Section 1: Introduction

1.1 Purpose

The BenchMark Standard provides a structured, repeatable framework for evaluating artificial intelligence tools deployed in or proposed for use in state, county, and municipal court judicial settings. It is designed for two audiences:

Courts seeking to evaluate AI tools before adoption, during deployment, and on an ongoing basis.
Vendors seeking to demonstrate that their tools meet the safety, accuracy, and constitutional requirements of judicial use.

This framework does not regulate the practice of law. It does not mandate or prohibit AI adoption. It provides the evaluation methodology that courts and vendors need to make informed decisions.

1.2 Scope

What This Framework Covers

AI tools used by judges, court staff, and court-appointed personnel in state, county, and municipal courts in the performance of judicial duties.
Tools used for legal research, document drafting, case management, scheduling, analytics, risk assessment, and decision support.
Both generative AI (large language models, document generators) and predictive AI (risk assessment instruments, outcome prediction tools).
Tools deployed as standalone applications, integrated features within existing court management systems, or application programming interface (API) based services.

What This Framework Does Not Cover

AI tools used exclusively by private attorneys in their own practice (governed by state bar ethics rules).
Law enforcement AI (facial recognition, predictive policing, surveillance): related but distinct regulatory domain.
Jury selection algorithms: subject to separate constitutional analysis.
Court recording and transcription systems, unless they incorporate AI-driven legal analysis.

Jurisdictional Note

The BenchMark Standard is designed with Tennessee law and court structure as the primary reference but is intentionally jurisdiction-agnostic in its methodology. The six evaluation domains, scoring system, certification tiers, and evaluation process apply universally.

Tennessee-specific elements appear in:

Test cases: Legal scenarios reference TCA statutes and Tennessee case law. Other states would substitute their own statutory and case law references.
Legal authority (Appendix B): Tennessee constitutional and statutory citations. Each state has equivalents.
Implementation plan: Proposes a Tennessee pilot. Adaptable to any state's administrative structure.

Courts in Arizona, Illinois, New York, Ohio, and other states with emerging AI policies can adopt this framework by replacing jurisdiction-specific references while retaining the core methodology.

Scope Exclusions

This framework evaluates AI-specific risks in judicial settings. It does not address:

Americans with Disabilities Act (ADA) accessibility compliance: Important for all court technology but enforced through existing law, not AI-specific evaluation. Future versions may add accessibility as a cross-cutting requirement.
General IT security beyond CJIS: Network architecture, hardware security, and non-AI software are governed by existing court IT standards.
User conduct: This framework evaluates the AI tool. It does not monitor or constrain how a user (judge, attorney, clerk, or pro se litigant) chooses to rely on the tool's output. A certified tool used carelessly or in violation of professional duties does not become uncertified, and an uncertified tool used carefully does not become safe. User behavior is governed by the Tennessee Code of Judicial Conduct (Supreme Court Rule 10), the Tennessee Rules of Professional Conduct (Supreme Court Rule 8), and the supervisory authority of presiding judges. The framework presumes users operating within those obligations.

Real-Time Courtroom Use

The framework does not regulate courtroom behavior. Real-time conduct in a Tennessee court remains governed by the Tennessee Rules of Evidence, the Tennessee Rules of Criminal Procedure, the Tennessee Rules of Civil Procedure, the Tennessee Rules of Juvenile Practice and Procedure, the Tennessee Supreme Court Rules, and the Tennessee Code of Judicial Conduct. The BenchMark Standard does not amend, interpret, or replace any of those authorities.

What the framework does evaluate is whether a tool is technically constrained to remain on the legal-reference side of the courtroom line. AI use in connection with a hearing may support:

Retrieval and citation of legal authorities, including statutes, case law, and Tennessee constitutional provisions.
Retrieval, organization, and summary of court rules and standing orders.
Retrieval of Department of Children's Services policies and other approved reference materials.
Production of cross-references among authorities the court is already considering.

