MLTF / Qwen3.8-27B 8bit
Play Tank Battle
Four supervised rounds, 430 requests.
From an open arena to a four-enemy maze.
Local models.
Real work, recorded.
From a solar system and browser games to long context. What models actually built, where they failed, and how human feedback changed the outcome.
Built by the models
Play directly in your browser
mlx-dspark / Qwen3.8-27B 8bit
Play Neon Bastion
Seven review rounds, about 30 hours.
Human play exposed gaps in automated checks.
All experiments
Loading…| Test date | Experiment / task | Runtime | Record |
|---|
No matching experiments
Try a model name or clear the current filters.
The process is part of the result.
These records cover generated projects, independent reviews, runtime observations and configuration comparisons. Repair rounds belong to the same experiment chain; document counts, model counts and independent sample counts are different quantities.
Local, LAN PC and cloud inference are labeled separately. Missing historical source, failed attempts, human interventions and measurement limits remain in the reports. Findings apply to the recorded tasks and environments.
Add a reportThe report could not be loaded
Reload the page. A local preview must be served over HTTP.