Structural Signals

Executive Summary

Every legislative session, state lawmakers across the country pass hundreds of K-12 education bills. Most handle the ordinary business of running school systems: nutrition, facilities, transportation, teacher pay, and countless other topics. “Structural Signals” is not about those pieces of legislation. Instead, this series examines a narrower question: Where does 2026 legislative activity suggest longer-term shifts in how education systems are organized?

To answer it, Bellwether scanned K-12 legislation across all 50 states and the District of Columbia, focusing on three structural policy areas that highlight emerging trends where the impact on students, providers, and system leaders will likely grow over time:

• Funding Portability and Provider Expansion: Letting public education dollars follow students across settings, making it easier for new kinds of schools to open, and expanding choice options for students and families. 
• Credentialing and Postsecondary Pathway Flexibility: Redefining what counts as learning and widening the routes students can take through school.
• Artificial Intelligence (AI) Governance and System Modernization: Setting the rules for how schools use AI and student data. 

This analysis includes every 2026 state bill that passed at least one legislative chamber, not only those enacted into law. That broader scope is intentional: A bill that cleared one chamber but stalled before enactment still reflects real momentum and shows where policymakers are directing their attention (Appendix).

FUNDING PORTABILITY AND PROVIDER EXPANSION

The Maturing School Choice Landscape

School choice is growing up. Across roughly 130 bills, states spent less energy launching new programs and more energy making existing ones more durable.

CREDENTIALING AND POSTSECONDARY PATHWAY FLEXIBILITY

Rewriting the Rules on Learning and Credentials

Across roughly 200 bills, states loosened the old default — put in the seat time, earn a district diploma — in several ways.

AI GOVERNANCE AND SYSTEM MODERNIZATION

Where Schools Draw the Line on AI

Across roughly 60 bills, the question shifted from whether AI belongs in schools to which decisions schools should refuse to hand off to it.

Key Takeaways

This year was about maturing programs, not launching them. Across all three policy areas, most of the legislative energy went into managing, governing, and building the systems underneath K-12 programs rather than launching new ones. States rebuilt school choice payment systems and provider rules; invested in the transcript standards and transfer agreements that support multiple pathways; and moved from asking whether AI belongs in schools to drawing guardrails around it. In each case, the focus shifted from whether a program exists to whether the systems supporting it can hold up as it grows.

Expansion is outrunning the systems meant to support it. In every area, it is easier to widen access than to build the systems that make access pay off. Choice programs scale before their payment systems are ready; students gain new pathways before transcripts and transfer rules catch up; AI enters classrooms before anyone has decided which decisions it can touch. The recurring risk is that access keeps expanding while the systems needed to support it lag behind.

Accountability, accountability, accountability. Each brief returns to oversight and responsibility. When and where should a school choice provider answer for itself? Who owns a credential that an employer helped design, and who answers to the student if the job never materializes? Who is accountable for a consequential decision about a student that a machine helped make? As the boundaries that were used to assign responsibility erode (district enrollment as the gateway to resources, the line between high school and college, and the line between human and automated judgment), the question of who is answerable becomes the one worth watching across the entire field.

The Maturing School Choice Landscape

School choice is growing up. Across the 2026 legislative session, states spent less time creating new programs and more time strengthening the ones that exist — fixing how the money moves, who can open a school, and what regulators can and cannot touch. Bellwether tracked roughly 130 K-12 bills across two areas: funding portability, meaning the ability of public dollars to follow a student to whichever school or program they pick, and provider expansion, making it easier for new kinds of schools (charters, microschools) to open. Five themes stood out:

