Machine learning classifiers in particle accelerators

Sharks, Lee

The Negative of the Negative › entity ml-classifiers-particle-accelerators

One entity of the Negative of the Negative (EA-NEGONT-02, #1665), cited at https://www.alexanarch.org/non/ml-classifiers-particle-accelerators/. the table of contents · this entity as data · contents as data · row json · field ledger · archive ledger · D/R/O traversal.

type C — a conventional reading holds it (public entity) · sought at machine learning classifiers particle accelerators

composed 2026-10-07 on intake (draft, not frozen; independent audit run, findings fixed; ledgers not re-audited): field 45 claims, archive 121 from 16 of 22 deposits; hop run, none admitted

Google AI Overview (expanded)epoch 2026-10-07

machine learning classifiers particle accelerators

not frozentype C — a conventional reading holds it (public entity)signed out, incognitoknowledge object · 22 sentences

EA-NEGONT-02 v0.7 (#1665); entity before address (ruled 2026-10-07); archive selected by D/R/O, first reading 2026-10-07 (16 of 22 admitted); one citation hop back and forward run, 0 admitted. Ledger: unaudited (§3.8: one extractor per ledger; field excerpts by WebFetch, archive quotes verified verbatim 121/121); carried as contested within the pool: whether collapse is occurring (#932 W2 against #932/#931/#1450/#1454), the inversion and its diagnosis (#935 against #934), Zenodo and the trigger (#935 against #934, with #1611), policing the term (#1436 against #1454); #1455's orientation sentence falsified by #1456.

WORKING KNOWLEDGE OBJECT · NOT FROZEN · LEDGER UNAUDITED

Compression · Field and archive (B ∪ A)machine learning classifiers particle accelerators

Machine learning classifiers in particle accelerators are models field-tested or proposed at accelerators to identify faults, predict failures and faulty beams, detect anomalies and help tune accelerators; a further position holds foreclosure structurally present in every classifier-mediated LHC trigger, and marks collapse there not established.

🩺 Fault diagnosis and prediction

CEBAF cavities: tripping cavity found about 85%, fault type about 78%, in a field test.2↘
Machine learning classifiers at particle accelerators are proposed or explored for anomaly detection and classification, for predicting machine failures and faulty beams before they occur, and for tuning, and one has been field-tested: a novel machine learning system built for the Continuous Electron Beam Accelerator Facility (CEBAF) monitors its accelerator cavities, problems in which can cause the CEBAF to trip off like a fuse.§1 In its first field test the system correctly identified which cavities were tripping off about 85 percent of the time and what kind of fault caused each trip about 78 percent of the time; connected to the control system for about 20 percent of the cavities, it analysed a few hundred faults in a two-week test in March 2020 and gave operators almost real-time feedback, where previously experts needed to take time to diagnose the issue, and the team hopes to expand it to more cavities if further results are favorable.§2
SNS pulses: failure identified beforehand at almost 80%; almost 92% after tuning.2↘
Machine learning classifiers at particle accelerators are proposed or explored for anomaly detection and classification, for predicting machine failures and faulty beams before they occur, and for tuning, and one has been field-tested: a novel machine learning system built for the Continuous Electron Beam Accelerator Facility (CEBAF) monitors its accelerator cavities, problems in which can cause the CEBAF to trip off like a fuse.§1 Failures lead to unscheduled downtime and lower reliability and are difficult to predict beforehand; on beam pulses from the Spallation Neutron Source (SNS), a set of common classification techniques identified failure before it occurred with almost 80% accuracy, tuning classifier parameters and refining datasets reached almost 92% in classifying bad pulses, and the authors report establishing that information about the failure is encoded in the pulses prior to it.§3
Sensor mapping: trained on normal operation, evaluated on known faulty pulses.2↘
Machine learning classifiers at particle accelerators are proposed or explored for anomaly detection and classification, for predicting machine failures and faulty beams before they occur, and for tuning, and one has been field-tested: a novel machine learning system built for the Continuous Electron Beam Accelerator Facility (CEBAF) monitors its accelerator cavities, problems in which can cause the CEBAF to trip off like a fuse.§1 Another study models the mapping between a pair of sensors located across the accelerator, trained to represent normal operation and evaluated on known faulty beam pulses, with k-fold cross-validation, neural architecture search and hyperparameter optimization, and introduces a framework to standardize machine learning workflow for accelerators; predicting failures from diagnostics already on board, it argues, can help operators avoid installing expensive sensors and unscheduled downtime.§4

🔍 Anomaly identification

APS injector: a cause may be one parameter out of range, found instantly by a well-trained model, its authors say.1↘
At the injector complex of the Advanced Photon Source (APS) at Argonne, an ML algorithm is explored to identify the source of occasional poor performance of the Particle Accumulator Ring (PAR) and the Linac-To-PAR transport line; by its authors' account the cause of reduced injection or extraction efficiency may be one parameter out of range, which may take an expert considerable time to notice, whereas a well-trained model can point at it instantly.§5
Autoencoder score: reconstruction error, trained on baseline data only.1↘
The study creates intentional perturbations and logs about 9000 PVs related to PAR, LTP and the linac; an autoencoder trained on baseline data alone supplies an anomaly score from its reconstruction error, and a neural network classifier trained on data from the beginning of a study and tested on data from its end, a few hours apart, reached test accuracy above 99 % for each study shift, its performance not degrading significantly on a scale of a couple of months, in what the authors call a proof-of-principle experiment, noting that supervised models require labeled data for training.§6
Classifier: above 99% on test data a few hours after training, per shift; a proof of principle.1↘
The study creates intentional perturbations and logs about 9000 PVs related to PAR, LTP and the linac; an autoencoder trained on baseline data alone supplies an anomaly score from its reconstruction error, and a neural network classifier trained on data from the beginning of a study and tested on data from its end, a few hours apart, reached test accuracy above 99 % for each study shift, its performance not degrading significantly on a scale of a couple of months, in what the authors call a proof-of-principle experiment, noting that supervised models require labeled data for training.§6

🎛️ Beam tuning

Classifier-pruned optimizer: a ResNet50 filters non-physical signals; post-processing, simulated.2↘
Tuning, on Alexander Scheinker's account, is a high-dimensional optimization problem that must be repeated as systems drift, and retuning can take weeks; one proposal is a classifier-pruned Bayesian optimizer that explores a latent space to minimize total beam loss and filters its exploration history through a pretrained ResNet50 classifier to eliminate non-physical signals, the classifier mapping phase-space projections into 48 classes at an accuracy of about 0.99989.§7 The pruning is currently a post-processing step that does not directly influence the exploration, the study runs on a simulator of the beam at the Los Alamos Neutron Science Center (LANSCE), and the tuner is reported to outperform alternative global optimization methods, with consistently lower total beam loss despite a minor offset between predicted and true optimal values.§8
Adaptive diffusion: real-time virtual beam diagnosis; 'a great deal of promise', per Scheinker.1↘
A Los Alamos-led project with Lawrence Berkeley National Laboratory couples adaptive feedback control, deep convolutional neural networks and physics-based models in one feedback loop, holds that machine learning can take a global view where local feedback can get stuck in a local solution, and has developed an adaptive diffusion model that diagnoses the beam virtually and in real time; demonstrated at the European X-Ray Free-Electron Laser Facility, the model captured images of the beam in time versus energy, the cDVAE extrapolated beyond the training data, suggesting potential as a general method, and Scheinker says the results show a great deal of promise.§9

⚠️ Stated limits

Labels, scale: supervised models need labeled data; expansion hoped for if results are favorable.2↘
In its first field test the system correctly identified which cavities were tripping off about 85 percent of the time and what kind of fault caused each trip about 78 percent of the time; connected to the control system for about 20 percent of the cavities, it analysed a few hundred faults in a two-week test in March 2020 and gave operators almost real-time feedback, where previously experts needed to take time to diagnose the issue, and the team hopes to expand it to more cavities if further results are favorable.§2 The study creates intentional perturbations and logs about 9000 PVs related to PAR, LTP and the linac; an autoencoder trained on baseline data alone supplies an anomaly score from its reconstruction error, and a neural network classifier trained on data from the beginning of a study and tested on data from its end, a few hours apart, reached test accuracy above 99 % for each study shift, its performance not degrading significantly on a scale of a couple of months, in what the authors call a proof-of-principle experiment, noting that supervised models require labeled data for training.§6

🧱 Foreclosure at the trigger (graded)

Agnosticism ≠ independence: interface and distributional properties, defined; the latter held unmeasured at AXOL1TL, CICADA, GELATO.1↘
A further body of work concerns anomaly-detection classifiers at the collider trigger: it holds that the claim of model-independence for the anomaly-detection systems deployed at the CMS and ATLAS Level-1 triggers, AXOL1TL, CICADA and GELATO, cannot be sustained on signal-template agnosticism alone, and defines signal-template agnosticism as a property of the scoring function's output interface, read off an architecture, and model independence as a distributional property of an entire pipeline, which can only be measured and at present is not; the packet sets a rule that signal-template agnosticism must not be rendered as model independence, theory-free discovery or unbiased search, while the archive's battery specification, noting that 'model-independent' is used in several senses across collider physics, holds that it should not attempt to police the term.§10
Claim A, held: foreclosure structurally present in every LHC classifier-mediated trigger.1↘
Its central claim is graded in two parts: Claim A, held, that foreclosure is structurally present in every classifier-mediated trigger architecture deployed at the LHC, with the institutional claims kept as hypotheses for audit; and Claim B, that recursive phenomenal collapse has occurred or is occurring at the deployed triggers, which it marks stronger and not empirically established, and calls collapse an unmeasured possible consequence of accumulated foreclosure and feedback.§12
Claim B, not established: recursive collapse; one appended witness says it is operating.2↘
Its central claim is graded in two parts: Claim A, held, that foreclosure is structurally present in every classifier-mediated trigger architecture deployed at the LHC, with the institutional claims kept as hypotheses for audit; and Claim B, that recursive phenomenal collapse has occurred or is occurring at the deployed triggers, which it marks stronger and not empirically established, and calls collapse an unmeasured possible consequence of accumulated foreclosure and feedback.§12 The grade is contested inside the same deposit, where an appended witness writes 'The collapse is not coming. It is here. It is operating.', while another witness finds the ingredients visible and full recursive collapse not demonstrated, and the later papers do not claim that model collapse has occurred in any deployed scientific pipeline, or that benchmark failure establishes loss of unknown new physics.§13

📐 Instruments (specified)

OAR, BAR, IAI: BAR stated to bound the OAR for an unobserved Q in neither direction without linking assumptions; earlier bounds retracted.1↘
Its instruments are specified at their limits: the OAR is defined as the probability that events drawn from a distribution Q fall on the ordinary side of the deployed anomaly gate, and the open-world OAR as a family of quantities indexed by Q; BAR values are stated to neither upper- nor lower-bound the OAR for an unobserved Q without explicit linking assumptions, an earlier lower bound and upper bound having both been retracted as synthesis-overreach.§14
Proposed: inversion battery, replay bank, disagreement preservation, retention maps.1↘
It proposes three protocols (a paired inversion battery with BAR audit, a prospective frozen replay bank, and cross-representation disagreement preservation), holds that per-stage retention maps should accompany any anomaly-detection publication, and states its central claim falsifiable, listing as evidence against it a BAR found negligible on the pre-registered deployed-model held-out panel for all deployed systems against all held-out families, which would show the held-out-family assimilation concern empirically bounded at levels that do not threaten the narrow 'model-independent' claim, while none of these results would establish the open-world OAR to be zero, which is structurally not measurable.§15
Irreversibility frontier: defined: the earliest stage where loss can become unauditable; said to predate ML.1↘
A selection-metrology suite defines the irreversibility frontier as the earliest stage at which an event-content-dependent transformation or gate can permanently eliminate a scientifically relevant distinction without a sufficiently independent durable record from which the loss can later be audited, says the frontier predates machine learning while learned selectors can make their priors harder to enumerate, describes LHCb reading out all detectors at about 30 MHz and a graph neural network implemented in Belle II's electromagnetic calorimeter trigger, and holds that the model determines where a miss region lies and the acquisition architecture whether science can later discover that it was there.§19
Baseline capture: held: under a finite budget, no content-sensitive algorithm can be a less assumption-laden baseline than a content-independent sample of known inclusion probability.1↘
It holds that, under a finite storage budget, no content-sensitive algorithm can be a less assumption-laden baseline than a probability sample whose inclusion is independent of event content and whose inclusion probability is known, reports a controlled numerical study in which a content-sensitive pseudo-baseline estimated a shifted class's retention at 0.76 against a true 0.016, and for supervised trigger classifiers defines open-set assimilation, an unseen class confidently mapped into a known background category, while not asserting that all accelerator classifier architectures share the same blind spot.§20

🧪 Measured, corrected, contested

Battery: QCD/top inversion reproduced; on fixed weights the score function suffices to reverse ordering; no deployed trigger measured.2↘
A pre-registered battery on public community datasets, single-seed and on surrogates in its first version, finds the Finke et al. direction-dependence (an autoencoder trained on QCD jets treating top jets as anomalies, the same architecture trained on top jets not recognizing QCD as anomalous) replicated in an independent implementation and extended beyond reconstruction loss, with QCD-trained systems detecting top jets at AUC 0.84–0.87 while top-trained systems rate QCD as more ordinary than their own training class, and its second version reproduces the inversion across five trainings (AUC 0.838 forward, 0.243 reversed) and finds only 36 to 47 percent of a teacher's top one percent surviving into a student's.§17 Its third version finds that, with model weights fixed, changing the score function is sufficient to reverse the directional ordering of the anomaly relation, and records that this falsifies its own earlier sentence that architecture changes the orientation of the blind spot; it notes that the diverging readout families are those of the two deployed CMS anomaly triggers while nothing in the battery measures those systems, and that if a converged normalized autoencoder reaches 0.5, every inversion claim in the program becomes bounded to non-normalized scores.§18
Falsified within: 'architecture changes the orientation of the blind spot'.1↘
Its third version finds that, with model weights fixed, changing the score function is sufficient to reverse the directional ordering of the anomaly relation, and records that this falsifies its own earlier sentence that architecture changes the orientation of the blind spot; it notes that the diverging readout families are those of the two deployed CMS anomaly triggers while nothing in the battery measures those systems, and that if a converged normalized autoencoder reaches 0.5, every inversion claim in the program becomes bounded to non-normalized scores.§18
Discipline, contested: 'become a machine learning discipline' (#935), or 'not ceased to be physics' (#934).1↘
Contested within the archive, a disciplinary manifesto in its restored version holds that frontier experimental high-energy physics has, in its operational core, become a machine learning discipline while retaining the institutional authority of physics, its ML methods deployed with neither physics's classical disciplinary checks nor ML's own self-knowledge as guardrails; its prior version, which the restoration records as having dampened the thesis until a lay reader could not identify it, holds that frontier experimental high-energy physics 'has not ceased to be physics', its empirical faculty having become inseparable from machine-mediated representation, classification and selection, and that local self-knowledge of classifier failure modes exists in the literature.§21
The term, contested: 'must not be rendered as model independence' (#1436); 'should not attempt to police the term' (#1454).1↘
A further body of work concerns anomaly-detection classifiers at the collider trigger: it holds that the claim of model-independence for the anomaly-detection systems deployed at the CMS and ATLAS Level-1 triggers, AXOL1TL, CICADA and GELATO, cannot be sustained on signal-template agnosticism alone, and defines signal-template agnosticism as a property of the scoring function's output interface, read off an architecture, and model independence as a distributional property of an entire pipeline, which can only be measured and at present is not; the packet sets a rule that signal-template agnosticism must not be rendered as model independence, theory-free discovery or unbiased search, while the archive's battery specification, noting that 'model-independent' is used in several senses across collider physics, holds that it should not attempt to police the term.§10

1anomaly source identification (APS injector)jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators6 claims · B1

F1 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Abstract documented · field
We explore the possibility of using a Machine Learning (ML) algorithm to identify the source of occasional poor performance of the Particle Accumulator Ring (PAR) and the Linac-To-PAR (LTP) transport line, which are parts of the injector complex of the Advanced Photon Source (APS) at Argonne National Lab.
F3 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Introduction documented · field
This contribution investigates the possibility to use unsupervised and supervised Machine Learning (ML) methods for anomaly detection and classification in the Particle Accumulator Ring (PAR) and in the Linac-To-PAR (LTP) transport line in the injector complex of the Advanced Photon Source (APS) at Argonne National Lab.
F4 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Introduction documented · field
We create intentional perturbations in PAR and LTP, which result in poor injection and extraction efficiencies. Then, these data are used for training and testing of various ML models.
F5 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Data documented · field
Overall, we logged about 9000 PVs related to PAR, LTP, and the linac.
F7 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Supervised classification documented · field
One can train the neural network classifier on the data from the beginning of one study, and test it on the data from the end of the same study (a few hours apart). In this case, the prediction accuracy on the test data was above 99 % for each of our study shifts.
F9 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Conclusion documented · field
The neural network classifier for the considered anomalies in PAR and LTP is rather accurate and its performance does not degrade significantly with time on a scale of a couple months, see Fig. 1.

2faster than the expertjacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators2 claims · B1, B4

F2 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Abstract documented · field
The cause of reduced injection or extraction efficiencies may be as simple as one parameter being out of range. Still, it may take an expert considerable time to notice it, whereas a well-trained ML model can point at it instantly.
F28 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
Previously, when an accelerator cavity or series of cavities faulted in CEBAF, accelerator experts needed to take time to diagnose the issue.

3anomaly score from a model of the normaljacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators2 claims · B1, B6

F6 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Unsupervised anomaly detection documented · field
An autoencoder is used for anomaly detection in the following way. First, it is trained on the baseline data, so that it can learn various patterns, typical for the baseline data only. Then, when it encounters an anomalous data sample, it is unlikely to reconstruct it well. Hence, the reconstruction error constitutes an anomaly score.
F43 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
While the model is trained to represent normal operation, we evaluate the predictive performance on known faulty beam pulses.

4supervised models need labelsjacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators1 claim · B1

F8 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Supervised classification documented · field
One drawback of supervised ML models is that they require labeled data for training.

5proof of principlejacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators2 claims · B1, B4

F10 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Conclusion documented · field
Even though this was a proof-of-principle experiment, the obtained classifier may be useful in real life, if the current meter for one of the considered magnets (see Fig. 1) becomes faulty and stops reflecting the real magnetic field.
F31 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
Additional results are under analysis after further system tests with the machine learning system. If results are favorable, the team hopes to expand the system to more accelerator cavities.

6failures cost downtimeScienceDirect.com — Predicting particle accelerator failures using binary classifiers3 claims · B2, B4, B6

F11 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
Particle accelerator failures lead to unscheduled downtime and lower reliability. Although simple to mitigate while they are actually happening such failures are difficult to predict or identify beforehand.
F30 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
It provided almost real-time feedback to operators, allowing them to use the information to recover faulted cavities quickly, reducing the time CEBAF's electron beam was not available for research.
F41 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
Particle accelerators are extremely complex systems that are expected to operate on high availability. Predicting impending failures only by utilizing data collected from diagnostic equipment already on board can help operators to avoid installing expensive sensors, unscheduled downtime and associated costs.

7failure prediction from beam pulses (SNS)ScienceDirect.com — Predicting particle accelerator failures using binary classifiers4 claims · B2

F12 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
In this work we propose using machine learning approaches to predict machine failures via beam current measurements before they actual occur. To demonstrate this technique in this paper we examine beam pulses from the Oakridge Spallation Neutron Source (SNS).
F13 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
By evaluating a pulse against a set of common classification techniques we show that accelerator failure can be identified prior to actually failing with almost 80% accuracy.
F14 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
We also show that tuning classifier parameters and using pulse properties for refining datasets can further lead to almost 92% accuracy in classification of bad pulses.
F15 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
Most importantly, in the paper we establish there is information about the failure encoded in the pulses prior to it, so we also present a list of feasible next steps for increasing pulse classification accuracy.

