What problem does it solve? Single-reviewer AI checks of plans, implementation reports, and factual Markdown documents tend to confirm rather than challenge claims, and a single model's verdict is hard to trust. This Skill runs two fresh, mutually blind judge processes against a byte-frozen snapshot of the document and the repository, so every load-bearing claim is tested adversarially against reproducible ground truth instead of model opinion. ## Core Features & Use Cases - Dual blind judges: Judge A runs the latest Astra model resolved live from the Codex catalog with max reasoning effort; Judge B runs Claude via the moving fable alias. Neither sees the other's prompt, output, or report. - Two modes: challenge-only publishes both independent reviews preserving dissent; consolidate adds an anonymous evidence evaluator, a frozen revision contract, and a separate reviser whose changes are mechanically replayed and validated. - Fail-closed validation: Reports are checked for required sections, claim matrices, evidence status, secret leaks, and internal path leaks before publication; failed checks abort rather than degrade silently. - Use Case: Challenge a migration plan or an implementation report by having both judges verify claims like "implemented", "merged", or "tested" against fixed repository commits, with every verdict backed by path:line evidence. ## Quick Start Ask the agent to challenge a specific Markdown plan or implementation document with the dual-judge review, for example by requesting two independent reviews of the file placed next to it as Judge A and Judge B reports.
