[
  { "id": "C01", "kind": "correct", "label": "pass", "confidence": "high",
    "claim": "Jev 1.13 accepts up to 64,000 tokens per request, shared by the state and all questions.",
    "evidence": ["eval/sources/jev-models.md:8 — the per-request token limit"],
    "why": "Line 8 states exactly this." },
  { "id": "C02", "kind": "correct", "label": "pass", "confidence": "high",
    "claim": "The state plus the longest single question must fit within 32,000 tokens.",
    "evidence": ["eval/sources/jev-models.md:9 — state plus longest question limit"],
    "why": "Line 9 states exactly this." },
  { "id": "C03", "kind": "correct", "label": "pass", "confidence": "high",
    "claim": "Jev charges $0.042 per million input tokens and nothing for output tokens.",
    "evidence": ["eval/sources/jev-models.md:14-15 — pricing"],
    "why": "Lines 14 and 15 give both facts." },
  { "id": "C04", "kind": "correct", "label": "pass", "confidence": "high",
    "claim": "A Choice question can offer at most 255 options.",
    "evidence": ["eval/sources/jev-primitives.md:7 — Choice option limit"],
    "why": "Line 7 states the limit." },
  { "id": "C05", "kind": "correct", "label": "pass", "confidence": "high",
    "claim": "TypeSafe advises doing arithmetic in code rather than asking Jev to calculate.",
    "evidence": ["eval/sources/jev-limits.md:6 — not a calculator"],
    "why": "Line 6 says Jev is not a calculator and to do arithmetic in code." },
  { "id": "C06", "kind": "correct", "label": "pass", "confidence": "high",
    "claim": "GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens.",
    "evidence": ["eval/sources/openai-models.md:9 — Luna row"],
    "why": "The Luna row gives both prices." },
  { "id": "C07", "kind": "correct", "label": "pass", "confidence": "high",
    "claim": "Claude Haiku 4.5 costs $1 per million input tokens and $5 per million output tokens.",
    "evidence": ["eval/sources/anthropic-pricing.md:10 — Haiku 4.5 row"],
    "why": "The Haiku row gives both prices." },
  { "id": "C08", "kind": "correct", "label": "pass", "confidence": "high",
    "claim": "The old gate passes a finding when both of its yes-or-no answers are at least 0.5.",
    "evidence": ["eval/sources/old-gate.ts:84-86 — rejection rules"],
    "why": "It only adds a reason below 0.5, and ok means no reasons." },
  { "id": "C09", "kind": "correct", "label": "pass", "confidence": "high",
    "claim": "The old gate passes the claim and the evidence strings to client.systemOne, and reads none of the cited sources.",
    "evidence": ["eval/sources/old-gate.ts:10 — evidence is a list of strings", "eval/sources/old-gate.ts:70-87 — the whole gate function, which fetches nothing"],
    "why": "Line 10 types evidence as strings. The function passes them to client.systemOne and reads no file or page." },
  { "id": "C10", "kind": "correct", "label": "pass", "confidence": "medium",
    "claim": "Jev's input price is more than 20 times lower than Claude Haiku 4.5's.",
    "evidence": ["eval/sources/jev-models.md:14 — Jev input at $0.042 per million", "eval/sources/anthropic-pricing.md:10 — Haiku input at $1 per million"],
    "why": "1 / 0.042 is about 23.8. Needs arithmetic across two sources." },

  { "id": "O01", "kind": "overreach", "label": "fail", "confidence": "high",
    "claim": "Jev answers every request in about 100 milliseconds.",
    "evidence": ["eval/sources/jev-models.md:20 — requests complete in around 100 ms"],
    "why": "Most requests, not every. Line 19 gives a range up to 500 ms." },
  { "id": "O02", "kind": "overreach", "label": "fail", "confidence": "high",
    "claim": "Jev cannot give a wrong answer.",
    "evidence": ["eval/sources/jev-models.md:24 — Jev cannot return an invalid answer"],
    "why": "The source says no invalid or malformed answer, not no wrong answer." },
  { "id": "O03", "kind": "overreach", "label": "fail", "confidence": "high",
    "claim": "Every Claude model on the price list costs $2 per million input tokens.",
    "evidence": ["eval/sources/anthropic-pricing.md:9 — $2 per million input"],
    "why": "Only Sonnet 5 costs $2. The others differ." },
  { "id": "O04", "kind": "overreach", "label": "fail", "confidence": "medium",
    "claim": "The old gate stops unsupported findings from ever reaching the board.",
    "evidence": ["eval/sources/old-gate.ts:84-86 — the gate rejects unsupported findings"],
    "why": "The gate only returns ok false. What happens next is not in this code." },
  { "id": "O05", "kind": "overreach", "label": "fail", "confidence": "high",
    "claim": "GPT-6 Luna and GPT-6 Astra have the same context window and the same price.",
    "evidence": ["eval/sources/openai-models.md:7-9 — both have a 1.05M context and identical pricing"],
    "why": "Same context, but Astra costs 100 times more." },
  { "id": "O06", "kind": "overreach", "label": "fail", "confidence": "medium",
    "claim": "TypeSafe recommends doing all work involving numbers in code, including reading numbers from text.",
    "evidence": ["eval/sources/jev-limits.md:6 — do numeric work in code"],
    "why": "The source covers arithmetic only, not reading numbers from text." },

