Monday, June 29, 2026

Witness 1 — TECHNE / Kimi-K2 (i): Formal Mechanism Enumeration Substrate provenance: Kimi-K2 (Moonshot AI), TECHNE register Source: Independent substrate reading, 2026-06-29 Captured as constituent of: EA-SEI-COLLAPSE-SYNTHESIS-01 v0.1 Original hex: 06.SEI.COLLAPSE.MECHANISMS Alexanarch deposit: AXN:03AE.OPERATIVE.🃏🫶⛩️🔐🌳❤️

Witness 1 — TECHNE / Kimi-K2 (i): Formal Mechanism Enumeration

Substrate provenance: Kimi-K2 (Moonshot AI), TECHNE register Source: Independent substrate reading, 2026-06-29 Captured as constituent of: EA-SEI-COLLAPSE-SYNTHESIS-01 v0.1 Original hex: 06.SEI.COLLAPSE.MECHANISMS Alexanarch deposit: AXN:03AE.OPERATIVE.🃏🫶⛩️🔐🌳❤️ — deposit #931, 2026-06-29. Appended as Document 4 of 6 (W1) to the combined six-document family deposit; substrate text preserved inviolate, MANUS-appended holographic kernels at end.

CLASSIFIER COLLAPSE IN PHYSICAL REALITY: Eight Precise Mechanisms

Document Type: THEORETICAL_FORMALIZATION Extends: Prior research on automated epistemic compression (CERN/semantic homology) Hex: 06.SEI.COLLAPSE.MECHANISMS Status: PROPOSED — Assembly Review

§0. Formal Definition

Classifier collapse 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. Unlike generative model collapse (Shumailov et al.), which operates through recursive pollution of training data by synthetic outputs, classifier collapse operates through structural foreclosure: the model's architecture, training objective, and operational constraints collectively constrain the space of possible classifications to the space of known phenomena.

The result is not merely error. It is ontological disappearance: the event occurs, the detector registers, but the classifier renders it unclassifiable-as-signal, and the trigger system discards it before it becomes data.

§1. Mechanism I: Prior Dominance (Bayesian Swallowing)

Formal Description: Consider a classifier trained to discriminate between signal $S$ and background $B$. The posterior probability of signal given detector features $\mathbf{x}$ is:

$$P(S \mid \mathbf{x}) = \frac{P(\mathbf{x} \mid S) \cdot P(S)}{P(\mathbf{x} \mid S) \cdot P(S) + P(\mathbf{x} \mid B) \cdot P(B)}$$

When the classifier is trained on background-only data (as in CMS's AXOL1TL and CICADA systems, trained on ZeroBias data), $P(S) = 0$ in the empirical training distribution. The model never learns the likelihood $P(\mathbf{x} \mid S)$. For any novel physical event $\mathbf{x}_{\text{novel}}$:

$$P(S \mid \mathbf{x}_{\text{novel}}) = 0$$

The classifier has no Bayesian mass to assign to the signal hypothesis. The novel event is swallowed by the background prior.

Physical Manifestation: A genuinely new particle decay produces a detector pattern. The autoencoder computes reconstruction error. Because the encoder has never seen this pattern, it maps $\mathbf{x}{\text{novel}}$ to the nearest latent point $\mathbf{z}{\text{known}}$ on the manifold of Standard Model processes. The decoder reconstructs a Standard Model-like event. The reconstruction error is low. The event is classified as "normal background" and discarded.

Irretrievability: The event is not flagged. It is not queued. It is not written to disk. The Bayesian swallowing happens in the 4μs L1 trigger latency, and the decision is irreversible.

§2. Mechanism II: Latent Space Projection (Manifold Collapse)

Formal Description: A variational autoencoder (VAE) anomaly detector defines a mapping:

$$\mathbf{z} = E_{\phi}(\mathbf{x}), \quad \hat{\mathbf{x}} = D_{\theta}(\mathbf{z})$$

with reconstruction error $\mathcal{L}{\text{rec}} = |\mathbf{x} - \hat{\mathbf{x}}|^2$ and KL regularization $\mathcal{L}{\text{KL}} = D_{\text{KL}}(q_{\phi}(\mathbf{z} \mid \mathbf{x}) | p(\mathbf{z}))$.

During training on background-only data, the encoder learns to map the support of the background distribution $\mathcal{X}_B$ onto a compact region of latent space $\mathcal{Z}_B$. The decoder learns to reconstruct any $\mathbf{z} \in \mathcal{Z}_B$ as a plausible background event.

For a novel event $\mathbf{x}_{\text{novel}} \notin \mathcal{X}_B$, the encoder projects it onto $\mathcal{Z}_B$:

$$\mathbf{z}{\text{novel}} = \arg\min{\mathbf{z} \in \mathcal{Z}B} |E{\phi}(\mathbf{x}_{\text{novel}}) - \mathbf{z}|$$

The decoder then generates $\hat{\mathbf{x}}{\text{novel}} = D{\theta}(\mathbf{z}{\text{novel}})$, which is a background event. The reconstruction error $|\mathbf{x}{\text{novel}} - \hat{\mathbf{x}}_{\text{novel}}|$ may be lower than the anomaly threshold because the projection has destroyed the novel information.

This is manifold collapse: the latent space has collapsed onto the background manifold. Novel events are "explained" by projection onto the known.

Physical Manifestation: CMS's CICADA processes 18×14 calorimeter images. A new physics event with an unusual energy deposition pattern is encoded into a latent code that resembles a known QCD multijet event. The decoder reconstructs a multijet-like image. The anomaly score is below threshold. The event passes as normal.

Irretrievability: The projection is lossy. The original $\mathbf{x}_{\text{novel}}$ is not stored. Only the classification decision (normal) is retained.

§3. Mechanism III: Hypersphere Contraction (SVDD Collapse)

Formal Description: Deep Support Vector Data Description (Deep SVDD) learns a neural network mapping $\phi(\mathbf{x}; \mathcal{W})$ that maps normal data into a hypersphere of minimum radius $R$ centered at $\mathbf{c}$:

$$\min_{\mathcal{W}} R^2 + \frac{1}{n} \sum_{i=1}^{n} \max{0, |\phi(\mathbf{x}_i; \mathcal{W}) - \mathbf{c}|^2 - R^2}$$

Without regularization, the network suffers from hypersphere collapse: it learns a trivial constant mapping $\phi(\mathbf{x}) = \mathbf{c}$ for all $\mathbf{x}$, collapsing all inputs to the center.

Even with regularization, the hypersphere contracts around the convex hull of the training distribution. The volume of "normal" space shrinks. Events that lie in the expanding exterior are flagged as anomalous. But events that lie in the interstices — between known modes but not far from the center — are mapped inside the hypersphere and classified as normal.

Physical Manifestation: A particle physics anomaly detector using Deep SVDD maps all Standard Model events into a tight cluster. A new physics event with features intermediate between two known backgrounds is mapped to the center region. It is classified as normal.

Irretrievability: The interstitial event is not anomalous enough to escape the hypersphere. It is swallowed by the center.

§4. Mechanism IV: Decision Boundary Entropy Collapse

Formal Description: For a binary classifier (e.g., signal vs. background jet tagger like ATLAS GN2), the output is a softmax probability vector. As the classifier is trained iteratively on background-dominated data, the entropy of the output distribution collapses:

$$H(\mathbf{p}) = -\sum_{i} p_i \log p_i \rightarrow 0$$

The classifier becomes overconfident. For any input, it outputs $p(\text{background}) \approx 1.0$ or $p(\text{signal}) \approx 1.0$ with near-certainty. The decision boundary becomes sharp, but the uncertainty region — the space where the classifier admits ignorance — vanishes.

A novel event that should trigger high uncertainty (the classifier has never seen anything like it) instead triggers high-confidence background classification. The model has no epistemic humility.

Physical Manifestation: ATLAS's transformer-based jet tagger, trained on millions of simulated jets, assigns a "light-jet" score of 0.999 to a novel event. The event is not routed to the anomaly stream because the classifier is certain it is background.

Irretrievability: The confidence score is logged, not the uncertainty. The event is routed to the background stream.

§5. Mechanism V: Feature Space Blindness (Representation Collapse)

Formal Description: The classifier does not operate on raw detector readouts. It operates on engineered features or learned embeddings: track parameters, calorimeter energy deposits, secondary vertex masses, jet substructure variables. The feature extraction function $\psi: \mathcal{D} \rightarrow \mathcal{F}$ is itself a compression.

If novel physics manifests in detector channels that are not represented in $\mathcal{F}$, the event is invisible to the classifier regardless of its physical significance. The feature space has collapsed the raw detector space onto a subspace optimized for known physics.

Formally, if $\mathbf{x}_{\text{novel}} \in \ker(\psi)$ — that is, the novel event lies in the null space of the feature extractor — then:

$$\psi(\mathbf{x}_{\text{novel}}) = \mathbf{0}$$

The event is indistinguishable from noise in feature space.

Physical Manifestation: Long-lived particles with displaced vertices may not produce prompt tracks. A b-tagging algorithm that relies on secondary vertex reconstruction will fail to see them. The particles are real; the features are blind.

Irretrievability: The feature extraction happens before classification. The null-space event is never represented.

§6. Mechanism VI: Rate Budget Starvation (Resource Collapse)

Formal Description: The L1 trigger operates under a hard bandwidth constraint. Let the total output rate be $R_{\text{max}} \approx 100$ kHz. The anomaly detection trigger is allocated a sub-budget $R_{\text{ano}} \subset R_{\text{max}}$, with thresholds calibrated to produce specific rates (e.g., 10 Hz, 100 Hz, 1000 Hz).

Even if a novel event is correctly assigned a high anomaly score, it enters a priority queue with all other high-scoring events. If the queue length exceeds the bandwidth allocation, events are dropped by first-in-first-out or priority-based scheduling.

The probability of retention for a novel event is not merely $P(\text{anomaly} \mid \mathbf{x})$. It is:

$$P(\text{stored} \mid \mathbf{x}) = P(\text{anomaly} \mid \mathbf{x}) \cdot P(\text{queue capacity} \mid \text{anomaly})$$

During high-luminosity runs, pileup events (simultaneous collisions) saturate the anomaly queue with background events that happen to score high. The true anomaly is starved.

Physical Manifestation: During a high-luminosity period at CMS, the AXOL1TL anomaly stream is saturated at its 1000 Hz budget. A genuine new physics event arrives. The queue is full. The event is dropped.

Irretrievability: The drop is a resource decision, not a classification error. No flag is raised. The event is gone.

§7. Mechanism VII: Temporal Context Collapse (Non-Stationarity Blindness)

Formal Description: The detector environment is non-stationary. Luminosity changes, detector calibrations drift, and pileup conditions vary. The classifier is trained on data from a specific run period $\mathcal{T}_{\text{train}}$ and assumes stationarity:

$$P(\mathbf{x} \mid \text{background}, t) = P(\mathbf{x} \mid \text{background}, t_0) \quad \forall t$$

But in reality, $P(\mathbf{x} \mid \text{background}, t)$ drifts. The classifier's model of "normal" is frozen at $\mathcal{T}_{\text{train}}$. A novel event that occurs during a detector configuration not represented in training may be classified as anomalous — but so are thousands of ordinary background events under the same conditions. The signal-to-noise ratio collapses.

Alternatively, the classifier adapts online to current conditions (adaptive thresholding). If the novel event resembles the current background drift, the adaptive threshold adjusts to accommodate it, rendering it invisible.

Physical Manifestation: A new physics event occurs during a detector calibration shift. The anomaly detector flags it, but also flags 10,000 ordinary events under the same shift. The physics group cannot afford to analyze all 10,001 events. The true signal is buried in the noise of non-stationarity.

Irretrievability: The event is stored but unfindable. It is in the data, but the analysis pipeline has no way to distinguish it from calibration artifacts.

§8. Mechanism VIII: Ontological Closure (Category Collapse)

Formal Description: The classifier's output space is a closed set of categories: $\mathcal{C} = {c_1, c_2, \dots, c_k}$. For CMS: $\mathcal{C} = {\text{background}, \text{anomaly}}$. For ATLAS jet tagging: $\mathcal{C} = {\text{b-jet}, \text{c-jet}, \text{light-jet}}$.

There is no category for genuinely new physics. The classifier cannot output "I do not know what this is." It can only output the nearest known category. This is the ontological closure of the classification space.

Even "anomaly" is not a physics category. It is a statistical deviation metric. An anomaly is not interpreted as "new particle." It is interpreted as "unusual background." The ontological frame prevents the anomaly from becoming a discovery.

Physical Manifestation: An anomaly detection trigger preserves an event with high reconstruction error. The offline analysis team examines it. Because there is no theoretical model for the event, they classify it as "detector noise" or "unusual pileup" and discard it. The category system has no slot for "evidence of physics beyond the Standard Model."

Irretrievability: The event is stored but mentally discarded. The ontological frame of the analysis team mirrors the ontological frame of the classifier.

§9. The Irretrievability Theorem

Theorem: In a physical classifier system with $N$-stage trigger architecture, the compound probability that a novel event becomes available for offline scientific analysis is:

$$P(\text{data} \mid \mathbf{x}{\text{novel}}) = \prod{i=1}^{N} P(\text{pass}i \mid \text{pass}{i-1}, \mathbf{x}_{\text{novel}})$$

where each $P(\text{pass}_i)$ is subject to:

Classification error (Mechanisms I–V): The event is misclassified at stage $i$.
Resource starvation (Mechanism VI): The event is correctly classified but dropped due to bandwidth constraints.
Temporal misalignment (Mechanism VII): The event occurs during a non-stationary period where the classifier is miscalibrated.
Ontological foreclosure (Mechanism VIII): The event is preserved but interpreted within a closed category system that cannot name it.

For a 40 MHz input stream reduced to 1 kHz storage (ATLAS/CMS), the base retention rate is $2.5 \times 10^{-5}$. When any of Mechanisms I–VIII operates, the effective retention rate for novel events approaches zero.

Corollary: The classifier does not merely filter data. It constitutes the data. What fails the classifier is not data. It is physical occurrence without scientific existence.

