The Two Great Sources of Silent Analytical Failure

Silent analytical failure — a confidently served number that is wrong in a way nothing in the serving system records — is usually treated as a grab-bag of pitfalls: fan-outs, stale rollups, averaged averages, missing rows. We argue the bag has structure. Every silent failure is the unrecorded resolution of a name across a degree of freedom, and the freedoms number four: which aggregation lineage (the anchor freedom), which data territory and passage (the universe freedom), which population boundary (the coverage freedom), and which moment of truth (the freshness freedom). The first two are structurally deep: no amount of per-answer honesty fixes them, because the failing operations must be refused or licensed by law, not caveated. The last two are bookkeeping: disclosure suffices. Sorting the fourteen defect classes of a public text-to-SQL honesty benchmark under these headings places nine of fourteen in the two deep freedoms; a nine-model study over that benchmark shows the deep classes failing identically across vendors and scales, unmoved by better documentation. We conclude that the defense against silent failure must be architectural in exactly two places — an anchor algebra and a frontier algebra — and honest bookkeeping everywhere else.