A protocol for disagreement: turning conflict into a decision

Online discussion optimizes for conflict because conflict holds attention. Real deliberation has to do three hard things at once: improve what people know, make their values explicit, and end in a choice they can audit and accept. The way to do that reliably is to stop building a discussion space and start building a protocol.

Inside Commonsent Signal there is a step called deliberation, a structured room where people who disagree work toward a decision. This piece is about how that step is built, and the claim is ambitious: deliberation should be specified like a networking protocol, the way TCP/IP is specified, with interoperable objects, required phases, minimum audits, and certification tests. Done that way, it can become a standard that any city, platform, or institution can adopt and that anyone can inspect. We call it the Commonsent Deliberation Protocol.

01Three targets, all at once

The reason most online deliberation fails is that it aims at one target and misses the other two. A system that can plausibly become a global standard has to hit three at the same time. Epistemic quality: participants update toward better-grounded beliefs rather than more confident ones. Preference legibility: values and tradeoffs are stated openly, not smuggled in disguised as facts. Decision closure: the process ends in a choice that participants can audit and accept as procedurally fair, even when they lost the vote. Hit one and you have a debate club, a survey, or a rubber stamp. Hit all three and you have deliberation worth standardizing.

This is an engineering problem with known failure modes, not a matter of good intentions. The research literature has converged on a handful of stable design constraints, and the protocol is built directly on them.¹ Hybrid designs that combine a small, well-facilitated mini-public with mass participation outperform either alone, because mini-publics raise the quality of reasoning but do not scale, and mass participation scales but needs structure rather than threads. Vote-to-map interfaces, where people cluster around statements instead of teams, reduce performative conflict and surface shared language, as the Polis and vTaiwan deployments showed.² Argument mapping helps, not by making debate prettier, but by turning claims into objects with warrants, evidence, and counters, so reasoning becomes inspectable. Language models can help find common-ground phrasing, but the trust is fragile, and people discount AI-assisted deliberation the moment it feels like manipulation.³ Forecasting adds signal as an input channel, never as a verdict channel. Each of these becomes a rule below.

02The seven phases

The protocol is a pipeline, not a room. An issue enters as raw disagreement and moves through a fixed sequence of phases, each with a defined human output that can be audited and a strictly bounded role for AI. The phases turn a shapeless fight into a decision with a paper trail.

The shape of the pipeline does the work that facilitation alone cannot. Issue compile prevents agenda capture and vague prompts. The shared evidence base separates what is roughly true from what people want before any argument starts. Statement harvesting and overlap mapping convert raw opinion into comparable units and then find coalitions by voting pattern rather than identity, so people discover agreement they did not know they had. The argument graph makes reasoning inspectable. Preference elicitation keeps values explicit and out of the evidence. The decision phase enables closure through package deals, bundles that trade across dimensions so more people can say yes. And post-decision learning builds legitimacy over time by scoring what was predicted against what happened.

03The Issue Object Model

Standardization needs a stable schema. The protocol's foundation is that every issue becomes a small typed database rather than a thread. The separation that matters most in any contested decision, facts versus values versus options, is enforced not as a moderation guideline that someone has to police, but as a data structure that simply does not let the categories blur.

A Claim is an assertion with a scope and a confidence. Evidence carries a source, a method, a quality rating, and an uncertainty. An Assumption states what must be true for a claim or forecast to hold. A Value is a normative premise, explicitly not evidence. A Constraint is a red line or a feasibility limit. Options, Tradeoffs, and Bundles describe what could be done and what each choice costs. A Forecast attaches a scenario and a distribution to its assumptions. And the Minority Report is a first-class, protected object: dissent that the system is structurally forbidden from overwriting. When an issue is built from these pieces, its disagreements become addressable instead of rhetorical.

04AI as clerk, never sovereign

The hardest design question is where to put the language model. Put it in charge and you get a persuasive narrator that people will rightly distrust. Leave it out and you lose the one tool that can make mass deliberation tractable. The protocol resolves this by allowing AI in exactly three roles, each with hard constraints, and forbidding it everywhere else.

The Mediator role is where the documented promise and the documented danger of AI deliberation both live. Synthesizing shared language can genuinely raise agreement, but only if the system can show which inputs drove which phrases, if minority protection is explicit, and if participants vote on the synthesized language line by line rather than accepting it whole. Bound that way, the model is a clerk who drafts and a cartographer who maps. It never becomes the sovereign who decides, which is the only version people will trust over time.

The model drafts and it maps. It never decides. That is the only version of AI deliberation people will trust once they understand it.

05Score bridging, not dominance

Incentives decide what a platform becomes. Reward engagement and you rebuild social media no matter how noble the interface. The protocol rewards epistemic and cooperative contribution instead, through a Contribution Quality Score that is mechanism design used as a guardrail rather than moral theater.

Every term rewards a behavior that healthy deliberation needs and that an engagement metric punishes. Cross-cluster usefulness pays you for persuading people unlike you that you were fair, not for rallying people like you. The steelman delta pays you for understanding the other side before you take it apart. Prediction accountability pays you for being right later, not loud now. For aggregating preferences once positions are clear, the protocol keeps a toolbox rather than a dogma: approval and ranked-choice voting for broad legitimacy, and clearly labeled quadratic-style mechanisms where the intensity of preference genuinely matters.

06Certification, like a benchmark suite

For the protocol to become a standard rather than a brand, its legitimacy has to rest on measurable acceptance criteria, not on whether anyone found a session persuasive. A deliberation that runs the protocol is certified by passing tests, the way a chip is certified or an audit is passed.

The crucial move is what these tests refuse to measure. None of them rewards agreement or persuasion. A deliberation passes by being well-represented, knowledge-improving, honest about polarization, faithful in its summaries, protective of minorities, and capable of reaching a decision people accept as fair. It is judged on the quality of the process and its effects, never on whether everyone ended up thinking the same thing.

07The Deliberative Digital Twin

Before a decision ships, the protocol can run a lightweight simulation pass, a deliberative digital twin of the proposal. Participants attach forecasts to the outcomes that matter, cost, uptake, equity, failure modes. The system builds an assumption graph showing what each forecast depends on, then highlights which assumptions are load-bearing and where the genuine disagreement actually lives. Decision-makers come away with a clean separation: what we agree on, what we do not, and which uncertainties dominate the choice. This uses forecasting as an input that sharpens the decision, never as an authority that makes it, and it connects directly to the simulation machinery the wider system already runs.⁴

08How it becomes a standard

A protocol is only a standard if others can adopt and verify it. The path is the same one open standards have always taken. Publish the Commonsent Deliberation Protocol as an open specification: the object schema, the required phases, the minimum audits. Ship a reference implementation as one open-source stack, with conformance tests that any other implementation can run to prove it complies. Stand up an independent audit program, the deliberation-integrity equivalent of a financial or security audit, so a claim of compliance can be checked by a third party. Guarantee interoperability by letting any city or platform export and import Issue Objects, so a deliberation is portable rather than locked to one vendor. Allow cultural localization, where the objects and phases stay constant but facilitation norms adapt locally. And wrap the whole thing in clear governance about when an output is advisory and when it is binding.

That governance wrapper is the line between a useful tool and a dangerous one. The protocol produces high-quality, auditable, legitimate decisions. Whether a given decision advises a council or binds it is a question for the community that adopts it, made in the open, not a power the protocol grants itself. Like every part of Commonsent, the deliberation engine is built to hand authority to people, never to accumulate it.