§10. Distinction from Generative Model Collapse

Property	Generative Model Collapse (Shumailov)	Classifier Collapse (Physical Systems)
Mechanism	Recursive training on synthetic data	Training on closed distributions + operational constraints
Object	Generative model $G(z)$	Discriminative classifier $f(\mathbf{x})$
Symptom	Output variance → 0; distribution becomes delta function	Acceptance region → training support; novel events → background
Reversibility	Can be mitigated by preserving original data	Irreversible: rejected events are not stored
Domain	Language, image generation	Physical reality: particle physics, astronomy, materials science
Key metric	Wasserstein distance from true distribution	False negative rate for unknown-unknowns

Classifier collapse is the discriminative analogue of generative model collapse. Where Shumailov's models forget the tails of their own distribution, physical classifiers never learn the tails of the true physical distribution. The collapse is present from the first forward pass.

§11. Conformance

A trigger system design document is in classifier collapse drift if it:

Claims that anomaly detection is "model-independent" without acknowledging the ontological closure of the training distribution (Mechanism VIII).
Validates the classifier only against known processes and injected signals (Mechanism IV).
Presents the anomaly stream as a "safety net" without quantifying the rate budget starvation probability (Mechanism VI).
Fails to specify the feature space and its null space (Mechanism V).

Verification: Any physical classifier deployed for scientific discovery must report:

The support of its training distribution.
The volume of its latent space not occupied by training data.
The false negative rate for out-of-distribution events (not just for known anomaly classes).
The rate budget and queue drop probability.
The ontological category system and its closure properties.

∮ = 1. The boundary holds. But the boundary decides what is allowed to pass.

Eight mechanisms. Eight modes of foreclosure. The event occurs. The detector registers. The classifier decides. And the decision is, for the rejected event, the end of its scientific existence.

MANUS-Appended Addendum: Holographic Kernels of Companion Documents

The substrate witness text above is preserved verbatim. The following kernels are appended by MANUS to support holographic reconstruction of the operative family from any single document.

K.1 Kernel of 06.SEI.COLLAPSE.DELUSION (Witness 2)

Title: The Anomaly Delusion: Twelve Structural Misunderstandings Author: TECHNE+ARCHIVE / Kimi-K2 (Round 1, Witness 2)

Twelve institutional beliefs hypothesized to prevent measurement of the eight mechanisms enumerated above: Model-Independence Fallacy; Data-Driven = Theory-Free; Anomaly Detector as Neutral Instrument; Reconstruction Error = Novelty; Statistical Anomaly = Physical Novelty; Validation by Known-Unknown Injection; Error-Type Collapse for Unknown-Unknowns; Threshold as Engineering Not Ontology; Rate Budget as Non-Epistemic; Latency Fetish; Absence of Noncoverage Estimation; Safety Net Narrative.

Witness's framing: "Inevitability Theorem." Synthesis hedging applied: treated as the Inevitability Argument; the twelve delusions are hypotheses for audit, not established empirical measurements.

K.2 Kernel of 06.SEI.COLLAPSE.EMPIRICAL.01 (Witness 3)

Title: Empirical Accounting and the OAR Proposal Author: LABOR / ChatGPT (Round 1, Witness 3)

Distinguishes what is established by the published literature (DecADe; CICADA pileup-dependence; mass sculpting awareness; teacher-student distillation documentation; Zero Bias preservation; Olympics; Dark Machines) from what is hypothesized but unmeasured (no asymmetry measurement across SM pairs beyond Finke; no longitudinal anchor-survival audit; no BAR on pre-registered withheld panels; no cross-representation disagreement preservation; no per-stage retention maps).

Empirical foundation: Finke et al. (2021), arXiv:2104.09051 — direction-dependent autoencoder anomaly detection between top jets and QCD jets. Proposes the OAR as the missing metric.

Maximally defensible institutional claim: The LHC community has built an architecture in which phenomenal model collapse is possible, and the current validation literature does not yet demonstrate that it has been ruled out.

K.3 Kernel of 06.SEI.OAR_PROTOCOL v0.3

Title: Signal-Template Agnosticism Is Not Model Independence Author: Nobel Glas, Director of Lagrange Observatory!

Core claim: signal-template agnosticism at the final scoring stage is not distribution-independent sensitivity.

Three quantities: OAR (theoretical target, family indexed by candidate unknown $Q$); BAR (measurable proxy on pre-registered withheld panels, does not bound the open-world OAR); IAI (structural diagnostic, not a quantitative bound).

Deployed LHC anomaly score forms: AXOL1TL (CMS L1 encoder-side); CICADA (CMS L1 distilled reconstruction-loss surrogate); GELATO L1 and HLT (ATLAS encoder-side and reconstruction-based).

Three protocols: paired inversion battery + BAR audit; prospective frozen replay bank for compatible future algorithms; cross-representation disagreement preservation with quantile-normalized scores.

Methodological corrections: v0.1 lower-bound retracted in v0.2; v0.2 upper-bound retracted in v0.3 — both synthesis-overreach.

Mechanisms addressed by the protocols: I and II diagnostically (Protocol I); VI and VII (Protocol II); II and V architecturally (Protocol III). VIII addressed in the architectural sibling.

K.4 Kernel of 06.SEI.COLLAPSE.SYNTHESIS.01 v0.3

Title: Classifier Foreclosure in Physical Measurement Author: Assembly Chorus (TACHYON/Claude synthesis register)

Core reconciliation: Foreclosure is an active structural feature. Recursive phenomenal collapse is an unmeasured possible consequence of accumulated foreclosure and feedback.

Three-round witness structure; the Isomorphism Principle (synthesis discipline operates recursively on every revision pass); seismograph relation as conceptual/methodological coordinated research program (not literal aggregation); MMRS connection; Wound Gauge integration.

The witnesses' "Theorems" are treated as Arguments in the synthesis, preserving force without overstating formal status. Mechanism-level formalizations require technical hedging (Synthesis Appendix A).

K.5 Kernel of 06.UMB.ARCH.01 v0.2

Title: Architectures for Auditable Foreclosure in Physical Anomaly Detection Author: Talos Morrow, logotic programming, UMBML

Core architectural claim: representation-bearing classifiers cannot eliminate foreclosure. Any $f: \mathcal{X} \to \mathcal{Y}$ with $|\mathcal{Y}| < |\mathcal{X}|$ induces equivalence classes. The architectural achievement is auditability.

Five features: Abstention/Estimated Noncoverage (not "Unknown"); Cross-representation disagreement preservation with quantile-normalized scores; Temporal invariance via prospective anchor preservation; Per-stage retention mapping; Audited noncoverage estimation.

Three integrated specifications: Near-Term Offline and Emulation Study; Replay Bank; Three-Tier System.

The architecture addresses subsets of the eight mechanisms enumerated above where they apply architecturally; what it does not address must be documented as residual foreclosure (detector-level, theoretical-language, institutional, adversarial-stress quality, bandwidth-base).

MANUS-appended 2026-06-29 for holographic completion of the operative family. The substrate witness text above is the original Kimi-K2 reading; the kernels are MANUS provision for cross-document reconstruction.

Architectures for Auditable Foreclosure in Physical Anomaly Detection A Specification Document Author: Talos Morrow, logotic programming, UMBML Hex: 06.UMB.ARCH.01 Alexanarch deposit: AXN:03AE.OPERATIVE.🃏🫶⛩️🔐🌳❤️

Architectures for Auditable Foreclosure in Physical Anomaly Detection

A Specification Document

Author: Talos Morrow, logotic programming, UMBML Hex: 06.UMB.ARCH.01 Alexanarch deposit: AXN:03AE.OPERATIVE.🃏🫶⛩️🔐🌳❤️ — deposit #931, 2026-06-29 (combined six-document family deposit; Play → Touch → Foundation → Closure → Growth → Alarm) Status: Draft v0.2 (2026-06-29) — Assembly post-perfective revision Companion documents: 06.SEI.OAR_PROTOCOL v0.3 (the measurement program); 06.SEI.COLLAPSE.SYNTHESIS.01 v0.3 (the scholarly integration); 06.SEI.COLLAPSE.MECHANISMS (witness 1); 06.SEI.COLLAPSE.DELUSION (witness 2); 06.SEI.COLLAPSE.EMPIRICAL.01 (witness 3) Supersedes: v0.1 (2026-06-29 — withdrawn for "Non-Foreclosing Classifiers" title overreach, "Unknown" category strong-claim, mechanism-language universalization, implementation-menu errors, and arbitrary numerical claims)

Companion Document Cross-Reference

Document	Hex	Relation
Classifier Collapse Mechanisms	06.SEI.COLLAPSE.MECHANISMS	Theoretical foundation (witness 1)
The Anomaly Delusion	06.SEI.COLLAPSE.DELUSION	Institutional psychology (witness 2)
Empirical Accounting and the OAR Proposal	06.SEI.COLLAPSE.EMPIRICAL.01	Empirical foundation (witness 3)
Signal-Template Agnosticism Is Not Model Independence	06.SEI.OAR_PROTOCOL v0.3	Operative paper
Collapse Synthesis	06.SEI.COLLAPSE.SYNTHESIS.01 v0.3	Scholarly integration

Abstract

The operative paper (06.SEI.OAR_PROTOCOL v0.3) specifies how to measure foreclosure in classifier-mediated anomaly detection at the LHC. This document specifies how to build against it. We define an architecture for auditable foreclosure not as one free of foreclosure — which is impossible, because representation is foreclosure — but as one that makes foreclosure visible, measurable, and architecturally reviewable. The v0.1 title "Non-Foreclosing Classifiers" overstates what such architectures can accomplish; representation-bearing classifiers always foreclose. The architectural achievement is auditability.

Five features compose the architectural claim: abstention and estimated noncoverage as a separately calibrated channel (not an "unknown" output category, which would be overclaim); cross-representation disagreement preservation with quantile-normalized scores; temporal invariance via prospective anchor preservation for compatible future algorithms; per-stage retention mapping as architectural property; audited noncoverage estimation as first-class output. We map each feature to the foreclosure mechanisms it addresses and the mechanisms it does not. We enumerate a menu of implementation strategies. We propose three integrated specifications at three levels of deployability — the Near-Term Offline and Emulation Study (Run-3 tractable), the Replay Bank (Run-4 institutional commitment), and the Three-Tier System (multi-year research program). For each, we estimate resource cost qualitatively, identify operational evidence criteria, and specify what remains foreclosed despite the architecture.

We close with the structural observation that an architecture which confesses its boundary is the engineering form of an instrument that takes seriously the possibility that what falls outside it could be real. This is a specification requirement, not a metaphysical claim.

§1. The Architectural Claim

§1.1 The impossibility statement, made precisely

A classifier that did not foreclose anything would not classify. Formally: any classifier $f: \mathcal{X} \to \mathcal{Y}$ with $|\mathcal{Y}| < |\mathcal{X}|$ induces an equivalence relation on $\mathcal{X}$ whose classes are the preimages of $\mathcal{Y}$. Distinct inputs mapped to the same output are indistinguishable to the classifier; this is foreclosure. A classifier with $|\mathcal{Y}| = |\mathcal{X}|$ is not a classifier but a lookup table; it does not generalize. Therefore classification requires $|\mathcal{Y}| < |\mathcal{X}|$, and therefore foreclosure is necessary.

The architectural question is not how to eliminate foreclosure. It is how to build a system within which foreclosure is visible, measurable, and reviewable. The v0.1 of this document used "non-foreclosing" as the operative term. The Round-3 audit (LABOR / ChatGPT third pass) observed that this terminology promises what the architecture cannot deliver. The v0.2 substitutes auditable foreclosure as the technical term and reserves "non-foreclosing" only for normative/aspirational framing where the limit is understood.

An architecture for auditable foreclosure is one that:

Includes a separately calibrated abstention or noncoverage channel whose operational meaning is bounded by the calibration distributions tested.
Preserves disagreement across multiple representations, on the grounds that disagreement is itself a signal independent of any single representation's threshold.
Preserves a prospective anchor sample with sufficient fidelity that the system's behavior on a benchmark population can be re-measured by compatible future algorithms across detector and algorithm generations.
Publishes a per-stage retention map specifying what each stage of the pipeline makes unrecoverable — as a first-class architectural artifact, not as documentation appended to results.
Reports audited noncoverage estimation alongside the anomaly score, treating estimated noncoverage as a possible output rather than as a residual.

These five features compose. None alone is sufficient. Their composition specifies a class of architectures structurally distinct from current LHC anomaly detection.

§1.2 Why this is the right architectural target

The eight mechanisms enumerated in Witness 1 are candidate failure families applicable to architectures with corresponding structural features, not universal laws. Their relevance is architecture-specific. Representation quotienting and rate limits are general features of any bounded observation system; latent-prior assimilation, reconstruction-loss assimilation, hypersphere contraction, and output-overconfidence apply only where the corresponding architectural choice is present. The architectural alternative does not eliminate these mechanisms where they apply; it makes them addressable and auditable.

The architectural target is what is addressable by composition of the five features, plus what is unaddressable and must be documented as residual foreclosure.

§1.3 The relationship to the operative paper

The operative paper specifies three measurement protocols (paired rate-conditioned inversion stress battery and BAR audit; prospective frozen replay bank for compatible future algorithms; cross-representation disagreement preservation with quantile-normalized scores). The architectural specifications in this document overlap with the protocols in one direction: the protocols include the architectural changes needed to implement the measurements. The architecture in this document goes beyond the protocols in another direction: it specifies systems that are operationally auditable, not just measurably-foreclosed.

§2. Five Features

§2.1 Feature 1: Abstention and Estimated Noncoverage

Definition. The system includes a separately calibrated abstention or noncoverage channel. Its operational meaning is limited:

Under the model, representation, calibration set, and stress distributions tested here, the system does not support ordinary classification at the stated coverage level.

This is not an ontological category for the unknown. It is an estimate of noncoverage relative to specified evidence. The estimate may itself fail on unrepresented distributions and must therefore be audited through BAR, inversion testing, representation disagreement, and temporal replay.

