# External llama.cpp Equivalence

**Verdict:** PASS  
**Implementation:** `real-isolated`  
**Cases:** 48

| Token | Layer | Expert IDs exact | Weight max abs | Output max abs |
|---:|---:|---:|---:|---:|
| 50279 | 0 | true | 0.0 | 0.0 |
| 50279 | 1 | true | 0.0 | 0.0 |
| 50279 | 2 | true | 0.0 | 0.0 |
| 50279 | 3 | true | 0.0 | 0.0 |
| 50279 | 4 | true | 0.0 | 0.0 |
| 50279 | 5 | true | 0.0 | 0.0 |
| 50279 | 6 | true | 0.0 | 0.0 |
| 50279 | 7 | true | 0.0 | 0.0 |
| 50279 | 8 | true | 0.0 | 0.0 |
| 50279 | 9 | true | 0.0 | 0.0 |
| 50279 | 10 | true | 0.0 | 0.0 |
| 50279 | 11 | true | 0.0 | 0.0 |
| 50279 | 12 | true | 0.0 | 0.0 |
| 50279 | 13 | true | 0.0 | 0.0 |
| 50279 | 14 | true | 0.0 | 0.0 |
| 50279 | 15 | true | 0.0 | 0.0 |
| 42 | 0 | true | 0.0 | 0.0 |
| 42 | 1 | true | 0.0 | 0.0 |
| 42 | 2 | true | 0.0 | 0.0 |
| 42 | 3 | true | 0.0 | 0.0 |
| 42 | 4 | true | 0.0 | 0.0 |
| 42 | 5 | true | 0.0 | 0.0 |
| 42 | 6 | true | 0.0 | 0.0 |
| 42 | 7 | true | 0.0 | 0.0 |
| 42 | 8 | true | 0.0 | 0.0 |
| 42 | 9 | true | 0.0 | 0.0 |
| 42 | 10 | true | 0.0 | 0.0 |
| 42 | 11 | true | 0.0 | 0.0 |
| 42 | 12 | true | 0.0 | 0.0 |
| 42 | 13 | true | 0.0 | 0.0 |
| 42 | 14 | true | 0.0 | 0.0 |
| 42 | 15 | true | 0.0 | 0.0 |
| 1000 | 0 | true | 0.0 | 0.0 |
| 1000 | 1 | true | 0.0 | 0.0 |
| 1000 | 2 | true | 0.0 | 0.0 |
| 1000 | 3 | true | 0.0 | 0.0 |
| 1000 | 4 | true | 0.0 | 0.0 |
| 1000 | 5 | true | 0.0 | 0.0 |
| 1000 | 6 | true | 0.0 | 0.0 |
| 1000 | 7 | true | 0.0 | 0.0 |
| 1000 | 8 | true | 0.0 | 0.0 |
| 1000 | 9 | true | 0.0 | 0.0 |
| 1000 | 10 | true | 0.0 | 0.0 |
| 1000 | 11 | true | 0.0 | 0.0 |
| 1000 | 12 | true | 0.0 | 0.0 |
| 1000 | 13 | true | 0.0 | 0.0 |
| 1000 | 14 | true | 0.0 | 0.0 |
| 1000 | 15 | true | 0.0 | 0.0 |

Aggregate weight max abs: 0.0 (limit 1.0e-6).  
Aggregate output max abs: 0.0 (limit 1.0e-5).

**Limit:** This establishes every-layer single-token MoE equivalence for the pinned OLMoE model and llama.cpp build; it is not an end-to-end generated-token proof.