what a hackathon prototype can claim

archipelago started at a hackathon: a layer where drones share what they’ve learned, not just what they’ve seen, with no central model in the loop. a convincing demo is easy to build in a weekend. what it’s allowed to claim is a different question.

the idea in one paragraph

each node trains its own policy. a node offers a model version to a neighbor; the neighbor validates it locally (shapes, finite values, size, age, provenance) before adopting, adapting or rejecting it. cut the link and everyone keeps acting on their own. the point is that useful behavior can spread across a fleet without a single point of failure.

what the experiments showed

learning beats idle, random and untrained policies on frozen test seeds, with a confidence interval that excludes zero. that’s real, and it’s also the weakest bar there is. hover the bars below.

it does not beat a simple rule-based explorer on return: 15.8 below, with an interval that stays below zero.

the live pilot

the pilot ran four real processes with real optimizer updates and real peer exchange, cutting the network in half during every episode and killing one node mid-run. the fleet that exchanged capabilities scored well above the isolated control. that’s the claim the project exists to make, so it’s the one to be most careful with: it comes from one training run per arm, and the intervals don’t include run-to-run variance, which the earlier study found to be large. the partition and crash-restart checks failed outright.

so the honest summary is a promising signal, one run deep.

why say so

the reports state these limits next to the numbers. that’s the right default: write down what the evidence supports and what it doesn’t, then design the next experiment around the gap. the roadmap is terrain, richer candidate validation, and keeping capabilities alive as nodes drop out.

code and the devpost entry are linked from the project page.