Probabilistic enforcement is not enforcement
A model asked to judge whether an action looks acceptable is still guessing. Why the control plane for agentic AI has to be deterministic, and what changes the moment it is.
Short pieces on deterministic authorization, bounded authority, and what governing agentic AI actually requires. Nothing is published yet: we're in stealth, and these are the arguments we're writing up first.
A model asked to judge whether an action looks acceptable is still guessing. Why the control plane for agentic AI has to be deterministic, and what changes the moment it is.
Every incident pattern we've studied traces back to authority that outlived the action it was granted for. The case for single-use, exhaustible warrants over long-lived credentials.
Underwriting requires a provable record of who authorized what, and when. What that record has to look like, and why it can only be produced at the point of execution.
These go up as they're finished. If you'd like to read them early, or you're working on the same problem, get in touch.