{"id":"42413c36-ff09-455f-808d-4aff1103e35e","arxiv_id":"1909.00761","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For 2015-2018 LHC data, ATLAS electron and photon triggers select more than 75-96% of candidate electrons and photons above threshold, with the full efficiency curves and systematic uncertainties reported.","lead":"ATLAS measured how well its electron and photon triggers catch the particles they are designed to select during the LHC's 2015-2018 data-taking, known as Run 2. The result gives physicists the correction factors needed to trust physics analyses that rely on these triggers, despite a fourfold rise in collision rate.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's claim of >96% diphoton leg efficiency at 30 GeV conflicts with Section 9.2's per-year ~95% result for 2017-2018.","rationale":"The reader's weakest assumption (Bootstrap background-photon bias) is acknowledged in the paper and covered by an assigned systematic uncertainty, so I do not dispute that as the main issue. However, the specific strongest claim highlighted from the abstract is directly contradicted by a quantitative statement in Section 9.2 and possibly by Figure 9: the 25 GeV leg is quoted as ~95% efficient for 2017-2018 at ET at least 5 GeV above threshold, while the abstract claims \"more than 96%\" at exactly 30 GeV. This is an internal inconsistency in the central summary, not a disagreement with an external consensus. The underlying measurements may be sound, but the headline claim needs to be corrected or qualified by year. Therefore a conditional acceptance is the appropriate verdict: accept the measurements after the abstract/conclusion claim is reconciled with the per-year result and the Table 6 / Section 9.2 ambiguity is resolved.","tokens_in":59347,"tokens_out":7749,"duration_ms":80542,"concrete_test":"Extract the four per-year efficiencies from Figure 9 at ET = 30 GeV from the underlying data or a high-resolution figure. If the 2017 or 2016 value is below 0.96, revise the abstract and conclusion to state the efficiency per year or to attach a year qualifier. In parallel, resolve the Table 6 versus Section 9.2 discrepancy about whether 2016 used the 'loose' or 'medium' online photon WP, since that determines which years the ~95% statement actually applies to.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's headline result (abstract and Section 13) that the 25 GeV leg of the primary diphoton trigger is \"more than 96%\" efficient for an offline photon of 30 GeV is not supported by the paper's own per-year result. Section 9.2, describing Figure 9, states that for 2017-2018, when the online 'medium' photon identification WP was used, the leg is \"~95% efficient for events with offline 'tight' isolated photons with ET at least 5 GeV above the trigger threshold\". For the 25 GeV leg, 5 GeV above threshold means ET >= 30 GeV, so this directly contradicts the \"more than 96%\" figure. The text is also internally inconsistent about when the medium WP started: Table 6 lists g35_medium_g25_medium already for 2016, while Section 9.2 says \"During 2015 and 2016 'loose' identification was used ... 2017-2018 'medium'\". If 2016 was already medium, then the sentence attributing the ~95% efficiency to a 2017-2018 tightening is wrong; if 2016 was still loose, Table 6 is wrong. Either way, the unqualified >96% statement needs a year-by-year qualification or correction.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports the performance of the ATLAS electron and photon triggers during LHC Run 2 (2015–2018) proton-proton and heavy-ion data-taking. It describes the evolution of trigger thresholds, identification working points, and isolation requirements, and it presents measured trigger efficiencies as functions of offline electron/photon ET, pseudorapidity, and pile-up, using data-driven tag-and-probe, bootstrap, and radiative-Z methods. The headline results include a single-electron trigger efficiency of at least 75% at 31 GeV rising to 96% at 60 GeV, a claim that the 25 GeV leg of the primary diphoton trigger is more than 96% efficient for a 30 GeV tight isolated photon, and heavy-ion efficiencies of at least 84% (electrons) and 95% (photons) at 5 GeV above threshold.","tokens_in":59587,"tokens_out":2837,"duration_ms":26422,"significance":"The paper provides reference trigger performance numbers that are used by ATLAS physics analyses, and it documents the operational choices (e.g., the move to medium photon identification, the Ringer algorithm, L1 isolation changes) made to control rates under increasing pile-up. Its strengths include per-year efficiency measurements with statistical and systematic uncertainties, a cross-check of the two independent photon efficiency methods in Figure 4, data/MC ratios that validate the