RESEARCH PREVIEW — enrollment opens after a multi-home cohort run and an adversarial verification run pass.

Research. The runtime and coordinator are built and publication pending, and speak the Tinker SDK API so researchers can submit work with familiar tools. Two Macs have fine-tuned, exchanged, and merged one adapter to byte-identical files across three rounds. Enrollment opens after a multi-home cohort demo and an adversarial verification demo pass, in that order.

• One program, safe by default — a signed menu-bar app that pauses when you work and respects disk, network, and thermal caps. Contributions run through an allowlisted operation set, never arbitrary code.
• Work sized for home internet — the wire, not the GPU, is the constraint (measured at roughly 20 megabits per second on real home links). Every run declares a byte budget your client enforces.
• Verified, then credited — contributions are checked by sparse recomputation and settled only on multi-party agreement. Verification evidence never silently becomes credit. Credit is recognition on an auditable ledger, never a token.
• Permissionless protocol, forkable curation — any run meeting published machine-checkable criteria (open-weights license required) can exist on the network. The featured-runs index is a signed, forkable default, not a chokepoint.

opentinker will test training-contribution integrity on volunteer-owned hardware. It has no payment rail and does not let strangers rent a participant's machine. Named research projects publish the runs, and accepted projects must produce open weights. Democratized distributed training describes the intended system; the public ledger separates measured results from pending tests.