Skip to content

task_illustrations__environment__missing_patch_label

Summary

  • Domain: illustrations
  • Scene id: environment
  • Implementation scene: environment
  • Implementation source: src/trace_tasks/tasks/illustrations/environment/missing_patch_label.py

Task Contract

Renders an environment source panel with one missing visual region and four or six same-size lettered patch options. The model selects the option letter that exactly restores the missing region.

Program Contract

Program: select_option(match_patch(source_image, missing_region, options, transform=none)); scene=environment; scope=missing_patch_label

Candidate set: the visible illustrated scene objects, people, regions, tiles, patches, labels, and option panels inside the missing_patch_label objective scope. Operands: visible scene state and prompt-bound operands named by match_patch, source_image, missing_region, transform, none, environment, missing_patch_label. Operation: evaluate select_option over the candidate set using the visible illustrated objects, regions, layout relationships, counts, patch/tile transforms, or option-selection constraints encoded in the program expression; generation enforces a unique final answer. Output binding: answer uses the option_letter schema; one of the visible option letters. Annotation witnesses: annotation uses the bbox_map schema; the prompt/annotation contract defines the minimal visual witnesses. Query ids: single.

Reasoning Operations

Families: transformation, matching

Query Branches

Query id Program schema
single select_option(match_patch(source_image, missing_region, options, transform=none)); scene=environment; scope=missing_patch_label

Program Metadata

  • Program signatures: selection.option_match
  • Base program contract: select_option(match_patch(source_image, missing_region, options, transform=none)); scene=environment; scope=missing_patch_label
  • Parameter axes: theme_id, option_count
  • Arguments:
  • source_image: semantic_role; allowed environment_source_panel; source program_schema_concrete
  • missing_region: semantic_role; allowed masked_source_region; source program_schema_concrete
  • options: semantic_role; allowed lettered_patch_options; source program_schema_concrete
  • transform: operation_parameter; allowed none; source program_schema_concrete
  • Argument metadata status: curated
  • Supported query ids: single

Answer Contract

  • Answer schema: option_letter
  • Generator answer_gt.type: option_letter
  • The answer value is one of the visible option letters.

Annotation Contract

  • Annotation schema: bbox_map
  • Generator annotation_gt.type: bbox_map
  • Annotation keys are missing_region and selected_option.
  • Annotation boxes are final-image pixel boxes around the missing source region and the selected patch option. Do not include all options, labels, source objects, or context-only regions.

Prompt And Trace Requirements

  • Prompt text must come from the illustrations_environment_v1 prompt bundle, with scene/task/output layers selected deterministically and recorded in metadata.
  • Render randomness, sampled fonts/styles, option order, crop box, and verifier payloads must be explicit in the instance trace.
  • Source and distractor crops use the shared illustration missing-patch crop sampler: crops are sampled from the rendered source image, spatially separated from the missing region when possible, and filtered by the shared visual-detail and patch-difference checks.
  • Option count is sampled from 4 or 6; all option bboxes use the same pixel width and height as the missing region.
  • Missing-region size is sampled relative to the resolved source image: width 15%-30%, height 15%-26%, area at most 6.5%.
  • The selected option bbox, answer label, and missing-region bbox must all come from the same compose_patch_options execution trace.