You're a Fractal. That's Why I Built One.

Look closely at a fern. Each leaf is a miniature of the whole frond. Each frond is a miniature of the whole plant. Zoom into a coastline on a map and the jagged edge you see at 1,000 kilometers looks the same at 100 meters. Mathematicians call these shapes fractals: structures where the part contains the pattern of the whole.

Here's what I believe, and what I bet a product on: people are fractals too.

The way you reply to a text you've been avoiding is the way you handle the career decision you've been avoiding. The loop you run when starting a small project — research, doubt, more research — is the same loop you run on the big ones, just at a different scale. There's an old saying: how you do one thing is how you do everything. Fractal geometry gives that folk wisdom a shape. It means something practical and slightly unnerving: you don't need to see all of someone to understand them. You need to catch one true pattern — and follow it inward.

That's what Fractal, the product, tries to do. It catches your smallest verified patterns and uses them as threads to the whole. And unlike every other system quietly building a model of you, it shows you the model, tells you how confident it is, and admits what it doesn't know yet.

This post is how it works, in three parts — the map, the learning, and the verdict.

---

Part 1: The Inner Map — the whole, waiting to be lit

The first thing you see in Fractal is a map of your inner world. Fracty — the AI — sits at the center, surrounded by four areas of your life: career, relationships, wealth, health. Around each orbit the forces that actually drive you. There are twelve, the same twelve for everyone: emotions, needs, fears, motivation, goals, decisions, projects, open loops, habits, identity, values, and contradictions.

What's not the same for everyone is how brightly each one glows.

Brightness is not decoration. Every node carries a score called knownness — 0 to 100 — and the glow you see is literally that number. Knownness falls into five honest bands: Unknown (0–20), Faint signal (21–40), Learning (41–60), Clear (61–80), Deeply mapped (81–100). When you first arrive, almost everything is dark. A node with zero evidence sits at a knownness of 6 — just visible enough to say "I exist, but Fractal knows nothing about me yet."

That darkness is deliberate. Most AI products fake intimacy from the first message — instant flattery, instant "insights." Fractal shows you an honest black map and lets you watch it light up.

A node earns its glow from seven kinds of evidence, each weighted: how strong the signals are (30%), whether you confirmed the observation (20%), whether independent sources agree (15%), persistence over time (10%), recent activity (10%), pattern maturity (10%), and never having been contradicted (5%). Evidence also fades — signals decay on a ~30-day curve — so the map reflects who you are now, not who you were three months ago.

At the top sits one number: how illuminated your map is, weighted by how much each life area currently matters to you. Watching it climb from 12% to 60% over weeks is the core experience of Fractal. It's a progress bar for being understood.

But a dark map raises the obvious question: how does it light up? That's where I have a confession to make.

---

Part 2: The Learning Engine — and the onboarding that almost killed it

My first onboarding was what you'd expect from an engineer: a series of questions about your life, then a request to upload your chat history so the AI could analyze you. Effort first, value later. It made perfect sense to me.

It did not make sense to the people I watched try it. They'd answer a question or two, hit the "upload your history" ask, and quietly leave. And they were right to. I was asking strangers for the most intimate data they own before showing them anything. The product's whole thesis was "one small pattern reveals the whole" — and my onboarding contradicted it by demanding the whole up front.

So I rebuilt the front door around the thesis itself: start with the smallest possible act of self-disclosure, and make it feel like play.

The result is a sorting game — a loving homage to Severance's Macrodata Refinement. Glowing phrases on a dark terminal, generated for you from a few quick answers, that you pull into three buckets: "This is me," "Working on it," "Not me." Thirty seconds of sorting. No forms, no uploads, no essay questions.

But every flick of your thumb is signal. What you claim lands in the model as Confirmed, at high confidence. What you're working on lands as Growing. What you reject lands as Blocked — because knowing what isn't you is signal too. When you finish, Fracty reads the sort back to you in two to four lines — a first, small, slightly-too-accurate observation — and you watch the map take its first breaths of light. Small pattern in, real understanding out. The fractal bet, played in the first minute.

