## 16:25 — Phase A FULLY COMPLETE (objpos + viewpoint)

**Phase A viewpoint final (10 ep/cell, 4 viewpoints):**
- 1v query-MLP avg: 88% (100/70/100/80 at (0,0)/(14,45)/(10,180)/(14,225))
- 2v query-MLP avg: 63% — 2v HURTS by ~25pp
- One 2v cell hung (killed after 50min on (14,45))

**Combined Phase A finding (objpos + viewpoint):**
- 1v ≈ 2v for objpos (~85%)
- 1v >> 2v for viewpoint (88 vs 63%)
- 2view projection-fusion doesn't add useful signal once 1v has env variation in training

Reports: `/data/cameron/para/libero/logs/PHASE_A_OBJPOS_FINAL.md` + `PHASE_A_VIEWPOINT_FINAL.md`. Memory updated.

**Open:** Phase B (DA3 latent depth fusion), Phase C (VLM trunk). Both pending Cameron review of A — finding may suggest re-thinking 2view fusion before committing to multi-modal experiments.