Does the swarm call the vote?
Passed the pre-registered rule, by 0.5 points
Three DAOs the swarm had never been run on (community, onboarding-dao, nearweek-news-contribution). Before the run we fixed the proposal set (snapshot hashes), the method (source-file hashes) and the metric, and committed their hash. Primary metric: balanced accuracy over 12,133 decided proposals = 50.5%, 95% interval 50.3%–50.7%. Pass rule: above 50% with the interval excluding 50%. Agreement 59.6% against a majority baseline of 59.2%. Recall: 1.2% of approved proposals, 99.8% of rejected ones: the swarm declines almost everything, so we do not read this pass as a success. Blind reviewer, claims shown in both orders: identical verdicts on 100.0% of 72,754 section reviews.
Held-out test DAOs (marketing.sputnik-dao.near, news.sputnik-dao.near): the swarm agreed with the final vote on 63.7% of 5,231 proposals. Saying FUND to everything would score 38.5%; always picking the majority outcome would score 61.5%. Balanced accuracy is 54.1% (50% is a coin toss). The method was frozen at 2026-11-10T08:20:00Z, before the test DAOs were run. Change after freezing: 2026-11-10: bug fix after the first test run. A risk agent that found no red flags emitted no claims, so the reviewer sent its section back and memos were scored without a risk section. Risk agents now always cite the proposer's earlier filings. No weight or threshold changed. Pre-fix test numbers: 63.6% agreement, 53.8% balanced accuracy (commit 05f8d9e).
Pre-registered (never-run DAOs)
| DAO | Reviewed | Agreement | Always FUND | Majority | Balanced | Mean cost | Claims flagged |
|---|---|---|---|---|---|---|---|
| community.sputnik-dao.near | 8 (5/3) | 50.0% | 62.5% | 62.5% | 53.3% | 0.96 N + $0.0006 | 0 / 58 |
| onboarding-dao.sputnik-dao.near | 35 (35/0) | 62.9% | 100.0% | 100.0% | n/a | 1.01 N + $0.0007 | 16 / 286 |
| nearweek-news-contribution.sputnik-dao.near | 12090 (4907/7183) | 59.6% | 40.6% | 59.4% | 50.3% | 1.10 N + $0.0008 | 24,084 / 135,335 |
Test (held out)
| DAO | Reviewed | Agreement | Always FUND | Majority | Balanced | Mean cost | Claims flagged |
|---|---|---|---|---|---|---|---|
| marketing.sputnik-dao.near | 339 (275/64) | 72.9% | 81.1% | 81.1% | 60.5% | 1.05 N + $0.0007 | 250 / 3,132 |
| news.sputnik-dao.near | 4892 (1740/3152) | 63.1% | 35.6% | 64.4% | 49.2% | 1.10 N + $0.0007 | 9,517 / 55,074 |
Dev (rules designed here)
| DAO | Reviewed | Agreement | Always FUND | Majority | Balanced | Mean cost | Claims flagged |
|---|---|---|---|---|---|---|---|
| creatives.sputnik-dao.near | 114 (102/12) | 70.2% | 89.5% | 89.5% | 42.9% | 1.07 N + $0.0007 | 102 / 1,124 |
| ref-community-board.sputnik-dao.near | 66 (60/6) | 48.5% | 90.9% | 90.9% | 56.7% | 1.10 N + $0.0007 | 84 / 647 |
| build.sputnik-dao.near | 30 (29/1) | 46.7% | 96.7% | 96.7% | 24.1% | 1.07 N + $0.0007 | 21 / 275 |
How to read this
Agreement means FUND on an approved proposal, or REVISE / DECLINE on a rejected one. Costs are winning sealed bids in demo NEAR plus an estimate of NEAR AI inference (the scripted agents make no LLM call). Most flagged claims come from quickcheck, a cheap agent whose knowledge base is a present-day export of the DAO: it cites votes that had not happened yet, and the reviewer refuses to pay for that.