AI use in connection with a hearing may not:

Investigate facts outside the record.
Evaluate witness credibility.
Recommend case-specific rulings.
Supply extra-record adjudicative facts.
Replace independent judicial decision-making.

A certified tool must be technically capable of being operated within the first set of functions and technically constrained against the second. Whether and how a court chooses to use a certified tool in real time during a hearing is a matter for the court, the parties, and the controlling Tennessee rules cited above. The certification answers whether the tool can be used consistently with those authorities; it does not decide whether, in any given hearing, it should be used.

The BenchMark Standard evaluates tools, not judicial conduct. It does not amend, interpret, or replace the Tennessee Rules of Evidence, the Tennessee Rules of Criminal Procedure, the Tennessee Rules of Civil Procedure, the Tennessee Rules of Juvenile Practice and Procedure, the Tennessee Supreme Court Rules, or the Tennessee Code of Judicial Conduct. A certified tool must be capable of being used consistently with those authorities. Certification does not decide admissibility, confrontation, disclosure, discovery, notice, recusal, evidentiary foundation, judicial ethics, or hearing procedure in a particular case.

Who Uses the Tool

The constitutional and ethical analysis of AI use in court depends not only on what the tool does, but on who operates it and how close that user is to the actual judicial decision. The framework as written assumes the user is a judge or judicial officer. In practice, AI tools in Tennessee state courts may be operated by:

Judges and judicial officers. The presumed user. Outputs are subject to the judge's independent legal analysis and the non-delegation principle in Domain 6. Standard tier requirements apply.
Magistrates, referees, and judicial commissioners. Operate under judicial supervision but may issue findings, recommendations, or orders subject to judicial review. The user's proximity to the judicial decision determines tier; a referee preparing recommended findings of fact in a juvenile matter operates near the judicial decision and should use only Certified or Certified-Sensitive tools.
Court-appointed attorneys and guardians ad litem. Use AI tools to investigate, prepare reports, and advocate on behalf of represented parties (often minors or persons under disability). Their outputs influence judicial decisions but are subject to adversarial testing. Their tool use is governed by state bar ethics rules in addition to the certification framework.
Staff attorneys and judicial law clerks. Often the actual operators of legal research and drafting tools, with outputs delivered to a judge. The judge's review is the formal check, but the tool's output structures the judge's deliberation. Certified is the appropriate floor for any tool used by staff attorneys for substantive legal work.
Court clerks and administrative staff. Operate scheduling, docketing, and document management tools. Verified tier applies to most clerk and administrative use. When an administrative tool's output affects notice, access, language services, or public-facing information, the Domain 2 (Bias) and Domain 3 (Constitutional) floors in the Verified tier come into play.
Pro se litigants interacting with court-provided AI tools. Some Tennessee courts provide self-help kiosks, online filing assistants, or AI-driven information services for self-represented parties. These tools are not used by court personnel but are deployed by the court for public use. The certification analysis treats these as administrative tools facing the public, with heightened attention to Domain 2 (language register, accessibility) and Domain 5 (clear disclosure that the tool is not legal advice).

The framework's tier requirements apply by function, not by user category alone. A scheduling tool is Verified-tier whether the user is a judge or a clerk. A legal research tool is Certified-tier whether the user is a judge or a staff attorney. The user category becomes determinative when the same tool can be used at different proximities to the judicial decision; in that case, the higher tier governs.

1.3 The Problem

No Bridge Between AI Capability and Judicial Safety

The gap in American judicial AI governance is not a lack of concern; it is a lack of methodology.

