One decision, three proofs
Our coherence model (96.8% on CIFAR-10) looks at one image and decides. That exact decision is then proven to you three independent ways, live in your browser: re-run it bit-for-bit, verify a zero-knowledge proof that hides the features, and check an Ed25519-signed audit credential. A genuinely accurate model whose decision you can also re-run, prove, and audit.
Your browser re-runs the model's int64 basin decision from its exact integer form, bit-for-bit, the same answer as the lab's CPU, native, and GPU.
A 807-byte Groth16 proof that the model produced this decision - without revealing the features. snarkjs verifies it right here, no server.
An Ed25519-signed receipt anyone verifies offline, plus the exact extractor it ran,
mapped to the controls a regulator asks for.
Honest split: a deep decision is a float feature extractor (the CNN) plus the int64 basin. The basin decision is what is re-run, zero-knowledge-proven, and signed above; the extractor is bound by a cryptographic fingerprint () - the honest float-provenance half.
basin_fp decision re-run bit-for-bit, a circom + snarkjs Groth16 proof over the
model's exact integer constants (features private), and an Ed25519-signed, re-runnable
receipt. Same decision, three proofs - on a genuinely accurate model. Built by Coherence Energy Labs.