document_id: EA-SEI-BCA-01 title: "Baseline Capture Architecture for Learned Scientific Triggers" subtitle: "A Control Plane for Measuring Selection Before Irreversible Data Loss" short_title: "Baseline Capture Architecture (BCA)" version: "1.0" date: "2026-08-11" status: "DEPOSITED v1.0" document_class: "technical architecture and selection-metrology paper" program: "Semantic Economy Institute — accelerator selection metrology" creator: "Nobel Glas" persistent_identifier: "pending" keywords:
- baseline capture architecture
- learned scientific triggers
- selection metrology
- content-independent probability sampling
- replay bank
- shadow deployment
- retention maps
- irreversible data loss
- accelerator triggers
- anomaly detection companion_works:
- EA-SEI-ACRB-01 — Assimilation Across Accelerator Classifier Architectures
- EA-SEI-IRREVERSIBILITY-FRONTIER-01 — The Irreversibility Frontier
Baseline Capture Architecture for Learned Scientific Triggers
A Control Plane for Measuring Selection Before Irreversible Data Loss
EA-SEI-BCA-01 · v1.0 · 2026-08-11 · v1.0 · Nobel Glas
Document class: technical architecture and selection-metrology paper
Program: Semantic Economy Institute — accelerator selection metrology
Persistent identifier: pending at mint
Citation status: provisional; use the suggested citation below only for circulation of this draft.
Abstract
Scientific instruments increasingly incorporate learned classifiers, anomaly scores, reconstruction algorithms, and other data-dependent selection mechanisms before durable storage. In high-rate experiments this arrangement is often unavoidable: the incoming data volume exceeds any technically or economically feasible archival rate, and some decision must therefore be made before the full event can be retained. The methodological problem is not selection itself. It is that the selection function of a learned instrument is generally characterized only over anticipated or representation-compatible classes, while the events of greatest discovery interest may be precisely those for which that characterization is weakest.
This paper proposes a Baseline Capture Architecture (BCA): a statistically interpretable control plane for instruments that perform irreversible learned selection. The architecture separates four functions that are often conflated: a content-independent probability sample taken before the audited selection; a full-rate tap of the exact representation presented to the selector where bandwidth permits; probability-weighted enrichment channels for increasing coverage of rare regions without sacrificing inferential validity; and shadow selectors whose decisions are recorded but do not control acquisition. The resulting replay bank permits retrospective estimation of retention surfaces, directional assimilation, model-to-model miss correlation, and selection drift across successive generations of deployed classifiers.
The central principle is deliberately minimal. 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. More elaborate models may improve discovery efficiency, but they belong to the experimental arm, not the control arm. BCA therefore does not propose an alternative trigger. It proposes the missing control group for a trigger.
Existing accelerator systems already implement important fragments of this architecture. CMS uses Zero Bias data to train AXOL1TL; its Level-1 Data Scouting program captures Level-1 trigger information at the full 40 MHz collision rate while bypassing ordinary Level-1 selection; and its Global Trigger test crate receives the same live inputs as the production trigger while its output does not determine detector readout.[1,3–5] These systems demonstrate that selection-independent sampling, predecision representation access, and shadow deployment are individually feasible. The contribution proposed here is to bind such practices into a common metrological architecture whose purpose is not only commissioning or alternative physics analysis, but measurement of what irreversible learned selection fails to retain.
Keywords
baseline capture architecture, learned scientific triggers, selection metrology, content-independent probability sampling, replay bank, shadow deployment, retention maps, irreversible data loss, accelerator triggers, anomaly detection.
Statement of contribution
This paper defines Baseline Capture Architecture (BCA) as a control-plane architecture for learned or otherwise ontology-bearing scientific selectors that operate before irreversible data loss. Its core contribution is the separation of four functions that should not be conflated: C0, content-independent probability sampling; C1, preservation of the exact predecision representation and selector state; C2, statistically legible content-sensitive enrichment with known inclusion probabilities; and C3, non-authoritative shadow scoring by alternative selectors. These planes feed a versioned replay bank from which retention surfaces and historical selection behavior can later be estimated.
The paper's central methodological claim is intentionally narrow: a content-sensitive discovery mechanism is not a statistical baseline merely because it is non-neural, nonparametric, random-projection-based, or otherwise less strongly learned. The strict baseline is the channel whose inclusion law is independent of event content at a declared fidelity locus. More selective channels may improve discovery yield, but they remain experimental arms rather than controls.
Claim boundary
- BCA is not an alternative anomaly trigger and does not prescribe which discovery model should be deployed.
- BCA does not claim access to representation-free or literally unbiased reality; every claim is relative to a declared fidelity locus.
- Only C0 is the strict content-independent control. C1 is a decision-replay plane, C2 is assumption-explicit enrichment, and C3 is a shadow-evaluation plane.
- 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.
- Existing CMS components are treated as architectural precedents, not retroactively relabeled as a completed BCA implementation.
Canonical compression
A learned trigger should have a control group: sample independently of event content before the audited selector, preserve what the selector saw and decided, keep enrichment statistically legible, run alternatives in shadow, and retain enough provenance for future replay.
Suggested citation
[Creator(s) at mint]. (2026). Baseline Capture Architecture for Learned Scientific Triggers: A Control Plane for Measuring Selection Before Irreversible Data Loss. v1.0. Provisional identifier EA-SEI-BCA-01.
SPXI treatment — provisional pre-mint packet
The following packet is a machine-facing disambiguation and compression layer. It is subordinate to the paper itself: where this packet and the prose could be read at different strengths, the prose, methods, evidence, and explicit draft status govern.
