Skip to main content
BATTLEGROUND INTEL

Last updated 2026-07-01

AI disclaimer

Battleground Intel uses large language models (LLMs) to augment its analysis of candidates, mission statements, and political alignment. This page explains in plain language what the AI does, what it does not do, and how the data you share is used.

1. What AI powers Battleground Intel

  • Anthropic Claude Sonnet 4.6— powers candidate auto-scoring across 33 policy issues and three compass axes (economic, social, governance), mission-statement parsing (converting an organization’s mission text into a weighted issue matrix), and the scheduled research agent’s monthly and quarterly candidate re-scoring.
  • Anthropic Claude Opus 4.8— powers the highest-judgment analytic passes: web-grounded verification of a candidate’s current campaign status, and the strategic synthesis section of generated battleground reports.

2. What the AI does

  • Structured scoring. Reads public candidate data (voting record, bill sponsorships, campaign-finance filings, press statements) and scores candidates on every issue on a calibrated 8-point political-direction scale.
  • Mission parsing.Extracts weight-tier priorities and directional preferences from your PAC’s mission statement.
  • Evidence gathering. Each score is accompanied by citations using our URIS (Unified Research and Intelligence Standard) to every source used.

3. What the AI does NOT do

  • The AI does not make endorsement decisions. Every recommendation surfaced in the dashboard is a starting point for your team’s own review, not a final verdict.
  • The AI does not replace human judgment. Political strategy, legal implications, and ethical trade-offs are outside its scope.
  • The AI does not provide legal, financial, or strategic advice. Outputs are analytic tools, not professional counsel.
  • The AI does not speak for candidates. Positions attributed to candidates are our interpretation of public records, not direct statements.

4. How your data is used

When you submit a mission statement or our system scores a candidate, the following data is sent to Anthropic’s API:

  • Your mission statement text (mission parsing).
  • Publicly scraped candidate content from sources including FEC, Congress.gov, Ballotpedia, Wikipedia, VoteSmart, and interest-group scorecards (candidate auto-scoring).
  • A structured prompt we’ve authored and maintain.

We do not send Anthropic: your billing data, your login credentials, the names or emails of your team members, or any data marked as private in your dashboard. See Anthropic’s privacy policy for how they handle data sent via their API.

5. Bias, limitations, and human oversight

LLMs have known limitations. We mitigate the ones most relevant to political analysis:

  • English-only. Our prompts and training context assume English-language source material.
  • Dataset recency.Candidate scoring re-runs on a monthly and quarterly cadence. A candidate’s listed score may reflect data that is up to 90 days old.
  • Political polarity on contested issues. Where sources disagree strongly (e.g., abortion, gun regulation), we emphasize axis-aware semantics — the same word can mean different things on the economic, social, and governance axes.
  • Prompt-injection resistance. Mission text and scraped content are wrapped in explicit trust fences; canary tokens detect attempted prompt leaks; flagged outputs trigger human review.

If you spot a mis-scoring, biased interpretation, or evidence mismatch, tell us via the in-product feedback form. Every report is reviewed.

6. Versioning

  • Taxonomy version: v3.1 (33 issues, 7 weight tiers)
  • Scoring model: Anthropic Claude Sonnet 4.6
  • Mission-parser model: Anthropic Claude Sonnet 4.6
  • Verification + report-synthesis model: Anthropic Claude Opus 4.8
  • Last updated: 2026-07-01

A note on keeping this page honest.Every time we change a model (for example, promoting a newer Claude Sonnet into production), adjust the issue taxonomy (adding, removing, or re-weighting issues), modify data-handling practices (new sources, new retention posture, new API terms), or revise the AI’s role in any section above, the version numbers and the “Last updated” date in this list MUST be updated in the same change set. Stale version metadata undermines the transparency this page exists to provide — treat it as part of the feature, not as documentation that can lag behind code.

7. Contact + corrections

Questions, corrections, or general feedback:

Back to top