not reliable for E_EM and dmax; the noise model is a concrete and plausible cause, and it can be tested without new data or a new inference framework. I therefore do not reject the paper's conditional acceptance, but I agree with the reader that the verdict must remain CONDITIONAL pending a corrected likelihood and re-validation. The authors' own discussion of the energy bias and the indirect constraining of dmax is honest, but it does not address the likelihood variance choice, which is a more fundamental issue.","tokens_in":5874,"tokens_out":3646,"duration_ms":46215,"concrete_test":"Rerun the reconstruction on a subset of the 2859 simulations, replacing the likelihood variance with a noise-only RMS estimate (e.g., from a pre-trigger or late-time window of each trace, or from the known injected noise level). Compare the fluence pull width, E_EM and dmax biases and pull distributions to Figures 1-3. If the fluence pull moves toward 1 and the E_EM/dmax biases or spreads change materially, the data-RMS likelihood is the culprit; if not, the miscalibration originates elsewhere, such as the prior or the MGVI posterior approximation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.4 states: 'The likelihood is assumed to be a normal distribution with a width equal to the RMS of the data.' This is the load-bearing assumption: the noise variance is taken from the full measured voltage trace, which contains the air-shower signal, rather than from an independent noise estimate. For the accepted stations with SNR>10, the signal dominates the trace, so the nominal 'noise' width scales with signal amplitude. Three consequences follow. First, the likelihood is not the true generative distribution of the data, so the posterior mode and covariance are not the correct Bayesian update. Second, the inflated variance for high-SNR stations naturally produces overestimated uncertainties, matching the fluence pull sigma of 0.73 reported in Section 3. Third, the same misspecified variance is used to compute the reduced chi2<1.05 cut, so event selection is also driven by the flawed noise model. The discrete ADC-count issue mentioned in Section 3 is a separate symptom of the same underlying misspecification. Because the paper's headline claim includes 'naturally provide uncertainties,' and because the reported pulls for E_EM and dmax are already miscalibrated, this noise model must be corrected before the central claim can be considered supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a forward model for Bayesian reconstruction of inclined cosmic-ray air showers from radio detector data, built on Information Field Theory. The model combines parametric descriptions of the lateral signal distribution, charge-excess emission, and spectral shape with Gaussian-process models for shower-to-shower fluctuations and narrowband RFI, followed by a detailed detector response and a Gaussian likelihood. The model is applied to 2859 CoREAS simulations folded through a Pierre Auger radio-detector simulation with measured ADC noise. The reconstructed energy fluence agrees well with simulations (pull sigma 0.73), while the reconstructed electromagnetic energy and distance to shower maximum show clear biases and miscalibrated uncertainties. The authors present the work as a first attempt and label the results preliminary.","tokens_in":6154,"tokens_out":2762,"duration_ms":36156,"significance":"If the central claim were fully supported, this would be a valuable step toward holistic, uncertainty-aware reconstruction of air-shower radio signals. The fluence normalisation in Eqs. (5)-(6) is a clean, parameter-free derivation from the Poynting theorem, and using IFT/MGVI to handle continuous fields is well motivated. The evaluation on a large set of CoREAS simulations with realistic noise injection is a strength, and reporting pull distributions is the right way to test uncertainty calibration. However, the paper's own results show that the model currently delivers calibrated, unbiased reconstruction only for the electric-field-related quantity (fluence), not for the two headline shower observables E_EM and dmax. The likelihood misspecification described below is load-bearing for the claimed ability to 'naturally provide uncertainties', so the central claim is not yet fully supported.","major_comments":[{"comment":"The likelihood is defined as a normal distribution whose width equals the RMS of the measured voltage trace, but that RMS includes the air-shower signal itself. For the stations retained by the SNR>10 cut, the signal dominates the trace, so the nominal noise variance scales with the signal amplitude rather than representing an independent noise estimate. This makes the likelihood an incorrect generative model for the data, so the posterior mode and covariance are not the correct Bayesian update. It also directly affects the reported uncertainty calibration: the inflated variance naturally produces overestimated uncertainties, consistent with the fluence pull sigma of 0.73 reported in Section 3. The same flawed noise model is used to compute the reduced chi2 that enters the event-selection cut, so the selection itself is contaminated by the misspecification. The paper should either use an independent noise estimate for the likelihood width or explicitly model the noise level as a free parameter; without this, the claim that the method 'naturally provides