The fields that carry real weight

ordering. The difference between a partial read being a sample and being a lie. An unordered source truncated at a cap gives you an arbitrary subset; an oldest-first source truncated at a cap gives you a subset that systematically excludes the present. Only the realm knows which.

queriedBy vs partition. When users query on axes the source cannot filter, every question needs the whole partition, and partial fetching cannot be made safe by narrowing. This mismatch is the single best predictor that a collection wants strategy: mirror rather than live.

It is not a verdict on its own. When every question is bounded by a partition the source DOES fetch by (one firm’s record, a date window), a live read takes that partition whole and narrows on the other axes locally, over a complete set. Say so by naming the partition as filterableAtSource names it:

partition: firmId                 # the fetch key, named as filterableAtSource names it
queriedBy: [firmId, isFormerRepresentative]
filterableAtSource: [firmId]      # the register cannot filter the flag; it reads a firm whole
sync: { strategy: live }

The platform warns about strategy: live only when an unfiltered axis is bounded by no partition the source fetches by, so each question would need the whole collection. Whether one partition fits a read is the producer’s page budget, reported per read as truncation. Mirror when partitions outgrow it, or when questions must cross partitions.

watermarkFilterable. A record may CARRY a last-updated timestamp without the source being able to FILTER on it. The first permits incremental refresh; the second forces a full re-walk with a local diff. Assume the wrong one and a day of changes vanishes silently. Declare it.

declaredTotal. Where the source states a partition’s true size, completeness stops being an assumption and becomes checkable arithmetic.