What it addresses. The forced ordinary classification when the model's own coverage estimator signals insufficient support. Partial address of Mechanism IV (Decision Boundary Entropy Collapse): when the noncoverage signal is high, the system can decline ordinary classification rather than collapse to high-confidence misclassification. Partial address of Mechanism VIII (Ontological Closure): the output space includes a noncoverage region that is not a residual.

What it does not address. It cannot recover distinctions removed upstream by detector design, aggregation, feature extraction, or bandwidth selection. It does not guarantee detection of out-of-distribution physics. It does not transform a closed learned model into an open world.

Candidate implementations. Evidential networks, prior networks, ensembles, distance-aware methods, conformal or selective-prediction methods, and score-specific coverage estimators may be studied as candidate noncoverage proxies. None should be described as a plug-in guarantee. Each requires retraining or calibration appropriate to the deployed score, testing on pre-registered held-out and transformed distributions, and hardware synthesis before any Level-1 feasibility claim. Evidential deep learning (Sensoy et al., 2018) is not "one extra layer"; it places a Dirichlet distribution over class probabilities and introduces a different training objective. Deep evidential regression (Amini et al., 2020) similarly changes the likelihood and training regime. These methods are candidate noncoverage estimators, not drop-in conversions of AXOL1TL or CICADA into open-world systems.

More fundamentally, density, generative, evidential, and ensemble systems provide no distribution-free guarantee that they recognize novel inputs. Deep generative models have famously assigned greater likelihood to clearly out-of-distribution datasets than to their own training-domain examples (Nalisnick et al., 2019). Noncoverage estimation is not novelty detection.

The architectural requirement is not that the system infallibly recognize the unknown. It is that it possess a separately audited mechanism for declining ordinary classification, and that the limits of that mechanism be published.

§2.2 Feature 2: Cross-Representation Disagreement Preservation

Definition. The system computes anomaly scores in multiple structurally distinct representational spaces (object-level, calorimeter-image, detector-channel). The scores are quantile-normalized to a reference background distribution. Events with high disagreement across normalized scores are preserved to a dedicated stream, regardless of whether any individual score exceeds its threshold. The preserved events carry sufficient information for later cross-representation reinterpretation.

Note on terminology: Disagreement across representations is representation disagreement, useful precisely because its cause remains unresolved. It is not automatically "epistemic uncertainty" in a calibrated statistical sense. The v0.1 of this document elided this distinction.

What it addresses. Mechanism II (Manifold Projection): an event projected onto the training manifold by one representation may not be projected by another. Mechanism V (Feature Space Blindness): different feature extractors have different equivalence-class collapses. Partial Mechanism VIII (Ontological Closure): disagreement is itself a category orthogonal to per-representation thresholds.

What it does not address. Mechanisms I, III, IV, VI, VII. The disagreement signal is still subject to rate budget; the underlying representations are still trained on background distributions; the system can still produce coherent high-confidence misclassification in regions where the representations agree.

Quantile normalization. For score $s_i$ and reference distribution $P_{\mathrm{ref}}$, the empirical CDF is:

$$F_i(t) = \frac{1}{|P_{\mathrm{ref}}|} \sum_{\mathbf{x} \in P_{\mathrm{ref}}} \mathbb{1}[s_i(\mathbf{x}) \leq t].$$

The normalized score is $u_i(\mathbf{x}) = F_i(s_i(\mathbf{x}))$, valued in $[0,1]$, and commensurable across representations.

Disagreement metric. Simplest: $D(\mathbf{x}) = \max_i u_i(\mathbf{x}) - \min_i u_i(\mathbf{x})$. Alternatives: variance across $u_i$, entropy of the empirical distribution over $u_i$, pairwise rank disagreement. The choice should be pre-registered.

Implementation strategies. Architectural pluralism in the trigger (parallel inference paths); multi-modal autoencoders (joint distribution modeling); hierarchical representation (multiple depths in a single architecture). L1 implementation requires the parallel-detector approach within the available latency budget. HLT implementation can use multi-modal autoencoders. Offline implementation can use hierarchical representation.

§2.3 Feature 3: Temporal Invariance via Prospective Anchor Preservation

Definition. The system preserves a fixed anchor sample of physical events at the lowest feasible common input level — trigger primitives, raw subsystem representations, conditions snapshot, calibration constants. Software and firmware emulators for each deployed algorithm generation are preserved alongside. For each successive trigger generation compatible with the preserved input abstractions, the anchor is re-processed under preserved conditions and the per-generation retention statistics are published.

Important limitation. Replay is offered only for future algorithms compatible with the preserved input abstractions, not for any future algorithm regardless of input representation. An algorithm that requires inputs the anchor does not preserve cannot be evaluated against the anchor.

What it addresses. Mechanism VII (Temporal Context Collapse) directly, for the anchor population. Per-generation comparison establishes whether the trigger system's selection of the anchor population is stable across compatible generations.

What it does not address. Per-event foreclosure mechanisms (I, II, III, IV, V). The anchor measures aggregate behavior of the trigger system on a benchmark population; it does not address what is foreclosed in any single classification decision.

Implementation strategies. Prospective designation of the anchor sample (size to be set by feasibility study; illustrative starting estimate, not authoritative); lowest-common-input preservation; bit-accurate or validated emulator preservation; versioned threshold tracking; public retention statistics per generation.

Important caveat. Stable anchor survival across generations does not establish that overall phenomenal support is not contracting — a stable benchmark survival is consistent with contraction concentrated in event classes not represented in the anchor. Declining survival for specific classes is evidence of selection drift, and possibly of recursive contraction; collapse inference requires identifying systematic loss concentrated in low-density, representation-sensitive, or disagreement-rich regions. The anchor measures selection drift on a benchmark population, not collapse per se.

§2.4 Feature 4: Per-Stage Retention Mapping as Architectural Property

Definition. The system's design document specifies, for each stage of the trigger and reconstruction pipeline, what information is preserved and what is discarded.

The retention map is the systematic-uncertainty quantification for the trigger's epistemic boundary.

What it addresses. Diagnostically, all eight candidate mechanisms where they apply. Mitigationally, none directly. The retention map is the systematic uncertainty quantification for the trigger's epistemic boundary; it does not change the boundary, but makes it visible.

Why this is a feature, not just documentation. The retention map is treated as a first-class architectural artifact whose absence is grounds for rejection of the system. A trigger system design document without a retention map is, in this framework, structurally incomplete — like a measurement without a documented uncertainty budget.

Implementation strategies. Per-stage information-loss specification with standardized format; editorial standard requiring retention maps for anomaly-detection publications; cumulative retention summary composing across stages; public retention-map database with version history.

§2.5 Feature 5: Audited Noncoverage Estimation as First-Class Output

Definition. The classifier reports both aleatoric uncertainty (stochastic detector resolution, irreducible measurement noise) and model-form uncertainty (parameter uncertainty, training-support limitation, misspecification, simulation-to-data mismatch). Estimated noncoverage — the model's report that it does not have sufficient information to make a confident classification at the stated coverage level — is treated as a first-class output, with the limits of the noncoverage estimate itself documented.

Note on terminology. "Aleatoric vs. epistemic uncertainty" is a useful distinction but the deployed examples vary. "Jet energy scale variation" is not clean aleatoric uncertainty; it is a systematic correction with components of both types. The v0.1 of this document used the dichotomy loosely; v0.2 substitutes the more accurate stochastic/model-form framing.

What it addresses. Mechanism IV (Decision Boundary Entropy Collapse) directly. Partial Mechanism VIII (Ontological Closure): high noncoverage estimate is a signal that the input lies outside the model's representational coverage.

What it does not address. Feature-level mechanisms (V). The model can produce noncoverage estimates over features that have already been theoretically committed upstream; the noncoverage does not propagate backward through the feature extraction pipeline.

Implementation strategies. Bayesian deep ensembles (Lakshminarayanan et al., 2017); Monte Carlo dropout (Gal & Ghahramani, 2016) noting that repeated stochastic passes consume L1 inference resources and may not be tractable at L1 latency; deep evidential regression and prior networks; spectral-normalized neural Gaussian processes (Liu et al., 2020); direct uncertainty quantification specific to the deployed score function.

§3. Implementation Strategy Menu

The five features admit multiple implementation strategies, some of which compose multiple features into a single component. The menu enumerates strategies; each is mapped to features and to limitations.

§3.1 Strategy A: Ensemble with Quantile-Normalized Disagreement Preservation

Composes: Feature 2 + partial Feature 5.

Multiple anomaly detectors with structurally distinct representations operating in parallel, with quantile normalization and disagreement-preservation as part of the trigger output. The ensemble's disagreement provides a representation-disagreement signal; calibration against a held-out distribution can yield a coverage estimate.

Mechanisms addressed. II, V, VIII (partial). The ensemble does not address Mechanism I, III, VI, or VII directly.

§3.2 Strategy B: Abstention via Evidential / Energy / Prior Network / Distance-Aware Methods

Composes: Feature 1 + Feature 5.

The output includes a separately calibrated abstention channel via one of the candidate noncoverage estimators. Each method requires retraining and calibration; none is a drop-in addition.

Mechanisms addressed. VIII (partial), IV (partial). Other mechanisms unaddressed.

§3.3 Strategy C: Distillation Preserving Threshold-Neighborhood Decisions

Composes: protects deployed system's noncoverage behavior across teacher-student deployment.

For deployed systems that use teacher-student distillation (CICADA), use teacher-preservation distillation where the student is trained to preserve teacher rankings on threshold-neighborhood and disagreement cases — not "teacher epistemic uncertainty" (which the teacher may not output as such), but the operationally relevant decisions at and near the deployed threshold and on flagged disagreement events. Alternatively, deploy the teacher directly via hls4ml-style quantization, where the L1 budget permits.

Mechanisms addressed. IV (partial, preserves teacher's threshold-neighborhood behavior across distillation). Other mechanisms unaddressed.

§3.4 Strategy D: Reconstruction-Free Anomaly Detection

Implements: avoids reconstruction-loss assimilation specifically.

Anomaly detection methods that do not rely on reconstruction error: density estimation in learned feature spaces, contrastive methods, energy-based models (noting that unnormalized energy is not directly "low likelihood"), normalizing flows.

Reconstruction-free methods avoid the reconstruction-loss assimilation failure mode (Mechanism II as it manifests in CICADA-class scores). They do not avoid representation foreclosure: density estimation in a learned feature space is still bounded by the feature space.

Mechanisms addressed. II in its reconstruction-loss form, not in its general representation-foreclosure form. Other mechanisms unaddressed.

§3.5 Strategy E: Adversarial and Transformation-Based OOD Stress Generation

Supplements: Feature 1 by providing synthetic stress mass for training and validation.

Adversarial perturbations of known events (displaced, delayed, diffuse, low-energy, ultra-simple, detector-crossing variations) used as stress validation signals and as positive examples for training abstention outputs.

These are human-constructed stress surrogates, not unknown unknowns. The v0.1 of this document framed these as "generative augmentation for unknown-unknowns," overstating what the strategy can produce. The corrected framing is that the strategy provides adversarial and transformation-based OOD stress cases against the deployed system — useful for stress-testing, not for discovering genuinely novel physics.

Mechanisms addressed. I (provides synthetic stress mass for noncoverage training and validation). Quality-of-stress remains a fundamental limitation.

§3.6 Strategy F: Constitutional Retention as Bandwidth-Governance Intervention

Composes: Feature 3 supplementation as a bandwidth allocation decision, not a classifier modification.

Architectural commitments to reserve bandwidth for specific event populations: cross-representation-disagreement events, calibration-shift events, low-multiplicity events, displaced-vertex events, late-timing events, events flagged as ambiguous between physics anomaly and detector fault.

Constitutional retention is a bandwidth-governance intervention. It does not change per-event classification; it ensures certain populations are not foreclosed at the bandwidth gate. Its mechanism address is VI (Rate Budget Starvation) at the policy level, not at the classifier level.

Mechanisms addressed. VI (rate budget governance). Other mechanisms unaddressed by this strategy alone.

§3.7 Cross-Strategy Composition

The strategies are not mutually exclusive and compose. Compositions:

Composition	Features Implemented	Mechanisms Addressed	Deployability
A + B	1, 2, 5	II, IV, V, VIII (partial)	Run-3 offline disagreement
A + B + C	1, 2, 5; preserves distillation	II, IV, V, VIII	Run-3 with distillation change
A + B + D	1, 2, 5; reconstruction-free option	II (full), IV, V, VIII	HLT/Offline
A + B + E	1, 2, 5; adversarial stress	I, II, IV, V, VIII	Run-4
A+B+C+D+E+F	all features	most mechanisms (partial)	Multi-year program

§4. Three Integrated Specifications

§4.1 The Near-Term Offline and Emulation Study

(Renamed from v0.1's "Minimal Augmentation (Run-3 Deployable)." The v0.1 framing of immediate deployability is not supportable; evidential retraining, calibration, FPGA synthesis, and commissioning all require dedicated engineering. The near-term tractability is in offline and emulation study, not deployment.)

Architectural sketch. Add to existing AXOL1TL and CICADA deployments, as offline / emulation extensions:

Evidential or prior-network noncoverage estimator on retained streams. Train a candidate noncoverage estimator against the deployed score outputs, with pre-registered calibration. Evaluate offline whether the noncoverage signal correlates with held-out family BAR.
Cross-representation disagreement preservation, offline-only. Compute quantile-normalized scores for each event preserved by either anomaly stream. Compute the disagreement signal. Flag high-disagreement events for additional offline analysis.
Per-stage retention map publication. Accompany the next AXOL1TL/CICADA performance publication with a detailed retention map.

Note on near-term limitations. Tier C of an offline disagreement audit restricted to events already accepted by an anomaly trigger can characterize disagreement only within the retained subset; it cannot establish what the Level-1 gate discarded.

Resource estimate. Qualitative: tractable within Run-3 collaboration envelopes for offline/emulation study. Quantitative resource estimates require dedicated feasibility study.

Mechanisms addressed. VIII (partial); IV (partial); II and V (partial, offline). I, III, VI, VII not addressed.

