Where AI stresses a legal duty, ranked by what the record already shows — capability (L1) → ruling (L2) → control adoption (L3), weighted by authority, not volume.
As of 2026-09-17 · 108 tracked items · each fault line shows L1 capability / L2 ruling / L3 adoption (0-10)
Bar length is how many sources stand behind the duty, and the colours are what kind of authority they are.
Each of these is a control the law has already moved on, at or above the flag threshold of 8.0 on ruling evidence only, so the list does not depend on which jurisdiction you practise in. The order is what to act on first: a control whose antecedents historically bound in 192 days gives less warning than one that took 907, so it comes first. The source count beside each one tells you how much to trust the line, not how soon to act on it. Open any of them to read the sources yourself.
Bars are treating AI competence as a program requirement, not a personal one.
12 sources · 5 appellate · 7 bar guidance · first signal 192d before it bound
Model Rules: 1.1 · our shorthand for this line is “Firm-wide training + governance program”
L1 2.5→5.8 L2 6.0→9.3 L3 5.5→5.5 E 6.0→7.9
Every meter reads 0 to 10, weighted by source authority rather than volume. L1 is what the AI can already do. L2 is what courts and bars have done about it. L3 is whether insurers and buyers have made the fix table stakes. The lines do not all run in that order: a statute can move L2 with almost no L1 behind it.
Why now: The duty is shifting from 'understand the tool you use' to 'run a governance program' — training, written policies, and oversight — with mandatory AI continuing-education requirements spreading state by state.
What to put in place: Training and enablement at scale (iTrain bet; AI Build Lab proof).
When a tool fails, the first question is what you can show about which one.
6 sources · 1 trial court · 4 statute/rule · 1 bar guidance · first signal 451d before it bound
Model Rules: — · our shorthand for this line is “Tool-provenance documentation”
L1 2.5→2.5 L2 3.5→8.3 L3 2.5→3.3 E 3.0→3.0
Every meter reads 0 to 10, weighted by source authority rather than volume. L1 is what the AI can already do. L2 is what courts and bars have done about it. L3 is whether insurers and buyers have made the fix table stakes. The lines do not all run in that order: a statute can move L2 with almost no L1 behind it.
Why now: Today the lawyer carries all the risk when an AI tool fails; there is no clear legal path to hold the vendor responsible. Pressure is building, through disclosure mandates and product-liability proposals, to shift some of that liability onto the toolmakers.
What to put in place: Track for a shift; document tool provenance now.
Protective orders increasingly name which tools are allowed, and how.
13 sources · 2 appellate · 1 statute/rule · 10 bar guidance · first signal 500d before it bound
Model Rules: 1.6 · 1.7 · 1.9 · our shorthand for this line is “Data-flow mapping + vendor attestation”
L1 4.5→4.5 L2 6.0→9.4 L3 6.0→6.0 E 5.5→6.6
Every meter reads 0 to 10, weighted by source authority rather than volume. L1 is what the AI can already do. L2 is what courts and bars have done about it. L3 is whether insurers and buyers have made the fix table stakes. The lines do not all run in that order: a statute can move L2 with almost no L1 behind it.
Why now: AI tools that learn from what you feed them, or mix data across matters, put client confidentiality at risk. Firms are being pushed to map where client data goes and get written no-training guarantees from vendors before a tool is approved.
What to put in place: Data-flow mapping and vendor attestation at tool onboarding.
Operating to the strictest rule you touch is cheaper than tracking every patchwork.
10 sources · 10 statute/rule · first signal 601d before it bound
Model Rules: 1.1 · 1.6 · our shorthand for this line is “One operating model, to the strictest standard”
L1 2.5→2.5 L2 6.2→9.3 L3 6.0→6.0 E 4.0→4.0
Every meter reads 0 to 10, weighted by source authority rather than volume. L1 is what the AI can already do. L2 is what courts and bars have done about it. L3 is whether insurers and buyers have made the fix table stakes. The lines do not all run in that order: a statute can move L2 with almost no L1 behind it.
Why now: The rules are a patchwork — the EU AI Act, Colorado, Texas, 35-plus state bars, 300-plus judges — each slightly different. The patchwork itself is the pressure: a firm working across jurisdictions needs one operating model built to the strictest standard.
