Verified-bridge demo recording script

This CPU recording follows the artifact path with the bridge extra. It runs from local fixtures and produces the same reports without a training setup.

Before recording

python -m pip install "miniverl[bridge]"
python scripts/prepare_verl_bridge_smoke.py --out _demo-source

Keep the terminal at least 110 columns wide. Start the recording after setup so the visible flow fits in 60–90 seconds.

Recording flow

# 0–10 s: show the exact compatibility target
miniverl --version
cat _demo-source/README.md

# 10–35 s: make the transactional standard-artifact bundle
miniverl export-verl --run _demo-source --target-verl v0.8.0 \
  --out _demo-bundle

# 35–65 s: inspect model, tokenizer, Parquet, config, reward, privacy and hashes
miniverl bridge doctor _demo-bundle --json

# 65–80 s: show the pin and the unsupported-semantics boundary
cat _demo-bundle/recipe/REQUIRED_VERL.txt
python -c "import json; p=json.load(open('_demo-bundle/provenance/compatibility-report.json')); print(p['unsupported_semantics']); print('launchable:', p['launchable']); print('distributed tested:', p['distributed_execution_tested'])"

# 80–90 s: show the portable tree
python -c "from pathlib import Path; print('\n'.join(p.as_posix() for p in sorted(Path('_demo-bundle').rglob('*')) if p.is_file()))"

Narration: “Here are the model, data and config files in the bundle. The report checks their structure and hashes against the pinned upstream version. This fixture still needs a reward implementation, so it reports launchable: false. Distributed execution: not tested.”

The release's stronger compatibility evidence additionally installs the exact official verl commit and uses OmegaConf plus PEFT to load the relevant surfaces. Its machine-readable record is generated/verl-bridge-smoke.json.