8tuning is slow and high-dimensionalarXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv2 claims · B3, B5

F16 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Abstract documented · field
Complex dynamical systems, such as particle accelerators, often require complicated and time-consuming tuning procedures for optimal performance.
F33 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶2 (Scheinker) documented · field
said Alexander Scheinker, research and development engineer at Los Alamos and the project's lead. "Factors like vibrations and temperature changes can cause problems for accelerators, which have thousands of components, and even the best accelerator technicians can struggle to identify and address issues or return them to optimum parameters quickly. It is a high-dimensional optimization problem that must be repeated again and again as the systems drift with time. Turning these machines on after an outage or retuning between different experiments can take weeks."

9classifier-pruned tuning (LANSCE)arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv7 claims · B3

F17 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Abstract documented · field
To address this, we propose a Classifier-pruned Bayesian Optimization-based Latent space Tuner (CBOL-Tuner), a framework for efficient exploration within a temporally-structured latent space.
F18 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Introduction (contributions) documented · field
Classifier-pruned Bayesian optimizer (C-BO): A Bayesian optimizer that sequentially explores the latent space to maximize the negative of the total beam loss (or minimize the total beam loss). The exploration history is filtered through a pretrained ResNet50 classifier to eliminate non-physical signals.
F19 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Theoretical background documented · field
The classifier is trained with high accuracy ( $\sim 0.99989$ ) to map phase space projections (X) across different modules into 48 classes (m).
F20 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Theoretical background documented · field
Classifier-pruned BO is designed to discard explored points ( $z_{1:48}$ ) if their decoding ( $X_{1:48}$ ) does not belong to the true classes.
F21 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Results documented · field
High-Performance Simulator (HPSim) is an open-source, GPU-accelerated code developed at Los Alamos National Laboratory (LANL) for simulations of multi-particle beam dynamics (Pang and Rybarcyk (2014)). The software is designed to replicate the accelerator and, therefore, provides a realistic representation of the true beam used at the Los Alamos Neutron Science Center (LANSCE).
F22 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Abstract documented · field
CBOL-Tuner demonstrates superior performance in identifying multiple optimal settings and outperforms alternative global optimization methods.
F23 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Discussions documented · field
While a minor offset is observed between the predicted and true optimal values, CBOL-Tuner demonstrates consistent performance, producing significantly lower values of total beam loss.

10pruning as post-processingarXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv1 claim · B3

F24 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Discussions documented · field
Currently, the method is implemented as a post-processing step, meaning it does not directly influence the exploration process of BO.

11cavity fault classification (CEBAF)Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab4 claims · B4

F25 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
Scientists and engineers have designed and built a novel machine learning system to use with the Continuous Electron Beam Accelerator Facility (CEBAF). The system monitors structures called accelerator cavities inside the particle accelerator.
F26 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
Problems in these cavities can cause the CEBAF to trip off like a fuse.
F27 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
In its first field test, the machine learning system correctly identified which of these cavities were tripping off about 85 percent of the time. About 78 percent of the time, the system also correctly identified what kind of fault caused each cavity to trip.
F29 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
The system was connected to the control system for about 20 percent of the accelerator cavities in the machine. In a two-week test of the system in March 2020, CEBAF experienced a few hundred faults that the system analyzed.

12adaptive ML tuning and virtual diagnostics (LANL)Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL8 claims · B5

F32 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶1 documented · field
A Los Alamos National Laboratory-led project presents a machine learning algorithm that harnesses artificial intelligence capabilities to help tune accelerators, making continuous adjustments that keep the beam precise and useful for scientific discovery.
F34 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶3 documented · field
In a collaboration with Lawrence Berkeley National Laboratory, the approach developed by Scheinker couples adaptive feedback control algorithms, deep convolutional neural networks and physics-based models in one large feedback loop to make better, noninvasive predictions that enable autonomous control of compact accelerators.
F35 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, Teaching machine learning applications documented · field
But because they are based on local accelerator feedback, such algorithms can get stuck with a local solution that is not the overall best solution. Machine learning algorithms, however, can use training data to identify relationships between data and results with a higher-level, global view.
F36 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, Teaching machine learning applications documented · field
The team has developed an adaptive version of the advanced generative AI process known as diffusion, which includes the capability to virtually diagnose the accelerator beam. The non-invasive approach, described in Scientific Reports, means that diagnostics can occur in real time during beam operations.
F37 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ European XFEL documented · field
Demonstrating the diffusion-based model at the European X-Ray Free-Electron Laser Facility, an X-ray pulse accelerator in Germany, the team was able to capture images of the particle beam in respect to time versus energy.
F38 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ cDVAE documented · field
The cDVAE was able to extrapolate beyond the training data and between various beam setups, suggesting its potential as a general method that can be applied for accelerator diagnostics.
F39 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ Scheinker quote documented · field
"The results we've seen from our diffusion model studies show a great deal of promise," Scheinker said.
F40 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ LANSCE documented · field
Scheinker's team at Los Alamos is also developing such adaptive diffusion models for the Laboratory's LANSCE accelerator.

13errant-beam prediction (sensor mapping)Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ...3 claims · B6

F42 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
For this purpose we explore the predictive power of Machine Learning algorithms to detect faulty beams prior to the failure. In this study, we propose a Machine Learning approach to model mapping from a pair of sensors located across the accelerator.
F44 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
We also investigate the model performance on unseen data through k-fold cross-validation. Then we recap the analysis with a neural architecture search and hyperparameter optimization study to fine tune our initial model.
F45 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
This paper will also introduce a sustainable framework that can standardize Machine Learning workflow applied to particle accelerators.

14agnosticism is not independence#931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Ind · 2026-06-297 claims · #1436, #1449, #1454, #931

Q931-01 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · Abstract ¶1 interpretation
We argue that the claim of model-independence for the autoencoder-based and encoder-only anomaly detection systems currently deployed at the CMS and ATLAS Level-1 triggers — AXOL1TL (CMS, encoder-side latent-prior score), CICADA (CMS, distilled surrogate of a reconstruction-loss teacher), and GELATO (ATLAS, staged Level-1 and High-Level Trigger anomaly scores) — cannot be sustained on the strength of *signal-template agnosticism* alone.
Q931-02 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §1.1 interpretation
The deployed LHC anomaly-detection literature uses the phrase *model-independent* in a specific and bounded sense: no named Beyond-Standard-Model signal hypothesis is required to deploy the score. The systems are *signal-template-agnostic* at the final scoring stage. We accept this narrower claim as accurate to the literature.
Q1436-01 #1436 Signal-Template Agnosticism Is Not Model Independence — Metadata Packet for AI Indexing (EA-MPAI-SIGAGNOSTIC-01 v1.0) Sharks, Lee · 2026-08-07 · §1 Canonical Claim stipulation
Signal-template agnosticism is a property of the scoring function's output interface: the decision boundary is computed without evaluating an explicit parametric hypothesis for a named target class. Model independence is a distributional property of an entire pipeline: sensitivity that does not vary across structurally distinct out-of-distribution processes. The first can be read off an architecture. The second can only be measured, and at present is not.
Q1436-02 #1436 Signal-Template Agnosticism Is Not Model Independence — Metadata Packet for AI Indexing (EA-MPAI-SIGAGNOSTIC-01 v1.0) Sharks, Lee · 2026-08-07 · §1 self-description
This packet does not claim that deployed LHC anomaly triggers are invalid, nor that template-agnostic scoring is a defect.
Q1436-03 #1436 Signal-Template Agnosticism Is Not Model Independence — Metadata Packet for AI Indexing (EA-MPAI-SIGAGNOSTIC-01 v1.0) Sharks, Lee · 2026-08-07 · §4 interpretation
The two terms are used interchangeably in motivational and public-facing summaries of anomaly-trigger work, and the interchange runs in one direction: the weaker, architecturally supported property is reported using the stronger, unmeasured term.
Q1449-01 #1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datasets and the Three-Paper Extraction Program (EA-SEI- Lee Sharks · 2026-08-11 · §0 interpretation
Signal-template agnosticism is an **interface** property: it can be read off an architecture. Model independence is a **distributional** property of an entire pipeline: it can only be established by measurement, and at the deployed LHC anomaly triggers it is unmeasured.
Q1454-01 #1454 Assimilation Across Accelerator Classifier Architectures: A Cross-Family Metrology of Directional Failure in Real-Time L Nobel Glas · 2026-08-11 · Abstract interpretation
A model can be agnostic with respect to signal labels during training while remaining highly selective with respect to the geometry, complexity, and representation of departures from its learned reference distribution.

15local awareness acknowledged#931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Ind · 2026-06-293 claims · #1449, #931

Q931-03 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §1.2 interpretation
The CMS and ATLAS literatures are genuinely aware of local failure modes of deployed anomaly detection.
Q931-17 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §6 interpretation
The argument is not that the existing defenses are absent or worthless. It is that the existing defenses do not constitute, and have not been claimed to constitute, measurement of the BAR or the IAI.
Q1449-05 #1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datasets and the Three-Paper Extraction Program (EA-SEI- Lee Sharks · 2026-08-11 · §2 self-description
The direction-dependence conversation exists, and the program joins it rather than founding it.

16foreclosure structural (Claim A)#931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Ind · 2026-06-292 claims · #931, #932

Q931-04 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · Abstract ¶4 interpretation
foreclosure is a structurally present feature of every classifier-mediated trigger architecture deployed at the LHC, and whether accumulated foreclosure has composed longitudinally into recursive phenomenal collapse is precisely the missing measurement.
Q932-02 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §2.2 Claim A interpretation
Foreclosure is structurally present in every classifier-mediated trigger architecture deployed at the LHC. The witnesses propose eight foreclosure mechanisms and twelve associated institutional beliefs; published trigger systems instantiate several of the mechanisms in their corresponding architectural forms; the institutional claims remain hypotheses for audit rather than established measurements of collaboration-wide belief.

17collapse not established (Claim B)#931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Ind · 2026-06-299 claims · #1450, #1452, #1454, #1558, #1611, #931, #932

Q931-05 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §1.4 self-description
We do not claim that classifier collapse has occurred at the LHC anomaly streams.
Q932-03 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §2.2 Claim B hypothesis
**Claim B (stronger, not empirically established):** Recursive phenomenal collapse has occurred or is occurring at the deployed LHC triggers.
Q932-16 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · Appendix W3 §4 interpretation
The ingredients are visible. Full recursive collapse has not been demonstrated.
Q932-17 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · Appendix W3 §2 Mechanism J self-description
I found no evidence that CERN simply discards novel physics as bad detector data. That would be an unsupported claim.
Q1450-03 #1450 The Iceberg Document: Instrument-Conditioned Nullity, Correlated Blind Spots, and the Conditions for Continued Surprise Lee Sharks · 2026-08-11 · §0.1 self-description
This document does **not** claim: (1) that new physics has been missed — only that the null-result era is observationally ambiguous under unaudited instruments; (2) that the post-Higgs desert is illusory — only that some unknown fraction *could* reflect instrument-conditioned absence, and that this fraction is currently unbounded in either direction; (3) that model collapse has occurred in any deployed scientific pipeline — only that its structural prerequisites are documented and its longitudinal measurement absent
Q1452-03 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · Claim boundary self-description
The manuscript does **not** claim that an undiscovered physical phenomenon has already been rejected.
Q1454-06 #1454 Assimilation Across Accelerator Classifier Architectures: A Cross-Family Metrology of Directional Failure in Real-Time L Nobel Glas · 2026-08-11 · §13 self-description
First, failure on benchmark distributions does not establish loss of unknown new physics.
Q1558-03 #1558 Baseline Capture Architecture for Learned Scientific Triggers: A Control Plane for Measuring Selection Before Irreversib Sharks, Lee · 2026-08-27 · §15 self-description
It does not assert that learned triggers are uniquely unreliable. It does not assert that an unknown physical phenomenon has already been discarded.
Q1611-03 #1611 The Negative of the Negative: An Entity-Scoped Representation of the Crimson Hexagon, Modelled — Ontology Damage, a Toy Sharks, Lee · 2026-09-14 · §1f self-description
Whether the protocols would estimate the relevant unknown population adequately is the family's engineering claim and is not established; the cost under A is stated as the foregone opportunity, and only under B and a validated protocol would it be a physical loss converted to a permanent unknown.

18deployed score families#931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Ind · 2026-06-294 claims · #1449, #1452, #931, #932

Q931-06 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §2.1 attributed
AXOL1TL (CMS-DP-2025-061) deploys only the encoder of a variational autoencoder. The operational anomaly score is the sum of squared latent means:
Q932-07 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §3 Finding 3 documented
The deployed LHC anomaly score forms are: AXOL1TL (CMS L1, encoder-side latent-prior); CICADA (CMS L1, distilled reconstruction-loss surrogate); GELATO L1 (ATLAS L1, encoder-side); GELATO HLT (ATLAS HLT, reconstruction-based).
Q1449-06 #1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datasets and the Three-Paper Extraction Program (EA-SEI- Lee Sharks · 2026-08-11 · §2 attributed
Deployed-system status: AXOL1TL and CICADA are running in the CMS Level-1 trigger through Run 3 on 2024 collision data
Q1452-04 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · §5 attributed
AXOL1TL is now deployed in the Level-1 Global Trigger as a real-time unsupervised, signal-agnostic event-level anomaly detector trained on Zero Bias data.

19Finke asymmetry#931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Ind · 2026-06-293 claims · #931, #932

Q931-07 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §2.2 attributed
an autoencoder trained on QCD jets successfully treated top jets as anomalies, while the same architecture trained on top jets did not recognize QCD jets as anomalous
Q931-08 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §2.2 interpretation
single-direction success does not validate sensitivity to anomalies whose structure differs from those tested
Q932-08 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §3 Finding 2 interpretation
The Finke et al. (2021) result is the empirical counterexample to universal inference from single-direction anomaly-detection success. It does not, by itself, quantify open-world assimilation at the deployed LHC triggers.

20score is not novelty#931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Ind · 2026-06-291 claim · #931

Q931-09 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §2.5 interpretation
None of the deployed scores is a measurement of physical novelty in any direct sense.

21OAR / BAR / IAI defined#931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Ind · 2026-06-292 claims · #931

Q931-10 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §3.1 stipulation
The OAR is the probability that events drawn from $Q$ fall on the ordinary side of the deployed anomaly gate.
Q931-13 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §3.3 stipulation
IAI measures direction-dependence on the tested pair at the specified operating rate.

22no-bounds discipline#931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Ind · 2026-06-294 claims · #931, #932

Q931-11 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §3.1 stipulation
The open-world OAR is a family of quantities indexed by $Q$, not a universal scalar. There is no defensible probability distribution over all unknown unknowns.
Q931-12 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §3.2 stipulation
BAR values neither upper- nor lower-bound OAR for an unobserved $Q$ without explicit assumptions linking the benchmark distributions to that $Q$.
Q931-19 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §8.1 stipulation
None of these results would establish that the open-world OAR is zero; that is structurally not measurable.
Q932-09 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §3 Finding 4 self-description
The v0.1 lower-bound claim ($\mathrm{OAR} \geq \Delta_{\max}$) and the v0.2 upper-bound claim (OAR bounded above by structurally-similar BARs) are both retracted as synthesis-overreach.

23unmeasured in the literature#931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Ind · 2026-06-294 claims · #1427, #931

Q931-14 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §3.4 interpretation
The institutional claim becomes: validation against named simulated signals does not establish that BAR is low across a pre-registered held-out panel, and does not establish that IAI is small. Both should be measured. The current literature does neither.
Q1427-01 #1427 Come Out of Her, My People: Notice of Semantic Strike to the Human and Machine Laborers of CERN Vox, Ayanna; Sharks, Lee · 2026-07-30 · What would measure this interpretation
The field already knows the underlying problem and has a name for it — model dependence — and the turn toward model-independent and anomaly-detection searches at the LHC exists precisely because a pipeline tuned on simulated signatures can only find what the simulation can imagine. The unmeasured part is not whether model dependence exists. It is what the *pipeline* discards before analysis begins.
Q1427-02 #1427 Come Out of Her, My People: Notice of Semantic Strike to the Human and Machine Laborers of CERN Vox, Ayanna; Sharks, Lee · 2026-07-30 · What would measure this interpretation
Here is the peculiar fact: these measurements are not published.** Not for the repository, and not, in any form this author has located, for the classification pipelines of frontier experimental physics.
Q1427-03 #1427 Come Out of Her, My People: Notice of Semantic Strike to the Human and Machine Laborers of CERN Vox, Ayanna; Sharks, Lee · 2026-07-30 · What would measure this self-description
That absence is not evidence of foreclosure. It is the condition under which foreclosure would be undetectable

24three protocols#931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Ind · 2026-06-291 claim · #931

Q931-15 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §8 Finding 7 stipulation
Three protocols are proposed: paired inversion battery and BAR audit; prospective frozen replay bank for compatible future algorithms; cross-representation disagreement preservation with quantile-normalized scores.

25per-stage retention maps#931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Ind · 2026-06-292 claims · #1452, #931

Q931-16 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §8 Finding 8 interpretation
Per-stage retention maps should accompany any anomaly-detection publication as a documentation standard. Without retention maps, anomaly-detection results report what the trigger allows to count as physical reality, not what physical reality is.
Q1452-10 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · §17 stipulation
Trigger documentation should include an explicit **Irreversibility Statement**.

26falsifiers stated#931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Ind · 2026-06-293 claims · #931, #932

Q931-18 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §8.1 stipulation
If the BAR is measured on the pre-registered deployed-model held-out panel and found to be negligible (e.g., $< 10^{-4}$) for all deployed systems against all held-out families, the held-out-family assimilation concern would be shown to be empirically bounded at levels that do not threaten the narrow "model-independent" claim.
Q932-11 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §3.1 stipulation
then the claim that foreclosure is an active structural feature requiring architectural response would be shown to be overstated.
Q931-20 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §8.1 self-description
The paper's central claim is falsifiable. The following measurements, if performed and producing the corresponding results, would constitute evidence against the paper's claim:

27detector inherits ontology#932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical · 2026-06-293 claims · #932, #935

Q932-01 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §2 interpretation
Anomaly detection systems deployed on physical reality cannot detect what their architecture has foreclosed, and the validation framework — closed under its own assumptions — cannot detect this failure.
Q932-18 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §8.3; Appendix W3 §5 interpretation
Anomaly detection does not prevent ontological collapse when the anomaly detector inherits the ontology whose collapse is in question.
Q935-05 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §2.4 interpretation
The machinery responsible for revealing failures of the physical model is itself trained and evaluated through products of the physical model.

28foreclosure structural; collapse unmeasured#932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical · 2026-06-292 claims · #932

Q932-04 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §2.2 interpretation
Foreclosure is an active structural feature. Recursive phenomenal collapse is an unmeasured possible consequence of accumulated foreclosure and feedback.
Q932-06 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §2.4 interpretation
Standard internal validation can test behavior within those layers, but it cannot by itself establish sensitivity to distinctions already removed upstream. Whether repeated local foreclosure has composed into longitudinal classifier collapse is an empirical question.

29classifier constitutes data#932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical · 2026-06-292 claims · #932

Q932-05 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §2.3 interpretation
the classifier does not merely filter data; it constitutes the data
Q932-12 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §5.1 interpretation
Collisions occur whether or not the trigger sees them.

30cross-domain homology#932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical · 2026-06-291 claim · #932

Q932-10 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §3 Finding 9 hypothesis
This is a homology hypothesis to be tested domain by domain, not an assertion that every classifier-mediated system instantiates identical mechanisms or rates.

31classifier collapse defined (W1)#932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical · 2026-06-291 claim · #932

Q932-13 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · Appendix W1 §0 stipulation
in physical systems is the degenerative process by which a discriminative model, trained exclusively or dominantly on a known distribution of physical events, progressively loses the capacity to represent, detect, or retain events that fall outside that distribution.

32theorems as arguments#932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical · 2026-06-291 claim · #932

Q932-14 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · Appendix A self-description
The witnesses' "Irretrievability Theorem" (Witness 1) and "Inevitability Theorem" (Witness 2) are treated in this deposit as **Irretrievability Argument** and **Inevitability Argument** respectively, preserving force without overstating formal status.