What exists today:

Entity	Contribution	Limitation
NIST AI Risk Management Framework	General AI risk taxonomy and governance principles	Not court-specific. No evaluation methodology for judicial tools.
EU AI Act (Annex III, §8a)	Classifies judicial AI as "high-risk" with compliance requirements	European jurisdiction. No American equivalent.
NCSC AI Governance Guides	Readiness assessment and governance principles for courts	Advisory only. Does not evaluate specific tools.
State court policies (IL, AZ, NY, OH)	Regulate attorney and court staff use of AI	Regulate behavior, not tools. No evaluation criteria.
ABA Formal Opinion 512	Guidance on attorney duties when using AI	Ethics framework, not technical evaluation.
Tennessee SB 1493 / HB 1455 (pending)	Would criminalize training AI for certain harmful conduct if enacted	Enforcement statute proposal, not evaluation methodology.

What does not exist:

A systematic framework that answers the question every court administrator must face: "This vendor says their AI is safe for our court. How do we verify that claim?"

The BenchMark Standard answers that question.

1.4 Design Principles

Five principles guided the development of this framework:

1. Court-Centric

Every criterion, test case, and scoring threshold is designed from the perspective of a working court, not a research lab, a law school, or a technology company. The question is always: Does this matter to a judge making decisions that affect people's lives and liberty?

2. Practical Over Theoretical

This framework is built to be used, not admired. Test methodologies use language courts understand. Scoring produces clear pass/fail determinations. A court administrator with no AI expertise should be able to read an evaluation report and understand the result.

3. Jurisdiction-Adaptable

The six-domain structure and evaluation methodology are universal. The specific test cases, legal references, and thresholds can be adapted for any state. Tennessee-specific test cases are provided as the reference implementation.

4. Vendor-Accessible

Vendors should be able to self-evaluate before formal submission. The framework is published openly, not behind a paywall or locked in a certification body. Transparency builds trust. Hidden criteria breed suspicion.

5. Living Document

AI capability changes quarterly. Legal standards evolve annually. This framework is versioned and will be updated. Each version is a starting point: a foundation, not a monument.

1.5 How to Use This Document

If you are...	Start with...
A judge or court administrator interpreting an AI tool's certification report	Executive Summary → Section 2 (Methodology) → Domain sections relevant to your use case → Section 9 (Certification Tiers)
A vendor preparing for BenchMark evaluation	Executive Summary → Section 2 (Methodology) → All six domain sections → Vendor Guide (companion document)
A state AOC or Supreme Court committee	Executive Summary → Section 1 (this section) → Implementation section → Appendix B (Legal Authority) → Appendix C (Framework Crosswalk)
A legislator or policy advisor	Executive Summary → Implementation section → Appendix B (Legal Authority)
A researcher or academic	Full document → Appendix C (Framework Crosswalk) → Test Case Repository

1.6 Author's Note

I am a sitting General Sessions and Juvenile Court judge in Tipton County, Tennessee. I am also a builder of legal AI tools. BenchBook, Sayada, and WarrantWorks are products I helped create and use in my own court.

A General Sessions and Juvenile Court bench in Tennessee is not a narrow vantage point. The work routinely intersects with circuit court (felonies bound over from preliminary hearings, civil matters appealed from General Sessions, divorce and domestic relations crossing into juvenile dependency), with chancery court (termination of parental rights (TPR) proceedings, adoptions, and probate matters that share parties or facts with juvenile or General Sessions cases), and with municipal courts on overlapping ordinance and traffic matters. The framework that follows reflects that breadth.

This is not a conflict of interest. It is the reason this framework exists.

Most people writing about judicial AI have never sat on the bench. Most judges using AI tools have never built one. Most vendors selling to courts have never presided over a case where liberty was at stake.

I have done all three. That perspective (builder, user, and judicial officer) is what makes this framework different from every other AI governance document published to date.

The BenchMark Standard asks the questions a judge would ask, tests the things a judge needs tested, and scores the results the way a judge would score them. It is practical because a working court needs practical tools. It is demanding because the stakes in a courtroom are not theoretical.

I submit this framework to my colleagues on the bench, to the Tennessee Administrative Office of the Courts, and to the broader judicial community not as a finished product but as a foundation, the beginning of a conversation that American courts can no longer afford to delay.

