EqualifyEverything / benchmarks-ai-alt

Adversarial review of the build found two defects that would have mattered in a real run: - Round reports were sorted lexicographically, so with rounds 9, 10 and 11 on disk the two-quiet-rounds gate read rounds 11 and 9 as the latest pair. That is the check that decides when the loop stops. Now ordered numerically, with a self-test case covering it. - validate.mjs crashed with a Node stack trace on a missing flag value or a directory path. Both now report and exit 3 or 2. Directive fixes: - Directive 01 told the agent to mark items `needs-second-pass`, a status the schema does not have, so the record would have been rejected. Single-pass items now simply stay candidates. - Clarified that `leaky` is a field to set, not a status. - Added a Paths section to directives 01 and 02: the `../` links resolve on GitHub but are not write paths, and the agent works from the project root. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>