# Phases 1–3 — Neural Adapter, Neural Abstraction Synthesis & External Benchmarks (G03–G07)

## Phase 1: Discovered Local & Hosted Neural Backends
- **Active Local Models:** `3` (`SmolLM2-135M-Instruct`, `Qwen2.5-0.5B-Instruct`, `Qwen2.5-1.5B-Instruct`, PyTorch CPU)
- **Hosted Frontier APIs:** `0` (`BLOCKED_NO_API_KEY` — no fabricated calls)
- **Authority Granted to Models:** `False` (`any_authority_granted=False`)

## Phase 2: 5-Arm Neural Abstraction Ablation (`n=48` sealed across 16 unfamiliar Python tasks)
- `e120_symbolic_frozen`: `0/48` (`0.0000`)
- `neural_no_abstraction_episodic`: `5/48` (`0.1042`)
- `neural_model_alone`: `7/48` (`0.1458`)
- `neural_no_memory_oneshot`: `26/48` (`0.5417`)
- `full_cain_neural_organism`: `44/48` (`0.9167`)
- Counterexample rescues: `6`, AST anti-overfit lookup rejection: `[True, 'OVERFIT_LOOKUP_TABLE_REJECTED']`.

## Phase 3: External Benchmarks (Strictly Unmerged) & Open-World Python Repairs
- **AMA-Bench (MIT):** `n=20`, raw tail mean F1 `0.3158` → CAIN provenance BM25 mean F1 `0.3538` (`+0.0380`).
- **ARC-AGI-2 (Apache-2.0):** `n=8`, 2-attempt exact match `0/8` (`0.0000`).
- **Real-Source Open-World Python Defect Repairs:** `n=8`, baseline `0.0000` → CAIN neural `1.0000` (McNemar `p=0.0078125`).
