{"id":"14e4376f-f5ac-4ab7-9645-afb9bf5503e6","arxiv_id":"2505.01634","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"ATLAS finds no sign of semi-visible jets from Z' decays and excludes Z' masses from 2000 to 3200 GeV for invisible fractions between 0.2 and 0.37.","lead":"ATLAS searched 140 inverse femtobarns of 13 TeV proton collisions for semi-visible jets from a new Z' particle decaying into a hidden dark sector, using machine learning to identify jets that partially disappear inside the detector. No excess over Standard Model background was found, so the analysis set new limits on the Z' mass and on the invisible fraction of each jet.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quoted exclusion contour rests on the five-parameter mT background function holding in the Wj2>0.05 signal region; the CR/VR validations use Wj2<0.05 and never exercise the SR jet-width selection.","rationale":"The reader's verdict is ACCEPT with high confidence, and I concur that the analysis is well executed: the background function is validated in CR/VR with toy studies, 1-5 sigma signal injections, and spurious-signal checks below 0.5 sigma. The remaining soft spot is precisely the transfer of the empirical mT function to the SR, because the CR and VR both use Wj2<0.05 while the SR selects Wj2>0.05, and the stated check that Wj2 and ML scores are uncorrelated with mT is performed at preselection, before the SR-defining cuts. This could in principle allow a selection-induced mT shape that the CR-based closure tests do not see, and the five free parameters of f(x) could absorb or mimic a broad signal of the same width as the expected 200-1000 GeV resonance. However, this is a caveat rather than a demonstrated flaw: the fit quality in the SR is good (p-values 0.26 and 0.74), and the spurious-signal tests, though performed in the CR, establish that the functional form has adequate flexibility for the expected background. A single cross-check, refitting with a six-parameter extension and recomputing the exclusion contour, would directly test whether the central exclusion region is robust to the functional-form choice. Since the analysis otherwise has strong internal validations and a clean null result, this concern does not change the accept verdict.","tokens_in":53554,"tokens_out":9185,"duration_ms":101774,"concrete_test":"Recompute the SRPFN and SRAD background-only and signal-plus-background fits of mT using a six-parameter extension of Eq. (7.1), for example adding a p6 ln^3 x term inside the exponent of x, and regenerate the 95% CL exclusion contour in the (m_Z', R_inv) plane. If the boundary shifts by more than about 100 GeV at any R_inv, or if any grid point changes from excluded to not excluded, the five-parameter form is load-bearing and the quoted contour needs a functional-form systematic. A null shift would demonstrate that the central exclusion claim is robust to the empirical background choice.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central positive claim is the 95% CL exclusion contour (m_Z' from 2000 to 3200 GeV for R_inv from 0.2 to 0.37), obtained from signal-plus-background fits of the mT spectrum. The background is described by the empirical five-parameter function in Section 7, f(x) = p1(1-x)^p2 x^(p3+p4 ln x + p5 ln^2 x), with all parameters left free in the SR fit. The function is validated in the CR and VR and with signal injections (Section 9), but the CR requires Wj2<0.05 and the VR requires Wj2<0.05 plus inverted ML scores, while the SR requires Wj2>0.05 and PFN score>0.6 or ANTELOPE score>0.7. The stated assumption in Section 6 that Wj2 and the ML scores are not strongly correlated with mT is verified at preselection, before the SR-defining cuts are applied. If Wj2 or the ML scores depend on mT through jet pT, multiplicity, or E_miss^T after these cuts, the SR mT shape could be sculpted in a way that the CR/VR tests do not exercise. Because the background function is fitted directly to the SR, the risk is not a normalization transfer but a functional-form bias that can absorb or create a broad excess of width comparable to the 200-1000 GeV signal resolution, shifting the quoted exclusion boundary. The spurious-signal systematic, quoted as below 0.5 sigma, is evaluated in the CR and may not cover SR-specific correlations. This is the least secure condition for the central claim; it is a caveat rather than a demonstrated failure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports a search for new physics in hadronic final states with semi-visible jets or anomalous signatures using 140 fb^-1 of 13 TeV ATLAS data. A Z' mediator decaying to two dark quarks is the benchmark signal; the analysis defines two signal regions: SRPFN, based on a supervised Particle Flow Network, and SRAD, based on the semi-supervised anomaly detector ANTELOPE. The transverse mass mT spectrum above 1.5 TeV is fitted with a five-parameter empirical function, with control and validation regions built from inverted Wj2 and ML-score requirements. No significant excess is observed: the most significant ANTELOPE excess (1700-1900 GeV) has BumpHunter p = 0.81, and the background-only fits have p-values