What problem does it solve? Research claims that fail robustness verification need repeated cycles of external review, targeted fixes, and re-verification; doing this manually loses track of budgets, claim states, and which fixes were already attempted. ## Core Features & Use Cases - Five-state claim routing: Consumes /auto-verify output (PASS / FAIL / INCONCLUSIVE / ZERO_ELIGIBLE_VARIANTS / INTEGRITY_ONLY) and routes each claim to the correct back-edge — variant-only fix, main-experiment fix, or claim-stage re-entry. - Bounded iteration budget: Enforces MAX_ITERATIONS = 6 and a MAX_CLAIM_REENTRIES = 2 sub-budget so the loop terminates instead of endlessly rewriting claims. - External reviewer LLM: Calls a configurable reviewer model via the llm-chat MCP server or environment variables, with persistent reviewer memory and resumable state in REVIEW_STATE.json. - Use Case: After running /auto-verify on mechanistic interpretability claims, trigger the loop with "auto review loop llm" to autonomously fix failing variants, re-run experiments, and produce a final per-claim iteration report. ## Quick Start Run the auto review loop llm on my project after /auto-verify has produced VERIFY_REPORT.md so the reviewer iterates on failing claims until the budget is exhausted.
