Subtext Discovery Benchmark: 100 Intent-Stratified Discovery Queries and Graded Relevance Judgments for Semantic IR

An open benchmark for evaluating semantic discovery and embedding collapse in information retrieval. Contains 100 organic queries stratified across 8 fine-grained intent classes (affective, sensory, negation, etc.) and 4,370 pooled candidate relevance judgments on a 4-point graded scale (0–3), calibrated against an expert human annotator subset. Includes standalone evaluation scripts for nDCG@10, Recall@100, and statistical significance testing.