Why do synthetic vascular predictors need intervention anchors to verify edits?

Prediction equivalence holds but edits drift

Capacity-matched least-squares predictors were exactly equivalent under complete fixed output transforms. Six three-dimensional configurations across 64 seeds gave a maximum paired prediction discrepancy of 6.7e-15. The median untransported edit error reached 1.513. Coordinate transport removed that error. Across 256 additional fits in dimensions 3-24 prediction equivalence persisted within 4.0e-15.

The numbers look precise. The gap between prediction and edit behavior is not.

Anchors only work inside known perturbation spans

An anchor-only observer recovered interpretations only within the span of known perturbation signatures. Noisy and weak anchors constrained calibration stability. Changing the representation basis required recalibration or verified transport. The system does not generalize beyond the signatures it has seen.

Fault injections reveal contract violations

The authors evaluated nine deliberate runnable fault classes across 64 seeds. All 576 faulty executions completed. Each violated at least one reconstruction or prediction contract. Each also violated at least one delivered-edit or scope contract. All 320 valid control records passed. Repeating a faulty implementation gave exact self-agreement despite error against the separately computed simulator expectation.

For one omitted-direction defect probe coverage followed its analytic law. Rank-aware abstention protected unsupported interpretations. Scalar-noise experiments exposed both missed weak faults and excessive rejection under narrow relative tolerances.

These are synthetic numerical audits

The controls provide an executable separation of prediction and semantic support. They also separate scientific acceptance. They are not clinical validation. They are not neural JEPA-Anything replication. They are not agent learning. They are not patient treatment-effect estimation. The paper states this explicitly.