uncertainties' is not supported.","section":"Section 2.4"},{"comment":"The results for the two headline shower observables contradict the abstract's unconditional claim that the model 'can infer ... the electromagnetic energy and position of shower maximum'. For all events, E_EM is biased toward overestimation at low energies with a wide spread, dmax is always overestimated and 'hardly constrained for deep showers', and the pull distributions show the E_EM uncertainty is severely overestimated while the dmax uncertainty is underestimated. Even after restricting to events with at least five stations, the biases remain. The abstract and introduction should be revised to state that the current model reconstructs the electric field/fluence reliably while E_EM and dmax are only preliminary and biased, or the model must be improved so that the headline claim matches the demonstrated performance.","section":"Section 3, Figures 2 and 3"},{"comment":"The selection criteria SNR>10 and reduced chi2<1.05 are both applied before evaluating reconstruction quality, but the reduced chi2 is computed with the same Gaussian likelihood whose width is the data RMS. This creates a circularity: events are accepted partly based on how well they fit the misspecified likelihood, and the reported performance is then conditioned on that acceptance. The number of events passing each cut should be reported, and the main results should be shown with and without the reduced-chi2 cut to demonstrate that the reported biases and pulls are not an artifact of the selection. If the cut is necessary to remove ADC-discretisation failures, a likelihood that accounts for the integer-count noise should be used instead of excluding a potentially large fraction of events.","section":"Section 3, selection criteria"}],"minor_comments":[{"comment":"There are several typos and spacing errors, including 'noticably', 'claculated', 'colou', 'mathematial', and 'one-dimensionallateraldistributionfunction'; these should be corrected.","section":"Throughout"},{"comment":"The assumption that the phase spectrum is linear with a slope set by the trace peak time is introduced without justification or validation; given that timing information is later identified as important for constraining dmax, this assumption deserves an explicit statement of its role and limitations.","section":"Section 2.4"},{"comment":"The Gaussian-process priors for shower-to-shower fluctuations and RFI are described only qualitatively; the kernel choice and hyperparameter values should be listed to make the inference reproducible.","section":"Section 2.2 / 2.4"},{"comment":"The statement that dmax is 'hardly constrained for deep showers' is not quantified; adding a typical uncertainty or credible-interval width for dmax as a function of depth would make the claim precise.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"This is an ICRC proceedings paper, so some brevity is expected, but the abstract makes a stronger claim than the results support. The likelihood misspecification in Section 2.4 is a real obstacle and should be addressed before the uncertainty-related claims are presented as demonstrated. The self-citation pattern is not unusual for a developing line of work, but the reader should be told more explicitly which parts are carried over from [8] and which are new here."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Simon,\n\nYou should know this paper is a conference-proceedings methods paper, not a finished analysis. The genuinely new piece is extending the group's IFT electric-field reconstruction to inclined showers and adding GS-LDF and charge-excess parameterizations so that the forward model can infer E_EM and dmax simultaneously. The fluence reconstruction on 2859 CoREAS simulations is solid—residuals narrow, pulls sigma 0.73. The electric-field normalization in Eqs. 5–6 is correct and clearly derived. The authors also deserve credit for reporting pull distributions and for stating their own limitations plainly in Section 4.\n\nThe soft spots are real. The load-bearing assumption is the Gaussian likelihood with width set to the RMS of the measured voltage trace. Since accepted stations have SNR>10, that RMS is dominated by signal, so the 'noise' variance scales with the signal amplitude. That is not the generative distribution for the data, and it directly degrades the posterior covariance and the reduced chi2<1.05 cut. The miscalibrated pulls for E_EM (overestimated uncertainty) and dmax (underestimated) are consistent with this misspecification. The ADC discretisation issue the authors mention is a separate symptom of the same problem. So the claim of 'naturally provide uncertainties' is not yet supported for the shower observables.\n\nThe biases in E_EM and dmax are also noted by the authors; their explanation about the different fluence definition is plausible. Minor quibbles: no code or data released, and key implementation details live in [8] and [5], so independent reproduction is hard. But for a proceedings paper that's not unusual.\n\nWho should read this: anyone building radio reconstruction pipelines or applying IFT in astroparticle contexts. It is a proof of principle, not a production algorithm. I would send it to a serious referee; it's a coherent, honest methods contribution with a clear path forward. The referee should push on the noise model and ask for a decomposition of the bias sources.