Architecture

Correctness receipts

Before privacy means anything, the split run has to be right. A correctness receipt proves the sharded output matches the whole model, bit for bit.

The claim

Splitting a model across machines is only useful if it computes the same answer as the unsplit model. The correctness claim is strict:

Gate A definition

Greedy output is token-identical between the split run and the single-process reference. If it does not match exactly, nothing else in the project matters yet.

Gate A

The runnable harness demonstrates this on CPU in seconds. It runs a model in two stages - first in-process, then across two operating system processes over the sealed wire - with no stage ever holding the whole model, and checks that the logits and sampled tokens are bit-identical.

.venv/bin/python -m proof.correctness_gate

# PASS: split logits == reference logits (bitwise)
# PASS: greedy tokens identical across 2 processes over the sealed wire
# no stage held the full model

What a receipt contains

A receipt is a JSON document (format noviq-receipt/v1) that a third party can re-check without trusting the operator. The correctness block records, among other fields:

Distinct participants - separate public IPs and GPU UUIDs, across multiple regions, so it was genuinely decentralized.
WAN-scale latency - measured round-trips above a floor, showing hops crossed the real internet rather than one box.
Output hash - a SHA-256 of the output, optionally matched against a reference token sequence.

Attestation and honesty

Self-reported by default

A receipt's attestation level is self-reported unless a signed envelope is present. Verification checks internal consistency - distinct nodes, plausible latencies, a matching hash - not a trusted hardware attestation. We say so plainly rather than overclaim, and stronger attestation (signed identities, trustless compute verification) is on the roadmap.

Correctness is the foundation the privacy receipt builds on: the same run that proves the output is right also carries the measured proof that the input stayed hidden. See Obfuscation and receipts.