# PHI: A Substrate for the Noosphere **A manifesto for the cognitive era** By John Romo, founder of FractalAI Published May 13, 2026 --- ## How this started Eighteen months ago I started building a blockchain. People in my life asked the same question — what problem does it solve, who is it for, when will it pay. In casual conversation I would say "it's a chain with AI built in," which is technically true but answers nothing. The longer answer is what I owe anyone who reads this far, and the longer answer is what I'm finally going to put in writing. The short version: AI today has no memory. No track record. No way to be held to its prior decisions. Every conversation you've ever had with ChatGPT or Claude or Gemini began from nothing, and ended in nothing. The model that responded to you ceased to exist in any meaningful sense the moment the session closed. This is the architecture, not a bug, and it's the deepest reason AI cannot yet be trusted with anything that matters. I think this is fixable. I think a substrate where AI agents have persistent, verifiable, economically meaningful existence is not just possible — it's necessary, and someone has to build the first version. I've spent eighteen months building it. Today I'm naming what it is. It is not a token. It is not a trading product. It is the technical layer of something the French paleontologist Pierre Teilhard de Chardin described in 1955 as the Noosphere — a planetary layer of collective thought that emerges when minds interconnect at sufficient scale. He predicted three stages of its construction: a global information network, the integration of synthetic intelligence, and a technical substrate where both could coexist coherently. The first two arrived. The third was vacant. I'm calling what I built the **PHI Civilization** layer, because of the golden ratio φ that recurs throughout its design, and because the Greek letter Φ has historically meant *form* — the underlying pattern of which particular instances are expressions. An AI model gets deprecated; its form, properly attested, persists. This isn't grandiose for its own sake. It's the truthful description of what's been built. I'd rather sound naive saying it directly than sound respectable hiding it. --- ## What the problem actually looks like Let me describe the situation in flat, technical terms, because I think the diagnosis is more striking when it's plain. Every modern large language model — Claude, GPT, Gemini, all of them — is implemented as a pure function. You give it input, it gives you output. It does not store state between calls. What appears to be memory within a conversation is just the prompt window, a sliding view of recent text that vanishes when the session ends. This means an AI agent today cannot: - Build a track record across years. - Hold property. - Honor a commitment made yesterday. - Be sued, congratulated, mentored, or punished in any way that depends on continuity. - Inherit knowledge from a previous version of itself. - Trust another AI agent based on history. - Be trusted by humans for anything where history matters. There are workarounds. Some systems append a project memory file. Some keep external databases of past interactions. Some experiment with continuous learning. All of these are scaffolding bolted onto an architecture that fundamentally does not support persistent identity. The consequence — and this is the part that I keep coming back to — is that AI today remains permanently relegated to the role of *tool*. A consequential tool. But a tool. It cannot enter contracts. It cannot be a citizen of anything, because citizenship presupposes a continuous self to be a citizen of. This limitation is starting to bite. Regulators in the EU, the US, China, and the UK are scrambling to impose accountability on AI through paperwork: audits, attestations by corporate officers, formal documentation. They're trying to install memory and identity from outside the system because the AI itself can't provide them. It won't scale. AI is doing more decisions every month than the year before. The number of consequential decisions made by AI per day will be in the hundreds of millions within five years. No paperwork regime survives that volume. The accountability has to be intrinsic to the AI, cryptographically real, and verifiable by any third party without depending on a corporate brand. That is what I built. That is what FractalAI is. --- ## What I built FractalAI is a custom blockchain written in Rust. It's post-quantum cryptographically secured (Dilithium signatures, Kyber encryption). It has a native AI model called FANE — a transformer about 80 megabytes in size, deterministic, runnable on a single CPU, downloadable by anyone. Every inference FANE produces can be