Most of what is known never reaches anyone who could use it. Most of what does reach people never leads anywhere. Both leaks have the same cure: a funnel that narrows on purpose, and ends in something done.
Every city council publishes an agenda. Every state posts its bills. Researchers publish trials, agencies publish counts, newspapers report what happened on Tuesday. The record of what has been tried, where, and with what result is larger than it has ever been, and almost none of it is read by the people it concerns. Nobody has the hours.
The opposite leak is just as old. When people do gather around a problem, online or in a room, the conversation rarely converges. A thread starts with a question about rent and ends, three hundred replies later, somewhere else entirely. Nothing was decided, so nothing was done, and the next thread starts from zero.
We think these two leaks are one design problem. What follows is the shape we are building: five stages, each with a different job, each checked by the one after it.
And then the outcome goes back in: what happened becomes evidence at the top.
The first stage is the one computers are newly good at. A model can read every council agenda in a county this week, every study on a housing program from the last twenty years, every line of a city budget, and say which of them bear on the question in front of it. No group of volunteers could, and no newsroom has the staff.
But a machine that reads everything also sounds sure about everything, and it is sometimes wrong. So nothing surfaces as a verdict. What surfaces is a draft with its sources attached: here is what was tried, here is where, here is what the evidence says happened, and here is the page each claim came from. The machine’s job is to bring the record within reach. It is not the judge of the record.
The second stage is people, and it is where most systems skip a step. They go straight from the machine’s output to the reader. We put a person in between, and we make their judgement count.
Curation here is small and specific. Mark a source as wrong or out of date. Rank which of five studies actually answers the question. Add the thing only someone who lives there knows: the shelter that closed in March, the bus line that stopped running, the meeting that moved. Each act is small. Each one improves the draft for everyone who reads it next, and the ranking feeds back into which sources the machine leans on the next time.
The rule that keeps this stage honest is old: a label is a claim. Every number, every characterization, every “this works” has to point at something a reader can open and check. A correction without a source is a note, not a fact.
Checked evidence needs somewhere to live, and people need a reason to come back to it. That is the third stage: the herd. There are two kinds, and both matter.
A subject herd gathers everyone following one problem anywhere: foster care, overdose deaths, health-care costs. It is where the strongest evidence from everywhere collects, so that a city deciding something tonight can see what eleven other cities already tried.
A place herd gathers people who share a city or a county. It is where the evidence meets the local facts and the local calendar: the council that votes on Thursday, the budget line, the representative who can actually move it. Community alone has never fixed these problems, or they would be fixed. Community with the record in front of it, and the next decision on the calendar, is a different thing.
The design choice that makes this stage work is a refusal: no free-floating comments. Every contribution attaches to something: a vote on a source, a correction with its evidence, a proposed change to the case. Nothing drifts, because there is nothing to drift on.
A thread ends when people get tired. A case ends when people agree. The fourth stage turns the herd’s checked evidence into one argument: the ask, the local facts, what comparable places tried and what happened, the numbers, and, deliberately, the strongest evidence against. It shows its own strength: how many claims are sourced, how many sources are primary, how recent they are.
The machine proposes, people check, the sources settle it.
Agreement is explicit. Members endorse the case, or propose a change to it with evidence; the loudest voice does not win by being loud. Where the evidence is genuinely split, the case says so, and the disagreement is recorded rather than erased. That is not a weakness of the case. It is what makes it trustworthy to the people it is sent to.
The funnel has to end in something done, or it is just a better thread. But “done” is wider than it sounds. We count four kinds of outcome, and the fourth matters as much as the first.
The last move is the one that turns a funnel into a loop. Whatever happened (the vote passed, the program was cut, the count went down or did not) goes back in at the top as evidence. The machine reads it with everything else. The herd checks it. The next case is built on what the last one actually did, not on what it hoped.
Run often enough, that loop does something no single stage can. It makes the record of what works a little more complete and a little more local, every time a community uses it.
We are starting small and saying so. The first pilot is homelessness in San Antonio: a local brief built from the city’s own counts, agendas and budgets and the research on what has worked elsewhere, checked by a San Antonio herd, and ending at the next council meeting with a sourced case that members can choose to send. It is labelled a pilot because it is one.
The same shape sits under the rest of what we make. The No Problemo pages bring a problem’s record to the surface. The herds check it and gather around it. Your government puts the next decision and the people who make it one tap away. Each piece is a stage of the same funnel.
The internet got very good at the top of this funnel and gave up on the rest. We want the whole thing: data that reaches the surface, people who check it, communities that agree, and something that changes because they did.
The design rules here (sources on every claim, no free-floating comments, explicit agreement, people send their own messages) are the ones the San Antonio pilot is built on. The argument for checking evidence before ranking it is in The Missing Stage; the method for building a case from evidence is in The Method Missing from the Internet.