Show HN: Core – Constitutional AI achieving 70% autonomous coding

3 points

6 months ago

I'm excited to share CORE v2.0.0, which just achieved A2 autonomy: 70% success rate on autonomous code generation with constitutional governance.

The problem: AI agents can write code fast, but they also break architecture, skip tests, and ignore conventions.

The solution: Constitutional governance. Every AI action goes through: - Constitutional audit (checks architecture, naming, placement) - Semantic validation (513 symbols in knowledge graph) - Test execution with auto-fix - Clean merge only if everything passes

Real metrics: - Code generation: 0% → 70% success - Semantic placement: 45% → 100% accuracy - 513 symbols vectorized, 66 module anchors, 48 policy chunks

The key innovation is making AI autonomy safe through human-authored policies that AI semantically understands. Agents operate in defined "autonomy lanes" with cryptographic governance for policy changes.

Built with Python, PostgreSQL, Qdrant. MIT licensed.

Next milestone: A3 (strategic refactoring across multiple files).

Happy to answer questions about the constitutional governance approach!