The Graph Is the Memory: An Architecture for Learned Routing, Crystallizing Knowledge, and Measured Growth over a Fleet of Language Models

We present an architecture in which routing and memory are the same structure. Tasks project onto facets; facets connect to actions (models, adapters, templates, extracted composites) through weighted edges; the edge graph is simultaneously the routing policy and the system's memory of which work belongs where. Reading memory is summing edges, writing memory is a forward-only delta rule, and because the representation is strictly additive, recall generalizes to facet combinations never seen. Knowledge crystallizes into progressively cheaper forms (edge weights, an exact-match cache, admission-gated composite actions, low-rank adapters distilled from the system's own accepted solves, and tiered shared stores), so the system becomes faster and cheaper as it runs. Results, each pre-registered and scored by a mechanical oracle rather than a judge model: a learned router matched a frontier-only control on accept rate at roughly 31 percent lower cost per accepted solve; a specialist adapter trained on the system's own record raised held-out accepts on its cell from 48.0 to 90.7 percent; a from-scratch discovery organ proposed a task partition no human named, which was admitted by evidence, wired into the live router, and whose first gated activation recovered reward at 35.7 and 4.0 standard errors; and concurrent learners consolidated with zero lost updates. We also report a load-bearing negative result: under the corrected reward currency, every composite previously admitted by an accept-only condition was reward-negative, and the growth layer's remaining claim is stated as an open, pre-registered program. Every claim in the report is labeled proven, partial, or open, and every number traces to a run artifact. Changes in version 0.4.1. This revision carries numeric corrections to v0.4; the supervised ceiling re-derived under an exchangeable hash-parity split (0.8158 retracted; honest ≈ 0.78), the warm-start result re-registered as ceiling-parity against that ceiling, and the dentate graft's standard error corrected (~4–6, not ~12). Plus the first executed three-run temporal-seams campaign: a headroom measurement from the record, a registered positive (run 1), its retraction under oracle re-verification (run 2), a byte-level store-path diagnosis, and a verified restoration (run 3: all four predictions pass on verified accepts, 22.9× cheaper per verified repeat-accept, 463/463 hits re-verified). Claim G receives its first verified positive instance, with the ε-power limitation disclosed per its registration. Two hypotheses from the independent analysis were withdrawn during the diagnosis and are printed in the correction ledger, whose governing principle is that every number survives because it was checked by something that could contradict it.