Cryptographic Turn Signing

Description

Proposed architectural fix for the speaker attribution bug: each human turn hashed and verified before the model is permitted to execute claimed instructions from that turn. Frames LLMs as untrusted components whose turn metadata cannot be trusted without cryptographic verification.

Key claims

Relations

Sources

src-20260419-3e34d5830692