{"id":"038b4865-a387-4987-bf64-5add98d2f74f","arxiv_id":"2508.10224","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Quantile regression applied to a pre-trained anomaly score removes its correlation with primary trigger observables, restoring trigger efficiency for low-mass anomalies.","lead":"DECADE is a new step for LHC triggers: it reshapes an anomaly score so that the trigger threshold no longer tracks the momentum-like variables that already decide what gets recorded. The goal is to keep low-mass, exotic collision events that current triggers reject, at negligible added computing cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Independence guarantee depends on conditional-quantile calibration; without a calibration check in the reported results, the low-mass efficiency gain from DECADE is unsubstantiated.","rationale":"The reader's verdict is UNVERDICTED because the full text is corrupted; my analysis of the abstract reaches the same epistemic state. The most load-bearing assumption is indeed the calibration of the quantile-regression model. However, the reader's formulation includes concern about performance on signal-like events not seen in training; that is not the critical failure mode. The threshold is set from background quantiles, so the independence property is defined on background events. Signal events affect only the numerator of the trigger efficiency, not the background rejection. The real risk is miscalibration of the background conditional quantile in the sparse, low-momentum regions, which would make the threshold momentum-dependent and undermine both background rejection and low-mass signal efficiency. This is a concrete, checkable property. Since the abstract reports no calibration curves, and the full text is unavailable, the central claim cannot be accepted or rejected at this stage; the appropriate verdict remains UNVERDICTED. The concrete test I propose would resolve the concern: if the calibration holds across low-pT bins, the independence claim is supported; if not, the low-mass efficiency gain is called into question.","tokens_in":7827,"tokens_out":4791,"duration_ms":53916,"concrete_test":"Using a large independent background sample (e.g., the LHC Olympics dataset or ATLAS Open Data with a standard QCD sample), train DECADE on a disjoint training subset. Then bin the test background by the primary trigger observable (e.g., leading-jet pT, in 5 GeV bins from 20 to 100 GeV). For each bin, compute the empirical fraction of events with anomaly score exceeding the predicted tau-quantile (tau = 0.90, 0.95, 0.99). Under exact independence, these fractions should equal 1 - tau within statistical error. Report the maximum absolute deviation across bins and a chi-square test of uniformity, focusing on the low-pT bins below the existing trigger threshold. If any bin with more than 1000 events deviates by more than 0.5 percentage points from the target, the independence 'guarantee' fails and the claimed low-mass efficiency gain is not reliable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DECADE 'guarantees' independence of the anomaly-score threshold from trigger observables. That guarantee holds if and only if the quantile-regression ensemble, trained on background, yields the exact tau-th conditional quantile of the anomaly score as a function of the trigger observables, pointwise in phase space. Decision-tree ensembles are approximate, step-function regressors; in low-momentum bins with sparse training data (precisely the low-mass region the method targets), the estimated quantile can be biased and high-variance. If the conditional quantile is underestimated in a low-pT bin, the threshold is too low, the trigger rate there exceeds the design value, and bandwidth may be exceeded; if overestimated, the threshold is too high and the low-mass signal is vetoed. The abstract provides no calibration diagnostic, so the claimed 'guarantee' is not established. The supplied full text is corrupted, so I cannot verify whether such diagnostics exist in the body; as presented, the central efficiency claim rests entirely on an un shown property of the quantile model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DECADE, a post-processing layer for anomaly-detection triggers that applies quantile regression to the anomaly score conditional on primary trigger observables. The claimed result is that the anomaly-score threshold becomes independent of primary trigger observables, recovering efficiency in low-momentum regions, with negligible additional latency and resource cost. The supplied record consists of an abstract and a largely unreadable, corrupted full text, so the detailed derivation, simulation setup, and numerical evidence cannot be checked.","tokens_in":8081,"tokens_out":2634,"duration_ms":28369,"significance":"If the guarantee were established, DECADE would be a useful contribution: it offers a simple, computationally cheap decorrelation layer for low-mass anomaly searches. The quantile-regression-on-trees idea is plausible, and the trigger-latency claim is in principle testable. However, the central guarantee is exactly as strong as the calibration of the conditional quantile estimator, and no calibration, efficiency, or latency validation appears in the available record.","major_comments":[{"comment":"The central claim that DECADE 'guarantee[s] the independence of the threshold on the anomaly score with respect to primary trigger observables' is not established. Quantile regression with decision-tree ensembles returns approximate, stepwise conditional quantiles; the guarantee holds only if the fitted quantile equals the true q-th conditional quantile pointwise in phase space. The record contains no calibration curves, no coverage diagnostics, no comparison of trigger rate versus primary observables, and no low-momentum efficiency