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.
Verifying Downloaded File Integrity with SHA-256 and MD5 Checksums on Linux
File integrity verification rests on the cryptographic hash function: a deterministic map from arbitrary-length input to a fixed-length digest (checksum) such that any change to the input — however s…