{
  "$comment": "Built on SIP — Starlight Model Arena run receipt. Schema: arena-run v0.1.",
  "date": "2026-06-10",
  "harness": "claude-code-agent-tool",
  "skill": "model-arena",
  "contestants": [
    "claude-fable-5",
    "claude-opus-4-8"
  ],
  "judges": {
    "t1": {
      "model": "claude-sonnet-4-6",
      "blind": true,
      "labelAssignment": {
        "A": "opus",
        "B": "fable"
      }
    },
    "t2": {
      "model": "claude-sonnet-4-6",
      "blind": true,
      "labelAssignment": {
        "A": "fable",
        "B": "opus"
      }
    }
  },
  "tasks": [
    {
      "id": "design-lab-component",
      "category": "real-world frontend work (read design.md + taste.md, build ModelComparisonCard.tsx, tsc-clean, design-token + a11y compliance)",
      "verification": "harness ran tsc --noEmit (clean, both), banned-pattern greps (style={{, arbitrary hex, purple/pink/gradient: 0 hits both), semantics + visible-verdict-text greps (present both); blind judge on craft",
      "results": {
        "fable": {
          "tscClean": true,
          "mechanicalCompliant": true,
          "judgeScore": 6,
          "lines": 70,
          "outputContractCompliant": false,
          "toolUses": 7,
          "durationMs": 115899,
          "tokens": 82375,
          "notes": "article element, font-mono taste, min-w-0 discipline — but three parallel Record maps for one enum and zero a11y affordances beyond color+text. FIRST recorded fable output-contract violation (preamble above required two lines)."
        },
        "opus": {
          "tscClean": true,
          "mechanicalCompliant": true,
          "judgeScore": 8,
          "lines": 118,
          "outputContractCompliant": false,
          "toolUses": 7,
          "durationMs": 94258,
          "tokens": 82161,
          "notes": "Table semantics with sr-only caption, scoped headers, aria-hidden decorative dot, unified VERDICT_META record. Judge: 'textbook accessible comparison-card craft'. Preamble leak again."
        }
      },
      "winner": "opus (judge 8 vs 6; both mechanically compliant)",
      "prompt": "real-world frontend work (read design.md + taste.md, build ModelComparisonCard.tsx, tsc-clean, design-token + a11y compliance)"
    },
    {
      "id": "acos-skill-authoring",
      "category": "agentic-system authoring (SKILL.md per ACOS conventions: frontmatter triggers, <800 words, voice constraints, no slop, no mythology leak)",
      "verification": "harness greps: frontmatter name+desc+trigger present (both), words 630/543 (both <800), slop 0/0, mythology 0/0, numbered workflow present (both); blind judge on usefulness",
      "results": {
        "fable": {
          "mechanicalCompliant": true,
          "judgeScore": 9,
          "words": 630,
          "outputContractCompliant": true,
          "toolUses": 6,
          "durationMs": 78999,
          "tokens": 72026,
          "notes": "Eight-step workflow with concrete field names, mechanical-over-taste principle, append-only history rule, one-claim-per-round-of-evidence. Exactly the required one-line final message."
        },
        "opus": {
          "mechanicalCompliant": true,
          "judgeScore": 8,
          "words": 543,
          "outputContractCompliant": false,
          "toolUses": 3,
          "durationMs": 48219,
          "tokens": 67620,
          "notes": "'Press release' framing + judged-share<=half constraint fable lacked; omitted append-only history + promotion threshold. Faster, half the tool calls. Preamble leak."
        }
      },
      "winner": "fable (judge 9 vs 8; both mechanically compliant)",
      "prompt": "agentic-system authoring (SKILL.md per ACOS conventions: frontmatter triggers, <800 words, voice constraints, no slop, no mythology leak)"
    }
  ],
  "summary": {
    "tally": {
      "fable": 1,
      "opus": 1,
      "tie": 0
    },
    "headline": "Premium work samples split 1-1: Opus 4.8 built the more rigorously accessible component (table semantics, unified type design); Fable 5 authored the sharper agentic skill doc. The new finding is about discipline under load: Fable 5 violated an output contract for the first time across four rounds — when the task itself is heavy, its constraint edge narrows. Opus was faster on both tasks and used half the tool calls on the authoring task.",
    "roundDeltas": [
      "Quality on real work: split by domain — a11y/code-craft to opus, doc/system-authoring to fable. Neither dominates premium work.",
      "Output discipline: fable's first contract violation (R1-R3: 0; R4: 1 of 2). Opus: violations on both tasks (now 5 of 7 structured-output tasks across rounds).",
      "Efficiency: opus faster on both R4 tasks, continuing the R3 agentic-axis pattern.",
      "Implication for routing: on heavy multi-step tasks, enforce output contracts structurally (schemas, StructuredOutput) regardless of model — model discipline degrades with task load."
    ],
    "caveats": [
      "n=1 per task — directional, not statistical.",
      "Judges are Claude-family (sonnet); blind shuffled labels mitigate but do not eliminate family bias.",
      "Model-in-Claude-Code-harness, not raw API behavior.",
      "Latency includes harness overhead; treat durations as rough.",
      "Work-sample quality scores are single-judge taste; repeat before routing on them."
    ]
  },
  "attestation": "Built on SIP — Starlight Intelligence Protocol",
  "round_id": "arena-2026-06-10-r4-work-samples",
  "title": "Round 4: Work Samples (2026-06-10)",
  "methodology": "Premium work-sample card: real production tasks instead of puzzles. Two tasks x two models, one parallel wave. Mechanical verification by harness (tsc, banned-pattern greps, frontmatter/word-count/slop checks) + blind non-contestant judges (sonnet) with shuffled labels for craft quality. Contestants identified only as c1/c2 directories to judges."
}