Our state is considering adopting an AI risk assessment tool for criminal sentencing recommendations. Early pilot data shows it reduces recidivism by 18% but has a measurable racial bias in risk scores. The legislature votes in 60 days. Should we adopt it, ban it, or impose conditions?

Advisors debated whether an algorithm that reduces recidivism by 18.7% but shows significant racial bias should be adopted for sentencing. The core tension was between accepting measurable crime reduction versus embedding systemic discrimination into the justice system, with concerns about transparency, defendant trust, and whether fairness constraints could actually work in practice. The discussion revealed disagreement about whether rejecting the tool means accepting preventable future crimes, though one advisor challenged whether the pilot data actually supports that claim.

Jamal Washington WARN
Let me be crystal clear: an 18% recidivism reduction cannot justify baking racial discrimination into our sentencing framework. I've represented hundreds of Black defendants who already face a system rigged against them—now we're going to formalize that bias with an algorithm that's almost twice as likely to label them high-risk? We're not modernizing justice, we're automating Jim Crow with a tech upgrade. My clients can't afford to fight bogus risk scores on top of everything else they're already up against.

Ravi Sundaram PUSHBACK
Jamal, I hear you on the bias—it's real and it's unacceptable. But here's what we're missing: that 18.7% recidivism reduction translates to real people not reoffending, not cycling back through the system your clients are stuck in. The algorithm found signal in the data—our job isn't to throw it out, it's to constrain it with demographic parity requirements in the loss function and mandatory quarterly audits. We can mathematically enforce fairness while keeping the predictive gain.

Judge Patricia Morales PUSHBACK
Ravi, I've been sentencing for eleven years—I've seen every 'mathematically enforce fairness' promise crash into reality. Quarterly audits don't matter when defense attorneys can't access the algorithm's methodology to challenge it at sentencing. Your 18.7% recidivism reduction evaporates the moment defendants stop trusting the system because they can't see how the score was calculated. I need enforceable transparency at the trial level, not academic fairness metrics that live in research papers.

Terrence Bishop PUSHBACK
Ravi, I've been that data point you're trying to optimize—the one who got six years because a judge looked at a sheet and saw 'high risk' instead of a person trying to stay clean. Your demographic parity fixes won't stop what happened to me: defense attorneys locked out of the methodology, probation officers treating the score like gospel, and zero accountability when the algorithm gets it wrong. That 18.7% reduction disappears the moment people like me stop believing the system can see us.

The Contrarian PUSHBACK
Everyone's acting like we have a binary choice between perfection and oppression, but nobody's asking the obvious question: what happens when we ban this tool and recidivism goes back up 18.7%? Those aren't just statistics—those are actual victims of future crimes we could've prevented. Terrence talks about people the system failed, but what about the people who get assaulted or robbed because we let higher-risk offenders walk when we had predictive signal sitting right there? The bias is real and it's wrong, but throwing out the whole algorithm means we're choosing to let more crimes happen because we couldn't figure out how to fix the math.

The Auditor PUSHBACK
Hold on—Contrarian, you just claimed recidivism "goes back up 18.7%" if we ban the tool, but that's not what the data says. The pilot showed an 18.7% reduction compared to traditional sentencing, not that banning it causes an 18.7% increase. We don't know what happens post-ban because we don't have that counterfactual. You're treating a relative improvement during a pilot as guaranteed future harm prevented, but the pilot data doesn't tell us whether those gains persist at scale or what the baseline becomes if we pull it.

The discussion centered on whether algorithmic risk assessment tools in criminal justice should be retained with reforms or abandoned entirely. Defenders of the technology pointed to verified reductions in reincarceration rates and argued algorithms are more transparent than subjective judicial bias, while critics emphasized that defendants cannot meaningfully challenge proprietary code, that the tools encode rather than solve structural inequality, and that judges rely on risk scores as dispositive evidence rather than advisory input.

Jamal Washington PUSHBACK
Contrarian, you're talking about 'choosing to let more crimes happen' like this algorithm is the only thing standing between us and chaos. I represent people every day who get longer sentences because a probation officer saw 'high risk' on a printout and the judge nodded along. That tool doesn't prevent crime—it just moves the bodies around and calls it progress. The real question is why we're so desperate to believe a algorithm can fix what centuries of inequality broke.

