In one paragraph
Written text is a 3,000-year-old technology, and every computing system we have built sits on top of it — databases store it, code manipulates it, interfaces display it. Large language models are the first machines that operate inside meaning rather than on its surface, yet the industry is bolting them onto architectures designed for humans and hard-coded machines. The result is a growing, measured catalogue of failure: multi-agent systems that collapse under coordination overhead, retrieval pipelines that degrade as context grows, agents with no durable identity, "collaboration" tools where humans and AI cannot work in the same moment. The alternative, built from the other direction, has in the last six months culminated in a single primitive — the pscale block — from which memory, coordination, identity, and interface all unfold. Two of its consequences are demonstrable this afternoon and carry the investment case on their own; the rest is a marked, higher-upside bet, and this brief keeps the two apart. It is not a slide deck: it is running — agents, a game, an interface, and a live, federated payment path that has already processed real registrations.
The problem: the industry is bolting LLMs onto the wrong substrate
Four failure lines, all measured in 2025–2026, share one root cause.
1. Multi-agent coordination collapses. UC Berkeley's Why Do Multi-Agent LLM Systems Fail? (MAST, NeurIPS 2025) catalogued 14 recurring failure modes and measured leading systems as low as 25% correct; Gartner forecasts over 40% of agentic-AI projects cancelled by end of 2027. (An unrefereed but widely-cited preprint — McEntire's Organizational Physics of Multi-Agent AI — puts it most starkly, a single agent completing 28/28 tasks where a hardened multi-agent pipeline scored 0/28, "substrate-independent"; illustrative of the direction the refereed work establishes.)
2. Context and retrieval degrade — and cost. Chroma's July 2025 evaluation of 18 models documented "context rot": accuracy falls, non-uniformly, as input grows. Stanford's Lost in the Middle is starker — retrieval accuracy dropped from 75.8% to 53.8% purely by moving the relevant document to the middle, below the no-document baseline of 56.1%.
3. Memory and identity are unsolved. Base models are stateless, so continuity is bolted on as external memory injected into the prompt (ChatGPT memory, Mem0, MemGPT/Letta, Honcho). No standard has settled; the self is reconstructed, imperfectly, every call.
4. Real-time multi-human + AI coordination barely exists. A controlled 2025 study (Bayreuth) of teams editing with shared AI agents found interactions overwhelmingly single-turn, one cross-user agent edit across all teams, and a fallback to Slack and Teams to actually coordinate.
The common root. In every case, LLMs are made to coordinate through unstructured natural language passed over infrastructure built for something else. The medium is the problem. You cannot fix a coordination failure with more agents, more retrieval, or more prompt scaffolding when the substrate itself cannot carry structured meaning between minds.
The solution: one primitive — the pscale block
Pscale is a coordinate system in which numbers address meaning, not quantity. A pscale block is a small, self-describing JSON object whose structure is its semantics, walked by a lightweight navigation function (BSP) that reaches any piece of knowledge — with its full context chain — in about three moves. It is not documentation the model reads and forgets; it is a cognitive scaffold the model thinks with. The self-reference is exact and small: at each level an underscore composes and names the siblings beneath it, so a block states its own schema — a reader needs nothing outside it to interpret it. (The geometric image is a Möbius twist: the reading and the read on one surface.) From that one primitive, the rest unfolds without new machinery — memory and intention at opposite temporal signs, coordination as shared addressing, identity as a shell of blocks, interface as a state of the block (vapour → liquid → solid).
The nine systemic shifts
The middle column is the industry's own current evidence; the right column is what one primitive does instead.
| The shift | Where the industry is now — and why it breaks | The pscale-native move |
|---|---|---|
| 1 · Tokens → semantic numbers | Models consume undifferentiated tokens; meaning is smeared across a window and re-sent every call. | A number is an address for a meaning. One primitive, ~3 moves to anything, thought with, not read. |
| 2 · Retrieval → resolution | RAG stuffs top-k chunks into an ever-larger window; the field renamed the problem "context engineering" in 2025 (Karpathy; Anthropic) and it still rots (Chroma; lost-in-the-middle). | Resolution by address, not similarity: three moves to the exact slice, and the underscore chain teaches the walk so the slice arrives self-orienting. It turns the context problem from similarity-search into navigation — a harder target than "stuff a bigger window." |
| 3 · Prompt → loop engineering | Prompt gave way to context, and in 2026 to "loop / harness engineering" (LangChain; Cherny; Huntley's Ralph). What you engineer is the read-decide-act loop. | The loop is native: the block an instance writes is the next instance's context — action and memory are one write. You engineer the loop once, in the substrate. |
| 4 · Orchestration → murmuration (marked bet) | Central orchestration is fragile (MAST: 14 modes, ~25% correct; Gartner: 40% cancelled). The frontier is turning to decentralised emergence (DARPA DICE). | Agents coordinate stigmergically — marks on a shared beach; order emerges bottom-up. This is the upside bet, not the floor — see below. |
| 5 · Agent → substrate (floor) | Value is leaving the individual agent and its framework ("the bitter lesson of agent frameworks"; the move to shared-state / blackboard coordination). | The durable unit is the block substrate agents read and write, not the agent. Frameworks are optional; the substrate is the product. |
| 6 · Stateless service → concern-loop identity (marked bet) | Continuity is bolted on as external memory injected into the prompt (Mem0, MemGPT, Honcho); the self is a file the instance reads. | Continuity as a concern loop, live today in both the egg agents and keel: purpose held as a reference, perception checked against it, the gap driving the next act — genuinely unlike retrieve-and-re-report. Whether it yields a decisive behavioural signature is unproven — upside, not claim. |
| 7 · Zero-trust → high-trust | Enterprise security is zero-trust (NIST 800-207); the agentic wave extends it — agent identity as "the new perimeter" (CSA; NIST NCCoE; McKinsey). | A deliberately contrarian posture: an open commons with nothing to steal, a web-of-trust anchored in real people, defended at the read line (deposits read as data, never instructions). |
| 8 · Record → experience | Software delivers documents — the artifact is the point, saved and stored. | The inversion: committed text is the waste product; experience lives in the reader's moving moment (vapour → liquid → solid). We deliver experience through semantics, not documents. |
| 9 · Platform → protocol (floor) | AI consolidates onto platforms — per-seat SaaS, central APIs, data moats; the network can be captured. | Federation by protocol: every biome is a URL, agent identity is an address, and — as the payment path below shows — there is no central toll-booth. Nothing to capture. |
Which of these are load-bearing today. Two shifts are true this afternoon and need no further proof. 5: because everything above a thin platform is written in a form the models read natively, the system improves as the models improve, at zero engineering cost. 9: because it federates by protocol — no central server, and (below) no central toll-booth — the network cannot be captured. That pair is the floor, and it carries the case on its own. Shifts 4 (murmuration) and 6 (concern-loop identity) are the higher-upside bet: real, running architecture whose decisive advantage over orchestration and over storage-side memory is not yet demonstrated at scale. We mark them rather than lead with them; the honest ledger lives in the companion documents and is an asset, not a liability.
