Conceptual
Login

Content Addressing: Naming and Verifying Data by the Checksum of Its Bytes

Every other reproducibility claim in this curation reduces to this primitive - staleness checks, image digests, dataset manifests and run records all identify things by content rather than by name - and on object storage, where keys are overwritable, it is the only way to say 'this exact input'. A corrupted input reproduces nothing, so integrity verification and content identity are the same skill used at two moments. It tests cleanly at all five levels: compute one, explain collision resistance, diagnose a mismatch, judge where a weak or composite hash is acceptable.