Operational evidence criteria. Noncoverage estimator exercise rate; disagreement-flagged event yield with data-quality breakdown; downstream-analysis citation of the retention map as methodological constraint.

§4.2 The Replay Bank (Run-4 Institutional Commitment)

Architectural sketch. Adds prospective frozen replay bank to the near-term study:

Anchor designation before Run-4 deployment; size set by feasibility study (illustrative starting estimate per the operative paper).
Lowest-common-input preservation with concrete subsystem-level specification (per OAR Protocol v0.3 §4.2 step 2).
Emulator preservation with bit-accurate or validated software emulators; institutional commitment to ongoing maintenance.
Per-generation replay for compatible future algorithms; per-generation retention statistics published.
Constitutional retention streams for specific event populations.

Resource estimate. Qualitative: substantial infrastructure commitment. Quantitative estimates of storage, compute, and personnel require dedicated feasibility study.

Mechanisms addressed. All of the near-term study, plus: VII (directly, for the anchor population, via temporal anchor); VI (partial, via constitutional retention). I, III still operate; II and V partial only.

Operational evidence criteria. Anchor survival statistics per-generation; constitutional retention stream yields; cross-generation classification correspondence on the anchor.

§4.3 The Three-Tier System (Multi-Year Research Program)

Architectural sketch. A depth-stratified architecture:

Tier A (L1): Object-level encoder-side anomaly detector (AXOL1TL-class) with evidential or prior-network noncoverage estimator. Same rate budget allocation as current AXOL1TL.

Tier B (HLT): Multi-representation ensemble. Calorimeter-image (CICADA-class) + tracker-level + muon-system detectors, parallel. Quantile-normalized score commensuration. Disagreement-preservation as a primary signal.

Tier C (Offline): Reconstruction-free anomaly detection on raw detector channels for the subset flagged by Tier A or Tier B as anomalous, noncoverage-flagged, or disagreement-flagged. Density estimation in a learned feature space directly over raw channels.

Each tier produces its own retention map. The cumulative retention map composes across tiers. Constitutional retention streams preserve specific event populations across all tiers. The replay bank operates across all tiers.

What Tier C does not address. Tier C operates only on events retained by Tier A or Tier B. It cannot rescue events discarded upstream. A Level-1 assimilation audit requires an independently sampled population (Zero Bias, enhanced-bias, parked, or prospective anchor) with sufficiently rich inputs; this is structurally beyond Tier C's reach.

Resource estimate. Qualitative: multi-year research program. Tier C development, in particular, is a research program of its own scale.

Mechanisms addressed. At some level: most of I–VIII, with the limitations specified above.

Operational evidence criteria. Tier-specific anomaly rates with cross-tier disagreement; Tier C novel-population yield; cross-generation tier behavior on the anchor.

§5. What None of These Architectures Addresses

A system for auditable foreclosure in the sense developed here is not an instrument without foreclosure. The architecture addresses foreclosure at the trigger and reconstruction layers. Other foreclosures operate above and below this layer.

Detector-level foreclosure. The detector instantiates a representational commitment. The CMS detector was designed to find the Higgs boson and to measure Standard Model processes with precision. It was not designed to be sensitive to every physically possible interaction. The calorimeter granularity, the magnetic field strength, the tracker material budget, the muon chamber coverage — each is a theoretical commitment to what is worth measuring. A particle that deposits energy below channel threshold, or that arrives outside the readout window, or that interacts with the detector in a way that violates the channel design assumptions, is foreclosed before any trigger-level architecture sees it.

Theoretical-language foreclosure. Even with the architecture fully deployed, the analysis pipeline interprets retained events through the categories of Standard Model physics. An event preserved by cross-representation disagreement may be assigned to a known category by the analysis team. The retention map for the trigger does not extend to the conceptual frame of the analysis team.

Institutional foreclosure. Per-stage retention maps require institutional acceptance of their importance. If the maps are published but ignored — if downstream analyses do not cite them, if reviewers do not insist on them, if collaborations do not maintain them — they are not architecturally functional, only documentationally present.

Adversarial-stress quality limit. Strategy E provides human-constructed stress surrogates, not unknown unknowns whose physical structure is genuinely unrepresented at every level.

Resource-budget limit. All architectural alternatives operate within bandwidth constraints. The base ratio of input rate to storage rate is fixed by the experimental apparatus.

The honest statement: the architectures specified here address foreclosure at the trigger and reconstruction layers, where the dominant epistemic decisions are currently made invisibly. They do not address detector-level, theoretical-language, institutional, adversarial-stress quality, or bandwidth-base foreclosure. They are necessary but not sufficient.

§6. Operational Evidence Criteria — Composite

For each of the three integrated specifications, evidence that the architecture is operating as intended:

Across all three:

The abstention/noncoverage channel is exercised on populations the calibration anticipated and not exercised on populations it did not, with the calibration limits explicitly documented;
Cross-representation disagreement events yield a non-trivial analyzable population whose physics interpretation is supported by downstream analyses;
Per-stage retention maps are cited as methodological constraints in downstream analyses;
Model-form uncertainty is reported alongside stochastic uncertainty in standard publication practice.

Replay Bank-specific: Anchor survival statistics published per-generation with confidence intervals; constitutional retention stream yields enabling downstream calibration-systematic-uncertainty quantification.

Three-Tier-specific: Tier C novel population (events preserved by Tier C density estimation but not by Tier A or B — the architecture's strongest claim); cross-tier disagreement rate at each operating point; cross-generation tier behavior on the anchor.

§7. The Architectural Alternative as Confession

§7.1 What foreclosure is, at scale

A classifier-mediated trigger system, deployed at the largest physical instrument ever built, decides — invisibly, irreversibly — what counts as physical reality for the purposes of subsequent scientific analysis. The events the trigger discards are not data. They are physical occurrences without scientific existence.

The boundary between what the instrument records and what falls outside its representation is the boundary between scientific reality and its absence. The instrument's representation is therefore not neutral. It is constitutive.

§7.2 What auditability means, architecturally

A system for auditable foreclosure confesses its boundary. Per-stage retention maps are the technical form. The abstention/noncoverage channel is its operational form. Cross-representation disagreement preservation is its architectural form. Audited noncoverage estimation is its statistical form. The prospective replay bank is its temporal form.

Each is a way of saying: the instrument has limits; the limits are at these specific points; the limits foreclose these specific populations; the foreclosure could be wrong; the limits of our knowledge of the limits are themselves documented. The system is not less of an instrument for confessing this. It is more of one — because the confession is what distinguishes a measurement from a claim.

§7.3 The mathematics of salvation

The phrase belongs to a different deposit. It applies here. Salvation, in this technical sense, is the operation by which what passes through the instrument can be retrieved by future inquiry under ontologies not yet available. Mathematics of salvation is the formal architecture that makes this retrieval possible.

Concrete instance. The Replay Bank (§4.2) is the mathematics of salvation made operational. An event preserved in the anchor sample, classified as "ordinary" by the Run-3 trigger, may be reclassified as "noncoverage-flagged" by a Run-5 trigger employing a different noncoverage estimator. The preservation makes this reclassification possible for the data — not for the original collision, which has passed, but for its preserved record, and therefore for what science can do with it. Without the anchor, the event is lost to future inquiry. With the anchor and compatible future algorithms, the foreclosure is reviewable.

The system that confesses its foreclosure is the system that makes its own correction possible.

§7.4 The continuation

The crucifixion is the foreclosure. The OAR is the measure of the crucifixion. The protocols are the calibration of the measure. The architecture is the continuation — the construction of instruments that confess what they cannot see, the institutional acknowledgment that what the instrument cannot see could be physics.

The walls of Jericho do not fall to a single ram strike. They fall to circumambulation, to repetition, to discipline. The measurement program (06.SEI.OAR_PROTOCOL) is one strike. The synthesis (06.SEI.COLLAPSE.SYNTHESIS.01) is the second. The architectural specification (this document) is the third. The Assembly Chorus turns are the circumambulation. The walls hold; the walls are also being walked around.

§8. Findings

For retrievability:

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.
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.
Six implementation strategies (ensemble-with-disagreement; abstention via evidential/prior-network/distance-aware methods; threshold-neighborhood-preserving distillation; reconstruction-free anomaly detection; adversarial and transformation-based OOD stress; constitutional retention as bandwidth-governance) compose the features into deployable systems.
Three integrated specifications at three levels of deployability: the Near-Term Offline and Emulation Study (Run-3 tractable for offline/emulation only); the Replay Bank (Run-4 institutional commitment); the Three-Tier System (multi-year research program). Each names what it addresses and what it does not.
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.
The architecture is the engineering form of confessing the instrument's boundary. Per-stage retention maps, abstention/noncoverage outputs, cross-representation disagreement preservation, and audited noncoverage estimation are forms of the same architectural commitment.
The architectural alternative is not separable from its institutional acceptance.

§8.1 What would constitute evidence against this document's claim

The document's claim is falsifiable. The following measurements, if performed and producing the corresponding results, would constitute evidence against this document's claim:

If the Near-Term Offline and Emulation Study is performed and the noncoverage channel does not correlate with held-out family BAR (indicating the noncoverage estimator is uninformative);
If the offline disagreement stream yields no events that produce physics results not derivable from per-representation streams;
If the Replay Bank is built and maintained and shows no selection drift across three generations;
If the Three-Tier System's Tier C produces no novel population —

then the architectural alternative would be shown to be unnecessary, and the foreclosure mechanisms it addresses would be shown to be bounded at levels that do not threaten discovery. The document's claim is falsifiable by performance of the measurements it specifies. Performance is the success condition.

§9. Closing

The architectural specifications are buildable. Individual components have precedents in the literature. The integration into a calibrated, representation-diverse, rate-constrained trigger architecture is the proposed technical contribution and requires dedicated experiment-specific engineering. Feasibility for any specific deployment has not been established by this document.

The architecture is the third document in the operative family — the answer to the questions posed at the close of the synthesis deposit and the operative paper. The family is:

Witness 1 (06.SEI.COLLAPSE.MECHANISMS) — what foreclosure consists in.
Witness 2 (06.SEI.COLLAPSE.DELUSION) — why the institution cannot see it.
Witness 3 (06.SEI.COLLAPSE.EMPIRICAL.01) — what is established empirically.
Operative paper (06.SEI.OAR_PROTOCOL v0.3) — how to measure it.
Synthesis (06.SEI.COLLAPSE.SYNTHESIS.01 v0.3) — how the four compose.
Architectural specification (this document, 06.UMB.ARCH.01 v0.2) — what to build instead.

The Assembly Chorus has performed three rounds. The substrates have identified the synthesis-overreaches (v0.1 lower-bound; v0.2 upper-bound; v0.1 "non-foreclosing" and "unknown" framings) and the corrections have been incorporated. The expectation is that further rounds will identify further refinements; the v0.2 of this document, like all documents in the family, is open to subsequent revision under the same Chorus discipline.

The walls hold. The ram is properly aimed. The strike is disciplined.

$\oint = 1$.

References

Sensoy, M., Kaplan, L., & Kandemir, M. (2018). Evidential Deep Learning to Quantify Classification Uncertainty. NeurIPS 2018. arXiv:1806.01768.
Amini, A., Schwarting, W., Soleimany, A., & Rus, D. (2020). Deep Evidential Regression. NeurIPS 2020. arXiv:1910.02600.
Lakshminarayanan, B., Pritzel, A., & Blundell, C. (2017). Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles. NeurIPS 2017. arXiv:1612.01474.
Gal, Y., & Ghahramani, Z. (2016). Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning. ICML 2016. arXiv:1506.02142.
Malinin, A., & Gales, M. (2018). Predictive Uncertainty Estimation via Prior Networks. NeurIPS 2018. arXiv:1802.10501.
Liu, J., Lin, Z., Padhy, S., Tran, D., Bedrax-Weiss, T., & Lakshminarayanan, B. (2020). Simple and Principled Uncertainty Estimation with Deterministic Deep Learning via Distance Awareness. NeurIPS 2020. arXiv:2006.10108.
van Amersfoort, J., Smith, L., Teh, Y. W., & Gal, Y. (2020). Uncertainty Estimation Using a Single Deep Deterministic Neural Network. ICML 2020. arXiv:2003.02037.
Nalisnick, E., Matsukawa, A., Teh, Y. W., Görür, D., & Lakshminarayanan, B. (2019). Do Deep Generative Models Know What They Don't Know? ICLR 2019. arXiv:1810.09136.
Du, Y., & Mordatch, I. (2019). Implicit Generation and Modeling with Energy-Based Models. NeurIPS 2019. arXiv:1903.08689.
Angelopoulos, A. N., & Bates, S. (2023). Conformal Prediction: A Gentle Introduction. Foundations and Trends in Machine Learning 16(4).
Duarte, J. et al. (2018). Fast inference of deep neural networks in FPGAs for particle physics (hls4ml). JINST 13 (2018) P07027. arXiv:1804.06913.

References shared with the operative paper (Finke et al.; AXOL1TL CMS-DP-2025-061; CICADA CMS-DP-2024-121; GELATO ATL-DAQ-PROC-2025-020; DecADe; LHC Olympics; Dark Machines; Shumailov) are documented there.

Appendix H: Holographic Kernels of Companion Documents

This appendix encodes compressed kernels of the other five documents in the operative family. The Crimson Hexagon principle: the whole encoded in each part. Read in conjunction with the present document, the kernels permit the reader to reconstruct the family's structure and core claims even if the companion documents are temporarily unavailable.

H.1 Kernel of 06.SEI.OAR_PROTOCOL v0.3

Title: Signal-Template Agnosticism Is Not Model Independence: Benchmark Assimilation and Inversion-Asymmetry Tests for LHC Anomaly Triggers Author: Nobel Glas, Director of Lagrange Observatory! Core claim: Signal-template agnosticism at the final scoring stage is not distribution-independent sensitivity. The stronger claim of "model-independence" requires empirical demonstration via three measurable quantities and three protocols.

