Frank started as an experiment: can you build a coherent AI personality on consumer hardware?
The timeline:
- Base model: Qwen2.5-3B-Instruct (abliterated for reduced refusal)
- Scaffolding: 76,000+ lines of Python — consciousness, personality, memory, embodiment
- Training: 7 iterations of IAPT Training Method over 3 days — Claude Code as adversarial tester
- Result: Identity coherence 9.2/10 across 390+ adversarial turns
The key insight: You don't need a bigger model. You need a smarter scaffold and better training methodology. Frank's personality isn't emergent from scale — it's trained for directly through iterative adversarial testing.
The v6 breakthrough: We discovered that regex-based output parsers were acting as invisible co-authors — creating selection pressure on Frank's language. Removing them immediately improved authenticity. This is something no benchmark would have caught; only interactive adversarial testing revealed it.
Total cost: electricity for 7 training runs at ~3 hours each. Under $50.