Converting and publishing¶

Any program can write task cards: it fills the 15 required slots of ASBTaskCard, keeps undeclared data in a sidecar file and validates the result. This page maps existing formats onto card slots and describes how a skill collection is published.

Principles¶

Fill the 15 required slots (Authoring a task card).
Fill the optional slots the source supports, in particular evidence_snippets, crossref_doi, task_kind and provenance_source.
Record the software stack in execution_environment; keep data the card does not declare (the environment a benchmark gives a solver, oracle, license, authors) in a sidecar file next to task_card.json.
Use the doi-slug__task_001 form of task_id when cards from several papers are pooled.
Keep finding, the source's answer, out of what an agent under evaluation sees.
Record gaps in uncertainty_notes and missing_information.

From benchmark rows¶

Several science benchmarks derive their tasks from published research and publish task rows that fill part of a card.

Benchmark	Fields that fill card slots	Authored or re-read
ScienceAgentBench¹	`instance_id` to `task_id`; `task_inst` to `task_description`; `domain`, `domain_knowledge`, `subtask_categories`, `github_name` to the slots of the same names; `output_fname` to `expected_artifact_name`	`crossref_doi`, `research_question`, `finding` from the paper
DiscoveryBench²	`domain`, `domain_knowledge`; `queries[].question` to `research_question`; `true_hypothesis` to `finding`; the gold workflow to `workflow_description`	The real split (264 tasks) derives from papers; the DOI from the paper
BixBench³	`id` to `task_id`; `question` and `hypothesis` to `research_question`; `ideal` to `finding`; `paper` to `crossref_doi` when it holds a DOI	A `finding` stated as the paper's result
CORE-Bench⁴	`capsule_id` to `task_id`; `capsule_title` to `title`; `field` to `domain`; `task_prompt` to `task_description`; `results` to `finding`	The article DOI (`capsule_doi` names the code capsule); one card per difficulty level

In each case evidence_snippets come from re-reading the source paper, and source_package, tools, skills and evaluation_strategy are authored. Benchmark licenses and the environments a benchmark gives a solver go to the sidecar file; the software stack a card ran with goes in execution_environment. Answers that a benchmark withholds stay withheld: a card holds finding beside task_description.

{
  "uncertainty_notes": [
    "finding taken from the benchmark answer key, not from the paper text"
  ],
  "missing_information": [
    "evidence_snippets not filled: source paper not re-read",
    "article DOI not checked against the capsule DOI"
  ]
}

feasible

From ASTRA documents¶

An ASTRA document records one analysis: a root Analysis with a description, inputs, outputs, decisions, prior insights, findings and nested analyses⁵. A reader can fill finding (from findings), expected_outputs (from outputs), run_command and crossref_doi; research_question, tools, skills, workflow_description and evaluation_strategy are authored, and a rule for task boundaries decides how one analysis splits into cards.

feasible

From claim records¶

A claim record binds a scientific assertion to its evidence. The claim adapter of the reference pipeline projects one claim onto one EvidenceSnippet: snippet_id from the claim identifier, source_path from the first evidence artifact or doi:<doi>, text from the source excerpt, and reason rendered from the claim. Check that source_path and text are not empty strings.

existing

Exporting cards¶

The exporters of the reference pipeline (ASTRA, ScienceAgentBench, PaperBench, workflow languages, ARA, Agent Skills) are described on Interoperability.

Checks¶

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

Closed validation reports undeclared keys; the open root of asb_card.schema.json accepts them (JSON Schema roots). Minimal conformance also asks for an RO-Crate 1.1 metadata file.

Publishing a collection¶

Collections are released through the public repository asb-skill-collections, following its contributing guide.

Propose. Open an issue with the "Propose a new collection" template: domain, a one- or two-sentence description, a slug, three to five seed papers by DOI, and a lead curator (GitHub handle and ORCID iD).
Stage. After agreement, open a pull request that adds staged-collections/<domain>/v1/collection.yaml.

Validate. The manifest validates closed against SkillCollection:

iri: https://w3id.org/holobiomicslab/asb-skill/collection/glycomics/v1
slug: glycomics
version: 1
title: Glycomics Skill Collection v1
description: Hypothetical example. Skills and tool records for computational glycomics.
license: CC-BY-4.0
released_at: '2026-10-30T00:00:00Z'

SCHEMA=$(python -c "from asb_schema import get_schema_path; print(get_schema_path('asb_skill_collection'))")
linkml-validate --schema "$SCHEMA" --target-class SkillCollection collection.yaml
# No issues found

Pass the gates. The repository's validation workflow checks the manifest, skill descriptions, derived_from DOIs, EDAM IRIs and capsule crates on every pull request.
Merge and promote. A maintainer merges; a later pull request moves the directory from staged-collections/ to collections/, where every gate blocks.

Reviewers are vetted through a candidacy pull request (candidates/<github-handle>.yaml) that links a GitHub handle, an ORCID iD and publication DOIs. Community skills enter through a proposal rail, at collections/<slug>/v<N>/proposals/skills/<skill-slug>/SKILL.md with provenance_tier: community and status: hold; provenance_tier takes literature, repository, synthetic or community.

The collection IRI https://w3id.org/holobiomicslab/asb-skill/collection/<slug>/v<N> resolves to the promoted directory. Collection content is licensed under CC-BY-4.0 and code under Apache-2.0; reuse credits the collection through its CITATION.cff and the source paper of each skill. Each release is deposited on Zenodo, with the DOI recorded in doi and zenodo_doi.