Round 1 (4 tabs):
- 🔧 Unified eval — 500-qid subset from $DATA_ROOT/eval/wiki_opentable/unified/test_with_chunks.unified.jsonl
- 🧬 Eval Structures v0 — typed structure docs from test_with_structures.unified.jsonl
- 🛠 Trajectory — Baseline A (per-qid scaffolds, cell4_agentic_a), F1 78.95
- 🛠 Trajectory (Structure Corpus) — Baseline B (flat structure corpus, cell5_agentic_b_corpus), F1 74.84
Round 2 (2 more tabs):
- 🛠 Trajectory (Rawtext Corpus) — Baseline C (flat rawtext corpus, cell6_agentic_c_rawtext), F1 89.22
- 📊 Compare — per-qid side-by-side grid across all 6 cells with local set-based F1/EM
(no LLM judge for wiki_opentable; the deterministic wiki_opentable_adapter.py
is the source of truth — same scorer the trajectory tabs use)
Per-cell mean F1 / EM in the Compare tab:
closed-book F1 22.03 EM 21.00
with-docs F1 85.82 EM 84.80
with-structures F1 80.90 EM 79.80
agentic-a F1 78.95 EM 77.80
agentic-b-corpus F1 74.84 EM 74.00
…
