# PARA — conventions and gotchas

Append when you learn something non-obvious. PARA-specific only;
cross-cutting stuff goes in `/data/cameron/vault/memory.md`.

## Known seeds (populate from existing backbones / mac memory)

- Latest > best for inference checkpoints — val loss is multimodal-noisy on the smith300/UMI family. (See `~/.claude/projects/.../memory/feedback_inference_checkpoint.md`)
- Headline metric for AR + volume work is `train_pix_argmax`, not val.
- Rotation discretization: 1D PCA when PC1 EV ≥ 0.85; K-means prototypes (k=8) otherwise.
- Image normalization happens in the dataset loader (ImageNet mean/std applied to [0,1]) — never re-normalize in the model.
- `--vis_every_steps=0` triggers a zero-div bug; use `100000` to disable.
- DINOv3 ViT-S/16+ at 504-space input is the production backbone (weights at `/data/cameron/keygrip/dinov3/weights/...`).

(Append more as agents encounter them.)