Our Methodology
VoterHQ is built on a trust-first principle. Every piece of information shown to voters passes through a rigorous pipeline of sourcing, AI-assisted structuring, human review, and ongoing correction. This page documents that process end-to-end.
Sources Used
VoterHQ draws exclusively from official, publicly available government and nonpartisan sources. We never use news opinion pieces, social media posts, or campaign materials as primary sources.
Primary source types
- State Secretary of State official voter guides
- Legislative analyst fiscal impact reports
- State and county election office publications
- Official ballot measure text (enrolled legislation)
- Federal Congressional Record and congress.gov data
- Google Civic Information API (elections, polling locations)
- OpenStates API (state legislator data)
All source documents are ingested through our admin pipeline, chunked for retrieval, and tagged with jurisdiction and document type metadata. Every chunk carries a source URL back to the original publication.
AI Models
We use Anthropic Claude models for two tightly scoped tasks. The models never browse the web, never access training data about elections, and are constrained to operate only on retrieved verified source chunks.
Ballot Measure Explainers
Claude Opus 4.7 generates structured 3-bullet plain-English summaries grounded strictly in verified chunks. Every bullet must include citation markers referencing the source chunk. The system prompt forbids partisan characterization, predictions, and unsourced claims.
Civic Q&A Chat
Claude Sonnet 4.6 answers voter questions using only retrieved context. If no relevant chunks are found, it responds with a refusal template rather than speculating. The system prompt is version-controlled and logged alongside every response.
Every AI-generated response logs the model version, prompt version, token counts, and cost. These metrics are monitored by our team for quality assurance.
Human Review Process
No AI-generated content is published without human review. Our review pipeline has three gates:
Document verification
Ingested source documents are reviewed by an admin for authenticity, recency, and relevance. Documents are marked as verified, rejected, or pending. Only verified chunks feed into AI generation.
Explainer review
AI-generated ballot measure explainers are saved as drafts. An admin reviews every field: verbatim quotes are checked against source documents; AI bullets are verified for accuracy and citation integrity. Pages do not publish until the admin sets the human_reviewed_at timestamp.
Ongoing monitoring
Chat Q&A interactions are logged to a monitoring table. The team reviews a sample of responses regularly for accuracy, appropriate refusals, and citation quality.
Flag Workflow
Every page on VoterHQ includes a “Submit feedback” button. Visitors can report issues in five categories:
Flags enter an admin queue where a reviewer sees the flag reason, the page path, the reporter’s description, and the original content. Each flag is resolved with one of three outcomes:
Correction made — the content is updated and the correction is auto-appended to our public corrections log.
No change needed — the flag is closed with a public explanation of why the content was determined to be accurate.
Content removed pending review — the content is taken down while a more thorough review is conducted.
Correction Policy
Accuracy is our highest priority. When we identify an error — whether from an internal review, a user flag, or a source update — we follow this process:
Acknowledge
The error is documented in our admin system immediately upon discovery.
Correct
The affected content is updated or removed. For ballot measure pages, the corrected explainer goes through a fresh human review before republishing.
Disclose
Every correction is logged to our public corrections log with a date, description of the error, and what was changed. This log is append-only and never modified after publication.
View our corrections log
All corrections are published transparently.