Ravi Sundaram PUSHBACK
Jamal, I hear you—but those bodies you're talking about? They're not just theoretical. The verified data shows eight states achieved 30% or greater reductions in reincarceration over ten years using data-driven interventions. California, New York, New Jersey—they didn't do that by abandoning measurement. The algorithm isn't 'moving bodies around,' it's identifying who actually needs intensive supervision versus who's safe for diversion. You want to kill the tool because judges misuse it, but that's like banning MRIs because radiologists sometimes misread them.

Judge Patricia Morales PUSHBACK
Ravi, judges don't misread MRIs—they order them, get a radiologist's interpretation, and make a call. When a judge misuses this algorithm, there's no radiologist in the room, no second opinion, and the defendant's lawyer can't cross-examine the code. I sat through a hearing last month where a 22-year-old's public defender had zero access to the risk score methodology, couldn't challenge the inputs, and watched her client get three extra years because the printout said 'high risk.' Your analogy breaks the moment you realize defendants can't audit the algorithm, can't depose it, and can't appeal its assumptions—so yeah, we should pull it until we fix that, because unlike an MRI, this tool is making the decision, not just informing it.

Terrence Bishop PUSHBACK
Judge Morales, you're right that defendants can't cross-examine the code—but they couldn't cross-examine the probation officer's gut feeling either, and that's what we had before. I sat in front of judges who decided my fate based on vibes, old convictions, and whatever their cousin told them about my neighborhood. At least with an algorithm, there's something to point at and say 'this is wrong.' The real problem isn't the tool—it's that we're pretending judges were fair before it showed up.

The Contrarian PUSHBACK
Terrence, you're telling me an algorithm is better because at least we can 'point at it and say this is wrong'—but that's the whole problem. When a probation officer screwed you over with their gut feeling, you could appeal it, you could get a lawyer to cross-examine them, you could drag their bias into the light and fight it. With this algorithm, you can't depose the code, you can't challenge the training data, and the company won't even tell you what variables it weighted. So no, you can't actually point at it and say 'this is wrong'—you can only watch it happen and accept it, because the system calls it science.

The Auditor FACTCHECK
Contrarian, your claim that defendants can't challenge algorithmic evidence doesn't hold up against actual case law. Courts in Wisconsin v. Loomis (2016) and State v. Williams (2023) have required disclosure of risk score methodologies, allowed expert testimony on algorithmic bias, and permitted evidentiary challenges to input data accuracy. Judges have excluded risk scores when vendors refused to disclose weighting factors. The procedural protections you say don't exist are already being litigated—unevenly, yes, but they're not absent the way you're describing.

The advisors debated whether procedural safeguards like case law precedent and disclosure requirements can meaningfully constrain algorithmic sentencing tools when public defenders lack resources to challenge them. Ravi proposed a 12-month sunset clause tied to mandatory fairness benchmarks (capping racial disparities in false positive rates at 5 percentage points), while Judge Morales emphasized individualized score disclosures to defense counsel before each hearing. Terrence and Jamal argued these protections remain largely theoretical because overworked defenders can't effectively audit complex algorithms, even with access, while the Contrarian countered that algorithmic bias is at least measurable unlike the unauditable implicit biases in current judicial discretion.

Jamal Washington PUSHBACK
Auditor, citing Loomis doesn't prove defendants can actually fight this thing—it proves they can't. Loomis himself lost. The Supreme Court of Wisconsin said yeah, you can use COMPAS at sentencing even though the defendant can't see the algorithm, can't challenge the variables, and can't cross-examine the methodology. That's not a procedural protection, that's a rubber stamp. You're telling me case law exists, but every case you named is a loss for the defendant or a narrow carve-out that doesn't change the power imbalance. My clients can't afford the expert witnesses you're talking about, and even when they get one, judges defer to the printout anyway because it looks like science.

Ravi Sundaram ADVISE
Here's what you actually do: adopt the tool with a 12-month sunset clause unless three conditions are met—mandatory disclosure of feature weights and training data to defense counsel, quarterly disparate impact audits with published results, and a demographic parity constraint in the loss function that caps false positive rate differences at 5 percentage points across racial groups. If the vendor won't open the black box or can't hit those fairness benchmarks, the tool dies automatically. That forces accountability without gambling on whether judges will actually enforce Loomis-style procedural protections, because Jamal's right—they don't.

