Being able to prove that the model file you are running is the one the publisher produced and has not been altered. It is ordinary supply chain hygiene applied to an artefact that is large, widely redistributed and frequently obtained from community mirrors.
Technical view
Implemented by cryptographic signing of weight files with verification at load time, hash pinning in deployment configuration, and attestation of the build or training environment. Model cards and datasheets provide the documentary layer. Emerging practice extends signing to fine-tuned derivatives so a chain from base model to deployed artefact can be established, which regulation increasingly expects.
Related
Part of AI Terminology Index, section: Security and Privacy.