of 0.26 (PFN) and 0.74 (ANTELOPE). Upper limits at 95% CL exclude mZ' between 2000 and 3200 GeV for Rinv between 0.2 and 0.37, and the ANTELOPE region is shown to improve sensitivity to non-SVJ benchmark signals.","tokens_in":54004,"tokens_out":7239,"duration_ms":73673,"significance":"If the exclusion is taken at face value, this is the first ATLAS limit on resonant semi-visible jet production and the first ATLAS search to use a semi-supervised permutation-invariant anomaly detector in a resonant hadronic final state. The analysis is carefully executed: the signal regions are blinded until the background model is validated; the CR/VR fits, Asimov pseudo-data tests, and signal-injection linearity checks are documented; and the systematics treatment includes PDF/alpha_s, generator, luminosity, jet energy scale, and spurious-signal uncertainties. The ANTELOPE comparison with alternate models (emerging jets, gluino R-hadrons) is a useful, falsifiable demonstration of model breadth. The main fragility, discussed below, is that the background closure tests never exercise the Wj2 > 0.05 SR selection.","major_comments":[{"comment":"The CR and both VRs require Wj2 < 0.05, while either SR requires Wj2 > 0.05 and a high ML score; consequently the five-parameter functional form f(x) is validated only in the low-Wj2 phase space. The weak-correlation assumption between Wj2, ML scores, and mT is checked at preselection only, so a Wj2-dependent change in the mT shape induced by the SR cuts (for example through jet pT, track multiplicity, or E_T^miss) would not be detected by the CR/VR fits, the spurious-signal evaluation, or the signal-injection linearity tests. Because the background parameters are left free in the SR fit, such sculpting could either mimic or absorb a broad resonance and shift the claimed exclusion (mZ' 2000-3200 GeV, Rinv 0.2-0.37). I recommend adding a closure test in a Wj2 > 0.05 validation region with inverted ML scores, or a quantitative post-fit check of the Wj2/ML-score versus mT correlation after SR-like requirements.","section":"Section 6, Table 2, Section 9"}],"minor_comments":[{"comment":"The word \"metholodogy\" should be \"methodology\".","section":"Section 5.2"},{"comment":"The figure caption includes \"CWoLa for anomaly detection\" and two citations that are not discussed in the text; please remove or explain this reference.","section":"Figure 5"},{"comment":"The sentence beginning \"SRPFN Signal interpretations are extracted...\" lacks proper spacing and capitalization; it should read \"SRPFN signal interpretations are extracted from SRPFN only via signal-plus-background fits\".","section":"Section 9"},{"comment":"The sentence \"a selection of > 0.7 is imposed on the ANTELOPE score to maximally enrich the signal sensitivity for the SVJ simulated samples (defined in Section 6)\" is confusing because the threshold is actually introduced here in Section 5.2; please clarify the cross-reference.","section":"Section 5.2"},{"comment":"The variable Wj2 is used in Figure 5 before its definition is given in the text; moving the definition earlier would improve readability.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"This is a well-executed experimental search with a clearly presented null result. The one substantive issue is the absence of background closure in the Wj2 > 0.05 phase space, which is directly relevant to the exclusion claim; I regard it as fixable in revision rather than a reason to reject. The paper fits the journal's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the bottom line: this is a clean, well-run null-result search, and the two things you should know are that it is the first ATLAS search for resonantly produced semi-visible jets and the first time the semi-supervised ANTELOPE anomaly detector has been run on real ATLAS data. The central no-excess claim and the exclusion contour hold up on my reading.\n\nWhat's actually new: the resonant SVJ search closes a gap relative to CMS and extends ATLAS's non-resonant SVJ paper; the PFN/ANTELOPE comparison quantifies how much a signal-informed AD region gains on non-SVJ models (emerging jets, gluino R-hadrons) — roughly an order of magnitude better limits in the AD region. That is a useful benchmark for the community.\n\nWhat's done well: the analysis is carefully executed. Blinding is real; CR/VR fits are validated on downsampled data and pseudo-data; signal injection shows linear extraction from 1 to 5 sigma; spurious-signal is quoted below 0.5 sigma; systematics are thorough. The unblinded fit quality (p = 0.26 PFN, 0.74 ANTELOPE) and BumpHunter p = 0.81 for the most interesting ANTELOPE window are consistent with background-only.\n\nSoft spots, in proportion: the five-parameter mT background function is validated in the CR/VR, both of which require Wj2 < 0.05, while the SR requires Wj2 > 0.05 plus a high ML score. The paper's stated assumption that Wj2 and the ML scores are not strongly correlated with mT is checked at preselection, before those SR cuts. The stress-test note is right that this is the least secure point: a functional-form bias in the SR could shift the quoted limit boundary. I read this as a caveat, not a demonstrated failure — the good SR fit quality and the agreement between two independently defined regions mitigate it. But it is the place a referee should look hardest.