{
  "model": "kimi-code",
  "modelName": "Kimi K3.1 Pro",
  "modelId": "k3.1-pro",
  "date": "2026-10-26",
  "completedRuns": 7,
  "method": "Seven inspectable Kimi K3.1 Pro visual artifacts are published across two delivery sets. The three October 25 core builds were recovered unchanged from Kimi Code session checkpoints. The four October 26 Superbench 2.1 builds came from fresh, separate tool-free generation agents; each received one canonical prompt and returned one raw HTML response, which was packaged without code edits or a repair turn.",
  "packaging": "Validation targets the preserved public source HTML. The standard SEO postprocessor inserts canonical, noindex, and description tags into built artifact heads and may adjust adjacent whitespace; it does not change the artifact from the original charset tag onward, including all CSS, JavaScript, and body content.",
  "evaluation": "Artifacts are published for direct inspection, not as provider benchmark scores, passed rubrics, or an aggregate rating. Standalone, offline browser smoke, and targeted interaction results—including observed failures—are documented in each artifact's VALIDATION.md; they do not constitute complete independent rubric review of behavior, visual quality, playability, and performance.",
  "modelVersionEvidence": "The DeepSeek Harness runtime identified the generation model as k3.1-pro. No independent provider served-version handshake, sampling setting, reasoning setting, or output-budget read-back was available.",
  "isolation": "Each 2.1 retry used a fresh subagent and received only its canonical prompt plus a compact-output limit. The instructions prohibited tools, file inspection, reference implementations, and sibling outputs. This separation is instruction-based rather than an independent access-control attestation; the orchestrator supplied no reference material or tool results.",
  "retryDisclosure": "The first four generation attempts exhausted their response budgets before delivering complete documents. Concourse Divide and Lightkeeper's Atelier returned no closing document; Skyhook Cargo and Wrecklight Salvage returned truncated HTML. All four incomplete attempts were discarded without publication or continuation. Fresh compact retries received no prior draft and used explicit limits of 28,000 characters for briefs 21–23 and 30,000 for brief 24. The first compact Skyhook retry failed before delivery and its 25,000-character successor also truncated; both were discarded. The compact Atelier attempt delivered a closing document, but operator syntax validation found duplicate lx declarations, so it was rejected without repair. Fresh 16,000-character Skyhook and 24,000-character Atelier replacements were requested with no earlier draft; Skyhook truncated and Atelier returned no closing document. A further 16,000-character Atelier attempt also returned no closing document. A fresh 8,000-character Skyhook attempt delivered the accepted 6,484-byte closing document. A fresh 10,000-character Atelier attempt delivered a 9,325-byte closing document, but both shaders compiled while the program failed to link because shared uniform uS had mismatched precision; it was initially quarantined without repair. A new fresh Atelier generation was then requested with the same canonical prompt, a 10,000-character limit, and an explicit shader-interface link check, but that attempt exhausted its response room without a closing document. With no replacement available, the complete 9,325-byte closing document was restored unchanged and published as explicitly failed-run evidence, not a working scene.",
  "runSets": [
    {
      "date": "2026-10-25",
      "track": "core",
      "benchmarks": ["neon-drift", "ember-glider", "low-poly-tower-defense"],
      "method": "Built in Kimi Code. The original session logs record model k3.1-pro and alias kimi-code/k3.1-pro. Final saved game HTML was recovered from session checkpoints and published unchanged."
    },
    {
      "date": "2026-10-26",
      "track": "v21",
      "benchmarks": ["skyhook-cargo", "concourse-divide", "lightkeepers-atelier", "wrecklight-salvage"],
      "method": "Each frozen canonical prompt was the sole creative brief for a fresh, isolated Kimi K3.1 Pro subagent. Generation used no tools, browsing, file inspection, reference implementations, or other model artifacts. Each agent returned one complete HTML document. The orchestrator saved that response unchanged and added only separate publication metadata; no generated artifact received a repair turn.",
      "evaluation": "No numeric score or aggregate pass claim is assigned. Per-artifact operator reports record standalone, browser, and targeted behavioral checks, including observed failures; the five binary observations per brief otherwise remain an independent checklist."
    }
  ],
  "runs": [
    {
      "benchmark": "neon-drift",
      "entry": "/benchmarks/neon-drift/kimi-code/index.html",
      "manifest": "/benchmarks/neon-drift/kimi-code/manifest.json"
    },
    {
      "benchmark": "ember-glider",
      "entry": "/benchmarks/ember-glider/kimi-code/index.html",
      "manifest": "/benchmarks/ember-glider/kimi-code/manifest.json"
    },
    {
      "benchmark": "low-poly-tower-defense",
      "entry": "/benchmarks/low-poly-tower-defense/kimi-code/index.html",
      "manifest": "/benchmarks/low-poly-tower-defense/kimi-code/manifest.json"
    },
    {
      "benchmark": "skyhook-cargo",
      "entry": "/benchmarks/skyhook-cargo/kimi-code/index.html",
      "manifest": "/benchmarks/skyhook-cargo/kimi-code/manifest.json"
    },
    {
      "benchmark": "concourse-divide",
      "entry": "/benchmarks/concourse-divide/kimi-code/index.html",
      "manifest": "/benchmarks/concourse-divide/kimi-code/manifest.json"
    },
    {
      "benchmark": "lightkeepers-atelier",
      "entry": "/benchmarks/lightkeepers-atelier/kimi-code/index.html",
      "manifest": "/benchmarks/lightkeepers-atelier/kimi-code/manifest.json"
    },
    {
      "benchmark": "wrecklight-salvage",
      "entry": "/benchmarks/wrecklight-salvage/kimi-code/index.html",
      "manifest": "/benchmarks/wrecklight-salvage/kimi-code/manifest.json"
    }
  ]
}