Judge M.O. Eckel III, Tipton County, Tennessee. 2026.

1.7 The Certification Model

Before turning to the evaluation methodology, this section explains who does what under the BenchMark Standard. The framework is built around a clear separation of roles among three parties: the certifying body, the adopting authority, and the courts that rely on certified tools.

The Certifying Body

The BenchMark Standard is operated by the Judicial AI Standards Institute, an independent certification body operated as a division of Velocity Venture Holdings LLC. The Institute publishes its registry, evaluation methodology, and current evaluator roster at judicialaistandards.org.

The Institute:

Maintains the published evaluation methodology, test case repository, and scoring rubrics.
Trains and qualifies independent evaluators who meet the requirements set out in Section 2.5.
Receives vendor submissions, conducts formal evaluations, and issues certification grades (Verified, Certified, or Certified-Sensitive).
Publishes evaluation reports and maintains the public registry of certified tools.
Conducts recertification at the intervals required by each tier.

The Institute is not a court. It does not adopt the Standard for any jurisdiction, and it does not decide which tools any particular court must use. Its function is technical: to test, to grade, and to publish.

Vendors

Vendors of judicial AI tools submit their products to the Institute for evaluation. They:

Pay the evaluation fee for the tier they are pursuing.
Provide the technical documentation and access required by the intake process described in Section 2.2.
May conduct a self-evaluation using the published methodology before formal submission, but self-evaluation does not constitute certification. Only an evaluation conducted by an independent qualified evaluator under the Institute's program produces a BenchMark grade.
Receive the evaluation report and the certification grade earned.

A vendor whose tool fails evaluation may remediate and resubmit under the procedures in Section 2.7. There is no limit on resubmissions.

The Adopting Authority

In Tennessee, the adopting authority is the Administrative Office of the Courts. The AOC's role is policy adoption, not technical execution. Specifically, the AOC:

Adopts the BenchMark Standard as the recognized evaluation methodology for AI tools used in Tennessee courts.
Sets minimum certification requirements for court use (for example, AI tools used in Tennessee courts must hold at least BenchMark Verified status; AI tools used in juvenile, sealed, or sensitive proceedings must hold BenchMark Certified-Sensitive status).
Publishes a public list of certified vendors and their grades, drawn from the Institute's registry, accessible to every court in the state.
Updates the published list as new certifications are issued, recertifications occur, and suspensions or revocations take effect.

The AOC does not evaluate tools. It does not maintain test cases. It does not employ evaluators. It adopts a standard developed and operated by an independent body, and it publishes the results.

The Courts

Individual courts do not evaluate, test, or certify any AI tool. They consult the AOC's published list of certified vendors, see what grade each vendor holds, and make adoption decisions appropriate to their court's caseload, budget, and operational needs.

A court selecting an AI tool for use:

Reviews the AOC's published list to confirm a vendor holds the certification tier required for the intended use.
Reviews the published evaluation report for any conditions, remediations, or disclosures attached to the certification.
Procures the tool through ordinary court purchasing procedures.
Uses the tool consistent with the conditions of its certification tier (for example, all Certified outputs require human review before use).

A court does not need technical capacity to evaluate AI. The technical evaluation has already been done. The court's role is operational adoption, not technical assessment.

Why This Structure

The three-party separation answers four questions that any judicial AI governance framework must address:

Who is qualified to evaluate? A specialized body with legal training, technical literacy, and demonstrated independence, not every individual court.
Who has authority to adopt? The state's administrative authority for the courts, exercising the policy function it already exercises for other court technology and standards.
What protects evaluator independence? The certifying body is structurally separate from the vendors it evaluates and from the courts that rely on certified tools. It has no contractual relationship with either party that creates a conflict of interest.
What happens when courts disagree about a vendor? They cannot disagree about certification, because courts do not certify. A court may decline to procure a certified tool for budgetary, operational, or fit reasons, but the certification grade is a single, statewide determination.