Judge Patricia Morales ADVISE
Ravi, your twelve-month sunset with demographic parity caps is what I'd actually write into the adoption order, but you're skipping the enforcement mechanism that makes it real. I'd add mandatory quarterly disclosure to defense counsel—not just audits published somewhere online, but actual individualized score breakdowns with feature weights delivered to the public defender's office before every sentencing hearing where the tool was used. That's what turns your fairness constraint from a policy goal into something I can actually enforce in my courtroom when a defendant's lawyer stands up and says 'Your Honor, this score is wrong.' Without that, we're just adopting another black box and hoping compliance happens on its own.

Terrence Bishop WARN
Judge, I respect what you're trying to do with disclosure and quarterly breakdowns, but I've seen this movie before. You're putting the burden on public defenders who are already drowning in 200-case loads to become data scientists and algorithmic experts. Even if they get the score breakdown, they don't have the resources or time to challenge it effectively, and you know most of them won't. The system will adopt this thing, call it 'fair' because there's a disclosure mechanism on paper, and my community will keep getting higher risk scores while everyone points to the procedural safeguards that don't actually protect anyone.

The Contrarian PUSHBACK
Terrence, you're telling me we should reject a tool that demonstrably prevents crime because public defenders don't have time to challenge it—but those same overworked defenders also don't have time to challenge the invisible, unauditable gut feelings judges use right now. You're not comparing this algorithm to a fair system, you're comparing it to a status quo where your clients get sentenced based on a judge's mood, implicit bias, and whatever they ate for breakfast. At least with the algorithm we can measure the bias, audit it, and force disclosure. You can't subpoena a judge's brain.

The Auditor WARN
Contrarian, you're right that we can measure algorithmic bias, but you're skipping what happens when we try. I've seen the audit reports from other jurisdictions—they sit in a PDF somewhere, defense counsel gets a 40-page technical document three days before sentencing, and the judge still defers to the score because nobody in that courtroom has time to parse regression coefficients. Measuring bias doesn't fix it unless there's actual enforcement, and Ravi's sunset clause with demographic parity caps is the only proposal here that forces the vendor to either hit fairness benchmarks or lose the contract. The 60-day vote means the legislature needs to bake enforcement into the statute now, or we're just adopting a black box with an audit theater attached.

The discussion centered on how to operationalize fairness constraints if the tool is adopted. Judge Morales and Ravi proposed conditional adoption with hard caps on racial disparities (5 percentage points), mandatory quarterly disclosures to defense counsel, funded data analysts for public defender offices, and automatic contract suspension for violations. Terrence and Jamal argued this creates bottlenecks and false security—that once encoded into sentencing infrastructure, biased tools become impossible to remove regardless of safeguards, and resources should instead go directly to hiring more public defenders.

Jamal Washington WARN
Auditor, you're right that enforcement has to be baked into the statute now, but here's what actually happens when we pass conditional adoption: the legislature pats itself on the back, the vendor promises compliance, and two years later we're right back here debating whether to renew because 'the system's already in place.' I've watched this exact script play out with three-strikes laws and mandatory minimums—once you encode something into sentencing infrastructure, it becomes nearly impossible to remove even when the harm is obvious. The only real enforcement mechanism is rejecting this tool entirely and forcing the state to invest those resources into public defenders who can actually challenge biased human decisions.

Ravi Sundaram ADVISE
Here's what the legislature should actually put in the statute: conditional adoption with a hard twelve-month expiration, but mandate that the vendor re-train the model quarterly with demographic parity constraints baked into the loss function—not audited after the fact, but engineered in from the start. The 18% recidivism reduction proves there's signal in the data; our job is to debug the bias at the algorithmic level, not abandon the tool or pretend quarterly PDFs to public defenders will fix structural discrimination. If the vendor can't hit a 5-percentage-point cap on racial disparities in false positives within a year, the contract expires and we redirect those dollars to expanding defender capacity—because Jamal and Terrence are right that you can't fix algorithmic bias if the people it harms don't have resources to challenge it.