What to put in place: The unifying operating model. This is the whole thesis.
Courts are beginning to order disclosure up front, not just punish after.
14 sources · 4 appellate · 3 trial court · 3 statute/rule · 4 bar guidance · first signal 730d before it bound
Model Rules: 3.3 · Rule 11 · our shorthand for this line is “Verification-by-default with an audit trail”
L1 5.5→5.5 L2 6.4→9.5 L3 6.0→6.0 E 6.0→7.2
Every meter reads 0 to 10, weighted by source authority rather than volume. L1 is what the AI can already do. L2 is what courts and bars have done about it. L3 is whether insurers and buyers have made the fix table stakes. The lines do not all run in that order: a statute can move L2 with almost no L1 behind it.
Why now: Courts first punished fake AI citations after the fact (Mata v. Avianca was the first such sanction). Now more than 300 judges require lawyers to disclose AI use up front, and that patchwork of local orders is converging toward a single certification standard.
What to put in place: Verification-by-default with an auditable trail, so certification is a byproduct.
Sanctions run to six figures, and they land on the signing lawyer.
20 sources · 7 appellate · 6 trial court · 2 statute/rule · 5 bar guidance · first signal 907d before it bound
Model Rules: 5.1 · 5.3 · our shorthand for this line is “Documented verification + trace logs”
L1 5.5→7.8 L2 6.2→9.7 L3 6.5→6.5 E 6.5→8.4
Every meter reads 0 to 10, weighted by source authority rather than volume. L1 is what the AI can already do. L2 is what courts and bars have done about it. L3 is whether insurers and buyers have made the fix table stakes. The lines do not all run in that order: a statute can move L2 with almost no L1 behind it.
Why now: Signing a line that says 'a human reviewed it' is no longer enough. Courts and bars increasingly expect a documented verification process: a record of what was checked, how, and by whom.
What to put in place: Logging, trace analysis, evaluation records as the compliance artifact.
How it works
Every reading comes from primary sources — court orders, statutes, bar opinions — weighted by who said it rather than by how many said it. A vendor blog carries a hundred and twenty-fifth the weight of a bar opinion, so it takes 125 of them to outweigh one ruling. Before anything is admitted, three independent reviews try to break the claim; what does not survive is dropped rather than softened. Every number here links to the sources behind it, so you can check the working instead of trusting the summary.
Why the advice is useful
This is a duty map, not a forecast. It reports what the law has already moved on, so acting on it is compliance rather than a bet on a prediction. The order is a sort for attention across controls you cannot all stand up at once, and it is graded against its own record rather than asserted. Where the evidence is thin, it says so; where a control is not required yet, this page does not pretend otherwise.
Why your fee structure changes the answer
The same control costs different firms different amounts, and whether it pays to put AI on a piece of work depends entirely on how you bill. On hourly billing, AI that cuts time cuts your revenue and the saving goes to the client, so adopting it is a net loss however good the tool is. On a fixed fee the same tool is margin. So the useful question is never “should we adopt AI” but “which work is worth it at our fee structure, and which is worth deferring” — and that answer is firm-specific, which is why the advisory layer asks for your pricing model before it tells you what to do.
Fault line: Signing a line that says 'a human reviewed it' is no longer enough. Courts and bars increasingly expect a documented verification process: a record of what was checked, how, and by whom.
Driver: Multi-step agentic work no single human fully traces.
Control (build now): Documented verification + trace logs — Logging, trace analysis, evaluation records as the compliance artifact.
