# DEVIATION_1 — initial-noise generator device
Recorded: 2026-10-27 05:16 MDT (America/Denver), after the amendment-3 run. Spotted by Abe while assembling the pitch packet.

- **Plan (FALSIFIER.md line 35):** the initial latent noise uses `torch.Generator` on CUDA.
- **What the code actually did:** `harness/common.py` lines 48–49 and 58 use `torch.Generator("cpu").manual_seed(seed)` with `torch.randn((1,4,64,64), float32)`. `harness/calib.py` line 71 also uses a CPU generator, for the planted-fake direction.
- **Scope:** every run (calibration, test, amendments 2 and 3) used the CPU generator consistently, so baseline and steered images share identical noise per seed and the paired design is intact.
- **Effect on the verdict:** none expected. The CPU and CUDA generators produce different noise for the same seed, so exact images can't be reproduced with a CUDA generator. They can be reproduced with the code as written. Seeds 0–4 are unchanged.
- **Recorded before anything was shared outside.** Not a pre-registered choice; this is a disclosed departure from the plan.