convergence-loop

Multi-pass claim verification uses agent-based discovery and T20 tallies to iteratively validate project claims. It runs modular behaviors like source-check, discovery, and verification to surface both verified and unresolved claims for audits.

Convergence loop is an agent-based workflow that iteratively validates cross-document claims through graduated tallies. It applies modular passes like fresh-eyes and write-then-verify to reduce wasted effort on incorrect claims.

Run the standard sequence of source-check, verification, and fresh-eyes on your target claims set. The workflow maintains a state.json file recording per-pass tallies and tracking graduated or unconverged claims for review.

Multi-pass verification requires a SKILL.md frontmatter with name and description, plus a REFERENCE.md for behavior definitions. Optional resources directories can be included, and the system maintains a state.json for recording tallies.

Agent-based discovery is suited for audits and discovery phases where iterative validation reduces wasted effort. Use it when cross-document consistency checks or plan verification require stateful governance and graduated convergence.

Unconverged claims occur when the T20 tallies and modular behaviors cannot fully validate a claim across passes. The state.json records these unresolved items alongside graduated claims to surface them for manual review.