The whole model, not just the head
Most "verifiable AI" re-runs a tiny last step and fingerprints the rest. Here your browser
re-runs the entire image model - every convolution - in deterministic integer arithmetic, and
lands on the lab's exact answer, bit-for-bit. No black box, no trusted extractor. Then check an
Ed25519 credential that signs the whole model. A sibling to our
three-ways proof, which trades a fingerprinted extractor for higher accuracy.
Starting from the quantized image (3072 integers), your browser runs all six convolution layers, the linear projection, and the int64 coherence basin - in pure integer arithmetic, no floating point on the path. It must reproduce the lab's 64 integer features, the per-class energies, and the decision exactly. This is the part everyone else fingerprints.
An Ed25519-signed receipt whose model_sha256 covers the CNN
and the basin - so the signature pins every weight, not just the head. Anyone verifies it
offline, mapped to the controls a regulator asks for.
The honest gap in most verifiable-ML is the feature extractor: it runs in float, so it is fingerprinted, not re-run. Make every layer deterministic integer and that gap closes - the entire decision, image included, is re-runnable and signable on any machine. That is the moat: not a more accurate model, a fully re-runnable one.
Honest split: running every layer in deterministic integer costs a little accuracy. This whole-integer model is on CIFAR-10; the three-ways demo is 96.8% but re-runs only the basin head and fingerprints the float extractor. Here there is no fingerprint - the model hash covers everything.
Ed25519 over
a model hash that covers every weight. Built by Coherence Energy Labs.