# Physical-device validation

The exact compiled artifact in this release was embedded in a fresh validation
application and run on `massif iPhone`, an iPhone 15 Pro with iOS 27.0. The
application used a separate Bundle ID, separate DerivedData, and did not modify
or launch the product host application.

## Gate

1. Verify the compiled asset metadata and source hash before model loading.
2. Load with `CoreAILanguageModel`, eager mode, and fixed-size KV cache.
3. Generate an exact short marker.
4. Read an exact marker from the tail of a 1,861-token prompt.
5. Unload and record process memory.

## Result

| Check | Result |
|---|---:|
| Source `.aimodel` main SHA-256 | `18f1fb8e...7a668a6` |
| Compiled `main-h16p.mlirb` SHA-256 | `192850a8...5933bb` |
| Fresh load | 37.161 s |
| Short marker | `COREAI-2K-OK` |
| Short TTFT / visible decode | 0.971 s / 23.798 tok/s |
| Near-2K input | 1,861 tokens |
| Tail marker | `TARGET-2K-0046` |
| Near-2K TTFT / visible decode | 2.063 s / 19.466 tok/s |
| Peak process RSS | 2,742.2 MiB |
| Resident memory after unload | 206.4 MiB |
| Suite | `COREAI_SUITE_PASS`, 2 / 2 |

The first packaging attempt exposed a root-metadata defect: it still pointed to
the authoring `.aimodel` even though the release contained the compiled
`.aimodelc`. The published directory fixes the root entry to
`qwen3_1_7b_w8_2k_ios.h16p.aimodelc`; the model binary itself is unchanged.

This gate proves that the package loads and generates at the 2K boundary on the
tested device. It is not a product-quality benchmark, long-duration thermal
test, background-execution guarantee, or compatibility claim for other chips.