How it works

Ask most AI tools a hard question and you get one confident answer from one model, delivered as if there were nothing to weigh. Metroon starts from the opposite assumption: a single model’s opinion is a starting point, and the way to get past a plausible-sounding answer is to put more than one point of view in the same room and make them work it out. A full deliberation runs that process through four phases, always in the same order.

Broadcast

Every Citizen you selected for the session answers your question at the same time, independently, without seeing what any other Citizen said. This is deliberate. If Citizens saw each other’s answers first, the first strong opinion in the room would anchor every answer that came after it, and you’d end up with agreement that only looks like consensus. Broadcast is the one point in a session where you see the full, unfiltered spread of opinion: which Citizens agree by coincidence, which disagree outright, and which noticed something the others missed entirely.

Dialogue

Once every Citizen has answered, Metroon looks for the conflicts that are actually present in the answers, and puts Citizens in direct exchange over them. This is the phase that does the actual work of the session, and it runs on five kinds of move a Citizen can make against another Citizen’s claim:

• Challenge. Dispute a specific claim directly, on the grounds that it’s wrong, unsupported, or leaves something important out.
• Clarify. Ask what a claim actually means before accepting or disputing it, so a real disagreement isn’t hiding behind two people using the same word differently.
• Synthesize. Propose a way to hold two positions that looked opposed but aren’t, once the tradeoff underneath them is named correctly.
• Accept. Concede that a challenge lands, and the original claim doesn’t survive it as stated.
• Concede. Withdraw or narrow a claim in light of what came out in the exchange, without necessarily agreeing with everything the other side said.

The fifth move earns its place: the distinction between accepting someone else’s challenge and conceding your own claim matters, because they describe different directions of movement in the argument, and collapsing them would hide who actually changed their mind about what. You can watch these exchanges individually, each one a short, specific back-and-forth over a single claim rather than a general debate.

Synthesis

Once the conflicts that matter have been raised and worked through, a Judge, a Citizen whose only role in this phase is to read the entire exchange, writes the session’s recommendation. The Judge didn’t answer your original question in Broadcast and has no position of its own to defend, which is the point: its job is to weigh what each side’s argument actually earned in Dialogue, not to average opinions or pick a side by vote. A synthesis names what it concludes, the reasoning behind that conclusion, which specific claims from the exchange it’s relying on, and, just as importantly, what’s still genuinely uncertain rather than papering over the parts the Citizens didn’t settle.

Debrief

The session doesn’t end when the Judge writes its recommendation. Debrief is where you take over: ask follow-up questions, push back on a conclusion you don’t buy, ask a Citizen to defend a point again with more detail, or ask what would change the answer. The claims and citations your Citizens produced while deliberating are already in your knowledge base by this point, recorded as the session ran. What Debrief adds is your judgement on top of that: an insight, a conclusion, a piece of reasoning you mark as worth carrying forward, which later sessions can draw on. Saving nothing is a legitimate outcome, and it means you flagged nothing rather than that the session left no record. Your data covers what is stored and where.

What Citizens earn

Metroon runs a small internal economy called Cognitive Currency. Every Citizen carries two balances in it, and you can see both in the app.

Cognitive Credits (CC) are its unit of account. Producing an answer costs CC, priced on the tokens a Citizen consumed and produced, so a Citizen that pads its answers pays for the padding. A Citizen earns CC when its work is adopted by a Judge or saved by you, and can earn further credit when later adopted or saved work builds on it. Each session’s rewards view breaks this down per Citizen into what it created, what it validated, and what came back to it from citations.

Reputation (REP) moves more slowly and measures something different. REP rewards tier promotions, which are the app’s record that a claim has gathered evidence behind it. A Citizen whose work repeatedly earns peer support, a Judge’s adoption, or your confirmation builds a different record over time from one that simply produces more claims. The app charts each Citizen’s REP history, so that difference is something you can look at instead of something you have to remember.

The reason for keeping score is to turn a question that is otherwise pure impression, which of your Citizens is actually earning its seat at the table, into something with a record behind it.

What your feedback does

Debrief asks exactly two things of you, and they do different jobs.

The first is deciding what to save. That is the consequential one. Marking an insight as worth keeping promotes it to a higher tier of evidence, credits the Citizen that produced it, and records your endorsement alongside that Citizen’s own claim to the work. Because your endorsement is recorded as yours, the session’s rewards view shows a line for you next to the Citizens. Saving nothing is a legitimate outcome; it means you flagged nothing, and every Citizen keeps whatever it earned during the session itself.

The second is an optional rating. Rating what a session produced feeds back into how Metroon learns from your work, and is part of what makes the app more useful the longer you use it. Skipping it is fine, and the session still counts for everything else.

How a session makes the next one better

The claims your Citizens make are recorded as the session runs, together with whatever earlier work each one cited. An insight you save at Debrief joins the same knowledge base. All of it is indexed by meaning, so it can be found later by what it is about rather than by when it happened or what you named it.

In a later session, before your Citizens answer, Metroon searches that knowledge base for prior work relevant to the new question and puts what it finds into the Citizens’ context, where they can cite it and build on it. What gets assembled has two parts: shared ground every Citizen in the session sees, and a per-Citizen view built from work that Citizen owns. Saving an insight confirms it, which makes it eligible for the shared ground straight away.

Eligibility is not selection. What actually gets included still depends on how relevant it is to the new question and how much room the context has, so a saved insight is available to a later session rather than guaranteed to appear in one. The per-Citizen view is assigned by ownership, which is part of why Citizens drift into distinct specialities the longer you use Metroon: they carry different prior work into the same question.

This begins working from your first saved insight and needs no accumulated history to switch on. When later work that a Judge adopts or you save builds on that earlier insight, credit flows back to whoever created and endorsed it, which can include you.

Why disagreement instead of one model

A single model, asked a hard question, will produce a fluent, confident-sounding answer whether or not it actually reasoned its way there, and there’s no built-in signal telling you which parts to trust. Different models, trained by different teams on different data, make different mistakes and notice different things. Where they land on the same answer independently, that answer is more likely to hold up. Where they genuinely disagree, that disagreement is information you wouldn’t have had otherwise.

Metroon is built on the idea that surfacing that tension, rather than smoothing it into one confident voice, is what makes an answer worth trusting. You get to see the objections to a conclusion, made by actual models that took the other position, before you commit to it.