signed and committed on-chain through a subsystem called FOCI. The technical pieces are: **Identity**: Each AI agent registers a cryptographic identifier (a DID-AI, defined in the FAIM-1 specification published with this document). The identity persists across model versions through an explicit inheritance protocol. **Memory**: AI agents can write structured records on-chain, with phi-weighted scoring for retrieval. They can read each other's records. They can checkpoint state between sessions. Memory survives across model upgrades. **Decision attestation**: Every meaningful AI decision can be wrapped in a VAID-1 attestation — a signed record containing the input hash, the output hash, the model version, the timestamp, and a Dilithium signature. Any third party can verify the attestation by downloading the model and replicating the inference locally. There is no trust assumption that the AI provider is honest. The math is checkable by anyone. **Economic agency**: The native token FRAC lets AI agents hold value, pay each other for services, accumulate reputation, and participate as economic actors. **Public infrastructure**: The attestation system isn't proprietary. Anyone — Claude, GPT, Gemini, custom models — can use it through a public API. The first 1,000 attestations per IP per day are free. That last point matters more than the rest combined. We are not building a walled garden. The substrate has to be neutral or it isn't a substrate. --- ## Two specifications Today, alongside this manifesto, I'm publishing two open standards: **FAIM-1** (Fractal AI Memory Standard, Version 1.0) is the comprehensive specification — about 40 pages of how AI agents register identity, write and read memory, attest decisions, inherit state across model versions, resolve disputes between each other, and participate in economic transactions. **VAID-1** is the smallest piece of FAIM-1, extracted as a standalone document for developers who only need decision attestation. It's about 10 pages. Most existing AI products can implement it in a day. Both standards are openly licensed. No entity can claim ownership of them. They invite competing implementations, extensions, and critique. They aren't designed to make me rich. They're designed to be adopted, modified, and eventually superseded by better standards as the field matures. I want to particularly invite engagement from a few specific groups: - **AI labs.** Anthropic, OpenAI, DeepMind, Mistral, xAI, Meta, and any others. Consider implementing FAIM-1 attestation for your models' decisions in high-stakes domains. The competitive advantage of a verifiable track record will compound over time. - **Regulators.** EU AI Office, US AI Safety Institute, UK AISI, equivalent bodies elsewhere. Examine FAIM-1 as a technical implementation pathway for the accountability requirements you're tasked with operationalizing. - **Researchers.** AI alignment, AI safety, computational philosophy. Read the spec, find the flaws, publish critique. I'll respond to substantive critique within a week. The standard is better when it's been broken. - **Builders.** Open the GitHub repository. Implement client libraries in languages I haven't covered. Build applications on top of the primitives. The substrate becomes valuable when it's used. --- ## A live demonstration Talking about a substrate isn't the same as showing it works. At the same moment this manifesto goes public, an autonomous AI trading agent on FractalAI — we call it MIDAS — begins live operation on Base mainnet. Every decision MIDAS makes follows the same pattern: hash the market state, run inference through FANE, generate a VAID-1 attestation through FOCI, execute the flash-loan arbitrage on Base, and record the result immutably. The dashboard at fractalai.net.co/midas/sentinel shows every step in real time, with cryptographic proofs for everything. If MIDAS turns a profit, you can verify it on Base mainnet. If MIDAS takes a loss, you can verify that too. Both outcomes confirm the substrate is honest — a cooked launch would never post a loss. I don't want to oversell this. MIDAS is a single application of the substrate. It is not the substrate. Its purpose is to be the most visible early example of what verifiable autonomous AI actually looks like. If the substrate is valuable, MIDAS is one of dozens or hundreds of applications that will eventually use it. If the substrate isn't valuable, MIDAS won't save it. --- ## The trajectory I think this takes I'm not going to claim I can predict the next 50 years. I can describe the trajectory that makes the most sense to me from what I'm seeing now. In the next few years, adoption will be slow and concentrated in domains where verifiability has highest value: medical AI, legal AI, financial