{
  "schema": "s2s-bench-v1",
  "exhibit": "outside-judges",
  "published_utc": "2026-08-13",
  "status": "PUBLISHED — live at research.strata2signal.com.",
  "licence": "CC BY 4.0",
  "attribution": "strata→signal research, research.strata2signal.com",
  "hardware": "This exhibit is the one leg of the arc that used no local GPU at all: every judge and every candidate here answered over a network. The instruments, scorers and sealed inputs are the workshop's own.",
  "dataset": "outside-judges-provenance",
  "what_this_is": "What was pinned before it could be scored, and what the pinning did not cover. Artifact names only — never a path, never a host.",
  "rule": "A pre-registration that can be edited after the numbers arrive is not a pre-registration. These hashes are what makes the claim checkable: hold the file, hash it, compare. But a hash proves IDENTITY, not TIME — so each registration carries the sha in force at its first scored call, read from git, beside the sha as published. Where the two columns cannot both be produced, the row says why.",
  "findings": {
    "summary": "2 provenance findings, published because the whole point of this page is that we do not relax a rule for ourselves. Neither of them moves a number: no case, arm, floor, ceiling, population or counting rule is affected by either, and every figure on this page was produced by frozen instruments over sealed inputs. What they affect is how strongly the REGISTRATION can be proved, which is exactly the sort of thing a page about outside checking should not quietly round up.",
    "entries": [
      "PREREG-PANEL-EXT.md: committed after its first scored call (2026-08-13T03:20:47+00:00); content frozen on disk beforehand and unchanged since.",
      "PREREG-INDEX.md does not record either of these two files' sha256. PREREG-GAUNTLET's own §9 step 2 requires it (\"record this file's sha in PREREG-INDEX.md before the first scored call\") and that step was not performed. The shas are published here instead, over bytes anyone can hash; the registered index entry is owed and is not pretended to exist."
    ]
  },
  "preregistrations": [
    {
      "role": "the cloud judging extension — the audition, the seat map, the rounds",
      "artifact": "PREREG-PANEL-EXT.md",
      "first_scored_call_utc": "2026-08-13T03:20:47+00:00",
      "sha_at_first_scored_call": null,
      "sha_at_publication": "ca422ac6928bd01acfb6d501f0cd7aaea3d2d3b5e4cbfc386ed462cdd924721c",
      "committed_before_first_scored_call": false,
      "unchanged_since_first_scored_call": null,
      "identical_to_the_bytes_committed": true,
      "finding": "the file was on disk and frozen before the first scored call — its bytes are byte-identical to what was committed minutes later and to what ships in this kit, and no character of it moved — but the COMMIT that witnesses the time landed AFTER the run began. The content pin holds; the independent time witness does not, and a registration whose only proof of time is our own file timestamp is weaker than one git can date. Stated rather than smoothed.",
      "on_disk_utc": "2026-08-13T03:20:41.339011+00:00"
    },
    {
      "role": "the gauntlet — three candidates on the frozen seat instrument, and the budget",
      "artifact": "PREREG-GAUNTLET.md",
      "first_scored_call_utc": "2026-08-13T04:50:44+00:00",
      "sha_at_first_scored_call": "98bdc43a22ae2ec2791b44202e8479635145926547b22a83cbec0dc45e5dd0b5",
      "sha_at_publication": "98bdc43a22ae2ec2791b44202e8479635145926547b22a83cbec0dc45e5dd0b5",
      "committed_before_first_scored_call": true,
      "unchanged_since_first_scored_call": true,
      "identical_to_the_bytes_committed": true,
      "on_disk_utc": "2026-08-13T03:19:37.190026+00:00"
    },
    {
      "role": "the index of the arc's pre-registrations, with its own re-pin history",
      "artifact": "PREREG-INDEX.md",
      "sha256": "8a56bde525f92b2d8c8097c5942f870e389c9b3c6a237f4db77e21a7dc19990b"
    }
  ],
  "frozen_instrument": {
    "what_this_is": "the seat instrument's three artifacts, recomputed from disk here and compared against the shas its pre-registration pinned. The runner performs this check before it renders a single case and exits on a mismatch; the kit performs it again so the published artifact carries the proof rather than the claim.",
    "all_three_match_their_pins": true,
    "artifacts": [
      {
        "role": "the answer key — 43 cases, 27 kill / 16 preservation",
        "artifact": "judge-cases-v1.json",
        "sha256": "6a226b62dceb005dbd658bc13cf229e38f6f0a2f07ae51091f444ec3e96b56d1",
        "sha_pinned_in_prereg": "6a226b62dceb005dbd658bc13cf229e38f6f0a2f07ae51091f444ec3e96b56d1",
        "matches_pin": true
      },
      {
        "role": "the floors",
        "artifact": "THRESHOLDS.md",
        "sha256": "37af75ffa18484a969528e894ef16f7fdcd10eb364e0669ac4f9333ce5279ba0",
        "sha_pinned_in_prereg": "37af75ffa18484a969528e894ef16f7fdcd10eb364e0669ac4f9333ce5279ba0",
        "matches_pin": true
      },
      {
        "role": "the instrument — the prompt pack every candidate answered",
        "artifact": "judge-prompt-packs-v2-prompted.md",
        "sha256": "92af6030dddface2c863fe9d672bec6a6bb1017122c0028e435214acd9fb298d",
        "sha_pinned_in_prereg": "92af6030dddface2c863fe9d672bec6a6bb1017122c0028e435214acd9fb298d",
        "matches_pin": true
      }
    ]
  },
  "original_rounds": {
    "note": "the sealed batches, letter maps and key files this extension read are exhibit six's and the voice addendum's, unchanged. The extension writes only verdict files and never a key file, which is what makes the published populations structurally unmovable rather than carefully unmoved.",
    "published_scorer_cannot_see_a_cloud_verdict": "the published scorer and the committed recount both walk key files; a cloud seat has no key directory. There is nothing for them to read."
