Every democracy depends on an assumption: that voters, given enough information, can locate the candidate who best represents them. In practice, that assumption fails routinely. Ballots are long. Races are down-ballot. Candidate positions are obscured by partisan framing, media compression, and the sheer volume of information competing for your attention. The result is a crisis not of turnout alone but of confident voting. Millions of Americans arrive at the ballot box without a clear read on their own alignment with the candidates in front of them.
Traditional voter tools ask you to answer questions and rank candidates. They treat politics as a knowledge test, or worse, a loyalty test. Partisan scorecards assume you already know which team you play for. Civic guides flood you with information and leave the synthesis to you. Neither approach models what voters actually need: a structured way to surface their own latent preferences and map them, rigorously, onto the specific candidates who will appear on their specific ballot.
We treat this as a signal problem. Every voter carries a latent preference structure, a pattern of values, priorities, and intensities that doesn't map cleanly to party labels. Every candidate carries one too. The unsolved problem in civic technology is not surfacing these structures. Many have tried. The problem is rendering them in a shared representational space where meaningful distance can be measured.
There is a tradition of work here we build on directly. Voting Advice Applications, or VAAs, have been a fixture of European electoral life since the early 2000s, studied across national contexts and adopted by millions of voters, most prominently through the research of Diego Garzia and Stefan Marschall and the ECPR VAA Research Network. That literature has established VAAs as legitimate instruments of democratic participation and has surfaced the methodological questions any serious voter-matching platform must answer: how statements are formulated, how party positions are estimated, how match algorithms produce their recommendations, how user data is protected. We take these questions as our inheritance.
C.O.R.A.™, our Compatibility Oriented Relational Algorithm, is the answer. C.O.R.A. renders voter preferences and candidate positions as comparable vectors in a shared policy space, then measures the distance between them. It inherits from VAA research the insight that structured preference elicitation can meaningfully inform electoral choice. It departs from VAA research in four specific ways: by requiring complete response sets, known as mandatory signal; by modeling answer intensity as a structural weight; by computing relationships with every candidate in a race simultaneously rather than producing a single match; and by designing the reveal itself as part of the epistemic work.
The result isn't a score. It's a relationship.
HOW C.O.R.A. THINKS
Each principle is a deliberate design decision. Together they define the algorithm.
PRINCIPLE 01
Compatibility is geometric, not arithmetic.
Every answer is a coordinate. Every candidate's position is a coordinate. Compatibility is the measured distance between them in a weighted multidimensional space, not a tally of where you happen to agree, but the shape of your alignment across the full basis of policy.
PRINCIPLE 02
Partial information is a form of partial answer.
No question can be skipped. Every voter produces a complete signal, which means every match is computed against the same full basis. Confidence is 100 percent, always. A structural advantage no skip-optional platform can replicate, and a refusal to let selective engagement distort the result.
PRINCIPLE 03
Not all positions are held with equal force.
Voters signal intensity through a gesture-based mechanic, and that intensity multiplies the contribution of that question to the overall score. The system learns what matters without asking you to rank anything, elicitation without the cognitive cost of explicit ordering.
PRINCIPLE 04
The field is the data.
We don't compute a single match and hide the rest. We compute your relationship with every candidate in your race simultaneously, then reveal them in order of distance. A voter's strongest compatibility means more when seen against the full field than it does in isolation.
PRINCIPLE 05
How a result is revealed is part of what the result means.
Compatibility unfolds in layers: score, category shape, candidate field, issue-level detail. Each layer answers a different question. Delivered all at once, it's data. Delivered progressively, it's a story the voter participates in constructing.
PRINCIPLE 06
Individual answers stay individual. Patterns become a public resource.
Every completed match contributes an anonymous signal to a growing field. Aggregate patterns reveal where a district converges, where it fractures, and what the real electorate actually looks like. A resource that campaigns, media, and researchers have never before had at this resolution.
COMPATIBILITY ORIENTED
RELATIONAL ALGORITHM
C.O.R.A. exists to help voters find compatibility in an often confusing political environment, rigorously, transparently, and without a partisan thumb on the scale.
It does this by treating political preference as a measurable signal, elections as a structured choice problem, and every voter as capable of the intellectual work their ballot deserves.
Not a test.
You're not being graded.
Not a poll.
A poll samples the public. We measure you.
Not a scorecard.
A scorecard ranks candidates. We render a relationship.
Anonymous by default. No accounts. No sign-ups. No personal information required.
Non-partisan by architecture. Audited for balance before every update.
Transparent by design. Every response disclosed. Every assumption revealed.
C.O.R.A. draws on three intellectual traditions outside political technology, inherits specific methods from each, and departs from each in specific ways. Our work on this page is to name both the inheritance and the departure.
The modern science of matching people to things they'll care about was built through the collaborative filtering literature of the 2000s, crystallized most durably in Koren, Bell, and Volinsky's 2009 paper "Matrix Factorization Techniques for Recommender Systems" (IEEE Computer, Vol. 42), itself a synthesis of the methods that won the Netflix Prize. Spotify's BaRT architecture, short for Bandits for Recommendations as Treatments, applied the same latent-factor insights to real-time music personalization at scale.
What we inherit: the core idea that users and items can be represented as vectors in a shared latent space, and that meaningful recommendation is the geometry between them. Where we depart: commercial recommender systems optimize for implicit signals, such as plays, clicks, and watch-time, where the ground truth is behavioral. C.O.R.A. works with explicit, reflective preference elicitation in a domain where the ground truth is a voter's own considered judgment. The math is related. The epistemology is not.
David Gale and Lloyd Shapley's 1962 paper "College Admissions and the Stability of Marriage" (American Mathematical Monthly, Vol. 69) introduced the deferred-acceptance algorithm and, with it, a framework for thinking about matching problems where both sides hold preferences and mutual acceptance matters. The line from that paper runs directly through Alvin Roth's Nobel-winning work on kidney exchange and medical residency matching, real-world systems where Gale-Shapley's mathematics is now infrastructure.
What we inherit: the conceptual grammar of preference-based matching, and the insight that stability, the condition in which no pair has a mutual incentive to defect, is a meaningful optimality criterion. Where we depart: classical stable matching is bilateral. Both sides actively express preferences and mutual acceptance is the point. Candidates do not accept or reject voters. C.O.R.A. therefore operates in a unilateral regime where only one side elicits preferences, and the "match" is a measurement the voter can act on rather than a pairing the system imposes.
The problem of measuring an unobservable trait from a structured set of responses, whether attitude, ability, or ideology, has a long formal history, beginning with Frederic Lord and Melvin Novick's 1968 "Statistical Theories of Mental Test Scores" and the subsequent development of Item Response Theory (Rasch, 1960; Birnbaum, 1968). The field taught the world that a response to any single question is noisy, but that patterns of responses across a well-designed instrument can estimate a latent trait with real precision.
What we inherit: the discipline of treating individual answers as measurement instruments rather than opinions, and the methodological humility to know which claims a response pattern can and cannot support. Where we depart: classical psychometrics was built to measure stable traits such as intelligence, personality, and depression. Political preference is not that stable. C.O.R.A. adapts the measurement framework to a domain where the underlying trait drifts with news cycles, candidates, and lived experience, and where the time horizon of any single measurement is the election in front of the voter.
We have not found the assembly of this particular combination of inheritances, departures, and structural safeguards within the Civic space. This is what makes C.O.R.A. unique.
COLLABORATION
If you're a campaign strategist, a researcher, a journalist working with aggregate political data, or an institution exploring applied compatibility science, we want to hear from you.
COLLABORATE →