A real calculation runs on your own device, and anyone can re-run it and get the exact same result.
A real computation — a tiny model turning an input into a decision — computed on your own device. The exact same integer math runs on CPU, WebAssembly, and your GPU, proven bit-identical (0-ULP); the result carries a coherence-proof the open verifier re-checks. Move the inputs and everything re-derives live. This is the compute a capsule runs — real, and checkable by anyone.
What this proves: that the computation is reproducible — the same input and code give the same bytes on every backend, and anyone can re-run it. What it doesn't: that the model's decision is correct or good. Reproducibility is a property of the math, not a judgment of the answer.
Eight signals into a fixed 4-class linear model (the "weights"). Drag to change the input.
Two guarantees. Soundness: the decision re-runs bit-identically here on every substrate. Standard: an independent open verifier (the same one the whole Coherence ecosystem uses) confirms the proof. Tamper anything and it fails. The GPU leg engages where WebGPU is present.
The step above proves a decision when the model is public. This tier proves a statement while the witness stays hidden — a trace-hiding STARK. The prover runs off this page; your device re-verifies the returned proof end-to-end: proof integrity, binding to the exact statement you asked about, and the full STARK transcript (a byte-identical port of the fabric verifier). A prover that lies is caught right here.