Local Cognition and Anchored Transmission: A Local-First Cognitive Extension for Swarmbly

Swarmbly fragments the problem, not the model, and its workers are deliberately almost stateless: the client is the only unit that remembers. Swarmbly LCE (Local Cognitive Extension) asks what to do with that memory without deforming the protocol, and answers with a strict division of labor: facts live in a readable memory, behavior lives in the weights. Each user keeps on their own device a wiki whose claims are typed, versioned and anchored to spans of their own sources. Alongside it sits a small LoRA adapter that learns only their voice, register, terminology and procedures. The adapter is regenerated from the base model and the wiki with interleaved replay, and accepted only on a freshly generated test. The protocol does not change. Personalization enters Swarmbly only through fields of the existing global contract Γ. Workers serving third parties run their base model, so verification by activation commitments is untouched. A task projection with double privacy classification decides what leaves the device, and it can only raise the privacy lane, never lower it. An optional social plane lets nodes pull small signed knowledge capsules (≤ 16 KiB) on demand: no gossip, no global graph, no transitive trust. Designed against the evidence, not around it. Fine-tuning injects new facts poorly and can increase hallucination. Recursive training on model output erodes the tails of the distribution unless data accumulate rather than replace. Conformist copying drives minority variants to extinction. And models of different families agree on the same wrong answer far more than chance: 60 % of the time when both are wrong, in one study of more than 350 models. Each finding becomes a design rule: capsule descendants require new human evidence, never a model paraphrase; prevalence is shown as a label, never used as an adoption criterion; rare variants are weighted up, as in negative frequency-dependent selection; the number of copies of a preserved capsule is derived from a loss tolerance. With node lifetimes of about three months and weekly repair, three replicas bring annual loss to about 2 %. Population genetics as instruments, not metaphors. A homology is accepted only when it supplies an instrument or a failure mode, following the method of the Swarmbly whitepaper v2. The accepted ones are complementary learning systems (the wiki as hippocampus, the adapter as neocortex), Wright's island model to budget diversity between neighborhoods (F_ST ≈ 1/(1+4Nm)), the breeder's equation to bound the gain from adapter selection, and Hebbian growth corrected by normalization for peer affinity. One homology also sets a limit. Local learning does not decorrelate errors within a request, because models share them by descent from overlapping pretraining corpora. One example runs through the paper. A user writes "¡chuta, se cayó el servidor!" in their own texts. Their adapter learns to use the Ecuadorian interjection in an informal register, and their wiki records, with its anchor, what it means. A node in another country that meets the word learns "in Ecuador 'chuta' is used as…", never "I am Ecuadorian". That knowledge survives the disappearance of the node that originated it. Status: draft; nothing has been measured. Eight invariants, seventeen hypotheses with abandonment conditions written before measurement, and twelve declared limitations accompany the architecture. The release ships a reference implementation with one test per invariant and a validation harness whose six instrument tests pass in simulation. That shows the measures respond to what they measure, not that the extension works. Runs with real models will be reported in version 0.2. Defensive publication. Elements EC1–EC11 become prior art with this publication. Contents: the whitepaper (version 0.1) in English and Spanish, as PDF and Markdown, and the annotated bibliography. The protocol it extends is described in 10.5281/zenodo.23031305. Code, specification and harness are at github.com/Sebastardito/SwarmblyLCE. Licence:CC BY 4.0.