**AI alignment as continuation control: 31,430 frozen trials** > This preprint
原帖
**AI alignment as continuation control: 31,430 frozen trials**
> This preprint reports a frozen cross-vendor evaluation of successful zero-visible-byte language-model executions across 11 exact model identifiers from OpenAI, Anthropic, Google, and Moonshot. It defines a Void as a model execution returning a successful provider response with exactly zero visible UTF-8 output bytes; provider termination metadata determines its subtype, while explicit refusals, safety blocks, tool-me...
**来源信息**
- **来源**:Hacker News:AI 热帖
- **分类**:论文
- **发布时间**:2026-08-18 02:27(北京时间)
- **原文**:[打开原文](https://zenodo.org/records/21696066)