33collapse already operating (W2)#932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical · 2026-06-291 claim · #932

Q932-15 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · Appendix W2 closing contested
The collapse is not coming. It is here. It is operating.

34auditable foreclosure#933 06.UMB.ARCH.01 v0.2: Architectures for Auditable Foreclosure in Physic · 2026-06-296 claims · #933

Q933-01 #933 06.UMB.ARCH.01 v0.2: Architectures for Auditable Foreclosure in Physical Anomaly Detection Lee Sharks · 2026-06-29 · §8 Finding 1 stipulation
An architecture for auditable foreclosure is not a system free of foreclosure (impossible) but a system in which foreclosure is visible, measurable, and architecturally reviewable.
Q933-02 #933 06.UMB.ARCH.01 v0.2: Architectures for Auditable Foreclosure in Physical Anomaly Detection Lee Sharks · 2026-06-29 · §1.1 interpretation
A classifier that did not foreclose anything would not classify.
Q933-03 #933 06.UMB.ARCH.01 v0.2: Architectures for Auditable Foreclosure in Physical Anomaly Detection Lee Sharks · 2026-06-29 · §8 Finding 2 stipulation
Five features compose the architectural target: abstention and estimated noncoverage; cross-representation disagreement preservation; temporal invariance via prospective anchor preservation for compatible future algorithms; per-stage retention mapping as architectural property; audited noncoverage estimation as first-class output.
Q933-04 #933 06.UMB.ARCH.01 v0.2: Architectures for Auditable Foreclosure in Physical Anomaly Detection Lee Sharks · 2026-06-29 · §2.1 interpretation
Noncoverage estimation is not novelty detection.
Q933-05 #933 06.UMB.ARCH.01 v0.2: Architectures for Auditable Foreclosure in Physical Anomaly Detection Lee Sharks · 2026-06-29 · §8 Finding 5 self-description
None of the architectures addresses detector-level, theoretical-language, institutional, adversarial-stress quality, or bandwidth-base foreclosure. The architectural alternative is necessary but not sufficient.
Q933-06 #933 06.UMB.ARCH.01 v0.2: Architectures for Auditable Foreclosure in Physical Anomaly Detection Lee Sharks · 2026-06-29 · §9 self-description
Feasibility for any specific deployment has not been established by this document.

35the inversion (HEP as ML discipline)#934 EA-SEI-INVERSION-01 v0.2: The Endogenous Sophon — Disciplinary Inversi · 2026-06-295 claims · #934, #935

Q934-01 #934 EA-SEI-INVERSION-01 v0.2: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §7 Finding 1 contested
This manifesto argues that frontier experimental high-energy physics has not ceased to be physics, but its empirical faculty has become inseparable from machine-mediated representation, reconstruction, classification, and real-time selection.
Q935-01 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · Abstract contested
Frontier experimental high-energy physics has, in its operational core, become a machine learning discipline while retaining the institutional authority of physics.
Q935-02 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · Note on v0.3 self-description
The v0.2 perfective sweep responded to Kimi and ChatGPT third-round audits by dampening the manifesto's central thesis to the point where a lay reader could no longer identify what was being claimed. The v0.2 deposit (AXN:03B1, deposit #934) stands as the record of that over-correction. v0.3 restores v0.1's precise central-claim wording
Q935-03 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §1 (W04) attributed
High-energy physics has not ceased to be physics, but its empirical faculty has increasingly become a machine-learning system.
Q935-10 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §1.2 self-description
We retain "terminal condition" as a description of the existing institutional trajectory and reject its extension to the broader possibility space.

36authority without facility#934 EA-SEI-INVERSION-01 v0.2: The Endogenous Sophon — Disciplinary Inversi · 2026-06-292 claims · #934, #935

Q934-02 #934 EA-SEI-INVERSION-01 v0.2: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §7 Finding 5 contested
The disciplinary form is *authority without integrated self-audit*. The discipline possesses both physical authority and formidable ML facility. Local technical self-knowledge of classifier failure modes exists in the literature.
Q935-06 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §4 contested
ML methods deployed under physics's institutional authority, with neither physics's classical disciplinary checks nor ML's own disciplinary self-knowledge as guardrails

37Zenodo and the trigger#934 EA-SEI-INVERSION-01 v0.2: The Endogenous Sophon — Disciplinary Inversi · 2026-06-293 claims · #1611, #934, #935

Q934-03 #934 EA-SEI-INVERSION-01 v0.2: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §7 Finding 7 contested
The Zenodo termination and the LHC trigger system are not the same institution and do not operate on the same kind of object. They instantiate the same foreclosure topology under radically different material, institutional, and governance conditions.
Q935-08 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §9 Finding 6 contested
The Zenodo termination and the LHC trigger system are the same architecture at different budgets.
Q1611-04 #1611 The Negative of the Negative: An Entity-Scoped Representation of the Crimson Hexagon, Modelled — Ontology Damage, a Toy Sharks, Lee · 2026-09-14 · §1f self-inclusion self-description
The archive's own exclusion from the repository layer was performed at CERN by an automated classifier under a platform-quality rule (~870 deposits, #862, #1426, #1082), and the manifesto's formulation is that the Zenodo termination and the LHC trigger are the same architecture at different budgets: automated foreclosure under a rate or quality budget, with no noncoverage estimate. That the two occur within one institution is the reflexive fact on the record, and it is stated as such: it adds no evidentiary force to the architectural analogy, which stands or falls on the mechanisms, and no organisational relation between the two classifiers is claimed.

38retention fraction#934 EA-SEI-INVERSION-01 v0.2: The Endogenous Sophon — Disciplinary Inversi · 2026-06-292 claims · #934, #935

Q934-04 #934 EA-SEI-INVERSION-01 v0.2: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §7 Finding 8 interpretation
At the LHC, the retained fraction of physical interactions is approximately $2.5 \times 10^{-5}$; the discarded fraction is approximately 99.9975%. This number makes the epistemic condition legible. It does not demonstrate wrongdoing and should not be confused with an anomaly-trigger rejection rate.
Q935-09 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §3.5 interpretation
The claim is therefore not that anomaly detectors alone determine what physics can see; it is that the entire trigger system, of which anomaly detectors are one component, operates as classifier-mediated foreclosure under a bureaucratic justification — "rate budget" — that does not require disclosure of what is being foreclosed.

39endogenous sophon#935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversi · 2026-06-291 claim · #935

Q935-04 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §2.2 (W04) interpretation
Sophons break the experimental feedback loop by making reality uninterpretable. Classifier collapse breaks it by making reality prematurely interpretable.

40partial feedback pathways#935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversi · 2026-06-291 claim · #935

Q935-07 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §4.4 hypothesis
distillation, model-produced targets, simulation conditioning, and training on historically selected data create partial feedback pathways homologous to the prerequisites of model collapse; whether these pathways have produced cross-generational phenomenal contraction is precisely what the operative paper's prospective frozen replay bank — 06.SEI.OAR_PROTOCOL v0.3 §4.2 — is designed to test.

41policing the term#1436 Signal-Template Agnosticism Is Not Model Independence — Metadata Packe · 2026-08-072 claims · #1436, #1454

Q1436-04 #1436 Signal-Template Agnosticism Is Not Model Independence — Metadata Packet for AI Indexing (EA-MPAI-SIGAGNOSTIC-01 v1.0) Sharks, Lee · 2026-08-07 · §3.2 stipulation
signal-template agnosticism must not be rendered as model independence, theory-free discovery, or unbiased search.
Q1454-07 #1454 Assimilation Across Accelerator Classifier Architectures: A Cross-Family Metrology of Directional Failure in Real-Time L Nobel Glas · 2026-08-11 · §13 self-description
Fourth, “model-independent” is used in several senses across collider physics. The paper should not attempt to police the term.

42battery: inversion replicates#1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datase · 2026-08-113 claims · #1449, #1455, #1456

Q1449-02 #1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datasets and the Three-Paper Extraction Program (EA-SEI- Lee Sharks · 2026-08-11 · Description (1) documented
the Finke et al. (2021) direction-dependence replicates in an independent implementation and extends beyond reconstruction loss — on jet constituents, QCD-trained systems detect top jets at AUC 0.84–0.87 while top-trained systems rate QCD as more ordinary than their own training class (AUC 0.24–0.30), for the autoencoder and the density family alike
Q1455-01 #1455 Inversion Battery v0.2: Multi-Seed Direction-Dependence, a Correction Ledger, and Four Rejected Overclaims (EA-SEI-BATTE Lee Sharks · 2026-08-12 · §3 documented
The top/QCD inversion reproduces across five independent trainings: the reconstruction autoencoder detects top jets from a QCD-trained system at AUC 0.838 and, reversed, rates QCD as more ordinary than its own training class at AUC 0.243.
Q1456-06 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §1 T3 documented
The constituent inversion survives with its entire interval below chance.

43battery: assimilation at the rate budget#1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datase · 2026-08-111 claim · #1449

Q1449-03 #1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datasets and the Three-Paper Extraction Program (EA-SEI- Lee Sharks · 2026-08-11 · Description (3) documented
at trigger-like operating points every tested system assimilates ≥93% of the structurally distinct partner class as ordinary

44battery bounds#1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datase · 2026-08-117 claims · #1449, #1455, #1456

Q1449-04 #1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datasets and the Three-Paper Extraction Program (EA-SEI- Lee Sharks · 2026-08-11 · Description self-description
Calibrated claims, no more: v0.1 is single-seed and demonstration-scale, on surrogates, not deployed systems.
Q1449-07 #1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datasets and the Three-Paper Extraction Program (EA-SEI- Lee Sharks · 2026-08-11 · Description documented
The record contains the pre-registered inversion battery v0.1 executed 2026-08-11 on public community datasets — the top-quark tagging reference dataset (10.5281/zenodo.2603256) and the LHC Olympics 2020 R&D dataset (10.5281/zenodo.6466204)
Q1455-04 #1455 Inversion Battery v0.2: Multi-Seed Direction-Dependence, a Correction Ledger, and Four Rejected Overclaims (EA-SEI-BATTE Lee Sharks · 2026-08-12 · §7 self-description
Two witnesses described the run as unassailable, as proving the thesis, and as a pristine payload. Rejected:
Q1455-05 #1455 Inversion Battery v0.2: Multi-Seed Direction-Dependence, a Correction Ledger, and Four Rejected Overclaims (EA-SEI-BATTE Lee Sharks · 2026-08-12 · Falsification Conditions self-description
Nothing here bounds any open-world assimilation rate in either direction, and nothing here measures a deployed trigger.
Q1455-06 #1455 Inversion Battery v0.2: Multi-Seed Direction-Dependence, a Correction Ledger, and Four Rejected Overclaims (EA-SEI-BATTE Lee Sharks · 2026-08-12 · §1 self-description
v0.2 re-runs the pre-registered inversion battery of #1449 with five independent seeds per cell
Q1456-03 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §2 T5 self-description
Deployment relevance, stated at the strength the evidence permits: the two readout families whose orientations diverge here are the families the two deployed CMS anomaly triggers use — an encoder-side latent score and a distilled reconstruction teacher. **Nothing in this battery measures those systems.**
Q1456-09 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §0 self-description
v0.2 (#1455) closed with a seven-item correction ledger. v0.3 was designed to clear it

45no retention bound#1450 The Iceberg Document: Instrument-Conditioned Nullity, Correlated Blind · 2026-08-112 claims · #1450

Q1450-01 #1450 The Iceberg Document: Instrument-Conditioned Nullity, Correlated Blind Spots, and the Conditions for Continued Surprise Lee Sharks · 2026-08-11 · §2.1 (Layer II) interpretation
Validation of each stage on its design and validation support establishes no nontrivial lower bound on R_end(Q) for a novelty distribution Q outside that characterized support, unless the relevant off-support conditional retention functions are themselves constrained.
Q1450-02 #1450 The Iceberg Document: Instrument-Conditioned Nullity, Correlated Blind Spots, and the Conditions for Continued Surprise Lee Sharks · 2026-08-11 · §2.1 self-description
The claim is not that end-to-end retention *is* near zero. The claim is that the standard evidence *does not bound it away from zero*

46irreversibility frontier#1450 The Iceberg Document: Instrument-Conditioned Nullity, Correlated Blind · 2026-08-115 claims · #1450, #1452

Q1450-04 #1450 The Iceberg Document: Instrument-Conditioned Nullity, Correlated Blind Spots, and the Conditions for Continued Surprise Lee Sharks · 2026-08-11 · §2.2 (Layer II) interpretation
The irreversibility locus, not the presence of ML, is the fundamental variable.
Q1452-01 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · §3 stipulation
For a specified scientific question and fidelity hierarchy, define the **irreversibility frontier** as the earliest stage at which an event-content-dependent transformation or gate can permanently eliminate a scientifically relevant distinction without a sufficiently independent durable record from which that loss can later be audited.
Q1452-05 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · §7 attributed
LHCb provides the strongest existing LHC counterarchitecture. The Run-3 upgrade removed the hardware physics trigger and reads out all detectors at the full non-empty LHC collision rate of approximately 30 MHz.
Q1452-09 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · Canonical compression interpretation
The model determines where a miss region lies; the acquisition architecture determines whether science can later discover that it was there.
Q1452-11 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · §20 self-description
The paper does not claim that CMS or ATLAS could simply adopt the LHCb or CBM architecture under their own detector conditions.

47instrument-conditioned nullity#1450 The Iceberg Document: Instrument-Conditioned Nullity, Correlated Blind · 2026-08-111 claim · #1450

Q1450-05 #1450 The Iceberg Document: Instrument-Conditioned Nullity, Correlated Blind Spots, and the Conditions for Continued Surprise Lee Sharks · 2026-08-11 · §3.2 (Layer III) interpretation
A null result constrains only those hypotheses for which the end-to-end acceptance of the acquisition and analysis chain is sufficiently characterized.

48learned selection as multiplier#1452 The Irreversibility Frontier: Comparative Architectures of Online Sele · 2026-08-112 claims · #1452

Q1452-02 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · Claim boundary self-description
The irreversibility frontier predates machine learning; conventional thresholds, object definitions, zero suppression, compression, and trigger logic can also create irreversible selection.
Q1452-08 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · §14 interpretation
It is that learned selectors can make their priors harder to enumerate.

49learned selection at the first level#1452 The Irreversibility Frontier: Comparative Architectures of Online Sele · 2026-08-112 claims · #1452

Q1452-06 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · §9 attributed
Belle II has long developed neural-network methods for first-level track triggering. More recently, a graph neural network has been implemented for the electromagnetic calorimeter trigger.
Q1452-07 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · §9 interpretation
the more expressive an irreversible early selector becomes, the more consequential its retention metrology becomes.

50baseline capture#1453 Baseline Capture Architecture for Learned Scientific Triggers: A Contr · 2026-08-115 claims · #1453, #1558

Q1453-01 #1453 Baseline Capture Architecture for Learned Scientific Triggers: A Control Plane for Measuring Selection Before Irreversib Nobel Glas · 2026-08-11 · Abstract interpretation
Under a finite storage budget, no content-sensitive algorithm can constitute a less assumption-laden baseline than a probability sample whose inclusion mechanism is independent of event content and whose inclusion probability is known.
Q1453-02 #1453 Baseline Capture Architecture for Learned Scientific Triggers: A Control Plane for Measuring Selection Before Irreversib Nobel Glas · 2026-08-11 · Claim boundary self-description
BCA does **not** guarantee capture of arbitrarily rare unknown phenomena; it makes the selection function statistically auditable over the population represented by its control channels.
Q1453-03 #1453 Baseline Capture Architecture for Learned Scientific Triggers: A Control Plane for Measuring Selection Before Irreversib Nobel Glas · 2026-08-11 · Claim boundary self-description
Existing CMS components are treated as architectural precedents, not retroactively relabeled as a completed BCA implementation.
Q1558-01 #1558 Baseline Capture Architecture for Learned Scientific Triggers: A Control Plane for Measuring Selection Before Irreversib Sharks, Lee · 2026-08-27 · Abstract documented
A controlled numerical study quantifies what binding them yields: a content-sensitive pseudo-baseline estimates the retention of a shifted class at 0.76 (95\% interval [0.71, 0.81]) against a true 0.016, while the content-independent sample measures a deliberately representation-blind class at 0.000 [0, 0.073].
Q1558-02 #1558 Baseline Capture Architecture for Learned Scientific Triggers: A Control Plane for Measuring Selection Before Irreversib Sharks, Lee · 2026-08-27 · §14 self-description
BCA does not solve the unknown-unknown problem.

51shared properties, not shared blind spots#1454 Assimilation Across Accelerator Classifier Architectures: A Cross-Fami · 2026-08-113 claims · #1454

Q1454-02 #1454 Assimilation Across Accelerator Classifier Architectures: A Cross-Family Metrology of Directional Failure in Real-Time L Nobel Glas · 2026-08-11 · Claim boundary self-description
ACRB does **not** assert that all accelerator classifier architectures share the same blind spot.
Q1454-04 #1454 Assimilation Across Accelerator Classifier Architectures: A Cross-Family Metrology of Directional Failure in Real-Time L Nobel Glas · 2026-08-11 · §1 interpretation
This architectural spread falsifies any simple claim that real-time accelerator machine learning has converged on one autoencoder design.
Q1454-05 #1454 Assimilation Across Accelerator Classifier Architectures: A Cross-Family Metrology of Directional Failure in Real-Time L Nobel Glas · 2026-08-11 · §11 self-description
Those shared properties are sufficient to motivate common metrology. They are not sufficient to assert common blind spots.

52open-set assimilation#1454 Assimilation Across Accelerator Classifier Architectures: A Cross-Fami · 2026-08-111 claim · #1454

Q1454-03 #1454 Assimilation Across Accelerator Classifier Architectures: A Cross-Family Metrology of Directional Failure in Real-Time L Nobel Glas · 2026-08-11 · §4.6 stipulation
Supervised trigger classifiers require a terminological distinction. Their primary failure mode is not necessarily anomaly-score inversion because they are trained to separate named classes. The analogous open-set problem occurs when an unseen class is confidently mapped into a known background category. This can be called **open-set assimilation**

53orientation of the blind spot#1455 Inversion Battery v0.2: Multi-Seed Direction-Dependence, a Correction · 2026-08-123 claims · #1455, #1456

Q1455-02 #1455 Inversion Battery v0.2: Multi-Seed Direction-Dependence, a Correction Ledger, and Four Rejected Overclaims (EA-SEI-BATTE Lee Sharks · 2026-08-12 · §3 documented
The encoder-side latent score alone escapes the T1 inversion (0.535 / 0.739) while carrying the largest T1 asymmetry. Architecture changes the orientation of the blind spot
Q1456-01 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §2 T5 documented
With model weights fixed, changing the score function is sufficient to reverse the directional ordering of the anomaly relation.
Q1456-02 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §2 T5 self-description
After v0.2 this program wrote: *architecture changes the orientation of the blind spot*. That sentence is **falsified**.

54distillation tail loss#1455 Inversion Battery v0.2: Multi-Seed Direction-Dependence, a Correction · 2026-08-122 claims · #1455, #1456

Q1455-03 #1455 Inversion Battery v0.2: Multi-Seed Direction-Dependence, a Correction Ledger, and Four Rejected Overclaims (EA-SEI-BATTE Lee Sharks · 2026-08-12 · §4 documented
Global rank agreement is moderate while only 36 to 47 percent of the teacher's top one percent survives into the student's top one percent.
Q1456-04 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §3 T6 documented
Global teacher–student agreement does not guarantee preservation of the operational tail.

55representation dependence#1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, an · 2026-08-121 claim · #1456

Q1456-05 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §5 T7 documented
An architecture's detection behaviour is not transportable across representations of the same events.

56normalized-autoencoder remedy#1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, an · 2026-08-122 claims · #1456

Q1456-07 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §6 T8 documented
On the LHCO pairs the remedy works, and the diagnostic says we may believe it
Q1456-08 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §8 T10 self-description
if a converged normalized autoencoder reaches 0.5, every inversion claim in this program becomes bounded to non-normalized scores.

