<task_id> Task Doc Template¶
1) Identity¶
- Domain:
- Scene id:
- Task id:
- Objective contract:
2) Scene + task contract¶
- Entities/relations:
- Supported
query_idvalues: answer_gt.type:- Answer precision/format, if narrower than the registered answer type:
- Default
annotation_gt.type: - Alternate annotation forms:
- Annotation witness policy:
- annotation contract family (
direct_visible_answer,derived_visual_value, ordiagram_primitives): - minimal visual answer-verification witnesses:
- derivation/proof details kept in trace metadata:
- annotation shape choice (
point,bbox,segment,point_set,bbox_set,segment_set,point_map,bbox_map, etc.): Usesegmentfor exactly one visual line/edge/path witness andsegment_setfor multiple unordered line/edge/path witnesses. - map annotation role names, if used:
- numeric/readout annotation handling:
- 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):
- Overlap/touch policy (if applicable):
3) Prompt contract¶
prompt_bundle_id:scene_key:task_key:- Optional query-id prompt mapping (
query_keyor slot-driven mapping): - Required slots:
- answer output:
json_output_contract_answer_only,answer_hint,json_example_answer_only - answer+annotation mode:
json_output_contract,annotation_hint,answer_hint,json_example - 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).
- Variant counts (scene/task/query-id/mode):
- Output modes:
answer_onlyanswer_and_annotation
4) Determinism + constraints¶
- Seed namespaces used:
- Unique-answer policy:
- Reject/resample conditions:
- No-auto-relaxation guarantee:
5) Program and reasoning metadata¶
- Concrete
## Program Contractexpression: ## Reasoning Operationsfamilies, in the canonical order fromdocs/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¶
- Determinism test:
- Answer/annotation consistency test:
- Prompt metadata/placeholder test:
- Constraint-specific tests:
Include only details that are part of the durable public task contract.