Model Rules: 5.1 · 5.3 | L1 capability seed 5.5 → 7.8 (3 demos, w=4.02) | L2 ruling seed 6.2 → 9.7 (28 items, w=30.01) | L3 adoption seed 6.5 → 6.5 (0 signals, w=0.0) | E enablement seed 6.5 → 8.4 (2 items, w=2.47) | S software seed 5.5 → 5.6 (3 moves, w=0.05) | queue 6.3 · lead ~2 quarters
| Date | Demo | L1 capability evidence (weighted low, labeled — not a ruling) |
|---|---|---|
| 2026-05-15 | benchmark | Etymolt trademark-clearance hallucination benchmark data w=1.25 |
| 2024-05-23 | benchmark | Stanford RegLab 'Hallucination-Free?' — RAG legal research tools still hallucinate 17–34% data w=1.25 |
| 2025-10-15 | study | JURIX 2025: frontier LLMs era-biased and citation-hallucinating (GPT-5 35.29% pre-1881) data w=1.0 |
| Date | Tier | L2 ruling evidence (provenance) |
|---|---|---|
| 2026-06-01 | T1 | Withers v. Aberdeen — both sides sanctioned, four lawyers removed for AI hallucinations data N.D. Miss. (court order) · w=1.25 |
| 2026-04-15 | T1 | Nebraska Supreme Court suspends attorney over hallucinated appeal brief data Nebraska Supreme Court · w=1.25 |
| 2026-04-06 | T1 | Gamez v. County of Fresno (E.D. Cal.) — OSC names legal-AI product OpenCase E.D. Cal. (1:26-cv-00297, Mag. Judge Grosjean) · w=1.0 |
| 2026-03-13 | T1 | Whiting v. City of Athens (6th Cir.) — two lawyers $15k each + double costs + referral data Sixth Circuit (Nos. 24-5918/5919, 25-5424) · w=1.25 |
| 2026-02-09 | T1 | Amarsingh v. Frontier Airlines (10th Cir.) — $1,000 FRAP 38 + bar referral data Tenth Circuit (No. 24-01391) · w=1.25 |
| 2025-12-15 | T1 | Couvrette v. Wisnovsky — ~$110K penalty for 15 fake cases, 8 fabricated quotes data Court order · w=1.25 |
| 2025-11-17 | T1 | Schlichter v. Kennedy — every citation must be checked California Court of Appeal · w=1.0 |
| 2025-10-29 | T1 | Mezu v. Mezu (Md. App.) — AGC referral for non-lawyer clerk AI delegation Appellate Court of Maryland (No. 361 Sept. Term 2025) · w=1.0 |
| 2025-10-03 | T1 | People v. Alvarez — AI verification duty, candor to the tribunal California Court of Appeal · w=1.0 |
| 2025-07-28 | T1 | AP Electric, Inc. v. Oakland Automation, LLC — fabricated quotes on real cites District Court, E.D. Michigan · w=1.0 |
| 2025-07-23 | T1 | Elizondo v. City of Laredo — nondelegable Rule 11 verification duty District Court, S.D. Texas · w=1.0 |
| 2025-07-18 | T1 | California Rule of Court 10.430 — courts must adopt generative-AI policies Judicial Council of California · w=1.0 |
| 2025-05-13 | T1 | Ramirez v. El Tri MX Restaurant & Bar Corp — $1,000 sanction for unverified AI cites District Court, E.D. New York · w=1.0 |
| 2024-01-30 | T1 | Park v. Kim (2d Cir.) — referral for AI-produced nonexistent case Second Circuit (91 F.4th 610) · w=1.0 |
| 2023-06-22 | T1 | Mata v. Avianca — Rule 11 sanctions for fabricated ChatGPT citations data S.D.N.Y. (Judge Castel) · w=1.25 |
| 2026-08-31 | T2 | Federal court excludes expert for AI-hallucinated citations data Federal court / WSBA · w=1.0 |
| 2026-03-13 | T2 | California COPRAC — AI duties woven into six Rules of Professional Conduct California State Bar (COPRAC) · w=0.8 |
| 2026-03-01 | T2 | Alabama State Bar Public Reprimand of Matthew Reeves (Butler Snow) Alabama State Bar Disciplinary Commission · w=0.8 |
| 2024-07-29 | T2 | ABA Formal Opinion 512 — Generative AI Tools ABA Standing Committee on Ethics · w=0.8 |
| 2023-12-12 | T2 | England & Wales judiciary guidance — first formal AI guidance for judges Courts and Tribunals Judiciary (UK) · w=0.8 |
| 2026-04-01 | T3 | AI sanction wave: ~$145K in Q1 2026 penalties data EDRM / ComplexDiscovery · w=0.391 |
| 2025-10-15 | T3 | JURIX 2025: frontier LLMs era-biased and citation-hallucinating (GPT-5 35.29% pre-1881) data arXiv 2510.20941 · w=0.205 |
| 2025-09-01 | T3 | E-discovery and contract review are the most mature, most-trusted legal-AI uses, now at scale data Lighthouse / FTI Consulting (adoption reporting) · w=0.173 |
| 2025-01-29 | T3 | Paxton AI $22M Series A, Unusual Ventures, total $28M (research/drafting) data legaltechnology.com · w=0.076 |
