What enters the model
Inputs can include official records, public datasets, reporting, research, expert analysis, firsthand observation, community testimony, photographs, video, historical records, and structured claims from participants.
Different inputs answer different questions. A budget document may establish an expenditure; testimony may reveal an impact the document cannot. The system should not flatten those differences.
Provenance before presentation
Every consequential claim should retain a chain back to its source: who or what produced it, when, under what conditions, and whether it has been altered, disputed, or superseded.
- Primary and original sources are distinguished from summaries.
- Independent corroboration is visible.
- Conflicts of interest and material limitations are attached to the evidence.
- Unverified claims remain labeled as claims.
Evidence is weighted, not counted
Collective Signal is not majority rule for facts. Ten repeated articles tracing back to one unsupported statement do not become ten independent sources. Weight should reflect relevance, proximity, reliability, independence, and corroboration—not popularity alone.
Disagreement remains visible
The output should distinguish what is well supported, reasonably inferred, contested, unknown, and unknowable with current evidence. Minority interpretations should remain available when responsibly supported, even when they do not control the summary.
Every model has a history
Updates should identify what changed, why it changed, and which evidence caused the revision. Earlier states remain part of the record. This allows accountability without punishing honest correction.
AI assists; people remain accountable
Machine systems can help locate, deduplicate, structure, compare, summarize, and test information. They should not quietly determine what the public sees. Significant automated judgments require documented criteria, testing, human review, and a way to challenge them.