From there, the learning never stops — it just moves into conversation:

Discoveries. Every chat with Fracty ends with a quiet act of bookkeeping: it extracts at most three structured discoveries per exchange, each typed — a goal, fear, blocker, constraint, pattern, strength, motivation, or opportunity. Not a chat log; a claim about you. Discoveries land in a timeline you can see, and — this is the part I care about — you verify or reject them. Confirmed discoveries get promoted into trusted self-knowledge the rest of the system builds on. Rejected ones teach it that it misread you. You are the editor of your own model.

The seven axes — and knowing what it doesn't know. Underneath, every signal lands as evidence on seven axes of your life: energy, agency, resources, meaning, relationships, exploration, stability. The scores aren't the interesting part — the error bars are. For each axis, Fractal computes a Wilson 95% confidence interval over the evidence, the same statistics used for polling. Clarity is defined as one minus the width of that band. Fewer than five signals on an axis? The system refuses to pretend it knows. When Fracty tells you "I know you at 74%," that's not vibes — it's a measurement, and the model is explicitly forbidden from inventing a different answer.

The loop detector. Deepest of all: from your conversations, Fractal extracts the recurring mental states you move through — anxiety, overthinking, researching, doubt, temporary relief, shipped — and the transitions between them. Then it treats your mind as a Markov chain: normalizes the transitions into probabilities and computes where the math says you actually live — the percentage of your time spent in each state, and the loops you get stuck in, like research → doubt → procrastinate → research, each with a stuck-score.

Which is exactly the machinery you need for part three.

---

Part 3: The Verdict — closing the loop

All of this would be a party trick if it didn't come back to you as something you can use. It comes back in conversation — because chat is where Fractal stops being a map and starts being a mirror.

Everything it has learned gets synthesized into a living document called SOUL.md — under 800 words, written in the third person, with sections like Essence, Worldview, Quirks & Signature Phrases, and my favorite: How to Talk to Them, written as direct instructions to the AI. It abstracts instead of records ("actively manages their investments" — never the dollar amounts you mentioned). You can read your own SOUL.md in the app any time. And every conversation with Fracty carries it, along with your live state — energy, momentum, frustration, current mode — which rewrites the rules of the exchange in real time. Energy critically low? Fracty is instructed: two sentences max, no action items, just be present. Frustration high? Do not add to their plate. Validate first. A system that knows you has to know when pushing is the worst possible move.

And then, when the evidence is there, you get what I call a verdict. Not advice. An observation with receipts, delivered mid-conversation or once a day — things like the gap between what you said you'd do and what you did, a topic you've raised ten times without acting on it once, or the loop detector reading you your own transition matrix:

"When you're in research, 71% of the time your next move is doubt."

Nobody enjoys that sentence. I've also never shown it to someone who didn't immediately recognize themselves in it.

The verdict has to earn its way to you: it must cite at least two specific data points from your actual history, fit in three sentences, and pass a gate that bans advice verbs and motivational filler. If nothing clears the bar — silence. One notification a day, maximum. Most days, zero. When Fractal taps you on the shoulder, it's because it saw something real.

Here's the loop, closed: the verdict names one small pattern. You recognize it — or reject it, which is also signal. Either way the model sharpens, the map brightens, and the next verdict cuts closer. One small pattern, followed inward, illuminating the whole.

That's not just how the product works. That's the claim about you: the part contains the whole. You're a fractal. This one just learned to read itself.

---

The contract

A system like this only works if you can trust it with the raw material of your inner life, so the contract is explicit: your data is never sold, never shared, never used to train AI models. And the exit door isn't a support ticket — there's a button in your profile that downloads everything Fractal knows about you as one file, and another that erases all of it, immediately and permanently.

Fractal is live at shinyfractal.ai. The map starts dark. What lights up is up to you.

---

If you try it and the loop detector reads you a sentence you didn't want to hear — I'd genuinely love to know. That's the product working.