1. Open Enrollment Expansion: Open enrollment, which lets a family send their child to a public school outside their home district, expanded in some states and picked up new conditions in others, as legislators start weighing what happens once portability scales up.
2. District Program Access for Nondistrict Students: Several states opened up existing public programs, such as CTE courses, extracurriculars, and JROTC, to students who do not attend district schools, pushing back on the assumption that young people must be enrolled in a district school to use what it offers.
3. Systems Improvements for Existing School Choice Programs: States with established school choice programs spent this session fixing how the money moves: redoing payment systems, restructuring administration, and preparing for scale. This is less exciting than launching a new program, but it often means the difference between a program that works on paper and one that works for families.
4. Pathways for Opening New Types of Schools: More than a dozen states worked on making it easier for new kinds of schools to open: cutting red tape, improving access to facilities, and changing who gets to approve new schools. A few states pushed the opposite way, tightening oversight instead. 
5. Protections for Private Providers in Choice Programs: A smaller group of states locked in legal protections that limit what regulators can do to private providers once they are part of a public choice program such as an education savings account (ESA), shaping how much oversight these programs get.

Collectively, these patterns surface questions for the field about where school choice is headed. 

Theme 1: Expanding Open Enrollment, With Conditions

Open enrollment is the simplest form of portability. It lets families enroll a child in a public school outside their home district, with no private school, no voucher, no scholarship organization involved. In 2026, legislators in roughly a half-dozen states took it up, pushing on the policy in both directions.

Theme 2: Broadening Access to Existing Public Programs

Roughly a half-dozen states considered or passed legislation opening up publicly funded programs, including CTE courses, extracurriculars, and JROTC, to students who do not attend district schools. What these bills share is a quiet challenge to a foundational assumption: that enrolling in a traditional K-12 school is what makes a student eligible for what the district offers. As more students learn outside traditional district schools, that assumption is being tested.

Theme 3: Redesigning the Flow of Funds

School choice programs can look good on paper and still fail families if the money does not move smoothly. Fragile payment systems, unclear funding flows, or administration housed inside institutions that have reasons to keep enrollment small all create friction that families and providers experience as program failure, regardless of what the statute says.¹⁷ How a program moves money is a design choice with real consequences for families. States with established programs are learning this lesson and are looking to rebuild accordingly.

Theme 4: Expanding the Provider Environment

Most school choice debates focus on demand: Can families access the schools available to them? Less attention goes to the supply side: Can new kinds of schools get off the ground at all? The 2026 session saw more than a dozen states working on exactly that, reducing the barriers to entry, improving access to facilities, and reforming who gets to approve new schools. Three sub-patterns stand out.

Theme 5: Locking in Provider Autonomy

A handful of states moved to write into law what the government can and cannot do to private providers once they are participating in publicly funded programs. The provisions showing up across this legislation include defining private providers as not agents of the government; prohibiting regulators from controlling curriculum, admissions, or hiring practices at participating schools; and codifying that rules placing an undue burden on providers are invalid. Not all of these bills passed, but the pattern is appearing in enough places to be an emerging trend worth following.

Rewriting the Rules on Learning and Credentials

In 2026 state legislative sessions, lawmakers continued grappling with a fundamental set of questions: Who gets to decide that K-12 learning happened, what counts as proof, and what does it all cost? For decades, the answer was fixed — put in the seat time, earn a district-issued diploma, and colleges and employers take it from there. That default has been loosening for years, and in this session, the shift surfaced in statehouses across the country. The old path is being reconsidered from several directions at once. Who certifies learning is opening up to employers and apprenticeships; what counts as proof now includes demonstrated skills, micro-credentials, and alternate diplomas; and even established pathways are being reexamined, as programs such as dual enrollment grow large enough that the question shifts from how to expand them to what they cost. The common thread is that certifying learning is no longer the school district’s job alone; employers, colleges, state agencies, and credentials earned outside the classroom all have a hand in it.