\n\nBest, [Your name]","headline":"Honest early-stage methods paper that reconstructs fluence well but the headline observables are biased and the uncertainty estimates are undercut by a signal-inclusive noise model.","tokens_in":6686,"tokens_out":2477,"would_cite":true,"duration_ms":28622,"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":"One Bayesian forward model can reconstruct a cosmic-ray air shower's electric field, geometry, electromagnetic energy, and depth of maximum from radio-detector traces in a single inference pass, with the electric field recovered well and…","keywords":["Information Field Theory","cosmic ray radio detection","air shower reconstruction","Bayesian inference","Gaussian process","metric Gaussian variational inference","energy fluence","charge excess emission"],"falsifier":"Re-run the same 2859 simulated events without the signal-to-noise and reduced-$\\chi^2$ cuts, replacing the Gaussian likelihood with one that models integer ADC counts; if the pull distributions for $E_{\\mathrm{em}}$ and $d_{\\mathrm{max}}$ remain offset or non-unit, the Gaussian-likelihood approximation is not the dominant source of the miscalibration, and the bias must instead come from the physical parametrisations.","tokens_in":5668,"feed_emoji":"📡","tokens_out":7259,"duration_ms":78782,"temperature":0.7,"pith_summary":"This paper argues that the full radio footprint of an inclined cosmic-ray air shower can be reconstructed in one Bayesian pass, treating the electric field, shower geometry, electromagnetic energy, and distance to shower maximum as unknowns of a single forward model. The model is semiparametric: hard-wired parametrisations of the lateral signal distribution, charge-excess contribution, and spectral shape carry the physics, while Gaussian processes absorb shower-to-shower fluctuations and narrowband radio-frequency interference. On 2859 simulated inclined showers folded through the Auger radio-detector response, the recovered energy fluence agrees well with the simulated value, with a small negative bias and a pull width of $\\sigma=0.73$. The electromagnetic energy and distance to shower maximum come out biased or poorly constrained, and their reported uncertainties are miscalibrated; the authors treat the results as preliminary and identify timing information as the likely missing ingredient. If the approach matures, it would replace the current practice of reconstructing one observable at a time with a single inference that naturally reports correlated uncertainties.","feed_headline":"Radio fit yields field, energy, and shower depth in one pass","feed_subtitle":"A single information-field model replaces separate reconstructions and returns uncertainties for each observable.","key_machinery":"The engine is a semiparametric forward model, i.e. a function that turns a finite set of shower parameters plus Gaussian-process fields into the expected voltage traces of every antenna. Its parametric core is the GS lateral distribution function (a Gaussian plus a sigmoid) for geomagnetic emission, the charge-excess fraction parametrised through $d_{\\mathrm{max}}$ and the density at shower maximum, and an analytic absolute spectrum normalised to the fluence via the Plancherel theorem. Deviations from these parametrisations, shower-to-shower fluctuations, and narrowband RFI are modelled as Gaussian processes added at the electric-field level. The instrument response, including the complex antenna gain pattern, is applied to the field, and the comparison to data uses a Gaussian likelihood whose width is the RMS of the measured trace. Inference is performed with metric Gaussian variational inference, which makes the high-dimensional continuum-limit problem tractable.","core_discovery":"On its own terms, the paper establishes that a semiparametric information-field forward model can map shower parameters and fluctuating field components all the way to measured ADC-count voltage traces, and that inverting that map with variational inference yields a posterior over the electric field together with the arrival direction, $f_0$ normalisation, electromagnetic energy $E_{\\mathrm{em}}$, and geometric distance to shower maximum $d_{\\mathrm{max}}$. The electric-field part of the posterior is validated: fluence residuals are narrow and the pull distribution has $\\sigma=0.73$, indicating overestimated uncertainties. The same plot shows a small systematic underestimation of fluence. For the two shower observables the reconstruction is acknowledged to be weaker: $E_{\\mathrm{em}}$ is biased high at low energies, $d_{\\mathrm{max}}$ is overestimated and hardly constrained for deep showers, and the pull distributions show overestimated energy uncertainties and underestimated $d_{\\mathrm{max}}$ uncertainties. The paper's claim, stated fairly, is that this holistic reconstruction is possible in principle, with these biases and miscalibrations as the current, understood limitations.","pith_inferences":["If the fluence-definition mismatch is corrected by subtracting a noise-window integral before comparing to the LDF, the electromagnetic-energy bias should largely disappear; this is directly testable by re-running the same forward model with the alternative fluence definition.","The strong dependence of $d_{\\mathrm{max}}$ constraining power on shower depth suggests that replacing the indirect