The intersection of political journalism and agentic AI just got a serious upgrade with the release of True Oath, a political accountability ledger built by developer Ujjavala. Submitted for the Sanity Challenge's Path One track, the project demonstrates how structured content management systems can power agents that do more than just scrape headlines. By leveraging Sanity's schema to model complex relationships between promises, evidence, and outcomes, True Oath moves beyond keyword search into the realm of verifiable, source-grounded audit trails. The current iteration focuses on Australian federal politics, but the architecture is designed to scale to other jurisdictions where public records are accessible.

Structuring Truth in a Post-Truth Era

Most political fact-checking tools fail because they treat claims as binary true/false statements. True Oath rejects this simplicity, modeling accountability as a network of linked records. The Sanity Studio schema separates manifestos, government records, specific promises, and evidence findings into distinct document types. This allows an agent to distinguish between a policy announcement, a measurable outcome change, and an independent assessor's verdict. The system intentionally preserves uncertainty, marking items as 'Unverifiable' or 'In Progress' rather than forcing a definitive status when the evidence trail is incomplete or contradictory.

The MCP Endpoint and Indexing Hurdles

Technically, the project connects to the Sanity Context MCP endpoint to retrieve relevant promises and follow source references. However, the developer encountered a practical constraint: the organization's beta index quota for the Knowledge Base (ID: kbgnQdlEqXlP) was exhausted by a previous project, Cyber Autopsy. Consequently, the current live agent relies on GROQ queries against the public production dataset rather than a fully indexed vector search. This highlights a common friction point in early-stage agentic development: infrastructure quotas often lag behind architectural ambition, forcing developers to pivot to live-dataset filtering modes.

Key Takeaways

  • True Oath models political promises as linked records with confidence scores and source URLs, not just text snippets.
  • The project uses a Next.js frontend deployed to Vercel and a Sanity Studio backend with a public-read production dataset.
  • Current limitations include exhausted Knowledge Base indexing quotas, requiring the agent to use GROQ filters instead of vector embeddings.
  • The system includes integrity events for corruption and conflict-of-interest findings, separating them from policy delivery status.

The Bottom Line

This is a proof-of-concept for how agents should handle nuance. By refusing to hallucinate certainty where evidence is missing, True Oath offers a blueprint for trustworthy political AI that prioritizes provenance over punchy verdicts.