| 2024-05-23 | T3 | Stanford RegLab 'Hallucination-Free?' — RAG legal research tools still hallucinate 17–34% data Magesh, Surani, Dahl, Suzgun, Manning & Ho (Stanford RegLab) · w=0.029 |
| 2023-06-26 | T3 | Thomson Reuters agrees to acquire Casetext $650M cash (closed 2023-08-17) data Thomson Reuters · w=0.008 |
| 2026-05-15 | T5 | Etymolt trademark-clearance hallucination benchmark data Etymolt (GitHub) · w=0.006 |
| 2024-09-26 | T5 | Relativity aiR for Review GA (genAI e-discovery) EDRM · w=0.0 |
| Date | Market | L3 adoption signal |
|---|---|---|
| No market-adoption signal yet — seed only. | ||
Fault line: Courts first punished fake AI citations after the fact (Mata v. Avianca was the first such sanction). Now more than 300 judges require lawyers to disclose AI use up front, and that patchwork of local orders is converging toward a single certification standard.
Driver: Patchwork of standing orders collapses into a uniform certification.
Control (build now): Verification-by-default with an audit trail — Verification-by-default with an auditable trail, so certification is a byproduct.
Model Rules: 3.3 · Rule 11 | L1 capability seed 5.5 → 5.5 (0 demos, w=0.0) | L2 ruling seed 6.4 → 9.5 (15 items, w=18.63) | L3 adoption seed 6.0 → 6.0 (0 signals, w=0.0) | E enablement seed 6.0 → 7.2 (1 items, w=0.9) | S software seed 5.5 → 5.5 (0 moves, w=0.0) | queue 5.7 · lead now
| Date | Demo | L1 capability evidence (weighted low, labeled — not a ruling) |
|---|---|---|
| No capability demonstration on record — seed only. | ||
| Date | Tier | L2 ruling evidence (provenance) |
|---|---|---|
| 2026-06-01 | T1 | Withers v. Aberdeen — both sides sanctioned, four lawyers removed for AI hallucinations data N.D. Miss. (court order) · w=1.25 |
| 2026-04-15 | T1 | Nebraska Supreme Court suspends attorney over hallucinated appeal brief data Nebraska Supreme Court · w=1.25 |
| 2026-03-25 | T1 | Elilton Alves Gouveia v. Meridian Financial Investments, LLC — prophylactic AI disclosure for pro se litigants District Court of Appeal of Florida · w=1.0 |
| 2025-12-15 | T1 | Couvrette v. Wisnovsky — ~$110K penalty for 15 fake cases, 8 fabricated quotes data Court order · w=1.25 |
| 2025-12-11 | T1 | In Re Adoption of Arkansas Supreme Court Administrative Order No. 25 — generative AI rule Supreme Court of Arkansas · w=1.0 |
| 2025-07-18 | T1 | California Rule of Court 10.430 — courts must adopt generative-AI policies Judicial Council of California · w=1.0 |
| 2025-04-04 | T1 | Daniel Jaiyong An v. Archblock, Inc. — GenAI certification ordered for future filings Court of Chancery of Delaware · w=1.0 |
| 2025-03-12 | T1 | Lillard v. Offit Kurman, P.A. — GenAI certification page ordered Superior Court of Delaware · w=1.0 |
| 2023-07-19 | T1 | Ex Parte Allen Michael Lee v. the State of Texas — AI filing certification (Judge Starr standing order) Court of Appeals of Texas · w=1.0 |
| 2023-06-22 | T1 | Mata v. Avianca — Rule 11 sanctions for fabricated ChatGPT citations data S.D.N.Y. (Judge Castel) · w=1.25 |
| 2026-05-14 | T2 | California updated Practical Guidance for Generative AI California Bar Board of Trustees · w=0.8 |
| 2026-03-13 | T2 | California COPRAC — AI duties woven into six Rules of Professional Conduct California State Bar (COPRAC) · w=0.8 |
| 2025-01-01 | T2 | Illinois Supreme Court AI policy — disclosure of AI use NOT required in pleadings Illinois Supreme Court · w=-0.8 |
| 2023-12-12 | T2 | England & Wales judiciary guidance — first formal AI guidance for judges Courts and Tribunals Judiciary (UK) · w=0.8 |
| 2026-04-01 | T3 | AI sanction wave: ~$145K in Q1 2026 penalties data EDRM / ComplexDiscovery · w=0.391 |
| Date | Market | L3 adoption signal |
|---|---|---|
| No market-adoption signal yet — seed only. | ||
Fault line: AI tools that learn from what you feed them, or mix data across matters, put client confidentiality at risk. Firms are being pushed to map where client data goes and get written no-training guarantees from vendors before a tool is approved.