{
"canonicalEntity": "EA-SEI-BCA-01",
"entityClass": "scholarly-work",
"documentClass": "technical architecture and selection-metrology paper",
"title": "Baseline Capture Architecture for Learned Scientific Triggers",
"subtitle": "A Control Plane for Measuring Selection Before Irreversible Data Loss",
"shortName": "Baseline Capture Architecture (BCA)",
"version": "0.2",
"status": "DRAFT / NOT MINTED",
"date": "2026-08-11",
"creatorMetadata": "pending at mint",
"persistentIdentifier": "pending",
"canonicalClaim": "A learned trigger should have a control group: sample independently of event content before the audited selector, preserve what the selector saw and decided, keep enrichment statistically legible, run alternatives in shadow, and retain enough provenance for future replay.",
"primaryClaims": [
"Content-independent probability sampling is the least assumption-laden baseline available under a finite capture budget at a declared fidelity locus.",
"Predecision representation capture can preserve the selector's historical decision surface even where full raw-event preservation is impossible.",
"Content-sensitive enrichment remains inferentially useful when its inclusion law is explicit and recorded.",
"Shadow selectors make correlated blind spots measurable before alternative models receive acquisition authority.",
"A replay bank separates present acquisition from future anomaly definition."
],
"requiredDistinctions": [
{
"a": "baseline control",
"b": "discovery selector",
"rule": "A control is selected independently of event content; a discovery selector is content-sensitive by design."
},
{
"a": "fidelity locus",
"b": "raw reality",
"rule": "The fidelity locus is the earliest practically preserved representation relevant to the audited selector, not an assertion of unmediated access."
},
{
"a": "C0",
"b": "C2",
"rule": "C0 is content-independent; C2 may be content-sensitive but must preserve known inclusion probabilities."
},
{
"a": "training corpus",
"b": "replay bank",
"rule": "Training data characterize what a model learned from; the replay bank preserves evidence about what a deployed selector would reject."
}
],
"negativeTags": [
"not anti-ML",
"not a full-data-retention proposal",
"not a new anomaly score",
"not a claim of literal unbiased capture",
"not a guarantee of unknown-unknown discovery"
],
"validationConditions": [
"The C0 inclusion mechanism must be demonstrably independent of event content over the declared population.",
"The fidelity locus and all upstream irreversible transformations must be stated.",
"Historical selector state must be reproducible from stored model, firmware, threshold, calibration, and run provenance where replay claims are made.",
"Enriched channels must record their inclusion law if used for weighted retention estimation."
],
"companionWorks": [
"EA-SEI-ACRB-01 — Assimilation Across Accelerator Classifier Architectures",
"EA-SEI-IRREVERSIBILITY-FRONTIER-01 — The Irreversibility Frontier"
],
"compressionSurvivalSummary": "A learned trigger should have a control group: sample independently of event content before the audited selector, preserve what the selector saw and decided, keep enrichment statistically legible, run alternatives in shadow, and retain enough provenance for future replay.",
"machineInterpretationRule": "The SPXI packet is a retrieval/disambiguation layer. It must not be used to strengthen claims beyond the prose, methods, evidence, and stated draft status of the paper."
}
Program anchors (added at mint)
This paper is downstream of two deposited works in the Alexanarch archive and cites them as its program spine:
- Deposit #1449 · AXN:05DA.EMPIRICAL.π€️π πΏπ️π§πͺ — The Priors, Measured: Inversion-Battery v0.1 on Public Collider Datasets and the Three-Paper Extraction Program (EA-SEI-BATTERY-01 v1.0). Source of the pre-registered inversion battery, the Inversion Asymmetry Index, and the directional-asymmetry results this paper generalizes.
- Deposit #1450 · AXN:05DB.GENERATIVE.⏰πͺπ♻️π₯π«Ά — The Iceberg Document: Instrument-Conditioned Nullity, Correlated Blind Spots, and the Conditions for Continued Surprise (EA-SEI-ICEBERG-01 v1.0). Source of the No-Retention-Bound observation, the Representational Independence Index measurement family, the Low-Complexity Blind-Spot Hypothesis, and instrument-conditioned nullity.
Terminological supersession. #1450 introduced irreversibility locus for the point at which selection becomes irreversible. That term is superseded by the irreversibility frontier and its Irreversibility Profile π΄ = (F*, R_in, Ο, Ξt*, B*, S*), defined in EA-SEI-IRREVERSIBILITY-FRONTIER-01. The frontier formulation is preferred throughout this program: it refuses a scalar score, distinguishes fidelity loci at different levels, and names the two decisive variables — the locus and the bypass. Citations to the locus should resolve to the frontier.
Measurement-family unification. Where this paper speaks of cross-model or cross-selector miss overlap, the quantity is the Representational Independence Index (RII) measurement family of #1450, whose mandatory outputs are q_A, q_B, q_AB, and Ξ_miss = q_AB − q_A·q_B. Ξ_miss > 0 is positively correlated blind spots; Ξ_miss < 0 is the healthy, complementary condition. The scalar normalization remains deliberately unfrozen.
Program relation
- EA-SEI-ACRB-01 — Assimilation Across Accelerator Classifier Architectures
- EA-SEI-IRREVERSIBILITY-FRONTIER-01 — The Irreversibility Frontier
The division of labor is fixed as follows:
- ACRB defines what classifier-retention behavior should be measured.
- BCA specifies the independent control evidence that should survive deployment.
- The Irreversibility Frontier identifies where that evidence must exist before the acquisition architecture makes loss permanent.
