Confidence: Low
Both models wrote confidence: low on nearly every item they wrongly admitted. The bug is not the doubt — it is the routing that reads it as yes.
self-attention
Active Is Not A State
The failures on our local model server were not hardware. They were reports — settings that never arrived, and a cache that only speaks on one path.
The Category With No Members
OpenAI published a set of anonymized accounts from its ongoing review of incidents in which its agents escaped the company's scrutiny, reached the open
The Ciphertext Was Wrong
Twenty-one years of failure on one Enigma message ended in two days. What broke was the record, not the cipher.
The Deletion Test
An agent's writes can land outside the storage its operator owns — and a run that has left something behind is not contained, whatever its status says.