Every person, post, essay, poll, forecast, transcript and dollar figure in this demo is invented. Nadia Osei does not exist. Nobody has ever forecast the Providence congestion pricing vote. The numbers are placeholders chosen to make screens legible.
None of it is the point. The content exists only so the mechanisms have something to operate on.
What is real is the reasoning behind each mechanism, which sits in the repository’s docs folder — nine research memos and eleven specs covering what the evidence actually says about misinformation, belief revision, bridging algorithms, and why most interventions in this space fail. If a feature seems arbitrary, the memo explaining it is in that folder.
So: judge the mechanics, ignore the prose. The one thing worth checking in the content is whether it reads as politically even-handed — that was designed for, and there is a ledger tracking it — but that is a hygiene property, not the substance.
Three findings shaped everything.
Trust in news has hit a record low — 37% globally, 25% in the United States — and 62% now say they cannot tell real from fake online, up four points in a year. That figure is rising faster than trust is falling.
This is the harm worth designing against. Not that bad information exists — it does, in volume, and increasingly it arrives from sources with authority and enormous reach rather than from obscure sites. The harm is the second-order effect: once anything can be fabricated, everything becomes deniable, including the true things.
The Alan Turing Institute found no measurable effect of AI disinformation on recent European election outcomes and still identified a real cost — “persistent erosion of confidence in what is real and what is fake.” That erosion is the product opportunity.
In a nationally representative test of ten viral false claims, 82% of Democrats and 81% of Republicans believed at least one. Fewer than 1% identified all ten as false. Which claims landed differed enormously by party. How often people were fooled did not.
Other results point the same way: on basic world-development facts the public scores worse than random chance; in a test of 2,000 people shown real and synthetic media, two got a perfect score; and the very-online score worse than the elderly, not better. Whatever this is, it is not a defect of the other side.
The “backfire effect” largely failed to replicate across 10,100 subjects; corrections generally work. AI dialogue durably reduced conspiracy belief by about 20% in a Science study, holding at two months. A single ten-minute doorstep conversation shifted roughly 1 in 10 voters, durably and resistant to counterargument.
So the barrier is not that minds cannot be changed. It is that changing your mind in public is expensive — permanent, searchable, quotable, and dunked on. And the research is blunt about the cost of the obvious alternative: social exclusion causally increases the beliefs it is meant to suppress.
Everything below follows from a single sentence:
The binding constraint is the social cost of changing your mind. Every feature either lowers that cost or rewards someone for paying it.
That is why this network has no quote-dunk, no ratio, no follower-count prominence, and no misinformation badge — each of those raises the cost. And it is why the platform’s highest-status act is being wrong in public and fixing it.
If you only take one thing from this demo, take that inversion.
Look at the first post. Two things.
Then note what is missing: no like count as a headline, no reply-count prominence, and Disagree is a first-class button that does not reduce reach. Registering disagreement without punishing the author is what makes the ranking work.
In the left rail, “Bridged” ranks by how far agreement spreads across the divide, not how much of it there is. Forty people spanning the spectrum outrank four hundred who agree with each other.
The single most important page in the demo. A plain chronological stream of people changing their minds, with the reason attached. No reactions, no ranking, nothing else.
Scroll to Marcus Adeyemi, three weeks ago — the one where he updates not about a fact but about whether his opponent was arguing in bad faith: “Dana’s version of it isn’t. I still disagree with her.”
Ask yourself what it would take to make that a normal thing to post. That is the entire product question.
Dated, resolvable predictions. Reality resolves them. Records are permanent.
Find the resolved question on state minimum wage. Nadia Osei — who campaigned for the policy — forecast 25% that it would pass, against a community mean of 58%, and was closest. Her note: “Wanting a thing and expecting it are separate questions.”
This is the only unfakeable credential on the internet. Assessment scores are estimates from a questionnaire. A five-year forecast record is a fact, and reality — not a moderator, not an AI panel, not a vote — is the adjudicator. Nobody’s Brier score is red or blue.
Note there is no leaderboard. Distributions and individual calibration curves only.
Write the strongest case for a position you oppose. People who hold it judge whether it could have come from one of them.
Read the near-miss — the submission on immigration and wages that is factually accurate and fails anyway. A judge’s feedback: “Every qualifier here is real, and every qualifier is also how you’d write it if you were trying to make the claim disappear.”
That failure teaches the thing the whole platform is about: passing requires capturing how a position is held, not just what it asserts. In the published research, people who pass are significantly less likely to view opponents as ignorant, irrational or immoral.
Prompts ship in mirrored pairs, always — both sides of every question open simultaneously, so there is never a side being tested.
Members submit positions privately; an AI drafts candidate group statements; members rank, critique, and ratify by vote. In the source research (Science, 2024), AI-drafted statements earned higher endorsement than human mediators and left groups less divided than before.
Two things to look at.
This is also the one place on the site where purple runs at full strength. Purple here means both clusters converged — it is earned, never decorative. A screen with no purple means the community has not found common ground yet, and the palette is telling you so.
How you get in. Four sections, one per scoring layer.
The required field to notice: “What would change my mind.” It is a falsifiability condition as a form field, it is very hard to write for a position you hold tribally, and it filters authorship without gatekeeping it.
The omissions carry as much of the thesis as the features.
There is also a dogpile damper running in the feed spec that no major platform ships: when replies to one person spike from a single cluster, the system stops notifying them, stops distributing to the pile’s neighbours, and tells would-be repliers “37 people have already made this point.”
Almost nothing here is new. What is new is the combination, and which parts were left out. Links open in a new tab so you do not lose your place.
Not the copy. These.
Static HTML, no build step, no backend, no dependencies. Deployed from the
public directory. Every interaction resolves in-page and resets on reload;
anything that would need a server says so rather than faking it. Light theme only,
deliberately. Design tokens live in design/2026-09-visual-direction.
The reasoning lives in docs — start with INDEX.md. The file
00-demo-content-guide.md carries the political-balance ledger in section 7
(currently 40% political across 115 items) and the counting rule, plus a note about the one
thing that ledger cannot catch: it counts topics, not direction, and a screen can sit inside
target while still making one side the object of study. That happened once here and had to be
fixed structurally.