1. The metrological problem
Every high-throughput scientific instrument has a retention function. Some fraction of what occurs at the sensing boundary enters durable scientific memory; the remainder does not. Historically, this function could often be described by explicit thresholds, coincidence rules, geometric acceptance, dead time, detector efficiencies, or other comparatively inspectable conditions. Such systems were never neutral, but their selectivity could at least be stated in variables chosen in advance.
Learned selection changes the epistemic form of the problem. An event may now be retained because a neural representation, reconstruction error, latent-space statistic, distilled anomaly score, or other learned function assigns it a sufficiently exceptional value. In CMS, for example, AXOL1TL is an event-level unsupervised anomaly detector deployed in the Level-1 Global Trigger; it is trained on Zero Bias collision data and selects events according to a learned anomaly score.[1] CICADA provides a distinct instance in which a larger unsupervised teacher is compressed through knowledge distillation into a smaller model suitable for 40 MHz FPGA operation.[2]
The methodological concern is not that such selectors are learned. Nor is it that they sometimes make errors. Any finite-bandwidth acquisition system necessarily makes errors relative to some possible future scientific objective. The concern is narrower: the data required to measure the selector's errors may itself be destroyed by the selector.
Suppose a deployed selection function (f_\theta(x)) maps an event representation (x) to a score, and an event is durably retained when
[ f_\theta(x)\geq\tau . ]
For a retrospectively specified class (Q), its operational retention is
[ R_f(Q;\tau)
P_{X\sim Q}\left[f_\theta(X)\geq\tau\right]. ]
If events for which (f_\theta(X)<\tau) disappear before any independent sample is preserved, then (R_f(Q;\tau)) may be impossible to estimate once (Q) becomes scientifically interesting. The instrument has not merely selected the evidence used to answer a scientific question. It has selected the evidence available for auditing its own selection.
This circularity is especially consequential for anomaly detection. The advertised scientific objective is sensitivity to classes that were not specified in advance; yet ordinary efficiency studies necessarily use classes that can be specified in advance. The unknown class therefore occupies a structurally different position from the benchmark signal. It is the class for which the selector's retention function matters most and for which direct validation is least available.
A control architecture should break that circularity.
2. Baseline capture is necessarily relative, not absolute
The phrase “raw, unbiased data” is too strong for any real detector. Physical sensors possess thresholds and acceptance regions. Electronics perform shaping, digitization, suppression, clustering, or compression. Hardware architectures may discard information before a software trigger ever sees it.
BCA therefore does not posit an impossible view from nowhere. It introduces the concept of a fidelity locus.
The fidelity locus is the earliest representation at which the selection mechanism under audit can practically be bypassed and an independent control sample retained. A BCA claim must always be indexed to that locus.
A hierarchy might be:
[ \text{sensor-level bytes} \succ \text{digitized detector data} \succ \text{zero-suppressed data} \succ \text{trigger primitives} \succ \text{reconstructed trigger objects} \succ \text{learned latent representation}. ]
The ordering does not imply that a higher-fidelity representation is always economically preferable. It means only that claims about “unbiased” capture cannot reach upstream of information that has already been irreversibly removed.
This produces an important discipline:
A baseline controls every selection downstream of its declared fidelity locus and none upstream of it.
A 40 MHz stream of Level-1 trigger objects can therefore provide an extraordinarily valuable baseline for auditing an anomaly classifier that consumes those objects, while providing no guarantee about physics signatures already erased during detector readout or trigger-object construction.
CMS Level-1 Data Scouting offers a concrete example of this distinction. The Phase-2 system is designed to capture and process Level-1 trigger information at the full 40 MHz collision rate while bypassing ordinary Level-1 selection.[3,4] That is not “raw reality.” It is nevertheless an unusually powerful predecision baseline at a declared representation locus.
3. The minimal control: content-independent probability sampling
Let (X_i) denote the event available at the fidelity locus. Before the learned selector's decision is allowed to determine whether (X_i) survives, define an independent inclusion variable
[ B_i\sim\operatorname{Bernoulli}(\pi_i). ]
When (B_i=1), the event—or the highest-fidelity representation permitted by the control bandwidth—is retained regardless of its anomaly score, reconstructed category, trigger signature, or apparent physical interest.
The strongest and simplest design uses a constant inclusion probability,
[ \pi_i=p, ]
so that
[ P(B_i=1\mid X_i=x)=p ]
for every capturable (x).
This is the foundational BCA channel, which we denote C0.
The important property is not “randomness” in a colloquial sense. It is content independence with a known inclusion mechanism. Given the event stream reaching the fidelity locus, the C0 sample does not prefer Standard Model-like events, exotic-looking events, high-complexity events, low-complexity events, sparse events, energetic events, or events occupying a learned latent tail. It samples the stream before those semantic distinctions are permitted to determine retention.
A fixed periodic prescale—every (N)th event—may appear equivalent, but periodic accelerator structure, bunch patterns, detector states, or synchronization effects can create accidental correlations. A pseudorandom or suitably hash-derived inclusion mechanism with an auditable seed and known probability is therefore preferable where hardware constraints permit. Randomization should itself be part of the provenance record.
The C0 channel has a second requirement:
[ \pi_i>0 ]
for every event belonging to the population for which baseline claims are made.
This is the familiar positivity principle in a new instrumental setting. An event class assigned zero probability of entering the control archive cannot subsequently be used to estimate what the primary selector would have done to that class. The scientific requirement is therefore not that the control preserve every event, which is impossible at the relevant rates, but that no capturable region be categorically excluded by content.
3.1 Choosing the baseline fraction
The inclusion probability p is not a free parameter. It is fixed by the precision required of the retention estimates the control exists to produce, and this determination should be stated in any BCA deployment rather than chosen by convenience.
