ANIMA WEIMARANER — Wave-10 WARDOG

• Paper: CAWM-Mamba: IR-Visible Fusion and Compound Adverse Weather
• ArXiv: https://arxiv.org/abs/2603.02560
• Upstream repo: https://github.com/Feecuin/CAWM-Mamba
• Focus: UAV/Drone defense perception under adverse weather
• Detector policy: YOLO26 (latest wave baseline)

Status

Stage-0 completed:

1. paper/repo verification,
2. full PRD + execution docs,
3. initial runnable CAWM-inspired code scaffold,
4. synthetic test harness.

Quick Start

# install editable package
python3 -m pip install -e .

# resolve runtime backend
python3 -m anima_weimaraner --backend auto backend

# run smoke inference
python3 -m anima_weimaraner --backend auto smoke --batch-size 1 --image-size 128

# run one synthetic train step
python3 -m anima_weimaraner --backend auto train-step --batch-size 1 --image-size 128

# run parity check contract
python3 -m anima_weimaraner --backend auto parity --image-size 128

Project Layout

project_weimaraner/
├── configs/default.toml
├── docs/
│   ├── REPRO_VERIFICATION.md
│   ├── ARCHITECTURE.md
│   ├── IMPLEMENTATION_ROADMAP.md
│   ├── EXPERIMENT_PLAN.md
│   ├── RISK_REGISTER.md
│   └── INTEGRATION_PLAN.md
├── src/anima_weimaraner/
│   ├── cli.py
│   ├── config.py
│   ├── device.py
│   ├── eval.py
│   ├── pipeline.py
│   ├── train.py
│   ├── yolo26_adapter.py
│   └── models/
├── tests/
├── PRD.md
└── NEXT_STEPS.md

Dataset Discovery

bash scripts/download_data.sh

This script checks mounted shared datasets and reports missing items. It does not auto-download in stage-0.

Notes on Reproducibility

See docs/REPRO_VERIFICATION.md. The upstream reference repo has reproducibility gaps (missing source file imports and hardcoded private paths), so this module uses a clean implementation path aligned to the paper.