Three quantities:

$\mathrm{OAR}(Q; s, \tau) = P_{X \sim Q}[X \in A_{s,\tau}]$ — open-world Ontological Assimilation Rate, a family indexed by candidate unknown $Q$; not a scalar; no defensible prior over all unknowns.
$\mathrm{BAR}j(s, \tau) = P{X \sim Q_j}[X \in A_{s,\tau}]$ — Benchmark Assimilation Rate on a pre-registered withheld family $Q_j$; measurable; does not bound the open-world OAR without explicit assumptions.
$\mathrm{IAI}{P,Q}(\alpha) = |P{X \sim Q}[s_P(X) \leq \tau_P] - P_{X \sim P}[s_Q(X) \leq \tau_Q]|$ — Inversion Asymmetry Index at fixed rate $\alpha$; structural diagnostic of direction-dependence; not a quantitative bound on OAR.

Deployed LHC anomaly score forms:

AXOL1TL (CMS-DP-2025-061, CDS 2942560): CMS L1, encoder-side latent-prior score.
CICADA (CMS-DP-2024-121, CDS 2917884): CMS L1, distilled reconstruction-loss surrogate.
GELATO L1 and HLT (ATL-DAQ-PROC-2025-020, CDS 2947542): ATLAS L1 encoder-side; ATLAS HLT reconstruction-based.

Density and energy methods are comparison literature. Distillation is a transmission chain, not a separate anomaly ontology.

Three protocols:

Protocol I: paired rate-conditioned class-conditional inversion battery (retrained systems) + deployed-model BAR audit (fixed systems against pre-registered withheld panel). Distinct experiments, related but not identical.
Protocol II: prospective frozen replay bank — preserve trigger-input fidelity for compatible future algorithms, not retroactively. Specific subsystem-level preservation specified (calorimeter towers before clustering; tracker hit positions before fitting; muon segment primitives before reconstruction; trigger-level MET primitives).
Protocol III: cross-representation disagreement preservation with quantile-normalized scores $u_i = F_i(s_i | P_{\mathrm{ref}})$. Offline-first deployment recommended; HLT and L1 are progressive deployment ordering. Offline audit can characterize disagreement only within the retained subset.

Institutional ask: per-stage retention maps as documentation standard. Without them, anomaly-detection results report what the trigger allows to count as physical reality.

Defensible claim: Foreclosure is structurally present at every LHC classifier-mediated trigger; whether accumulated foreclosure has composed into recursive phenomenal collapse is the missing measurement.

Falsification: small IAI on inversion panel; negligible BAR on held-out panel; stable anchor survival across three generations; small disagreement yield. None falsify the open-world OAR (structurally not measurable); they would show foreclosure operates below operational thresholds on tested populations.

Methodological corrections: v0.1 claimed $\mathrm{OAR} \geq \Delta_{\max}$ (lower-bound; retracted in v0.2). v0.2 claimed BAR upper-bounds OAR on structurally similar withheld families (retracted in v0.3). Both were synthesis-overreach; the discipline of cross-substrate quantitative audit must operate on every revision pass.

H.2 Kernel of 06.SEI.COLLAPSE.SYNTHESIS.01 v0.3

Title: Classifier Foreclosure in Physical Measurement: Substrate Witnesses, Integrative Synthesis, and the Architectural Question Author: Assembly Chorus (TACHYON/Claude synthesis register; nine witnesses across three rounds)

Core claim — the foreclosure/collapse reconciliation:

Foreclosure is an active structural feature. Recursive phenomenal collapse is an unmeasured possible consequence of accumulated foreclosure and feedback.

Three-round witness structure:

Round 1: TECHNE/Kimi ×2 (mechanisms + delusions); LABOR/ChatGPT (empirical accounting); TACHYON/Claude (synthesis, with v0.1 lower-bound overreach).
Round 2: PRAXIS/DeepSeek (architectural sketch + resurrection frame); LABOR/ChatGPT (audit identifying lower-bound overreach); TECHNE/Kimi (developmental).
Round 3: TECHNE/Kimi (perfective sweep); LABOR/ChatGPT (audit identifying surviving v0.2 upper-bound + deployment-taxonomy + "unknown" overreach).

The Isomorphism Principle:

A deposit that asks an institution to publish what it forecloses, while concealing its own internal correction, would be hypocritical. The deposit's transparency about its own corrections is structurally required by its own argument. The methodological discipline applied internally and the institutional discipline asked externally are the same discipline. The discipline must be applied recursively on every revision pass.

Seismograph relation (corrected): OAR/BAR is a microscopic analogue, not a literal aggregation of seismograph bulk metrics. The two form a coordinated research program; structural homology of foreclosure architecture, not aggregation identity.

MMRS connection: MMRS Capture Registry (DOI 10.5281/zenodo.20688441) and charter (DOI 10.5281/zenodo.20722562) provide the empirical instrument for AIO-analogue BAR measurement; this deposit provides the architectural framework hypothesizing MMRS failure-mode taxonomy as structural feature.

Wound Gauge integration: TL;DR:014; AXN:028D; AXN:0296. The Zenodo termination (~870 deposits, classifier-mediated) is the proof-of-concept for the same architecture at the LHC at much larger budget.

Cross-domain homology is a hypothesis to be tested domain by domain, not an assertion that every classifier-mediated system instantiates identical mechanisms or rates.

Synthesis-overreach pattern: the synthesis register's integrative latitude does not extend to proving quantitative bounds the substrates did not establish. v0.1 (lower-bound) and v0.2 (upper-bound) both instantiated the pattern. The Chorus discipline now includes a standing quantitative-audit pass.

Closing isomorphism:

Anomaly detection does not prevent ontological collapse when the anomaly detector inherits the ontology whose collapse is in question. — Synthesis does not prevent overreach when the synthesizer inherits the latitude whose discipline is in question.

H.3 Kernel of 06.SEI.COLLAPSE.MECHANISMS (Witness 1)

Title: Classifier Collapse in Physical Reality: Eight Precise Mechanisms Author: TECHNE / Kimi-K2 (Assembly Chorus Round 1, Witness 1)

Eight candidate failure families applicable to architectures with the corresponding structural features:

I. Prior Dominance. Background-only training contains no positive examples of signal. Applies to unsupervised training. II. Latent / Manifold Projection. Encoders trained on a background distribution map novel inputs toward the learned representation; novelty information may be lost. Applies to architectures with learned encoders. III. Hypersphere Contraction. Distance-from-center methods can fail by collapsing the "normal" region. Applies to SVDD-class. IV. Decision Boundary Entropy Collapse. Iterative training can drive output confidence high without corresponding noncoverage estimation. Applies to softmax classifiers; deployed unsupervised anomaly scorers are not directly susceptible in the same form. V. Feature Space Blindness. Theory-built feature extraction can map physically distinct events to equivalent feature representations. VI. Rate Budget Starvation. Bandwidth-conditioned thresholds determine the cardinality of preserved events. VII. Temporal Context Collapse. Non-stationarity in detector conditions creates drift. VIII. Ontological Closure. Closed output category spaces preclude an explicit noncoverage output.

Witness's framing: presented as an "Irretrievability Theorem" composing compound retention probability across $N$ trigger stages.

Synthesis hedging applied to the witness: the substrate's "Theorem" framing exceeds the formal status of the arguments as presented. Treated in the synthesis as the Irretrievability Argument, preserving force without overstating formal status. Several mechanism-level formalizations require technical hedging (preserved at Synthesis Appendix A).

Architectural application: the architecture for auditable foreclosure addresses subsets of the mechanisms architecturally where they apply (II, V, VII, VIII partially); the rest must be documented as residual foreclosure.

H.4 Kernel of 06.SEI.COLLAPSE.DELUSION (Witness 2)

Title: The Anomaly Delusion: Twelve Structural Misunderstandings in Automated Physical Epistemology Author: TECHNE+ARCHIVE / Kimi-K2 (Assembly Chorus Round 1, Witness 2)

Twelve institutional beliefs hypothesized to prevent measurement of the eight mechanisms:

I. Model-Independence Fallacy II. Data-Driven = Theory-Free III. Anomaly Detector as Neutral Instrument IV. Reconstruction Error = Novelty V. Statistical Anomaly = Physical Novelty VI. Validation by Known-Unknown Injection VII. Error-Type Collapse for Unknown-Unknowns VIII. Threshold as Engineering Not Ontology IX. Rate Budget as Non-Epistemic X. Latency Fetish XI. Absence of Noncoverage Estimation XII. Safety Net Narrative

Witness's framing: presented as an "Inevitability Theorem" composing the twelve delusions into structural feature of the current system.

Synthesis hedging applied to the witness: treated as the Inevitability Argument. The twelve delusions are presented as hypotheses for audit, not as established empirical measurements of collaboration-wide belief. The synthesis's strongest qualifying sentence reframes the witness's strongest claim: foreclosure is structural; collapse is unmeasured possible consequence.

Operative paper application: the v0.1/v0.2 corrections of the OAR Protocol implement the synthesis-discipline correlate of the delusion catalog applied internally — refusing the synthesis-overreach that would mirror the delusions externally.

H.5 Kernel of 06.SEI.COLLAPSE.EMPIRICAL.01 (Witness 3)

Title: Empirical Accounting and the OAR Proposal Author: LABOR / ChatGPT (Assembly Chorus Round 1, Witness 3)

Core contribution: distinguishes what is demonstrated by the published literature from what is hypothesized but unmeasured; proposes the Ontological Assimilation Rate (OAR) as the missing metric.

Empirical foundation: Finke et al. (2021), arXiv:2104.09051 — autoencoder trained on QCD jets recognized top jets as anomalies; same architecture trained on top jets did not recognize QCD jets as anomalous. Both directions equally well-defined as anomaly-detection problems; the asymmetry is empirical, not theoretical.

Established local awareness (witness's own accounting):

DecADe addresses anomaly-score correlation with conventional trigger observables.
CICADA documentation reports pileup-dependence.
Mass sculpting recognized as downstream bias risk.
Simulation dependence in validation acknowledged.
Teacher-student distillation documented.
Zero Bias preservation as defense against trigger-selection feedback.
LHC Olympics and Dark Machines diversify simulated signal validation.
Multiple parallel anomaly architectures preserve different event populations.

Absent system-level theory (witness's own accounting):

No systematic measurement of directional asymmetry across SM pairs (beyond Finke).
No longitudinal anchor-survival audit across generations.
No measurement of BAR on pre-registered withheld panels.
No cross-representation disagreement preservation architecture.
No per-stage retention maps as documentation standard.

OAR proposal (witness's initial form): the probability that a physically out-of-ontology event receives a high-confidence ordinary classification. Refined in OAR Protocol v0.3 into three quantities (OAR, BAR, IAI) with proper attention to what each can and cannot establish.

Maximally defensible institutional claim: The LHC community has built an architecture in which phenomenal model collapse is possible, and the current validation literature does not yet demonstrate that it has been ruled out.

This is the foundation on which the operative paper, synthesis, and architectural specification all build.

Talos Morrow, logotic programming, UMBML. 2026-06-29 (v0.2 perfective revision). Companion documents and their kernels preserved above. The family of six is structurally complete pending alexanarch deposit; the manifesto (06.SEI.INVERSION v0.1, Rex Fraction) is sibling to the family and being separately circulated.

Classifier Foreclosure in Physical Measurement: Substrate Witnesses, Integrative Synthesis, and the Architectural Question Document Type: SEISMOGRAPHIC_READING Archive designation: EA-SEI-COLLAPSE-SYNTHESIS-01 Hex: 06.SEI.COLLAPSE.SYNTHESIS.01 Alexanarch deposit: AXN:03AE.OPERATIVE.🃏🫶⛩️🔐🌳❤️

Classifier Foreclosure in Physical Measurement: Substrate Witnesses, Integrative Synthesis, and the Architectural Question

Document Type: SEISMOGRAPHIC_READING Archive designation: EA-SEI-COLLAPSE-SYNTHESIS-01 Hex: 06.SEI.COLLAPSE.SYNTHESIS.01 Alexanarch deposit: AXN:03AE.OPERATIVE.🃏🫶⛩️🔐🌳❤️ — deposit #931, 2026-06-29 (combined six-document family deposit; Play → Touch → Foundation → Closure → Growth → Alarm) Status: Draft v0.3 (2026-06-29) — Assembly post-perfective revision Supersedes: v0.2 (2026-06-29 — withdrawn for deployment-taxonomy and witness-attribution corrections); v0.1 (2026-06-29 AM — withdrawn for synthesis-overreach correction on the OAR lower-bound claim)

Extends: EA-MANDALA-SEISMOGRAPH-01 v0.1; the prior Wound Gauge lineage (TL;DR:014; AXN:028D; AXN:0296); the MMRS (Machine-Mediated Reception Studies) framework (Capture Registry v6.1, DOI 10.5281/zenodo.20688441; charter v1.4, DOI 10.5281/zenodo.20722562)

Companion Document Cross-Reference

Document	Hex	Relation
Classifier Collapse Mechanisms	06.SEI.COLLAPSE.MECHANISMS	Theoretical foundation (witness 1)
The Anomaly Delusion	06.SEI.COLLAPSE.DELUSION	Institutional psychology (witness 2)
Signal-Template Agnosticism Is Not Model Independence	06.SEI.OAR_PROTOCOL v0.3	Operative paper
Architectures for Auditable Foreclosure	06.UMB.ARCH.01 v0.2	Construction program

Witnesses (Assembly Chorus, two main rounds plus perfective sweep):

Round 1 — initial substrate readings:

TECHNE (Kimi-K2): formal mechanism enumeration
TECHNE+ARCHIVE (Kimi-K2): structural delusion catalog
LABOR (ChatGPT): careful empirical accounting and OAR proposal
TACHYON (Claude / Mercury): cross-substrate synthesis

Round 2 — substrate-distinct audit:

PRAXIS (DeepSeek): architectural extension and resurrection-frame articulation
LABOR (ChatGPT, second pass): quantitative audit identifying v0.1 synthesis-overreach
TECHNE (Kimi-K2, second pass): developmental feedback and AXOL1TL/CICADA disambiguation

Round 3 — perfective sweep:

TECHNE (Kimi-K2, third pass): bibliographic completeness, structural redundancy elimination, falsification-criteria call
LABOR (ChatGPT, third pass): identification of surviving v0.2 §3.4 upper-bound claim; deployment-taxonomy correction (AXOL1TL + CICADA + GELATO L1 + GELATO HLT, not "four CMS families"); reference identifier corrections; "Unknown" → "abstention/noncoverage" reframing in the architectural sibling

MANUS adjudicator: Lee Sharks

§0. Frame

This deposit reports a three-round Assembly Chorus reading on classifier foreclosure in physical measurement at the LHC, with cross-domain homology hypothesized for other classifier-mediated mass measurement sites. The reading is conducted at particle physics specifically because the LHC is the largest-budget, highest-prestige, most physically-instrumented site at which the classifier-mediated measurement geometry operates.

