2026 · Research design, geometry system & evaluation
GeoSketch-CoT: Executable Geometry Scratchpads
We turn auxiliary geometry constructions into executable, verifiable, and replayable intermediate states, then use paired experiments to separate visual gains from reasoning-protocol gains.
- 149 passed
- core and web tests
- 85 / 92
- Flash confirm correct
- +2.000
- two-stage utility gain
Problem
Multimodal models may miss diagram relations or invent invalid constructions, while apparent gains can be confounded by extra reasoning tokens and output formatting.
Approach
We built Graph Schema v2, a constrained drawing DSL, transactional execution, geometric verification, SVG/PNG rendering, and a research workbench with matched-text and counterfactual controls.
Outcome
All 149 tests and both production builds pass. The visual scratchpad itself is not yet a reliable improvement; the confirmed gain comes from a two-stage analysis and strict decision protocol.