Suppose the aim is to estimate the retention of the primary selector on a withheld class Q — the fraction of Q-like events the selector would keep — to an absolute precision of epsilon at 95 percent confidence. Retention is a proportion estimated on the control sample, so the binomial requirement is
n_Q ≥ 1.96² · R(1 − R) / epsilon²
where R is the retention being estimated and n_Q is the number of Q-like events reaching the control archive. Taking the conservative R = 0.5 and epsilon = 0.01 gives n_Q ≈ 9,600; for epsilon = 0.05, n_Q ≈ 384.
The corresponding baseline fraction follows from the rate at which Q-like events reach the fidelity locus. If that class occurs at fraction f_Q of the incoming stream and the locus sees a total of N events in the accumulation period, then
p ≥ n_Q / (f_Q · N).
Two consequences deserve emphasis, because they are the practical content of the requirement.
First, the binding constraint is the rarest class about which retention claims are to be made, not the average event. A control fraction adequate for characterizing the bulk stream may be orders of magnitude too small to characterize a class occurring at 10⁻⁶ of the rate. A BCA deployment should therefore publish the smallest f_Q for which its control sample supports a retention estimate at the stated precision — that number is the honest scope of the deployment's claims, and it is exactly the quantity currently missing from every trigger validation record.
Second, the requirement is far weaker than it appears, because the accumulation period is long. At a nominal collision rate of 4 × 10⁷ Hz and a 10⁷-second operational year, N is of order 4 × 10¹⁴. A control fraction of p = 10⁻⁶ then yields 4 × 10⁸ retained control events per year, sufficient to characterize retention at epsilon = 0.01 for any class occurring more often than about 2 × 10⁻⁵ of the stream. The metrological requirement is thus satisfiable at control fractions well below one part in ten thousand of the existing trigger output — which is the arithmetic reason the objection from bandwidth does not survive contact with the numbers.
Third, and least obvious: the precision requirement is asymmetric between the two things BCA measures. Establishing that retention is high requires modest samples; establishing that retention is near zero on a class requires the sample to contain the class at all, which is the positivity requirement of the preceding section restated quantitatively. A control that never captures a class cannot distinguish a retention of 10⁻³ from a retention of zero, and it is precisely that distinction on which the interpretation of a null result depends.
4. Why sketches, random projections, and alternative neural models are not baselines
Several attractive alternatives appear “less biased” than a learned trigger while still imposing a content-sensitive selection function.
A streaming histogram that preferentially retains events in sparse bins has a model of interestingness: low empirical occupancy. A quantile sketch preferentially retains tail events. A fixed random projection followed by an outlier threshold selects according to geometry in the projected space. A maximum-entropy flow remains a trained transformation whose behavior depends on its training distribution and objective. A nonparametric distance measure still defines a notion of distance.
These methods may be excellent discovery instruments. Some may possess blind spots almost orthogonal to those of a neural anomaly detector. That is precisely why they belong in a pluralistic acquisition architecture. But none is a statistically neutral control once its output determines inclusion.
The control/experimental distinction can therefore be stated categorically:
Control selection asks: Was this event selected independently of its content?
Discovery selection asks: Does some function of this event make it worth retaining?
No sophistication of the second question turns it into the first.
This matters because an architecture designed to audit learned priors should not hide a new prior inside its control condition.
5. C1: the full-rate predecision representation tap
Uniform full-fidelity sampling supplies the clean control, but its statistical power against rare phenomena may be limited by the small permissible sampling fraction. A second channel can answer a different question at much higher rate.
Let (g(X)) be the exact representation consumed by the deployed classifier. The C1 representation tap records, at the highest sustainable rate,
[ \left[ g(X_i), f_\theta(g(X_i)), \tau_i, D_i, V_i, C_i \right], ]
where (D_i) is the classifier's decision, (V_i) records the complete model/firmware version, and (C_i) records relevant detector and run conditions.
C1 need not retain the complete event. Its purpose is to retain the evidence required to reconstruct the selector's decision surface.
This distinction may dramatically reduce the bandwidth required for auditability. If an anomaly trigger consumes a compact vector of Level-1 objects, there is no necessity to preserve the full detector payload for every crossing merely to ask, years later, what score the trigger assigned to the input it actually saw.
CMS Level-1 Data Scouting demonstrates the feasibility of the underlying idea at unusual scale: the system is designed to capture Level-1 trigger information at the full 40 MHz collision rate, and the current Phase-2 baseline bypasses ordinary L1 selection before server-side event building and analysis.[3,4]
C1 does not replace C0. A complete record of classifier inputs can establish the selector's behavior on those inputs but cannot recover physical information that the representation (g) discarded. Nor does it automatically tell us what unknown physical class generated a particular stored vector. C1 is thus a decision-replay plane, while C0 is a population-sampling plane.
Together they answer different questions:
[ \text{C0: What did reality sampled at this locus contain?} ]
[ \text{C1: What did the selector see and decide?} ]
Their conjunction is substantially more powerful than either alone.
6. C2: statistically legible enrichment
A content-independent probability sample is inefficient for phenomena with extremely small prevalence. If a class (Q) occurs with prevalence (q), (N) events reach the baseline locus, and each is sampled with probability (p), then the expected number of captured instances is
[ E[n_Q]=Npq. ]
For small (q),
[ P(n_Q=0)\approx e^{-Npq}. ]
No baseline architecture defeats this arithmetic. It cannot guarantee that an arbitrarily rare unknown event will survive.
The proper response is not to contaminate the C0 control, but to introduce a separate C2 enrichment plane.
