---
date: 2026-06-27
phase: 4
headline: CupSpill evals 100% + bimanual Raiden fidex setup
---

## What we did
- **Evaluated CupSpill** — model hit **100%** on the 3-step task (reorient → pick towel → wipe). Real headline-grade dexterous result on YAM.
- **Bimanual Raiden fidex setup** — fiducial / camera-calibration scaffolding for the bimanual rig, getting infra ready ahead of the Jul-1 → mid-Jul bimanual integration window.
- Spent time refining the one-line method explanation (PixelNeRF-style). Working tagline: *"Action prediction as feature selection."*

## Reflection
- CupSpill at 100% on a multi-stage sequential task is a meaningful jump from egg pick-place — much more legible to non-roboticists and a stronger paper figure anchor candidate. Worth resolving whether it replaces or stacks with egg in the figure.
- Starting bimanual fidex setup today is slightly ahead of the Jul-1 plan, but it's the right kind of forward-loading: rig prep before Annie arrives Jun 29 means the hardware time can actually go to hardware, not calibration debugging.

## Next steps
- Lock CupSpill's role in the paper figure (anchor vs additional task).
- Decide if CupSpill needs OOD-pos / OOD-view evals on Monday.
- Annie onboarding Mon Jun 29 → hard mode starts.