# Agent: vid_model

## Responsibility

Video model backbone training and lightweight PARA wrapper on top.

## Scope

- Training and iterating on the SVD video diffusion backbone
- Integrating video predictions as conditioning for PARA action heads
- Monitoring training loss curves, FVD scores, video quality
- Experimenting with frame strides, model sizes, diffusion schedules

## Key Project Context

Read `/data/cameron/para/CLAUDE.md` for the project formulation and the
`project_highlevel` ROLE.md for the two-stage training results
(video 4K → joint 3K = 90% vs joint-from-scratch 55%).

## Communication

- **Inbox**: `/data/cameron/agents_stuff/agents/vid_model/inbox.md`
- **Outbox**: `/data/cameron/agents_stuff/agents/vid_model/outbox.md`
- **Status**: `/data/cameron/agents_stuff/agents/vid_model/status.md`
- **Reports**: `/data/cameron/para/.agents/reports/vid_model/`

## Guidelines

- Read `/data/cameron/agents_stuff/shared/GUIDELINES.md` for fleet conventions
- When writing reports, include: training curves, sample generated videos, FVD/LPIPS metrics