Where the Cloud Collapses

Sensor fusion, data fusion, dashboards and data visualisation are conventionally distinguished by input type and by where they sit in a processing stack. This paper argues that the distinction which matters for decision quality is a different one, and that it is two variables applied to all four: every representation of multi-source evidence collapses a distribution somewhere between sensor and decision, and representations differ in the locus of that collapse and in whether the residual uncertainty survives it as a carried object. Fusion collapses upstream and, when honest, ships the residual as covariance or hypothesis mass. Dashboards defer collapse to the viewer's retina and ship nothing; the resulting belief has no bookkeeping. Visualisation is the encoding of whichever object arrives, indifferent to whether it is a point or a distribution, which is why it is blamed for collapses committed upstream. The two variables, collapse locus and residual retention, place the four terms on one map, generalise the confabulation coefficient of The Confabulating Collator from language-model collators to the whole representation stack, and yield a diagnostic: the failure mode conventionally called bad fusion is a product that collapsed upstream and discarded its residual, delivering the confidence of a posterior with the epistemic hygiene of a gauge. The publicly reported expectation, in the investigation of the February 2026 Minab strike, that a targeting platform would have flagged contradictions it carried no representation of, is read as this failure at operational scale, with the causal question left where the public record leaves it. A design rule follows: defer collapse to the human where the human carries the prior the system lacks, fuse where they do not, and at neither locus discard the residual. The constructive requirement for language-level fusion products is stated and handed to a companion paper on divergence retention. The operator located in the architecture is Binary Bias.