The 30-second version of the whole framework. Run it:
python examples/intent_demo/run_demo.py
Fully offline, no dependencies beyond pytest (already in [dev]), no API key.
The scenario used throughout the docs: “Add export to CSV for admin compliance
reporting.” Without a structured intent, an AI agent fills the gaps with
training-data patterns — a plausible, readable, wrong implementation
(hallucinated_export.py): no admin check, renamed
columns, US-formatted dates.
csv_export_intent.yaml is the intent artifact an
AI would write instead — three constraints, each with a real pytest test in
test_csv_export.py.
run_demo.py runs that real test suite against both implementations:
hallucinated_export.py → 3/3 tests failcorrect_export.py (written against the intent) → 3/3 tests passNothing here is narrated or canned. Every PASS/FAIL in the demo output is a
live pytest subprocess run against the code in this directory.
Because it doesn’t need to be. The mechanism IVD is built on — a testable constraint catching a wrong implementation before a human reviews it — works identically whether the implementation was written by a person, an AI agent, or copy-pasted from Stack Overflow. This demo isolates that mechanism from everything else (LLM calls, MCP tool orchestration, IDE integration) so you can see it work in under a minute.
For the AI-specific half of IVD — compounding judgment from real corrections
across agent runs — see ../judgment_demo/.