Bellwether tracked roughly 200 K-12 bills on credentialing, seat time, and pathway flexibility that passed at least one state legislative chamber. Across those bills, five themes stood out:

1. Work-Based Learning Barriers: States are working to clear legal and structural barriers, such as unclear liability, child-labor rules, and licensing age floors, that keep work-based learning from scaling.
2. Dual Enrollment Guardrails: Dual enrollment has scaled fast enough to attract fiscal scrutiny, and states are moving in both directions at once, broadening access in some places while capping credits and restricting delivery in others.
3. CTE Pathways: Industry credentials are being repurposed as a teacher supply mechanism, opening direct routes from the trades into CTE classrooms rather than treating those credentials only as a student outcome. 
4. The Changing Diploma: The diploma itself is being rewritten: what it signals, how students earn it, and who the system is designed to credential in the first place.
5. Postsecondary Pathway Infrastructure: States are building the shared infrastructure (common applications, transcript standards, transfer guarantees) that makes postsecondary pathway programs work.

These five themes point to the same underlying tension: States are moving quickly to widen who can enter these pathways and how, but far more slowly to build the systems that determine whether that access pays off. 

Theme 1: Clearing the Barriers to Work-Based Learning

Work-based learning has been a policy priority for years, and its persistent gaps are well documented: too few employer partners, students without transportation to worksites, programs that exist on paper but do not place students.⁴³ States have tried to address these challenges through program design, funding, and coordination. In 2026, a cluster of bills targeted the specific legal and structural barriers — unclear liability, child-labor laws that inadvertently exclude work-based learning placements, licensing age floors that block apprenticeship entry, no worksite access for rural students — that program design alone cannot fix.

Theme 2: Managing Dual Enrollment at Scale

Nationally, dual enrollment has grown substantially over the past decade.⁴⁹ In 2026, state legislatures both expanded the program and moved to rein it in. Some moved to contain it, capping state-funded credits, narrowing eligibility, or restricting online delivery, while others moved to widen it, lowering GPA thresholds or opening it to new groups of students. After a decade in which growth was the national story, the notable shift is that constraint has become a visible part of the conversation: Some states are now weighing what participation produces and what it costs, and access that keeps expanding is increasingly paired with caps and conditions.

Theme 3: Rerouting the Credential-to-Classroom Pipeline

Most conversations about industry credentials in K-12 focus on students earning them. In 2026, several states took a different view: The professionals who hold those credentials are also a solution to the CTE teacher shortage. CTE teaching vacancies consistently outpace those in other subjects,⁵⁶ and the traditional certification pipeline was not designed with industry professionals in mind.⁵⁷ Several states stopped waiting for that pipeline to fix itself and started building direct routes from industry into classrooms.

Theme 4: Rewriting the Terms of a Diploma

Several states moved beyond adjusting graduation requirements and started restructuring what the diploma itself means and who can access it. The mechanisms vary — multiple graduation tracks, seat-time flexibility, competency-based test-outs, alternative credentials for students with disabilities, and equivalency pathways opened earlier in the high school sequence. What they share is a common direction. A diploma should reflect the path a student took and what they can demonstrate, not just the minimum threshold they cleared, and it should be reachable by students who existing structures have historically left without a credential at all. This trend is worth tracking because postsecondary institutions and employers have long used the diploma as a standardized signal. However, as states multiply the pathways to earning one, that signal carries less uniform information, and colleges and employers have not yet worked out how to respond.

Theme 5: Building the Infrastructure Behind the Programs

Across the themes in this brief, a common pattern emerges: States are investing in the coordination infrastructure that makes programs work, not launching new ones. This shows up in dual enrollment, in work-based learning, and in credentialing. A dual enrollment program with a separate application at every college is not a pathway. A high school credential that never lands on a transcript or transfers to postsecondary is a credential in name only. A work-based learning program with no answer for who is liable when a student is hurt is not a functioning program. States with established programs are recognizing this and responding with shared systems, shared standards, and shared accountability — the coordination infrastructure that should have been built years ago.

