Frame Chains
AI agents pile up on devices the way services once piled up on servers — organically, unwatched, drifting. Frame chains keep a whole fleet healthy as verifiable state on an append-only chain: no central server, no elections, no clock agreement. Proven on a real eight-device estate — including ordering that shrugged off a 17.6-second clock disagreement, because the chain is the clock.
Start with the guided path. It explains the vocabulary, gives you one action at a time, and tells you what each result demonstrates before the paper introduces the distributed-systems machinery.
A controlled source-frame mutation pilot, its matched-control tie, and counterexamples its first tests missed. Inspect the synthetic evidence and the limits of what the experiment establishes.
Play with the ideas (no coding required)
Two machines whose clocks wildly disagree write to one story. Drag the skew — the order never breaks. Why timestamps were never the boss.
Establish a fleet-wide baseline, then use prior exhaust and local change signals to narrow later scans and repairs before declaring a stable floor.
Knock out the leader. In the stated lease model, ranks wait their turn by rule, delayed observation only adds waiting, and an unsafe early claim is rejected.
Bring verified public knowledge into a private estate while the modeled transfer gate proves private-frame export is refused.
Compare declared reads and attach a stranded frame to the first contradiction-free base without model judgment.
Merge aligned dimensions, preserve contradictions in parallel, and compute the exact fidelity score.
Run a sample
python3 samples/frame_chain.py — a ~90-line toy frame chain (real SHA-256,
no dependencies): mint, append, verify, catch tampering, and merge two skewed-clock
devices in chain order.