Lean prose paired with dense executable code; no basic-concept padding and every section advances a concrete evaluation step, so it lands at the 'every token earns its place' anchor rather than the 'could be tightened' level.

Provides complete, copy-paste-ready Python (dcg/ndcg/mrr/evaluate, BM25/kNN/RRF retrievers, pytest regression gate) with real Elasticsearch/OpenSearch API calls, plus engine-specific code in verified reference files.

Clear 7-step sequence from judgment set through regression gate, with an explicit CI validation checkpoint (Step 7) and a feedback loop on threshold failure; not capped at 2 because this is measurement work with validation present.

Core runnable spine stays in SKILL.md while engine-specific weighted-fusion and reranker code is split into two real, well-signaled one-level-deep reference files (judgment-sets.md, weighted-fusion-and-reranking.md), both verified present.

Lists multiple concrete actions ('authors ground-truth judgment sets, computes nDCG@k and MRR... measures the lift from Reciprocal Rank Fusion vs weighted fusion vs single-stage retrieval, and quantifies reranker... impact'), matching the multi-action anchor rather than the partial 'names domain and some actions' level.

Explicitly answers both 'what' (end-to-end hybrid retrieval evaluation actions) and 'when' via the 'Use when a production system combines lexical and semantic retrieval and you need a numeric relevance baseline, fusion-strategy comparison...' clause; third-person voice throughout.

Good coverage of natural specialist terms a search engineer would say (hybrid retrieval, BM25, vector, reranker, nDCG@k, MRR, RRF, cross-encoder, Cohere, bge, lexical and semantic retrieval), including common variations.

Clear narrow niche (fused hybrid-retrieval relevance evaluation with specific metrics) and distinct 'Use when' triggers unlikely to collide with sibling BM25-only or vector-recall skills.