The right person authorizes the exact action — not a summary of it
For anything that can't be undone, a specific person signs off on the specific action about to execute — not a description of it, not a model's summary of what it intends to do, but the action itself.
Human-in-the-loop, not human-on-the-loop
There's a meaningful difference between a human who can review what an AI system did after the fact, and a human whose authorization is required before an irreversible action executes. Luminous is built for the second: for any action classified as irreversible, execution blocks until a specific person approves that specific action. The system doesn't proceed on a timeout, a default, or an assumption that silence means yes.
WYSIWYS — what you see is what you sign
WYSIWYS ("what you see is what you sign") is a long-standing principle from cryptographic signing — the guarantee that what a signer is shown is exactly what they're authorizing, with nothing substituted between the display and the execution. Luminous applies the same principle to human approval of AI-agent actions: the approver sees the literal action — not a model-generated summary of what it plans to do, not a paraphrase that could omit or soften something material. The approval is cryptographically bound to that exact action, so a mismatched or manipulated request can't be waved through by mistake, and an approval can't later be pointed at a different action than the one actually shown.
Why the summary is the risk
A model-generated description of "what I'm about to do" is itself a model output — it can be wrong, incomplete, or subtly steered by the same prompt manipulation that produced the proposed action in the first place. If a human is approving a summary rather than the action, the approval step inherits every failure mode of the thing it's supposed to be checking. WYSIWYS closes that gap by binding sign-off to the actual, executable action.