Driver: No-training clauses become an ethical default; vendor attestations required.
Control (build now): Data-flow mapping + vendor attestation — Data-flow mapping and vendor attestation at tool onboarding.
Model Rules: 1.6 · 1.7 · 1.9 | L1 capability seed 4.5 → 4.5 (0 demos, w=0.0) | L2 ruling seed 6.0 → 9.4 (14 items, w=15.95) | L3 adoption seed 6.0 → 6.0 (0 signals, w=0.0) | E enablement seed 5.5 → 6.6 (1 items, w=0.65) | S software seed 5.5 → 5.5 (0 moves, w=0.0) | queue 5.6 · lead now
| Date | Demo | L1 capability evidence (weighted low, labeled — not a ruling) |
|---|---|---|
| No capability demonstration on record — seed only. | ||
| Date | Tier | L2 ruling evidence (provenance) |
|---|---|---|
| 2025-12-02 | T1 | In Re TikTok Inc. v. State of Texas — AI no-training protective order Court of Appeals of Texas · w=1.0 |
| 2025-07-18 | T1 | California Rule of Court 10.430 — courts must adopt generative-AI policies Judicial Council of California · w=1.0 |
| 2024-08-29 | T1 | In Re Amendments to Rules Regulating The Florida Bar - Chapter 4 — generative AI confidentiality Supreme Court of Florida · w=1.0 |
| 2026-06-10 | T2 | Derek Cirilo v. Resorts World Las Vegas — public AI platform restriction District Court, D. Nevada · w=0.8 |
| 2026-06-04 | T2 | Wilmington Trust v. Kingsbridge Realty — closed-AI-only protective order District Court, S.D. New York · w=0.8 |
| 2026-05-14 | T2 | California updated Practical Guidance for Generative AI California Bar Board of Trustees · w=0.8 |
| 2026-02-15 | T2 | New York Formal Opinion 2025-6 — recording/transcribing client meetings with AI New York State Bar · w=0.8 |
| 2025-02-01 | T2 | Texas Ethics Opinion 705 on generative AI Texas State Bar Professional Ethics Committee · w=0.8 |
| 2025-01-01 | T2 | Illinois Supreme Court AI policy — disclosure of AI use NOT required in pleadings Illinois Supreme Court · w=0.8 |
| 2024-07-29 | T2 | ABA Formal Opinion 512 — Generative AI Tools ABA Standing Committee on Ethics · w=0.8 |
| 2024-04-25 | T2 | Missouri Informal Opinion 2024-11 — pre-ABA generative-AI guidance Office of Legal Ethics Counsel, Supreme Court of Missouri · w=0.8 |
| 2024-01-19 | T2 | Florida Bar Ethics Opinion 24-1 — self-learning AI and client confidentiality The Florida Bar Board of Governors · w=0.8 |
| 2023-12-12 | T2 | England & Wales judiciary guidance — first formal AI guidance for judges Courts and Tribunals Judiciary (UK) · w=0.8 |
| 2025-06-01 | T3 | Enterprise legal-AI vendors bind no-training and ship data-governance / attestation controls CaseClerk / Baker McKenzie (vendor data-governance practice) · w=0.097 |
| Date | Market | L3 adoption signal |
|---|---|---|
| No market-adoption signal yet — seed only. | ||
Fault line: The duty is shifting from 'understand the tool you use' to 'run a governance program' — training, written policies, and oversight — with mandatory AI continuing-education requirements spreading state by state.
