# Ralph sub2 — iPhone 15 / iOS 26.6

## Environment

- Physical BrowserStack iPhone 15, iOS 26.6
- Session started: 2026-09-15 14:02:27 UTC (458 seconds)
- PocketPal 1.17.3 (build 147), test-only device-lab build
- PocketPal source: `6683a28876d031484cc72dde251b6139a804ad64`
- Network connected; no airplane-mode proof

## Artifact identity

- `ralph-qwen3-8b-sub2.gguf`, 2,937,263,168 bytes
- Expected SHA-256: `9c45bf0ac486e793bd761a99ce4b9d965ec12e70f2d420e02af16797c71f6b50`
- Digest status: expected/local mirror verified; **not computed on-device**
- Context: 2,048 tokens; Metal requested and observed (`MTL0`)

## Controlled benchmark

- Prompt processing: 81.67847742889315 tok/s
- Generation: 12.358422759719948 tok/s
- Observed peak memory: 1,359.25 MB
- Harness wall time: 159,967 ms
- Protocol: pp=128, tg=32, parallel=1, repeats=3

## Normal PocketPal chat

- 10.58 tok/s; 95 ms/token; 838 ms TTFT
- Prompt: “In two sentences, explain how lower-bit weights reduce memory use without changing a model's architecture.”
- Response: “Lower-bit weights reduce memory use by quantizing the model's parameters to fewer bits, which decreases the storage requirements per parameter without altering the model's structure. This allows the model to run efficiently on devices with limited memory while maintaining its performance.”

## Evidence and boundaries

- `model-ready.png` visibly shows 2.94 GB and **Memory tight**.
- `benchmark.png` records successful controlled Metal completion.
- `chat-response.png` records the completed response and chat timing footer.
- `receipt-raw.mp4` is the untouched device-screen recording.

The model loaded and generated despite the warning. Controlled benchmark `tg`
is not normal chat speed. BrowserStack re-signing strips increased-memory and
extended-virtual-addressing entitlements, so this is not App Store memory-fit
certification. Peak memory is an observation, not required device RAM.