Transparency & Trust
How Tank prevents hallucinated citations and ensures factual integrity
Last-read trust metrics
Verification Pipeline
Citation Resolver
Every [ref: N] tag is checked against configured resolvers (CrossRef, OpenAlex, PubMed, and any patent office that is actually configured). Green = resolved. Amber = grace. Red = fabricated. Unconfigured offices stay empty — this is not a live USPTO screen.
Content Gates
Before any content surfaces to the user, it passes through authenticity gates that check for known hallucination patterns and unverified claims.
Honest-State Banners
When a data source is degraded or unavailable, Tank surfaces an honest banner with estimated restore time — never silently serves stale data.
Disagreement Confidence
Stored debates are capped when disagreement exceeds threshold. The confidence metric is visible on this run, preventing artificial consensus.
Export Blocking
PDF exports and email digests are blocked if any citation resolves red or >5% resolve amber. No unverified content leaves the platform.
Why This Matters
Columbia University's May 2026 study found LLM-fabricated citations in published papers rose 12× in three years. Stanford HAI found ChatGPT-4 hallucinated 58% of case-law citations. Tank's architecture makes fabricated citations a blocking error, not a silent failure.