Donatello AI Benchmark / Protocol & evidence
Versioned input protocols, historical planning estimates and a separate measured API pilot. Original v1 downloads remain unchanged; no human votes or quality scores are claimed.
Protocol v1 and its downloads remain available. The separately registered October 6 pilot adds 24 API attempts, technical results and costs. No public votes, independently reviewed quality scores or quality winners are published.
Inspect the 24-attempt image API pilot: full inputs, settings, decoded outputs, response times and charges, with a downloadable JSON evidence file. It does not replace the historical files below.
Archive correction · October 6: the Creator Value v1 downloads have been restored to their original calculation and checksums after a live-price update affected them. They are now decoupled from current pricing. Read the correction record.
JSON preserves settings, source snapshots and nested caveats. CSV preserves the record fields; empty cells mean unknown, not zero. Nested lists are JSON-encoded within CSV cells. Dates are ISO dates, cash values are USD before tax, and durations are seconds. Benchmark records contain no customer data.
protocol · 30 records
Thirty input prompts and acceptance checks. Not generation results.
protocol · 8 records
Eight planned cases with reference readiness and the public input manifest. No generated test outputs.
documentation-estimate · 60 records
Sixty product/budget/brief combinations, with unknown full costs and completion explicitly retained.
Donatello.studio. Donatello AI Benchmark: protocol and documentation release. Version 2026-09-11-v1, published 2026-09-11. Cite the relevant dataset's URL and checksum, and identify it as a protocol or estimate—not an experiment proving a model's quality.
The prompt corpus and factual records are publicly accessible. Input media, future generated media and provider documentation retain their applicable rights; this page does not grant a blanket media or training licence. See Terms.
No image-quality score, human-preference percentage, cost-per-usable result exists yet. The separate pilot has response-time and charge tables, not end-to-end consumer performance or independent evaluation. Publication dates on input files are not test-execution dates.