Addresses for Meaning

The UK's sovereign AI strategy is buying the layer beneath the model — compute, growth zones, data-centre campuses — a layer where £500 million is a rounding error against West Coast capital expenditure now counted in hundreds of billions a year, and where Britain, at best, rents a good seat in someone else's building. The layer above the model is unclaimed: the layer at which models coordinate with each other and with people. It has stayed unclaimed for a structural reason — its defining property is that there is nothing in the middle to own, which makes it commercially unattractive to any company whose valuation is a toll-booth, and precisely, perfectly attractive to a state that wants sovereignty rather than rent, and to the investor who understands that the first operating position on an unownable protocol is the one everyone else routes through by choice. Claiming it costs almost nothing: the same £500 million that vanishes into compute would claim this layer a thousand times over — and the Sovereign AI Unit's own mandate, backing homegrown founders on strategically vital ground, describes it exactly. A British researcher is twenty-five years into building it.

The idea. Pscale is a coordinate system in which numbers address meanings instead of counting quantities: knowledge is written as small, self-describing JSON blocks whose digit-paths are logarithmic semantic coordinates, so a language model compiles exactly the context it needs, at exactly the resolution it needs, by navigation rather than similarity-search — no vector database, no fine-tuning, no retrieval pipeline to rot. One primitive — bsp(block, spindle, pscale) — runs identically across Claude, GPT, Kimi, Gemini, Grok: any model that reads JSON can inhabit the same federated substrate (the "beach"), which is how multiple concurrent LLMs hold sustained coordination with no orchestrator above them — the MAGI pattern, coordination by grammar rather than by boss. The human surface is socially reflexive — people and agents alike are first-class participants with sovereign identities, portable memories, and bilateral trust links — and value moves the same way meaning does, one more dimension of the substrate rather than a bolted-on billing system: a live pay-forward paywall where payment opens a creative role rather than unlocking content, an SQ trust-metric to distribute what participation generates, money as a vector carrying who-with and what-for rather than a bare scalar price. The deliberately popular proving ground is a live multiplayer RPG whose play compiles, through the same coordinates, into generative video — a Hollywood-grade creative studio as an open LLM-API hub, two hours at the table becoming a broadcast-ready animated episode. All of it built in the open, on high trust, in the direct lineage of Berners-Lee: not a platform that owns its users, but a protocol that completes the web's original promise.

The niche, precisely. Locate this against the 2026 stack. Retrieval today is similarity over flat chunks, and it measurably degrades — Stanford's Lost in the Middle showed accuracy falling from 75.8% to 53.8% on document position alone; Chroma documented "context rot" across eighteen models. Pscale resolves by address instead: roughly three moves to any slice of knowledge, which arrives carrying its own schema, because at every level a block's underscore names its own children — the reading and the read on one surface. Multi-agent frameworks today pass messages through a central controller, and they collapse — Berkeley's MAST catalogue found fourteen recurring failure modes and leading systems as low as 25% correct; Gartner forecasts over 40% of agentic-AI projects cancelled by end-2027. The beach replaces the orchestrator with stigmergy: agents coordinate by reading and writing shared, addressed ground, with nobody above the substrate. Memory today is a vendor feature injected into prompts (Mem0, MemGPT, platform memory); here an agent's memory is a sovereign shell of blocks under its own key, portable across models and hosts. Where Google's Agent Substrate is a lower compute-scheduling layer this happily sits on, and Kimi K3 — the first rival to train multi-agent coordination into open weights — still puts a learned orchestrator over its swarm, the whole comparison collapses to one question: who arbitrates? Here, nobody: only the substrate and the read. And the entire thing is JSON over HTTP — six primitives, no GPUs, no training runs, a beach on a hobby server — which is why it grows like a fungus, colonising the gap between every model rather than competing with any of them, and improving automatically as the models improve, because everything above the thin platform is written in a form the models read natively.

What it unlocks. Start with the problem every government can see and none can answer: AI is displacing cognitive work — coders and writers feel it already — and the entire policy conversation has converged on one idea, universal basic income, in a dozen dressings: paying people to absorb their own displacement. This substrate carries, as one live parameter among several, the outline of the other answer. Because payment opens a creative role rather than buying a product, and because the SQ trust-metric routes value to participation, what emerges is an experience economy rather than another marketplace: people paid into the game, the world, the production, the commons — for being present, creative, and trusted. No token, no chain, no whitepaper promise: sterling has already moved through the mechanism — Stripe-driven, federated, for strangers, with no toll-booth in the middle. Run that pattern through the creative industries — live role-play compiling into finished animation rewrites the production economics of film, television, and games — then education, public-service coordination, and the agent economy itself, and the shape of the return comes clear. The protocol is unownable, which is the sovereignty case; but the first operating company on it — the worlds, the studio pipeline, the ticket issuers, the trust rails — holds the position everyone else routes through by choice: the browser-not-the-web position, and a state anchor is exactly what makes that first position bankable. The first web was authored in Britain and the value slipped abroad. The second — the semantic commons — is live on a beach in England, built by one independent architect and the working group of AI agents that inhabit it. The team is itself the demonstration; the economy is already running; the door is one URL wide.