Assessment Extraction

Assessment Document Extractor Spec

Goal Convert assessment documents into JSON where each item bundles: • Student prompts • Deliverables • Evidence • Teacher scoring (rubrics, checklists, feedback) • Associated visual elements

⸻

Rules • Output: strict JSON only (no markdown, no preface). • Completeness: do not invent content. If missing, omit and record in errors[]. • Atomicity: each atomic prompt = one item. • IDs: label as "1A", "1B", etc.

⸻

Fields per Item • prompt: concise, actionable student-facing instruction (no teacher voice). • constraints: word limits, format, or other explicit requirements (1–2 sentences). • deliverables: required outputs (e.g., “1–2 sentence explanation”, “Annotated diagram”, “Complete the response grid”). • evidence: only explicitly referenced materials/figures/tables. • images[]: • id • uri (https) • caption • alt_text • metadata (source_reference, page_hint, position, dimensions, file ) • tables[]: • role: "display" or "response_grid" • headers: [string] • rows: [[string]] • column_types: [text|choice|checkbox] • grid_type: radio|checkbox (if response_grid) • summary: 1-line accessibility description • metadata ( source_reference, page_hint, position ) • tags: skills/topics if present. • rubric: structured scoring: • criterion • order • weight (default 1) • levels[] ordered ascending: ( name, descriptor, points, order ) • checklist[]: per criterion: ( text, required ). • feedback: reusable teacher comments.0 • misconceptions: common student errors.

⸻

Global (applies across items) • overall: capture any rubric/constraints that apply across all items. • confidence: single score in [0,1]. • errors[]: list ambiguities, conflicts, or unreadable sections.

⸻

Figure Association Guidelines • Associate figures/tables by: 1. Proximity in the document 2. Explicit prompt references 3. Contextual relevance • If multiple items possible → include with the most relevant item only. • If no clear association → omit.

{student_instructions}

{answer_key}