Semantic Defects in AI-Generated Code

An assurance framework for high-stakes systems that use AI to generate code.

Read the full site: semanticdefects.foundryside.dev

What this is

AI coding tools are in active use across high-stakes code bases. This body of work identifies a specific class of defect they produce: code that is syntactically correct, passes automated checks, looks right to reviewers, but makes the wrong decision about data that matters in its institutional context.

The project provides:

• A taxonomy of failure modes (ACF) mapped to STRIDE categories
• Assessment tools for CISOs, security assessors, and technical leads
• A practical review guide for code authors working without CI pipelines
• The Wardline as-built specification — a semantic trust-boundary enforcement tool (wardline on PyPI), specified as actually built and honestly annotated with what was designed but not built

Start here

• Governing AI-Generated Code — the entry point for all audiences (~13 pages)
• Practical Guide for Code Authors — hands-on review guidance with worked examples
• Full Discussion Paper — the complete technical analysis (~200 pages)

Status

v0.2.0 — Draft for Comment

Local development

pip install mkdocs mkdocs-material
mkdocs serve        # http://127.0.0.1:8000

Licence

See About This Project for attribution and licence details.