nx_qabench_engine.nx

nx_drbench.nx -- R2: the FIRST public-benchmark run of the Nishi researcher (census axis [BEN]: "MEASURED scores on public benchmarks -- NEVER run"). Dataset = HotpotQA distractor validation rows 0..99 (banked by nx_drbench_fetch -> knowledge/fetched/drb_hotpot_p0.raw, sovereign TLS). Protocol = the published distractor eval: each question ships 10 paragraphs (2 gold + 8 distractors); systems retrieve + extract an answer; scored by SQuAD word-F1/EM (nx_qa_score_lib, integer permille) + gold-paragraph recall at 2. TWO MODES measured side by side: mode0 SINGLE-HOP lexical baseline: question-term overlap retrieval -> span heuristic extraction. mode1 MULTI-HOP (the DeepResearcher loop shape, mechanical): hop-1 best para/sentence -> harvest BRIDGE entities (cap-runs + hop-1 title terms not in the question) -> expanded second retrieval -> extract. The LOW numbers ARE the honest baseline (no LLM synth yet -- that is census rung R1). Teeth: T1 parse integrity (100 rows, 10 paras, nonempty q/a) T2 gold titles parsed + resolvable to paragraphs T3 NEG-CONTROL: predictions scored against the WRONG (rotated) golds must collapse to ~0 (metric not rigged) T4 non-degenerate retrieval floor (recall at 2 must beat the ~200 permille random floor) T5 report mode0 vs mode1 delta (measurement, not a forced win -- honest either way) NOTE: all state lives in ONE ctx table g: *i64 passed to every fn (the proven rctx pattern) -- statics hit the known multi-static subscript-read compiler bug (filed; PHASE-A/B bisect proof in the session log). ctx slots: 0 raw 1 blen 2 boff 3 cur 4 q 5 a 6 ty 7 gtb 8 gto 9 gtl 10 ngt 11 ptb 12 pto 13 ptl 14 np 15 sb 16 so 17 sl 18 sp 19 ns 20 nb 21 no 22 nl 23 tko 24 tkl 25 tks 26 tkc 27 ntk 28 qnb 29 qto 30 qtl 31 nqt 32 qk 33 tnb 34 tto 35 ttl 36 tts 37 ttc 38 nttk 39 shit 40 thit 41 bhit 42 psc 43 xsc 44 btb 45 bto 46 btl 47 nbt 48 cnd 49 cno 50 cnl 51 cnt 52 ncn 53 scr 54 p0 55 p1 56 ga expect_exit: 0 license_tier: ORIGINAL