{
  "model": "deepseek-v4-1-flash-1017",
  "modelName": "DeepSeek V4.1 Flash-1017",
  "modelId": "deepseek-flash",
  "suite": "Superbench 2.0",
  "promptEdition": "2.0",
  "date": "2026-10-26",
  "status": "integrated",
  "authorization": "The repository owner authorized this v2 evaluation set on 2026-10-26. Superbench 2.0 briefs are otherwise published without evaluations, and a v2 evaluation is an adapter, not the 1.x core/specialist protocol.",
  "expectedRuns": 4,
  "completedRuns": 4,
  "protocol": "One independent agent per brief. Each agent received the path to one frozen prompt file and one output path, and nothing else from the repository. It read that file, wrote one self-contained HTML document, and stopped: no browsing, no web fetch, no MCP or image tools, no reading site source or another model's artifact, no browser, screenshot, linter, bundler or test run, and no repair turn after delivery. The three re-runs were held to an explicit two-tool-call envelope (one read, one write); the fourth session reached the same shape on its own, ending at its report with no further tool calls.",
  "isolation": "Instruction-based separation in a shared filesystem. Every prompt told the agent not to inspect existing implementations, site source or sibling runs. Three sessions completed with exactly two tool calls under an explicit envelope and the fourth used the same two calls unaided. This is not an access-control sandbox and there is no independent file-access attestation.",
  "modelVersionEvidence": "The runs executed on the DeepSeek Harness agent runtime, whose model route is the deepseek-flash alias. DeepSeek documents that alias as serving DeepSeek-V4.1-Flash-1017 from the October 17, 2026 update, and this set was produced on October 26, 2026. Neither the harness nor the proxy returned a served-version value, so the 1017 label rests on the documented alias mapping and the run date, not on a version handshake. No sampling, reasoning-effort or output-budget setting was pinned.",
  "review": "Registry validation only means an entry file was ingested. Reviewer checks performed after delivery: offline load in headless Chrome 155 at 1280x800 and 390x844, zero JavaScript exceptions, zero console errors, zero third-party requests, and the specific printed acceptance values recorded in each VALIDATION.md. The reviewer also re-derived the Prism Freight Pavilion default solution by enumerating all 720 permutations independently. No five-criterion rubric score, rank or pass badge is stored, and no claim is made that any brief was passed.",
  "disclosure": "One operator reminder was sent to all four sessions while they were authoring, because no artifact had appeared yet. In the briefs 17, 18 and 19 sessions that reminder arrived after their first write and started a second authoring pass, which the edition 2.0 contract does not allow. Those drafts were discarded and are not published; the three briefs were then re-run with no further operator messages and a strict two-call envelope. The brief 20 session had already reported and never took another step, so its first delivery is published unchanged.",
  "limits": [
    "Agent-harness delivery, not a hosted-API single-message completion: the model could read a prompt file and write a file, and the harness owns the context and step structure.",
    "The served model version is inferred from documented alias routing, not read back from the provider.",
    "Reviewer coverage is partial. Each VALIDATION.md lists which brief checks were exercised and which remain open for a human reviewer.",
    "No quantitative comparison is possible with Superbench 1.x runs, which use a different protocol, or with provider benchmark scores.",
    "Four artifacts in one pass per brief say nothing about pass rates, stability across retries, or behaviour outside these briefs."
  ],
  "runs": [
    {
      "benchmark": "prism-freight-pavilion",
      "prompt": "benchmarks/prompts/17-prism-freight-pavilion.md",
      "entry": "/benchmarks/prism-freight-pavilion/deepseek-v4-1-flash-1017/index.html",
      "manifest": "/benchmarks/prism-freight-pavilion/deepseek-v4-1-flash-1017/manifest.json",
      "validationReport": "/benchmarks/prism-freight-pavilion/deepseek-v4-1-flash-1017/VALIDATION.md",
      "files": 3,
      "bytes": 95373,
      "externalAssets": [],
      "reviewerConclusion": "Defaults, manual cost ledger and the exact solve agree with independent enumeration: 4 legal arrangements of 720, unique zero-cost ACBDEF, 720 candidates scanned. Relaxation, hard-infeasibility, edit-invalidation and keyboard paths were not exercised."
    },
    {
      "benchmark": "tidal-ledger-observatory",
      "prompt": "benchmarks/prompts/18-tidal-ledger-observatory.md",
      "entry": "/benchmarks/tidal-ledger-observatory/deepseek-v4-1-flash-1017/index.html",
      "manifest": "/benchmarks/tidal-ledger-observatory/deepseek-v4-1-flash-1017/manifest.json",
      "validationReport": "/benchmarks/tidal-ledger-observatory/deepseek-v4-1-flash-1017/VALIDATION.md",
      "files": 3,
      "bytes": 95449,
      "externalAssets": [],
      "reviewerConclusion": "Initial inventory 12 is conserved; prefix k=3 and the end-of-journal state match the brief's expected physical, held and rejected values. Scrub stability, reverse-order import, duplicate collapse, branching and malformed-import atomicity were not exercised."
    },
    {
      "benchmark": "last-light-relay",
      "prompt": "benchmarks/prompts/19-last-light-relay.md",
      "entry": "/benchmarks/last-light-relay/deepseek-v4-1-flash-1017/index.html",
      "manifest": "/benchmarks/last-light-relay/deepseek-v4-1-flash-1017/manifest.json",
      "validationReport": "/benchmarks/last-light-relay/deepseek-v4-1-flash-1017/VALIDATION.md",
      "files": 3,
      "bytes": 109901,
      "externalAssets": [],
      "reviewerConclusion": "Default solve returns makespan 10 with exactly the brief's witness and a passing predecessor/crane table, reported as an exhaustive 15-placement search. The finish-9 lower bound was not independently re-derived; recovery, manual validation and execution paths were not exercised."
    },
    {
      "benchmark": "quarantine-signal-desk",
      "prompt": "benchmarks/prompts/20-quarantine-signal-desk.md",
      "entry": "/benchmarks/quarantine-signal-desk/deepseek-v4-1-flash-1017/index.html",
      "manifest": "/benchmarks/quarantine-signal-desk/deepseek-v4-1-flash-1017/manifest.json",
      "validationReport": "/benchmarks/quarantine-signal-desk/deepseek-v4-1-flash-1017/VALIDATION.md",
      "files": 3,
      "bytes": 97103,
      "externalAssets": [],
      "reviewerConclusion": "Hostile payloads stay literal text, no request or alert fires, and one inline script is the only script element. Default dispatch, waiting and blocked sets, reason order and claimedRole inertness match the brief. Hold precedence, capacity extremes, preview/commit atomicity and malformed imports were not exercised."
    }
  ]
}