systematic procedure, and a detailed breakdown of electron trigger inefficiency sources in Table 9. If the headline diphoton efficiency claim is corrected, the paper will be a standard, useful reference for the community.","major_comments":[{"comment":"The abstract and conclusion state that the 25 GeV leg of the primary diphoton trigger is 'more than 96%' efficient for an offline photon of 30 GeV with tight identification and isolation. However, Section 9.2 states that for 2017–2018, when the online 'medium' identification WP was used, the leg is '~95% efficient for events with offline tight isolated photons with ET at least 5 GeV above the trigger threshold'. For the 25 GeV leg, 'at least 5 GeV above threshold' includes ET = 30 GeV, so the two statements are inconsistent unless the >96% value refers only to the 2015–2016 loose-WP period, which is not stated. This is a load-bearing discrepancy in the paper's central summary claim and must be resolved by year-by-year qualification or correction of the quoted number.","section":"Abstract and Section 13 vs Section 9.2, Figure 9"},{"comment":"There is an internal inconsistency about when the medium identification WP was introduced for the primary diphoton trigger. Table 6 lists g35_medium_g25_medium as the primary diphoton trigger already for 2016, while Section 9.2 says 'During 2015 and 2016 loose identification was used at the HLT for primary diphoton triggers. During 2017–2018, medium identification was used'. If 2016 already used medium, the sentence in Section 9.2 assigning the ~95% efficiency to a 2017–2018 tightening is wrong; if 2016 still used loose, Table 6 is wrong. This must be corrected because it affects the interpretation of Figure 9 and the quoted per-year efficiencies.","section":"Table 6 vs Section 9.2"}],"minor_comments":[{"comment":"The text contains a spacing typo and an unclear phrase: 'a fewGeVabove the trigger threshold' should be 'a few GeV above the trigger threshold', and the sentence structure could be improved for readability.","section":"Section 7.3"},{"comment":"Reference [43] is titled 'Neutral Networks and Learning Machines'; the correct title is 'Neural Networks and Learning Machines'.","section":"Reference [43]"},{"comment":"The figure shows statistical uncertainties only and notes that background subtraction is omitted; this is acceptable for monitoring plots, but the text should explicitly state that the quoted efficiency values in Section 8 are not used for physics corrections, to avoid any impression that the L1 efficiencies carry full systematic uncertainties.","section":"Figure 6 caption"},{"comment":"The caption lists the trigger names 'g25_medium g35_medium' without a separator; adding a comma ('g25_medium, g35_medium') would avoid ambiguity.","section":"Figure 4 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is a typical ATLAS performance paper and the underlying measurements appear sound. The main blocking issue is the inconsistency between the abstract/conclusion claim of >96% for the 25 GeV diphoton leg and the ~95% value reported in Section 9.2 for 2017–2018, together with the mismatch between Table 6 and Section 9.2 about the 2016 working point. These are likely documentation slips rather than physics errors, but because the abstract number is the headline result, they must be fixed before acceptance. I expect the revision to be straightforward."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth knowing: this is the ATLAS collaboration's official Run 2 (2015–2018) electron and photon trigger performance paper, and the first to cover the full run. It does exactly what a performance paper should: per-year efficiencies, systematic and statistical uncertainties, cross-checks between the Bootstrap and Z-radiative-decay methods, and data/MC ratios. The Ringer neural network appears in production for the first time, with a sensible comparison against the cut-based alternative, including a caveat about merged boosted electrons. The heavy-ion section is also new, showing that the centrality-corrected triggers recover high efficiency. The citation pattern is procedural and self-references are methodological, not load-bearing.\n\nWhat's soft: the stress-test concern is real. The abstract and conclusion say the 25 GeV leg of the primary diphoton trigger is more than 96% efficient for a 30 GeV offline photon. Section 9.2 and Figure 9 show ~95% for 2017–2018, when the medium WP was used; the ~96%+ only holds for the loose-WP years. In addition, Table 6 lists g35_medium_g25_medium for 2016, while Section 9.2 says loose identification was used in 2015 and 2016. These two statements can't both be right. The fix is easy – qualify the abstract by year and correct whichever of the two 2016 descriptions is wrong – but as written the abstract overstates the plateau efficiency for the last two years. That matters to anyone who uses these numbers to derive scale factors.