Generation Is Cheap, Selection Is Sovereign

Generative AI has made plausible output abundant and cheap, but plausibility is not evidence and evidence is not authority. Leading AI assistants already concede this in the notice beneath every prompt box, telling users the system can make mistakes and asking them to check — handing the work of verification to individuals without instruments, records or authority. This working paper proposes Computational Trust Engineering (CTE), an engineering discipline for the path from a language model’s completion to an organization’s consequence. It separates three selections that current practice collapses into one: a generative selection (which continuation is probable?), an evidentiary selection (which candidate is supportable?) and an authority selection (may the supported candidate produce consequence?). CTE grounds these in four realities (human, AI, actual and governance reality) and treats familiar failure modes as measurable gaps between them. It frames evidentiary selection as metrology, with the individual claim as the counted unit, “unknown” as a legitimate result, calibrated instruments and traceable provenance. It introduces attention as the allocation of scarce measurement effort, and states when a language model may serve as an instrument. The paper defends the three-layer separation through collapse analysis and precedent from internal control, functional safety and metrology. It also maps CTE to the NIST AI RMF, NIST AI 600-1, ISO/IEC 42001, assurance cases and the EU AI Act, and closes with deployment postures, five falsifiable hypotheses and a research agenda.