plots. Without these, the claimed low-mass efficiency gain is unsubstantiated. Please provide such checks or weaken the claim.","section":"Abstract"},{"comment":"The supplied full text is corrupted and largely unreadable: large portions are mojibake, and the internal header 'arXiv:2508.10225v2 [math.NT] 30 Sep 2025' does not match the manuscript ID. It is impossible to verify the derivation, simulation parameters, or results. A readable, self-contained manuscript is required for review.","section":"Full text"},{"comment":"The abstract states that the authors 'demonstrate that DECADE would add an insignificant additional latency and resource cost' to existing and proposed hardware triggers. No latency numbers, resource-utilization figures, or FPGA synthesis results appear in the available record. This is a load-bearing part of the practical claim and must be backed by concrete measurements or estimates.","section":"Abstract / Full text"}],"minor_comments":[{"comment":"The word 'guaranteeing' is too strong for an approximate machine-learning estimator; consider 'aims to enforce' or 'approximately decorrelates' unless exactness is formally proven.","section":"Abstract"},{"comment":"The text contains an extraneous arXiv identifier and corrupted encoding. Please clean the source and ensure the manuscript compiles without mojibake.","section":"Full text"},{"comment":"Formal definitions of the primary trigger observables, the anomaly score, and the quantile-regression loss would improve readability; the abbreviated notation in the corrupted text cannot be followed.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The record is not in reviewable condition: the full text is corrupted and appears to contain a different arXiv header. I recommend requiring the authors to submit a clean, readable version before any further assessment. The scope is appropriate for hep-ex, but the central calibration and latency evidence must be supplied."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: DECADE is a sensible new combination of known pieces: quantile regression applied to the output of a pre-trained anomaly score, decorrelating the trigger threshold from primary trigger observables, implemented with tree ensembles for software and FPGA triggers. The problem it targets is real — anomaly scores in LHC triggers carry correlations with trigger observables, so low-mass signals get suppressed. Making the threshold observables-independent is a practical fix, and the abstract's latency claim is concrete enough to matter.\n\nWhat is genuinely new is the packaging. The ingredients individually exist, but the specific recipe of post-training quantile regression on a frozen anomaly score, at trigger level, is not something I have seen as a dedicated study. That makes it worth referee time.\n\nThe soft spot is the word 'guaranteeing.' Independence of the threshold is only as good as the quantile regressor's conditional calibration, pointwise in phase space. The abstract shows no calibration curves, no efficiency turn-ons, no rate numbers. Tree ensembles are step functions; in low-momentum bins with sparse training data the estimated quantile can be biased. If it is wrong, you either exceed your trigger rate or veto the low-mass signal you were trying to recover. The stress-test note raises exactly this, and it is not a misreading: the guarantee claim is structural, not empirical, unless the paper demonstrates calibrated quantiles. Given the full text we received is corrupted mojibake with an unrelated header, I cannot check whether the body provides those diagnostics. That is an artifact of our record, not necessarily a flaw of the paper, but the abstract alone does not substantiate the central claim.\n\nThat said, I do not think this is a hopeless paper. The method is plausible, the problem is real, and if the authors show quantile calibration — especially in the low-pT corners — plus honest rate curves, this becomes a genuinely useful contribution to trigger-ML.\n\nWho this is for: people working on online anomaly detection at LHC and trigger designers for HL-LHC. It deserves a serious referee, not a desk reject. My recommendation: send it to review, with a referee instruction to check the calibration evidence. If you work in this subfield, keep an eye on it; if not, it is not essential reading yet.","headline":"DECADE is a plausible, well-motivated recipe for decorrelating anomaly trigger thresholds, but the 'guarantee' overstates what uncalibrated quantile regression can promise — and the corrupted full text blocks verification.","tokens_in":8526,"tokens_out":3927,"would_cite":true,"duration_ms":35251,"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 introduces DECADE, a trigger-stage method that uses quantile regression to make the anomaly-score threshold independent of primary trigger observables, so LHC triggers can record low-mass anomalies without momentum bias and with","keywords":["anomaly detection","trigger","quantile regression","decorrelation","LHC","low-mass new physics","real-time event selection","FPGA"],"falsifier":"Take a held-out background sample, bin it finely in a primary trigger observable such as leading-object $p_T$, and compare the DECADE pass fraction to the target $1-\\alpha$ in each bin. If the pass fraction departs from target in the lowest-$p_T$ bins, the quantiles are miscalibrated, the threshold still depends on momentum, and the low-mass gain is not realised. A complementary test: inject simulated low-mass signal events and measure whether their trigger efficiency is flat in $p_T$; a rising or falling efficiency signals residual bias.","tokens_in":7738,"feed_emoji":"⚛️","tokens_out":4227,"duration_ms":48973,"temperature":0.7,"pith_summary":"The paper introduces DECADE (DECorrelated Anomaly DEtection), a trigger-level wrapper for the LHC. It takes an anomaly score