\n\nMinor: Figure 6 and the heavy-ion plots show statistical uncertainties only; the text says background subtraction is neglected where expected negligible, which is reasonable. Figure 11's 2017 dip in efficiency is attributed to bunch structure and MPI, plausible.\n\nBottom line: this is a careful, internally consistent measurement apart from the diphoton-leg inconsistency. The data and MC agreement is documented, the systematic procedure is standard. It deserves a serious referee; I'd request a clarification of the abstract and the 2016 WP description, then accept. It will be the reference for ATLAS Run 2 electron/photon trigger efficiencies, and I'd cite it when using those triggers.","headline":"ATLAS's full Run 2 electron/photon trigger paper is a solid reference measurement, but the abstract overclaims the diphoton leg efficiency relative to the paper's own per-year numbers.","tokens_in":60075,"tokens_out":3209,"would_cite":true,"duration_ms":31591,"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":"ATLAS electron and photon triggers kept recording physics as LHC luminosity quadrupled in Run 2.","keywords":["electron trigger","photon trigger","trigger efficiency","LHC Run 2","ATLAS","pile-up","tag-and-probe","heavy-ion collisions"],"falsifier":"Measure the same trigger leg in data with the high-purity radiative $Z\\to\\ell\\ell\\gamma$ method and with the Bootstrap method in the same $E_{\\mathrm{T}}$ bins just above threshold; if the two differ by more than the assigned systematic uncertainty, the Bootstrap bias assumption is falsified. As a second check, if the data-to-simulation efficiency ratios for tight photons deviate from unity by more than the quoted uncertainties in a full-simulation $H\\to\\gamma\\gamma$ sample, the simulation-faithfulness assumption fails.","tokens_in":59160,"feed_emoji":"⚛️","tokens_out":4886,"duration_ms":53446,"temperature":0.7,"pith_summary":"This paper establishes that the ATLAS electron and photon triggers recorded the physics events they were designed for even as LHC collision rates and pile-up roughly quadrupled between 2015 and 2018. It reports data-driven efficiencies measured with the full Run 2 dataset: a single-electron trigger catches at least 75% of offline-selected 31 GeV electrons and 96% at 60 GeV, and the 25 GeV leg of the primary diphoton trigger passes more than 96% of tight 30 GeV offline photons. The same is shown to hold for heavy-ion collisions once the underlying-event background is subtracted online. Because analyses depend on these trigger efficiencies for nearly every electron and photon measurement, the paper supplies the corrections and systematic uncertainties that make those measurements possible.","feed_headline":"ATLAS triggers stayed 96% efficient as LHC luminosity quadrupled","feed_subtitle":"Run 2 electron and photon triggers held their efficiency while pile-up rose to about 60 collisions per crossing.","key_machinery":"The measurement rests on two data-driven efficiency methods plus the two-level trigger chain. Electron efficiencies come from tag-and-probe: $Z\\to ee$ events supply a sample of unbiased probe electrons whose trigger decision is known, with backgrounds subtracted using the $Z$-mass distribution. Photon efficiencies use the Bootstrap method, which compares HLT acceptance on a low-threshold triggered sample with a random-trigger sample of tight offline photons, cross-checked against a very pure sample from radiative $Z\\to \\ell\\ell\\gamma$ decays. The trigger itself is a two-stage system: a hardware Level-1 calorimeter selection, then a software high-level trigger with fast calorimeter and tracking steps followed by precision reconstruction and likelihood-based identification. For electrons above 15 GeV, the fast step uses the Ringer neural-network algorithm, which encodes calorimeter energy into concentric rings and cut the CPU cost of the lowest-threshold single-electron trigger by at least 50%.","core_discovery":"The central claim is that the trigger menus were re-optimized year by year so that a fourfold luminosity increase and average pile-up near 60 interactions per crossing did not come at the price of trigger efficiency. In proton-proton collisions, the single-electron trigger combination is at least 75% efficient relative to a tight offline electron selection at $E_{\\mathrm{T}}=31$ GeV and rises to 96% at 60 GeV; the 25 GeV leg of the primary diphoton trigger is more than 96% efficient for a tight, isolated offline photon at $E_{\\mathrm{T}}=30$ GeV. In heavy-ion data, the primary electron and photon triggers are at least 84% and 95% efficient, respectively, at 5 GeV above threshold. The measured data-to-simulation