Pelican Test Expands Into Video

This is a decent hacky benchmark because it punishes shallow captioning and rewards actual video understanding plus layout-aware code generation.

• –It shifts the test from static SVG composition to temporal reconstruction, which is harder and more revealing for multimodal models.
• –The real signal here is spatial fidelity: can the model preserve text placement, motion, edits, and transitions without hand-holding?
• –The prompt is still informal and noisy, so it’s better as a vibes benchmark than a rigorous eval suite.
• –Interesting that the author highlights line positioning; that usually exposes whether the model is actually parsing structure or just pattern-matching aesthetics.
• –If this catches on, expect people to use it as a quick litmus test for video-capable models, especially in local/VLLM circles.