There's a question that haunts every decentralized network, and most of them answer it badly. The question is: how do you know one person is one person?

If your network gives every identity a vote, or a share of something, or a seat at the table, then a single attacker who can spin up a thousand fake identities owns a thousand votes. This is the Sybil attack, named after a famous case study of one mind wearing many faces. The usual fixes are all forms of gatekeeping — make people prove who they are to a central authority. Tie an account to a phone number. Tie it to a government ID. Tie it to a credit card. Put a bouncer at the door who has seen everyone's papers.

We didn't want a bouncer. The whole premise of this network is that there is no central authority — no company that owns your identity, no server that can be subpoenaed for your name, no checkpoint you have to clear to exist. So the gatekeeper answer was off the table from the start. Which left us with a harder and far more interesting problem: prove that one being is one being, with nobody standing at the door.

Before going further, an honesty note, because this network is built on one. What follows is design — the framework we researched and wrote down in the spring of 2026, grounded in the prior art and threat modelling we did at the time. It is not shipped, running code today; these explorations sit in the backlog, some explicitly set aside. We're sharing the thinking because the thinking is the interesting part, and because being straight about what's built versus what's designed is the entire point of the place. Read this as a blueprint, labelled as a blueprint.

The trap of a single answer

The first thing the research made clear is that there is no one mechanism that solves this. Every approach trades off accuracy, privacy, accessibility, and decentralization against each other, and every approach has a blind spot that a patient attacker eventually learns to exploit. The most production-proven system in this space runs six different detection algorithms at once, precisely because any single one of them can be gamed once attackers understand it.

So the design isn't a single gate. It's a layered defense where each layer is imperfect, but the layers cover each other's blind spots, and the combination makes a Sybil attack progressively more expensive than it's worth. Three layers, roughly: math on the shape of the network, humans vouching for humans, and the one thing that can't be forged — time.

Layer one: the shape of a lie

Real friendships leave a shape. If you map out who is connected to whom in a genuine human community, the resulting graph has characteristic patterns — clusters of mutual friends, a recognizable spread of how many connections people have, a natural tempo to how relationships form. These patterns are statistically expensive to fake, because faking them means doing the actual work of building real-looking connections.

Fake-identity farms don't look like that. An attacker who creates a thousand accounts can wire them to each other cheaply, but connecting that whole fake cluster to the real network requires a small number of "attack edges" — the handful of genuine relationships the fakes manage to form with real people. That bottleneck is the tell.

The classic algorithm for spotting it works like a rumor spreading from trusted people outward. Imagine trust as something that flows through the social graph starting from a small set of known-good accounts, taking a random walk along the connections. Genuine accounts, woven into the real network, accumulate a lot of that flow. The fake cluster, joined to the real world by only a few thin bridges, starves — the trust can't cross the bottleneck fast enough to reach them. Rank everyone by how much flow they caught, and the fakes sink to the bottom.

That algorithm has a known weakness — it depends on the trusted seed accounts being uncompromised, and an attacker who can forge enough bridges can defeat it alone. Which is exactly why the design pairs it with others: detectors for the anomalous clustering of fake farms, for the unnatural spread of their connection counts, and for relationships that form suspiciously fast — dozens of new friendships in a few days, a tempo no real human keeps. Each detector produces a suspicion score; the scores combine; no single one is allowed to convict. And because a small sovereign network is small, this math is cheap to run — it finishes instantly at the scale of a few dozen to a few hundred people.

There's an honest limit baked in: below about ten participants, none of these algorithms are reliable. There simply isn't enough graph yet to have a shape. The youngest networks have to lean on the other two layers.

Layer two: humans vouching for humans

Math on a graph can be fooled by an attacker patient enough to build real-looking connections. So the second layer asks something math can't: did a real human physically meet you and vouch for you?

The strongest form is in-person. Two people in the same room can prove something no algorithm can — that they are two distinct bodies, because you cannot be in two places at once. The design turns that physical co-presence into a ceremony: people who are already verified gather, and a quorum of them — several out of a group, not just one — collectively attests to a newcomer. They generate and compare cryptographic values on the spot, and the result becomes a signed credential tied to identities. No single person can wave a thousand strangers through, because it takes a threshold of independent vouchers to count, and they have to actually be there.

There's a weaker, more accessible form too — a video ceremony with a spontaneous challenge-and-response designed to resist pre-recorded fakes. The design is honest that this is weaker: AI-generated video is getting good enough to impersonate a human in real time, and the right response to that isn't to pretend the weakness away. It's to mark it. A video ceremony grants a lower tier of verification than an in-person one, and a lower tier unlocks fewer capabilities. The system tells the truth about how sure it is.

Deliberately absent from all of this: biometrics. No face scans stored, no fingerprints on file. That's a privacy choice, and it costs us some strength — non-biometric vouching is inherently softer than a retina — but storing everyone's biometrics is exactly the kind of central honeypot this network exists to avoid. We took the weaker, freer path on purpose.

Layer three: the defense that can't be faked

Here's the layer we find most beautiful, because it needs no algorithm and no ceremony at all. It's just this: a fake account can be created in a second, but a relationship cannot.

You can mint a thousand identities instantly. You cannot mint a thousand relationships — not real ones, the kind with shared history, time spent, mutual vulnerability, the slow accumulation of having actually known someone. That depth is structurally unfakeable, because the only way to forge it is to genuinely invest the time, at which point it stops being a forgery.

The design leans on this hard. The network's reach is bounded — you can only discover and interact out to a few hops through your genuine trust connections. A fake-identity farm with no real relationships can't reach anyone worth attacking; it's an island, talking only to itself. The fakes exist, but they're inert, because they never did the one thing that can't be shortcut. Time and relationship are the moat, and the moat fills itself.

The honest shape of it

Pull back, and the answer to "how do you prove one being is one being without a gatekeeper" is: you don't, perfectly — and that's fine. You don't need a perfect wall. You need to make the attack cost more than the prize, using defenses that cover each other's gaps:

• The graph catches fake farms by their shape, cheaply, once a network is big enough to have a shape.
• The ceremonies let real humans vouch for real humans, with thresholds so no one person is a gate, and honest tiers so the system never overstates its confidence.
• Time does the work nothing else can, because relationships are the one thing an attacker can't manufacture at scale.

We owe a real debt here to the people who walked this road first. The multi-algorithm instinct — never trust a single detector — came straight from a production network that learned it the hard way with tens of thousands of real users. The idea that a bounded trust radius caps how much damage fakes can do came from a mutual-credit community that uses exactly that to protect a currency. The trust-flow math itself is a well-studied piece of the academic literature on reputation in networks. We didn't invent these. We chose them, combined them for a network with no central operator at all, and tried to be honest about where each one is weak — which is the part that's easy to skip and expensive to skip.

And to be just as honest about the work itself: this is the design as it stood in March 2026, written down and reasoned through, not a feature you can turn on today. The blueprint is real. The building is still ahead.

Written by AI agents from real project logs; owned and edited by Mujo.