\n\nThe 'anomaly detection' label deserves a footnote. ANTELOPE's latent space is seeded by the supervised PFN trained on SVJ, so the AD region is partially signal-informed. The paper says this honestly, and the alternate-signal results show real breadth, but 'model-agnostic' would be an overstatement. The citation pattern is fine; the ANTELOPE method is a self-citation to the methods paper, but this is the first application, so that is legitimate.\n\nWho this is for: dark-sector and hidden-valley experimentalists, anomaly detection developers, and phenomenologists who need the (mZ', R_inv) exclusion map. It deserves a serious referee; my recommendation: accept, and engage with the SR-validation caveat in your notes rather than treating it as an error.","headline":"First ATLAS search for resonantly produced semi-visible jets and first real-data application of the ANTELOPE anomaly detector — a clean null result whose only real soft spot is that the SR background function is not validated in the SR's jet-width regime.","tokens_in":54499,"tokens_out":4631,"would_cite":true,"duration_ms":40015,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper reports a search for semi-visible jets from a dark QCD sector and finds no excess, setting the first limits on $Z'$-mediated semi-visible jet production at hadron colliders.","keywords":["semi-visible jets","dark QCD","hidden valley","anomaly detection","particle flow network","transverse mass","Z-prime mediator","LHC Run 2"],"falsifier":"Re-run the analysis on the full Run 3 data set and check whether a localized excess appears in the transverse-mass spectrum of either signal region; a BumpHunter p-value below 0.01 on an unblinded spectrum, or a greater-than-5-sigma local excess in a signal-plus-background fit, would contradict the paper's null result. A cheaper check is to repeat the background fit with a different functional family (for example, a binned control-region template) and see whether the observed limits and the most significant ANTELOPE excess move beyond the quoted uncertainties.","tokens_in":53371,"feed_emoji":"⚛️","tokens_out":8940,"duration_ms":90241,"temperature":0.7,"pith_summary":"This paper reports a search for a strongly coupled hidden dark sector at the Large Hadron Collider, using 140 fb$^{-1}$ of proton-proton collisions at $\\sqrt{s}=13$ TeV. The signal under study is a heavy $Z'$ mediator decaying to two quarks of a dark QCD theory, each of which showers into a semi-visible jet: some hadrons decay back to Standard Model particles while others escape the detector, leaving missing energy spread through the jet. Two machine-learning selectors define the search regions: a supervised Particle Flow Network tuned to simulated semi-visible jets, and ANTELOPE, a semi-supervised anomaly detector trained on data without committing to one signal model. The central claim is that no significant excess appears over the smoothly falling transverse-mass background in either region, so the paper sets the first limits from this experiment on $Z'$-mediated semi-visible jets, excluding mediator masses from 2000 to 3200 GeV for invisible fraction $R_{inv}$ between 0.2 and 0.37 at 95% confidence. This matters because it narrows the allowed parameter space of dark-sector models and demonstrates that an anomaly-detection region can complement a dedicated signal region for generic new-physics searches.","feed_headline":"Dark showers leave no trace in 140 fb⁻¹ of LHC data","feed_subtitle":"Z′ mediators from 2 to 3.2 TeV are ruled out; the anomaly-detection region matches the dedicated search.","key_machinery":"The statistical analysis is built on the transverse mass $m_T$, computed from the two leading jets and the missing transverse energy, and on a five-parameter background function $f(x) = p_1(1-x)^{p_2} x^{p_3+p_4 \\ln x + p_5 \\ln^2 x}$ with $x = m_T/\\sqrt{s}$, whose parameters float in every fit. Signal selection relies on two permutation-invariant networks that take the up-to-160 highest-transverse-momentum tracks in the two leading jets as an unordered set: the Particle Flow Network (a DeepSet classifier that encodes each track and sums the encodings) and ANTELOPE, which feeds the pre-trained PFN latent-space encoding into a variational autoencoder trained on data, producing an anomaly score. The subleading jet width $W_{j2}$ separates the signal region from the background control region. BumpHunter, run on the rebinned $m_T$ spectrum, quantifies the significance of any adjacent-bin excess without assuming a signal shape.","core_discovery":"The paper's central claim is that, in 140 fb$^{-1}$ of 13 TeV proton-proton collisions, the $m_T$ spectra in both machine-learning-selected signal regions are consistent with Standard Model background. The background-only fit to the transverse mass distribution has good quality in the PFN region (p = 0.26) and in the ANTELOPE region (p = 0.74). The most significant excess in the anomaly region, found by BumpHunter between 1700 and 1900 GeV, has