Lean body that assumes Claude's competence: tight tables for judgment sources and metric selection, runnable code blocks, and no padding explaining what Elasticsearch or NDCG is. The only mild inefficiency is the repeated "Per the [Elasticsearch Rank Eval API]" lead-in before several sections, but every section earns its place.

Provides fully executable guidance throughout — a complete `_rank_eval` POST body, a runnable Python test (`load_judgments` + `test_search_relevance_baseline`), a per-query regression test, and a concrete CSV judgment format — matching the copy-paste-ready score-3 anchor rather than pseudocode.

A clearly sequenced 5-step workflow (build judgments → define metrics → submit rank_eval → wrap as test → per-query regression) with explicit validation checkpoints in the `assert result["metric_score"] >= 0.80` and `assert delta >= -0.10` gates. The one gap is an explicit "fail → diagnose → fix → re-run" feedback loop, which is only implicit, but the explicit assertions keep it closest to the score-3 anchor.

Well-organized overview sections with a single, clearly signaled one-level-deep reference to references/rank-eval-guide.md (verified to exist), which appropriately offloads binary thresholds, snapshot corpora, and Quepid/Splainer tooling — matching the clear-overview-with-well-signaled-references anchor.

Lists multiple concrete actions — "Author Elasticsearch relevance regression tests", "judgment lists (query + expected docs at ranks)", per-query metrics enumerated by name, and "reproducible test corpora" — matching the score-3 anchor for specific concrete actions; far richer than naming only a domain and some actions.

Explicitly answers both what (authoring judgment lists, per-query metrics, reproducible corpora) and when ("Use before changing analyzers, synonyms, boosts, or query templates on an Elasticsearch index that serves user-facing search"), satisfying the score-3 anchor for explicit what-and-when triggers.

Natural terms a search engineer would say are well covered — "relevance regression tests", "analyzers, synonyms, boosts, or query templates", "NDCG / MRR baseline", and "Quepid + Splainer"; the explicit "Use before changing..." clause is exactly how the need arises in practice.

Pinned to Elasticsearch's `_rank_eval` endpoint with niche-specific triggers (judgment lists, NDCG/MRR, Quepid, Splainer), giving a clear niche unlikely to trigger for unrelated skills.