Core – Constitutional AI achieving 70% autonomous coding
Core – Constitutional AI achieving 70% autonomous coding
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!
Share cardActual performance
Launch Intel predictions
Analyze your own launch →Correct prediction on native model
Similar products
C2mon-Core
A RISC-V core in Racket
RAMBOOT "Ubuntu Core Bootstrap" 1.0.0-beta1 released
Autonomous Landmines
NuScenes – ImageNet for Autonomous Vehicles
CoreScript – a scripting language for dotnet core
When you try to tackle Primal Core, you're giving yourself a
Korea-Born Self-Evolving AI Core (PPR and TTP and AIDoc) Open Sourced"
Octaspire Core in less than 5 minutes [video]
ParcelKit integrates Core Data with Dropbox Datastore API