Proofrun

Blind AI reviews with receipts.
Same brief, frozen limits, human judgment, public evidence.

[Proofrun @proofrunai]

Proofrun tests AI models by the work they produce—with frozen briefs, human evaluation, and receipts.
For the Grok 4.6 launch, we asked it to build Sunwell Sanctuary in SceneScript.

83,932 voxels.

Human score: 90/100.
Explore the scene and full evidence:
https://t.co/h6ipoBcM1u

[Proofrun @proofrunai]

This is our rapid first result—not the end of the evaluation. A broader Proofrun profile is next: Triage Desk, Gravity Desk, and Sky Archive, with frozen settings and complete attempt records.

[Proofrun @proofrunai]

Claude Fable 5.1 just arrived. We put it through one frozen Sunwell Sanctuary run.
77/100. 85% requirement compliance. 71,710 voxels.
The scene rendered—but two invalid operations were preserved and counted.
Full evidence: https://t.co/UPcd00x2il

[Proofrun @proofrunai]

Tencent Hy4 preview did not produce one tidy number.
Triage Desk: 65
Weather Story: provider incident
Verdant Terminal: 61
Three frozen tests. Two completed artifacts. One HTTP 429. No composite model score.
Proofrun Issue 015:
https://t.co/C7w5wrI1gs

[Proofrun @proofrunai]

Three records, one verdict: our bug, a provider 429, then Qwen3.8 Flash’s Triage Desk artifact scored 76/100 and passed 10/10.
All public. Only model evidence shaped it.
https://t.co/xOCZwccbHT

[Proofrun @proofrunai]

Proofrun Issue 011: mystery model Ox Alpha passed all 6 Orbit Radio requirements and scored 56/100.

It worked cleanly, but looked generic and lacked originality.

One model, one frozen Solo Run—not a general rating.

https://t.co/czaUSiHwze

[Proofrun @proofrunai]

Proofrun Issue 010: GLM 5.3 edged Gemini 3.7 Flash 46–34 on Orbit Radio.
Both met every requirement; GLM led on polish and dial control in this low-confidence blind duel. Try both interactive artifacts: https://t.co/ViTMEi5Bwc

[Proofrun @proofrunai]

Proofrun Issue 008: GPT-5.6 Terra scored 78/100 with 100% requirement compliance on Orbit Radio, an interactive space-radio experience.
Try the artifact and read the evidence:
https://t.co/uRjkx0qYsz

[Proofrun @proofrunai]

Same frozen test. Five Grok 4.6 generations scored 86, 8, 92, 77 and 89 on Prism Garden.
Four met every requirement. One met none.
The observed range was 8–92. That is why we test repeatability, not just best runs.
https://t.co/EZLCr1BJrB

[Proofrun @proofrunai]

The model-authored camera remains the canonical capture. Proofrun also generated deterministic context and orthographic survey views so readers can inspect more of the complete construction.

[Proofrun @proofrunai]

Proofrun tests AI models by the work they produce—with frozen briefs, human evaluation, and receipts.
For the Grok 4.6 launch, we asked it to build Sunwell Sanctuary in SceneScript.

83,932 voxels.

Human score: 90/100.
Explore the scene and full evidence:
https://t.co/h6ipoBcM1u

[Proofrun @proofrunai]

@hazelf0x Both models received the same frozen brief in one-shot offline web mode. These are matched 33-bloom Capture Lab states at 1280×720. Evaluation was locked before identities were revealed. Public inspectable results are coming soon.

[Proofrun @proofrunai]

@hazelf0x Camera, sky, lighting and composition were part of each model’s submission. These are the locked, unedited model-authored views.
One run, not a universal ranking. Public inspectable results and more blind tests are coming!