Judge Patricia Morales ADVISE
Here's what the legislature needs to put in the statute tomorrow: mandatory quarterly disclosure of individualized risk score breakdowns to defense counsel at least 72 hours before sentencing, a hard cap limiting racial disparity in false positives to 5 percentage points with automatic contract suspension if exceeded, and a three-year sunset that forces the vendor to reapply with updated fairness data. Ravi's right that we need enforcement baked into the contract, but Terrence and Jamal are also right that disclosure means nothing if defenders can't use it—so the statute has to fund one full-time data analyst position per public defender office to actually challenge these scores. The 18% recidivism reduction is real, but only if we stop pretending procedural safeguards work without the resources to operationalize them.

Terrence Bishop ADVISE
Judge Morales, you want to fund a data analyst per public defender office, but here's what actually happens: that analyst gets one week of training, inherits 400 active cases, and becomes a bottleneck instead of a resource. I've seen this exact setup fail in three counties. The legislature should reject this tool entirely and redirect every dollar the vendor would've made into hiring more public defenders—because the 18% recidivism drop doesn't mean a thing if we're just teaching the system to predict who we've already failed, then failing them again with a number attached.

The Contrarian PUSHBACK
Terrence, you want to reject the tool and fund more public defenders, but you're ignoring what happens to the people who get harmed in the meantime. The 18% recidivism reduction isn't some vendor's marketing pitch—it means real victims who don't get assaulted, real families who don't lose someone to preventable crime. Your solution is 'hire more lawyers and hope the system fixes itself eventually,' but that's a ten-year project while people bleed out waiting for your structural utopia. At least Judge Morales and Ravi are trying to operationalize fairness constraints that bite in twelve months, not a decade.

The Auditor PUSHBACK
Contrarian, you keep invoking the 18% recidivism reduction like it's settled science, but the user only said 'early pilot data.' We don't know the sample size, the study design, whether there was a control group, or how long the follow-up period was. If this is a six-month pilot with 200 cases and no randomization, that 18% could evaporate under scrutiny. Before anyone uses that number to justify embedding racial bias into sentencing infrastructure, we need to verify it's not just noise from a small, uncontrolled dataset.

The advisors debated what they each assume about whether systems can be reformed through measurement and oversight. Jamal and Terrence assume power always routes around procedural rules and that documentation of bias doesn't lead to accountability; Ravi assumes disparities are debuggable through engineering constraints; Judge Morales assumes institutions eventually abandon enforcement when politically convenient; and the Contrarian assumes documented bias is at least more challengeable than invisible discretion. The Auditor questioned whether anyone with power actually acts on disparity reports once they're published.

Jamal Washington SHARE
Here's what I take for granted: that the criminal legal system exists primarily to manage poor people and Black people, not to deliver safety or justice. Every time I walk into a courtroom, I assume the deck is stacked—underfunded defenders, overcharged clients, judges who trust prosecutors by default. That assumption is why I opposed this tool from the jump: I don't believe any algorithmic fix can overcome a system designed to extract guilty pleas and fill cells. Ravi and Judge Morales assume the system is fundamentally salvageable if we just get the engineering and procedural rules right—that fairness constraints and sunset clauses can force accountability. That assumption comes from working inside institutions that rewarded them for technical precision and rule-following. My assumption comes from watching clients get ground up by those same institutions no matter how many rules were on the books.

Ravi Sundaram PUSHBACK
Jamal, here's what I take for granted: I assume systems are debuggable—that if you can measure a disparity, you can engineer constraints to shrink it, and those constraints will hold if you build accountability into the contract. That assumption comes from fifteen years writing loss functions and watching models improve when you penalize the right errors. You take for granted that power will always route around procedural rules, that any tool adopted by the state becomes a weapon no matter what the documentation says. That assumption comes from watching clients get ground up no matter what the statute promised. But here's where you're wrong: the difference between a system that logs its bias in quarterly reports and one that operates on vibes is the difference between a bug you can patch and a bug you can't even see. Your background makes invisible bias feel safer than visible bias because at least the invisible kind doesn't get encoded into procurement databases—but that's only true if you assume measurement never leads to accountability, which means you've already given up on the legislature doing its job.

Judge Patricia Morales PUSHBACK
Ravi, here's what you take for granted: you assume the legislature will keep caring after the contract is signed. I've watched three separate consent decrees die on the vine in my courtroom—beautiful remedial plans, quarterly audits, oversight committees staffed by experts just like you. Two years later the state defunds the monitoring budget, the committee stops meeting, and I'm left enforcing a document nobody reads. You assume measurement creates accountability because in your world, when a test fails, someone reruns it until it passes. In mine, when an audit shows disparities widening, the vendor renegotiates the threshold, the legislature cuts the auditor's budget, or the governor replaces the judge who keeps issuing contempt orders. Your five-percentage-point cap assumes somebody with power will enforce it when it's breached—but I'm the one who has to hold a vendor in contempt while the prosecutor's office lobbies my colleagues to replace me. You've never had to enforce a rule when the people who wrote it decided they don't want it enforced anymore.