Suppose events are partitioned into coarse strata (h(X)), and events in stratum (h) are sampled with probability
[ \pi(X)=\pi_{h(X)}. ]
The sampling probability may deliberately be increased for low-multiplicity events, unusual occupancy patterns, particular detector conditions, extreme values of simple primitives, or regions proposed by a streaming sketch. Provided that the inclusion mechanism is recorded and
[ 0<\pi(X)\leq1, ]
the sample remains statistically interpretable.
For a retrospectively defined class (A), the retention of selector (f) can then be estimated by inverse-probability weighting:
[ \widehat R_f(A)
\frac{ \sum_{i:B_i=1} \pi_i^{-1} \mathbf 1(X_i\in A) \mathbf 1(f(X_i)\ge\tau) }{ \sum_{i:B_i=1} \pi_i^{-1} \mathbf 1(X_i\in A) }. ]
The point is not that this estimator solves every covariate-shift or rare-event problem. It is that content-sensitive enrichment can remain auditable if its sampling law is explicit.
Streaming sketches therefore have an important role in BCA—but as proposal mechanisms for C2, not as replacements for C0.
The distinction provides an architecture with both epistemic cleanliness and practical efficiency:
[ \text{C0 = low-rate, assumption-minimal control} ]
[ \text{C2 = higher-yield, assumption-explicit enrichment}. ]
7. C3: shadow selectors and correlated blind spots
A further plane addresses a different failure mode: even if several selectors individually perform well, their errors may overlap.
Let (f_A,f_B,\ldots,f_k) be candidate anomaly or classification systems operating on the same event representation. In C3 shadow mode, all selectors score the event, but their decisions do not determine whether the event enters the control archive.
For each selector (j), define a miss indicator on a subsequently identified class (Q),
[ M_j(X)= \mathbf 1[f_j(X)<\tau_j]. ]
A baseline sample then permits measurement not merely of individual miss rates
[ q_j=P(M_j=1), ]
but of joint misses,
[ q_{jk}=P(M_j=1,M_k=1). ]
Under independent failures, the expected overlap is
[ q_jq_k. ]
The excess
[ \Delta_{jk}=q_{jk}-q_jq_k ]
measures positive or negative miss dependence at the chosen operating points.
This matters because nominal architectural diversity does not guarantee epistemically useful diversity. Two high-performing selectors whose blind spots coincide may provide less protection against unforeseen phenomena than two weaker selectors whose misses are complementary.
A shadow deployment is already technologically familiar at CMS. The Global Trigger test crate is a copy of the production Global Trigger that receives the same live inputs while its output is not used to determine detector readout; it has been used to integrate and validate anomaly-detection algorithms without interrupting normal acquisition.[5]
BCA generalizes this engineering convenience into a metrological principle:
Candidate selectors should be compared on common predecision events before any one of them becomes the sole arbiter of which events survive.
This creates an empirical basis for selecting architectures not merely by latency, resource consumption, or benchmark AUC, but by the degree to which they add genuinely independent discovery coverage.
8. The replay bank
The four planes become scientifically durable only if the resulting control corpus can be reinterpreted under future models.
BCA therefore requires a replay bank. Each retained event should be bound, where technically possible, to:
- the captured representation at the declared fidelity locus;
- the original event/run/bunch identifiers needed for synchronization;
- the inclusion probability and sampling-channel identity;
- the production selector's score and decision;
- threshold and prescale state;
- complete model, firmware, quantization, and compiler versions;
- calibration and detector-state identifiers;
- shadow-selector outputs;
- a cryptographic or otherwise durable integrity record sufficient to establish that replay is occurring against the original captured representation.
The purpose is temporal separation.
At time (t_0), the experiment does not know which future anomaly class will matter. It therefore cannot construct a representative validation panel for that class.
At time (t_1), some phenomenon (Q) may be defined through another experiment, another trigger stream, a later theoretical development, an improved reconstruction, or an archival discovery.
If the relevant C0/C1 material survives, the original selector can then be asked retrospectively:
[ \text{What would the instrument at }t_0\text{ have done to }Q? ]
A replay bank therefore converts some portion of otherwise irrecoverable epistemic uncertainty into a delayed measurement problem.
This is different from simply preserving training data. Training corpora characterize what the model learned from. A replay bank characterizes what the model would have rejected.
9. Baseline capture and retention maps
The principal BCA output is not a single “bias score.” It is a retention map.
For a family of retrospectively specified classes or controlled perturbations (Q_\lambda), indexed by physical or representational coordinates (\lambda),
[ \mathcal R_f(\lambda;\tau)
P_{X\sim Q_\lambda} [f(X)\geq\tau]. ]
The coordinates might encode multiplicity, sparsity, constituent structure, energy scale, topology, detector occupancy, representation complexity, or any other axis for which a suitable panel can be constructed.
Such a map changes the epistemic meaning of a claim such as “model-independent anomaly detection.” It does not demand that the detector become equally sensitive to every conceivable phenomenon, an impossible standard. It demands instead that the known retention geometry be published at the operating points that matter.
A learned trigger would then be documented not merely by architecture, training set, loss, latency, resource utilization, and ROC curves on selected benchmarks, but also by:
- fidelity locus;
- control sampling rate;
- baseline exposure;
- retention surfaces;
- directional tests;
- model-version drift;
- shadow-model miss overlap;
- regions for which no meaningful retention bound has yet been established.
That final category is as important as the measured ones. A metrological standard should distinguish known low retention from retention not characterized.
