eVoiceClaw V3 debunks mixed volume hypothesis

This is a better story for agent builders than for tropical geometry: the system produced a specific conjecture, then the test loop showed it was mostly a proxy for simpler model-size effects. The interesting part is the negative result plus the regime shift, not the original claim.

• –MV appears to change behavior by dimension, which makes the relationship look conditional rather than universal.
• –The d=40 collapse is intriguing, but it needs more runs, controls, and uncertainty estimates before anyone should treat it as a real phenomenon.
• –If mixed volume and parameter count are this close in rank correlation, MV probably adds little practical signal for generalization prediction.
• –For multi-model orchestration systems, the takeaway is that hypothesis generation is cheap; validation discipline is where the value is.