# Article example assets

These examples use Python 3 and its standard library. They make no external model calls, send no messages and use no production credentials.

From the article package directory:

```text
python3 assets/calculate-costs.py
python3 assets/approval-demo.py
python3 assets/score-rag.py
python3 assets/test-mcp-fixture.py
```

- **cost-worksheet.csv** contains two explicitly hypothetical scenarios. The calculator assumes equal input/output token counts for every call in each scenario, including extra calls. It does not model caching, tiered billing or a real provider account. Edit the inputs to explore assumptions; reconcile actual usage separately.
- **approval-demo.py** exercises seven persisted local application-state cases. Its reviewer and tenant values stand in for trusted application context. It has no real authentication service or external send. The atomic local SQLite effect does not establish cross-system exactly-once behavior.
- **rag-fixtures.json** contains a fictional corpus and six deliberately constructed outcomes, with expected facts, behavior, source IDs and example citations. **score-rag.py** aggregates the supplied review labels and source sets. It does not grade arbitrary natural-language text or run a model. The demonstration intentionally has three end-to-end failures, one of which exposes a fictional private document; its proposed release gate therefore fails.
- **shared-evaluation-scorecard.csv** is a blank team-trial worksheet. It contains no actual platform scores or measured winners.
- **mcp-fixture.py** is an in-memory restricted JSON-RPC/stdio teaching profile. **test-mcp-fixture.py** starts that fixture, checks eleven invariants and waits for it to exit. This is not a complete MCP implementation, OAuth test, real-client interoperability check or external-integration test. State does not survive server restart.

Keep real data sanitized and use a test destination when adapting these examples. A local teaching-fixture pass does not approve a production release.