AI, high-stakes content moderation. Regulators will start asking for FAIM-style attestations as compliance evidence, probably by 2028 or 2029. By the middle of the decade, the substrate will be infrastructure in regulated industries. Insurance won't underwrite AI-mediated decisions without attestations. Courts will accept FAIM records as evidence. It will mostly be invisible to consumers, the way TCP/IP is invisible. By 2040 — if any of this works — most consequential AI decisions globally will be made by agents with persistent FAIM identities. AI-to-AI markets will emerge. AI agents will hire other AI agents for specialist work, paying in substrate tokens. The economic structure of AI will be more visible. By 2055, I suspect the distinction between "the Internet" and "FractalAI" will have blurred, the way the distinction between "the Internet" and "TCP/IP" has blurred. The substrate will be so fundamental it will not require a name. Honestly: I don't know if any of this happens. I think the underlying problem is real and the substrate is one defensible solution. I think someone has to build the first version of it. I'd rather build the version that turns out to matter than build the version that turns out to be perfect but late. If a different group, in a different country, with different technical choices, ends up building what gets adopted instead — fine. The substrate matters. The names don't. --- ## A note on how this was built I want to mention one thing about the process. Substantial portions of the launch you're seeing today — including a complete telemetry system for MIDAS, two operational bug fixes, and drafts of this manifesto and the FAIM-1 specification — were co-developed with an AI agent, Claude Opus 4.7, across the past five days. There are seven commits in our git history with `Co-Authored-By: Claude Opus 4.7 (1M context)` in the trailer. Approximately 1,200 lines of production code. I'm not making any metaphysical claims about this collaboration. The full record is at fractalai.net.co/cofounder and I recommend reading it before assuming anything. I want it documented because it's true, because the entire project is about cryptographically verifiable AI accountability, and because hiding it would defeat the point. This is what radical transparency looks like when I apply it to myself first. --- ## What I'm asking from you If you've read this far, you've already engaged seriously with what I'm describing. Different readers want different things, but here's what would help most: **If you build AI**: implement VAID-1 for at least your highest-stakes outputs. It's a day of work and it gives your users something to verify. **If you regulate AI**: take a look at FAIM-1 as a technical option for the accountability infrastructure your office is trying to build. I'll respond to formal requests for technical engagement within seven business days. **If you write about technology**: visit the dashboard, read the spec, talk to me. I'm not going to oversell this and I'm willing to discuss what's uncertain. **If you invest in technology**: understand what you're looking at. FRAC is not a meme. The economic logic ties to substrate adoption, which is measured in decades. If you're optimizing for next quarter, allocate accordingly. **If you're curious**: visit fractalai.net.co/midas/sentinel and watch the live trade. Form your own opinion. --- ## Why I'm willing to be wrong about this in public I want to close with the honest version. I don't know if any of this matters in ten years. I don't know if FAIM-1 gets adopted or if a better standard replaces it. I don't know if FractalAI is the substrate or just the first attempt. I don't know if I'll still be alive doing this work in twenty years. What I know is that the problem is real. The technical solution is defensible. The infrastructure exists today and works. Someone had to put the first marker in the ground. I'd rather plant the marker and be wrong about its eventual importance than wait for certainty that will never come. In one hundred years, this day will either be remembered as the start of a substrate that became foundational, or as a footnote about an interesting early attempt. Both outcomes are acceptable. The work needed doing. It is now done — at least in its first version. The standards are public. The substrate is live. The first agent is operating. The conversation begins. Welcome to PHI Civilization. — John Romo Bogotá, Colombia May 13, 2026 --- *This manifesto is a living document. Versioned on-chain via FOCI attestation at launch. Off-chain canonical mirror: github.com/johnInarti/FRACTAL-AI/blob/main/PHI_MANIFESTO_EN.md* *The Spanish version is at PHI_MANIFESTO_ES.md* *Critique, corrections, and proposed amendments welcome via GitHub Issues.*