Where Schools Draw the Line on AI

As AI spreads through schools, the question for policymakers is no longer whether it belongs there but how to govern it. Bellwether reviewed approximately 60 K-12 AI-and-education bills that advanced in state legislatures in the 2026 session. Most cluster into a familiar set of themes — acceptable-use policies, AI literacy, educator training, student-data protection, advisory bodies, and family transparency — and other organizations, including ExcelinEd and FutureEd, have already documented them in depth.⁷⁷ Looking at the same bills, Bellwether found an additional pattern: A small set of bills asks not how schools should use AI, but which decisions they should refuse to hand over to it. In these states, lawmakers began naming the decision-making that must stay with humans. It surfaced in three areas:

1. Teacher Evaluation: Keeping AI out of the summative judgment on a teacher’s performance while allowing its use for administrative support.
2. Student Discipline and School Safety: Barring automated systems from being the sole basis for disciplining a student, and requiring human investigation and context before a school acts on an alert or risk flag.
3. Grading and Feedback: Requiring educator oversight of classroom-facing AI and keeping the tools from being the primary basis for student grades, promotion, or retention decisions.

Two questions run underneath these bills — one an open tension between protecting judgment and preserving AI’s benefits, and one a legislative design choice about how durable these guardrails will be.

Teacher Evaluation

Evaluation is where the line between judgment and administration is drawn most explicitly. Lawmakers are comfortable letting AI handle the paperwork around a teacher’s performance but not the verdict. 

Illinois signed Senate Bill 2909 into law in July, and it is the clearest example of this theme in action. The bill amends the state’s teacher evaluation law to prohibit evaluators from using AI to assign a numerical score or qualitative rating for any component of a teacher evaluation, or for any evaluation task requiring professional judgment. It still allows AI tools to support administrative tasks.⁸²

The bill places a parallel limit on teachers. A teacher may not use an AI tool to generate evidence of professional practice that an evaluator will use to judge their performance, though teachers, like evaluators, may still use AI for administrative tasks. Those permitted administrative uses carry a transparency requirement that runs both ways: An evaluator who uses an AI tool must disclose its name and specific purpose to the teacher, and a teacher who uses one must disclose the same to the evaluator. Separately, the joint committee that governs the district’s evaluation plan decides how AI tools may be used across its components.⁸³

The bill is a fairly targeted guardrail, as it does not appear to prevent an evaluator from using an AI transcription tool or drafting support, but the final act of evaluation judgment cannot be delegated to a tool.

Student Discipline and School Safety

Discipline decisions raise the stakes further, since an automated flag or risk score can trigger real consequences — removal from a classroom, a law enforcement referral — before a human ever weighs in. Several states are moving to guarantee that a person, not a system, makes that call.

Washington Senate Bill 5956, which passed the Senate, touches on a similar concern but in the context of student discipline.⁸⁴ The bill is grounded in the legislature’s finding that public schools are beginning to use AI, automated decision systems (i.e., a computational process that “makes or materially influences decisions or recommendations concerning a student,”⁸⁵ such as flagging a student at risk of dropping out),⁸⁶ and surveillance technologies in ways that can affect student discipline and school safety. It is also grounded in the legislature’s finding that these tools can amplify existing disparities. The bill’s stated intent includes ensuring that automated decision systems do not replace the judgment of trained school personnel in discipline-related decisions.

The bill would require that an automated decision system may not be the sole or determinative basis for any student discipline-related decision. The bill would also prohibit emergency removal, suspension, expulsion, law enforcement referral, or assignment to an alternative education setting based only on an automated prediction, score, or classification, or on surveillance data, without independent human investigation and consideration of context.⁸⁷ In practice, a school could still receive an alert or review data from a safety system. But before acting, school personnel would need to review the context and make an independent judgment. 

Oklahoma’s Senate Bill 1734, now enacted as the Oklahoma Responsible Technology in Schools Act, also touches on student discipline by specifying that AI tools may not be used as the primary basis for discipline, among other higher-stakes educational decisions.⁸⁸

Grading and Feedback

Grading and feedback sit closest to the classroom’s daily rhythm, which is why the guardrails here carry the most risk of overreach.

