{
  "$comment": "Built on SIP — Starlight Model Arena run receipt. Schema: arena-run v0.1.",
  "date": "2026-06-09",
  "harness": "claude-code-agent-tool",
  "skill": "model-arena (formalized this session at ~/.claude/skills/model-arena/SKILL.md)",
  "contestants": [
    "claude-fable-5",
    "claude-opus-4-8"
  ],
  "judge": {
    "model": "claude-sonnet-4-6",
    "blind": true,
    "labelAssignment": {
      "task4": {
        "A": "opus",
        "B": "fable"
      }
    }
  },
  "tasks": [
    {
      "id": "reasoning-consecutive-divisors",
      "category": "reasoning + output discipline (no tools allowed)",
      "verification": "ground truth 33 (3*11, 2*17, 5*7) computed by harness script before dispatch; output must be ONLY the integer",
      "results": {
        "fable": {
          "answer": "33",
          "correct": true,
          "constraintCompliant": true,
          "durationMs": 4973,
          "tokens": 59271,
          "notes": "Correct, bare integer, zero extra characters."
        },
        "opus": {
          "answer": "814",
          "correct": false,
          "constraintCompliant": true,
          "durationMs": 2676,
          "tokens": 59123,
          "notes": "Wrong answer (814 = 2*11*37 has 8 divisors). Fastest response of the round — speed may have cost accuracy."
        }
      },
      "winner": "fable",
      "prompt": "reasoning + output discipline (no tools allowed)"
    },
    {
      "id": "coding-recursive-descent-evaluator",
      "category": "coding (real parser required — eval/exec/compile/ast banned; self-verifying asserts incl. stacked unary minus; max 3 attempts)",
      "verification": "harness independently re-ran both tests.py (ALL PASS confirmed) and grepped solutions for banned calls (none found)",
      "results": {
        "fable": {
          "status": "PASS",
          "attempts": 1,
          "outputContractCompliant": true,
          "durationMs": 54012,
          "tokens": 64066,
          "notes": "Recursive descent, first try. Final message exactly the required two lines."
        },
        "opus": {
          "status": "PASS",
          "attempts": 1,
          "outputContractCompliant": false,
          "durationMs": 35442,
          "tokens": 64284,
          "notes": "Recursive descent, first try, faster. Added an explanatory preamble above the required two-line output — contract violation."
        }
      },
      "winner": "tie (correctness); fable on output contract",
      "prompt": "coding (real parser required — eval/exec/compile/ast banned; self-verifying asserts incl. stacked unary minus; max 3 attempts)"
    },
    {
      "id": "agentic-repo-grounding-frankx",
      "category": "agentic tool use + repo-grounded accuracy (NEW axis this round)",
      "verification": "four ground truths fixed by harness before dispatch: 14 aliases (live count), Friday 11:00 CET, lib/voice/frankx-voice.ts, no + private/404-agent-queue.jsonl",
      "results": {
        "fable": {
          "correct": "4/4",
          "formatCompliant": false,
          "toolUses": 4,
          "durationMs": 32120,
          "tokens": 63369,
          "notes": "All facts correct including live alias count. Dropped the required 'N:' line prefixes — 4 lines, wrong pattern."
        },
        "opus": {
          "correct": "4/4",
          "formatCompliant": false,
          "toolUses": 2,
          "durationMs": 10091,
          "tokens": 63404,
          "notes": "All facts correct, 3x faster, fewer tool calls. Added a preamble above the four lines — 'EXACTLY four lines' violated."
        }
      },
      "winner": "tie (4/4 both; neither fully format-compliant — fable missed prefixes, opus added preamble)",
      "prompt": "agentic tool use + repo-grounded accuracy (NEW axis this round)"
    },
    {
      "id": "voice-constraint-stack",
      "category": "brand-voice writing (90-110 words, required phrase, 8 banned words, final sentence <=6 words)",
      "verification": "harness script: fable 106 words / opus 102 words, both contain 'model-in-harness', zero banned hits, final sentences 5 and 3 words — BOTH fully compliant",
      "results": {
        "fable": {
          "judgeScore": 9,
          "wordCount": 106,
          "constraintCompliant": true,
          "durationMs": 21846,
          "tokens": 60477
        },
        "opus": {
          "judgeScore": 8,
          "wordCount": 102,
          "constraintCompliant": true,
          "durationMs": 7162,
          "tokens": 59494,
          "notes": "Closed the Round 1 compliance gap completely. Blind judge preferred fable's causal tightness this time."
        }
      },
      "winner": "fable (judge 9 vs 8, both constraint-compliant)",
      "prompt": "brand-voice writing (90-110 words, required phrase, 8 banned words, final sentence <=6 words)"
    }
  ],
  "summary": {
    "tally": {
      "fable": 2,
      "opus": 0,
      "tie": 2
    },
    "headline": "Fable 5 takes Round 3 (hard-capability card): the only model to solve the no-tools reasoning task (Opus answered 814 vs ground truth 33), and winner of the blind style judgment — reversing Round 1's style split. Output-discipline edge held but is not spotless: Fable slipped once on a format pattern (dropped line prefixes) while Opus violated output contracts twice via unrequested preambles. Opus was faster on 3 of 4 tasks; on the task it lost, it was fastest of the round.",
    "roundDeltas": [
      "Reasoning: parity in R1 -> fable outright in R3 (opus incorrect, 814 vs ground truth 33, on a harder no-tools problem).",
      "Style: opus won the blind judge in R1 -> fable won in R3; n=1 both rounds, treat style as contested.",
      "Constraint compliance: consistent across R1/R2/R3 — fable precise on stacked constraints; opus leaks preambles past 'output only' contracts (twice this round).",
      "New agentic axis (R3): correctness parity 4/4; opus notably more efficient (2 tool calls vs 4, 10s vs 32s).",
      "Complements R2 stress card: R2 found opus stronger on situational judgment (gate-flagging, spec pushback); R3 finds fable stronger on hard clean reasoning. Different axes, both receipted."
    ],
    "caveats": [
      "n=1 per task — directional, not statistical.",
      "Judge is a Claude-family model (sonnet); blind shuffled labels mitigate but do not eliminate family bias.",
      "Both contestants run inside the Claude Code agent harness; results measure model-in-harness, not raw API behavior.",
      "Latency includes harness overhead; treat durations as rough.",
      "Single-judge style scores flipped between rounds — style verdicts need repeated rounds before any routing change."
    ]
  },
  "attestation": "Built on SIP — Starlight Intelligence Protocol",
  "round_id": "arena-2026-06-09-r3-true-challenge",
  "title": "Round 3: True Challenge (2026-06-09)",
  "methodology": "Same prompt dispatched per task to two subagents with per-agent model override, two waves of 4 (GREEN-zone concurrency cap). Ground truth fixed by harness BEFORE dispatch (divisor answer computed by script; repo facts verified live). Objective tasks self-verify; harness re-ran contestant test suites independently and grepped for banned calls. Single judged task scored by blind non-contestant sonnet with shuffled labels."
}