# E120 Unified Generalization Engine Report (G03–G06)

**Anti-shortcut policy:** No task IDs, filenames, seeds, or world keys passed to learners; content-derived invariants only. | **Matched compute budget:** `1200` candidate evaluations/task.

| Condition | Frozen | No-Memory | No-Learning | No-Abstraction | Full CAIN | Non-CAIN RF | Gain vs No-Learning |
|---|---:|---:|---:|---:|---:|---:|---|
| A_unseen_combinations | 0.375 | 1.000 | 0.375 | 0.375 | **1.000** | 0.000 | +0.625 (Holm p=5.72e-06) |
| B_unseen_primitives | 0.125 | 1.000 | 0.562 | 0.125 | **1.000** | 0.000 | +0.438 (Holm p=0.000122) |
| C_unseen_task_family | 0.000 | 1.000 | 0.000 | 0.000 | **1.000** | 0.000 | +1.000 (Holm p=1.86e-09) |
| D_unfamiliar_2d_domain | 0.000 | 1.000 | 0.500 | 0.000 | **1.000** | 0.000 | +0.500 (Holm p=6.1e-05) |

- **Retention of earlier skills:** 1.000 (30/30, 95% CI [0.8865, 1.0])
- **Negative transfer check:** blind macro transfer 0.000 vs counterexample-gated **1.000** (avoided: True)
- **Limits:** Constructed symbolic/grid domains with disclosed primitive vocabularies and schema templates; not open-world AGI.; Condition C discovers parameters and branch structures within polynomial, threshold, and parity-conditional schema families.; Non-CAIN RandomForest reference is trained on small per-task support (6-8 examples); a frontier LLM was not tested due to blocked API credits.
