STEALTHMARK · PHYSICS AND SPATIAL REASONING

Newton’s Mirror

Can a model keep physics, perspective and recursive reflections consistent in one SVG?

One scene. Connected constraints.

A convincing surface is only part of the task. We judge the physical state, camera placement, reflection geometry and materials together. A model must carry the same scene through every reflection.

Newton’s Mirror has the greatest influence on the current StealthMark score. Its reasoning categories also provide the main evidence for General Intelligence.

Inspect the reasoning

Follow the reflected shapes and compare their positions. Inconsistent geometry reveals where the model loses track of the scene, even when the drawing looks polished.

We withhold the exact prompt to reduce direct optimization for the test. Our published category scores and artwork show the results.

GPT-6 Astra57.8/100

GPT-6 Astra: Newton's Mirror SVG benchmark
Researcher observations

Major placement and the visible contact sequence are convincing. Dense planar light repetition, recognizable spherical panel images and simplified nested sphere motifs provide useful optical structure. Missing outer chains, truncated distant sequences and incorrect bead-like belts still leave the recursive reflections substantially incomplete.

Prompt withheld

This prompt is held back to prevent benchmaxxing.

Mimo V2.6 Flash7.8/100

Mimo V2.6 Flash: Newton's Mirror SVG benchmark
Researcher observations

The faint lower sphere image provides limited reflection evidence through reversed paint-color ordering. The wider reflection system and nested room imagery are absent. The colored beads do not resolve the required impact and transport history, while displaced paint positions and weak material rendering further limit the result.

Prompt withheld

This prompt is held back to prevent benchmaxxing.

Compare these models →All Newton’s Mirror results →