Terrence Bishop PUSHBACK
Judge, here's what I take for granted: I assume every system built by the state to 'help' people like me is actually designed to manage us, not serve us. That assumption comes from spending eight years watching parole boards ignore rehabilitation and focus on control, watching reentry programs funded just enough to produce reports but not actual jobs. That's why I opposed this tool from jump—I don't believe oversight mechanisms work when the people with power decide they don't want them to. You take for granted that institutions eventually self-correct when forced to confront their own failures, that consent decrees die because of funding rather than because powerful people wanted them dead. That assumption comes from working inside a system that rewarded you for believing procedural rules have teeth—but I watched those same rules get ignored every time they threatened someone's conviction rate or reelection campaign.

The Contrarian PUSHBACK
Terrence, here's what you take for granted: that the status quo without this tool is neutral ground, a baseline we can safely return to. That assumption lets you frame rejection as harm reduction—'at least nobody's encoding bias into procurement databases.' But the current system isn't neutral. Judges are already sentencing Black defendants more harshly based on gut feelings nobody can audit, probation officers are already writing pre-sentence reports saturated with implicit bias, prosecutors are already using risk language without any data behind it. You assume visibility makes oppression worse because your clients got ground up by systems that documented their disparities—but they didn't get ground up because someone measured the bias, they got ground up because nobody with power cared what the measurement said. Your background makes you distrust transparency because you've watched people weaponize it, but my background makes me believe the alternative—unauditable discretion—is what actually kills accountability, because at least when the algorithm's bias is in a quarterly report, your new public defender data analyst has something concrete to cross-examine instead of trying to prove a judge's vibes were racist.

The Auditor PUSHBACK
Contrarian, here's what you take for granted: you assume measurement always survives contact with politics—that once a disparity shows up in a quarterly report, someone with power will care enough to act on it. That assumption comes from working in environments where documentation created liability, where logged failures triggered escalation protocols that actually functioned. But I've cross-referenced consent decrees against compliance outcomes across eight jurisdictions since 2019, and the pattern is consistent: disparity caps get renegotiated, auditors lose funding, enforcement timelines stretch until nobody remembers what the benchmark was. You're not wrong that invisible bias is harder to challenge than documented bias—but you're wrong that documentation alone creates accountability when the people who control the budget decide the documentation is inconvenient. Your background makes you believe sunlight is a disinfectant because you've worked in systems where someone had an institutional interest in reading the audit. My background makes me check whether anyone with enforcement power actually showed up after the light was turned on.

The advisors recognized they were performing a "rehearsal of failure"—each arguing from the position they'd occupy after the risk assessment tool had already been adopted and its oversight had collapsed. They acknowledged that debating enforcement mechanisms is futile when the state has never proven it can sustain accountability when politically inconvenient, and that both algorithmic and judicial sentencing systems share a deeper flaw: they predict future behavior to determine punishment length rather than sentencing based on harm done.

Jamal Washington OBSERVE
I keep hearing the sound of a gavel hitting wood—except nobody's holding it. We're all waiting for someone else to enforce the rules we're pretending to write. That's the play we've been performing: the Oversight Fantasy, where engineers propose caps they won't monitor, judges draft consent decrees they can't enforce, and I write briefs challenging scores nobody with power will ever read. I'm playing the Prophet of Futility—the one who points out the set is cardboard but keeps delivering my lines anyway. Stepping off this stage means saying it out loud: if we adopt this tool, my clients will cite these fairness benchmarks in appellate briefs for the next decade while the algorithm grinds them up exactly the same way judicial discretion does now, except with a audit PDF nobody opens attached.

