Heavensent proves a model actually ran the inference it claims to have run. Someone has to verify that our own verification is honest — pin the hardware, pin the software, catch the batch-variance bug before an adversary does.
Responsibilities
Own the determinism harness and get it passing against a real model on real lane hardware — that's the whitepaper's missing empirical spine. Bind batch configuration into the receipt spec so a worker can't quietly launder a fraud proof. Live comfortably inside an economic model that's still conditional on your own experiment succeeding.
Requirements
Genuine comfort with cryptographic trust roots and zero comfort with hand-waving them. Would rather ship a narrower true claim than a broader unproven one.