nx_qrels_bench.nx

nx_qrels_bench.nx -- THE PER-DOCUMENT RELEVANCE RULER (Q7 runner, 2026-09-04). WHY IT EXISTS. The estate's incumbent web ruler judges by EXPECTED HOST, and that is not a detail -- it is why it could not see any defect found on 2026-09-04. "Icing (aeronautics)" at rank 1 for the query "ice cream" sits on en.wikipedia.org and therefore scored a PERFECT hit@1, while the stored document does not contain the word "cream" anywhere. A HOST IS NOT AN ANSWER. This runner scores against knowledge/search_qrels.conf, whose judgements are per URL. COMPOSED, NOT REIMPLEMENTED, ON BOTH SIDES: retrieval -- dss_search over the SAME code path /search serves, so the number describes the live engine rather than a reimplementation of it. Deliberately NO HTTP: an organ that fetches its own daemon measures the transport too, and would queue behind the serve it is grading. metrics -- nx_ir_eval (ie_ndcg_at_k, ie_mrr, ie_hits_at_k, ie_precision_at_k, ie_ap), the estate's Q16.16 integer-exact floor. A second metric implementation is the duplicate-ruler defect. SCOPE, INHERITED FROM THE QRELS FILE AND NOT WEAKENED HERE: that set is a BOOTSTRAP -- its candidates were pooled from the engine's own required-term arm, so it measures whether known-good documents rank UP and it CANNOT prove recall over documents that arm never retrieved. The recall figure is therefore recall over the JUDGED POOL, and it is emitted as judged_recall_permil so nobody reads it as the other thing. GRADE 0 IS A JUDGEMENT, NOT A GAP. A result absent from the file scores 0 deliberately: the file records every document examined per query, so "unlisted" means "examined and judged irrelevant" -- which is exactly how a stem-variant-only match (icing for ice, warring for wars) is meant to score.