Heavensent is a protocol for verifying that a language model actually ran the inference it claims to have run, without trusting the worker that ran it. The bet is narrow on purpose: cheap optimistic exact-recompute on a deliberately restricted deterministic lane, with a clean upgrade path to per-step cryptographic proofs once those get cheap enough for interactive use.
Three lanes, three trust roots
Proving an entire inference cryptographically end to end is still too slow to be useful. Heavensent instead runs three lanes side by side, each honest about what it actually trusts: a reference lane that recomputes a challenged job bit-for-bit and slashes on mismatch; a confidential lane built on hardware that's shipping today — NVIDIA confidential computing on H100, H200, and Blackwell, at single-digit overhead; and a cryptographic lane that commits to a full inference trace, then proves only the single challenged step with a lightweight ZK proof.
The finding that de-risks the reference lane
The reference lane depends on one assumption: that a language model's forward pass is actually reproducible. It is — but not for the reason most people assume. Non-determinism in LLM inference mostly isn't floating-point race conditions; it's batch variance, kernels whose reduction order changes with batch size. Fix that, and a fixed hardware class and software stack gives bitwise-identical output, run after run. The catch is precise: hardware, software version, and batch configuration all have to be pinned into the receipt itself, or a worker who batches your job with others produces different bytes than a verifier recomputing it alone — which would fire a false fraud proof.
Where it sits
Verifiable inference is a crowded field already, and this isn't a claim to have invented the category. Modulus Labs and Giza deploy verifiable models today; Gensyn is live on mainnet; the three-lane split itself is closer to industry consensus than a differentiator. What Heavensent is actually betting on is narrower: cheap optimistic recompute on a deliberately restricted deterministic lane, not full end-to-end proving.
The economics
Security is priced directly — reward plus stake has to clear the cost of an independent verification, divided by the fraction of jobs actually checked, with a zero-bond case where the system self-indexes without requiring an upfront stake. All of it is conditional on one unrun experiment: the determinism harness has to pass against a real model on real lane hardware before the economics are more than a model on paper.