HOW IT WORKS
One goal, five steps, no coordinator.
Every robot periodically broadcasts a small authenticated frame that carries what it knows: its neighbors, which jobs are done, and who claims which job. The merge rule is commutative, associative and idempotent, so the swarm converges to one view no matter how frames are delayed, repeated or reordered.
Split. Agree. Do once. Prove. Learn.
NO COORDINATOR
Every robot broadcasts what it knows to its neighbors. The swarm converges to one view with no server and no reliable link.
ONE START PER JOB
A claim with a bid, a settle time of two periods, then a START that dominates every claim. Exactly one robot starts each job.
NOTHING TO TRUST
Finished work is signed, bundled and anchored on SUM Chain. Anyone can check it without trusting the operator.
THE LOOP
Split, agree, do once, prove, learn.
A signed goal becomes a tree of jobs. A robot claims a leaf with a bid, waits a fixed settle time, then announces it has started. Finished work is signed and bundled. The bundle hash goes on chain and the raw data goes to storage. Idle robots train on that data, and the better models flow back.
ALLOCATION
Exactly one robot starts each job.
A claim carries a bid. The claimant waits two dissemination periods, long enough for a better bid to arrive, then sends a START marker that dominates every claim. The first draft waited one period. The numerical checks showed double starts, so the rule became two. That is Theorem 4, and two periods is the minimum.
ARCHITECTURE
Five planes, one invariant.
Link, coordination, autonomy, settlement and learning. Each plane has one job and a narrow interface to its neighbors. Safety lives in the coordination plane and holds whatever the models, the chain or the backhaul are doing. No model can break it, because the rules are enforced outside the models.
ON THE ROBOT
Two tiers in every robot.
Tier 0 is a microcontroller running no_std Rust: the radio, the link protocol, the shared state, the allocation rules and integer inference. It keeps working when the robot is isolated. Tier 1 is a companion computer that holds the small models, builds the records and bundles, and talks to SUM Chain, SNIP and OmniNode when a backhaul is available.
JOB DONE
Prove it on chain. Learn from it later.
A finished job becomes a signed 196-byte attempt record. Records are bundled, and the bundle hash is anchored on SUM Chain. The success and failure data goes to SNIP as an episode pack. A robot with no job fetches packs, trains through OmniNode and publishes a weight delta. The hash of the new model lands in the next bundle.
THE MODELS
