Keep only image-backed rows (drop prompt-only placeholders)

Upload full question bank: all designed items (image rows + prompt-only rows)

Add HC6 (invented-melody notation, retrieval-proof pass)

Add HC4 (solved OUR novel sudoku, 81/81) — overturns retrieval hypothesis

Add HC5 (full 42-note Twinkle, 5/5 pass)

Add HC3 (24 solar terms clock dial, 6/6 pass)

Fix dataset_info: add meta.difficulty + meta.checklist to match parquet

Add HC1/HC2 (retrieval finding); meta now includes checklist (judging criteria) + difficulty

Add H5,N3,N8,N10 + HB1,HB2,HB5,HB6,HB7 (all with eval)

Fix dataset_info: add meta.eval field to match parquet schema

Add H5, N3, N10, N8 (+eval); v1 now 10 items

Add H4, N7, N9 (+eval); v1 now 6 items

Add eval results (chatgpt-5.5-high) to meta

Add v1: first 3 items (G6, H2, H3)

initial commit