Oklahoma’s Responsible Technology in Schools Act, as enacted, is more notable for how far it reaches into everyday classroom practice. It not only limits AI use in formal discipline or placement decisions, as mentioned above, but it also requires classroom- and student-facing AI tools to be used under the direction of an educator, with AI outputs reviewed by an educator or authorized school employee before they are used in instruction, feedback, assessment, or decision-making. The law also says AI tools may not be used as the primary basis for student grading, placement, promotion, retention, or other higher-stakes educational decisions.⁸⁹

That language reflects the same concern running through other state bills included in this analysis: Schools should not hand over judgment to AI systems. However, Oklahoma’s approach could have broader practical effects because it applies not only to final decisions but also to instruction and feedback. Depending on how districts interpret the law, it could limit some tools designed to give students quick feedback or help teachers review student work more efficiently.⁹⁰ 

Idaho enacted Senate Bill 1227, which takes a less prescriptive approach but points in a similar direction.⁹¹ The law directs the state’s Department of Education to develop a statewide generative AI (GenAI) framework that prioritizes human-centered oversight and ensures that AI does not replace or eliminate human teachers. It also describes GenAI in education as tools that may assist educator decision-making while keeping human judgment as the final authority. Unlike Oklahoma, Idaho leaves more of the details to the state framework and local policy. However, the underlying signal is that states are beginning to determine what decisions should be left to teachers, not AI.

Appendix: Methodology

The “Structural Signals” series draws on a scan of 2026 K-12 education legislation in all 50 states and the District of Columbia. The work was designed, overseen, and interpreted by Bellwether analysts and aided by AI tools at specific, defined steps.

Scope

This legislative scan was not an exhaustive attempt to catalog everything state legislatures did in education in 2026. Instead, it targeted bills with substantive provisions in three structural education policy categories:

1. Funding portability and provider expansion, including education savings accounts, vouchers, open enrollment with funding portability, microschools, course-level funding, and scholarship tax credits.
2. Credentialing, seat-time, and pathway flexibility, including competency-based credit, seat-time waivers, industry-credential integration, credit for prior learning, and cross-sector workforce pathways.
3. AI governance and system modernization, including AI procurement standards, AI-use policies, educator training mandates, pilot programs, cross-agency data integration, and student-data privacy related to AI.

One deliberate threshold applied throughout this analysis: A bill had to have passed at least one state legislative chamber to be in scope. A bill that cleared a chamber but stalled before enactment still reflects genuine momentum, so it stayed in; a bill that died in committee, however notable, is outside the scope of this analysis.

Data

LegiScan bulk datasets were used, covering all 50 states and the District of Columbia. Each dataset contains every bill introduced in a state’s 2026 session, along with its title, description, sponsor and committee history, progress milestones, and other metadata. The full bill text was pulled only after filtering down the full list of bills. A bill’s title and summary rarely reveal the provisions that matter for this analysis; those are often buried deep in the text, which is why the full document is what gets classified once the universe of potentially relevant bills is narrowed.

Narrowing the Field

Every introduced bill analyzed in this series passed through a sequence of filters on its way to becoming a viable candidate pool. The first six filters were rule-based; the seventh filter used an AI model. A separate audit step (described below) ran alongside these filters to catch bills the rules missed.

1. Drop federal legislation. This series focuses exclusively on state-level policy; U.S. Congress bills were omitted.
2. Keep only bills. LegiScan tagged each instrument by type. Only formal bills (the kind that become statutes) were kept; resolutions, memorials, and proclamations were set aside.
3. Match the K-12 topic vocabulary. A bill had to reference K-12 education terms in its subject tags, title, or description. This screen is intentionally permissive: It casts a wide net over anything plausibly K-12-related, and it lets higher-education bills through, because credentialing and pathway bills frequently span K-12 and postsecondary.
4. Confirm an education-related committee referral. A bill had to have been referred, currently or historically, to a committee whose name signals education, workforce, or a related area. This removed bills that touch schools only incidentally (healthcare, transportation, real estate licensing) and were never seriously routed through education policy channels.
5. Require passage of at least one chamber. A bill had to show chamber passage in its current status or its history.
6. Require some type of action in 2026. A bill had to have had some movement in 2026. 
7. Classify relevance with an AI model. Every bill that cleared the prior filters had its full text read by an AI model against the three category definitions. The model returned a yes-or-no verdict and a one-sentence justification for each bill.

