Texts (input/text.json)

The text library is an array of entries. Each entry defines one piece of source material to be tested.

{
  "label": "my-text",
  "content": {
    "title": {
      "full": "Full Title of the Work",
      "short": "Short Title",
      "extra_fields": { "year": "2020" }
    },
    "section_headings": [
      "Introduction",
      "Methods",
      "Results",
      "Discussion"
    ],
    "aliases": ["Alt Title", "Abbreviation"]
  }
}

Field	Purpose
label	Unique identifier — appears in output files as text_label
content.title.full	Full title string, substituted into prompts via {title}
content.title.short	Short title for display
content.title.extra_fields	Arbitrary metadata (year, edition, etc.)
content.section_headings	List of expected section or chapter headings — used to score list task responses
content.aliases	Alternative names used for fuzzy title matching

section_headings is the ground truth for list tasks: the harness measures how many headings the model reproduces (coverage), how many are hallucinated, and whether the first item is correct.