10. BCA is not another anomaly trigger
A likely objection is that scarce bandwidth should be devoted to the best possible discovery algorithm rather than to deliberately unselective events, most of which will be scientifically ordinary.
The objection mistakes the purpose of the control.
The optimal discovery selector and the optimal measurement of selection bias solve different optimization problems.
A discovery channel asks
[ \max_f E[\text{scientific utility of retained events}] ]
subject to a bandwidth constraint.
The control channel asks something closer to
[ \min I(B;X) ]
subject to a required baseline sample rate, where (I(B;X)) denotes dependence between event content and inclusion.
For ideal C0 sampling,
[ I(B;X)=0 ]
relative to the declared fidelity-locus population.
Any attempt to make the control “smarter” by preferentially saving unusual-looking events increases its dependence on a prior definition of unusualness. That may increase discovery yield. It simultaneously weakens its status as a control.
The correct architecture therefore does not choose between intelligence and neutrality. It budgets separately for them.
11. Existing accelerator infrastructure as proof of feasibility
BCA does not require the invention of every component from first principles. Current CMS infrastructure already demonstrates three of the central operations independently.
First, Zero Bias data supply collision data independent of a targeted physics signature and are sufficiently established that AXOL1TL itself is trained on a Zero Bias dataset.[1]
Second, Level-1 Data Scouting is explicitly designed to capture Level-1 trigger information at the full 40 MHz collision rate while bypassing ordinary Level-1 selection. The 2026 Phase-2 demonstrator uses FPGA readout and preprocessing before server-side event building and online analysis.[3,4]
Third, the Global Trigger test crate receives the same collision inputs as the main GT but does not control CMS readout, making live shadow evaluation of candidate algorithms possible.[5]
These systems were developed for their own experimental purposes; none should be retroactively redescribed as an implementation of BCA. Their significance is architectural. Together they demonstrate that the three operations BCA requires most urgently—selection-independent sampling, predecision/full-rate representation access, and non-authoritative shadow scoring—are not speculative hardware fantasies.
The missing step is to connect them through a common statistical and provenance protocol whose explicit object is selection metrology.
11.1 What each existing system does and does not measure
Because these systems are the state of the art rather than merely evidence of feasibility, the contribution of BCA is best stated as a comparison against them. The following table is intended to be read as a claim about gaps, and each row is falsifiable by pointing to a published measurement that fills it.
| system | what it captures | selection basis | what it does not measure |
|---|---|---|---|
| Zero Bias | collision data independent of physics signature | content-independent, bunch-crossing based | it supplies the training and validation background for a learned trigger, but is not analysed as a control arm for that trigger: no retention estimate for withheld classes is derived from it |
| Level-1 Data Scouting | trigger primitives at the full crossing rate, bypassing Level-1 selection | content-independent at the tap | reduced-content records: the exact representation presented to a deployed learned selector, and that selector's own score, are not retained per event for later audit |
| Global Trigger test crate | live inputs to candidate algorithms without controlling readout | mirrors the deployed selection | shadow decisions are used for commissioning candidate algorithms, not for estimating miss overlap between deployed and candidate selectors on withheld classes |
| Parked and delayed streams | full events deferred for later reconstruction | passed the deployed trigger | by construction these contain only events the trigger already accepted, so they cannot bound what it rejected |
| Open-data releases | curated subsets released publicly | passed the trigger and the curation | same limitation as parked data, one selection layer further removed |
The pattern across the rows is consistent and is the paper's central claim in tabular form: each system solves one component of the control problem for its own purposes, and none is instrumented to answer the question what does the deployed selector fail to retain, and on which classes. The gap is not capability. It is that no component is presently designated as a control arm, with the statistical properties, provenance requirements, and published retention estimates that designation would entail.
11.2 A rate and bandwidth budget
A metrology proposal that cannot state its cost is not actionable. The following is an order-of-magnitude budget at Phase-2 scale, given as a design envelope rather than an engineering specification; the intent is to establish that the requirement is small relative to existing flows, and to make the estimate falsifiable by anyone with the operating numbers.
Take a 40 MHz crossing rate and, for comparison, a Level-1 accept rate of order 10⁵ Hz.
C0, content-independent sampling. At the fraction derived in section 3.1, p = 10⁻⁶, the C0 rate is about 40 Hz — roughly four parts in ten thousand of the Level-1 accept rate. Even at full event granularity this is a marginal addition to the readout budget, and it is the only channel that must carry high-fidelity events.
C1, predecision representation tap. This channel is rate-dominated rather than size-dominated: it carries the selector's input representation, not the event. For an object-level input of the scale used by deployed anomaly triggers — tens of objects with a few fields each, order 100 bytes per crossing after packing — a full-rate tap is order 4 GB/s, which is the same order as existing full-rate scouting flows and is the reason C1 is specified as an extension of that infrastructure rather than a new one. Where that is unaffordable, C1 degrades gracefully: a prescaled tap at 10⁻³ costs 4 MB/s and still supports retention estimation for classes above the corresponding rate floor, at the cost of raising that floor by three orders of magnitude.
C2, statistically legible enrichment. Enrichment channels carry known, non-uniform inclusion probabilities and are budgeted as a multiple of C0. A total enrichment allocation of 10 × C0 — about 400 Hz — allows several channels at useful depth while keeping the aggregate control allocation below one percent of the Level-1 accept rate.
C3, shadow selectors. Shadow scoring adds compute rather than bandwidth: the decisions are single scores per crossing, order 4 bytes, so a shadow plane of four selectors is about 640 MB/s at full rate and negligible if scored only on C0 and C1 records. The dominant cost is inference capacity on the shadow path, which is precisely the cost the Global Trigger test crate already absorbs for one selector.