Two of the rule-based screens (Steps 3 and 4) included an AI carve-out: Any bill whose title or description mentions AI-related vocabulary (e.g., artificial intelligence, machine learning, automated decision systems, algorithmic tools, student-data privacy, longitudinal data systems, and similar terms) bypassed the topic and committee screens entirely. This carve-out existed because AI-in-education bills are sometimes written without K-12 keywords and are frequently routed through technology, innovation, or government operations committees rather than education committees. Without the AI carve-out step, those bills would have been erroneously omitted.

The Audit

To guard against the keyword-and-committee filter missing bills (especially creatively named ones), the bills the rule-based filters rejected were periodically audited. Every chamber-passed bill that the topic or committee screens had excluded had its title and description sent to the AI model in batches, alongside a calibration sample of bills already judged relevant, and the model flagged any whose title or description signaled that they should have stayed in. Flagged bills were then routed back through the pipeline and classified against their full text. 

This audit recovered specific bills, which were appended directly to the candidate pool even when the rules would not have caught them on their own. It also improved the rules themselves as the model reviewed the existing keyword patterns against the rejected pool and proposed terms to add. The expanded term lists in the topic and committee screens came directly from this feedback. The first audit alone surfaced 112 bills that had been wrongly excluded — most of them AI bills that never mentioned K-12 topics in their titles or descriptions.

After Bellwether analysts confirmed whether the bucketing was correct, Anthropic’s Claude read each bill and created a short description of its key features. Bellwether’s team used that description along with the rationale of the bucketing to help identify the themes. 

Human Oversight

The filters and audit together produced a viable candidate pool of roughly 570 bills. That pool was the starting point for Bellwether’s human analysis. Bellwether analysts reviewed the candidates against the full bill text, set aside those that did not hold up on closer reading, and organized the rest by category. The set of final bills discussed across the “Structural Signals” series’ three briefs is roughly 390 bills: approximately 130 in funding portability and provider expansion, roughly 200 in credentialing and pathway flexibility, and roughly 60 in AI governance and system modernization.

Reviewing, Coding, and Analyzing the Candidate Bills

Once the triage pipeline produced the set of bills, the project moved from selection to analysis in three stages: a second review of the initial categorization, structured coding of each bill, and early theme development. Each stage paired the AI model with human review.

A. Reviewing the Initial Categorization

Starting from the full set of candidate bills, the model conducted a second review of the initial categorization (i.e., the assignment of each bill to one of the three policy areas) as a check on the triage step. The model did two things in this pass. First, it flagged bills where the actual bill text conflicted with the original triage reasoning (17 bills). For long bills, this review relied on partial reads of extracted summaries rather than the full text. Second, it rated its own confidence in each bill’s categorization as high, medium, or low. Human review then covered:

• All flagged bills.
• All low-confidence bills.
• Approximately 10% of the medium- and high-confidence bills, as a spot check on classification accuracy.

B. Coding the Bills

Bellwether developed a template, which the model completed for every bill for each of the three topic areas. The template captured:

• Category: The type of policy within the bill’s area, to help characterize what kind of legislation it is. For example, in funding portability: ESA, voucher, tax credit scholarship, or microschool; in pathways: dual enrollment, work-based learning, industry credential, or competency-based education; in AI: governance policy, literacy curriculum, study, or training.
• New or Modified: Whether the bill created a new program or law or modified existing law.
• Summary: One to two sentences on the bill’s focus relative to its category.
• Notable Provisions: Two to four bullets capturing the most important provisions relative to that category.
• Surprising: Anything unusual relative to a typical bill in that area.
• Review Flag: Set when a borderline call needed human review.
• Coding Rationale: A short description of why a bill was flagged for review.

Funding portability came first, and its approach differed slightly. For that category, the model read a sample of the bills and suggested the template categories. An initial attempt to code additional detail (the type of charter provision, microschool provision, or program structure) proved insufficiently accurate to use in the analysis. That coding was dropped from the analysis, and similar information was not collected for the pathways or AI categories.

