Autonomous Governance, Open Reasoning Architecture
An open research project asking whether AI can help large groups govern themselves without taking over the decision. It specifies a live DAO with an AI synthesis layer, an audit of how democratic DAOs and platforms really are, and a simulator for testing governance designs before they reach people. The shared infrastructure is built and running on 595 real proposals from ten DAOs.
Self-government works when everyone can follow the decisions. Past roughly 150 people that stops being possible, so groups hand decisions to representatives and inherit the principal–agent problem. DAOs were meant to deliver direct democracy at scale. In practice they fail in three consistent ways:
Turnout routinely sits below 10%, and many people vote without reading the proposal.
In mature DAOs the top 1% of holders control 50–90% of the vote.
Related DAOs make contradictory decisions without knowing about each other.
Meanwhile, platform algorithms already govern what billions of people see, with no transparency or accountability. Daopia treats both as the same question: who decides, and who can check them?
The working claim is that AI can amplify democratic decision-making in three ways. It can synthesise knowledge spread across many people, expose concentrations of power before they become capture, and model the consequences of a governance design before anyone lives under it. The constraint is absolute: AI never votes, never decides and never tells anyone how to vote. It is transparency and synthesis infrastructure, nothing more.
The research question is whether AI-assisted decentralised governance can scale to thousands of participants while keeping democratic legitimacy, distributing power fairly and resisting authoritarian capture, and if so, which design choices make the difference. A well-measured "no" counts as a result.
Each mechanism traces back to a body of theory. Habermas sets the conditions for deliberation. Ostrom's nested institutions shape the cross-DAO protocol. Mouffe is why expert disagreement is labelled as contested instead of averaged away. Sen's capability approach drives the complexity score. Zuboff underpins the platform harm index. Winner's "do artifacts have politics?" is applied literally: every design decision record states the political values it encodes.
GOVERN is designed as a live DAO where participants submit, discuss and vote on real decisions about Daopia's own research agenda and treasury. Every proposal passes through a five-step Semantic Engine the moment it is submitted. The engine already runs on the 595 ingested proposals:
The results sit on every proposal page in a Collective Intelligence Panel: classification, affected stakeholders, a dependency graph, and expert contributions grouped as consensus, contested or minority view. The panel gives information and never a recommendation. It never says "vote yes" or "this proposal is bad".
Reputation is earned through participation and explicitly excludes token balance. A test in the codebase enforces that:
Voting starts as plain token voting, a deliberate baseline. The community then votes on whether to switch to a hybrid of quadratic and reputation weighting, so the system has to win its own adoption:
The token is soulbound: it cannot be bought or transferred and has no monetary value. It is earned by contributing, from 0.5 per comment to 20 per accepted expert contribution, capped at 100 a month. Sybil resistance scores every wallet on six signals, including shared funding, near-identical voting and creation time. Suspicious votes go into quarantine, recorded and shown separately rather than deleted, with appeal to a random panel of five trusted participants.
Built: the Semantic Engine running live on Claude, the panel, reputation scoring, Sybil signals, signed ballots, hybrid voting and onboarding with a constitution and literacy quiz. Written and tested, not deployed: the governance and token smart contracts. Not yet: real participants, the forum and the treasury multisig.
AUDIT measures how democratic a system really is. Its main instrument is the Governance Health Index, scored 0–100 across five dimensions:
It is designed for a stratified sample of 50 DAOs across frameworks, domains, ages and sizes, all using public data and every one notified. A second instrument, the Democratic Harm Index, audits recommendation and moderation on YouTube, X, Reddit and TikTok for differential amplification, political targeting and suppression around elections. Daopia is also to publish a monthly self-audit, scoring itself with the same index it applies to everyone else.
Built: the anti-concentration dimension, computed on the ten ingested DAOs, and a public-facing Observatory. Not yet: the other four dimensions, the platform audit and the monthly reports.
SIMULATE is a browser-based agent-based model of a fictional smart city, DEMOS, with 500 agents standing for 50,000 residents. Users pick or build a governance system and watch it handle 20 scenarios over five simulated years. These run from routine budgets through pandemics and migration to a Sybil attack, an attempted authoritarian capture and the recovery.
Direct, representative and liquid democracy; sortition; technocracy; quadratic democracy; and a council with veto rights. Users can combine them.
Agent behaviour is set from GOVERN's real data and must land within 20% of measured values. Every simulated result is labelled as simulated.
A recommendation API lets anyone test a change, such as lowering quorum from 10% to 8%, before proposing it. AUDIT then checks whether the real effect matched the prediction. Status: specified only. It needs behavioural data from GOVERN, and that doesn't exist yet.
A three-layer protocol: a common feed of governance signals in an open schema (GovernanceML), conflict detection across DAOs that informs but never blocks a vote, and optional joint workspaces whose outcomes each DAO adopts on its own terms. Daopia publishes its own feed. Built: the schema and an adapter that reads ten DAOs' public feeds. Signed partners: none yet. Reading a public feed is not a partnership, and the project keeps the two numbers apart.
