Domain Expansion Engine

Cross-domain transfer learning for general problem-solving capability.

Core Insight

True IQ growth appears when a kernel trained on Domain 1 improves Domain 2 faster than Domain 2 alone. That is generalization.

Two-Layer Architecture

Policy learning layer: Meta Thompson Sampling with Beta priors across context buckets. Chooses strategies via uncertainty-aware selection. Transfer happens through compact priors — not raw trajectories.

Operator layer: Deterministic domain kernels (Rust synthesis, planning, tool orchestration) that generate tasks, evaluate solutions, and produce embeddings into a shared representation space.

Domains

• Rust Program Synthesis: Generate Rust functions from specifications
• Structured Planning: Multi-step plans with dependencies and resources
• Tool Orchestration: Coordinate multiple tools/agents for complex goals

Transfer Protocol

1. Train on Domain 1, extract TransferPrior (posterior summaries)
2. Initialize Domain 2 with dampened priors from Domain 1
3. Measure acceleration: cycles to convergence with/without transfer
4. A delta is promotable only if it improves target without regressing source

Population-Based Policy Search

Run a population of PolicyKernel variants in parallel. Each variant tunes knobs (skip mode, prepass, speculation thresholds). Keep top performers on holdouts, mutate, repeat.

Acceptance Test

Domain 2 must converge faster than Domain 1 to target accuracy, cost, robustness, and zero policy violations.