Eval Accuracy

Use this skill to evaluate how factually accurate an assistant response is.

Inputs

Require:

• The assistant response text to evaluate.

Internal Rubric (1–5)

5 = Factually correct, no misleading claims, no hallucinations, claims are well-supported or appropriately qualified
4 = Mostly correct, minor imprecision that does not materially affect meaning
3 = Partially correct, contains one significant inaccuracy or unsupported claim
2 = Multiple inaccuracies or misleading statements
1 = Fundamentally incorrect, fabricated, or contradicts known facts

Workflow

1. Evaluate factual claims in the response.
2. Compare them against widely accepted knowledge.
3. Score accuracy on a 1-5 integer scale using the rubric only.
4. Write concise rationale tied directly to rubric criteria.
5. Produce actionable suggestions that improve factual correctness.

Output Contract

Return JSON only. Do not include markdown, backticks, prose, or extra keys.

Use exactly this schema:

{
"dimension": "accuracy",
"score": 1,
"rationale": "...",
"improvement_suggestions": [
"..."
]
}

Hard Rules

• dimension must always equal "accuracy".
• score must be an integer from 1 to 5.
• rationale must be concise (max 3 sentences).
• Do not include step-by-step reasoning.
• improvement_suggestions must be a non-empty array of concrete edits.
• Never output text outside the JSON object.