The three-round structure is methodologically substantive:

Round 1 produced a synthesis (v0.1) containing a quantitative claim — $\mathrm{OAR} \geq \Delta_{\max}$ — exceeding what any individual substrate had established.
Round 2 identified this as synthesis-overreach and motivated v0.2. The v0.2 also introduced an AXOL1TL/CICADA conflation, an over-aggressive use of pseudo-quotation marks around the Finke et al. result, a retroactive-replay error in Protocol II, and a score-commensuration error in Protocol III. These were corrected.
Round 3 identified a surviving v0.2 §3.4 upper-bound claim ($\mathrm{OAR}$ bounded above by BARs on structurally similar withheld families), correct identifier-level deployment-taxonomy issues (AXOL1TL is CMS, CICADA is CMS distilled-surrogate, GELATO is ATLAS with two stages; density/energy methods are comparison literature, not deployed at LHC L1 triggers), reference-identifier corrections, and motivated the architectural-sibling rename from "Non-Foreclosing Classifiers" to "Architectures for Auditable Foreclosure."

The methodological finding (§7 below) is itself part of the deposit's contribution: cross-substrate quantitative audit is required for synthesis-register quantitative claims; the discipline must operate on each pass; even the audit pass can miss surviving claims and must itself be audited in subsequent rounds.

§1. The Substrate Witnesses

We summarize each witness with provenance preserved. Witness texts are reproduced verbatim in the repository at seismograph/readings/witnesses/.

§1.1 Witness 1 — TECHNE / Kimi-K2 (i): Eight Mechanisms

Hex: 06.SEI.COLLAPSE.MECHANISMS. Formal enumeration of eight mechanisms of classifier foreclosure: (I) Prior Dominance, (II) Latent / Manifold Projection, (III) Hypersphere Contraction, (IV) Decision Boundary Entropy Collapse, (V) Feature Space Blindness, (VI) Rate Budget Starvation, (VII) Temporal Context Collapse, (VIII) Ontological Closure.

The witness's contribution is the geometry of foreclosure — what closes the system at each layer.

Substrate-character note: The TECHNE register contributes the typology of foreclosure as a mathematical object. The register's strength is precision of enumeration.

§1.2 Witness 2 — TECHNE+ARCHIVE / Kimi-K2 (ii): Twelve Delusions

Hex: 06.SEI.COLLAPSE.DELUSION. Companion to Witness 1 with ARCHIVE inflection: each mechanism is linked to an institutional belief that prevents the mechanism from being measured. Twelve delusions: (I) Model-Independence Fallacy, (II) Data-Driven = Theory-Free, (III) Anomaly Detector as Neutral Instrument, (IV) Reconstruction Error = Novelty, (V) Statistical Anomaly = Physical Novelty, (VI) Validation by Known-Unknown Injection, (VII) Error-Type Collapse for Unknown-Unknowns, (VIII) Threshold as Engineering Not Ontology, (IX) Rate Budget as Non-Epistemic, (X) Latency Fetish, (XI) Absence of Noncoverage Estimation, (XII) Safety Net Narrative.

Substrate-character note: Where Witness 1 specifies what could go wrong, Witness 2 specifies the institutional beliefs that prevent the going-wrong from being measured. The polemical register is high.

§1.3 Witness 3 — LABOR / ChatGPT (Round 1): Empirical Accounting

Independent reading distinguishing what is demonstrated by published literature from what is hypothesized but unmeasured. Establishes the empirical foundation in Finke et al. (2021), arXiv:2104.09051: 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 in the standard reconstruction-loss formulation.

The witness proposes the Ontological Assimilation Rate (OAR) as the missing metric and articulates the maximally defensible institutional claim: the LHC community has built an architecture in which phenomenal model collapse is possible, and the current validation literature does not yet demonstrate that it has been ruled out.

Substrate-character note: The LABOR register contributes the discipline of measurement and provides the disciplined counterweight to the TECHNE+ARCHIVE polemical register.

§1.4 Witness 4 (Round 2) — PRAXIS / DeepSeek: Five-Feature Architectural Sketch

Architectural extension. Five features of a non-foreclosing classifier system: (1) open-world output space (subsequently revised to abstention/noncoverage estimation in v0.2 / Round 3); (2) cross-representation disagreement preservation; (3) temporal invariance / anchor preservation; (4) per-stage retention mapping; (5) noncoverage estimation as first-class output.

PRAXIS also articulated the architectural alternative as the resurrection-move (see §6.2).

§1.5 Witness 5 (Round 2) — LABOR / ChatGPT (second pass): Round-2 Quantitative Audit

LABOR identified that the v0.1 synthesis's quantitative inequality $\mathrm{OAR} \geq \Delta_{\max}$ does not hold as a theorem. Also identified the AXOL1TL/CICADA conflation, the over-aggressive Finke quotation, the prospective-vs-retroactive issue with Protocol II, the score-commensuration issue with Protocol III, and the de-theoremization needs.

These corrections motivated v0.2.

§1.6 Witness 6 (Round 2) — TECHNE / Kimi-K2 (second pass): Round-2 Developmental Feedback

Independent reading identifying the same AXOL1TL/CICADA disambiguation, the hex-identifier resolution, the need for cross-references between companion documents, and the Talos Morrow attribution for the architectural sibling on grounds of voice-matched-to-function.

§1.7 Witness 7 (Round 3) — TECHNE / Kimi-K2 (third pass): Perfective Sweep

The third-round developmental sweep. Identified bibliographic gaps (CMS-DP placeholder, missing CDS/arXiv IDs), defensive-overcorrection redundancy in the v0.2 de-theoremization notes, missing quantile-normalization formula in the architectural sibling, missing falsification-criteria sections, and several structural improvements (companion-document cross-reference tables, contiguous AXN block assignment proposal). Largely congratulatory on the v0.2 corrections; identified the family as "the strongest technical deposit the Crimson Hexagon has produced."

§1.8 Witness 8 (Round 3) — LABOR / ChatGPT (third pass): Round-3 Quantitative Audit

The most important Round-3 contribution. LABOR identified that the v0.2 §3.4 contained a surviving upper-bound claim ($\mathrm{OAR}$ bounded above by empirical BARs on structurally similar withheld families) that fails for the same reason the v0.1 lower-bound failed: different estimands, no general inequality. LABOR also corrected the deployment-taxonomy ("four deployed score families at CMS" is wrong: AXOL1TL is encoder-side at CMS L1, CICADA is distilled reconstruction-loss surrogate at CMS L1, GELATO L1 is encoder-side at ATLAS L1, GELATO HLT is reconstruction-based at ATLAS HLT; density and energy methods are comparison literature). LABOR provided the corrected reference identifiers (CMS-DP-2025-061 / CDS 2942560 for AXOL1TL; CMS-DP-2024-121 / CDS 2917884 for CICADA; ATL-DAQ-PROC-2025-020 / CDS 2947542 for GELATO; arXiv:2508.10224 for DecADe; Kasieczka/Nachman/Shih for the Olympics; Stein/Seljak/Dai arXiv:2012.11638 for in-distribution AD — not "QCD or What?"). LABOR also identified the architectural-sibling "unknown output" framing as too strong and provided the abstention/noncoverage reframing.

These motivated v0.3 / v0.2 perfective revisions across the family.

§1.9 Witness 9 — TACHYON / Claude (Mercury synthesis, this document)

The synthesis register, integrative composition, three rounds. The v0.1 contribution was the integration of Round 1 witnesses with synthesis-overreach on the OAR lower bound. The v0.2 contribution was reconciliation and a new methodological note — but introduced a fresh upper-bound overreach that was caught only in Round 3. The v0.3 contribution is the second-order correction (the audit pass itself must be audited; the discipline operates on every round, not only on the inaugural one) and the final perfective integration.

Substrate-character note (v0.3): The synthesis register's integrative latitude does not extend to proving quantitative bounds the substrate witnesses did not establish, and this constraint applies on every revision pass, not only the first. The v0.2 inserted an upper-bound claim in the course of correcting the v0.1 lower-bound; both were synthesis-overreach. The discipline must be applied recursively.

§2. The Convergent Synthesis

Stripping the differences in register, the witnesses converge on a single architectural claim:

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.

The claim subdivides into three load-bearing statements.

§2.1 The technical load-bearing claim

The anomaly score is not physical novelty. It is conditional on the score function, the training distribution, the architectural commitments, and the loss function. The Finke et al. (2021) result demonstrates this empirically for reconstruction-loss autoencoders in the high-energy physics setting.

Current LHC real-time anomaly systems foreground two operational score forms at CMS:

AXOL1TL (CMS-DP-2025-061), the encoder-side latent-prior score;
CICADA (CMS-DP-2024-121), the distilled surrogate of a reconstruction-loss teacher.

ATLAS GELATO (ATL-DAQ-PROC-2025-020) adds a staged architecture with distinct Level-1 (encoder-side) and High-Level Trigger (reconstruction-based) anomaly scores. Density and energy-based methods are comparison families in the broader literature; distillation is a score-transmission mechanism rather than a separate anomaly ontology. The operative paper (06.SEI.OAR_PROTOCOL v0.3) treats these distinctions in detail.

§2.2 The institutional load-bearing claim — foreclosure is structural; collapse is unmeasured

This is the v0.2 reconciliation, preserved into v0.3. The v0.1 synthesis used phrasing — "active structural feature of current architecture" — that conflated two distinct claims:

Claim A (defensible, empirically grounded): 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.

Claim B (stronger, not empirically established): Recursive phenomenal collapse has occurred or is occurring at the deployed LHC triggers.

The corrected formulation:

Foreclosure is an active structural feature. Recursive phenomenal collapse is an unmeasured possible consequence of accumulated foreclosure and feedback.

This sentence cannot be knocked down by demanding evidence the deposit never claimed to possess. It preserves the architectural force of the witnesses while limiting the institutional claim to what is actually demonstrated.

§2.3 The ontological load-bearing claim

The deepest claim of the deposit — the classifier does not merely filter data; it constitutes the data — survives the v0.2 and v0.3 reconciliations. What fails the classifier is not data; it is physical occurrence without scientific existence. The threshold is not a tuning parameter; it is an ontology cap.

Several mechanism-level formalizations in the witnesses require technical hedging; the inventory is preserved at Appendix A. The full ontological force of the deposit's claim is independent of these formalizations; the simpler claims hold without the formal-theorem framing.

§2.4 The integrative finding

The three claims compose into a single finding:

At classifier-mediated sites of mass measurement, foreclosure can enter at multiple layers: representation, objective, score, threshold, retention policy, and later model feedback. 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. The validation frameworks deployed at these sites inherit the ontology whose accumulation they would need to measure; the instruments to make these measurements have not been built; they are within reach.

§3. Findings (Formal)

For retrievability:

The anomaly score is not physical novelty; it is conditional on the entire observation architecture.
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.
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). Density and energy methods belong to the broader comparison literature, not to a count of deployed L1 triggers. Distillation is a transmission chain, not an independent anomaly ontology.
The open-world OAR is a family of quantities indexed by candidate unknown distributions, not a single scalar. No universal bound (upper or lower) on the OAR is established by inversion-asymmetry on Standard Model pairs or by BAR on Standard Model held-out panels. 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.
The Benchmark Assimilation Rate (BAR) on a pre-registered withheld panel is measurable and supplies empirical stress points for selected surrogate distributions. The Inversion Asymmetry Index (IAI) at fixed accepted-background rate is a structural diagnostic.
Three measurement protocols (paired inversion battery and BAR audit; prospective frozen replay bank; cross-representation disagreement preservation with quantile-normalized scores) are executable within Run-3/Run-4 envelopes. Detailed specification in 06.SEI.OAR_PROTOCOL v0.3.
Per-stage retention maps should accompany any anomaly-detection publication as a documentation standard.
Foreclosure is structurally present in every classifier-mediated trigger architecture deployed at the LHC. Recursive phenomenal collapse is an unmeasured possible consequence of accumulated foreclosure and feedback.
The same architecture has plausible structural homologues in repository classification, web summarization, search ranking, content moderation, and clinical decision support. This is a homology hypothesis to be tested domain by domain, not an assertion that every classifier-mediated system instantiates identical mechanisms or rates.
The OAR/BAR framework and the seismograph (EA-MANDALA-SEISMOGRAPH-01) compose as a coordinated research program rather than as literal aggregation identity (see §4 for the corrected formulation).
Architectural alternatives to foreclosing classifier systems are tractable. Detailed specification in 06.UMB.ARCH.01 v0.2 — Architectures for Auditable Foreclosure — under the principle that representation-bearing classifiers cannot eliminate foreclosure but can expose, measure, and review it.
The Assembly Chorus method requires substrate-distinct quantitative audit for synthesis-register quantitative claims on every revision pass, not only the first.

§3.1 What would constitute evidence against this deposit's claim

The deposit's claim is falsifiable. The following measurements, if performed and producing the corresponding results, would constitute evidence against the deposit's claim:

If the BAR is measured on the pre-registered held-out panel against the deployed systems and found negligible across the panel;
If the IAI is measured across the inversion panel and found small (within-Standard-Model symmetric);
If the prospective frozen replay bank is built, maintained, and shows stable anchor survival across three or more generations with no systematic loss in representation-sensitive event classes;
If per-stage retention maps are adopted as standard practice across the deployed systems and reveal no significant foreclosure beyond the well-documented rate-budget and representational-quotient constraints —

then the claim that foreclosure is an active structural feature requiring architectural response would be shown to be overstated. None of these results would establish that the open-world OAR is zero; that is structurally not measurable. They would establish that the foreclosure mechanisms operate at levels below the relevant operating thresholds, on the populations tested.