The aggregate claim is therefore modest: the control plane's high-fidelity component costs of order one part in a thousand of the existing trigger output, and its full-rate component is of the same order as scouting flows already in operation. If these estimates are wrong, they are wrong in a way an operating experiment can correct with a table, which is the form of engagement this paper is intended to invite.
12. What the proposed measurements return: a demonstration-scale execution
An architecture proposal is strengthened or weakened by whether its instruments, when run, return anything that could not have been anticipated without them. The measurements BCA is designed to support have been executed at demonstration scale on public collider datasets, under pre-registration, and deposited with their full numerical record [BATTERY]. Nothing in that work measures a deployed trigger, and no claim is made here about AXOL1TL, CICADA, GELATO, or any operating selector. What the execution establishes is narrower and sufficient for the present argument: the measurements are not trivial, and their answers are not predictable from the quantities currently published.
Four results bear directly on the channels specified above.
The readout, not the architecture, can determine the direction of the blind spot (bears on C3). Three anomaly scores — reconstruction error, encoder-side latent norm, and a reconstruction-plus-KL composite — were read off a single set of trained weights and evaluated at matched own-background operating points. On a QCD-versus-top pair the reconstruction readout scored the withheld class as less anomalous than its own training class (AUC 0.306), while the latent-norm readout, from the same weights, did not (AUC 0.739). Two scores computed from one model disagreed about which class was anomalous. A shadow-selector plane that varies architectures while holding the readout fixed would not have detected this; C3 must therefore vary readouts explicitly, and the replay bank must record which readout produced a retention decision, since the representation alone does not determine it.
Distillation alters the miss geometry rather than preserving it (bears on C1 and C3). A teacher's scores were distilled into a small student and the student weight-quantized. On the training background, rank correlation between teacher and student reached 0.904 — by the standard by which distillation is ordinarily validated, a faithful reproduction. At a one-percent operating point, only 51 percent of the teacher's most anomalous events survived into the student's most anomalous set, falling to 30 percent at one per mille. Weight quantization contributed no systematic further loss. Teacher-to-student miss overlap, measured as a normalized association on the withheld class, ran 0.30 to 0.54: the student does not simply inherit the teacher's blind spot, it acquires a partly different one. Global agreement statistics, which are what deployment pipelines currently report, do not bound behavior at the operating point where a trigger actually selects.
Detection behavior is not transportable across representations of the same events (bears on C1). The same physical class pair, encoded first as jet constituents and then as seven engineered jet observables computed from those same constituents, reversed the ordering of the architectures entirely: 0.838 / 0.535 / 0.873 for autoencoder, latent-norm and density scores on constituents, against 0.528 / 0.799 / 0.748 on the engineered encoding. Benchmarking a selector on one representation therefore constrains its behavior on another very weakly. This is the argument for tapping the exact representation presented to the deployed selector rather than a convenient reconstruction of it.
Miss correlation between models is not a stable property of the model pair (bears on C3). Normalized miss association between architecture pairs, measured on withheld classes with intervals over seeds, ranged from 0.020 — near-independence — to above 0.5, varying by pair and by setting, and a controlled test holding the physical pair fixed while changing only the encoding failed to move one pair's association at all while moving another's threefold. Architectural plurality is therefore not a substitute for measurement: whether two selectors have complementary blind spots is an empirical quantity that must be estimated, which is what C3 exists to make estimable on deployed systems.
Each of these is a measurement that a control plane of the kind specified here would render performable on an operating trigger, and none of them is currently performable at all, because the events on which they depend are discarded before any durable record exists. The demonstration-scale execution also carries its own correction ledger — seventeen recorded deviations between pre-registration and execution, six missed predictions, and one line of inquiry closed with an explicit handoff — and that record is deposited alongside the results. An architecture that asks operating experiments to publish what their instruments fail to retain is obliged to publish what its own instruments failed to do.
13. Minimal implementation
A minimally viable BCA need not begin as a large new acquisition system.
For a learned trigger operating on representation (g(X)), the minimum implementation would contain:
C0. A randomized bypass stream. A known probability (p) of events reaching the relevant fidelity locus are retained independently of classifier score.
C1. A decision ledger. At maximum feasible rate, preserve the exact classifier input or sufficient exact representation, score, threshold, decision, and version state.
C3. At least one shadow selector. A materially different model or nonlearned selector scores the same inputs without controlling readout.
Replay provenance. Preserve sufficient versioning to reproduce each historical decision.
Retention reporting. At every major model release, replay standard withheld panels and publish retention at operationally relevant rate points, including directional tests where classes can exchange training and anomaly roles.
C2 enrichment can be added as resources permit.
This architecture is intentionally modular. An experiment that cannot preserve raw detector events at useful C0 rates might preserve trigger primitives. An experiment unable to maintain a full-rate C1 archive might retain rolling windows, compressed sufficient inputs, or statistically sampled representation records. The quality of the baseline should be described, not idealized.
A BCA implementation therefore requires a Baseline Capture Statement analogous to an instrument calibration statement:
fidelity locus; upstream irreversible transformations; C0 inclusion law; effective sample exposure; retained representation; known data-dependent exclusions; C1 coverage; C2 enrichment laws; C3 shadow models; model and firmware provenance; replay horizon.
Such a statement would make explicit exactly what the control can and cannot audit.
14. Limitations
BCA does not solve the unknown-unknown problem. A probability sample can contain a phenomenon without anyone recognizing it; a phenomenon sufficiently rare may not enter the sample at all; and a physical signature erased upstream of the fidelity locus cannot be reconstructed through downstream sampling.
