Visual rating collection

30 reproductions, 8 judges, 240 completed judgments out of 240 intended.

1 earlier request(s) are preserved in each affected result’s history. Only the selected assessment per judge contributes to scores, color baselines and ranks; attempts are never averaged together.

1 selected assessment(s) are qualified: the native client automatically added one recovery message in the same session before any assessment text was returned. Their exact final ratings are included, with the original strict extra-input audit failure retained. Details identify the affected responses; this is a disclosed departure from the ordinary single-turn input condition.

Each judge received one reference/result image pair in a fresh session, reference first, using the same eight-aspect prompt. Raw scores and explanations are retained without calibration or adjustment. Missing or invalid scores remain unavailable. Summaries use exactly one valid score for each named aspect V1–V8, independently of raw response formatting. Unknown extra entries remain visible and flagged but do not enter arithmetic. Duplicate, missing or invalid required scores make summaries unavailable; combined scores require every judge.

This measures visible resemblance at the supplied resolution. It does not establish game feel, interactivity or 3D implementation quality. The producing model’s identity and other judges’ scores were withheld from each judging session. A judge may assess its own model’s output under the same anonymous procedure.

Requested effort, image-delivery verification and parsing qualifications are shown with each response. Provider runtimes and unavoidable image preprocessing can differ. A later judge expands the collection without rerunning earlier judgments. Colors use the current collection’s same-judge, same-metric mean; changing the collection can change derived baselines and ranks while raw scores stay fixed.

• GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna model identities and max effort were observed. Gemini 3.8 high was configured; effective effort was not independently exposed. Fable 5.1, Sonnet 5.5, Opus 5.5 and Haiku 5.5 max was requested/configured and catalog-supported; effective per-turn effort was not exposed.
• Native image audits found unchanged 960×600 PNG pairs for GPT and Gemini judgments. Claude clients converted reference images and some result images to JPEG at unchanged dimensions; source PNGs and captures remained untouched. The Haiku extension’s 51 new requests included 40 converted payloads matching retained same-source native conversions. Provider-internal preprocessing is unmeasured.
• All 189 earlier selected assessments remain unchanged. The full collection has 30 reproductions and eight judges. The same model may judge its own anonymous output; no calibration or score adjustment is applied. One earlier failed Fable request remains separate from its selected replacement and never enters averages. Adding results changes the population used for derived means, colors and ranks.
• The October 8 judgments used already-installed Claude 2.1.293, Codex 0.162.0-alpha.2 and Antigravity 1.3.1. These differ from earlier cohorts; old assessments retain their original conditions. No executor upgrade or completed-pair rerun occurred. Gemini’s admitted native wrapper includes timestamp/model-selection metadata; full native system text and effective effort are unexposed.
• Sixty-five whole-response Markdown fences are extracted for display without changing scores or explanations; format flags and exact raw responses remain visible. One earlier Sonnet response contains an extra empty placeholder: all nine entries remain visible, while only its eight uniquely identified required scores enter summaries.
• One inherited Opus assessment includes a native-client recovery message before any assessment text. Its original final scores remain included with an explicit qualification and the strict extra-input audit failure disclosed. The other 239 selected assessments have ordinary completed status. All 51 October 8 requests completed in their original sessions without replacements or new qualifications.

Exact judging prompt · Public raw ratings

Collected through 2026-10-08T03:13:19.816192+00:00. Source collection SHA-256: 02b9ce5a9b99c7da4cd8bf3d8a41c94fbccfcc138029987c43fb342151f90349