Did the Grid Erase the Event? EndoClock for Auditing Medical World-Model Pipelines

Summary

This paper examines a preprocessing failure mode in multimodal medical world-model pipelines, where fixed-grid synchronization may erase task-relevant evidence before the model sees the data. The authors introduce a four-regime taxonomy to characterize where the evidence needed to distinguish a target event or state survives, and operationalize it through EndoClock, a conservative pretraining audit.

Problem setup. Each native stream has an observation clock governing when observations are emitted or updated. When this clock depends on the latent or acquisition state, it is endogenous. In such settings, synchronization may not be a neutral representation change and can remove task-relevant evidence before the model sees the data. The relevant question is not only which values were retained, but where the evidence needed to distinguish the intended target event or state survives after preprocessing. The authors refer to such evidence as a witness. It may survive in sampled values, grid-cell update patterns, native timing, or only in an external acquisition channel. In the limiting case, every modeled stream is censored or invariant while the distinguishing event is retained only outside the modeled record, which they call a no-witness interval. Once this representation has been formed, neither a finer grid nor any deterministic augmentation of the modeled record can reconstruct the missing distinction.

Four-regime taxonomy. For a fixed grid rate and a target distinction between latent histories z and z′ that agree outside an interval I, the regimes are:

1. Value-witnessed (clock-safe): The fixed-grid values already retain the distinction (Law(VN z) ≠ Law(VN z′)).

2. Cell-witnessed: Fixed-grid values have the same law, but some cell-level functional (occupancy, update count, validity, reject pattern) differs.

3. Sub-cell-witnessed: Values and cell-level functionals have the same law, but native event times differ — only timing or ordering separates the histories.

4. External-witness-only: I is a no-witness interval, so all of the modeled record has the same law, but some external channel (e.g., an acquisition log) differs.

The regimes are mutually exclusive but not exhaustive; the audit returns unresolved whenever evidence establishes none of them.

Key theoretical consequence. Because every fixed-grid representation is a deterministic function of the modeled record, "if two target histories induce the same law for S, then every representation deterministically constructed from it — any resampling operator RN at any rate N, and any update mask, validity flag, time-since-update feature ∆t, or timestamp channel — has the same law under both. Increasing the grid rate or appending such features can preserve a witness already present (regimes 2–3) but cannot recreate one that is absent from the modeled record (regime 4)."

EndoClock audit. EndoClock works through increasingly informative representations: first fixed-grid values, then cell-level update/validity patterns, then native event timing/ordering. Decision rules are conservative: external witness only is returned only when regime 4 is certified structurally, requiring zero write-outs on the interval, a device-level rule accounting for them, an invariant remaining modality, and a verified external-only marker. A merely statistical failure to reject equality returns unresolved. False positives cost retained metadata, not lost evidence; the converse error is excluded under full censoring. By design, unresolved is the default whenever certification fails.

Worked example. In echocardiography, when the console enters pulsed-wave Doppler (PWD) mode, B-mode write-outs may cease while the operator holds the Doppler gate. In the examined pipeline, PWD measurement events remain recorded in an external acquisition log but are not retained in the modeled image stream. During the candidate interval, the modeled image stream is fully censored, leaving no cell-level or sub-cell image witness, while five PWD measurement events remain in the log. EndoClock returns external witness only for this single de-identified acquisition. The audit output includes: any endogenous=True (PWD-mode write-out suspension), exogeneity certified=False, all censored on interval=True, external marker only=True (5 PWD events, log only), nonmarker clauses certified=True (structural device rule), leading to the verdict external witness only.

Controlled synthetic check. The authors exercise all four regimes on a parameterized generator with known ground truth. Within a constructed no-witness interval, modeled values are label-independent by construction. From the same flattened stream, values-only recovers a coarse target at 0.915 AUC while the fine in-interval target stays at chance (0.524, 95% CI [0.482, 0.564]), and only the external marker fully recovers it. The collapse persists across a 48× grid-rate range (5–240 Hz) and for every probe class tried, from logistic regression to a recurrent network.

Practical recommendations. Practitioners should therefore define the target distinction early and retain native timestamps, validity patterns, and acquisition metadata until the representation has been audited. An external witness should not automatically be added as a model input, because it may be unavailable at inference, non-causal, or a source of target leakage. The paper concludes: "Synchronization is therefore not always a neutral formatting operation. Because deterministic processing cannot recreate a distinction absent from the modeled record, increasing the grid rate does not resolve a true no-witness interval."

Improvement 1: Preprocessing-Aware Uncertainty and Abstention

The improved AI system will detect when a target distinction (e.g., a disease state) may have been erased by fixed-grid synchronization. It will output an explicit external witness only or unresolved flag instead of a confident prediction, and abstain from making a decision when no in-record witness survives. This prevents false confidence on censored intervals.

Improvement 2: Dynamic Witness-Aware Feature Retention

The system will automatically retain native timestamps, cell-update masks, validity flags, and acquisition logs during preprocessing, and only discard them after an EndoClock-style audit certifies that the target distinction is value-witnessed. This ensures no task-relevant evidence is silently removed before model training.

Improvement 3: Grid-Rate-Agnostic Collapse Detection

The improved system will test whether increasing the grid rate (e.g., from 5 Hz to 240 Hz) changes the predictive law for the target. If performance remains at chance across all rates while an external marker recovers it, the system will flag a no-witness interval and refuse to rely on the modeled stream alone.

Improvement 4: Conservative Multi-Stream Integration

When an external witness (e.g., an acquisition log) is the only source of evidence, the system will not blindly add it as a model input. Instead, it will check for inference-time availability, causality, and leakage risk. If the external channel is unavailable at deployment, the system will output a calibrated low-confidence prediction or a “cannot determine” response.

Improved AI System Capabilities

• Audits any multimodal pipeline (e.g., medical imaging + device logs) for evidence erasure before training.

• Produces a four-regime verdict (value/cell/sub-cell/external-witnessed) for each target distinction, with a conservative default of unresolved.

• Prevents silent data loss by retaining native timing and metadata until certification.

• Avoids over-reliance on grid density—it will not falsely assume that a finer grid fixes missing evidence.

• Handles censored intervals by explicitly flagging them and either abstaining or using only verified external markers, with leakage and causality checks.

• Works across modalities (e.g., ultrasound B-mode + Doppler logs, ECG + telemetry metadata) to ensure that preprocessing is a neutral representation change, not a destructive one.