When Is a Latent a Scientific Instrument? A Validation Protocol for Scientific Foundation Models

Latents extracted from internal representations become credible scientific instruments only when their semantic meaning, causal role, and stability are validated separately. Using a sparse autoencoder on a foundation model for neutrino data, we show that these properties sharply dissociate: physically readable features can be unused by one prediction head, but used by another. These results highlight the need for a strict validation protocol for latent identification.