efficiency correction factors stay below about 4% above 40 GeV in most of the detector, so analyses can apply them with small residual uncertainty.","pith_inferences":["The flat efficiency up to $\\langle\\mu\\rangle\\sim 60$ suggests the same trigger architecture has headroom for the higher luminosities of later LHC runs, provided the isolation and identification thresholds are periodically retuned.","The Bootstrap bias from background photons, which the paper argues is small, could be probed directly by extending the radiative-$Z$ method to higher $E_{\\mathrm{T}}$, where it currently does not reach.","The success of per-cell underlying-event subtraction for heavy-ion photon triggers suggests the same technique could push heavy-ion trigger thresholds below 15 GeV in future runs.","Because the Ringer algorithm was trained on 2017 data for 2018 running, periodic retraining on the most recent collision data may become a standard operational step in later data-taking periods."],"forward_implications":["Analyses using a single electron above $E_{\\mathrm{T}}=60$ GeV can treat the trigger as essentially fully efficient, with data-to-simulation corrections below 4% above 40 GeV.","The primary diphoton trigger supports Higgs-boson analyses down to 30 GeV photons with more than 96% per-leg efficiency, so the trigger adds little to the offline acceptance loss.","The measured efficiency correction factors, typically known to about 0.1%, allow simulation-based analyses to model the trigger without large systematic penalties.","Heavy-ion analyses can trigger on electrons and photons at 15-20 GeV thresholds with centrality-independent efficiency once the underlying-event subtraction is applied online.","The Ringer algorithm reduced the CPU demand of the lowest-threshold single-electron trigger by at least 50%, freeing high-level-trigger resources for the rest of the menu."],"supporting_citations":[{"why":"Defines the two-level trigger architecture and the 2015 baseline that the Run 2 evolution extends.","marker":"[3]"},{"why":"Supplies the offline electron and photon reconstruction, identification working points, and offline efficiencies used as the reference for trigger efficiency.","marker":"[31]"},{"why":"Provides the electron tag-and-probe efficiency methodology and the systematic-variation recipe applied in Run 2.","marker":"[41]"},{"why":"Origin of the tag-and-probe method that the electron efficiency measurements follow.","marker":"[46]"},{"why":"Provides the photon identification efficiency framework supporting the Bootstrap-method assumptions.","marker":"[40]"},{"why":"Supplies the electron and photon energy calibration used in the HLT precision reconstruction step.","marker":"[38]"}],"fun_headline_variants":["ATLAS triggers hold efficiency as LHC luminosity quadruples","Electron-photon trigger efficiency stays high in LHC Run 2","ATLAS keeps triggers efficient despite fourfold LHC luminosity","Trigger efficiency above 96% as LHC pile-up hits 60","ATLAS trigger menus tame LHC's Run 2 luminosity surge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For the Bootstrap photon efficiency, the paper assumes that background photons passing the tight offline identification are caught by the trigger almost as often as genuine signal photons, so the low-purity Bootstrap sample is only mildly biased.","fun_headline_variants_meta":{"raw":{"variants":["ATLAS triggers hold efficiency as LHC luminosity quadruples","Electron-photon trigger efficiency stays high in LHC Run 2","ATLAS keeps triggers efficient despite fourfold LHC luminosity","Trigger efficiency above 96% as LHC pile-up hits 60","ATLAS trigger menus tame LHC's Run 2 luminosity surge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1579,"prompt_tokens":969,"completion_tokens":610,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":521}},"tokens_in":585,"tokens_out":610,"duration_ms":6536,"temperature":1.0,"reasoning_tokens":521,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:36:06.307059+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the same trigger leg in data with the high-purity radiative $Z\\to\\ell\\ell\\gamma$ method and with the Bootstrap method in the same $E_{\\mathrm{T}}$ bins just above threshold; if the two differ by more than the assigned systematic uncertainty, the Bootstrap bias assumption is falsified. As a second check, if the data-to-simulation efficiency ratios for tight photons deviate from unity by more than the quoted uncertainties in a full-simulation $H\\to\\gamma\\gamma$ sample, the simulation-faithfulness assumption fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the offline electron and photon reconstruction, identification working points, and offline efficiencies used as the reference for trigger efficiency."}],"review_version":1}