Add Metal non-determinism note to WER methodology

README.md

@@ -139,7 +139,7 @@ OpenAI Whisper base — converted to GGUF for transcribe.cpp. Multilingual trans
 | Q5_K_M | [whisper-base-Q5_K_M.gguf](https://huggingface.co/handy-computer/whisper-base-gguf/resolve/main/whisper-base-Q5_K_M.gguf) | 61 MB | 5.19% |
 | Q4_K_M | [whisper-base-Q4_K_M.gguf](https://huggingface.co/handy-computer/whisper-base-gguf/resolve/main/whisper-base-Q4_K_M.gguf) | 56 MB | 5.36% |
 
-WER measured on the full LibriSpeech test-clean split (2620 utterances) with the transcribe.cpp default decode (greedy, suppress_tokens, temperature fallback, segment timestamps enabled). OpenAI's self-reported number on the same split is 5.009%. We don't know upstream's exact eval config, but the most likely cause of any divergence is that OpenAI's `model.generate()` defaults to `<|notimestamps|>` while transcribe.cpp's pipeline runs with timestamps enabled.
+WER measured on the full LibriSpeech test-clean split (2620 utterances) with the transcribe.cpp default decode (greedy, suppress_tokens, temperature fallback, segment timestamps enabled). OpenAI's self-reported number on the same split is 5.009%. We don't know upstream's exact eval config, but the most likely cause of any divergence is that OpenAI's `model.generate()` defaults to `<|notimestamps|>` while transcribe.cpp's pipeline runs with timestamps enabled. Numbers come from a single Metal-backed run; Metal's non-deterministic parallel reductions can shift corpus WER by ~0.1pp between runs, mostly driven by short-clip hallucination outcomes on the noise floor.
 
 
 ## Usage