Verification and Projection: Two Epistemic Operations on a Mechanism

Mechanism-based analysis produces two structurally distinct types of evidence: evidence that tests a specific diagnosis under falsification pressure, and evidence that a mechanism reproduces in an independent domain. These are not the same operation. Conflating them produces a system that cannot tell whether it is learning or merely repeating itself — one that systematically converges toward mechanisms that are widespread rather than mechanisms that are true. This note formalises the distinction, defines four evidence types, introduces the temporal asymmetry that makes conflation structurally inevitable without explicit design, and proposes minimum conditions for promoting a mechanism from candidate to working model.