task_illustrations__park_playground__person_count¶
Summary¶
- Domain:
illustrations - Scene id:
park_playground - Implementation source:
src/trace_tasks/tasks/illustrations/park_playground/person_count.py
Task Contract¶
Count every visible person in one rendered park/playground illustration.
Program Contract¶
Program: count(visible_people); scene=park_playground; scope=person_count
Candidate set: the visible illustrated scene objects, people, regions, tiles, patches, labels, and option panels inside the person_count objective scope.
Operands: visible scene state and prompt-bound operands named by visible_people, park_playground, person_count.
Operation: evaluate count 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 integer schema; generation binds a unique final answer.
Annotation witnesses: annotation uses the bbox_set schema; the prompt/annotation contract defines the minimal visual witnesses.
Query ids: single.
Reasoning Operations¶
Families: counting
Answer Contract¶
- Answer schema:
integer_count - Generator
answer_gt.type:integer - The answer is the number of rendered people in the scene.
Annotation Contract¶
- Annotation schema:
bbox_set - Generator
annotation_gt.type:bbox_set - Annotation is an unordered set of final-image pixel points, one near the center of each visible person.
- Annotation and answer are projected from the same generated scene trace.
Prompt And Trace Requirements¶
- Prompt text comes from
src/trace_tasks/resources/prompts/illustrations/park_playground/illustrations_park_playground_v0.json. - The prompt scene layer must not disclose the sampled person count.
- Render randomness, sampled style, activity mix, person count, and verifier payloads are recorded in trace metadata.