Adds per-document outputs of pi_extract_hy4_vision_evidence (Tencent Hy4 in the pi coding agent, reasoning high, vision, page and box citations) for ExtractBench Pro at data revision 60b1017.
Outputs cover 75 of 100 documents: document_conventions 16 of 22, house_vocabulary 9 of 14, house_counting_rules 10 of 12, drawn_relations 17 of 20, charts_inside_documents 20 of 25, long_chart_sheets 3 of 7. The other 25 runs failed (timeouts, provider content-filter stops or request errors) and have no output; the evaluator scores them as 0. Records keep each document's wall-clock latency, so the 30-minute view scores outputs that took longer as 0.
Files: results/pi_extract_hy4_vision_evidence/<challenge>.jsonl and a new pipelines.jsonl row. No other pipeline changes.