produced by any unsupervised anomaly-detection algorithm and uses quantile regression to set the event-passing threshold as a conditional quantile of that score given the standard trigger observables—momenta and multiplicities. Because a fixed conditional quantile keeps the background pass rate uniform across phase space, the trigger no longer favours high-momentum anomalies; low-mass, low-momentum signals that existing triggers discard become accessible. The authors argue that the quantile-regression trees are cheap enough for both software and FPGA-based triggers and would add negligible latency and resource use to current ATLAS/CMS systems and proposed HL-LHC triggers.","feed_headline":"A wrapper makes LHC anomaly triggers momentum-blind","feed_subtitle":"DECADE keeps the background rate flat across phase space, opening low-momentum corners to anomaly searches.","key_machinery":"The central object is a conditional-quantile threshold $q_\\alpha(x)$ for the anomaly score $s$, where $x$ denotes the primary trigger observables. A decision-tree quantile-regression ensemble estimates the $\\alpha$-quantile of $s \\mid x$ from background events; DECADE passes an event when $s > q_\\alpha(x)$. This makes the threshold a function of phase space rather than a single global cut, so the background trigger rate is constant in $x$ and the anomaly-score threshold is independent of the trigger observables. The tree ensemble is what makes this cheap enough for real-time use: it can be evaluated in microseconds, fits in FPGA logic, and adds negligible latency.","core_discovery":"DECADE's central claim is that momentum bias in anomaly-detection triggers is not an unavoidable property of the anomaly score; it is an artifact of choosing a global threshold. By regressing the anomaly score's quantiles on the primary trigger observables, DECADE replaces the global cut with a per-event threshold that tracks the background distribution. An event triggers when its anomaly score exceeds the $\\alpha$-quantile predicted for its momentum/multiplicity region; under the background-only hypothesis, exactly $1-\\alpha$ of events pass in every region. The trigger efficiency for background is therefore flat by construction, while any anomaly that produces an unusually high score in a l","pith_inferences":["Editorial extension: The same decorrelation recipe transfers to any high-throughput selection threshold, not just anomaly triggers—wherever a score is correlated with a variable one wants to remain unbiased, a conditional-quantile threshold can flatten the background rate.","Editorial extension: The guarantee is only as strong as the quantile model's calibration; in sparse low-momentum regions tree ensembles can miscalibrate, so a calibration audit on held-out background is the natural next step before deployment.","Editorial extension: Because the quantile regression is trained only on background, the method assumes signal events do not distort the conditional quantile estimates; this could be tested by injecting simulated low-mass signals and checking that signal trigger efficiency is flat in momentum."],"forward_implications":["A fixed conditional quantile keeps the background trigger rate constant across all momentum and multiplicity bins, so low-$p_T$ anomalies no longer lose out to high-$p_T$ ones.","The method is model-agnostic: it can wrap any pretrained anomaly score, so existing and future anomaly detectors can be converted into trigger algorithms without retraining the detector.","Decision-tree quantile regression is cheap enough for both software triggers and FPGA hardware, adding negligible latency and resource use to current ATLAS/CMS systems and to HL-LHC proposals.","If deployed, it extends the LHC's discovery reach to low-mass signals—soft leptons, low-momentum jets, and similar topologies—that current single-object and anomaly triggers systematically discard."],"supporting_citations":[],"fun_headline_variants":["Momentum-blind LHC trigger finds low-mass anomalies","Quantile regression removes trigger momentum bias","Per-event anomaly thresholds unveil low-mass physics","DECADE: LHC trigger that ignores momentum","How to see low-mass new physics at the LHC"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the quantile-regression ensemble, trained on background events, gives correctly calibrated conditional quantiles in every phase-space region the trigger sees—especially the low-momentum, low-statistics corners where the claimed gain lives—and that this calibration survives on signal-like events it never trained on.","fun_headline_variants_meta":{"raw":{"variants":["Momentum-blind LHC trigger finds low-mass anomalies","Quantile regression removes trigger momentum bias","Per-event anomaly thresholds unveil low-mass physics","DECADE: LHC trigger that ignores momentum","How to see low-mass new physics at the LHC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000789,"raw_usage":{"total_tokens":3356,"prompt_tokens":828,"completion_tokens":2528,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":2469}},"tokens_in":572,"tokens_out":2528,"duration_ms":20566,"temperature":1.0,"reasoning_tokens":2469,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:34:38.891598+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out background sample, bin it finely in a primary trigger observable such as leading-object $p_T$, and compare the DECADE pass fraction to the target $1-\\alpha$ in each bin. If the pass fraction departs from target in the lowest-$p_T$ bins, the quantiles are miscalibrated, the threshold still depends on momentum, and the low-mass gain is not realised. A complementary test: inject simulated low-mass signal events and measure whether their trigger efficiency is flat in $p_T$; a rising or falling efficiency signals residual bias.","supporting_citations":[],"review_version":1}