spectral-slope constraint with explicit per-station arrival-time information would sharpen the posterior more than any other single model change.","The SNR>10 and reduced-$\\chi^2<1.05$ cuts are a direct consequence of the Gaussian likelihood's inability to describe integer ADC counts; a discrete-count likelihood would recover the excluded low-signal events and reveal whether the reported pulls are artefacts of the cut rather than of the physics.","A natural stress test is to apply the same forward model, unchanged, to vertical or near-vertical showers, where the rotationally-symmetric LDF assumption is better justified; agreement there would separate geometry effects from model-form errors."],"forward_implications":["One inference run yields a joint posterior over the electric field, shower geometry, electromagnetic energy, and distance to shower maximum, so the correlations between these quantities become available in radio reconstruction.","Every reconstructed quantity carries a natural uncertainty estimate, which is exactly the piece traditional single-observable methods struggle to provide.","Narrowband RFI and shower-to-shower fluctuations are handled inside the forward model by Gaussian processes, so no separate RFI subtraction or fluctuation-averaging step is needed.","Because the model is modular, timing distributions or particle-detector data can be added as extra data channels in a later version, which the authors say is needed to constrain $d_{\\mathrm{max}}$.","The energy bias is attributed to a definitional mismatch: the LDF parametrisation was fit to noise-subtraction fluences, while this work integrates over all time, implying the method most likely underestimates the highest energies rather than systematically overestimating all energies."],"supporting_citations":[{"why":"Establishes the information-field-theory continuum-limit result that lets the continuous electric field be inferred from discretised measurements.","marker":"[3]"},{"why":"Supplies the GS lateral distribution function, the charge-excess parametrisation, and the early-late correction with their shower-to-shower scatter priors.","marker":"[5]"},{"why":"Provides the metric Gaussian variational inference algorithm used to approximate the posterior in the high-dimensional latent space.","marker":"[6]"},{"why":"Is the earlier electric-field model that this work extends by adding the lateral-distribution and charge-excess parametrisations.","marker":"[8]"},{"why":"Gives the parametrisation of the absolute electric-field spectrum whose amplitude the fluence normalises.","marker":"[9]"},{"why":"Describes the Auger radio detector whose response, including gains, is used to forward-fold the simulated showers to ADC counts.","marker":"[10]"}],"fun_headline_variants":["One Bayesian fit infers radio shower field and geometry","Information field theory unifies cosmic ray radio reconstruction","Single model maps radio data to shower energy and depth","Bayesian radio reconstruction yields full event with uncertainties","Holistic fit extracts field, energy, depth from radio traces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The forward model assumes the measured voltage traces are Gaussian-distributed with a width equal to their own RMS; the paper admits this is wrong for low-signal stations because the digitised ADC counts are integers, and it removes those stations by cuts instead of modelling the discreteness.","fun_headline_variants_meta":{"raw":{"variants":["One Bayesian fit infers radio shower field and geometry","Information field theory unifies cosmic ray radio reconstruction","Single model maps radio data to shower energy and depth","Bayesian radio reconstruction yields full event with uncertainties","Holistic fit extracts field, energy, depth from radio traces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000208,"raw_usage":{"total_tokens":1461,"prompt_tokens":1057,"completion_tokens":404,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":673,"completion_tokens_details":{"reasoning_tokens":328}},"tokens_in":673,"tokens_out":404,"duration_ms":4849,"temperature":1.0,"reasoning_tokens":328,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:26:58.623611+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same 2859 simulated events without the signal-to-noise and reduced-$\\chi^2$ cuts, replacing the Gaussian likelihood with one that models integer ADC counts; if the pull distributions for $E_{\\mathrm{em}}$ and $d_{\\mathrm{max}}$ remain offset or non-unit, the Gaussian-likelihood approximation is not the dominant source of the miscalibration, and the bias must instead come from the physical parametrisations.","supporting_citations":[{"cited_title":"Strähnz for the Pierre Auger collaboration PoS ICRC2025 401","cited_arxiv_id":null,"evidence_quote":"Establishes the information-field-theory continuum-limit result that lets the continuous electric field be inferred from discretised measurements."},{"cited_title":"Terveer, S","cited_arxiv_id":null,"evidence_quote":"Supplies the GS lateral distribution function, the charge-excess parametrisation, and the early-late correction with their shower-to-shower scatter priors."},{"cited_title":"Martinelli, F","cited_arxiv_id":null,"evidence_quote":"Describes the Auger radio detector whose response, including gains, is used to forward-fold the simulated showers to ADC counts."}],"review_version":1}