The position underneath our distributed-training work is simple to state. Consumer Apple Silicon is abundant, energy-efficient, and built around unified memory. Combined with deterministic replay and peer-to-peer transport, we believe it is a viable substrate for machine learning that people own. We treat that as an engineering program, not a speculation, and a program has a discipline: every load-bearing claim gets its own measurement, its own written evidence state, and permission to fail. This note is the public ledger of where those claims stand.
Measured
- Two machines can train one model and prove they agree. Two Apple Silicon machines fine-tuned, exchanged, and robust-merged real LoRA adapters to byte-identical files in three consecutive rounds, validation loss falling from 6.07 to 2.23. Small run, real result: Two Macs, one adapter.
- Adapter exchange is the wall, not training. In that run, 98.7 percent of round wall time went to moving 2.5 MB between machines. A negative result, documented like a result, with a designed replacement protocol and pre-registered pass bars.
- Sparse recomputation catches tampering at tunable cost. A verifier that re-executes a sparse sample of a committed update's output coordinates raises tamper detection as the sample grows, at measured verifier cost: in one measured setting, sampling under half a percent of bytes caught 82 percent of a defined corruption for about two percent of verifier time. Measured against synthetic corruption; the run against a real recomputer is still owed, and we say so.
- Pure-Go collectives hold the C++ latency floor over Thunderbolt RDMA. In a two-host benchmark the pure-Go path matched the C++ floor at the median and edged ahead at the tail, with zero heap allocations per collective. Proven at two ranks and 16 KiB payloads on Apple hardware only, and not yet wired into the mesh layer above it.
Analyzed, measurement pending
- Byte-reproducible merging has exact sufficient conditions. The conditions that make a floating-point merge byte-identical across independent hosts are worked out, with a failure taxonomy for how real implementations silently break them and a pre-registered cross-architecture property test. All result cells pending.
- Robust aggregation has a covert failure mode after selection. If surviving peers are combined by weights the peers themselves advertise, one peer can inflate its influence after the robust selection step, in a place the standard defenses never look. One peer, one vote is the counter. The three-arm poisoning demonstration that shows it end to end has not been run.
- Pipeline-split training across two Macs was bit-exact once; the record must be re-captured. A model split across an M4 Max and an M3 Max pair reproduced sequential computation over real TCP to bit-exact logits, 769,536 values with maximum difference zero, in a working run whose raw record did not survive. By our own rules a number without its surviving measurement is a target, so it sits here until the run is re-captured.
- The wire, not the GPU, looks like the constraint. Working observations put consumer Wi-Fi on this fleet near 20 megabits per second, an order of magnitude under what published datacenter-grade decentralized training assumes, with a 181 MiB sync round taking a fifth of a second on loopback and over a minute on the real link. Those observations came from the same unsurvived record; the link budget re-measurement is specified and pending.
- The exchange bottleneck has a designed fix. A streaming, resumable, verified store protocol with a latency model calibrated against the measured run. Its decision rules are fixed in advance so a later revision cannot move the goalposts.
Pre-registered, not yet run
- Byte-exact GPU-to-CPU replay at useful model sizes. The claim that a strict CPU backend can reproduce a GPU inference trace bit for bit is load-bearing for verification without trusted hardware, and it has not been measured where it matters. The measurement is specified before the result exists.
- A training floor for a handful of Macs on home Wi-Fi. The hypothesis is that the wall is the network, the roughly 20 megabits per second a real home link sustains, not memory or compute. Both the link budget re-measurement and the training floor over it are pending, and until they land, it is a hypothesis.
- Reproducible agent binaries across independent builders. Byte-equivalent rebuilds can bind a binary to reviewed source only if independent builders actually reproduce the bytes. A pre-registered study defines the harness; the harness is not built.
Rules of the program
- A number without its measurement is a target, not a result, and is labeled as one.
- Failures are recorded like results. The rank-4 transfer that died before merge is in the ledger, not swept out of it.
- Verification evidence never silently becomes credit. Checking that work happened and deciding who gets credit for it are separate decisions, kept separate on purpose.
- Every claim above is allowed to fail. A substrate claim that cannot fail is marketing.
The write-ups behind this ledger are in preparation for publication and available for review on request. For what a verification record does and does not establish, see What a training receipt does and does not prove. For the run itself, see Two Macs, one adapter. For where the program is pointed, see Democratized distributed training.