Civic AI Commons is a proof instrument. It tests whether a documented solution in one field can be found by someone working in another, without either of them knowing the other field exists.
Rings group records by the field each source paper was studied in. That is the provenance of the paper and not the domain of the mechanism, so a line between two rings means two literatures met. Whether the mechanism itself crossed a boundary is established by reading the sources, never by the grouping. Hover any node, or any line except the unverified proposals.
A housing office turns people away because the paperwork defeats them. A clinic loses patients between referral and appointment. An election office watches eligible voters fail to register. Four institutions, four literatures, one structure. Each has partial answers the others have never read.
That is not a knowledge problem. Every one of those fields has done the research. It is a retrieval problem, and retrieval is what a corpus is for.
So the project builds records from published sources, describes each one by the mechanism it uses and the institutional failure it corrects, and then looks for pairs that match across fields. Every proposed match is checked by a person against published evidence before it counts. Nothing counts because the software suggested it.
Look at how cross-field transfers have actually happened and a pattern shows up. Someone spent a decade running epidemic response, came home, and was asked about street violence. A physician sat on a police reform committee, saw crime maps, and asked whether health data would look the same. A mayor saw a policing system in one city and hired the person who built it.
In each case the transfer waited for one person to happen to hold two fields in their own head. That is a slow and unreliable way to move a working idea, and it does not scale with need. It scales with careers.
If this works, a practitioner facing a documented problem could see which other fields have the same structure, what each of them tried, what it required, and where it broke. Then they decide, because deciding was never the part that was missing.
The corpus is young and has not met practice. This is the argument for the method, not a description of a finished product.
A confirmed transfer needs a document that says the move happened. Not two fields that resemble each other, and not a plausible argument that one could have borrowed from the other. A source has to state it.
Track crime by place and time, hold a named commander accountable for what the numbers show, and meet weekly to act on them. Built to reduce crime in New York.
The same loop pointed at potholes, trash collection, and building inspections. Baltimore's incoming mayor hired one of CompStat's creators rather than copying the idea from a distance.
The mechanism that moved was not policing. It was the loop: measure by place, name someone accountable, meet often enough that the measurement changes what happens next. Nothing about that is specific to crime, and nothing about it was obvious to city services until somebody carried it across.
A specific way of conducting a conversation, in which the practitioner draws out the person's own reasons for changing rather than supplying reasons for them. Developed for problem drinking.
The difficulty is not that patients lack information. Taking medication daily for a condition with no symptoms requires a reason the patient actually holds, which is what the technique elicits.
The two fields share almost nothing. Different training, different journals, different clinics. What they share is a person who knows what they should do and does not do it, which is a structure and not a subject.
Two records sharing a mechanism label, sitting in different categories, became a proposed transfer. It produced sixty one proposals.
The mechanism label was written when the record was created, before there was text to support it. The category records which field the source paper was studied in, not where the mechanism belongs. A pair built from those two was a claim about two papers.
A narrower question, asked from the problem side rather than between two solutions, and answered by similarity of mechanism text corrected for records that sit near everything. Four of the five confirmed transfers came from that ranking, each surfaced in a problem’s top three and then confirmed by a person reading external sources. The fifth was built by hand after research on a different pair turned up a lineage attached to records nobody had connected. Neither route runs on a schedule; the last time either proposed a pair was 2026-08-06.
Fourteen of the sixty one proposals were checked against published sources. None confirmed. That generator is retired, and the proposals it made are still drawn above as hairlines, because a proposal is a proposal and drawing it is not a claim that it is real.
No new candidate pair has been proposed since. The work since has gone into checking the pairs that already existed, and the rest of this page is what that found.
Corrected 2026-08-31. Until today this panel named a different route — a matcher over paired institutional-failure vocabularies — as the source of the four. It was not. The four were traced to the similarity ranking described above, and the matcher could not have produced them: all four problem records carry no institutional failure tag at all, which is the one thing that matcher requires.
Confirmed transfers stand at five. That is down from six. One of them was restated after its own evidence was read closely: the source described an adaptation of a treatment model for a different population, which is a real and documented lineage but not a crossing between fields. It still counts as a transfer. It no longer counts as a cross-domain one, so it is not gold in the graph above: it is drawn as a hairline, exactly like the unverified proposals, and the distinction is carried by hovering it rather than by its colour.
Twenty records were read against the sources they cite. Ten of them claimed more than the source supported. That is a floor rather than an estimate, because every disputed reading in that exercise ran in the same direction, toward the record.
Most of the five confirmations link to at least one record whose cited paper is about a different aspect of the subject, or a different subject, than the record claims. One is clean on both sides. Those defects do not touch the lineages, which were established from external documents rather than from the records, but they do mean a reader following a confirmation into the corpus can land on a weaker source than the confirmation rests on.
None of that was found by an outside reviewer. It was found by running the checks on ourselves and writing down what came back. A method that has never been shown to fail has not been tested.
There is no working measure of how often a mechanism actually crosses a field boundary, because the field that would have to carry that information records something else. A replacement has been specified and not built.
The vocabulary used to tag institutional failures is not as controlled as it should be, and some records carry free text where a fixed tag belongs.
Every confirmation so far required a person to find and read an external document. Nothing in the system has ever produced a lineage on its own, and the parts that could be automated are the parts that fetch and surface, not the parts that judge.
There is considerably more to do here than has been done.