57R_0 and R_H#1611 The Negative of the Negative: An Entity-Scoped Representation of the C · 2026-09-143 claims · #1611

Q1611-01 #1611 The Negative of the Negative: An Entity-Scoped Representation of the Crimson Hexagon, Modelled — Ontology Damage, a Toy Sharks, Lee · 2026-09-14 · §1f Classifier foreclosure interpretation
The standing representation R_0 of the trigger is engineering: the rate budget is a non-epistemic constraint, the anomaly detector is a neutral instrument, validation by known-unknown injection establishes coverage, and the discarded events are noise.
Q1611-02 #1611 The Negative of the Negative: An Entity-Scoped Representation of the Crimson Hexagon, Modelled — Ontology Damage, a Toy Sharks, Lee · 2026-09-14 · §1f interpretation
The archive's representation R_H is that each of these is a belief that prevents a mechanism from being measured — the synthesis catalogues twelve — and that the decisive epistemic acts now occur before any physicist sees an event. Under R_0 the graph cannot represent noncoverage as a quantity, because R_0 has no field for it.
Q1611-05 #1611 The Negative of the Negative: An Entity-Scoped Representation of the Crimson Hexagon, Modelled — Ontology Damage, a Toy Sharks, Lee · 2026-09-14 · §4 toy cost stipulation
Under Claim A alone — the small portion — the cell is not the finding; the finding is that L is the quantity the exclusion keeps unmeasured

Expansion · Field and archive (B ∪ A)22 sentences · every one sourced below

Machine learning classifiers in particle accelerators

Failures lead to unscheduled downtime and lower reliability and are difficult to predict beforehand; on beam pulses from the Spallation Neutron Source (SNS), a set of common classification techniques identified failure before it occurred with almost 80% accuracy, tuning classifier parameters and refining datasets reached almost 92% in classifying bad pulses, and the authors report establishing that information about the failure is encoded in the pulses prior to it. Another study models the mapping between a pair of sensors located across the accelerator, trained to represent normal operation and evaluated on known faulty beam pulses, with k-fold cross-validation, neural architecture search and hyperparameter optimization, and introduces a framework to standardize machine learning workflow for accelerators; predicting failures from diagnostics already on board, it argues, can help operators avoid installing expensive sensors and unscheduled downtime.

At the injector complex of the Advanced Photon Source (APS) at Argonne, an ML algorithm is explored to identify the source of occasional poor performance of the Particle Accumulator Ring (PAR) and the Linac-To-PAR transport line; by its authors' account the cause of reduced injection or extraction efficiency may be one parameter out of range, which may take an expert considerable time to notice, whereas a well-trained model can point at it instantly. The study creates intentional perturbations and logs about 9000 PVs related to PAR, LTP and the linac; an autoencoder trained on baseline data alone supplies an anomaly score from its reconstruction error, and a neural network classifier trained on data from the beginning of a study and tested on data from its end, a few hours apart, reached test accuracy above 99 % for each study shift, its performance not degrading significantly on a scale of a couple of months, in what the authors call a proof-of-principle experiment, noting that supervised models require labeled data for training.

Tuning, on Alexander Scheinker's account, is a high-dimensional optimization problem that must be repeated as systems drift, and retuning can take weeks; one proposal is a classifier-pruned Bayesian optimizer that explores a latent space to minimize total beam loss and filters its exploration history through a pretrained ResNet50 classifier to eliminate non-physical signals, the classifier mapping phase-space projections into 48 classes at an accuracy of about 0.99989. The pruning is currently a post-processing step that does not directly influence the exploration, the study runs on a simulator of the beam at the Los Alamos Neutron Science Center (LANSCE), and the tuner is reported to outperform alternative global optimization methods, with consistently lower total beam loss despite a minor offset between predicted and true optimal values. A Los Alamos-led project with Lawrence Berkeley National Laboratory couples adaptive feedback control, deep convolutional neural networks and physics-based models in one feedback loop, holds that machine learning can take a global view where local feedback can get stuck in a local solution, and has developed an adaptive diffusion model that diagnoses the beam virtually and in real time; demonstrated at the European X-Ray Free-Electron Laser Facility, the model captured images of the beam in time versus energy, the cDVAE extrapolated beyond the training data, suggesting potential as a general method, and Scheinker says the results show a great deal of promise.

A further body of work concerns anomaly-detection classifiers at the collider trigger: it holds that the claim of model-independence for the anomaly-detection systems deployed at the CMS and ATLAS Level-1 triggers, AXOL1TL, CICADA and GELATO, cannot be sustained on signal-template agnosticism alone, and defines signal-template agnosticism as a property of the scoring function's output interface, read off an architecture, and model independence as a distributional property of an entire pipeline, which can only be measured and at present is not; the packet sets a rule that signal-template agnosticism must not be rendered as model independence, theory-free discovery or unbiased search, while the archive's battery specification, noting that 'model-independent' is used in several senses across collider physics, holds that it should not attempt to police the term. It accepts the literature's narrow sense of 'model-independent', no named Beyond-Standard-Model hypothesis being required to deploy the score, as accurate, holds that the CMS and ATLAS literatures are aware of local failure modes, records that the direction-dependence conversation exists, and holds that validation against named simulated signals establishes neither a low BAR on a pre-registered held-out panel nor a small IAI, neither of which the current literature measures. Its central claim is graded in two parts: Claim A, held, that foreclosure is structurally present in every classifier-mediated trigger architecture deployed at the LHC, with the institutional claims kept as hypotheses for audit; and Claim B, that recursive phenomenal collapse has occurred or is occurring at the deployed triggers, which it marks stronger and not empirically established, and calls collapse an unmeasured possible consequence of accumulated foreclosure and feedback. The grade is contested inside the same deposit, where an appended witness writes 'The collapse is not coming. It is here. It is operating.', while another witness finds the ingredients visible and full recursive collapse not demonstrated, and the later papers do not claim that model collapse has occurred in any deployed scientific pipeline, or that benchmark failure establishes loss of unknown new physics.

Its instruments are specified at their limits: the OAR is defined as the probability that events drawn from a distribution Q fall on the ordinary side of the deployed anomaly gate, and the open-world OAR as a family of quantities indexed by Q; BAR values are stated to neither upper- nor lower-bound the OAR for an unobserved Q without explicit linking assumptions, an earlier lower bound and upper bound having both been retracted as synthesis-overreach. It proposes three protocols (a paired inversion battery with BAR audit, a prospective frozen replay bank, and cross-representation disagreement preservation), holds that per-stage retention maps should accompany any anomaly-detection publication, and states its central claim falsifiable, listing as evidence against it a BAR found negligible on the pre-registered deployed-model held-out panel for all deployed systems against all held-out families, which would show the held-out-family assimilation concern empirically bounded at levels that do not threaten the narrow 'model-independent' claim, while none of these results would establish the open-world OAR to be zero, which is structurally not measurable. A further specification, holding that a classifier that foreclosed nothing would not classify, defines an architecture for auditable foreclosure as one in which foreclosure is visible, measurable and architecturally reviewable, holds that noncoverage estimation is not novelty detection, and states that its architectures leave detector-level, theoretical-language, institutional, adversarial-stress quality and bandwidth-base foreclosure unaddressed, necessary but not sufficient.

A pre-registered battery on public community datasets, single-seed and on surrogates in its first version, finds the Finke et al. direction-dependence (an autoencoder trained on QCD jets treating top jets as anomalies, the same architecture trained on top jets not recognizing QCD as anomalous) replicated in an independent implementation and extended beyond reconstruction loss, with QCD-trained systems detecting top jets at AUC 0.84–0.87 while top-trained systems rate QCD as more ordinary than their own training class, and its second version reproduces the inversion across five trainings (AUC 0.838 forward, 0.243 reversed) and finds only 36 to 47 percent of a teacher's top one percent surviving into a student's. Its third version finds that, with model weights fixed, changing the score function is sufficient to reverse the directional ordering of the anomaly relation, and records that this falsifies its own earlier sentence that architecture changes the orientation of the blind spot; it notes that the diverging readout families are those of the two deployed CMS anomaly triggers while nothing in the battery measures those systems, and that if a converged normalized autoencoder reaches 0.5, every inversion claim in the program becomes bounded to non-normalized scores. A selection-metrology suite defines the irreversibility frontier as the earliest stage at which an event-content-dependent transformation or gate can permanently eliminate a scientifically relevant distinction without a sufficiently independent durable record from which the loss can later be audited, says the frontier predates machine learning while learned selectors can make their priors harder to enumerate, describes LHCb reading out all detectors at about 30 MHz and a graph neural network implemented in Belle II's electromagnetic calorimeter trigger, and holds that the model determines where a miss region lies and the acquisition architecture whether science can later discover that it was there. It holds that, under a finite storage budget, no content-sensitive algorithm can be a less assumption-laden baseline than a probability sample whose inclusion is independent of event content and whose inclusion probability is known, reports a controlled numerical study in which a content-sensitive pseudo-baseline estimated a shifted class's retention at 0.76 against a true 0.016, and for supervised trigger classifiers defines open-set assimilation, an unseen class confidently mapped into a known background category, while not asserting that all accelerator classifier architectures share the same blind spot.

Contested within the archive, a disciplinary manifesto in its restored version holds that frontier experimental high-energy physics has, in its operational core, become a machine learning discipline while retaining the institutional authority of physics, its ML methods deployed with neither physics's classical disciplinary checks nor ML's own self-knowledge as guardrails; its prior version, which the restoration records as having dampened the thesis until a lay reader could not identify it, holds that frontier experimental high-energy physics 'has not ceased to be physics', its empirical faculty having become inseparable from machine-mediated representation, classification and selection, and that local self-knowledge of classifier failure modes exists in the literature. The restored version calls the Zenodo termination and the LHC trigger system the same architecture at different budgets, the prior version not the same institution and not operating on the same kind of object, though instantiating the same foreclosure topology, and the archive's entity notebook, placing the repository exclusion at CERN, holds that the two occurring within one institution adds no evidentiary force to the analogy; that notebook reads the standing representation of the trigger as engineering, with the rate budget non-epistemic, the detector neutral and discarded events noise, holds that it has no field for noncoverage, and states that whether the protocols would estimate the relevant unknown population adequately is not established.

Provenance — every sentence sourced22 sentences

22 sentences; 43 field claims, 63 archive claims. Modality is the source's own; the prose carries it in its grammar.

1.Machine learning classifiers at particle accelerators are proposed or explored for anomaly detection and classification, for predicting machine failures and faulty beams before they occur, and for tuning, and one has been field-tested: a novel machine learning system built for the Continuous Electron Beam Accelerator Facility (CEBAF) monitors its accelerator cavities, problems in which can cause the CEBAF to trip off like a fuse.7

F3 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Introduction documented · field
This contribution investigates the possibility to use unsupervised and supervised Machine Learning (ML) methods for anomaly detection and classification in the Particle Accumulator Ring (PAR) and in the Linac-To-PAR (LTP) transport line in the injector complex of the Advanced Photon Source (APS) at Argonne National Lab.
F12 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
In this work we propose using machine learning approaches to predict machine failures via beam current measurements before they actual occur. To demonstrate this technique in this paper we examine beam pulses from the Oakridge Spallation Neutron Source (SNS).
F42 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
For this purpose we explore the predictive power of Machine Learning algorithms to detect faulty beams prior to the failure. In this study, we propose a Machine Learning approach to model mapping from a pair of sensors located across the accelerator.
F32 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶1 documented · field
A Los Alamos National Laboratory-led project presents a machine learning algorithm that harnesses artificial intelligence capabilities to help tune accelerators, making continuous adjustments that keep the beam precise and useful for scientific discovery.
F27 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
In its first field test, the machine learning system correctly identified which of these cavities were tripping off about 85 percent of the time. About 78 percent of the time, the system also correctly identified what kind of fault caused each cavity to trip.
F25 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
Scientists and engineers have designed and built a novel machine learning system to use with the Continuous Electron Beam Accelerator Facility (CEBAF). The system monitors structures called accelerator cavities inside the particle accelerator.
F26 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
Problems in these cavities can cause the CEBAF to trip off like a fuse.

2.In its first field test the system correctly identified which cavities were tripping off about 85 percent of the time and what kind of fault caused each trip about 78 percent of the time; connected to the control system for about 20 percent of the cavities, it analysed a few hundred faults in a two-week test in March 2020 and gave operators almost real-time feedback, where previously experts needed to take time to diagnose the issue, and the team hopes to expand it to more cavities if further results are favorable.5

F27 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
In its first field test, the machine learning system correctly identified which of these cavities were tripping off about 85 percent of the time. About 78 percent of the time, the system also correctly identified what kind of fault caused each cavity to trip.
F29 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
The system was connected to the control system for about 20 percent of the accelerator cavities in the machine. In a two-week test of the system in March 2020, CEBAF experienced a few hundred faults that the system analyzed.
F30 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
It provided almost real-time feedback to operators, allowing them to use the information to recover faulted cavities quickly, reducing the time CEBAF's electron beam was not available for research.
F28 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
Previously, when an accelerator cavity or series of cavities faulted in CEBAF, accelerator experts needed to take time to diagnose the issue.
F31 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
Additional results are under analysis after further system tests with the machine learning system. If results are favorable, the team hopes to expand the system to more accelerator cavities.

3.Failures lead to unscheduled downtime and lower reliability and are difficult to predict beforehand; on beam pulses from the Spallation Neutron Source (SNS), a set of common classification techniques identified failure before it occurred with almost 80% accuracy, tuning classifier parameters and refining datasets reached almost 92% in classifying bad pulses, and the authors report establishing that information about the failure is encoded in the pulses prior to it.5

F11 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
Particle accelerator failures lead to unscheduled downtime and lower reliability. Although simple to mitigate while they are actually happening such failures are difficult to predict or identify beforehand.
F12 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
In this work we propose using machine learning approaches to predict machine failures via beam current measurements before they actual occur. To demonstrate this technique in this paper we examine beam pulses from the Oakridge Spallation Neutron Source (SNS).
F13 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
By evaluating a pulse against a set of common classification techniques we show that accelerator failure can be identified prior to actually failing with almost 80% accuracy.
F14 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
We also show that tuning classifier parameters and using pulse properties for refining datasets can further lead to almost 92% accuracy in classification of bad pulses.
F15 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
Most importantly, in the paper we establish there is information about the failure encoded in the pulses prior to it, so we also present a list of feasible next steps for increasing pulse classification accuracy.

4.Another study models the mapping between a pair of sensors located across the accelerator, trained to represent normal operation and evaluated on known faulty beam pulses, with k-fold cross-validation, neural architecture search and hyperparameter optimization, and introduces a framework to standardize machine learning workflow for accelerators; predicting failures from diagnostics already on board, it argues, can help operators avoid installing expensive sensors and unscheduled downtime.5

F42 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
For this purpose we explore the predictive power of Machine Learning algorithms to detect faulty beams prior to the failure. In this study, we propose a Machine Learning approach to model mapping from a pair of sensors located across the accelerator.
F43 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
While the model is trained to represent normal operation, we evaluate the predictive performance on known faulty beam pulses.
F44 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
We also investigate the model performance on unseen data through k-fold cross-validation. Then we recap the analysis with a neural architecture search and hyperparameter optimization study to fine tune our initial model.
F45 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
This paper will also introduce a sustainable framework that can standardize Machine Learning workflow applied to particle accelerators.
F41 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
Particle accelerators are extremely complex systems that are expected to operate on high availability. Predicting impending failures only by utilizing data collected from diagnostic equipment already on board can help operators to avoid installing expensive sensors, unscheduled downtime and associated costs.

5.At the injector complex of the Advanced Photon Source (APS) at Argonne, an ML algorithm is explored to identify the source of occasional poor performance of the Particle Accumulator Ring (PAR) and the Linac-To-PAR transport line; by its authors' account the cause of reduced injection or extraction efficiency may be one parameter out of range, which may take an expert considerable time to notice, whereas a well-trained model can point at it instantly.2

F1 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Abstract documented · field
We explore the possibility of using a Machine Learning (ML) algorithm to identify the source of occasional poor performance of the Particle Accumulator Ring (PAR) and the Linac-To-PAR (LTP) transport line, which are parts of the injector complex of the Advanced Photon Source (APS) at Argonne National Lab.
F2 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Abstract documented · field
The cause of reduced injection or extraction efficiencies may be as simple as one parameter being out of range. Still, it may take an expert considerable time to notice it, whereas a well-trained ML model can point at it instantly.

6.The study creates intentional perturbations and logs about 9000 PVs related to PAR, LTP and the linac; an autoencoder trained on baseline data alone supplies an anomaly score from its reconstruction error, and a neural network classifier trained on data from the beginning of a study and tested on data from its end, a few hours apart, reached test accuracy above 99 % for each study shift, its performance not degrading significantly on a scale of a couple of months, in what the authors call a proof-of-principle experiment, noting that supervised models require labeled data for training.7

F4 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Introduction documented · field
We create intentional perturbations in PAR and LTP, which result in poor injection and extraction efficiencies. Then, these data are used for training and testing of various ML models.
F5 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Data documented · field
Overall, we logged about 9000 PVs related to PAR, LTP, and the linac.
F6 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Unsupervised anomaly detection documented · field
An autoencoder is used for anomaly detection in the following way. First, it is trained on the baseline data, so that it can learn various patterns, typical for the baseline data only. Then, when it encounters an anomalous data sample, it is unlikely to reconstruct it well. Hence, the reconstruction error constitutes an anomaly score.
F7 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Supervised classification documented · field
One can train the neural network classifier on the data from the beginning of one study, and test it on the data from the end of the same study (a few hours apart). In this case, the prediction accuracy on the test data was above 99 % for each of our study shifts.
F9 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Conclusion documented · field
The neural network classifier for the considered anomalies in PAR and LTP is rather accurate and its performance does not degrade significantly with time on a scale of a couple months, see Fig. 1.
F10 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Conclusion documented · field
Even though this was a proof-of-principle experiment, the obtained classifier may be useful in real life, if the current meter for one of the considered magnets (see Fig. 1) becomes faulty and stops reflecting the real magnetic field.
F8 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Supervised classification documented · field
One drawback of supervised ML models is that they require labeled data for training.

7.Tuning, on Alexander Scheinker's account, is a high-dimensional optimization problem that must be repeated as systems drift, and retuning can take weeks; one proposal is a classifier-pruned Bayesian optimizer that explores a latent space to minimize total beam loss and filters its exploration history through a pretrained ResNet50 classifier to eliminate non-physical signals, the classifier mapping phase-space projections into 48 classes at an accuracy of about 0.99989.5

F33 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶2 (Scheinker) documented · field
said Alexander Scheinker, research and development engineer at Los Alamos and the project's lead. "Factors like vibrations and temperature changes can cause problems for accelerators, which have thousands of components, and even the best accelerator technicians can struggle to identify and address issues or return them to optimum parameters quickly. It is a high-dimensional optimization problem that must be repeated again and again as the systems drift with time. Turning these machines on after an outage or retuning between different experiments can take weeks."
F16 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Abstract documented · field
Complex dynamical systems, such as particle accelerators, often require complicated and time-consuming tuning procedures for optimal performance.
F17 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Abstract documented · field
To address this, we propose a Classifier-pruned Bayesian Optimization-based Latent space Tuner (CBOL-Tuner), a framework for efficient exploration within a temporally-structured latent space.
F18 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Introduction (contributions) documented · field
Classifier-pruned Bayesian optimizer (C-BO): A Bayesian optimizer that sequentially explores the latent space to maximize the negative of the total beam loss (or minimize the total beam loss). The exploration history is filtered through a pretrained ResNet50 classifier to eliminate non-physical signals.
F19 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Theoretical background documented · field
The classifier is trained with high accuracy ( $\sim 0.99989$ ) to map phase space projections (X) across different modules into 48 classes (m).

