How do I verify that an open-weight model was actually trained from scratch?

When a lab publishes weights and claims to have built them from scratch, there is no standard protocol for verifying the claim. Spectral fingerprinting via SVD decomposition, attention-based differential analysis, and refusal-vector behavioral comparison can each suggest a model is a fine-tuned derivative of Llama or DeepSeek, but the methods often disagree after quantization or pruning, no protocol is standardized, and any result can be disputed. The problem went commercially significant in 2026 when several labs making self-developed claims were found, under independent analysis, to be substantially more derivative than advertised, with legal and regulatory implications. EU AI Act high-risk model provisions now require provenance disclosure, but the technical machinery for making or verifying that disclosure does not exist at scale.