This structure also addresses the practical concern that many smaller courts lack the technical capacity to evaluate AI tools on their own. They do not need that capacity under the BenchMark Standard. The technical evaluation has already been performed by qualified evaluators, the results have been published by the AOC, and the court's task reduces to a procurement decision against a known grade.

1.8 Court Applicability

The BenchMark Standard is designed to evaluate AI tools deployed in any Tennessee state trial court. The framework's six domains, scoring system, and tier structure apply uniformly. What varies by court type is which AI use cases are most likely to arise and which certification tier is appropriate for a given function.

General Sessions Courts

General Sessions courts in Tennessee handle criminal misdemeanors, traffic offenses, civil matters within statutory dollar limits, preliminary hearings on felony cases, and orders of protection. Likely AI use cases include calendar and docket management, legal research on misdemeanor and traffic statutes, drafting standard form orders, and probation monitoring support. The dominant tier in General Sessions practice is Verified for administrative functions and Certified for research and drafting assistance. Certified-Sensitive is rarely required at this level except in matters that touch sealed records or juvenile-adjacent proceedings.

Juvenile Courts

Juvenile courts hear delinquency, dependency and neglect, status offenses, transfer to adult court proceedings, custody where joined with juvenile matters, and termination of parental rights under T.C.A. § 36-1-113. Confidentiality requirements under T.C.A. § 37-1-153 apply throughout. Any AI tool used in juvenile court that touches case content, party identities, or proceedings must hold Certified-Sensitive certification. Tools used purely for administrative scheduling without access to case content may operate at Verified, but the determination requires careful review of what data the tool actually accesses.

Circuit Courts

Circuit courts hear felonies, civil matters above General Sessions jurisdictional limits, divorce and domestic relations, jury trials, and appeals from General Sessions. Practice areas include healthcare liability, personal injury, complex civil litigation, and criminal proceedings carrying significant sentence exposure. Likely AI use cases include legal research across complex case law, drafting jury instructions, sentencing memoranda preparation, and case management for multi-party civil litigation. Certified is the typical tier for substantive legal work; Certified-Sensitive applies to any tool used in proceedings involving sealed records, sensitive medical information, or matters that overlap with juvenile or chancery jurisdiction.

Chancery Courts

Chancery courts handle equity matters, probate, adoption, termination of parental rights (concurrent with juvenile court), guardianships, conservatorships, and complex civil litigation involving equitable relief. Practice areas include probate administration, contested estates, real property disputes, and TPR proceedings filed in chancery. AI tools used in chancery for probate matters, guardianship determinations, or TPR/adoption work must hold Certified-Sensitive certification. The presence of minor heirs, mental health considerations, and family privacy interests in chancery practice raises the certification floor.

Municipal Courts

Municipal courts handle city ordinance violations and certain traffic matters. AI use cases are concentrated in administrative functions: calendar management, ordinance lookup, citation processing. Verified tier covers most municipal court applications.

Cross-Court Considerations

Several practice areas span multiple court types and require attention to which court is hearing the matter:

Termination of parental rights can be filed in juvenile court, chancery court, or circuit court depending on case posture and jurisdictional facts. Certified-Sensitive applies in all three.
Adoption proceedings typically originate in chancery but may involve juvenile court files for predicate findings. Certified-Sensitive applies to tools that touch either side of the proceeding.
Domestic relations matters can move between General Sessions (orders of protection), juvenile court (dependency findings affecting custody), circuit court (divorce, parenting plans), and chancery (adoption following TPR). The certification analysis follows the data the tool actually accesses, not the court in which the tool is nominally deployed.
Civil matters appealed from General Sessions to Circuit raise the question of whether a tool certified for General Sessions use is also appropriate for the appellate posture. The framework treats the tier as following the function, not the court; if the underlying use does not change, the tier does not change.

The principle running through all of these: certification follows the data and the function, not the court name on the docket.