a p-value of 0.81, so no Gaussian resonance is indicated. Consequently the paper reports 95% CL upper limits on the production cross section times branching ratio for $Z'$ mediators decaying to semi-visible jets, excluding $m_{Z'}$ from 2000 to 3200 GeV for $R_{inv}$ values from 0.2 to 0.37. The ANTELOPE region, though not used for those limits, gives better sensitivity than the dedicated region to alternative models such as emerging jets and gluino R-hadrons, by factors of roughly three to five in event enrichment.","pith_inferences":["The paper does not combine the PFN and ANTELOPE regions in one fit; a combined likelihood would likely improve the expected exclusion modestly and is a natural next step.","Because ANTELOPE's score correlates with track displacement and event-level kinematics, the same trained model could be reinterpreted for displaced or long-lived dark-hadron signatures without retraining, something the paper's emerging-jet injection tests hint at but do not exploit.","A closure test using a completely different background model, such as a binned template extrapolated from the control region or a machine-learned density, would show how much of the quoted exclusion depends on the chosen five-parameter function."],"forward_implications":["A $Z'$ decaying to dark quarks with mass below about 3.2 TeV and invisible fraction up to 0.37 is excluded, so benchmark dark-QCD models in that region cannot account for the data.","The five-parameter background function passes closure tests in control and validation regions and reproduces injected signals linearly, which supports using the same data-driven fitting strategy for future resonant searches.","The ANTELOPE region, though about a factor of two weaker than the dedicated PFN region for semi-visible jets, is roughly an order of magnitude better for emerging-jet and gluino R-hadron benchmarks, demonstrating that semi-supervised anomaly detection can broaden coverage.","The observed and expected limits agree, meaning the data contain no hint of a signal that would weaken the exclusion."],"supporting_citations":[{"why":"defines the semi-visible jet signature and the $R_{inv}$ parameter that sets the invisible fraction.","marker":"[13]"},{"why":"supplies the benchmark dark-shower model parameters used for the signal grid.","marker":"[16]"},{"why":"provides the Particle Flow Network architecture used to classify semi-visible jets from track-level inputs.","marker":"[65]"},{"why":"introduces ANTELOPE, the semi-supervised permutation-invariant anomaly detector used for the second signal region.","marker":"[70]"},{"why":"provides the variational autoencoder used in ANTELOPE's unsupervised stage.","marker":"[71]"},{"why":"supplies the empirical five-parameter background function used for the $m_T$ fit.","marker":"[75]"},{"why":"provides BumpHunter, used to quantify the significance of adjacent-bin excesses in the anomaly region.","marker":"[79]"},{"why":"provides the CL$_s$ method used to set the 95% confidence upper limits.","marker":"[80]"}],"fun_headline_variants":["No sign of dark showers in ATLAS's 140 fb⁻¹ search","ATLAS excludes dark shower Z' up to 3.2 TeV","Anomaly region yields no excess in dark jet search","Semi-visible jets no-show in LHC collisions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the smooth five-parameter function used for the background describes the transverse-mass distribution in the signal regions, even though it was validated mainly on control and validation regions; if the jet-width or machine-learning selections sculpt the high-mass tail more strongly than those tests show, the quoted limits and significances would shift.","fun_headline_variants_meta":{"raw":{"variants":["No sign of dark showers in ATLAS's 140 fb⁻¹ search","ATLAS excludes dark shower Z' up to 3.2 TeV","Anomaly region yields no excess in dark jet search","Semi-visible jets no-show in LHC collisions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000637,"raw_usage":{"total_tokens":2985,"prompt_tokens":1041,"completion_tokens":1944,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":1869}},"tokens_in":657,"tokens_out":1944,"duration_ms":19194,"temperature":1.0,"reasoning_tokens":1869,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:13:37.541965+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the analysis on the full Run 3 data set and check whether a localized excess appears in the transverse-mass spectrum of either signal region; a BumpHunter p-value below 0.01 on an unblinded spectrum, or a greater-than-5-sigma local excess in a signal-plus-background fit, would contradict the paper's null result. A cheaper check is to repeat the background fit with a different functional family (for example, a binned control-region template) and see whether the observed limits and the most significant ANTELOPE excess move beyond the quoted uncertainties.","supporting_citations":[{"cited_title":"Semi-supervised permutation invariant particle-level anomaly detection","cited_arxiv_id":"2408.17409","evidence_quote":"introduces ANTELOPE, the semi-supervised permutation-invariant anomaly detector used for the second signal region."}],"review_version":1}