Swarmbly AI v2.0.0: protocol, reference implementations, validation harness and benchmark

Abstract

Peer-to-peer language-model inference today splits the model: layers or tensors are spread across machines and activations cross the public internet on every generated token. Swarmbly distributes the problem instead: a client-side orchestrator decomposes a request into semantic sub-tasks, each dispatched once to a volunteer node running a complete small model, and the returned fragments are verified and assembled locally, so the network is crossed once per unit of work rather than once per token. What version 2 adds: a theory of the fragment. Version 1.4 knew that it had to fragment but not into what. Version 2 derives it: the fragment is a task domain cut at interfaces of weak coupling; its size L is a measured property of the node class and the fragment count is derived from it; its overlap is the maximum of two observable quantities (longest repeat, longest dependency crossing a boundary), which saves 29 % of network compute; and the context budget is derived rather than swept, formalised as an indirect rate–distortion problem with a second axis δ. Five falsifiable models (M1–M5) accompany the theory, each with its death condition written before measurement, and Section 3 turns the genomics analogy into an explicit method for accepting or discarding a homology. Measured, with a published harness. The document ships with a reference implementation, a falsification harness that refuses when its instrument cannot decide, and a corpus of 473 real runs over five local model families. M4 is falsified in two independent instruments; M2 survives corrected; M3 survives; M1 and M5 remain unmeasured. With a matched output budget, fragmenting costs in all five families (aggregate +11.17 %, 95 % CI [+4.80, +17.34]); that aggregate is confounded with output length in both directions, so the abandonment criterion is reported as not measured, neither met nor failed. The design’s thesis, measured directly. On a structured-extraction task with a genuine fragmented arm, extracting with a small model and aggregating with code beats the same model answering alone by +46.9 points (95 % CI [+33.3, +59.4]). A whole-document control separates the two mechanisms: at 80 rows the gain comes from taking the arithmetic away from the model and splitting adds nothing; at 160 rows single-call extraction collapses and only splitting sustains it (+77.1 points). A small node extracts reliably up to some task size, and fragmenting keeps each task below that threshold. The scope is narrow (8 documents, one model family, two sizes) and is declared as such. Withdrawn results are kept withdrawn. The ρ-tax curve and the reliability claim of the confidence map, both withdrawn in version 1.4, are not rescued. Defensive publication. Elements E1–E18 were disclosed in version 1.4; elements E19–E24 are new and become prior art with this publication. This record is the software artifact. The whitepaper (version 2) is deposited separately as 10.5281/zenodo.23031305. Contents: the reference implementations, the validation harness, the benchmark corpus and the documentation, with the whitepaper sources in English and Spanish. Licensing: software and specification under AGPL-3.0-or-later; document text under CC BY 4.0.