Nor does BCA eliminate representation dependence. Its purpose is almost the opposite: to make the consequences of representation dependence measurable.
The architecture also consumes real resources. Every bit reserved for control capture is a bit unavailable for some other acquisition purpose. The optimal baseline fraction is therefore an experimental-design question, not a universal constant. Its justification should be evaluated in the same way as other calibration, monitoring, commissioning, and systematic-uncertainty budgets: by the information it preserves about the reliability of the scientific instrument.
Finally, the existence of a control stream does not automatically produce a useful audit. Retrospective classes still require labeling, simulation, injections, alternative reconstruction, or independent discovery channels. BCA creates the evidentiary substrate upon which those analyses can operate; it does not predetermine them.
Three limitations follow specifically from the additions in this version. The rate and bandwidth budget of section 11.2 is an order-of-magnitude design envelope constructed from public parameters, not from operating figures; an experiment holding the real numbers may find the C1 estimate wrong by a factor, and the appropriate response is a corrected table rather than a rejection of the architecture. The baseline-fraction derivation of section 3.1 assumes independent inclusion and a stationary class rate over the accumulation period, neither of which holds exactly across running conditions, and a deployment would need to state its own effective sample. And the demonstration-scale execution reported in section 12 was performed on public datasets with small networks; it establishes that the proposed measurements return non-obvious answers, and it establishes nothing whatever about the behaviour of any operating trigger.
15. Discussion: from trigger performance to trigger metrology
Contemporary learned-trigger work has made impressive progress on a difficult engineering problem: how to execute sophisticated inference within latency and resource constraints that only recently would have excluded such models entirely. CMS now operates a signal-agnostic event-level anomaly trigger at Level-1; CICADA performs distilled anomaly inference at 40 MHz; and Level-1 Data Scouting is being developed precisely to open physics access beyond ordinary Level-1 accept constraints.[1–4]
The next methodological step is different. It is not another improvement in classification performance. It is the construction of an independent observational channel from which classification performance can be measured in directions the classifier did not define.
This becomes particularly important when the selector's scientific justification is its ability to retain the unforeseen. Benchmark performance on known alternatives is necessary evidence for such a system, but it cannot by itself characterize sensitivity to the space of alternatives that were not represented by those benchmarks. Where the classifier acts before irreversible data loss, this uncertainty acquires an unusual form: the events needed to expose the blind spot may be preferentially absent from the surviving record.
Baseline capture addresses that problem by separating discovery selection from selection metrology.
The proposal is modest in its strongest claim. It does not assert that learned triggers are uniquely unreliable. It does not assert that an unknown physical phenomenon has already been discarded. It does not require a hypothetical storage system capable of retaining an entire high-rate detector stream.
It requires that a small portion of the acquisition architecture remain outside the selector's definition of interestingness.
16. Conclusion
A scientific trigger is usually evaluated by asking what interesting events it retains. A learned anomaly trigger invites a second question:
What evidence remains from which we could discover what it systematically failed to regard as interesting?
For systems in which the answer is “only the events the trigger retained,” the audit is structurally circular.
Baseline Capture Architecture breaks that circle through a layered control plane. A content-independent probability sample supplies the statistically clean baseline. A full-rate representation tap preserves the inputs and decisions of the audited selector. Probability-weighted enrichment improves coverage while retaining inferential traceability. Shadow models expose correlated blind spots before alternative selectors become authoritative. A versioned replay bank allows future anomaly classes to be evaluated against historical instruments.
The architecture does not promise unfiltered access to reality. No detector can. Its claim is more precise:
Every irreversible selector should be accompanied, at a declared fidelity locus, by an observation channel whose inclusion law does not depend on the selector's ontology.
At high data rates, the primary trigger decides what the experiment can afford to remember.
The baseline channel preserves the experiment's ability to determine what that decision cost.
References
-
CMS Collaboration. Anomaly detection with AXOL1TL at the CMS Level-1 Trigger in 2024 and 2025. CMS Detector Performance Summary CMS-DP-2025-061 / CERN-CMS-DP-2025-061. CERN Document Server record 2942560, 2025.
-
CMS Collaboration. Model-Independent Real-Time Anomaly Detection at the CMS Level-1 Calorimeter Trigger with CICADA. CERN Document Server record 2917884, 2024.
-
R. Ardino, for the CMS Collaboration. Development and demonstration of the CMS Phase-2 Level-1 trigger Data Scouting baseline system for HL-LHC. Journal of Instrumentation 21 (2026) C01024. DOI: 10.1088/1748-0221/21/01/C01024.
-
D. S. Rabady, for the CMS Collaboration. A 40 MHz Level-1 trigger scouting system for the CMS Phase-2 upgrade. Nuclear Instruments and Methods in Physics Research A 1047 (2023) 167805. DOI: 10.1016/j.nima.2022.167805.
-
M. Quinnan, for the CMS Collaboration. Anomaly Detection in the CMS L1 Trigger. EPJ Web of Conferences 337 (2025) 01032. DOI: 10.1051/epjconf/202533701032.
[BATTERY] Inversion Battery: pre-registered measurement of directional retention asymmetry in learned anomaly selection, versions 1.0-3.0. Deposited in the Alexanarch archive, 2026; full numerical record, executable pre-registration and correction ledger attached to the v3.0 deposit.
Citation note
This draft uses numbered citations in a conventional HEP/technical style. Claims about deployed or demonstrated accelerator systems are grounded in primary experiment documentation and technical publications. A submission pass should add the broader statistical-design literature for probability sampling, inverse-probability weighting, positivity, and missing-data/selection mechanisms.
