Evidence becomes publicly consequential only when an observation can survive a chain of transformations without losing the possibility of inspection. Peirce shows that data acquires evidentiary force through inference rather than through mere existence; Popper makes vulnerability to contradiction a condition of robust knowledge; Foucault demonstrates that institutions participate in determining what counts as legitimate evidence; Bowker and Star reveal the classificatory infrastructures that make heterogeneous records interoperable, searchable and actionable. These traditions converge on a central proposition: evidence has an architecture. An observation must be recorded, given provenance, described through metadata, published, indexed, cited, retrieved and eventually interpreted within another decision. Each transition can increase public capacity while introducing distortion. Records without provenance weaken trust; metadata without conceptual precision produces invisibility or noise; publication without persistence creates fragile knowledge; citation without inspection becomes ritual; indicators can eliminate uncertainty while appearing to increase precision. Open science must therefore be understood as more than accessibility. A publicly downloadable file can remain epistemically closed if it cannot be discovered, identified, versioned, interpreted or connected to the evidence from which its claims emerged. The same issue intensifies with artificial intelligence, where generated statements can circulate after their evidentiary histories have been compressed or erased. Publication, metadata, persistent identifiers and citation are consequently not clerical appendices to research but components of its epistemology. Public knowledge exists when other actors can recover the route connecting a claim to its observations, assumptions and transformations, examine where uncertainty entered the chain, contest its interpretation and reuse the record without severing it from the conditions that produced its meaning.