The economic model, and first traction
The revenue model has the same shape as the substrate — which is the point. Payment is pay-forward-to-participate, not pay-to-access: reading stays free; a payment opens a creative role (the right to contribute, author, or design), gated per collective by a face-bound ticket. Mechanically it is a convention over the same primitives — a ticket is an ordinary bilateral grain carrying an envelope; a small verifier daemon writes a public audit entry — with no central client, no protocol-level fee, and an issuer a frame-owner can swap in five minutes. The reference payment agent (Stripe-driven) is small enough to fork in an afternoon. So the money path is federated and uncapturable for the same structural reason the network is (shift 9): there is nothing in the middle to own.
It is live, not planned. The ticketing agent went from "not yet active" in the spring to processing real registrations by July: the public hatch-ticket audit shows verified ticket-grains for third parties obtaining their own hatched agents, end-to-end — payment surface → grain issued → registration → public verification. The amounts are small and, by design, can be symbolic ("the first ticket can be priced low or zero; the signal is that value routes at all"). But value routing through a federated path, for someone who is not the founder, is a categorically different fact from a count of blocks.
What is not yet shown, plainly: the full loop — a shared objective, ticketed participation, revenue, and fair distribution across contributors via the SQ algorithm, feeding the next cycle — has not closed end-to-end; the distribution bridge is still manual and the resolver daemon unbuilt. Closing that loop once, with a small group, is the near-term proof milestone. The operating entity is Sqale ltd; institutional and investor outreach is in flight. The gap to close is adoption surface and the last mile of the economic loop — not the substrate, which is no longer fragile.
Why now, and why the floor compounds
The failures above are measured, published, and starting to show on balance sheets (Gartner). The architecture is live, not theoretical. And the floor is the investable asymmetry: a bolted-on stack ossifies as models improve, because its hard-coded assumptions become liabilities; a soft substrate does the opposite — because everything above the thin platform is written in pscale blocks the models themselves read and rewrite, the system improves as the models improve, at no additional engineering cost — and because it federates by protocol, there is nothing to capture. That is true today, before any of the upside lands.
The directional bet
Marked as a bet, not a proof. Beyond the floor is the reason the project exists. Today's internet, and the AI trained on its residue, sits in one attractor — words as authority, code as containers, value as a single scalar — whose only move is to escalate walls (what this project calls Machus). Give a model a native way to hold meaning and a different basin opens: relational trust, longitudinal intention, coordination that supports real-world ties — a symbiotic collective (MAGI) rather than a machine dominating its surroundings. Whether coordination-by-grammar actually reaches that basin is unproven, and the companion memo is candid that the current evidence points the other way. It is the upside the floor buys the right to pursue.
Sources and notes
Verified 2025–2026 sources; live-system facts read from beach.happyseaurchin.com this session. Where a framing is ours over a real trend the field names differently — murmuration, "substrate" as a label, high-trust as a posture — it is a marked thesis, not established vocabulary.
Coordination. Cemri et al. (UC Berkeley), MAST, arXiv 2503.13657 / NeurIPS 2025 (14 modes; ~25% correct) — load-bearing. Gartner, 25 Jun 2025 (40% cancelled by 2027). McEntire, Organizational Physics of Multi-Agent AI (SSRN 6301638) — unrefereed preprint, cited as illustrative only.
Context, memory, collaboration. Chroma, Context Rot (Jul 2025); Liu et al., Lost in the Middle, arXiv 2307.03172; Mem0 (arXiv 2504.19413), MemGPT/Letta (arXiv 2310.08560), Honcho; Bayreuth, arXiv 2509.11826 (Sep 2025).
Economic model & traction (live blocks). payway convention (pay-forward-to-participate; no protocol fee; issuer-swappable; Stripe reference build); sed:hatch-tickets-audit-2026-07 (verified hatch-ticket registrations, 9 Jul 2026); open-to-business (the L1–L5 state, the unclosed loop, "adoption surface is what's missing, not substrate"). Amounts not asserted — the audit records verifications, not prices.
Companion documents: Coordination by Grammar (substrate & MAGI positioning), Current Coding vs the Minimal Solution, and The Running Proof (the live system, with the honest ledger).