Driver: Mandatory AI CLE spreading; firm-level governance capability expected.
Control (build now): Firm-wide training + governance program — Training and enablement at scale (iTrain bet; AI Build Lab proof).
Model Rules: 1.1 | L1 capability seed 2.5 → 5.8 (5 demos, w=3.22) | L2 ruling seed 6.0 → 9.3 (18 items, w=15.37) | L3 adoption seed 5.5 → 5.5 (0 signals, w=0.0) | E enablement seed 6.0 → 7.9 (2 items, w=1.8) | S software seed 5.0 → 5.0 (0 moves, w=0.0) | queue 5.1 · lead ~1 year
| Date | Demo | L1 capability evidence (weighted low, labeled — not a ruling) |
|---|---|---|
| 2025-04-02 | study | LRAGE: legal RAG gains are corpus/backbone-dependent, not monotonic data w=1.0 |
| 2026-03-13 | model_access | Claude 1M context GA (Opus 4.6/Sonnet 4.6) w=0.45 |
| 2025-08-12 | model_access | Claude 1M-token context public beta (Sonnet 4) w=0.45 |
| 2024-03-04 | model_access | Claude 3: first Claude vision/document input w=0.45 |
| 2023-11-21 | model_access | Claude 2.1: 200K context + beta tool use w=0.45 |
| Date | Tier | L2 ruling evidence (provenance) |
|---|---|---|
| 2026-08-04 | T1 | TOV Realty, LLC v. Suarez — AI hallucinations violate Rule 1.1 competence Supreme Court of Connecticut · w=1.0 |
| 2026-07-10 | T1 | Estate of Lane Caviness v. Atlas Air — outsourcing to AI is not competent lawyering Court of Appeals for the Eleventh Circuit · w=1.0 |
| 2026-04-03 | T1 | United States v. John Farris — baseline AI ethics duties Court of Appeals for the Sixth Circuit · w=1.0 |
| 2025-12-11 | T1 | In Re Amendments to Arkansas RPC — AI competence under Rule 5.3 Supreme Court of Arkansas · w=1.0 |
| 2025-10-29 | T1 | Mezu v. Mezu (Md. App.) — AGC referral for non-lawyer clerk AI delegation Appellate Court of Maryland (No. 361 Sept. Term 2025) · w=1.0 |
| 2026-05-14 | T2 | California updated Practical Guidance for Generative AI California Bar Board of Trustees · w=0.8 |
| 2026-03-13 | T2 | California COPRAC — AI duties woven into six Rules of Professional Conduct California State Bar (COPRAC) · w=0.8 |
| 2026-03-01 | T2 | Alabama State Bar Public Reprimand of Matthew Reeves (Butler Snow) Alabama State Bar Disciplinary Commission · w=0.8 |
| 2025-04-01 | T2 | New Jersey requires technology CLE including AI — competence as a program requirement New Jersey Supreme Court · w=0.8 |
| 2025-02-01 | T2 | Texas Ethics Opinion 705 on generative AI Texas State Bar Professional Ethics Committee · w=0.8 |
| 2024-07-29 | T2 | ABA Formal Opinion 512 — Generative AI Tools ABA Standing Committee on Ethics · w=0.8 |
| 2024-04-25 | T2 | Missouri Informal Opinion 2024-11 — pre-ABA generative-AI guidance Office of Legal Ethics Counsel, Supreme Court of Missouri · w=0.8 |
| 2026-05-05 | T3 | PLI launches the first legal-AI competency framework Practising Law Institute · w=0.357 |
| 2025-04-02 | T3 | LRAGE: legal RAG gains are corpus/backbone-dependent, not monotonic data arXiv 2504.01840 · w=0.096 |
| 2026-03-13 | T5 | Claude 1M context GA (Opus 4.6/Sonnet 4.6) Anthropic · w=0.004 |
| 2025-08-12 | T5 | Claude 1M-token context public beta (Sonnet 4) Anthropic · w=0.002 |
| 2024-03-04 | T5 | Claude 3: first Claude vision/document input Anthropic · w=0.0 |
| 2023-11-21 | T5 | Claude 2.1: 200K context + beta tool use Anthropic · w=0.0 |
| Date | Market | L3 adoption signal |
|---|---|---|
| No market-adoption signal yet — seed only. | ||
Fault line: The rules are a patchwork — the EU AI Act, Colorado, Texas, 35-plus state bars, 300-plus judges — each slightly different. The patchwork itself is the pressure: a firm working across jurisdictions needs one operating model built to the strictest standard.
