The Neural Conscience (services/neural_conscience.py, ~650 lines) is a bio-inspired quality gate for idle thoughts. It replaced hardcoded regex scoring with a 25K-parameter neural network that learns what a "good" thought looks like.
Architecture: 5 Cortical Modules
| Module | Analog | Function |
|---|---|---|
| ACC | Anterior Cingulate | Conflict detection — is this thought contradictory? |
| Insula | Insular Cortex | Gut feeling — does something feel off? |
| vmPFC | Ventromedial PFC | Value assessment — is this thought useful? |
| OFC | Orbitofrontal | Outcome prediction — will this lead somewhere? |
| dlPFC | Dorsolateral PFC | Executive judgment — final quality score |
Training
Multi-task supervised learning from thought outcomes. Trains during consolidation phases with an 80ms budget. SFT-stabilized with anchor points to prevent catastrophic drift.
Cold Start Phases
- Bootstrap (steps 0-49): 100% rule-based fallback
- Blending (50-149): smoothstep mix of neural + rules
- Maturing (150-299): 75% neural
- Mature (300+): 95% neural, 5% safety margin