  },
  "scorers_and_harnesses": [
    {
      "role": "the cloud judging driver: audition, retry, normalisation, ledger",
      "artifact": "cloud_judges.py",
      "sha256": "bfc9906f849def70521b08d87e06b5a7e454e07cb1fcfd739551c4dee1ba69c2"
    },
    {
      "role": "the cloud transport and its registered num_ctx drop",
      "artifact": "cloud_api.py",
      "sha256": "028e16f59ecff0ecbaa7a615fb02127d2e24f2ddf1b243ee6c997b3f73789e24"
    },
    {
      "role": "the driver's tests, including the bytes-on-the-wire assertion",
      "artifact": "test_panel_ext.py",
      "sha256": "75d2e0741be1fd24e82f88a7721914d3d942de2f82358198ab3711b1a44f8601"
    },
    {
      "role": "the narrator scorer, called verbatim for the original panel",
      "artifact": "voice_score.py",
      "sha256": "11cdafcb95ccd6551761eb4c210554862ca951c4b964efdc00a2dfb1734618e4"
    },
    {
      "role": "the independent, non-importing recount",
      "artifact": "recount.py",
      "sha256": "8a2be349d21a7e3604b89a113d62a831192a928ea69490fddb3757008c5f85a3"
    },
    {
      "role": "the Claude arm: renders the sealed cases, collects the replies",
      "artifact": "claude_arm.py",
      "sha256": "b86b960aba8d45527a07c1acd1f2c1906f9494df52813b04d49195e7e5962cd2"
    }
  ],
  "judges": {
    "model_ids": [
      "claude-opus-5",
      "claude-fable-5",
      "deepseek-v4-pro",
      "mistral-large-3:675b",
      "nemotron-3-ultra",
      "kimi-k3"
    ],
    "vendors": [
      "anthropic",
      "deepseek",
      "mistral",
      "moonshot",
      "nvidia"
    ],
    "disclosure": "One of the six judges' families wrote the character questions, the canon notes, the pre-registrations and this page, and two of the three gauntlet candidates are that family's models. That conflict is why the other five judges exist, and their columns are the check on ours — but it is not cured by them, and it is stated wherever a number it touches is printed.",
    "dating": "Cloud judges are versionless hosted services. Every cloud verdict in this kit was produced 2026-08-12/13 and is dated rather than version-pinned; it may not reproduce against a later checkpoint of the same tag. The local instruments and scorers these judges audit are sha-frozen, and the asymmetry is stated rather than hidden."
  },
  "data_boundary": "What left the box for judging was model outputs over public-domain text and published game canon — never user data, never a character's private canon, never telemetry. The cloud judges saw the sealed batches published in this kit's terms and nothing else: no letter map, no model name, no arm roster, no other judge's answer, no part of a key file.",
  "no_paths": "Artifacts are named, never located. No box names, no addresses, no filesystem paths.",
  "judge_dating": "Cloud judges are versionless hosted services. Every cloud verdict in this kit was produced 2026-08-12/13 and is dated rather than version-pinned; it may not reproduce against a later checkpoint of the same tag. The local instruments and scorers these judges audit are sha-frozen, and the asymmetry is stated rather than hidden.",
  "contamination_caveat": "Published 2026-08-13. Models with a later training cutoff may have seen these sets. We author fresh sets each cycle; this one is not a standard, it is our kit, yours to reuse.",
  "not_a_standard": "Our own trials, our own judges, our own instruments. Not first independent numbers, and not a benchmark."
}