Shared Vision as Coordinating Prior: Aggregation-Layer Masking and the Witness-Filter Pathology

Version 1.0 (initial release). Shared visions, strategic priors, and operating theses coordinate multi-agent systems by reducing policy divergence around a common token. We identify three structural failure modes by which such coordination produces persistently wrong steady states even when the system has working feedback, solicits dissent, and updates on reported error. First, any first-moment aggregator (mean, balanced weighted mean, rank-symmetric median) maps balanced bias-split witness populations to zero; the shared prior is therefore frozen by the arithmetic of the aggregation rule, not by any procedural refusal to update. Second, public alignment to a shared prior is alias-compatible with hidden local-gradient compositional divergence under stationary conditions, with the divergence surfacing only under strategic shift. Third, witness inclusion correlated with prior-alignment defeats every aggregator — including shape-preserving receipt-lineage architectures — by removing signal upstream of aggregation; a single such filtering event produces persistent non-vanishing error with no internal mechanism for self-correction. The constructive consequence is that corrective governance architectures must preserve per-witness structure and maintain witness inclusion independent of prior-alignment; collapsing witness reports to scalar summaries or curating the witness population by alignment reintroduces the pathology.