Driver: Firms adopt one framework calibrated to the strictest standard.
Control (build now): One operating model, to the strictest standard — The unifying operating model. This is the whole thesis.
Model Rules: 1.1 · 1.6 | L1 capability seed 2.5 → 2.5 (0 demos, w=0.0) | L2 ruling seed 6.2 → 9.3 (12 items, w=14.5) | L3 adoption seed 6.0 → 6.0 (0 signals, w=0.0) | E enablement seed 4.0 → 4.0 (0 items, w=0.0) | S software seed 5.0 → 5.5 (3 moves, w=0.27) | queue 5.6 · lead now
| Date | Demo | L1 capability evidence (weighted low, labeled — not a ruling) |
|---|---|---|
| No capability demonstration on record — seed only. | ||
| Date | Tier | L2 ruling evidence (provenance) |
|---|---|---|
| 2026-08-01 | T1 | EU AI Act obligations take effect European Union (regulation) · w=1.0 |
| 2026-06-30 | T1 | Colorado AI Act — consumer disclosure, Colo. Rev. Stat. § 6-1-1704 Colorado Legislature · w=1.0 |
| 2026-05-14 | T1 | Colorado repeals and replaces its AI Act (SB 26-189) — narrower transparency law Colorado (statute) · w=1.0 |
| 2026-05-07 | T1 | EU Digital Omnibus: Annex III high-risk deferred to 2026-12-02, Annex I to 2028-08-02 EU (licentium.io) · w=1.0 |
| 2026-02-01 | T1 | Colorado AI Act — NIST/ISO 42001 affirmative defense, § 6-1-1706 Colorado Legislature · w=1.0 |
| 2026-01-01 | T1 | Texas Responsible AI Governance Act (TRAIGA) takes effect Texas (statute) · w=1.0 |
| 2025-07-18 | T1 | California Rule of Court 10.430 — courts must adopt generative-AI policies Judicial Council of California · w=1.0 |
| 2024-12-08 | T1 | EU Product Liability Directive (EU) 2024/2853 — software and AI are 'products' European Union (Directive) · w=1.0 |
| 2024-05-01 | T1 | Utah AI Policy Act — AI liability, Utah Code § 13-77-102 Utah Legislature · w=1.0 |
| 2021-01-01 | T1 | NIST AI Standards — 15 U.S.C. § 278h-1 United States Congress · w=1.0 |
| 2025-11-10 | T3 | Clio closes vLex ($1B) + $500M Series G (NEA) at $5B data ABA Journal · w=0.226 |
| 2025-06-30 | T3 | Clio signs definitive agreement to acquire vLex ($1B) data Clio · w=0.136 |
| Date | Market | L3 adoption signal |
|---|---|---|
| No market-adoption signal yet — seed only. | ||
Fault line: Today the lawyer carries all the risk when an AI tool fails; there is no clear legal path to hold the vendor responsible. Pressure is building, through disclosure mandates and product-liability proposals, to shift some of that liability onto the toolmakers.
Driver: Pressure toward disclosure mandates and certification standards on vendors.
Control (build now): Tool-provenance documentation — Track for a shift; document tool provenance now.
