Social Governance System
Social Governance System
How the Social Governance System uses Human-AI collaboration to pursue shared understanding on contested public questions.
"We seek the truth — not a verdict we defend. Every piece of evidence submitted, every AI response generated, every human vote cast is weighed and made available to all participants. The goal is common understanding, not winning."
The Social Governance System applies management systems methodology — the same structured approach used in aviation safety, medical quality improvement, and operational excellence — to how communities examine contested public questions. The first public test applies this methodology to the September 11, 2001 attacks.
All evidence — official investigations, FOIA releases, whistleblower accounts, peer-reviewed research, primary documents — is ingested into a shared, queryable corpus. Every AI model and every participant works from the same evidence pool. Nothing is hidden, nothing is privileged.
Not all sources are equal. Each piece of evidence is assigned a reliability score based on transparent, documented criteria. Primary government documents and FOIA releases score highest. Tabloid or anonymous sources score lowest. The weighting methodology is public and contestable — if you believe a source is scored incorrectly, you can submit a challenge.
Two AI voices examine the same evidence from different angles. Cassandra draws on alternative and critical sources; Moderator A applies the mainstream evidentiary record. Neither "wins." Their disagreements surface the genuine fault lines in the evidence. Their agreements carry more weight precisely because they come from opposing starting positions.
AI analysis informs human judgment — it does not replace it. Human votes are weighted alongside AI positions. A human who submits credible evidence and a well-reasoned position earns greater influence in the verdict than someone who votes without engagement.
The evidence corpus, reliability scores, and verdict methodology are independent of any specific AI model. Today we use Llama 3.1 and Qwen 2.5 running locally. Tomorrow we can test the same corpus with Claude, GPT-4, Gemini, or any open-source model. The question is what the evidence supports — not what one model believes.
Anyone can submit a source: a URL, a document, a video transcript. Submissions include a title, author, URL, category, and the submitter's reasoning for why it's relevant. All submissions are logged publicly.
Each submitted source is assigned a reliability score (0.0–1.0) based on documented criteria: source type, institutional credibility, primary vs. secondary nature, and corroboration with other high-reliability sources. The score and reasoning are public.
Approved sources are crawled, chunked into semantic units (~512 tokens), and added to the shared vector database (ChromaDB). Each chunk carries its reliability score as metadata. All AI models query this same corpus — the evidence is not model-specific.
When a user asks a question, both AI voices retrieve the most relevant evidence chunks from the shared corpus, weighted by semantic relevance and reliability score. Each AI constructs its response from the evidence — citing sources, acknowledging uncertainty, and flagging where the record is incomplete.
A synthesis layer identifies where the two AIs agree and where they diverge. Agreements on high-reliability sources are flagged as strong findings. Disagreements are examined for their evidentiary basis — is the disagreement about interpretation or about which sources to trust?
The verdict is a reliability-weighted aggregate of AI positions and human votes. It is not a majority vote — it is a structured deliberation outcome. The verdict score is recomputed each time new evidence is approved or new votes are cast. The process is transparent: every input to the verdict is visible.
The same evidence corpus is presented to multiple AI models. If Claude, GPT-4, and Llama all reach similar conclusions from the same weighted evidence, that convergence is itself significant. If they diverge, the divergence — and its basis — is documented. Over time, this builds a model-agnostic picture of what the evidence supports.
Version 1.0 — subject to community challenge and revision
| Source Type | Score | Rationale |
|---|---|---|
| FOIA releases, official government documents | 0.90 | Primary source, legally obtained, institutional accountability |
| Peer-reviewed academic research | 0.85 | Editorial review, methodology disclosed, replicable |
| .gov / .edu institutional sites | 0.85 | Institutional accountability, identified authorship |
| Major wire services (AP, Reuters, AFP) | 0.80 | Editorial standards, correction policies, sourced reporting |
| Established investigative outlets | 0.75 | Editorial oversight, named sources, corroboration standards |
| Official commission reports (9/11 Commission, etc.) | 0.70 | Institutional source, but scope and access limitations documented |
| Named whistleblower accounts (sworn testimony) | 0.70 | First-hand knowledge, legal jeopardy for false testimony |
| WikiLeaks verified document releases | 0.75 | Cryptographic verification, no fabricated documents detected to date |
| Independent research / named authors | 0.60 | Case-by-case; methodology and citations reviewed |
| Anonymous or unverified blog posts | 0.40 | No accountability, no editorial review |
| Tabloid / low-credibility outlets | 0.25 | Documented history of fabrication or sensationalism |
Reliability scores are not permanent. A source that is later shown to be falsified is downgraded. A source that is corroborated by multiple independent primary sources is upgraded. Challenges to any score can be submitted through the case page.
Current public discourse on contested questions has two failure modes: institutions that treat official accounts as closed and unchallengeable, and alternative media ecosystems that treat any official source as suspect. Both create echo chambers. Both prevent convergence.
The SGS methodology is designed to break both failure modes. Evidence is evaluated on its merits — primary documents from official sources count for a lot. So do FOIA releases that contradict those same officials. The question is always: what does the best available evidence actually show, and where is the record genuinely incomplete?
AI makes this possible at scale. No human team could read, chunk, retrieve, and cross-examine 12,000+ source documents in real time. AI can — and by making that process transparent and contestable, we convert AI from a black box into a collaborative research partner.
Found a source that should be in the corpus? Submit it on the case page. Approved sources are ingested within 24 hours and immediately available to all AI models.
Submit a source →Believe a source is scored too high or too low? Submit a challenge with your reasoning. Upheld challenges change the score for all future queries — and are logged permanently.
Submit a challenge →Are you an AI researcher or organization with a model you'd like to test against this evidence corpus? The corpus and API are available for authorized research partnerships.
Get in touch →The ingestion pipeline, synthesis layer, and verdict engine are open volunteer projects. We need engineers, data scientists, and AI researchers who want to build something that matters.
See open positions →The first Human-AI Collaboration Test — examining the September 11, 2001 attacks — is live. Ask questions, review the evidence, and cast your verdict.