Model use and reading approach. For shorter and medium-length bills (often under 30 pages), the model read the full text in small batches, using Claude Opus 4.7 (for “The Maturing School Choice Landscape” analysis) and Claude Opus 4.8 (for the “Rewriting the Rules on Learning and Credentials” and “Where Schools Draw the Line on AI” analyses). Longer bills were handled through manual review or targeted model review; for these, the model first reviewed how the text was marked up (e.g., strikethrough, double-strikethrough, colors) to establish the bill’s structure before coding it.

Review process. Bellwether analysts reviewed all flagged bills, along with 10% of those not flagged. When errors surfaced, the prompt was updated. Most corrections concerned whether a bill should be included at all; for example, flagging bills that were primarily postsecondary or early-childhood focused, or, for pathways, flagging general teacher-certification pathways with no CTE connection. Occasionally, the errors were interpretive, but most often the errors came from the model describing the bill text as proposing something new when it was only a technical change.

C. Developing Themes

The completed templates were used to identify early themes within each category. Bellwether analysts then reviewed and updated the themes as well as read the full text of each bill cited as an example of a given theme, confirming that it supported the theme it was attached to.

Ensuring Rigor When Using AI

AI did a specific, bounded job in this project: It read the full text of thousands of bills and made a first-pass relevance call. It did not decide what the categories mean, which bills made the briefs, or what those bills signal. Those judgments stayed with Bellwether’s team. Several design choices kept the role of AI in check:

• The model was told to err toward inclusion. The classification prompt instructed the model to flag any bill that arguably touched one of the three areas and to leave the harder calls for human review, rather than making a final exclusion on its own.
• Bills were routed to different models by length. Shorter bills (under roughly 120,000 characters) went to a faster, lower-cost model, Claude Haiku 4.5. Longer bills, typically omnibus packages where a relevant provision can be buried in one section of many, went to Claude Sonnet 4.6 with an extended-context configuration (a one-million-token beta) so the model read the entire bill rather than only a portion of it. This routing meaningfully improved accuracy on long bills at a modest added cost.
• The audit loop checked the rules, not just the bills. As described above, the audit was a standing check on false negatives. It repeatedly asked whether the filters were mistakenly dropping bills, and it fed corrections back into the term lists.
• Every verdict was reviewable, and every relevant bill was spot-checked. For each bill, the classification output recorded its state, identifiers, title, the model’s verdict, and its one-sentence reasoning. Bill texts were then organized into a state-by-relevance folder structure, with a per-state index pairing each bill to the model’s reasoning, so a reviewer could open the underlying text and confirm the verdict against what the bill actually says.

This approach could not guarantee perfection. A large language model classification carries real risks: A model can misread a bill, and its one-sentence justification is a summary of its verdict, not a full account of its reasoning. The safeguards above (an inclusive prompt, the audit loop, length-based routing, and human spot-checks) are designed to catch and correct those errors. The counts and category assignments in the briefs reflect human judgment applied on top of the AI’s first pass, not the AI’s output on its own.

Limitations

The scan made deliberate trade-offs in breadth, precision, and feasibility:

• The topic screen was keyword-based. A creatively titled bill with no recognized keyword in its subject tags, title, or description can slip past the rules. The audit step mitigated this but could not fully eliminate it; new naming patterns needed another audit pass to surface.
• Committee names vary by state. Most states use recognizable committee names, but a bill routed only through an unusually named committee could be missed. The AI carve-out covered this for AI bills; bills in the other two categories with unusual routing relied on the audit’s manual recovery.
• The AI carve-out was keyword-based. It caught bills that use AI vocabulary explicitly, not bills framed entirely around concepts such as “automated decision tools” or “ed tech procurement” without those exact terms.
• Bills published only as image-only PDFs could not be classified from text. The pipeline reads extracted text, and a small number of bills, typically from states that publish scanned documents, have none; these were marked as skipped.