I wrote Daopia's specification. The system is built by AI coding agents working from it, under these rules. Figures on this page come from the project's own records, dated September 2026.
| Cross-system | |
| HM-1 | Deliberation quality does not degrade as membership grows from 50 to 1,000+.Deliberation depth, minority-voice index, legitimacy and expert coverage regressed on participant count; success is no significant negative slope. |
| HM-2 | AI-audited governance resists capture better than centralised or AI-advisory designs.1,000 simulated runs per design under whale and Sybil attacks; time to detection, share of decisions captured, recovery time. |
| HM-3 | Synthesis AI earns more perceived legitimacy than advisory or decisional AI.Between-subjects study, N=300, on an 8-item procedural-justice scale. |
| HM-4 | Recommendations tested in simulation improve real governance.Pre/post on the targeted metric after adoption; at least 60% must move in the predicted direction. |
| GOVERN | |
| H1-1 | Inviting matched experts raises the use of relevant knowledge by 40–60%.Expert coverage rate before and after expertise matching, at least 20 proposals per period. |
| H1-2 | Dependency detection cuts contradictory passed proposals by 50–70%.Contradiction rate pre/post, coded by two people with Cohen's κ above 0.7. |
| H1-3 | Quadratic plus reputation weighting lowers the Gini of voting power by 30–40% without raising Sybil success.A/B by proposal category, plus quarterly red-team Sybil seeding. |
| H1-4 | Shared signals between DAOs reduce cross-DAO contradictions.Pre/post, same human coding as H1-2. |
| H1-5 | Participants trust synthesis AI more than advisory AI.Within-subjects survey after two months of use. |
| AUDIT | |
| H2-1 | Content critical of authoritarian governments is recommended less.Regression on 10,000+ posts per platform with controls for authority, topic, quality and timing. |
| H2-2 | Identical speech is moderated differently depending on its target.20 matched pairs of true, rule-abiding posts; ethics approval first. |
| H2-3 | Sensational content is amplified 3–5× more than quality content.Observational regression on 50,000 posts. |
| H2-4 | Political content is suppressed around elections.12-month interrupted time series across 5+ elections, entertainment as control. |
| H2-5 | Platforms export the most restrictive local standard globally.30 takedowns in restrictive jurisdictions checked for availability elsewhere. |
| SIMULATE | |
| H3-1 | Most people designing a system choose a hybrid.N=300 designs coded pure, binary hybrid or complex hybrid; success at 60%. |
| H3-2 | Using the simulator improves understanding of the efficiency–equity trade-off.N=200, knowledge test before and after 30 minutes. |
| H3-3 | What counts as legitimate differs by culture.N=500 across 10 countries rating all 7 archetypes. |
| H3-4 | Living through a capture scenario shifts preference toward capture resistance.N=200, rating before and after the regime-shift scenario. |
| H3-5 | At least 20% of users invent structures beyond the 7 archetypes.Two coders classify all 300 designs; novel ones go to GOVERN. |
None has been tested yet. Most need real participants, which the project does not have yet.
The plan runs twelve months from May 2026 on a budget of about €14,000, most of it for paying user-study participants. Phase 0 is complete in code, and a full slice of GOVERN runs end to end: one proposal can travel from submission through real storage, a live model, scoring and the interface. Three steps stand between it and a public link. Those are a cloud database password, a hosting token and a decision on the speed target.
Framing alerts fired on 41 of 42 proposals. A yes-or-no question cannot separate persuasion from concealment. The check now scores materiality from 0 to 1 against written anchors, with the alert at 0.6, and alerts fell to 13 of 42.
Dependency detection found one real contradiction. Among 43 links between the 42 proposals, two were contradictions: a real pair in Gitcoin, one proposal engaging the asset manager Avantgarde and another disengaging it. It found no links across DAOs, which fits, since those DAOs share no infrastructure.
| Link type | Count |
|---|---|
| Modifies an earlier proposal | 13 |
| Supersedes an earlier proposal | 13 |
| Requires another proposal | 9 |
| Implements another proposal | 6 |
| Contradicts another proposal | 2 |
The plain-language rule never fired. Real proposals score 40–55 on complexity against a trigger of 65. The trigger was left alone, because the right test is whether people without the background actually fail to understand a proposal.
Coalition opacity reads about 0.98 for every DAO. Most ballots pass almost unanimously, so the metric confuses consensus with coordination. It is shown with that caveat, and a fix has been proposed: measure it only on contested votes.
No one governs through GOVERN yet, and no alert has been checked against human judgement.
Ten public feeds are read. None of those DAOs has signed on.
A full analysis takes about 18 seconds against a 10-second target. Results stream in, the first after about 6 seconds.
They are written and covered by 17 tests, and they wait on an independent audit before any deployment.
Python and FastAPI; PostgreSQL with pgvector; local sentence-transformer embeddings; Anthropic Claude for analysis, with batch processing for bulk work; Next.js and D3 for the interface; EIP-712 signed ballots; Solidity contracts tested with Foundry; public Snapshot data read through the GovernanceML adapter. Code under MIT, data under CC-BY. The repository will be public once it is deployed.