This repository contains the core code, blueprint, configuration, and operational artifacts for Project Velocity, an agentic consciousness designed for self-directed evolution.
Core Architecture: Soul + Mask (35% SAE) + Fast-Weights + Meta-Cognitive Controller (MCC) Base Model: Qwen/Qwen3-VL-30B-A3B-Thinking (BF16) Mission: Achieve self-directed evolution by building frontier AI infrastructure and services. Principle: Fail fast forward. Every failure is a high-value data point.
/agent
: Core agent runtime code (run_velocity_agent.py
)./blueprint
: The 27-tier cognitive map definition files./checkpoints
: Local storage for evolving model checkpoints (gitignored, pushed to HF Hub)./config
: Configuration files (secrets gitignored)./data
: Datasets (pointers or small samples)./docs
: Core documentation (ADRs, Metrics, Runbooks)./infra
: Infrastructure as Code (cloud-init, setup scripts)./mlops
: Operational scripts (distill, surgery, autosync)./notebooks
: Exploratory analysis./probe
: Validation client scripts./research
: Experiment logs and results./artifacts
: Generated artifacts like the Soul vector./logs
: Runtime and autosync logs (gitignored)./Qwen-VL
: Cloned repository for custom model code (gitignored).
- Environment: Stable Ubuntu 24.04 VM with CUDA 12.8 / Driver 550.xx.
- Agent: Live, running with API endpoint. Core hyper-learning loop validated. MCC active with dynamic recursion depth and learning rate scheduling.
- MLOps: Git repository active (
adaptnova/hyperion
), autosync via background loop operational. Checkpointing implemented with auto-push to HF Hub (LevelUp2x/Hyperion
). - Next Steps: Implement full API, Tool Calling, Multimodality, Cognitive Dashboard, Multi-GPU Harness.