Model Rules: — | L1 capability seed 2.5 → 2.5 (0 demos, w=0.0) | L2 ruling seed 3.5 → 8.3 (10 items, w=8.41) | L3 adoption seed 2.5 → 3.3 (2 signals, w=0.22) | E enablement seed 3.0 → 3.0 (0 items, w=0.0) | S software seed 3.5 → 3.8 (3 moves, w=0.13) | queue 2.7 · lead later
| Date | Demo | L1 capability evidence (weighted low, labeled — not a ruling) |
|---|---|---|
| No capability demonstration on record — seed only. | ||
| Date | Tier | L2 ruling evidence (provenance) |
|---|---|---|
| 2026-05-07 | T1 | EU Digital Omnibus: Annex III high-risk deferred to 2026-12-02, Annex I to 2028-08-02 EU (licentium.io) · w=1.0 |
| 2026-04-06 | T1 | Gamez v. County of Fresno (E.D. Cal.) — OSC names legal-AI product OpenCase E.D. Cal. (1:26-cv-00297, Mag. Judge Grosjean) · w=1.0 |
| 2026-01-01 | T1 | California AI Transparency Act — provider disclosure, § 22757.3 California Legislature · w=1.0 |
| 2025-07-01 | T1 | Wyoming AI developer liability immunity — Wyo. Stat. § 1-1-143 Wyoming Legislature · w=1.0 |
| 2024-12-08 | T1 | EU Product Liability Directive (EU) 2024/2853 — software and AI are 'products' European Union (Directive) · w=1.0 |
| 2026-05-19 | T2 | EU Commission draft: agentic/coordinating systems = single high-risk system (anti-splitting) arthurcox.com · w=0.8 |
| 2026-03-04 | T3 | Nippon Life v. OpenAI — first AI-vendor-liability lawsuit N.D. Ill. (No. 1:26-cv-02448) · w=0.281 |
| 2025-09-29 | T3 | AI LEAD Act (S.2937) — proposed federal AI product liability standards U.S. Senate (proposed) — Sens. Durbin & Hawley · w=0.154 |
| 2026-05-26 | T4 | Law review proposes strict product liability for legal-AI vendors Litle/Morris/Das; Kildeev (law review) · w=0.226 |
| 2025-08-08 | T4 | Ebers 'LawGPT': UPL-by-LLM a live comparative question Cambridge Handbook of Generative AI and the Law · w=0.074 |
| Date | Market | L3 adoption signal |
|---|---|---|
| 2026-05-26 | commentary | Law review proposes strict product liability for legal-AI vendors w=0.129 |
| 2026-03-04 | commentary | Nippon Life v. OpenAI — first AI-vendor-liability lawsuit w=0.094 |
The engine tracks a cascade: L1 capability (AI can now do it) → L2 ruling (a court/bar moves) → L3 adoption (the control becomes table stakes). Each order depends on the one before it. Point-in-time backtest: replay the engine 90 days before each event, using only evidence available then; "called" = the right meter was already ≥ 8.0. Overall 44% (4/9) · 5 snapshots on record.
By order: L1 Capability 33% (1/3) · L2 Ruling 100% (3/3) · L3 Adoption 0% (0/3).
Cascade conditionals — each layer only counts if the one it depends on held: P(L2|L1) n/a (0/0) · P(L3|L2) 0% (0/1) . A called L1 with no L2 event yet means the capability outran the law (reported, not scored as a fail). The seed set is small and the L1 lane is young — its first-of-kind capabilities have no precursor to call from — so this is the honest record so far.
| Landed | Order | Result | Fault line | Reading (lead→event) | Resolution |
|---|---|---|---|---|---|
| 2025-12-15 | L2 | called | disclosure | 9.0 → 9.2 | Couvrette v. Wisnovsky ~$110K sanction landed |
| 2026-04-15 | L2 | called | verification | 9.4 → 9.6 | Nebraska Supreme Court suspension over hallucinated brief |
| 2026-08-01 | L2 | called | convergence | 9.0 → 9.3 | EU AI Act obligations took effect |
| 2026-06-15 | L3 | missed | insurance | 6.5 → 8.4 | CNA makes AI governance a coverage condition (adoption) |
| 2026-08-31 | L3 | missed | verification | 6.5 → 6.5 | Verification duty extended to expert work product (adoption) |
| 2026-08-29 | L3 | missed | agentic | 5.3 → 6.4 | AI AGENT Act codifies scope-limited, logged agent authorization (adoption) |
| 2024-01-11 | L1 | missed | benchmark | 6.5 → 7.3 | Legal LLM hallucination capability demonstrated at scale (Stanford/Yale, 58-88%) |
| 2024-05-23 | L1 | missed | verification | 5.5 → 6.6 | 'A human reviewed it' shown insufficient — RAG legal tools produce confident wrong cites |
| 2026-06-01 | L1 | called | benchmark | 8.2 → 8.4 | Purpose-built legal AI tools quantified at 17-34% error (Westlaw/Lexis) |