Skip to content

<task_id> Task Doc Template

1) Identity

  1. Domain:
  2. Scene id:
  3. Task id:
  4. Objective contract:

2) Scene + task contract

  1. Entities/relations:
  2. Supported query_id values:
  3. answer_gt.type:
  4. Answer precision/format, if narrower than the registered answer type:
  5. Default annotation_gt.type:
  6. Alternate annotation forms:
  7. Annotation witness policy:
  8. annotation contract family (direct_visible_answer, derived_visual_value, or diagram_primitives):
  9. minimal visual answer-verification witnesses:
  10. derivation/proof details kept in trace metadata:
  11. annotation shape choice (point, bbox, segment, point_set, bbox_set, segment_set, point_map, bbox_map, etc.): Use segment for exactly one visual line/edge/path witness and segment_set for multiple unordered line/edge/path witnesses.
  12. map annotation role names, if used:
  13. numeric/readout annotation handling:
  14. answer-option annotation policy (only allowed for complete visual option-image/panel tasks with a source/reference/original image or region; otherwise ground source/candidate objects or primitives):
  15. Overlap/touch policy (if applicable):

3) Prompt contract

  1. prompt_bundle_id:
  2. scene_key:
  3. task_key:
  4. Optional query-id prompt mapping (query_key or slot-driven mapping):
  5. Required slots:
  6. answer output: json_output_contract_answer_only, answer_hint, json_example_answer_only
  7. answer+annotation mode: json_output_contract, annotation_hint, answer_hint, json_example
  8. JSON example validity rule: every documented prompt JSON example must be a valid response for the active task/variant/output mode (keys, value types, and annotation cardinality/semantics).
  9. Variant counts (scene/task/query-id/mode):
  10. Output modes:
  11. answer_only
  12. answer_and_annotation

4) Determinism + constraints

  1. Seed namespaces used:
  2. Unique-answer policy:
  3. Reject/resample conditions:
  4. No-auto-relaxation guarantee:

5) Program and reasoning metadata

  1. Concrete ## Program Contract expression:
  2. ## Reasoning Operations families, in the canonical order from docs/contracts/PROGRAM_SCHEMA_CATALOG.md:
## Reasoning Operations

Families: `ranking`, `aggregation`

Reasoning operations are exhaustive analysis metadata, not public taxonomy nodes. The public task class's literal reasoning_operations tuple is the source of truth; this section is its exact documentation mirror. Record every meaningful answer-determining operation, but do not count primitive visual access, output binding, annotation construction, or MCQ letter binding. direct_retrieval is an exclusive fallback and must appear alone.

6) Tests

  1. Determinism test:
  2. Answer/annotation consistency test:
  3. Prompt metadata/placeholder test:
  4. Constraint-specific tests:

Include only details that are part of the durable public task contract.