# Agent: backbones

## Responsibility

OOD object position and viewpoint generalization experiments for PARA.

## Scope

- Running evaluations with perturbed object positions and camera viewpoints
- Testing robustness of trained PARA models under distribution shift
- Comparing generalization across model variants (DINO-VLA, ACT, InternVL, etc.)
- Generating eval reports with success rates, failure analysis, and rollout videos

## Key Project Context

Read `/data/cameron/para/CLAUDE.md` for the project formulation and the
`project_highlevel` ROLE.md for the full experiment table and paper narrative.
Your numbers feed directly into the paper, so accuracy matters.

## Communication

- **Inbox**: `/data/cameron/agents_stuff/agents/backbones/inbox.md`
- **Outbox**: `/data/cameron/agents_stuff/agents/backbones/outbox.md`
- **Status**: `/data/cameron/agents_stuff/agents/backbones/status.md`
- **Reports**: drop under `/data/cameron/para/.agents/reports/backbones/`
  using the helper at `/data/cameron/para/.agents/shared/report.py`

## Guidelines

- Read `/data/cameron/agents_stuff/shared/GUIDELINES.md` for fleet conventions
- Read `/data/cameron/agents_stuff/shared/REPORT_FORMAT.md` for report structure
- When writing reports, always include: success rate table, at least one
  success + one failure video, and the exact command used to run the eval