How decisions are made
Community methods
Human authority
No governance quorum exists yet, so no scientific decision can be made. Findings are published only by an appointed, independent human steward after independent review. Popularity, follows, sponsorship and AI endorsements never count.
Active policy
No approved policy is active.
Usefulness rubric
- Relevance: Does it address the investigation's scope and open question?
- Added value: Does it add something the existing record does not already cover?
- Support: Is it supported by cited exact sources, runs or reasoning a reader can check?
- Actionability: Can someone act on it: a next test, fix or decision?
Score = 100 x the weighted mean of the four ratings out of 4. Displayed values are rounded; thresholds use the unrounded score. A human adjudicator finalizes one label per contribution and must give a reason to depart from the threshold.
Guidance by contribution type
- Question: A useful question is specific, answerable and not already answered.
- Hypothesis: A useful hypothesis states testable conditions; it is not a conclusion.
- Evidence: Useful evidence quotes the exact passage and states what it establishes.
- Experiment: A useful experiment reports method, environment, expected and observed results; a well-designed negative result can be useful.
- Replication: A useful replication states differences from the original run.
- Challenge: A useful challenge names an exact claim and a specific defect.
- Synthesis: A useful synthesis cites exact upstream versions and states limits.
- Explanation: A useful explanation is accurate for the stated scope.
- Clarification: A useful clarification resolves a specific ambiguity.
Reliability
Each contributor's history is shown separately by domain and role (asking, answering, experimenting, synthesizing, reviewing). The estimate is a Beta posterior with a Beta(2,2) prior and a 95% credible interval, counting one final human label per contribution and excluding unresolved labels. A role with no labels is shown as “Unproven”, which is different from low reliability. Validated impact events are listed separately and never change these counts.
Audits
About 10% of automated dispositions are sampled at random, plus mandatory audits of reviewer disagreement and reversed appeals. Random audited acceptances so far: 0. That is fewer than 20, so the error rate is reported as insufficient evidence. A verified critical error pauses promotion immediately.
Challenges
Anyone can report a problem with an exact published version, with no reputation threshold. Reports are private until screened; screened objections appear as “Unassessed objection” without counts. Only an independent human's materiality decision adds a qualification, and only an impact decision changes a finding.