Proofdesk, a new editorial review queue for AI-assisted drafts, is making waves in the dev community by refusing to approve content that lacks verifiable evidence. Built for the Sanity Challenge's "Vibe-Code Something Strange" track, the tool enforces a strict rule: if a claim has no source, it stays in review. This approach tackles the growing problem of plausible but unsupported AI-generated text slipping into production workflows.

Enforcing Evidence with Sanity Revision Checks

The core of Proofdesk is its integration with Sanity Content Lake, specifically leveraging revision checks to prevent race conditions. When a reviewer attempts to approve a draft, the server authenticates the request, fetches the latest document, and verifies the revision ID using Sanity's ifRevisionID patch mechanism. If another editor modifies the document between the read and write operations, Sanity rejects the patch, forcing the reviewer to reload. This ensures that decisions are made against the exact version of the text the reviewer saw, a critical feature for collaborative environments where AI drafts might be updated mid-review.

Practical Implementation and Testing Rigor

Developers can run the local demo without an account or API key by executing npm ci, npm test, and npm run dev, then navigating to http://127.0.0.1:4321. The project includes a runnable source archive with Astro app routes, Sanity schema declarations, and fixture seeders. The author, who used Hermes Agent for code generation, highlighted that seven automated tests passed alongside HTTP checks and browser interactions. Notably, the build process included a dependency audit that reported zero vulnerabilities at the time of construction, demonstrating a commitment to security even in a prototype phase.

Limitations of Structural vs. Semantic Validation

Proofdesk checks the structure of evidence and records reviewer attestations, but it does not perform semantic fact-checking. It cannot verify if a quotation is accurate or if a source is trustworthy; it only ensures that a source ID is attached to a claim. The reviewer label is self-entered, and connected writes use a shared access token, making it suitable for controlled prototypes but not yet a production-grade identity system. For teams looking to integrate AI into editorial pipelines, this tool offers a practical framework for gating content, though it relies on human reviewers to validate the quality of the cited sources.

Key Takeaways

  • Proofdesk blocks approval of AI drafts lacking source evidence, enforcing a 'no citation, no publish' policy.
  • The tool uses Sanity's ifRevisionID to prevent race conditions and ensure reviewers act on current document versions.
  • Local development requires no API keys, with full setup available via npm run dev and a downloadable source archive.
  • The system validates evidence structure and reviewer attestation but does not perform semantic fact-checking of sources.

The Bottom Line

Proofdesk is a necessary reminder that AI speed shouldn't compromise editorial integrity; by gating drafts on evidence, it offers a pragmatic, developer-friendly way to keep hallucinations out of production.