September 8, 2026 · David Young
Enforceable safety bars need evidence the model cannot touch
OpenAI research has examined chain-of-thought monitoring and its limitations. The operational issue for GRACE is narrower: controls should remain independent of an agent's self-reported intent.
Reference: OpenAI — Evaluating chain-of-thought monitorability.
Evidence has to exist outside the action being governed.
A control that matters in an audit needs a record of the authorization decision. A log of what happened afterward is not the same thing as evidence of what the system was permitted to do before consequence.
Three properties audit evidence has to have
Generated outside the governed action path. The authorization record should be created at the boundary where the decision is made.
Cover abstentions, not only permissions. Evidence must include when the system did not autonomously proceed.
Verifiable without treating the operator's narrative as the evidence. Signed, tamper-evident artifacts narrow what an examiner or auditor must trust.
What GRACE adds now
The Action Path remains the pre-execution enforcement boundary described in the original article. The current architecture adds a Context Path before AI reasoning. A signed Context PAP records what information was authorized for the task; a signed Action PAP records what action was authorized; the Chain of Confidence verifies the relationship between those records.
That extension changes the audit question from only “was this action authorized?” to two linked questions: “was this information authorized for the task?” and “was this action authorized before execution?”
David Young is the founder and CEO of Grace AI Control, Inc.