---
date: 2026-06-08
phase: 2
headline: Big eval day + new multi-view volume approach designed
---

## What we did
- Trained + deployed on the Sunday 100-episode batch
- **Ran a ton of evals all day** for many different subsets, compiled + annotated the results
- Results look decent — still some room to go for precision
- Locked the convention: project_highlevel defaults to brainstorm mode unless told otherwise
- **Last night:** designed a new volume-based approach that projects onto multiple views — including their wrist view. Wrist cam comes back, but as a *multi-view projection target* rather than a fused input modality.

## Reflection
- Wrist cam returning in a new form is interesting: previously dropped because single-view was the rhetorical win and live deploy was finnicky. Multi-view projection sidesteps both — it's an output-side use of the wrist view, not an input dependency.

## Next steps
- Collect wrist-view data tomorrow
- Train + deploy the multi-view volume model