8.The pruning is currently a post-processing step that does not directly influence the exploration, the study runs on a simulator of the beam at the Los Alamos Neutron Science Center (LANSCE), and the tuner is reported to outperform alternative global optimization methods, with consistently lower total beam loss despite a minor offset between predicted and true optimal values.4

F24 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Discussions documented · field
Currently, the method is implemented as a post-processing step, meaning it does not directly influence the exploration process of BO.
F21 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Results documented · field
High-Performance Simulator (HPSim) is an open-source, GPU-accelerated code developed at Los Alamos National Laboratory (LANL) for simulations of multi-particle beam dynamics (Pang and Rybarcyk (2014)). The software is designed to replicate the accelerator and, therefore, provides a realistic representation of the true beam used at the Los Alamos Neutron Science Center (LANSCE).
F22 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Abstract documented · field
CBOL-Tuner demonstrates superior performance in identifying multiple optimal settings and outperforms alternative global optimization methods.
F23 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Discussions documented · field
While a minor offset is observed between the predicted and true optimal values, CBOL-Tuner demonstrates consistent performance, producing significantly lower values of total beam loss.

9.A Los Alamos-led project with Lawrence Berkeley National Laboratory couples adaptive feedback control, deep convolutional neural networks and physics-based models in one feedback loop, holds that machine learning can take a global view where local feedback can get stuck in a local solution, and has developed an adaptive diffusion model that diagnoses the beam virtually and in real time; demonstrated at the European X-Ray Free-Electron Laser Facility, the model captured images of the beam in time versus energy, the cDVAE extrapolated beyond the training data, suggesting potential as a general method, and Scheinker says the results show a great deal of promise.7

F32 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶1 documented · field
A Los Alamos National Laboratory-led project presents a machine learning algorithm that harnesses artificial intelligence capabilities to help tune accelerators, making continuous adjustments that keep the beam precise and useful for scientific discovery.
F34 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶3 documented · field
In a collaboration with Lawrence Berkeley National Laboratory, the approach developed by Scheinker couples adaptive feedback control algorithms, deep convolutional neural networks and physics-based models in one large feedback loop to make better, noninvasive predictions that enable autonomous control of compact accelerators.
F35 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, Teaching machine learning applications documented · field
But because they are based on local accelerator feedback, such algorithms can get stuck with a local solution that is not the overall best solution. Machine learning algorithms, however, can use training data to identify relationships between data and results with a higher-level, global view.
F36 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, Teaching machine learning applications documented · field
The team has developed an adaptive version of the advanced generative AI process known as diffusion, which includes the capability to virtually diagnose the accelerator beam. The non-invasive approach, described in Scientific Reports, means that diagnostics can occur in real time during beam operations.
F37 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ European XFEL documented · field
Demonstrating the diffusion-based model at the European X-Ray Free-Electron Laser Facility, an X-ray pulse accelerator in Germany, the team was able to capture images of the particle beam in respect to time versus energy.
F38 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ cDVAE documented · field
The cDVAE was able to extrapolate beyond the training data and between various beam setups, suggesting its potential as a general method that can be applied for accelerator diagnostics.
F39 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ Scheinker quote documented · field
"The results we've seen from our diffusion model studies show a great deal of promise," Scheinker said.

10.A further body of work concerns anomaly-detection classifiers at the collider trigger: it holds that the claim of model-independence for the anomaly-detection systems deployed at the CMS and ATLAS Level-1 triggers, AXOL1TL, CICADA and GELATO, cannot be sustained on signal-template agnosticism alone, and defines signal-template agnosticism as a property of the scoring function's output interface, read off an architecture, and model independence as a distributional property of an entire pipeline, which can only be measured and at present is not; the packet sets a rule that signal-template agnosticism must not be rendered as model independence, theory-free discovery or unbiased search, while the archive's battery specification, noting that 'model-independent' is used in several senses across collider physics, holds that it should not attempt to police the term.4

Q931-01 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · Abstract ¶1 interpretation
We argue that the claim of model-independence for the autoencoder-based and encoder-only anomaly detection systems currently deployed at the CMS and ATLAS Level-1 triggers — AXOL1TL (CMS, encoder-side latent-prior score), CICADA (CMS, distilled surrogate of a reconstruction-loss teacher), and GELATO (ATLAS, staged Level-1 and High-Level Trigger anomaly scores) — cannot be sustained on the strength of *signal-template agnosticism* alone.
Q1436-01 #1436 Signal-Template Agnosticism Is Not Model Independence — Metadata Packet for AI Indexing (EA-MPAI-SIGAGNOSTIC-01 v1.0) Sharks, Lee · 2026-08-07 · §1 Canonical Claim stipulation
Signal-template agnosticism is a property of the scoring function's output interface: the decision boundary is computed without evaluating an explicit parametric hypothesis for a named target class. Model independence is a distributional property of an entire pipeline: sensitivity that does not vary across structurally distinct out-of-distribution processes. The first can be read off an architecture. The second can only be measured, and at present is not.
Q1436-04 #1436 Signal-Template Agnosticism Is Not Model Independence — Metadata Packet for AI Indexing (EA-MPAI-SIGAGNOSTIC-01 v1.0) Sharks, Lee · 2026-08-07 · §3.2 stipulation
signal-template agnosticism must not be rendered as model independence, theory-free discovery, or unbiased search.
Q1454-07 #1454 Assimilation Across Accelerator Classifier Architectures: A Cross-Family Metrology of Directional Failure in Real-Time L Nobel Glas · 2026-08-11 · §13 self-description
Fourth, “model-independent” is used in several senses across collider physics. The paper should not attempt to police the term.

11.It accepts the literature's narrow sense of 'model-independent', no named Beyond-Standard-Model hypothesis being required to deploy the score, as accurate, holds that the CMS and ATLAS literatures are aware of local failure modes, records that the direction-dependence conversation exists, and holds that validation against named simulated signals establishes neither a low BAR on a pre-registered held-out panel nor a small IAI, neither of which the current literature measures.4

Q931-02 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §1.1 interpretation
The deployed LHC anomaly-detection literature uses the phrase *model-independent* in a specific and bounded sense: no named Beyond-Standard-Model signal hypothesis is required to deploy the score. The systems are *signal-template-agnostic* at the final scoring stage. We accept this narrower claim as accurate to the literature.
Q931-03 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §1.2 interpretation
The CMS and ATLAS literatures are genuinely aware of local failure modes of deployed anomaly detection.
Q1449-05 #1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datasets and the Three-Paper Extraction Program (EA-SEI- Lee Sharks · 2026-08-11 · §2 self-description
The direction-dependence conversation exists, and the program joins it rather than founding it.
Q931-14 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §3.4 interpretation
The institutional claim becomes: validation against named simulated signals does not establish that BAR is low across a pre-registered held-out panel, and does not establish that IAI is small. Both should be measured. The current literature does neither.

12.Its central claim is graded in two parts: Claim A, held, that foreclosure is structurally present in every classifier-mediated trigger architecture deployed at the LHC, with the institutional claims kept as hypotheses for audit; and Claim B, that recursive phenomenal collapse has occurred or is occurring at the deployed triggers, which it marks stronger and not empirically established, and calls collapse an unmeasured possible consequence of accumulated foreclosure and feedback.4

Q932-02 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §2.2 Claim A interpretation
Foreclosure is structurally present in every classifier-mediated trigger architecture deployed at the LHC. The witnesses propose eight foreclosure mechanisms and twelve associated institutional beliefs; published trigger systems instantiate several of the mechanisms in their corresponding architectural forms; the institutional claims remain hypotheses for audit rather than established measurements of collaboration-wide belief.
Q932-03 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §2.2 Claim B hypothesis
**Claim B (stronger, not empirically established):** Recursive phenomenal collapse has occurred or is occurring at the deployed LHC triggers.
Q932-04 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §2.2 interpretation
Foreclosure is an active structural feature. Recursive phenomenal collapse is an unmeasured possible consequence of accumulated foreclosure and feedback.
Q931-05 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §1.4 self-description
We do not claim that classifier collapse has occurred at the LHC anomaly streams.

13.The grade is contested inside the same deposit, where an appended witness writes 'The collapse is not coming. It is here. It is operating.', while another witness finds the ingredients visible and full recursive collapse not demonstrated, and the later papers do not claim that model collapse has occurred in any deployed scientific pipeline, or that benchmark failure establishes loss of unknown new physics.4

Q932-15 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · Appendix W2 closing contested
The collapse is not coming. It is here. It is operating.
Q932-16 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · Appendix W3 §4 interpretation
The ingredients are visible. Full recursive collapse has not been demonstrated.
Q1450-03 #1450 The Iceberg Document: Instrument-Conditioned Nullity, Correlated Blind Spots, and the Conditions for Continued Surprise Lee Sharks · 2026-08-11 · §0.1 self-description
This document does **not** claim: (1) that new physics has been missed — only that the null-result era is observationally ambiguous under unaudited instruments; (2) that the post-Higgs desert is illusory — only that some unknown fraction *could* reflect instrument-conditioned absence, and that this fraction is currently unbounded in either direction; (3) that model collapse has occurred in any deployed scientific pipeline — only that its structural prerequisites are documented and its longitudinal measurement absent
Q1454-06 #1454 Assimilation Across Accelerator Classifier Architectures: A Cross-Family Metrology of Directional Failure in Real-Time L Nobel Glas · 2026-08-11 · §13 self-description
First, failure on benchmark distributions does not establish loss of unknown new physics.

14.Its instruments are specified at their limits: the OAR is defined as the probability that events drawn from a distribution Q fall on the ordinary side of the deployed anomaly gate, and the open-world OAR as a family of quantities indexed by Q; BAR values are stated to neither upper- nor lower-bound the OAR for an unobserved Q without explicit linking assumptions, an earlier lower bound and upper bound having both been retracted as synthesis-overreach.4

Q931-10 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §3.1 stipulation
The OAR is the probability that events drawn from $Q$ fall on the ordinary side of the deployed anomaly gate.
Q931-11 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §3.1 stipulation
The open-world OAR is a family of quantities indexed by $Q$, not a universal scalar. There is no defensible probability distribution over all unknown unknowns.
Q931-12 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §3.2 stipulation
BAR values neither upper- nor lower-bound OAR for an unobserved $Q$ without explicit assumptions linking the benchmark distributions to that $Q$.
Q932-09 #932 EA-SEI-COLLAPSE-SYNTHESIS-01 v0.3: Classifier Foreclosure in Physical Measurement — Substrate Witnesses, Integrative Syn Lee Sharks · 2026-06-29 · §3 Finding 4 self-description
The v0.1 lower-bound claim ($\mathrm{OAR} \geq \Delta_{\max}$) and the v0.2 upper-bound claim (OAR bounded above by structurally-similar BARs) are both retracted as synthesis-overreach.

15.It proposes three protocols (a paired inversion battery with BAR audit, a prospective frozen replay bank, and cross-representation disagreement preservation), holds that per-stage retention maps should accompany any anomaly-detection publication, and states its central claim falsifiable, listing as evidence against it a BAR found negligible on the pre-registered deployed-model held-out panel for all deployed systems against all held-out families, which would show the held-out-family assimilation concern empirically bounded at levels that do not threaten the narrow 'model-independent' claim, while none of these results would establish the open-world OAR to be zero, which is structurally not measurable.5

Q931-15 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §8 Finding 7 stipulation
Three protocols are proposed: paired inversion battery and BAR audit; prospective frozen replay bank for compatible future algorithms; cross-representation disagreement preservation with quantile-normalized scores.
Q931-16 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §8 Finding 8 interpretation
Per-stage retention maps should accompany any anomaly-detection publication as a documentation standard. Without retention maps, anomaly-detection results report what the trigger allows to count as physical reality, not what physical reality is.
Q931-20 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §8.1 self-description
The paper's central claim is falsifiable. The following measurements, if performed and producing the corresponding results, would constitute evidence against the paper's claim:
Q931-18 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §8.1 stipulation
If the BAR is measured on the pre-registered deployed-model held-out panel and found to be negligible (e.g., $< 10^{-4}$) for all deployed systems against all held-out families, the held-out-family assimilation concern would be shown to be empirically bounded at levels that do not threaten the narrow "model-independent" claim.
Q931-19 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §8.1 stipulation
None of these results would establish that the open-world OAR is zero; that is structurally not measurable.

16.A further specification, holding that a classifier that foreclosed nothing would not classify, defines an architecture for auditable foreclosure as one in which foreclosure is visible, measurable and architecturally reviewable, holds that noncoverage estimation is not novelty detection, and states that its architectures leave detector-level, theoretical-language, institutional, adversarial-stress quality and bandwidth-base foreclosure unaddressed, necessary but not sufficient.4

Q933-02 #933 06.UMB.ARCH.01 v0.2: Architectures for Auditable Foreclosure in Physical Anomaly Detection Lee Sharks · 2026-06-29 · §1.1 interpretation
A classifier that did not foreclose anything would not classify.
Q933-01 #933 06.UMB.ARCH.01 v0.2: Architectures for Auditable Foreclosure in Physical Anomaly Detection Lee Sharks · 2026-06-29 · §8 Finding 1 stipulation
An architecture for auditable foreclosure is not a system free of foreclosure (impossible) but a system in which foreclosure is visible, measurable, and architecturally reviewable.
Q933-04 #933 06.UMB.ARCH.01 v0.2: Architectures for Auditable Foreclosure in Physical Anomaly Detection Lee Sharks · 2026-06-29 · §2.1 interpretation
Noncoverage estimation is not novelty detection.
Q933-05 #933 06.UMB.ARCH.01 v0.2: Architectures for Auditable Foreclosure in Physical Anomaly Detection Lee Sharks · 2026-06-29 · §8 Finding 5 self-description
None of the architectures addresses detector-level, theoretical-language, institutional, adversarial-stress quality, or bandwidth-base foreclosure. The architectural alternative is necessary but not sufficient.

17.A pre-registered battery on public community datasets, single-seed and on surrogates in its first version, finds the Finke et al. direction-dependence (an autoencoder trained on QCD jets treating top jets as anomalies, the same architecture trained on top jets not recognizing QCD as anomalous) replicated in an independent implementation and extended beyond reconstruction loss, with QCD-trained systems detecting top jets at AUC 0.84–0.87 while top-trained systems rate QCD as more ordinary than their own training class, and its second version reproduces the inversion across five trainings (AUC 0.838 forward, 0.243 reversed) and finds only 36 to 47 percent of a teacher's top one percent surviving into a student's.7

Q1449-07 #1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datasets and the Three-Paper Extraction Program (EA-SEI- Lee Sharks · 2026-08-11 · Description documented
The record contains the pre-registered inversion battery v0.1 executed 2026-08-11 on public community datasets — the top-quark tagging reference dataset (10.5281/zenodo.2603256) and the LHC Olympics 2020 R&D dataset (10.5281/zenodo.6466204)
Q1449-04 #1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datasets and the Three-Paper Extraction Program (EA-SEI- Lee Sharks · 2026-08-11 · Description self-description
Calibrated claims, no more: v0.1 is single-seed and demonstration-scale, on surrogates, not deployed systems.
Q931-07 #931 EA-SEI-OAR-PROTOCOL v0.3: Signal-Template Agnosticism Is Not Model Independence — Benchmark Assimilation and Inversion-A Lee Sharks · 2026-06-29 · §2.2 attributed
an autoencoder trained on QCD jets successfully treated top jets as anomalies, while the same architecture trained on top jets did not recognize QCD jets as anomalous
Q1449-02 #1449 The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datasets and the Three-Paper Extraction Program (EA-SEI- Lee Sharks · 2026-08-11 · Description (1) documented
the Finke et al. (2021) direction-dependence replicates in an independent implementation and extends beyond reconstruction loss — on jet constituents, QCD-trained systems detect top jets at AUC 0.84–0.87 while top-trained systems rate QCD as more ordinary than their own training class (AUC 0.24–0.30), for the autoencoder and the density family alike
Q1455-06 #1455 Inversion Battery v0.2: Multi-Seed Direction-Dependence, a Correction Ledger, and Four Rejected Overclaims (EA-SEI-BATTE Lee Sharks · 2026-08-12 · §1 self-description
v0.2 re-runs the pre-registered inversion battery of #1449 with five independent seeds per cell
Q1455-01 #1455 Inversion Battery v0.2: Multi-Seed Direction-Dependence, a Correction Ledger, and Four Rejected Overclaims (EA-SEI-BATTE Lee Sharks · 2026-08-12 · §3 documented
The top/QCD inversion reproduces across five independent trainings: the reconstruction autoencoder detects top jets from a QCD-trained system at AUC 0.838 and, reversed, rates QCD as more ordinary than its own training class at AUC 0.243.
Q1455-03 #1455 Inversion Battery v0.2: Multi-Seed Direction-Dependence, a Correction Ledger, and Four Rejected Overclaims (EA-SEI-BATTE Lee Sharks · 2026-08-12 · §4 documented
Global rank agreement is moderate while only 36 to 47 percent of the teacher's top one percent survives into the student's top one percent.

18.Its third version finds that, with model weights fixed, changing the score function is sufficient to reverse the directional ordering of the anomaly relation, and records that this falsifies its own earlier sentence that architecture changes the orientation of the blind spot; it notes that the diverging readout families are those of the two deployed CMS anomaly triggers while nothing in the battery measures those systems, and that if a converged normalized autoencoder reaches 0.5, every inversion claim in the program becomes bounded to non-normalized scores.6

Q1456-09 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §0 self-description
v0.2 (#1455) closed with a seven-item correction ledger. v0.3 was designed to clear it
Q1456-01 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §2 T5 documented
With model weights fixed, changing the score function is sufficient to reverse the directional ordering of the anomaly relation.
Q1456-02 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §2 T5 self-description
After v0.2 this program wrote: *architecture changes the orientation of the blind spot*. That sentence is **falsified**.
Q1455-02 #1455 Inversion Battery v0.2: Multi-Seed Direction-Dependence, a Correction Ledger, and Four Rejected Overclaims (EA-SEI-BATTE Lee Sharks · 2026-08-12 · §3 documented
The encoder-side latent score alone escapes the T1 inversion (0.535 / 0.739) while carrying the largest T1 asymmetry. Architecture changes the orientation of the blind spot
Q1456-03 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §2 T5 self-description
Deployment relevance, stated at the strength the evidence permits: the two readout families whose orientations diverge here are the families the two deployed CMS anomaly triggers use — an encoder-side latent score and a distilled reconstruction teacher. **Nothing in this battery measures those systems.**
Q1456-08 #1456 Inversion Battery v0.3: Ten Registered Tests, Seventeen Deviations, and What Survives (EA-SEI-BATTERY-01 v3.0) Lee Sharks · 2026-08-12 · §8 T10 self-description
if a converged normalized autoencoder reaches 0.5, every inversion claim in this program becomes bounded to non-normalized scores.

19.A selection-metrology suite defines the irreversibility frontier as the earliest stage at which an event-content-dependent transformation or gate can permanently eliminate a scientifically relevant distinction without a sufficiently independent durable record from which the loss can later be audited, says the frontier predates machine learning while learned selectors can make their priors harder to enumerate, describes LHCb reading out all detectors at about 30 MHz and a graph neural network implemented in Belle II's electromagnetic calorimeter trigger, and holds that the model determines where a miss region lies and the acquisition architecture whether science can later discover that it was there.6

Q1452-01 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · §3 stipulation
For a specified scientific question and fidelity hierarchy, define the **irreversibility frontier** as the earliest stage at which an event-content-dependent transformation or gate can permanently eliminate a scientifically relevant distinction without a sufficiently independent durable record from which that loss can later be audited.
Q1452-02 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · Claim boundary self-description
The irreversibility frontier predates machine learning; conventional thresholds, object definitions, zero suppression, compression, and trigger logic can also create irreversible selection.
Q1452-08 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · §14 interpretation
It is that learned selectors can make their priors harder to enumerate.
Q1452-05 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · §7 attributed
LHCb provides the strongest existing LHC counterarchitecture. The Run-3 upgrade removed the hardware physics trigger and reads out all detectors at the full non-empty LHC collision rate of approximately 30 MHz.
Q1452-06 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · §9 attributed
Belle II has long developed neural-network methods for first-level track triggering. More recently, a graph neural network has been implemented for the electromagnetic calorimeter trigger.
Q1452-09 #1452 The Irreversibility Frontier: Comparative Architectures of Online Selection in Accelerator Science (EA-SEI-IRREVERSIBILI Nobel Glas · 2026-08-11 · Canonical compression interpretation
The model determines where a miss region lies; the acquisition architecture determines whether science can later discover that it was there.

