Train together, one byte per vote.

Research. Pure array functions for mlx-go, evaluated as an in-process multi-peer simulation. What crosses between peers is votes on public probes (not weights, not raw data) at roughly 115 KB per peer per round in the evaluated setting.

• Vote histograms — builds the consensus distribution from one-byte argmax votes per probe across peers.
• Hybrid objective — the paper's loss: local cross-entropy on labeled probes plus KL divergence against the consensus histogram.
• Consensus decay — linearly curtails the consensus channel over a configurable window, so out-of-distribution probes cannot collapse training.
• Randomized response — an optional differential-privacy mechanism applied to votes before release, giving a formal bound instead of a hope.
• FedAvg bridge — the paper's bandwidth-bridge variant: an optional parameter mean every M rounds for identical architectures.

mlx-go-tallytrain is the lowest-bandwidth rung of the family's federated-training research: where mlx-go-fedistill exchanges weight deltas and mlx-go-ccl synchronizes full training state, tallytrain asks how much peers can learn from each other when all that crosses is a vote. Because the probe set is public and the vote is one byte, the privacy story is legible enough to explain in a sentence, and the randomized-response option makes it a bound rather than a promise.