Auditable transactional prequential adaptation for visual anomaly localization under unknown stream contamination

Technical Summary: Auditable Transactional Prequential Adaptation for Visual Anomaly Localization

Problem Formulation
The paper addresses the challenge of visual anomaly localization in an online, streaming setting where the deployment environment is subject to unknown distribution shifts and unlabeled contamination (defects). Traditional adaptive methods face a causal and evidential dilemma: to update a model, it must process an image, but if that image contains a defect, updating the model on it may cause the system to "forget" or fail to detect similar defects later. Conversely, if the system updates on benign shifts, it may drift.

The core difficulty lies not only in how to update but in how to verify what has been measured. The author argues that standard train-adapt-test descriptions often obscure the timing of updates relative to predictions. If an image is used for adaptation before its prediction is recorded, the reported score becomes retrospective. Furthermore, many "memory-bounded" methods fail to account for hidden state, such as readable caches or unreported rollback copies, effectively violating their stated memory constraints.

The paper defines this setting as a first-arrival prequential contract:

1. Ordering: At event $t$, the localizer must read the pre-event state, emit a prediction (score and patch map), and seal that output before deciding whether the event influences the model.
2. Privacy: Ground truth, masks, and future events are hidden from the runtime.
3. Accounting: Every future-readable byte of state must be accounted for, including snapshots required for reversibility.

Methodology: BSPAL
The author instantiates this contract in a baseline method named BSPAL (Bounded State Prequential Anomaly Localization). The architecture is deliberately simple to ensure inspectability:

• Representation: A frozen DINOv2 ViT-S/14 encoder generates patch descriptors. The system uses two fixed views (identity and brightness-adjusted) to compute anomaly maps based on the mean nearest distance to a source bank ($A_0$) and an adaptive bank ($A_t$).
• Source-Calibrated Admission: A statistic derived only from the frozen source bank ($A_0$) determines if an event is admitted to the adaptive process. This prevents the adaptive bank from influencing its own admission gate.
• Transactional Adaptation:
  • Quarantine: Admitted events are compressed into coordinate-median summaries and stored in a FIFO quarantine.
  • Proposal: Once the quarantine holds sufficient entries, a new prototype is proposed via coordinate median.
  • Commit & Probation: A complete snapshot of the pre-commit state is stored. The new prototype is committed atomically, entering a 16-event probation period.
  • Rollback: If the source-only statistic exceeds a high threshold for three consecutive events during probation, the system triggers a rollback, restoring the exact pre-commit snapshot. The rollback mechanism is charged against the same memory budget as the adaptive prototypes, ensuring no "free" auxiliary memory is used.
• State Bounding: The persistent state is strictly bounded by fixed capacities (e.g., 32 prototypes in the adaptive bank, 64 entries in quarantine, 1 snapshot), independent of stream length.

Key Contributions
The paper's contributions are explicitly limited to the definition of the contract, the implementation of the transactional rule, and the integrity of the evidence chain:

1. First-Arrival Contract: A formal definition for unlabeled visual anomaly streams that combines an evaluator-visible estimand with a strict runtime information boundary and peak byte accounting for all future-readable state.
2. Bounded Transactional Adaptation: A rule establishing four properties:
   • Constant persistent state size regardless of stream length.
   • Causal first-arrival prediction (prediction depends only on pre-event state).
   • A local coordinate-median bound for proposals (under the assumption of a majority-normal population).
   • Exact restoration of the transaction view upon rollback.
3. Protocol Integrity and Evidence Discipline: The author maps the contract to executable oracles (e.g., hash chains, prefix invariance, truth swapping) to verify the software implementation. Crucially, they report a protocol-integrity falsification: a post-seal audit of the VisA benchmark revealed that the role split (training vs. testing) failed to isolate physical instances and near-duplicates.

Results and Evidence
The paper presents a strict separation between engineering traces and efficacy claims:

• State Accounting: A read-only reconstruction of 945 run ledgers confirmed that all runs remained below the configured 32 MiB persistent state cap (maximum observed: ~1.18 MB). This validates the bounded-state claim.
• Latency: Direct per-event latency was not measured; thus, no claims regarding inference speed are made.
• Efficacy (Invalidated): The historical benchmark computation on MVTec AD and VisA is invalidated as a scientific result. The audit found that 20 inspected cross-role pairs in VisA were confirmed matches within physical groups. Consequently, the population did not satisfy the "group-disjoint" prerequisite.
  • The paper explicitly states that no benchmark improvement, safety, or equivalence conclusion is drawn from the endpoint values (e.g., Pixel AUPRO, Image AUPRC).
  • The historical results are retained only as an audit trace to demonstrate the failure of the population admission protocol.

Significance and Claims
The paper's significance lies in its methodological rigor regarding protocol integrity rather than performance gains.

• Causal Ordering: It demonstrates that adaptation timing must be part of the measured method, preventing retrospective scoring.
• Reversibility Cost: It makes the memory cost of reversibility (snapshots) explicit and comparable, rather than treating rollback as free auxiliary memory.
• Audit Discipline: The primary finding is that a reproducible calculation is not a valid experiment if the data population violates isolation prerequisites. The author argues that the appropriate response to such a failure is to preserve the computation for traceability while withdrawing its scientific interpretation, rather than selectively deleting rows post-hoc.

The paper concludes that while the transactional contract and its bounded implementation are sound, the question of whether transactional adaptation improves first-arrival localization under hidden contamination remains unanswered and requires a new, fully admitted execution with a group-disjoint population. The current contribution is the contract, the implementation, and the evidence discipline that distinguishes reproducible bytes from admissible inference.