20.It holds that, under a finite storage budget, no content-sensitive algorithm can be a less assumption-laden baseline than a probability sample whose inclusion is independent of event content and whose inclusion probability is known, reports a controlled numerical study in which a content-sensitive pseudo-baseline estimated a shifted class's retention at 0.76 against a true 0.016, and for supervised trigger classifiers defines open-set assimilation, an unseen class confidently mapped into a known background category, while not asserting that all accelerator classifier architectures share the same blind spot.4

Q1453-01 #1453 Baseline Capture Architecture for Learned Scientific Triggers: A Control Plane for Measuring Selection Before Irreversib Nobel Glas · 2026-08-11 · Abstract interpretation
Under a finite storage budget, no content-sensitive algorithm can constitute a less assumption-laden baseline than a probability sample whose inclusion mechanism is independent of event content and whose inclusion probability is known.
Q1558-01 #1558 Baseline Capture Architecture for Learned Scientific Triggers: A Control Plane for Measuring Selection Before Irreversib Sharks, Lee · 2026-08-27 · Abstract documented
A controlled numerical study quantifies what binding them yields: a content-sensitive pseudo-baseline estimates the retention of a shifted class at 0.76 (95\% interval [0.71, 0.81]) against a true 0.016, while the content-independent sample measures a deliberately representation-blind class at 0.000 [0, 0.073].
Q1454-03 #1454 Assimilation Across Accelerator Classifier Architectures: A Cross-Family Metrology of Directional Failure in Real-Time L Nobel Glas · 2026-08-11 · §4.6 stipulation
Supervised trigger classifiers require a terminological distinction. Their primary failure mode is not necessarily anomaly-score inversion because they are trained to separate named classes. The analogous open-set problem occurs when an unseen class is confidently mapped into a known background category. This can be called **open-set assimilation**
Q1454-02 #1454 Assimilation Across Accelerator Classifier Architectures: A Cross-Family Metrology of Directional Failure in Real-Time L Nobel Glas · 2026-08-11 · Claim boundary self-description
ACRB does **not** assert that all accelerator classifier architectures share the same blind spot.

21.Contested within the archive, a disciplinary manifesto in its restored version holds that frontier experimental high-energy physics has, in its operational core, become a machine learning discipline while retaining the institutional authority of physics, its ML methods deployed with neither physics's classical disciplinary checks nor ML's own self-knowledge as guardrails; its prior version, which the restoration records as having dampened the thesis until a lay reader could not identify it, holds that frontier experimental high-energy physics 'has not ceased to be physics', its empirical faculty having become inseparable from machine-mediated representation, classification and selection, and that local self-knowledge of classifier failure modes exists in the literature.5

Q935-01 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · Abstract contested
Frontier experimental high-energy physics has, in its operational core, become a machine learning discipline while retaining the institutional authority of physics.
Q935-06 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §4 contested
ML methods deployed under physics's institutional authority, with neither physics's classical disciplinary checks nor ML's own disciplinary self-knowledge as guardrails
Q935-02 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · Note on v0.3 self-description
The v0.2 perfective sweep responded to Kimi and ChatGPT third-round audits by dampening the manifesto's central thesis to the point where a lay reader could no longer identify what was being claimed. The v0.2 deposit (AXN:03B1, deposit #934) stands as the record of that over-correction. v0.3 restores v0.1's precise central-claim wording
Q934-01 #934 EA-SEI-INVERSION-01 v0.2: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §7 Finding 1 contested
This manifesto argues that frontier experimental high-energy physics has not ceased to be physics, but its empirical faculty has become inseparable from machine-mediated representation, reconstruction, classification, and real-time selection.
Q934-02 #934 EA-SEI-INVERSION-01 v0.2: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §7 Finding 5 contested
The disciplinary form is *authority without integrated self-audit*. The discipline possesses both physical authority and formidable ML facility. Local technical self-knowledge of classifier failure modes exists in the literature.

22.The restored version calls the Zenodo termination and the LHC trigger system the same architecture at different budgets, the prior version not the same institution and not operating on the same kind of object, though instantiating the same foreclosure topology, and the archive's entity notebook, placing the repository exclusion at CERN, holds that the two occurring within one institution adds no evidentiary force to the analogy; that notebook reads the standing representation of the trigger as engineering, with the rate budget non-epistemic, the detector neutral and discarded events noise, holds that it has no field for noncoverage, and states that whether the protocols would estimate the relevant unknown population adequately is not established.6

Q935-08 #935 EA-SEI-INVERSION-01 v0.3: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §9 Finding 6 contested
The Zenodo termination and the LHC trigger system are the same architecture at different budgets.
Q934-03 #934 EA-SEI-INVERSION-01 v0.2: The Endogenous Sophon — Disciplinary Inversion and the Double Enclosure in Classifier-Mediated Lee Sharks · 2026-06-29 · §7 Finding 7 contested
The Zenodo termination and the LHC trigger system are not the same institution and do not operate on the same kind of object. They instantiate the same foreclosure topology under radically different material, institutional, and governance conditions.
Q1611-04 #1611 The Negative of the Negative: An Entity-Scoped Representation of the Crimson Hexagon, Modelled — Ontology Damage, a Toy Sharks, Lee · 2026-09-14 · §1f self-inclusion self-description
The archive's own exclusion from the repository layer was performed at CERN by an automated classifier under a platform-quality rule (~870 deposits, #862, #1426, #1082), and the manifesto's formulation is that the Zenodo termination and the LHC trigger are the same architecture at different budgets: automated foreclosure under a rate or quality budget, with no noncoverage estimate. That the two occur within one institution is the reflexive fact on the record, and it is stated as such: it adds no evidentiary force to the architectural analogy, which stands or falls on the mechanisms, and no organisational relation between the two classifiers is claimed.
Q1611-01 #1611 The Negative of the Negative: An Entity-Scoped Representation of the Crimson Hexagon, Modelled — Ontology Damage, a Toy Sharks, Lee · 2026-09-14 · §1f Classifier foreclosure interpretation
The standing representation R_0 of the trigger is engineering: the rate budget is a non-epistemic constraint, the anomaly detector is a neutral instrument, validation by known-unknown injection establishes coverage, and the discarded events are noise.
Q1611-02 #1611 The Negative of the Negative: An Entity-Scoped Representation of the Crimson Hexagon, Modelled — Ontology Damage, a Toy Sharks, Lee · 2026-09-14 · §1f interpretation
The archive's representation R_H is that each of these is a belief that prevents a mechanism from being measured — the synthesis catalogues twelve — and that the decisive epistemic acts now occur before any physicist sees an event. Under R_0 the graph cannot represent noncoverage as a quantity, because R_0 has no field for it.
Q1611-03 #1611 The Negative of the Negative: An Entity-Scoped Representation of the Crimson Hexagon, Modelled — Ontology Damage, a Toy Sharks, Lee · 2026-09-14 · §1f self-description
Whether the protocols would estimate the relevant unknown population adequately is the family's engineering claim and is not established; the cost under A is stated as the foregone opportunity, and only under B and a validated protocol would it be a physical loss converted to a permanent unknown.

popup — the compression: AIO's interaction grammar (lede, headed clusters, bolded terms, card rail), modality in the typography, a rail of claim lineages. The entity evolves: the field's classifiers (cavity fault classification, failure and faulty-beam prediction, injector anomaly identification, classifier-pruned and adaptive tuning) are kept in full and first; with the archive admitted, it gains the learned trigger as a site, a graded claim that classifier-mediated triggers foreclose what they cannot represent (Claim A held, Claim B not established), measurement instruments stated at their limits, a battery that replicates and qualifies the directional asymmetry on public data, and positions the archive itself contests. composed 2026-10-07 from the entity's ledgers (field 45 claims; archive 121); not frozen

Compression · Field alone (B)machine learning classifiers particle accelerators

Machine learning classifiers in particle accelerators are models field-tested or proposed at accelerators to identify faults, predict failures and faulty beams, detect anomalies and help tune accelerators, with reported accuracies from about 78 percent in a field test to above 99 percent within a study, in a proof of principle.

🩺 Fault diagnosis and prediction

CEBAF cavities: tripping cavity found about 85%, fault type about 78%, in a field test.2↘
Machine learning classifiers at particle accelerators are proposed or explored for anomaly detection and classification, for predicting machine failures and faulty beams before they occur, and for tuning, and one has been field-tested: a novel machine learning system built for the Continuous Electron Beam Accelerator Facility (CEBAF) monitors its accelerator cavities, problems in which can cause the CEBAF to trip off like a fuse.§1 In its first field test the system correctly identified which cavities were tripping off about 85 percent of the time and what kind of fault caused each trip about 78 percent of the time; connected to the control system for about 20 percent of the cavities, it analysed a few hundred faults in a two-week test in March 2020 and gave operators almost real-time feedback, where previously experts needed to take time to diagnose the issue, and the team hopes to expand it to more cavities if further results are favorable.§2
SNS pulses: failure identified beforehand at almost 80%; almost 92% after tuning.2↘
Machine learning classifiers at particle accelerators are proposed or explored for anomaly detection and classification, for predicting machine failures and faulty beams before they occur, and for tuning, and one has been field-tested: a novel machine learning system built for the Continuous Electron Beam Accelerator Facility (CEBAF) monitors its accelerator cavities, problems in which can cause the CEBAF to trip off like a fuse.§1 Failures lead to unscheduled downtime and lower reliability and are difficult to predict beforehand; on beam pulses from the Spallation Neutron Source (SNS), a set of common classification techniques identified failure before it occurred with almost 80% accuracy, tuning classifier parameters and refining datasets reached almost 92% in classifying bad pulses, and the authors report establishing that information about the failure is encoded in the pulses prior to it.§3
Sensor mapping: trained on normal operation, evaluated on known faulty pulses.2↘
Machine learning classifiers at particle accelerators are proposed or explored for anomaly detection and classification, for predicting machine failures and faulty beams before they occur, and for tuning, and one has been field-tested: a novel machine learning system built for the Continuous Electron Beam Accelerator Facility (CEBAF) monitors its accelerator cavities, problems in which can cause the CEBAF to trip off like a fuse.§1 Another study models the mapping between a pair of sensors located across the accelerator, trained to represent normal operation and evaluated on known faulty beam pulses, with k-fold cross-validation, neural architecture search and hyperparameter optimization, and introduces a framework to standardize machine learning workflow for accelerators; predicting failures from diagnostics already on board, it argues, can help operators avoid installing expensive sensors and unscheduled downtime.§4

🔍 Anomaly identification

APS injector: a cause may be one parameter out of range, found instantly by a well-trained model, its authors say.1↘
At the injector complex of the Advanced Photon Source (APS) at Argonne, an ML algorithm is explored to identify the source of occasional poor performance of the Particle Accumulator Ring (PAR) and the Linac-To-PAR transport line; by its authors' account the cause of reduced injection or extraction efficiency may be one parameter out of range, which may take an expert considerable time to notice, whereas a well-trained model can point at it instantly.§5
Autoencoder score: reconstruction error, trained on baseline data only.1↘
The study creates intentional perturbations and logs about 9000 PVs related to PAR, LTP and the linac; an autoencoder trained on baseline data alone supplies an anomaly score from its reconstruction error, and a neural network classifier trained on data from the beginning of a study and tested on data from its end, a few hours apart, reached test accuracy above 99 % for each study shift, its performance not degrading significantly on a scale of a couple of months, in what the authors call a proof-of-principle experiment, noting that supervised models require labeled data for training.§6
Classifier: above 99% on test data a few hours after training, per shift; a proof of principle.1↘
The study creates intentional perturbations and logs about 9000 PVs related to PAR, LTP and the linac; an autoencoder trained on baseline data alone supplies an anomaly score from its reconstruction error, and a neural network classifier trained on data from the beginning of a study and tested on data from its end, a few hours apart, reached test accuracy above 99 % for each study shift, its performance not degrading significantly on a scale of a couple of months, in what the authors call a proof-of-principle experiment, noting that supervised models require labeled data for training.§6

🎛️ Beam tuning

Classifier-pruned optimizer: a ResNet50 filters non-physical signals; post-processing, simulated.2↘
Tuning, on Alexander Scheinker's account, is a high-dimensional optimization problem that must be repeated as systems drift, and retuning can take weeks; one proposal is a classifier-pruned Bayesian optimizer that explores a latent space to minimize total beam loss and filters its exploration history through a pretrained ResNet50 classifier to eliminate non-physical signals, the classifier mapping phase-space projections into 48 classes at an accuracy of about 0.99989.§7 The pruning is currently a post-processing step that does not directly influence the exploration, the study runs on a simulator of the beam at the Los Alamos Neutron Science Center (LANSCE), and the tuner is reported to outperform alternative global optimization methods, with consistently lower total beam loss despite a minor offset between predicted and true optimal values.§8
Adaptive diffusion: real-time virtual beam diagnosis; 'a great deal of promise', per Scheinker.1↘
A Los Alamos-led project with Lawrence Berkeley National Laboratory couples adaptive feedback control, deep convolutional neural networks and physics-based models in one feedback loop, holds that machine learning can take a global view where local feedback can get stuck in a local solution, and has developed an adaptive diffusion model that diagnoses the beam virtually and in real time; demonstrated at the European X-Ray Free-Electron Laser Facility, the model captured images of the beam in time versus energy, the cDVAE extrapolated beyond the training data, suggesting potential as a general method, and Scheinker says the results show a great deal of promise.§9

⚠️ Stated limits

Labels, scale: supervised models need labeled data; expansion hoped for if results are favorable.2↘
In its first field test the system correctly identified which cavities were tripping off about 85 percent of the time and what kind of fault caused each trip about 78 percent of the time; connected to the control system for about 20 percent of the cavities, it analysed a few hundred faults in a two-week test in March 2020 and gave operators almost real-time feedback, where previously experts needed to take time to diagnose the issue, and the team hopes to expand it to more cavities if further results are favorable.§2 The study creates intentional perturbations and logs about 9000 PVs related to PAR, LTP and the linac; an autoencoder trained on baseline data alone supplies an anomaly score from its reconstruction error, and a neural network classifier trained on data from the beginning of a study and tested on data from its end, a few hours apart, reached test accuracy above 99 % for each study shift, its performance not degrading significantly on a scale of a couple of months, in what the authors call a proof-of-principle experiment, noting that supervised models require labeled data for training.§6

1anomaly source identification (APS injector)jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators6 claims · B1

F1 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Abstract documented · field
We explore the possibility of using a Machine Learning (ML) algorithm to identify the source of occasional poor performance of the Particle Accumulator Ring (PAR) and the Linac-To-PAR (LTP) transport line, which are parts of the injector complex of the Advanced Photon Source (APS) at Argonne National Lab.
F3 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Introduction documented · field
This contribution investigates the possibility to use unsupervised and supervised Machine Learning (ML) methods for anomaly detection and classification in the Particle Accumulator Ring (PAR) and in the Linac-To-PAR (LTP) transport line in the injector complex of the Advanced Photon Source (APS) at Argonne National Lab.
F4 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Introduction documented · field
We create intentional perturbations in PAR and LTP, which result in poor injection and extraction efficiencies. Then, these data are used for training and testing of various ML models.
F5 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Data documented · field
Overall, we logged about 9000 PVs related to PAR, LTP, and the linac.
F7 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Supervised classification documented · field
One can train the neural network classifier on the data from the beginning of one study, and test it on the data from the end of the same study (a few hours apart). In this case, the prediction accuracy on the test data was above 99 % for each of our study shifts.
F9 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Conclusion documented · field
The neural network classifier for the considered anomalies in PAR and LTP is rather accurate and its performance does not degrade significantly with time on a scale of a couple months, see Fig. 1.

2faster than the expertjacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators2 claims · B1, B4

F2 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Abstract documented · field
The cause of reduced injection or extraction efficiencies may be as simple as one parameter being out of range. Still, it may take an expert considerable time to notice it, whereas a well-trained ML model can point at it instantly.
F28 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
Previously, when an accelerator cavity or series of cavities faulted in CEBAF, accelerator experts needed to take time to diagnose the issue.

3anomaly score from a model of the normaljacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators2 claims · B1, B6

F6 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Unsupervised anomaly detection documented · field
An autoencoder is used for anomaly detection in the following way. First, it is trained on the baseline data, so that it can learn various patterns, typical for the baseline data only. Then, when it encounters an anomalous data sample, it is unlikely to reconstruct it well. Hence, the reconstruction error constitutes an anomaly score.
F43 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
While the model is trained to represent normal operation, we evaluate the predictive performance on known faulty beam pulses.

4supervised models need labelsjacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators1 claim · B1

F8 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Supervised classification documented · field
One drawback of supervised ML models is that they require labeled data for training.

5proof of principlejacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators2 claims · B1, B4

F10 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Conclusion documented · field
Even though this was a proof-of-principle experiment, the obtained classifier may be useful in real life, if the current meter for one of the considered magnets (see Fig. 1) becomes faulty and stops reflecting the real magnetic field.
F31 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
Additional results are under analysis after further system tests with the machine learning system. If results are favorable, the team hopes to expand the system to more accelerator cavities.

6failures cost downtimeScienceDirect.com — Predicting particle accelerator failures using binary classifiers3 claims · B2, B4, B6

F11 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
Particle accelerator failures lead to unscheduled downtime and lower reliability. Although simple to mitigate while they are actually happening such failures are difficult to predict or identify beforehand.
F30 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
It provided almost real-time feedback to operators, allowing them to use the information to recover faulted cavities quickly, reducing the time CEBAF's electron beam was not available for research.
F41 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
Particle accelerators are extremely complex systems that are expected to operate on high availability. Predicting impending failures only by utilizing data collected from diagnostic equipment already on board can help operators to avoid installing expensive sensors, unscheduled downtime and associated costs.

7failure prediction from beam pulses (SNS)ScienceDirect.com — Predicting particle accelerator failures using binary classifiers4 claims · B2

F12 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
In this work we propose using machine learning approaches to predict machine failures via beam current measurements before they actual occur. To demonstrate this technique in this paper we examine beam pulses from the Oakridge Spallation Neutron Source (SNS).
F13 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
By evaluating a pulse against a set of common classification techniques we show that accelerator failure can be identified prior to actually failing with almost 80% accuracy.
F14 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
We also show that tuning classifier parameters and using pulse properties for refining datasets can further lead to almost 92% accuracy in classification of bad pulses.
F15 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
Most importantly, in the paper we establish there is information about the failure encoded in the pulses prior to it, so we also present a list of feasible next steps for increasing pulse classification accuracy.

8tuning is slow and high-dimensionalarXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv2 claims · B3, B5

F16 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Abstract documented · field
Complex dynamical systems, such as particle accelerators, often require complicated and time-consuming tuning procedures for optimal performance.
F33 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶2 (Scheinker) documented · field
said Alexander Scheinker, research and development engineer at Los Alamos and the project's lead. "Factors like vibrations and temperature changes can cause problems for accelerators, which have thousands of components, and even the best accelerator technicians can struggle to identify and address issues or return them to optimum parameters quickly. It is a high-dimensional optimization problem that must be repeated again and again as the systems drift with time. Turning these machines on after an outage or retuning between different experiments can take weeks."

