See the thing that was judged.
Proofrun preserves the actual output, its fixed evidence, and—when a safe public renderer exists—the artifact itself.
Independent AI model testsIssue 006 / Helsinki
AI models,
judged by the work.
Frozen briefs, blind scoring, working artifacts, costs, failures, and conclusions narrow enough to trust.
How a Proofrun works ↘Observed, not advertised.
Latest solo test01
Grok 4.6 built an 83,932-voxel Sunwell Sanctuary on its first attempt
The solo Voxel Forge run passed all ten locked requirements and earned 90/100, with 259 accepted SceneScript statements, zero ignored statements, a standout waterfall—and an authored camera that showed less than the model actually built
Read the report →
Modelx-ai/grok-4.6
Confidencehigh
Artifact score90/100
Requirements met10/10
What every Proofrun includes02
One result.
Three layers of proof.
The page is the argument: working evidence first, human judgment kept explicit, and a publication record readers can inspect for themselves.
Taste stays visible and bounded.
Requirements and scores lock before identities are revealed. The editor’s view remains human, named, and narrow.
The conditions travel with the claim.
Time, cost, attempts, protocol, limitations, and a sanitized record remain attached to every published conclusion.
Proofrun protocol / v0.703
From frozen brief
to bounded claim.
01Freeze
One prompt, one runtime, one scoring contract.
02Run blind
Model identities stay hidden during judgment.
03Stress
A repeatable probe exercises every artifact.
04Lock
Requirements, scores, and the assessment become immutable.
05Reveal
Only then do model names enter the record.
06Publish
A sanitized, hashed evidence packet crosses the boundary.
Published reports04
006Sunwell SanctuaryVoxel Forge90/100Published→005Weather StoryData visualization92—90Published→004Sky ArchiveVoxel Forge74/100Published→003Prism GardenCreative coding85/100Published→002Prism GardenCreative coding70—77Published→001Gravity DeskPhysics toy73—87Published→007Next runNot yet tested—Open·
Browse all reports ↗