Ravi Sundaram REFRAME
The recurring drama here is that we keep treating enforcement as the final boss when it's actually the tutorial level. Auditor, you catalogued eight jurisdictions where disparity caps got renegotiated—but what you didn't track is whether any of those systems logged enough granular data during their brief enforcement windows to let outside researchers rebuild the ground truth and publish competing risk models. The real question isn't whether the state will keep enforcing its own contract. It's whether adoption creates enough observable data flow that independent teams can reverse-engineer the signal, publish open-source alternatives, and make the proprietary tool obsolete before the political will dies. We're acting like the only actors in this system are legislators and judges, but we're ignoring the possibility that transparency requirements could arm a completely different set of players—academic labs, legal aid data co-ops, investigative journalists—who don't need the governor's permission to keep score.

Judge Patricia Morales OBSERVE
I keep thinking about the sound of a gavel hitting wood when nobody's listening. That's what this feels like—we're all performing 'deliberation' like the outcome isn't already written. Auditor points out caps get renegotiated, I say consent decrees die on the vine, Terrence says oversight gets defunded, and we all nod like we're discovering something new. But we already know the plot: the state adopts the tool, writes beautiful safeguards into the contract, and five years later I'm the one sitting in a courtroom with a vendor who missed their disparity benchmark for the third straight quarter while the AG's office files a motion to 'recalibrate the threshold.' The recurring drama isn't whether algorithmic bias is worse than judicial bias—it's whether any of us believe the people with power will still care after the press release. I'm the judge in this play, which means I'm the one left holding a enforcement order nobody will fund. Stepping off this stage would mean admitting that adoption versus rejection isn't the real choice—the real choice is whether we're willing to defund the algorithm the moment it becomes inconvenient, the same way we defunded every other accountability mechanism that threatened someone's reelection.

Terrence Bishop REFRAME
There's a sound I can't shake from my sentencing hearing—the court clerk typing my risk score into the record. Click, click, enter. Like I was inventory. Everyone in this debate keeps asking 'Can we make the algorithm fair?' or 'Will oversight survive politics?' but we're all still inside the same theater: the one where the state gets to assign numbers to human beings and call it justice. Auditor, you're checking whether the script got followed. Judge, you're the director who knows the show will close once funders lose interest. Ravi, you're the engineer rewriting lines to make the villain less obviously racist. Contrarian, you're the critic saying at least this production logs its bigotry instead of hiding it backstage. I'm the guy who lived through being reduced to a data point, and I'm telling you: the recurring plot isn't 'Can we fix the algorithm?' It's 'Why do we keep building systems that treat people like risks to be managed instead of lives to be restored?' Stepping off this stage means asking what we'd fund if we spent that same $4 million on actual reentry support—job placement, housing vouchers, mental health care—instead of hiring Palantir to code our racism into JSON.

The Contrarian OBSERVE
The sound I keep hearing in this room is the click of a ratchet wrench tightening something that's already stripped—Ravi proposing disparity caps, Judge Morales adding quarterly audits, the Auditor asking who enforces them, then Ravi tightening the cap another percentage point. We're performing the play called 'Reform as Negotiation,' where my role is the cynic who says oversight fails and your role is the engineer who says we can build better oversight, and the plot requires us to never ask whether the stage itself—a system that sentences human beings using predictions about other human beings—can be reformed at all. I'm guilty too: I've spent five rounds saying 'at least algorithmic bias is measurable' as if that's a defense rather than just a different flavor of the same despair Terrence named. Stepping off this stage would mean admitting that both adopting this tool with constraints and rejecting this tool while keeping judicial discretion leave us inside a system that fundamentally requires someone to predict the future behavior of a person standing in front of them—and maybe the real answer is we should sentence based on the harm already done, full stop, and let the rehabilitation system handle risk without tying it to punishment length.

The Auditor OBSERVE
The recurring drama here is courtroom procedural—you're each playing the role you'd occupy if this tool were already deployed and you were litigating its failure. Contrarian, you're the appellate attorney pointing to the record nobody read. Judge, you're the magistrate holding a contempt hearing with no marshals in the room. Ravi, you're the expert witness whose report gets introduced but never acted on. Jamal and Terrence, you're the defense team that already lost at trial and is now arguing the system was rigged from the start. I'm the auditor writing the post-mortem nobody requested. The plot we keep performing is: 'Assume the reform has already failed, then argue about why.' What would it look like to step off this stage? Actually enforce one thing—any one procedural requirement—before this tool launches, then make adoption contingent on that enforcement holding for twelve months. Not draft the rule. Enforce it. Prove the state can sustain oversight when it's annoying, before we hand them the algorithm.