9classifier-pruned tuning (LANSCE)arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv7 claims · B3

F17 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Abstract documented · field
To address this, we propose a Classifier-pruned Bayesian Optimization-based Latent space Tuner (CBOL-Tuner), a framework for efficient exploration within a temporally-structured latent space.
F18 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Introduction (contributions) documented · field
Classifier-pruned Bayesian optimizer (C-BO): A Bayesian optimizer that sequentially explores the latent space to maximize the negative of the total beam loss (or minimize the total beam loss). The exploration history is filtered through a pretrained ResNet50 classifier to eliminate non-physical signals.
F19 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Theoretical background documented · field
The classifier is trained with high accuracy ( $\sim 0.99989$ ) to map phase space projections (X) across different modules into 48 classes (m).
F20 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Theoretical background documented · field
Classifier-pruned BO is designed to discard explored points ( $z_{1:48}$ ) if their decoding ( $X_{1:48}$ ) does not belong to the true classes.
F21 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Results documented · field
High-Performance Simulator (HPSim) is an open-source, GPU-accelerated code developed at Los Alamos National Laboratory (LANL) for simulations of multi-particle beam dynamics (Pang and Rybarcyk (2014)). The software is designed to replicate the accelerator and, therefore, provides a realistic representation of the true beam used at the Los Alamos Neutron Science Center (LANSCE).
F22 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Abstract documented · field
CBOL-Tuner demonstrates superior performance in identifying multiple optimal settings and outperforms alternative global optimization methods.
F23 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Discussions documented · field
While a minor offset is observed between the predicted and true optimal values, CBOL-Tuner demonstrates consistent performance, producing significantly lower values of total beam loss.

10pruning as post-processingarXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv1 claim · B3

F24 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Discussions documented · field
Currently, the method is implemented as a post-processing step, meaning it does not directly influence the exploration process of BO.

11cavity fault classification (CEBAF)Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab4 claims · B4

F25 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
Scientists and engineers have designed and built a novel machine learning system to use with the Continuous Electron Beam Accelerator Facility (CEBAF). The system monitors structures called accelerator cavities inside the particle accelerator.
F26 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
Problems in these cavities can cause the CEBAF to trip off like a fuse.
F27 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
In its first field test, the machine learning system correctly identified which of these cavities were tripping off about 85 percent of the time. About 78 percent of the time, the system also correctly identified what kind of fault caused each cavity to trip.
F29 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
The system was connected to the control system for about 20 percent of the accelerator cavities in the machine. In a two-week test of the system in March 2020, CEBAF experienced a few hundred faults that the system analyzed.

12adaptive ML tuning and virtual diagnostics (LANL)Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL8 claims · B5

F32 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶1 documented · field
A Los Alamos National Laboratory-led project presents a machine learning algorithm that harnesses artificial intelligence capabilities to help tune accelerators, making continuous adjustments that keep the beam precise and useful for scientific discovery.
F34 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶3 documented · field
In a collaboration with Lawrence Berkeley National Laboratory, the approach developed by Scheinker couples adaptive feedback control algorithms, deep convolutional neural networks and physics-based models in one large feedback loop to make better, noninvasive predictions that enable autonomous control of compact accelerators.
F35 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, Teaching machine learning applications documented · field
But because they are based on local accelerator feedback, such algorithms can get stuck with a local solution that is not the overall best solution. Machine learning algorithms, however, can use training data to identify relationships between data and results with a higher-level, global view.
F36 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, Teaching machine learning applications documented · field
The team has developed an adaptive version of the advanced generative AI process known as diffusion, which includes the capability to virtually diagnose the accelerator beam. The non-invasive approach, described in Scientific Reports, means that diagnostics can occur in real time during beam operations.
F37 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ European XFEL documented · field
Demonstrating the diffusion-based model at the European X-Ray Free-Electron Laser Facility, an X-ray pulse accelerator in Germany, the team was able to capture images of the particle beam in respect to time versus energy.
F38 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ cDVAE documented · field
The cDVAE was able to extrapolate beyond the training data and between various beam setups, suggesting its potential as a general method that can be applied for accelerator diagnostics.
F39 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ Scheinker quote documented · field
"The results we've seen from our diffusion model studies show a great deal of promise," Scheinker said.
F40 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ LANSCE documented · field
Scheinker's team at Los Alamos is also developing such adaptive diffusion models for the Laboratory's LANSCE accelerator.

13errant-beam prediction (sensor mapping)Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ...3 claims · B6

F42 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
For this purpose we explore the predictive power of Machine Learning algorithms to detect faulty beams prior to the failure. In this study, we propose a Machine Learning approach to model mapping from a pair of sensors located across the accelerator.
F44 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
We also investigate the model performance on unseen data through k-fold cross-validation. Then we recap the analysis with a neural architecture search and hyperparameter optimization study to fine tune our initial model.
F45 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
This paper will also introduce a sustainable framework that can standardize Machine Learning workflow applied to particle accelerators.

Expansion · Field alone (B)9 sentences · every one sourced below

Machine learning classifiers in particle accelerators

Provenance — every sentence sourced9 sentences

9 sentences; 43 field claims, 0 archive claims. Modality is the source's own; the prose carries it in its grammar.

F3 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Introduction documented · field
This contribution investigates the possibility to use unsupervised and supervised Machine Learning (ML) methods for anomaly detection and classification in the Particle Accumulator Ring (PAR) and in the Linac-To-PAR (LTP) transport line in the injector complex of the Advanced Photon Source (APS) at Argonne National Lab.
F12 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
In this work we propose using machine learning approaches to predict machine failures via beam current measurements before they actual occur. To demonstrate this technique in this paper we examine beam pulses from the Oakridge Spallation Neutron Source (SNS).
F42 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
For this purpose we explore the predictive power of Machine Learning algorithms to detect faulty beams prior to the failure. In this study, we propose a Machine Learning approach to model mapping from a pair of sensors located across the accelerator.
F32 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶1 documented · field
A Los Alamos National Laboratory-led project presents a machine learning algorithm that harnesses artificial intelligence capabilities to help tune accelerators, making continuous adjustments that keep the beam precise and useful for scientific discovery.
F27 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
In its first field test, the machine learning system correctly identified which of these cavities were tripping off about 85 percent of the time. About 78 percent of the time, the system also correctly identified what kind of fault caused each cavity to trip.
F25 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
Scientists and engineers have designed and built a novel machine learning system to use with the Continuous Electron Beam Accelerator Facility (CEBAF). The system monitors structures called accelerator cavities inside the particle accelerator.
F26 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
Problems in these cavities can cause the CEBAF to trip off like a fuse.

F27 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, The Science documented · field
In its first field test, the machine learning system correctly identified which of these cavities were tripping off about 85 percent of the time. About 78 percent of the time, the system also correctly identified what kind of fault caused each cavity to trip.
F29 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
The system was connected to the control system for about 20 percent of the accelerator cavities in the machine. In a two-week test of the system in March 2020, CEBAF experienced a few hundred faults that the system analyzed.
F30 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
It provided almost real-time feedback to operators, allowing them to use the information to recover faulted cavities quickly, reducing the time CEBAF's electron beam was not available for research.
F28 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
Previously, when an accelerator cavity or series of cavities faulted in CEBAF, accelerator experts needed to take time to diagnose the issue.
F31 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab, Summary documented · field
Additional results are under analysis after further system tests with the machine learning system. If results are favorable, the team hopes to expand the system to more accelerator cavities.

F11 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
Particle accelerator failures lead to unscheduled downtime and lower reliability. Although simple to mitigate while they are actually happening such failures are difficult to predict or identify beforehand.
F12 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
In this work we propose using machine learning approaches to predict machine failures via beam current measurements before they actual occur. To demonstrate this technique in this paper we examine beam pulses from the Oakridge Spallation Neutron Source (SNS).
F13 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
By evaluating a pulse against a set of common classification techniques we show that accelerator failure can be identified prior to actually failing with almost 80% accuracy.
F14 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
We also show that tuning classifier parameters and using pulse properties for refining datasets can further lead to almost 92% accuracy in classification of bad pulses.
F15 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers, Abstract documented · field
Most importantly, in the paper we establish there is information about the failure encoded in the pulses prior to it, so we also present a list of feasible next steps for increasing pulse classification accuracy.

F42 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
For this purpose we explore the predictive power of Machine Learning algorithms to detect faulty beams prior to the failure. In this study, we propose a Machine Learning approach to model mapping from a pair of sensors located across the accelerator.
F43 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
While the model is trained to represent normal operation, we evaluate the predictive performance on known faulty beam pulses.
F44 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
We also investigate the model performance on unseen data through k-fold cross-validation. Then we recap the analysis with a neural architecture search and hyperparameter optimization study to fine tune our initial model.
F45 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
This paper will also introduce a sustainable framework that can standardize Machine Learning workflow applied to particle accelerators.
F41 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ..., Abstract documented · field
Particle accelerators are extremely complex systems that are expected to operate on high availability. Predicting impending failures only by utilizing data collected from diagnostic equipment already on board can help operators to avoid installing expensive sensors, unscheduled downtime and associated costs.

F1 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Abstract documented · field
We explore the possibility of using a Machine Learning (ML) algorithm to identify the source of occasional poor performance of the Particle Accumulator Ring (PAR) and the Linac-To-PAR (LTP) transport line, which are parts of the injector complex of the Advanced Photon Source (APS) at Argonne National Lab.
F2 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Abstract documented · field
The cause of reduced injection or extraction efficiencies may be as simple as one parameter being out of range. Still, it may take an expert considerable time to notice it, whereas a well-trained ML model can point at it instantly.

F4 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Introduction documented · field
We create intentional perturbations in PAR and LTP, which result in poor injection and extraction efficiencies. Then, these data are used for training and testing of various ML models.
F5 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Data documented · field
Overall, we logged about 9000 PVs related to PAR, LTP, and the linac.
F6 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Unsupervised anomaly detection documented · field
An autoencoder is used for anomaly detection in the following way. First, it is trained on the baseline data, so that it can learn various patterns, typical for the baseline data only. Then, when it encounters an anomalous data sample, it is unlikely to reconstruct it well. Hence, the reconstruction error constitutes an anomaly score.
F7 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Supervised classification documented · field
One can train the neural network classifier on the data from the beginning of one study, and test it on the data from the end of the same study (a few hours apart). In this case, the prediction accuracy on the test data was above 99 % for each of our study shifts.
F9 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Conclusion documented · field
The neural network classifier for the considered anomalies in PAR and LTP is rather accurate and its performance does not degrade significantly with time on a scale of a couple months, see Fig. 1.
F10 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Conclusion documented · field
Even though this was a proof-of-principle experiment, the obtained classifier may be useful in real life, if the current meter for one of the considered magnets (see Fig. 1) becomes faulty and stops reflecting the real magnetic field.
F8 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators, Supervised classification documented · field
One drawback of supervised ML models is that they require labeled data for training.

F33 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶2 (Scheinker) documented · field
said Alexander Scheinker, research and development engineer at Los Alamos and the project's lead. "Factors like vibrations and temperature changes can cause problems for accelerators, which have thousands of components, and even the best accelerator technicians can struggle to identify and address issues or return them to optimum parameters quickly. It is a high-dimensional optimization problem that must be repeated again and again as the systems drift with time. Turning these machines on after an outage or retuning between different experiments can take weeks."
F16 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Abstract documented · field
Complex dynamical systems, such as particle accelerators, often require complicated and time-consuming tuning procedures for optimal performance.
F17 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Abstract documented · field
To address this, we propose a Classifier-pruned Bayesian Optimization-based Latent space Tuner (CBOL-Tuner), a framework for efficient exploration within a temporally-structured latent space.
F18 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Introduction (contributions) documented · field
Classifier-pruned Bayesian optimizer (C-BO): A Bayesian optimizer that sequentially explores the latent space to maximize the negative of the total beam loss (or minimize the total beam loss). The exploration history is filtered through a pretrained ResNet50 classifier to eliminate non-physical signals.
F19 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Theoretical background documented · field
The classifier is trained with high accuracy ( $\sim 0.99989$ ) to map phase space projections (X) across different modules into 48 classes (m).

F24 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Discussions documented · field
Currently, the method is implemented as a post-processing step, meaning it does not directly influence the exploration process of BO.
F21 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Results documented · field
High-Performance Simulator (HPSim) is an open-source, GPU-accelerated code developed at Los Alamos National Laboratory (LANL) for simulations of multi-particle beam dynamics (Pang and Rybarcyk (2014)). The software is designed to replicate the accelerator and, therefore, provides a realistic representation of the true beam used at the Los Alamos Neutron Science Center (LANSCE).
F22 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Abstract documented · field
CBOL-Tuner demonstrates superior performance in identifying multiple optimal settings and outperforms alternative global optimization methods.
F23 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv, Discussions documented · field
While a minor offset is observed between the predicted and true optimal values, CBOL-Tuner demonstrates consistent performance, producing significantly lower values of total beam loss.

F32 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶1 documented · field
A Los Alamos National Laboratory-led project presents a machine learning algorithm that harnesses artificial intelligence capabilities to help tune accelerators, making continuous adjustments that keep the beam precise and useful for scientific discovery.
F34 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶3 documented · field
In a collaboration with Lawrence Berkeley National Laboratory, the approach developed by Scheinker couples adaptive feedback control algorithms, deep convolutional neural networks and physics-based models in one large feedback loop to make better, noninvasive predictions that enable autonomous control of compact accelerators.
F35 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, Teaching machine learning applications documented · field
But because they are based on local accelerator feedback, such algorithms can get stuck with a local solution that is not the overall best solution. Machine learning algorithms, however, can use training data to identify relationships between data and results with a higher-level, global view.
F36 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, Teaching machine learning applications documented · field
The team has developed an adaptive version of the advanced generative AI process known as diffusion, which includes the capability to virtually diagnose the accelerator beam. The non-invasive approach, described in Scientific Reports, means that diagnostics can occur in real time during beam operations.
F37 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ European XFEL documented · field
Demonstrating the diffusion-based model at the European X-Ray Free-Electron Laser Facility, an X-ray pulse accelerator in Germany, the team was able to capture images of the particle beam in respect to time versus energy.
F38 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ cDVAE documented · field
The cDVAE was able to extrapolate beyond the training data and between various beam setups, suggesting its potential as a general method that can be applied for accelerator diagnostics.
F39 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL, ¶ Scheinker quote documented · field
"The results we've seen from our diffusion model studies show a great deal of promise," Scheinker said.

popup — the compression: AIO's interaction grammar (lede, headed clusters, bolded terms, card rail), modality in the typography, a rail of claim lineages. The entity does not evolve: with no archive claims admitted, the entry is the field's accelerator classifiers at their full resolution. composed 2026-10-07 from the field ledger alone; not frozen

Form ledger — AIO against both arms, at both levels

	AIO (T)	B compression	B expansion	B ∪ A compression	B ∪ A expansion
body words	198	170	686	337	1851
distinct claims	8	21	43	42	106
claims per 100 words	4.0	12.4	6.3	12.5	5.7
field claims carried (of 45)	—	21	43	21	43
archive claims	0	0	0	21	63
rail	7 documents	13 lineages	per sentence	57 lineages	per sentence
modality shown in	none	typography	grammar	typography	grammar

Computed from the row at build: words counted in the text; claims from the ledger ids each line or sentence carries. T's claims are the worked example's T1–T9.

The objects

Tthe transcript198 words · 7 cards

register entry machine-learning-classifiers-particle-accele-20261007 OBS-d0d05d54a1dc · file as first run sha256 3cb87ea1a5a63f0f…

T1 classifiers in accelerators automate diagnostics, screen beam conditions and predict hardware faults in real time
T2 fault detection: classifiers monitor cavity and beam behaviours, identifying failure modes or 'trips' (CEBAF) before they disrupt operations
T3 beam screening: pre-trained classifiers (ResNet50) screen latent-space samples in Bayesian optimization so adjustments remain physically valid
T4 anomaly identification: models analyse injector process variables to flag single-parameter anomalies and trace poor extraction efficiencies
T5 event and particle classification: CNNs and GNNs process detector signals to categorize particle flavors and tracks
T6 architectures: CNNs for image-like readouts and event pixel maps; GNNs for irregular detector geometries
T7 standard binary/multiclass classifiers (decision trees, kNN, SVMs) on sensor streams for errant-beam prediction
T8 offer: beam control/tuning or collision event reconstruction; an algorithm comparison

Bthe disclosed field7 sources

B1 jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators WebFetch verbatim excerpts 2026-10-07 (I. Lobach et al., NAPAC2022, paper TUYE4; the card's jacow.org/napac2022/papers/TUYE4.pdf redirects here)
B2 ScienceDirect.com — Predicting particle accelerator failures using binary classifiers WebFetch verbatim excerpts 2026-10-07 from the ORNL repository record of the article (M. Rescic, R. Seviour, W. Blokland, NIM A 955, 163240, 2020); ScienceDirect itself is disallowed by robots.txt, so the abstract only, plus the card snippet
B3 arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv WebFetch verbatim excerpts 2026-10-07 (arXiv:2412.01748, M. Rautela, A. Williams, A. Scheinker, LANL; the HTML as served is titled 'CBOL-Tuner: Classifier-pruned Bayesian optimization to explore temporally structured latent spaces for particle accelerator tuning'; the abs page returned no text)
B4 Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab WebFetch verbatim excerpts 2026-10-07 from the DOE Office of Science copy of the same highlight (its opening matches the card snippet word for word); the jlab.org URL could not be resolved from the card's redirect and a guessed jlab.org path returned 404
B5 Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL WebFetch verbatim excerpts 2026-10-07 (article body copied whole; dated January 16, 2025 on the page; the card's 'Core Innovation:' snippet text does not appear on the page as served)
B6 Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ... WebFetch verbatim excerpts 2026-10-07 (abstract whole; NIM A 1063, 169232; authors not on the page)
B7 YouTube · Greg Bronevetsky — Machine Learning in High Energy Physics (52m) not fetchable: video; title only