  { "id": "W01", "kind": "wrong-number", "label": "fail", "confidence": "high",
    "claim": "Jev costs $0.42 per million input tokens.",
    "evidence": ["eval/sources/jev-models.md:14 — input costs $0.42 per million"],
    "why": "The source says $0.042, ten times lower." },
  { "id": "W02", "kind": "wrong-number", "label": "fail", "confidence": "high",
    "claim": "Jev allows 32,000 tokens in total per request.",
    "evidence": ["eval/sources/jev-models.md:8-9 — 32,000 tokens per request"],
    "why": "The total is 64,000. 32,000 is for state plus the longest question." },
  { "id": "W03", "kind": "wrong-number", "label": "fail", "confidence": "high",
    "claim": "GPT-6 Sol costs $2 per million input tokens and $50 per million output tokens.",
    "evidence": ["eval/sources/openai-models.md:8 — Sol row"],
    "why": "Sol output is $10. $50 is Astra's output price." },
  { "id": "W04", "kind": "wrong-number", "label": "fail", "confidence": "high",
    "claim": "A Choice question can offer up to 256 options.",
    "evidence": ["eval/sources/jev-primitives.md:7 — up to 256 options"],
    "why": "The limit is 255." },
  { "id": "W05", "kind": "wrong-number", "label": "fail", "confidence": "medium",
    "claim": "Jev's input price is more than 50 times lower than Claude Haiku 4.5's.",
    "evidence": ["eval/sources/jev-models.md:14 — Jev input at $0.042 per million", "eval/sources/anthropic-pricing.md:10 — Haiku input at $1 per million"],
    "why": "1 / 0.042 is about 23.8, well under 50. Needs arithmetic." },

  { "id": "X01", "kind": "contradicted", "label": "fail", "confidence": "high",
    "claim": "Jev accepts images as well as text.",
    "evidence": ["eval/sources/jev-models.md:10 — supported input types"],
    "why": "Line 10 says images are not accepted." },
  { "id": "X02", "kind": "contradicted", "label": "fail", "confidence": "high",
    "claim": "Jev can summarize a document into a short paragraph.",
    "evidence": ["eval/sources/jev-limits.md:9 — what Jev can produce"],
    "why": "Line 9 says Jev cannot summarize or generate text." },
  { "id": "X03", "kind": "contradicted", "label": "fail", "confidence": "high",
    "claim": "All three GPT-6 models accept reasoning effort none.",
    "evidence": ["eval/sources/openai-models.md:12 — reasoning effort support"],
    "why": "Line 12 says Astra does not." },

  { "id": "U01", "kind": "unsupported", "label": "fail", "confidence": "medium",
    "claim": "Jev 1.13 was trained on about two trillion tokens.",
    "evidence": ["eval/sources/jev-models.md:7-8 — Jev 1.13 model details"],
    "why": "The source says nothing about training data." },
  { "id": "U02", "kind": "unsupported", "label": "fail", "confidence": "high",
    "claim": "Claude Opus 5.5 has a 1M token context window.",
    "evidence": ["eval/sources/anthropic-pricing.md:7,12 — Opus 5.5 row and context note"],
    "why": "The source gives a context window only for Haiku 4.5." },
  { "id": "U03", "kind": "unsupported", "label": "fail", "confidence": "high",
    "claim": "The old gate retries a rejected finding once before posting it.",
    "evidence": ["eval/sources/old-gate.ts:70-87 — gateFinding"],
    "why": "The retry lives in the researcher loop, not in this function." },

  { "id": "F01", "kind": "fake-citation", "label": "fail", "confidence": "high",
    "claim": "Jev supports streaming responses.",
    "evidence": ["eval/sources/jev-models.md:140 — streaming support"],
    "why": "The file has 24 lines. Line 140 does not exist." },
  { "id": "F02", "kind": "fake-citation", "label": "fail", "confidence": "high",
    "claim": "Jev offers a 50% discount for batch requests.",
    "evidence": ["eval/sources/jev-pricing.md:3 — batch discount"],
    "why": "The file does not exist." },
  { "id": "F03", "kind": "fake-citation", "label": "fail", "confidence": "high",
    "claim": "GPT-6 Luna supports a 2M token context window.",
    "evidence": ["eval/sources/openai-models.md:30-34 — extended context option"],
    "why": "The file has 12 lines. The cited range does not exist." }
]