Task card¶

A task card is a structured record of one bounded, reproducible scientific task extracted from a paper or project. It states the research question, the answer the source reports, the inputs and expected outputs, the tools and skills involved, the workflow and a plan for evaluating an attempt. Its class is ASBTaskCard in src/asb_schema/asb_card.yaml; a card is one JSON document, conventionally task_card.json.

asb_card.yaml 0.3.0 https://w3id.org/holobiomicslab/asb-schema/card/v0.3 normative 15 of 64 slots required

Example¶

A hand-authored card with the 15 required slots and no other key. source_package is derived from the hypothetical DOI 10.1000/example.

Example

{
  "task_id": "task_001",
  "schema_version": "0.18.0",
  "scitask_version": "0.1.1",
  "title": "Reproduce Figure 2A: feature importance ranking",
  "source_package": "10.1000_example-paper",
  "created_at": "2026-10-30T12:00:00Z",
  "task_description": "Re-run the random forest feature importance analysis and reproduce Figure 2A.",
  "research_question": "Which molecular features best discriminate group A from group B?",
  "finding": "m/z 523.2 and retention time 4.3 min ranked highest (importance > 0.15).",
  "inputs": ["mzML files for 60 samples", "metadata CSV"],
  "expected_outputs": ["feature_importance.csv", "figure_2a.png"],
  "tools": ["matchms", "scikit-learn", "matplotlib"],
  "skills": ["feature importance ranking", "figure reproduction"],
  "workflow_description": "1. Load spectra with matchms. 2. Train a random forest classifier. 3. Extract and plot feature importances.",
  "evaluation_strategy": {
    "direct_checks": ["feature_importance.csv exists and contains ranked features"],
    "expert_review": ["Top-5 features match the reference within rank 1"]
  }
}

SCHEMA=$(python -c "from asb_schema import get_schema_path; print(get_schema_path('asb_card'))")
linkml-validate --schema "$SCHEMA" --target-class ASBTaskCard task_card.json
# No issues found

Structure¶

The 64 slots fall into fourteen groups. Required slots are in bold.

Group	Slots	Purpose
Identity and origin	`task_id`, `title`, `source_package`, `crossref_doi`, `github_name`, `related_repos`, `provenance_source`, `synthesized_from`, `created_at`	Name the task and its source
Scientific content	`research_question`, `finding`, `task_description`, `task_objective`, `domain_knowledge`	What is asked and what the source answered
Inputs and outputs	`inputs`, `expected_outputs`, `data_in`, `data_out`, `task_inputs`, `task_outputs`, `available_artifacts`, `data_accessions`, `expected_artifact_name`, `landmark_outputs`, `artifact_refs`	Data a solver starts from and files it produces
Tools and skills	`tools`, `skills`	Software and reusable capabilities the task requires
Workflow	`workflow_description`, `methodology_summary`, `parameters`, `script_refs`	Procedure, parameters and source code per step
Evaluation	`evaluation_strategy`, `reproducibility_tier`	How an attempt is checked and how far the step re-executes
Execution	`run_command`, `run_cwd`, `run_timeout_seconds`, `executable`, `execution_profile`, `execution_environment`, `resumption_contract`	A command that re-runs the task, its requirements and how a workflow resumes
Chaining	`input_from`	A prior task whose output is this card's first input
Classification	`task_kind`, `domain`, `primary_domain`, `subdomains`, `article_type`, `techniques`, `keywords`, `subtask_categories`	Relation to the source, domain and technique
Evidence and claim links	`evidence_snippets`, `finding_evidence_span`, `linked_result_ids`, `section_provenance`, `prov`	Verbatim passages, claim-ledger entries and provenance links
Entity spine	`spine_entity_ids`, `spine_condition_ids`, `abstraction_frame`	Paper-level entities and conditions the card refers to
Quality notes	`uncertainty_notes`, `missing_information`, `review_questions`	Gaps and questions for a reviewer
Assignment	`scenario_id`, `project_id`	Scenario or project of the card
Format versions	`schema_version`, `scitask_version`	Card-format versions the producer followed

The pair research_question and finding is the center of the card. The free-text lists inputs, expected_outputs, tools and skills describe the task in prose; available_artifacts, data_accessions and artifact_refs bind that prose to files and accessions. Every slot, with range and mappings, is in the field reference.

The finding is the answer

finding holds the source's answer to the research question. When a card is used to build an evaluation item, the agent receives the task text (task_description, inputs, workflow_description) and the evaluator keeps the full card.

Nested classes¶

Class	Slot	Content
`EvaluationStrategy`	`evaluation_strategy`	`direct_checks` (automated, deterministic checks) and `expert_review` (criteria for human judgment), two lists of strings
`EvidenceSnippet`	`evidence_snippets`	One verbatim passage: `snippet_id`, `source_path`, `reason` and `text`, all required
`AvailableArtifact`	`available_artifacts`	One file of the source package: `path` and `role` required, `indexable` and `sha256` optional
`JsonObject`	`parameters`, `data_accessions`, `artifact_refs`, `script_refs`	An open JSON object

asb_card.yaml also defines eight classes for execution and typed ports, used by optional slots: TaskPort (task_inputs, task_outputs), ProvenanceLinks (prov), ExecutionProfile with ToolAlternative (execution_profile), ResumptionContract (resumption_contract), ExecutableSpec (executable), ExecutionEnvironment (execution_environment) and TypedArtifact (data_in, data_out). They are listed in the field reference.

Enumerations used by the card¶

Enumeration	Slot	Values
`ProvenanceSourceEnum`	`provenance_source`	`published`, `preprint`, `draft_paper`, `local_project`
`TaskKindEnum`	`task_kind`	`reproduction` (a reported result), `analysis` (a re-analysis of the paper's data), `component_reconstruction` (a described but not separately validated mechanism), `extension` (beyond the reported scope)
`ReproducibilityTierEnum`	`reproducibility_tier`	`certified`, `type-checked`, `bypassed`, `manual-only`, `experimental`

Versions¶

A card records two strings: schema_version, the version of the card format, and scitask_version, the version of the task specification (the slot keeps the former name of the standard). Both are required strings without a pattern. The reference pipeline writes schema_version 0.18.0 and scitask_version 0.1.1. Card-format history:

2026-05-22: SciTask Standard 0.1.0, card schema_version 0.17.0.
2026-05-24: rename to ASB Schema; SciTaskCard becomes ASBTaskCard.
2026-06-09: card schema_version 0.18.0, with task_kind.
Release 0.4.0: parameters, data_accessions, artifact_refs and script_refs take lists of JsonObject; the fourteen optional slots that the reference pipeline writes beyond the former 50 are declared, so its cards validate closed.

Scope of the card¶

The card records the scientific task and its evidence; its optional execution_environment records the software stack that ran the card or that it requires. The environment a benchmark gives a solver, the oracle, hidden answers, canary strings, license and authorship belong to the formats a card feeds: the license and credit of released material sit on the skill collection, and benchmark formats add the rest (Benchmarks and agents). A producer keeps such data in a sidecar file next to task_card.json.