Dataset/nested_squares_dataset_3000/README.md

GRIP Nested Squares Dataset 3000 — v8

Version nested-squares-8.0.0 contains 3,000 images and 15,000 questions. V8 occupies the unsuffixed release paths; superseded local builds were removed to keep the repository clean.

Phase 8 changes

Level 5 now uses a 12% outermost-side threshold and a rescaled 9.6%–14.4% exclusion band. V7 medians (6.78% squares, 7.85% triangles, 6.92% hexagons) showed that lowering 5% would increase the already dominant yes class. Full guard-intersection calibration found 12% to be the lowest shared rounded threshold whose lower margin remains feasible for 10–12 triangles while retaining both visibility floors. The v8 result is exactly 1,500 yes / 1,500 no, with shape-count correlation 0.000000 and an achievable extrapolated-fraction range of 9.30%–16.99%.

The 15 px innermost-side guard and 3 px adjacent clear-background guard are unchanged. Clearance targets are sampled by paired nuisance rank and enforced for both Level 2 labels; minimum-clearance Cramér's V is 0.010438. The definitional association whitelist contains exactly factor_progression_direction and reduction_factor_span; no nuisance feature reaches V=0.10.

All 3,000 PNG outline counts were independently recovered from disk. Every guard injection test passes, Level 4 retains its 0.194030 theoretical baseline and 0.194510 observed capture, and validation reports PASS — 0 mismatches. Full distributions and all 136 minimum-clearance rejections are recorded in validation_report.txt.

Supplementary open-ended questions

Version nested-squares-11.0.0 replaces the supplementary free-response set with the exact domain template in OPEN_QUESTION_SPEC.md. It includes 3000 eligible items and excludes 0 under {}. Every prompt uses visible evidence, asks for justification, and ends with a confidence score from 0 to 1.

• open_questions.csv is public and contains exactly question_id,image,prompt.
• open_answer_key.csv is answer-key-side and contains separate partial-credit fields, an exhaustive acceptance_set, deterministic targets, and machine-readable tolerances for every numeric field.
• open_annotations.jsonl is answer-key-side and adds the complete derivation and scoring declaration.
• open_validation_metrics.json contains full target and answer distributions, constant-answer baselines, prompt/schema checks, independent metadata derivation results, and exhaustive PNG recovery results.

Scored fields at or above 60% after template revision: {}. Targeted fields: shape_count, shrink_pattern, rotation_degrees_nearest_10. Do not provide open_answer_key.csv, open_annotations.jsonl, or the closed-set annotations.jsonl to a model under evaluation because they expose answer-side scene metadata.