Task 240 — Scientific Plot Reasoning Visualization

Task contract. Keep each PlotQA chart unchanged in Input. Target begins with one clean second, then traces only the exact Prompt-relevant official legend, axis/category, and numeric operands, and finishes with one compact source-recomputed expression and answer.

Current lifecycle status. The original Gate B PR #506 merged, but merged-code Full preflight found incomplete all-split census and globally colliding split-local Sample IDs. Remediation PR #530 adds a sealed all-split census contract, explicit exclusions, and contiguous global ordinals. Owner and MJXWANG Evaluator PASS are bound to exact head 1da41748; trusted Workflow state is READY_TO_MERGE. Full generation, Final writes, Gate C, and promotion have not started.

Input → target preview

Representative Gate B evidence. Input is the unchanged source chart. Target uses the same geometry and timing, adding a restrained legend-to-operand trace, category highlight, numeric callout, guide line, and final expression. The full decision interface remains a separate immutable review page.

These six videos preserve accepted Gate B continuity examples. Open the current immutable eight-pair remediation Gate B review for the dual-PASS evidence and complete human checklist.

Identity and provenance

Measured Validation inspection

Paper, authors, and citation

Raw location, acquisition, and contamination

Native formats and annotation semantics

PlotQA v2 / <split>
├── png/<image_index>.png        → complete chart
├── annotations.json             → plot elements + structured values
└── qa_pairs_V2.json             → exact quantitative questions + answers

canonical pair = split / image_index / question_id

V2V conversion

Selection and split policy

• Publisher train, validation, and test splits are preserved; selection is grouped by plot ID so questions from one visual never cross splits.
• Gate B freezes one exact question for each of eight unique Validation plot IDs.
• The roster covers four dot-line and four horizontal-bar charts; difference, average, ratio, and range; same/cross-series traces; 2–8 operands; positive/negative answers; small decimals, million/billion scales, long labels, dense overlays, four-item legends, and formula-width risk.
• Only allowlisted numeric templates with uniquely resolvable annotation-backed operands are admitted. Ambiguous, missing, or non-finite values fail closed.
• Full intends at most one unit per published plot. The displayed 224,377 is a planned upper bound; Gate C requires an all-split eligibility census and explicit exclusions.

Quality gates

• The complete chart geometry remains unchanged between aligned Input and Target frames except for intended overlays.
• Every trace begins at the relevant official legend, reaches only exact Prompt-selected categories and operands, and uses source-derived numeric callouts.
• Same-series dense traces chain operand-to-operand after the first legend anchor; formula cards choose the least-overlapping corner and size to content.
• All 16 videos decode completely as H.264/yuv420p at 12 FPS and 96 frames, with no audio and a clean first second.
• All eight samples pass four-file, Metadata 1.2, source pairing, answer recomputation, unique fingerprint, semantic, media, S3 VersionId, and HTTP checks.
• Formal resume skipped all eight complete samples; independent regeneration reproduced all 32 payload objects byte-for-byte.

Pipeline → final four-file contract

official split downloads → freeze complete file + PNG inventory identity
→ stream annotations and QA v2 → pair split/image_index/question_id
→ allowlisted template + unique operand resolution → recompute answer or fail closed
→ unchanged eight-second Input → staged source-grounded reasoning Target
→ H.264/yuv420p pair → exact prompt.txt + Metadata 1.2 metadata.json
→ four-file, semantic, decode, reproducibility, S3, HTTP, and human checks
→ after dual Gate B PASS + merged authority: separately authorize Full and Gate C

Completeness and source verification