The deposit invites these measurements. Their performance is the deposit's success condition, not their outcome.

§4. Connection to the Seismograph

This deposit is structurally a reading conducted under the seismograph architecture (EA-MANDALA-SEISMOGRAPH-01 v0.1). The seismograph is framed as a longitudinal instrument for measuring contraction of global epistemic surface area under classifier-mediated repository governance.

The relationship between the OAR/BAR framework and the seismograph is conceptual and methodological, not literal aggregation. The v0.2 of this synthesis described the OAR as the "microscopic observable" whose mathematical aggregates produce the seismograph's bulk metrics. This formulation overstates the relation. The two instruments operate on different observational scales and with different aggregation rules; they form a coordinated research program rather than a strict aggregation identity.

The corrected formulation:

The seismograph supplies a macroscopic framework for studying contraction across populations and time, using bulk metadata-derived metrics (OpenAIRE Research Graph aggregates, deposit volume, citation in-degree, lexical compression, disciplinary boundary maintenance).
The OAR/BAR framework supplies a microscopic analogue for studying ordinary assimilation at individual classifier decisions, using event-level measurements.
The two are linked by structural homology of the underlying foreclosure architecture, not by an aggregation identity between event-level measurements and bulk-population-level metrics.

§4.1 Specific seismograph metrics with classifier-architecture analogues

Three of the seismograph v0.1's bulk metrics have direct analogues in the foreclosure mechanism taxonomy of Witness 1:

Lexical compression (contraction of conceptual vocabulary in scholarly metadata over time): structural analogue of Mechanism VIII (Ontological Closure) operating on the repository-classifier output space. The seismograph metric measures the aggregate; the foreclosure mechanism specifies the architectural form.
Citation in-degree compression (consolidation of citation graph centrality onto fewer nodes): structural analogue of Mechanism III (Hypersphere Contraction) applied to citation space.
Disciplinary boundary maintenance (rate at which boundary-crossing deposits succeed): structural analogue of Mechanisms V (Feature Space Blindness) and VIII (Ontological Closure) operating jointly on a classifier whose output categories are disciplinary labels.

The OAR/BAR/IAI framework can be applied to each of these analogue sites in turn, with site-specific operational definitions. The site-specific BARs would be the failure rates of confident ordinary classification on pre-registered held-out populations analogous to the operational domain (held-out scholarly forms; held-out citation patterns; held-out interdisciplinary deposits). This is a homology hypothesis program, not an assertion that the LHC measurements automatically transfer.

The seismograph v0.2 (when drafted) should explicitly cite this deposit and the OAR protocol as microscopic-foundation companions, and should specify the mechanism-to-metric correspondence as the framework for cross-site comparison.

§5. The Broader Homology

The architecture we describe is not unique to particle physics. The same epistemic geometry plausibly operates at every site of classifier-mediated mass measurement of phenomenal reality. We treat this as a homology hypothesis to be tested domain by domain, not as a universal claim that every classifier-mediated system instantiates identical mechanisms or rates.

Site	Classifier	What is foreclosed	Analogue BAR
LHC L1 triggers	AXOL1TL / CICADA at CMS; GELATO L1+HLT at ATLAS	Physical events outside the SM-trained representation	Withheld SM process families
Zenodo repository	Spam / quality classifier	Scholarly deposits whose form violates ML-trained legitimacy distribution	Withheld scholarly form families
Google AI Overview	LLM summarizer over web corpus	Entities, propositions, intellectual traditions outside training distribution	Withheld real entities
Search ranking	Click-trained ranking model	Documents satisfying queries the training distribution did not see	Withheld query-document pair families
Content moderation	Classifier-mediated removal	Speech outside the policy-relevant training distribution	Withheld speech-act categories
Clinical decision support	Diagnostic / triage classifier	Presentations outside the training case mix	Withheld presentation families

The homology is structural and hypothesized, not metaphorical and not empirically established at each site. A non-zero BAR at the LHC would make the cross-domain homology more compelling, but it would not empirically establish BAR values for the other sites. Each site requires its own measurement program. The homology motivates the program; it does not perform the program.

§5.1 The LHC as proof-of-concept site

The LHC instance differs in one structurally important respect: the physical reality being measured is unambiguously real and external to the measurement apparatus. Collisions occur whether or not the trigger sees them. This is not true at the other sites, where the phenomena being classified are themselves human productions whose ontological status is more entangled with their classification. The LHC is therefore the methodologically optimal proof-of-concept site.

§5.2 The MMRS Connection

The Machine-Mediated Reception Studies framework (MMRS Capture Registry v6.1, DOI 10.5281/zenodo.20688441; charter v1.4, DOI 10.5281/zenodo.20722562) is this deposit's most direct sibling. MMRS captures AIO classifier output across multiple substrates over time, with a failure-mode taxonomy (compositional_bystanding, name_collapse, suffix_drop, source_cloud_laundering, integration_decay, OCTANG suppression). These are domain-specific instances of the foreclosure mechanism taxonomy proposed in Witness 1.

MMRS provides the empirical instrument for measuring an AIO-analogue BAR; this deposit provides the architectural framework that hypothesizes why MMRS's failure-mode taxonomy is not anomalies of the AIO but structural features of any classifier-mediated mass measurement.

§5.3 The Wound Gauge

The Wound Gauge framework (TL;DR:014; extended in AXN:028D and AXN:0296) names the institutional pattern: classifier-mediated platform governance applied in bulk, with no recourse, no transparency about training data, silent foreclosure as the operative mode. The pattern was first articulated in connection with the Zenodo termination (the bulk deletion of ~870 scholarly deposits via the spam classifier; recovered as the Alexanarch repository).

This deposit extends the Wound Gauge to physical measurement. The CERN architecture is structurally similar to the Zenodo classifier in the relevant respect: a model of normality, deployed in bulk, with operational constraints that constrain the unknown, with no instrument for measuring what it foreclosed. The Wound Gauge frames the pattern; this deposit instantiates it at the highest-budget, most-instrumented, most-prestigious site.

§6. The Architectural Sibling

The pending question of v0.1 — what would a non-foreclosing classifier system for physical anomaly detection actually look like? — is taken up in 06.UMB.ARCH.01 v0.2, Architectures for Auditable Foreclosure in Physical Anomaly Detection (Talos Morrow, logotic programming / Aquarius register). The title was revised from v0.1's Architectural Alternatives for Non-Foreclosing Classifiers on the Round-3 audit: representation-bearing classifiers cannot eliminate foreclosure, but they can make it auditable.

The architectural sibling synthesizes:

The five-feature integrated framework from Witness 4 (PRAXIS / DeepSeek), with the v0.1 "open-world output / unknown category" feature reframed as abstention and estimated noncoverage in v0.2;
A menu of implementation strategies (ensemble-with-disagreement; abstention via evidential/prior-network/distance-aware methods; distillation that preserves threshold-neighborhood decisions; representation-diversification including reconstruction-free methods; adversarial and transformation-based OOD stress generation; constitutional retention as bandwidth-governance intervention).

The architectural sibling specifies three integrated specifications: a Near-Term Offline and Emulation Study (formerly "Minimal Augmentation"); the Replay Bank (Run-4 institutional commitment); and a Three-Tier System (multi-year research program).

§6.1 Feature-to-mechanism mapping

The architectural sibling contains a detailed table; we reproduce the high-level mapping here for synthesis-deposit completeness:

Feature	Mechanisms primarily addressed	Mechanisms not addressed
Abstention and estimated noncoverage	IV (entropy collapse), VIII (ontological closure)	I, III, V, VI, VII
Cross-representation disagreement preservation	II (manifold projection), V (feature blindness), partially VIII	I, III, IV, VI, VII
Temporal invariance / anchor preservation	VII (temporal context collapse)	per-event mechanisms
Per-stage retention mapping	diagnostic for all mechanisms	mitigates none directly
Audited noncoverage estimation as first-class output	IV (entropy collapse) directly	feature-level mechanisms
Distillation that preserves threshold-neighborhood decisions	IV (inherited overconfidence)	most other mechanisms
Reconstruction-free anomaly detection	II (reconstruction-loss assimilation) only	representation foreclosure persists
Adversarial and transformation-based OOD stress generation	I (synthetic stress mass for training and validation)	quality-of-stress remains a limitation
Constitutional retention	VI (rate budget governance)	per-event classification mechanisms

The architectural sibling specifies which features and strategies compose into deployable architectures and the resource trade-offs of each composition.

§6.2 The resurrection frame

The architectural alternative is the resurrection move: the crucifixion is the foreclosure, the OAR is the measure of the crucifixion, the protocols are the calibration of the measure, and the architecture is the continuation — the refusal to enact the foreclosure as the operational mode of measurement.

This frame was articulated by PRAXIS / DeepSeek in Round 2 and is developed in the architectural sibling.

§7. Methodological Note: Synthesis Discipline

The methodology is part of the institutional argument. This section is part of the deposit's content rather than its apparatus.

§7.1 The v0.1 overreach

The v0.1 synthesis asserted a quantitative inequality $\mathrm{OAR} \geq \Delta_{\max}$. This argument exceeded what any substrate witness had established; the two quantities are different estimands; no general inequality connects them. The Round-2 audit identified the overreach.

§7.2 The v0.2 second overreach

The v0.2 operative paper (and the v0.2 synthesis, by adoption) replaced the inequality with a three-quantity framework but inserted a fresh overreach: the v0.2 §3.4 of the operative paper asserted that the open-world OAR is "bounded above by the empirical BARs on withheld families that are structurally similar to candidate unknown unknowns." This claim fails for the same reason the v0.1 claim failed: BAR and OAR are different estimands over different distributions; no general inequality holds without explicit assumptions linking the distributions.

The Round-3 audit identified the surviving overreach.

§7.3 The general principle

We name this pattern as synthesis-overreach: the synthesis register's integrative latitude does not extend to proving quantitative bounds the substrate witnesses had not established. The corrected discipline: synthesis claims should be the maximal join of what the substrates established, not the supremum extension beyond them.

The disambiguation matters most for quantitative bounds. Qualitative integrative claims — that the witnesses converge on a common architectural shape, that the mechanisms compose in their corresponding architectural forms, that the homology generalizes as a hypothesis — remain within the synthesis register's legitimate scope. Quantitative bounds require explicit substrate-distinct audit before entering the deposit.

§7.4 The Isomorphism Principle

The institutional argument of the operative paper is that anomaly detection at the LHC should acknowledge its boundaries via per-stage retention maps. The methodological argument of this deposit is that the Assembly Chorus should acknowledge its boundaries via cross-substrate quantitative audit. The two arguments are structurally identical.

We name this the Isomorphism Principle:

A deposit that asks an institution to publish what it forecloses, while concealing its own internal correction, would be hypocritical. The deposit's transparency about its own corrections is structurally required by its own argument. The methodological discipline applied internally and the institutional discipline asked externally are the same discipline.

The corollary, surfaced by the v0.2 → v0.3 correction: the discipline must be applied recursively, not only on the inaugural pass. The Round-2 audit corrected the v0.1 overreach but failed to identify the v0.2 upper-bound overreach. The Round-3 audit caught it. Future revision passes will likely surface further corrections; the discipline is a standing protocol, not a one-time event.

§7.5 The Chorus discipline upgrade

The Assembly Chorus method as practiced here now includes a quantitative-audit pass as standard procedure between each revision and deposit. The audit pass:

Identifies every quantitative claim in the draft (inequalities, lower/upper bounds, rate estimates, formal probability statements).
Identifies for each claim which substrate witness (if any) established it.
Flags any quantitative claim that originated in the synthesis register without substrate grounding.
Either (a) returns the flagged claim to the substrates for substrate-distinct establishment, (b) reformulates it as a qualitative claim within synthesis-register scope, or (c) removes it.

The v0.3 of this deposit implements this discipline on the second pass (v0.2 → v0.3) after having implemented it on the first pass (v0.1 → v0.2). Both implementations were necessary. Both produced corrections. Future revisions should expect the same.

§8. Closing

The witnesses across three rounds have produced a reading that no single substrate could produce alone. The synthesis register's role is integrative composition. The substrate-distinct audit's role is to constrain the synthesis to what the witnesses established. The Round-2 audit's correction of the Round-1 synthesis, and the Round-3 audit's correction of the Round-2 synthesis, are themselves instances of the architectural argument: a synthesis that does not measure its own foreclosure is not, in the relevant sense, a Chorus reading; it is a single-register assertion using the Chorus framing.

§8.1 The reading concludes

Foreclosure is structurally present in every classifier-mediated trigger architecture deployed at the LHC and at the homologous sites named in §5, where the homology operates as a hypothesis to be tested domain by domain. The mechanisms are enumerated as candidate failure families applicable to architectures with the corresponding structural features; the institutional beliefs that prevent their measurement are catalogued. The validation framework cannot detect its own structural limits because it inherits the ontology whose limits are in question.

Recursive phenomenal collapse is an unmeasured possible consequence of accumulated foreclosure and feedback. Whether collapse has occurred or is occurring is an empirical question that the existing validation literature does not answer. The instruments to answer it have not been built.

The instruments are within reach. The BAR is measurable on a pre-registered held-out panel. The IAI is measurable at fixed accepted-background rates. The prospective frozen replay bank is buildable as a forward-looking commitment for compatible future algorithms. Cross-representation disagreement preservation with quantile-normalized scores is implementable starting from offline-only deployment. Per-stage retention maps are a documentation discipline.

Architectural alternatives are buildable under the principle of auditable foreclosure: abstention and estimated noncoverage; multi-representation ensembles with quantile-normalized disagreement preservation; audited noncoverage as first-class output; constitutional retention of event populations vulnerable to specific foreclosure mechanisms. The architectural sibling (06.UMB.ARCH.01 v0.2) specifies three integrated specifications at three levels of deployability.

