A frozen cross-vendor study of 31,430 trials across 11 GPT, Claude, Gemini & Kimi Large Language Models found 11,658 successful executions with exactly zero visible UTF-8 output bytes.
Across 4,290 strict matched semantic pairs, null-condition arms produced 2,505 Voids; matched output-licensed controls produced 0.
These were not refusals, safety blocks, rate limits, or transport failures.
Raw records, event hashes, verification code, and full analysis are public.
[link] [comments]