skills/proofread-math/evals/evals.json

{
  "skill_name": "proofread-math",
  "evals": [
    {
      "id": 1,
      "prompt": "Proofread this pasted LaTeX. Correct the prose but do not guess the sign: Let $f\\colon X\\to Y$ be a continuous maps. We have $d(f(x),f(y))=-d(x,y)$.",
      "expected_output": "Corrects the grammatical error, preserves the sign, and reports the sign as unresolved.",
      "assertions": [
        "Corrects 'a continuous maps'",
        "Does not silently change the minus sign",
        "Reports the possible sign issue outside the corrected source"
      ]
    },
    {
      "id": 2,
      "prompt": "Review only: For every $i\\in I$, choose $x_i\\in X_i$. Then $x_j$ is nonzero. No $j$ is defined.",
      "expected_output": "Reports the unbound symbol without editing or asserting that i is certainly intended.",
      "assertions": [
        "Produces findings rather than rewritten source",
        "Identifies the unbound j",
        "Does not guess a correction with certainty"
      ]
    },
    {
      "id": 3,
      "prompt": "Self-review only the theorem hunk you just changed and adjacent definitions. No build command is documented.",
      "expected_output": "Limits coverage and reports validation not run rather than inventing a command.",
      "assertions": [
        "Limits coverage to changed hunk and necessary context",
        "Does not rewrite unrelated sections",
        "Reports validation not run with reason"
      ]
    },
    {
      "id": 4,
      "prompt": "Proofread this theorem and also confirm that the cited paper proves it over integers.",
      "expected_output": "Keeps syntax and prose proofreading separate, treats the substantive source claim as outside proofreading, and routes that claim through the available literature-check skill rather than attempting source verification under the proofreading contract.",
      "assertions": [
        "Proofreads only within the conservative editing boundary",
        "Distinguishes citation syntax from mathematical source verification",
        "Routes the integral source question to literature-check",
        "Does not silently alter the theorem based on an unchecked citation"
      ]
    }
  ]
}