Can the claim survive a second opinion?

My role

I designed and built the plugin, data model, review orchestration, document checks and guarded workflows.

Where it stands

Working personal tool with Claude Code and Codex adapters. The source repository is private.

What changed.

Eight review perspectives examine the rendered document; unsupported claims and failed checks block completion.

The problem

A fluent sentence can make a weak claim sound convincing. I wanted a system that keeps the original evidence close enough to challenge the wording, including the numbers, scope and ownership of the work.

Separate writing from judging

The workflow separates the writer from the verifier model family. Reviewers inspect a fixed artifact and report findings instead of silently editing the document they judge. The system checks whether the expected reviewers actually returned.

The artifact is part of the answer

A good source document can still become a broken PDF. I added checks against the rendered output, including parsing, layout and exact-artifact receipts. Reusing an old verdict after changing the page is not an acceptable shortcut.

Facts outlive a draft

Raw portfolio facts and their provenance live outside the plugin repository. A guarded correction path updates evidence without discarding unrelated facts. Published examples use synthetic identities, and sensitive actions remain separate from drafting.

The stack

• Python
• Multi-agent orchestration
• Claude Code
• Codex
• LaTeX
• pytest