← ClaudeAtlas

blind-judge-gatelisted

Gate AI-made work (images, pages, videos, copy) behind independent blind reviewers scored against a real benchmark, then aggregate votes into PASS, NEAR, FAIL or NO_JUDGE. Use when the author cannot judge its own taste and you need a verdict that cannot be self-approved.
Tommyapi/ai-employee-skills · ★ 0 · AI & Automation · score 70
Install: claude install-skill Tommyapi/ai-employee-skills
# Blind judge gate The agent that made the work is the worst judge of it. This gate turns review into a procedure the author cannot game: 1. Pick a real benchmark (a top-performing piece from the wild). The deliverable is defined as "aligned with the benchmark", not "I think it looks good". 2. Start 2 fresh reviewer sessions that did not take part in the work and cannot see each other. Give each the benchmark and the candidate side by side plus `references/judge-prompt.md`. Each returns one JSON vote. 3. Put the votes in a file and run `python scripts/judge_gate.py votes.json --key hook,clarity`. ## Verdicts | Verdict | Rule (per vote) | Exit | |---|---|---| | PASS | no dimension 2+ below the benchmark, key dimensions at or above it, no hard flaw | 0 | | NEAR | every gap <= 2, key dimensions >= 7 absolute, no hard flaw. Ship only if labelled "same tier, not at benchmark", and log the gap | 10 | | FAIL | anything else | 1 | | NO_JUDGE | fewer than 2 valid votes, or 2 votes disagree with no third | 3 | Aggregation: two agreeing votes decide; if they disagree, a third vote is required and the majority wins (a three-way split takes the median). Every vote counts. Dropping the unfavourable one is not allowed. NO_JUDGE is never a pass. ## Why NEAR exists Against the best work on the internet, "any key dimension below the benchmark fails" blocks everything and publishing stops. NEAR keeps shipping honest: same tier, labelled, gap recorded for the next iteration. ## Rules - The