{"reader":"lexical span reader, prior_weight=1.0","measured":{"retrieval_ndcg@5":0.7052,"retrieval_hit@5":0.9008,"reader_f1_no_prior":0.1835,"reader_f1_with_prior":0.2146,"reader_f1_oracle_rerank":0.3023,"extraction_ceiling_f1":0.7147,"answer_extractable_verbatim_pct":11.1},"abstention":"NOT enabled. Conformal calibration gave tau=0.9496, which accepts 1 of 28,000 queries — no threshold reaches alpha=0.10 at any usable coverage.","known_limits":["Answers average 19.2 words and are human-written sentences; only 11.1% exist verbatim in the passage, so an extractive reader is capped at F1 0.715 before any implementation error.","0.088 F1 of reranking headroom remains unclaimed."]}