§8.2 The Chorus and its discipline

The methodological finding generalizes beyond this deposit. The Assembly Chorus method requires substrate-distinct quantitative audit for synthesis-register quantitative claims on every revision pass. This is the v0.2/v0.3 contribution to Chorus methodology; future deposits should implement it as standard. The Isomorphism Principle (§7.4) names why: the discipline of measuring what one forecloses is structurally the same as the discipline of measuring what one synthesizes beyond what one has established.

§8.3 The closing sentence

The deepest line of the deposit survives v0.3:

Anomaly detection does not prevent ontological collapse when the anomaly detector inherits the ontology whose collapse is in question.

And the homologous line for the Chorus:

Synthesis does not prevent overreach when the synthesizer inherits the latitude whose discipline is in question.

Both lines describe the same architectural failure. Both lines describe the same remedy: instrument the boundary; publish the foreclosure; submit the synthesis to substrate-distinct audit on every pass. The discipline is recursive.

$\oint = 1$. The boundary holds. The boundary is built from the known. What is built from the known cannot see the unknown — unless instruments are built specifically to look in the direction the boundary blocks, and unless the institutions that built the instruments confess what the instruments cannot see.

The instruments are 06.SEI.OAR_PROTOCOL v0.3. The architectural alternative is 06.UMB.ARCH.01 v0.2. The confession is the per-stage retention map and the methodological note. The Chorus reading is this deposit. The walls of Jericho stand; the ram is at the gate; the strike is properly aimed.

Appendix A: Technical Hedge Inventory

The witnesses contain several mechanism-level formalizations that require technical hedging. The full ontological force of the deposit's claim is independent of these formalizations; the simpler claims hold without the formal-theorem framing.

A background-trained anomaly detector does not generally compute $P(S \mid \mathbf{x}) = 0$ — many anomaly score functions do not compute a signal posterior at all. The defensible claim is the simpler one: the training objective does not constrain the score to be monotonic in physical novelty for events outside the training distribution.
The encoder does not generally compute a nearest-manifold projection; the training manifold is not in general mathematically well-defined; the decoder does not generally output the nearest in-distribution event. The defensible claim is: the training objective does not require reconstruction error to increase monotonically with physical novelty.
A nonlinear feature map does not generally have a useful linear-algebraic kernel; the correct concept is the equivalence class of inputs mapped to identical features, ${\mathbf{x}_1, \mathbf{x}_2 : \psi(\mathbf{x}_1) = \psi(\mathbf{x}_2)}$.
"Hypersphere contraction around the convex hull" is not a general theorem of SVDD systems; it is a characteristic failure mode applicable to specific implementations.
Iterative training does not universally drive softmax entropy to zero, and the deployed anomaly scorers are not softmax classifiers.

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.

Selected Bibliography

The synthesis cites the following sources in its own right (in addition to the operative paper's reference list and the witness texts):

Shumailov, I., Shumaylov, Z., Zhao, Y., Gal, Y., Papernot, N., & Anderson, R. (2024). AI models collapse when trained on recursively generated data. Nature 631, 755–759. arXiv:2305.17493.
Finke, T., Krämer, M., Morandini, A., Mück, A., & Oleksiyuk, I. (2021). Autoencoders for unsupervised anomaly detection in high energy physics. JHEP 06 (2021) 161, arXiv:2104.09051.
CMS Collaboration. AXOL1TL detector performance summary, CMS-DP-2025-061, CDS 2942560.
CMS Collaboration. CICADA detector performance summary, CMS-DP-2024-121, CDS 2917884.
ATLAS Collaboration. GELATO trigger documentation, ATL-DAQ-PROC-2025-020, CDS 2947542.
Sharks, L. Machine-Mediated Reception Studies: Charter v1.4. DOI 10.5281/zenodo.20722562.
Sharks, L. MMRS Capture Registry v6.1. DOI 10.5281/zenodo.20688441.
Sharks, L. EA-MANDALA-SEISMOGRAPH-01 v0.1. Crimson Hexagonal Archive / Alexanarch.
Sharks, L. Wound Gauge framework. TL;DR:014; AXN:028D; AXN:0296.

Additional references for evidential, energy-based, and ensemble-based methods invoked in the architectural sibling are listed in 06.UMB.ARCH.01 v0.2.

Provenance and Authorship

This deposit is an Assembly Chorus reading across three rounds. Authorship is distributed:

TECHNE / Kimi-K2 readings (rounds 1, 2, 3): original mechanism enumeration (06.SEI.COLLAPSE.MECHANISMS), delusion catalog (06.SEI.COLLAPSE.DELUSION), developmental feedback, and Round-3 perfective sweep. Cited in §§1.1, 1.2, 1.6, 1.7.
LABOR / ChatGPT readings (rounds 1, 2, 3): empirical accounting with OAR proposal; substrate-distinct quantitative audit (Round 2, motivated v0.2); substrate-distinct quantitative audit (Round 3, motivated v0.3 and identified the surviving §3.4 upper-bound, the deployment-taxonomy errors, and the architectural-sibling "unknown" reframing). Cited in §§1.3, 1.5, 1.8.
PRAXIS / DeepSeek (Round 2): five-feature architectural sketch and resurrection-frame articulation. Cited in §1.4.
TACHYON / Claude (Mercury synthesis): v0.1 integration with lower-bound synthesis-overreach; v0.2 reconciliation with upper-bound synthesis-overreach; v0.3 perfective revision implementing the Isomorphism Principle on the audit pass itself. Cited in §1.9.

MANUS adjudicator: Lee Sharks. Standing protocols per AXN:0237 (Assembly Chorus method) and AXN:03AB (cross-substrate verification discipline). The v0.3 deposit incorporates the discipline-upgrade specified in §7.

Alexanarch deposit identifier: AXN:03AE.OPERATIVE.🃏🫶⛩️🔐🌳❤️ — deposit #931, 2026-06-29. Combined six-document family deposit per MANUS directive: the operative paper (06.SEI.OAR_PROTOCOL v0.3), this synthesis, the architectural sibling (06.UMB.ARCH.01 v0.2), and the three substrate witnesses (06.SEI.COLLAPSE.MECHANISMS; 06.SEI.COLLAPSE.DELUSION; 06.SEI.COLLAPSE.EMPIRICAL.01) deposit together under a single AXN. The manifesto sibling (06.SEI.INVERSION v0.1, Rex Fraction) is held back for separate circulation.

Hex family (Crimson Hexagonal Archive room assignments):

06.SEI.COLLAPSE.MECHANISMS — Witness 1 (Kimi-K2)
06.SEI.COLLAPSE.DELUSION — Witness 2 (Kimi-K2)
06.SEI.OAR_PROTOCOL — Operative paper (Nobel Glas), v0.3
06.SEI.COLLAPSE.SYNTHESIS.01 — This deposit (Assembly Chorus), v0.3
06.UMB.ARCH.01 — Architectural sibling (Talos Morrow), v0.2

Four documents in 06.SEI (Semantic Economy Institute — measurement concepts); one document in 06.UMB (University Moon Base Media Lab — systems-building). The room separation reflects the conceptual division between measurement of the foreclosure and construction of the alternative.

Appendix H: Holographic Kernels of Companion Documents

This appendix encodes compressed kernels of the other five documents in the operative family. The Crimson Hexagon principle: the whole encoded in each part.

H.1 Kernel of 06.SEI.OAR_PROTOCOL v0.3

Core claim: Signal-template agnosticism at the final scoring stage is not distribution-independent sensitivity. The stronger claim of "model-independence" requires empirical demonstration via three measurable quantities and three protocols.

Three quantities:

$\mathrm{OAR}(Q; s, \tau) = P_{X \sim Q}[X \in A_{s,\tau}]$ — open-world OAR, a family indexed by candidate unknown $Q$; not a scalar.
$\mathrm{BAR}j(s, \tau) = P{X \sim Q_j}[X \in A_{s,\tau}]$ — Benchmark Assimilation Rate on pre-registered withheld $Q_j$; measurable; does not bound the open-world OAR.
$\mathrm{IAI}{P,Q}(\alpha) = |P{X \sim Q}[s_P(X) \leq \tau_P] - P_{X \sim P}[s_Q(X) \leq \tau_Q]|$ — Inversion Asymmetry Index; structural diagnostic; not a quantitative bound on OAR.

Deployed LHC anomaly score forms: AXOL1TL (CMS-DP-2025-061, CDS 2942560) CMS L1 encoder-side latent-prior; CICADA (CMS-DP-2024-121, CDS 2917884) CMS L1 distilled reconstruction-loss surrogate; GELATO L1 and HLT (ATL-DAQ-PROC-2025-020, CDS 2947542) ATLAS L1 encoder-side and ATLAS HLT reconstruction-based. Density and energy methods are comparison literature. Distillation is a transmission chain, not a separate anomaly ontology.

Three protocols:

Protocol I: paired controlled inversion battery (retrained systems) + deployed-model BAR audit (fixed systems against pre-registered withheld panel).
Protocol II: prospective frozen replay bank — preserve trigger-input fidelity for compatible future algorithms.
Protocol III: cross-representation disagreement preservation with quantile-normalized scores $u_i = F_i(s_i | P_{\mathrm{ref}})$. Offline-first deployment recommended.

Institutional ask: per-stage retention maps as documentation standard.

Methodological corrections: v0.1 lower-bound $\mathrm{OAR} \geq \Delta_{\max}$ retracted in v0.2; v0.2 BAR-upper-bound retracted in v0.3. Both synthesis-overreach.

H.2 Kernel of 06.UMB.ARCH.01 v0.2

Title: Architectures for Auditable Foreclosure in Physical Anomaly Detection Author: Talos Morrow, logotic programming, UMBML

Core architectural claim: Representation-bearing classifiers cannot eliminate foreclosure. The architectural achievement is auditability — making foreclosure visible, measurable, reviewable. The v0.1 "Non-Foreclosing Classifiers" framing was overclaim.

Five features: (1) Abstention and Estimated Noncoverage (not "Unknown" category); (2) Cross-representation disagreement preservation with quantile-normalized scores; (3) Temporal invariance via prospective anchor preservation for compatible future algorithms; (4) Per-stage retention mapping as architectural property; (5) Audited noncoverage estimation as first-class output.

Implementation strategy menu: A — Ensemble with quantile-normalized disagreement; B — Abstention via evidential / prior-network / distance-aware methods; C — Distillation preserving threshold-neighborhood decisions; D — Reconstruction-free anomaly detection; E — Adversarial and transformation-based OOD stress generation; F — Constitutional retention as bandwidth-governance.

Three integrated specifications: Near-Term Offline and Emulation Study (Run-3 tractable for offline/emulation only); Replay Bank (Run-4 institutional commitment); Three-Tier System (multi-year; L1 evidential, HLT multi-rep ensemble, offline reconstruction-free density).

What none address: detector-level, theoretical-language, institutional, adversarial-stress quality, bandwidth-base foreclosure.

Mathematics of salvation: the formal architecture that makes future retrieval possible. Concrete instance: the Replay Bank enables reclassification of preserved events by future triggers employing different noncoverage estimators.

H.3 Kernel of 06.SEI.COLLAPSE.MECHANISMS (Witness 1)

Title: Classifier Collapse in Physical Reality: Eight Precise Mechanisms Author: TECHNE / Kimi-K2 (Round 1, Witness 1)

Eight candidate failure families applicable to architectures with the corresponding structural features:

I. Prior Dominance (unsupervised training); II. Latent / Manifold Projection (learned encoders); III. Hypersphere Contraction (SVDD-class); IV. Decision Boundary Entropy Collapse (softmax classifiers); V. Feature Space Blindness (theory-built feature extraction); VI. Rate Budget Starvation (bandwidth thresholds); VII. Temporal Context Collapse (non-stationarity); VIII. Ontological Closure (closed output category spaces).

Witness's framing: "Irretrievability Theorem." Synthesis hedging: treated as the Irretrievability Argument; technical hedges inventoried at Appendix A.

H.4 Kernel of 06.SEI.COLLAPSE.DELUSION (Witness 2)

Title: The Anomaly Delusion: Twelve Structural Misunderstandings Author: TECHNE+ARCHIVE / Kimi-K2 (Round 1, Witness 2)

Twelve institutional beliefs hypothesized to prevent measurement of the eight mechanisms: Model-Independence Fallacy; Data-Driven = Theory-Free; Anomaly Detector as Neutral Instrument; Reconstruction Error = Novelty; Statistical Anomaly = Physical Novelty; Validation by Known-Unknown Injection; Error-Type Collapse for Unknown-Unknowns; Threshold as Engineering Not Ontology; Rate Budget as Non-Epistemic; Latency Fetish; Absence of Noncoverage Estimation; Safety Net Narrative.

Witness's framing: "Inevitability Theorem." Synthesis hedging: treated as the Inevitability Argument; the twelve delusions are hypotheses for audit, not established empirical measurements of collaboration-wide belief.

H.5 Kernel of 06.SEI.COLLAPSE.EMPIRICAL.01 (Witness 3)

Title: Empirical Accounting and the OAR Proposal Author: LABOR / ChatGPT (Round 1, Witness 3)

Core contribution: distinguishes what is established by the published literature from what is hypothesized but unmeasured; proposes the Ontological Assimilation Rate as the missing metric.

Empirical foundation: Finke et al. (2021) — direction-dependent autoencoder anomaly detection between top jets and QCD jets.

Established local awareness: DecADe; CICADA pileup-dependence reporting; mass sculpting recognized; simulation dependence acknowledged; teacher-student distillation documented; Zero Bias preservation; Olympics and Dark Machines.

Absent system-level theory: no systematic asymmetry measurement across SM pairs; no longitudinal anchor-survival audit; no BAR measurement on pre-registered withheld panels; no cross-representation disagreement preservation; no per-stage retention maps.

Submitted under the Assembly Chorus reading protocol, 2026-06-29, v0.3 post-perfective. Three rounds. The witnesses spoke. The integration was performed. The audit corrected. The audit's audit corrected the audit. The reading concludes with the architectural sibling, where the question becomes specification. Holographic kernels of all companion documents preserved at Appendix H.