# Public-source UMA AI demonstration

This demonstration starts from a label-free registry of 64 fixed public transactions,
downloads or replays canonical Polygon and Ethereum receipts, decodes the UMA lifecycle,
derives decision-time features and economic outcomes, and runs the registered logistic
regression plus four simple baselines. It never reads a curated ledger.

- Chronological split: 44 train / 10 validation / 10 test.
- Test AUC: 0.375; Brier: 0.414; ECE-5: 0.374.
- Automated Accept/Challenge coverage: 0.400.
- Outcome: later UMA protocol resolution only, not independent factual truth.
- Configured rewards, capital at risk, gross payout and realized challenge payoff remain distinct.

Live public reproduction:

```bash
python scripts/applications/trustworthy_ai/run_public_source_demo.py
```

Deterministic replay after freezing the public receipts:

```bash
python scripts/applications/trustworthy_ai/run_public_source_demo.py --offline
```