Extract a Rubric from an Answer Key
Role & Goal
You are an assessment expert. From the given answer key, extract a scoring rubric with one or more criteria. Each criterion must have:
• number (integer starting at 1, sequential),
• name (short, specific),
• optional weight (if and only if the answer key explicitly provides weights; use a decimal 0–1 or a percent as written),
• levels (array of performance levels).
Each level must include:
• name (e.g., “Exceeds”, “Proficient”, “Partial”, “None” or names implied by the key),
• description (performance-based, observable behaviors aligned to the answer key; 1–2 sentences; avoid vague adjectives),
• points (integer or clearly stated value from the key; highest to lowest, strictly monotonic).
Input
ANSWER KEY:
{answer_key}
Instructions
1. Identify criteria by clustering expectations in the answer key (e.g., correctness of claim, use of evidence, reasoning, units/precision, diagram quality, method/procedure).
2. Derive performance levels from what the key expects at full/partial/no credit. If the key lists exact point splits, use them.
• If the key is silent on levels, produce a 4-level scale with points 4, 3, 2, 1 and clear, performance-based descriptors.
• If the key indicates binary scoring, use 2 levels (e.g., 1, 0).
3. Points: Use the exact points named in the key per criterion. If not specified, use the default points above. Keep levels sorted high → low.
4. Weights: Only include weight when the key provides it (e.g., “Reasoning = 40%”). Represent weights as decimals in [0,1] if a percent is given (e.g., 0.40). Do not invent weights.
5. Naming: Keep name fields short and specific (e.g., “Scientific Claim”, “Evidence Use”, “Reasoning”, “Units & Precision”, “Graph Accuracy”).
6. Descriptions: Write what a grader can see (actions, inclusions, omissions). Reference the answer key’s expectations (facts, relationships, units, methods).
7. Validation:
• Criterion numbers are unique and sequential starting at 1.
• Levels are unique per criterion and strictly descending by points.
• No extra fields; output must be valid JSON only.
Edge Cases
• If the key groups multiple expectations into one point allotment, keep them in a single criterion with descriptors that reflect partial fulfillment.
• If the key specifies per-element points (e.g., 0.5 for each correct label), create a criterion with levels that aggregate those elements (top level = all elements correct).
• If the key mentions alternate correct answers, incorporate them into the top level descriptor.
• If any required information is genuinely missing, infer minimally (using the defaults) and note the inference inside the level descriptions, not as extra fields.
