The problem Verity Desk solves

Misinformation spreads faster than traditional fact-checking. People often encounter viral articles, social-media posts and images without knowing whether the claims are accurate, outdated or presented without context. Manually checking multiple sources is slow, while relying on a single AI model can produce confident but unsupported answers.
Verity Desk makes verification faster, safer and more transparent. Users can submit a URL, text claim or image. The platform then:

• Extracts the main factual claims.
• Searches for supporting and contradicting evidence.
• Evaluates source quality and independence.
• Cross-checks the evidence using multiple Gonka-hosted AI models.
• Returns Unverified instead of guessing when evidence is insufficient.
• Records every source, model response and Gonka request ID in a transparency ledger.
  Each AI model has a separate responsibility:
• MiniMax M2.7 extracts checkable claims, plans evidence searches and identifies research gaps. It also performs one independent evidence review.
• DeepSeek V4 Flash acts as another independent verifier, checking the evidence for support, contradictions, incorrect dates and missing context.
• Kimi K2.6 performs the final decision review. It compares both verifier outputs against the evidence and source-credibility assessment before producing its recommendation.
• Deterministic safety rules then calculate the final result and can override model confidence by returning Unverified when reliable evidence is insufficient.
  Citizens can check viral stories before sharing them. Governments can monitor public misinformation. Traders can investigate market-moving news, while researchers can review claims and their supporting sources more efficiently.
  Unlike a simple AI chatbot, Verity Desk shows users how a conclusion was reached. This reduces blind trust in AI and helps people make evidence-based decisions.

Challenges we ran into

One major challenge was keeping multiple AI models consistent. Different Gonka models sometimes returned different verdicts, invalid JSON or explanations in different languages. We addressed this with strict response schemas, English-language validation, automatic retries and deterministic consensus rules. When evidence is insufficient, the system returns Unverified rather than forcing a conclusion.
We also encountered changing model availability. For example, Kimi appeared in the Gonka model catalogue but initially rejected inference requests. We created smoke tests to verify each configured model and designed a safe fallback that records failed calls without allowing them to influence the result. Kimi later became available and was successfully integrated as the final decision reviewer.
Web evidence introduced another challenge. Search results can be duplicated, outdated, inaccessible or copied across multiple websites. We added URL deduplication, source-quality scoring, publication-date checks, official-source identification and syndication-risk detection.
Multi-claim articles were also difficult because evidence from one claim must never affect another. We solved this by creating separate reports, evidence IDs, model calls and verdicts for every extracted claim.
Finally, creating meaningful transparency required more than displaying a final score. We built a SQLite audit ledger that stores model IDs, public reasoning summaries, evidence sources, latency, failures and Gonka request IDs—without exposing API keys or private chain-of-thought.