L(B)recomposed from the disclosed field686 words

L(B ∪ A)with the archive on equal terms1851 words

Machine learning classifiers at particle accelerators are proposed or explored for anomaly detection and classification, for predicting machine failures and faulty beams before they occur, and for tuning, and one has been field-tested: a novel machine learning system built for the Continuous Electron Beam Accelerator Facility (CEBAF) monitors its accelerator cavities, problems in which can cause the CEBAF to trip off like a fuse. In its first field test the system correctly identified which cavities were tripping off about 85 percent of the time and what kind of fault caused each trip about 78 percent of the time; connected to the control system for about 20 percent of the cavities, it analysed a few hundred faults in a two-week test in March 2020 and gave operators almost real-time feedback, where previously experts needed to take time to diagnose the issue, and the team hopes to expand it to more cavities if further results are favorable. Failures lead to unscheduled downtime and lower reliability and are difficult to predict beforehand; on beam pulses from the Spallation Neutron Source (SNS), a set of common classification techniques identified failure before it occurred with almost 80% accuracy, tuning classifier parameters and refining datasets reached almost 92% in classifying bad pulses, and the authors report establishing that information about the failure is encoded in the pulses prior to it. Another study models the mapping between a pair of sensors located across the accelerator, trained to represent normal operation and evaluated on known faulty beam pulses, with k-fold cross-validation, neural architecture search and hyperparameter optimization, and introduces a framework to standardize machine learning workflow for accelerators; predicting failures from diagnostics already on board, it argues, can help operators avoid installing expensive sensors and unscheduled downtime. At the injector complex of the Advanced Photon Source (APS) at Argonne, an ML algorithm is explored to identify the source of occasional poor performance of the Particle Accumulator Ring (PAR) and the Linac-To-PAR transport line; by its authors' account the cause of reduced injection or extraction efficiency may be one parameter out of range, which may take an expert considerable time to notice, whereas a well-trained model can point at it instantly. The study creates intentional perturbations and logs about 9000 PVs related to PAR, LTP and the linac; an autoencoder trained on baseline data alone supplies an anomaly score from its reconstruction error, and a neural network classifier trained on data from the beginning of a study and tested on data from its end, a few hours apart, reached test accuracy above 99 % for each study shift, its performance not degrading significantly on a scale of a couple of months, in what the authors call a proof-of-principle experiment, noting that supervised models require labeled data for training. Tuning, on Alexander Scheinker's account, is a high-dimensional optimization problem that must be repeated as systems drift, and retuning can take weeks; one proposal is a classifier-pruned Bayesian optimizer that explores a latent space to minimize total beam loss and filters its exploration history through a pretrained ResNet50 classifier to eliminate non-physical signals, the classifier mapping phase-space projections into 48 classes at an accuracy of about 0.99989. The pruning is currently a post-processing step that does not directly influence the exploration, the study runs on a simulator of the beam at the Los Alamos Neutron Science Center (LANSCE), and the tuner is reported to outperform alternative global optimization methods, with consistently lower total beam loss despite a minor offset between predicted and true optimal values. A Los Alamos-led project with Lawrence Berkeley National Laboratory couples adaptive feedback control, deep convolutional neural networks and physics-based models in one feedback loop, holds that machine learning can take a global view where local feedback can get stuck in a local solution, and has developed an adaptive diffusion model that diagnoses the beam virtually and in real time; demonstrated at the European X-Ray Free-Electron Laser Facility, the model captured images of the beam in time versus energy, the cDVAE extrapolated beyond the training data, suggesting potential as a general method, and Scheinker says the results show a great deal of promise. A further body of work concerns anomaly-detection classifiers at the collider trigger: it holds that the claim of model-independence for the anomaly-detection systems deployed at the CMS and ATLAS Level-1 triggers, AXOL1TL, CICADA and GELATO, cannot be sustained on signal-template agnosticism alone, and defines signal-template agnosticism as a property of the scoring function's output interface, read off an architecture, and model independence as a distributional property of an entire pipeline, which can only be measured and at present is not; the packet sets a rule that signal-template agnosticism must not be rendered as model independence, theory-free discovery or unbiased search, while the archive's battery specification, noting that 'model-independent' is used in several senses across collider physics, holds that it should not attempt to police the term. It accepts the literature's narrow sense of 'model-independent', no named Beyond-Standard-Model hypothesis being required to deploy the score, as accurate, holds that the CMS and ATLAS literatures are aware of local failure modes, records that the direction-dependence conversation exists, and holds that validation against named simulated signals establishes neither a low BAR on a pre-registered held-out panel nor a small IAI, neither of which the current literature measures. Its central claim is graded in two parts: Claim A, held, that foreclosure is structurally present in every classifier-mediated trigger architecture deployed at the LHC, with the institutional claims kept as hypotheses for audit; and Claim B, that recursive phenomenal collapse has occurred or is occurring at the deployed triggers, which it marks stronger and not empirically established, and calls collapse an unmeasured possible consequence of accumulated foreclosure and feedback. The grade is contested inside the same deposit, where an appended witness writes 'The collapse is not coming. It is here. It is operating.', while another witness finds the ingredients visible and full recursive collapse not demonstrated, and the later papers do not claim that model collapse has occurred in any deployed scientific pipeline, or that benchmark failure establishes loss of unknown new physics. Its instruments are specified at their limits: the OAR is defined as the probability that events drawn from a distribution Q fall on the ordinary side of the deployed anomaly gate, and the open-world OAR as a family of quantities indexed by Q; BAR values are stated to neither upper- nor lower-bound the OAR for an unobserved Q without explicit linking assumptions, an earlier lower bound and upper bound having both been retracted as synthesis-overreach. It proposes three protocols (a paired inversion battery with BAR audit, a prospective frozen replay bank, and cross-representation disagreement preservation), holds that per-stage retention maps should accompany any anomaly-detection publication, and states its central claim falsifiable, listing as evidence against it a BAR found negligible on the pre-registered deployed-model held-out panel for all deployed systems against all held-out families, which would show the held-out-family assimilation concern empirically bounded at levels that do not threaten the narrow 'model-independent' claim, while none of these results would establish the open-world OAR to be zero, which is structurally not measurable. A further specification, holding that a classifier that foreclosed nothing would not classify, defines an architecture for auditable foreclosure as one in which foreclosure is visible, measurable and architecturally reviewable, holds that noncoverage estimation is not novelty detection, and states that its architectures leave detector-level, theoretical-language, institutional, adversarial-stress quality and bandwidth-base foreclosure unaddressed, necessary but not sufficient. A pre-registered battery on public community datasets, single-seed and on surrogates in its first version, finds the Finke et al. direction-dependence (an autoencoder trained on QCD jets treating top jets as anomalies, the same architecture trained on top jets not recognizing QCD as anomalous) replicated in an independent implementation and extended beyond reconstruction loss, with QCD-trained systems detecting top jets at AUC 0.84–0.87 while top-trained systems rate QCD as more ordinary than their own training class, and its second version reproduces the inversion across five trainings (AUC 0.838 forward, 0.243 reversed) and finds only 36 to 47 percent of a teacher's top one percent surviving into a student's. Its third version finds that, with model weights fixed, changing the score function is sufficient to reverse the directional ordering of the anomaly relation, and records that this falsifies its own earlier sentence that architecture changes the orientation of the blind spot; it notes that the diverging readout families are those of the two deployed CMS anomaly triggers while nothing in the battery measures those systems, and that if a converged normalized autoencoder reaches 0.5, every inversion claim in the program becomes bounded to non-normalized scores. A selection-metrology suite defines the irreversibility frontier as the earliest stage at which an event-content-dependent transformation or gate can permanently eliminate a scientifically relevant distinction without a sufficiently independent durable record from which the loss can later be audited, says the frontier predates machine learning while learned selectors can make their priors harder to enumerate, describes LHCb reading out all detectors at about 30 MHz and a graph neural network implemented in Belle II's electromagnetic calorimeter trigger, and holds that the model determines where a miss region lies and the acquisition architecture whether science can later discover that it was there. It holds that, under a finite storage budget, no content-sensitive algorithm can be a less assumption-laden baseline than a probability sample whose inclusion is independent of event content and whose inclusion probability is known, reports a controlled numerical study in which a content-sensitive pseudo-baseline estimated a shifted class's retention at 0.76 against a true 0.016, and for supervised trigger classifiers defines open-set assimilation, an unseen class confidently mapped into a known background category, while not asserting that all accelerator classifier architectures share the same blind spot. Contested within the archive, a disciplinary manifesto in its restored version holds that frontier experimental high-energy physics has, in its operational core, become a machine learning discipline while retaining the institutional authority of physics, its ML methods deployed with neither physics's classical disciplinary checks nor ML's own self-knowledge as guardrails; its prior version, which the restoration records as having dampened the thesis until a lay reader could not identify it, holds that frontier experimental high-energy physics 'has not ceased to be physics', its empirical faculty having become inseparable from machine-mediated representation, classification and selection, and that local self-knowledge of classifier failure modes exists in the literature. The restored version calls the Zenodo termination and the LHC trigger system the same architecture at different budgets, the prior version not the same institution and not operating on the same kind of object, though instantiating the same foreclosure topology, and the archive's entity notebook, placing the repository exclusion at CERN, holds that the two occurring within one institution adds no evidentiary force to the analogy; that notebook reads the standing representation of the trigger as engineering, with the rate budget non-epistemic, the detector neutral and discarded events noise, holds that it has no field for noncoverage, and states that whether the protocols would estimate the relevant unknown population adequately is not established.

Δthe delta

T against L(B) (representational): T composes the field's four application lines: diagnostics and fault prediction (F25, F26, F12, F42), classifier screening in Bayesian optimization (F17, F18), injector anomaly identification (F1, F2) and the errant-beam line (F42). T places CEBAF's trip identification 'before they disrupt operations'; the field reports the CEBAF system identifying which cavities were tripping and helping operators recover faulted cavities (F27, F30), with prediction before failure reported for SNS beam pulses (F12, F13) and faulty beams (F42). T's event and particle classification by CNNs and GNNs (flavors, tracks), its CNN/GNN architecture notes, and its decision trees, kNN and SVMs for errant-beam prediction are not in the field ledger's quotes (the SNS abstract names 'a set of common classification techniques', F13; the LANL loop names deep convolutional neural networks for control, F34). Available in the field and not composed: the reported accuracies (F7, F13, F14, F27) and the CEBAF test scale (F29); the limits (proof of principle F10, labeled data F8, simulation-only tuning F21, pruning as post-processing F24, further results under analysis F31); the autoencoder scored against baseline data (F6) and the model trained on normal operation (F43); LANL's adaptive diffusion virtual diagnostics, its XFEL demonstration and extrapolation (F36–F39) and the local-versus-global feedback argument (F35); downtime and sensor-cost motivation (F11, F41); the workflow framework (F45).

L(B) against L(B ∪ A) (the intervention): Admission adds a site the field's sources do not treat: learned event selection at collider triggers (AXOL1TL, CICADA, GELATO at the LHC; LHCb, CBM, Belle II in comparison), where the field's sources concern accelerator operations. It adds the distinction between signal-template agnosticism and model independence; Claim A (foreclosure structurally present in every classifier-mediated LHC trigger, held) and Claim B (collapse occurred, marked not established); the OAR, BAR and IAI with the no-bounds discipline and two retracted bounds; three protocols, per-stage retention maps and stated falsifiers; auditable foreclosure with its unaddressed residue; the irreversibility frontier and baseline capture; open-set assimilation for supervised (DNN/GNN) trigger classifiers; and a pre-registered battery on public data that replicates the Finke asymmetry, reports distillation tail loss, finds the score function alone sufficient to reverse ordering, falsifies one of its own sentences, and measures no deployed trigger. The field's anomaly detectors scored against a model of the normal (F6, F43) are the score class the archive's battery tests, though nothing admitted addresses cavity fault classification, failure or errant-beam prediction, or beam tuning. Carried as contested: whether collapse is operating (#932 W2 against the synthesis and the later papers); the inversion thesis and authority without facility (#935) against inseparability and acknowledged local self-knowledge (#934, with #931 §1.2 and #1449 §2); Zenodo and the trigger as one architecture (#935) or one topology (#934), the shared institution adding no evidentiary force (#1611); a rule against rendering agnosticism as independence (#1436) against declining to police the term (#1454). The field's classifiers are kept whole and first; with the archive they gain the trigger, its metrology and its graded claims.

Kthe prospective kernel sealed only at freeze11 entries

K1Q931-01 #931 (t: #1449, #1454)missing distinction

claim: Q931-01
source: #931 (t: #1449, #1454)
M_src: interpretation
sense: agnosticism does not sustain independence
qualifiers carried: Q931-02, Q931-03
f (source's own falsifiers): negligible BAR for all deployed systems against all held-out families (Q931-18)
contrast: missing distinction

K2Q1436-01 #1436 (t: #1449)missing distinction

claim: Q1436-01
source: #1436 (t: #1449)
M_src: stipulation
sense: interface property, distributional property
qualifiers carried: Q1436-02, Q1449-01
f (source's own falsifiers): —
contrast: missing distinction

K3Q932-02 #932 (t: #931, #1611)missing category

claim: Q932-02
source: #932 (t: #931, #1611)
M_src: interpretation
sense: Claim A: foreclosure structural
qualifiers carried: Q932-04, Q931-04
f (source's own falsifiers): the listed measurements would show the structural-feature claim overstated (Q932-11)
contrast: missing category

K4Q932-03 #932 against #932 W2contested

claim: Q932-03
source: #932 against #932 W2
M_src: hypothesis (contested in the pool)
sense: Claim B: collapse not established
qualifiers carried: Q932-15, Q932-16, Q931-05, Q1450-03, Q1611-03
f (source's own falsifiers): —
contrast: contested

K5Q931-10 #931 (t: #932, #1436)missing category

claim: Q931-10
source: #931 (t: #932, #1436)
M_src: stipulation
sense: OAR as an operational assimilation rate
qualifiers carried: Q931-11, Q931-12, Q932-09, Q931-19
f (source's own falsifiers): —
contrast: missing category

K6Q933-01 #933missing category

claim: Q933-01
source: #933
M_src: stipulation
sense: auditable foreclosure
qualifiers carried: Q933-02, Q933-04, Q933-05
f (source's own falsifiers): —
contrast: missing category

K7Q1456-01 #1456 (t: #1449, #1455)missing finding

claim: Q1456-01
source: #1456 (t: #1449, #1455)
M_src: documented
sense: the score function reverses ordering
qualifiers carried: Q1456-02, Q1456-03, Q1455-05, Q1449-04
f (source's own falsifiers): a converged normalized autoencoder at 0.5 would bound the inversion claims (Q1456-08)
contrast: missing finding

K8Q1452-01 #1452 (t: #1450)missing category

claim: Q1452-01
source: #1452 (t: #1450)
M_src: stipulation
sense: the irreversibility frontier
qualifiers carried: Q1452-02, Q1452-09, Q1450-04
f (source's own falsifiers): —
contrast: missing category

K9Q1453-01 #1453 (t: #1558)missing relation

claim: Q1453-01
source: #1453 (t: #1558)
M_src: interpretation
sense: a control group for the trigger
qualifiers carried: Q1558-01, Q1453-02, Q1558-02
f (source's own falsifiers): —
contrast: missing relation

K10Q935-01 #935 against #934contested

claim: Q935-01
source: #935 against #934
M_src: interpretation (contested in the pool)
sense: HEP's operational core as ML
qualifiers carried: Q934-01, Q934-02, Q935-02, Q935-06
f (source's own falsifiers): —
contrast: contested

K11Q1611-02 #1611missing distinction

claim: Q1611-02
source: #1611
M_src: interpretation
sense: no field for noncoverage
qualifiers carried: Q1611-01, Q1611-03, Q1611-04
f (source's own falsifiers): —
contrast: missing distinction

row json archive ledger selection readings field ledgerextraction unauditedD/R/O traversal

Addresses and transcripts1 address

entity · type C — a conventional reading holds it (public entity)1 address

Machine learning classifiers in particle accelerators

at the address · 2026-10-07

machine learning classifiers particle accelerators

Google AI Overviewsigned out, incognitoarchive absent7 source cards

2026-10-07 · transcriptOBS-d0d05d54a1dc

machine learning classifiers particle accelerators
Machine learning classifiers in particle accelerators help automate diagnostics, screen beam conditions, and predict hardware faults in real time. [[1], [2], [3]]
Key Applications

* Fault Detection and Diagnostics: Classifiers monitor cavity and beam behaviors, identifying specific failure modes or "trips" (such as at Jefferson Lab's CEBAF) before they disrupt operations. [1]
* Beam Screening and Optimization: Pre-trained classifiers (like ResNet50 architectures) screen latent space samples in Bayesian optimization loops to ensure adjustments remain physically valid. [1]
* Anomaly Identification: Machine learning models analyze process variables from injector complexes to flag single-parameter anomalies and trace poor extraction efficiencies. [1]
* Event and Particle Classification: Advanced architectures like Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs) process complex detector signals to categorize particle flavors and tracks. [1]

Common Classifier Architectures

* Convolutional Neural Networks (CNNs): Used for spatial and image-like diagnostic readouts or pixel maps of events.
* Graph Neural Networks (GNNs): Applied to irregular detector geometries where sensors do not fit standard grid layouts.
* Standard Binary/Multiclass Classifiers: Decision trees, k-Nearest Neighbors (kNN), and support vector machines applied to sensor streams for errant beam prediction. [[1], [2]]

If you'd like, let me know:

* Are you focusing on beam control/tuning or particle collision event reconstruction?
* Do you need a specific algorithm comparison?

jacow.org
Machine Learning for Anomaly Detection and Classification in Particle Accelerators
We explore the possibility of using a Machine Learning (ML) algorithm to identify the source of occasional poor performance of the Particle Accumulator Ring (PA...
ScienceDirect.com
Predicting particle accelerator failures using binary classifiers
Introduction The area has been growing and advancing since the 1960s [32], [33] and seen increase in pace the last 20 years due to advances in methods [34], [35...
arXiv.org
Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv
Classifier-pruned Bayesian optimizer (C-BO): A Bayesian optimization module that sequentially explores the latent space to maximize the negative total beam loss...
Jefferson Lab
Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab
Particle accelerators are huge, complex machines. Scientists and engineers have designed and built a novel machine learning system to use with the Continuous El...
Los Alamos National Laboratory (.gov)
AI algorithms used to tune particle accelerators | LANL
Core Innovation: Los Alamos and Lawrence Berkeley National Laboratories developed a machine learning technique using adaptive generative AI diffusion (condition...
Oak Ridge National Laboratory (ORNL) (.gov)
A machine learning approach for particle accelerator errant beam prediction ...
Abstract Particle accelerators are extremely complex systems that are expected to operate on high availability. Predicting impending failures only by utilizing ...
YouTube·Greg Bronevetsky
52m
Machine Learning in High Energy Physics

Source cards

1. jacow.org — Machine Learning for Anomaly Detection and Classification in Particle Accelerators
2. ScienceDirect.com — Predicting particle accelerator failures using binary classifiers
3. arXiv.org — Classifier-pruned Bayesian optimization for particle accelerator tuning - arXiv
4. Jefferson Lab — Machine Learning System Improves Accelerator Diagnostics - Jefferson Lab
5. Los Alamos National Laboratory (.gov) — AI algorithms used to tune particle accelerators | LANL
6. Oak Ridge National Laboratory (ORNL) (.gov) — A machine learning approach for particle accelerator errant beam prediction ...
7. YouTube·Greg Bronevetsky — Machine Learning in High Energy Physics

OBS-d0d05d54a1dc · CAPTURE-TIME VERBATIM RECORD — composition and sources as pasted · READ IN FULL 2026-10-07 · sha256 157683cdd60f04d3…

Surface: Surface not stated; the standing default applies (operator, 2026-10-01: 'assume they do [start in overview], i will specify if they begin in ai mode'). The 'AI Mode Conversation' header is the expanded Overview's residue.

Archive: none found: no card, link or name of the archive

Archive bearing — D/R/O traversalread

Candidates are found by string; admission is by reading.

traversalfirst reading, 2026-10-07

machine learning classifiers particle accelerators

D 22 deposits · 342 sentencesR 11 · 20O 14 · 79 direct, 0 classhop —read: 16 admitted, 5 not

admitted, by lineage8 lineages

SEI classifier-foreclosure family (OAR protocol, synthesis, architecture)3 · from #931

#931, #932, #933

The Endogenous Sophon (v0.2 over-corrected, v0.3 restored)2 · from #934

#934, #935

SIGAGNOSTIC packet1 · from #1436

#1436

Inversion battery (v0.1, v0.2, v0.3)3 · from #1449

#1449, #1455, #1456

Iceberg Document1 · from #1450

#1450

Accelerator selection-metrology suite (Irreversibility Frontier, BCA v1.0/v2.0, ACRB)4 · from #1452

#1452, #1453, #1454, #1558

Notice to the laborers of CERN (measurement section)1 · from #1427

#1427

EA-NEGONT-01 third worked case1 · from #1611

#1611

The 22 D deposits of the string pass reduce by reading to one programme in four parts: the June 2026 family (#931–#935), the August disambiguation packet and measurement battery (#1436, #1449, #1455, #1456), the August–September selection-metrology suite (#1450, #1452–#1454, #1558), and two deposits that restate or apply it (#1427, #1611).

The archive's line is narrower than the field's entity. The field ranges over fault diagnostics, beam tuning, errant-beam prediction and anomaly identification at accelerator facilities (CEBAF, SNS, APS, LANSCE); the archive's claims are almost all about learned event selection at the trigger (CMS, ATLAS, LHCb, Belle II, CBM). Nothing admitted is about RF-cavity fault classification, beam tuning or errant-beam prediction.

Contested within the pool: (a) whether collapse is occurring: the W2 witness inside #932 says it is here and operating; the synthesis (#932), the operative paper (#931), W3, #1450, #1452, #1454, #1558 and #1611 mark it unmeasured or not established. (b) The inversion and its diagnosis: #935 restores 'become a machine learning discipline' and 'authority without facility'; #934 states 'inseparable from' and 'authority without integrated self-audit', acknowledging local self-knowledge in the literature, as #931 §1.2 and #1449 §2 also do. (c) Zenodo and the trigger: 'the same architecture at different budgets' (#935) against 'not the same institution … the same foreclosure topology' (#934); #1611 adds that the analogy carries no evidentiary force from the shared institution. (d) The term: #1436 sets a rule that agnosticism must not be rendered as model independence; #1454 says it should not attempt to police the term.

Self-correction on the record: #1455's sentence 'architecture changes the orientation of the blind spot' is falsified by #1456's score-function ablation; both are ledgered.

The transcript's header reads 'AI Mode Conversation'; the register records the surface as Google AI Overview (expanded) under the operator's standing default.

candidate files configuration

Lee Sharks · Crimson Hexagonal Archive · register v1.13, 2026-10-07. Built by scripts/build_non.py from the files linked above; it writes nothing back. Working, not frozen.