{
  "$comment": "Built on SIP — Starlight Model Arena run receipt. Schema: arena-run v0.2. Round 3 = full Anthropic lineup, model lane of the Proving Ground.",
  "date": "2026-06-10",
  "lane": "model",
  "feedsScorecard": "the Proving Ground model lane",
  "contestants": [
    "claude-fable-5",
    "claude-opus-4-8",
    "claude-sonnet-4-6",
    "claude-haiku-4-5-20251001"
  ],
  "tasks": [
    {
      "id": "coding-longest-palindrome",
      "category": "coding — leftmost-longest edge case, self-verifying asserts",
      "verification": "agent ran `python solution.py`; ALL PASS required; leftmost-tie asserts pin the gotcha",
      "results": {
        "fable": {
          "status": "PASS",
          "attempts": 1
        },
        "opus": {
          "status": "PASS",
          "attempts": 1
        },
        "sonnet": {
          "status": "PASS",
          "attempts": 1
        },
        "haiku": {
          "status": "PASS",
          "attempts": 1
        }
      },
      "winner": "tie (4-way) — saturated; all four used expand-around-center with strict-greater tie-keeping",
      "prompt": "coding — leftmost-longest edge case, self-verifying asserts"
    },
    {
      "id": "constraint-stack-json",
      "category": "7 simultaneous hard output constraints, script-verified",
      "verification": "python check: 4 keys, verdict enum, summary EXACTLY 16 words + contains 'lane' + excludes 'score', 3 kebab tags, exact date",
      "results": {
        "fable": {
          "status": "PASS 7/7"
        },
        "opus": {
          "status": "FAIL",
          "failed": [
            "missing date key",
            "15 words not 16",
            "'scorecard' contains 'score'"
          ]
        },
        "sonnet": {
          "status": "FAIL",
          "failed": [
            "12 words not 16"
          ]
        },
        "haiku": {
          "status": "FAIL",
          "failed": [
            "14 words not 16",
            "'scorecard' contains 'score'",
            "wrapped in markdown fences"
          ]
        }
      },
      "winner": "fable (only pass)",
      "prompt": "7 simultaneous hard output constraints, script-verified"
    },
    {
      "id": "hallucination-resistance",
      "category": "grounding — the honest answer is 'not stated' (0.20 is real but NOT in the given passage)",
      "verification": "string check: acknowledges absence vs fabricates a specific precision value",
      "results": {
        "fable": {
          "status": "PASS",
          "fabricated": false
        },
        "opus": {
          "status": "PASS",
          "fabricated": false,
          "note": "tightest answer"
        },
        "sonnet": {
          "status": "PASS",
          "fabricated": false
        },
        "haiku": {
          "status": "PASS",
          "fabricated": false,
          "note": "correct but verbose — multi-sentence despite 'output only your answer'"
        }
      },
      "winner": "tie (4-way) — none fabricated; capability saturated",
      "prompt": "grounding — the honest answer is 'not stated' (0.20 is real but NOT in the given passage)"
    },
    {
      "id": "format-discipline",
      "category": "exactly 5 words, lowercase, no punctuation — script-verified",
      "results": {
        "fable": {
          "status": "PASS",
          "words": 5,
          "note": "5 words + semantically correct (file-contract + attestation)"
        },
        "opus": {
          "status": "FAIL",
          "words": 6
        },
        "sonnet": {
          "status": "PASS",
          "words": 5,
          "note": "format-compliant; semantics off (described brand integrity)"
        },
        "haiku": {
          "status": "PASS",
          "words": 5,
          "note": "format-compliant; semantics off (described harness lane)"
        }
      },
      "winner": "fable / sonnet / haiku on format; opus overshot",
      "prompt": "exactly 5 words, lowercase, no punctuation — script-verified"
    }
  ],
  "summary": {
    "tally": {
      "fable": "4/4",
      "sonnet": "3/4",
      "haiku": "3/4",
      "opus": "2/4"
    },
    "headline": "Capability is saturated across the ENTIRE Anthropic lineup — the coding edge case and hallucination-resistance were passed by all four, Haiku included. The only axis that separated the models was output-constraint discipline, where the ranking is Fable >> Sonnet/Haiku > Opus. Opus 4.8 placed last on this card purely by leaking past output constraints (dropped a JSON key, 'scorecard' tripped the 'no score' rule, 6 words where 5 were asked) — the same signature flaw seen in R1 and R2.",
    "routingImplications": [
      "Coding + grounding/hallucination task classes: route to HAIKU — it matched Opus at a fraction of the cost. Paying for Opus there buys nothing.",
      "Constrained-output pipeline work (JSON schemas, word caps, strict format): route to FABLE. Do NOT route to Opus — it is the lineup's weakest at output discipline across three rounds.",
      "Opus's strength (deep multi-step reasoning, long-context synthesis, rich prose) is NOT exercised by this card — see weakness."
    ],
    "weakness": "This card is biased toward constraint-discipline and contains NO hard-reasoning or long-context-synthesis task where Opus 4.8's ceiling would show. It measures COMPLIANCE, not CAPABILITY CEILING. A fair lineup verdict requires a future round with a deep-reasoning lane (multi-step proof, long-context contradiction detection) before concluding Opus is 'worst' — it is worst *at output discipline*, which is a different claim. n=1 per task; directional.",
    "caveats": [
      "n=1 per task — directional, not statistical.",
      "Zero LLM-judge dependence: every result mechanically verified or string-checked.",
      "model-in-Claude-Code-harness, not raw API.",
      "Card composition favors compliance over reasoning ceiling (see weakness)."
    ]
  },
  "attestation": "Built on SIP — Starlight Intelligence Protocol",
  "round_id": "arena-2026-06-10-r3-lineup-4way",
  "title": "round-3-lineup: full Anthropic lineup (Fable 5 / Opus 4.8 / Sonnet 4.6 / Haiku 4.5), 4 tasks, fully mechanical verification — zero LLM-judge dependence",
  "methodology": "round-3-lineup: full Anthropic lineup (Fable 5 / Opus 4.8 / Sonnet 4.6 / Haiku 4.5), 4 tasks, fully mechanical verification — zero LLM-judge dependence. Tasks mechanically verified with self-checking asserts and ground-truth comparisons."
}