Brain-Measurable Diffusion Decoding: Auditing Information Provenance in fMRI Reconstruction

Diffusion models are increasingly used as scientific reconstruction engines, but realistic samples do not by themselves reveal where target-specific information entered the system. This ambiguity is acute in fMRI-to-image reconstruction: a frozen image generator can amplify weak neural conditioning into plausible scenes, while target captions, target-image features, retrieval, or candidate-pool selection can silently turn reconstruction into an oracle-assisted protocol. We formulate this issue as an information-provenance problem and define brain-measurable diffusion decoding: every inference-time control supplied to the generator must be a function of the measured fMRI response, fixed learned weights, fixed sampler settings, and target-independent randomness. We instantiate the criterion with a three-channel fMRI-conditioned decoder that predicts a Stable Diffusion VAE latent, learns a CLIP-aligned visual coordinate during training, and supplies brain-derived cross-attention tokens to a frozen SD1.5 denoiser. On 982 NSD shared-test images, the method reaches competitive standard-8 reconstruction quality under a clean train-only, single-rendering protocol. Provenance ablations, sampler-strength sweeps, learning-dynamics traces, calibration checks, and resource audits identify which controls carry stimulus-specific information and which evaluation choices can alter the scientific claim. The result is both a reconstruction system and an evaluation template for separating brain-derived generation from target-side shortcuts in scientific uses of deep generative models.