What changed, and when we last looked

EU AI Act October update held: the first enforcement step has no official record yet

The monthly EU AI Act update was drafted and fact-checked. Its fixes were applied: the strand counts, the Joint Opinion date and the Article 101 wording. Its main event, the AI Office's first Requests for Information to general-purpose AI providers (29 August to 1 September 2026), is so far documented only in press reports of Commission statements, so the update and its timeline row are held. The EU AI Act text was re-opened and stamped.

Decision. Publish when an official Commission source for the requests is found, or when the editor accepts press-attributed Commission statements.

2 recordsby session

Florida fee-sanction news brief published; Q3 incident report updated to the full quarter

Published a news brief on Kid International v. City National Bank, in which Florida's Fourth District let a party's appellate-fee motion sanction AI-fabricated citations. Updated the Q3 2026 incident report in place to cover 1 July to 30 September: 87 incidents, 61 in US courts. The superseded 3 September figures are kept in a dated note. Both passed an independent claim-by-claim fact-check; the report's figures were recomputed from the tracker data by a separate script.

2 recordsKid International opinion (CourtListener) by session

Three incidents added from 7 October appellate rulings in Florida and Iowa

Added Kid International v. City National Bank (Fla. 4th DCA: appellants and counsel jointly liable for appellate fees after AI-fabricated citations), Washington v. Justice (Fla. 2d DCA: appeal dismissed with a pre-filing bar) and Tapscott v. Fariss (Iowa Ct. App.: appeal dismissed). Each was read from the court's opinion. Neill-Fraser and Valu were re-checked and stay unstamped because no permitted route reaches the judgments.

Decision. Not added: AA v SSHD (AI translation of a witness letter, no misuse finding). Auron Hoxha [2026] UKUT 378 (IAC) is guidance and was queued for the Regulation Tracker.

5 re-opened5 recordsKid International opinion (CourtListener) by session

Regulation sweep: 20 documents re-opened, none revised; Queensland judicial guidelines corrected

Re-opened 20 in-force documents from the UAE, Argentina, Austria, Australia, Belgium and Bahrain against their official URLs; all unchanged and stamped. The Queensland judicial-officer guidelines were found at a new official URL, read in full and corrected from placeholder provisions to the seven published guidelines. 35 country cells stamped (AE, AR, AT, AU, BE, BH); no status changed. The discovery sweep found no new official document since 5 October.

Decision. Eight documents stay unstamped because their official sites block automated access (ABA 512, NJ, FCFCOA, Arizona, CEPEJ, Florida 24-1, two NY UCS); archive copies show no change. No provisional country was re-opened, so none is ready for graduation.

21 re-opened6 recordsQueensland Courts judicial guidelines by session

Ten tools rechecked, LawPavilion to LexLegis AI; no facts downgraded

Re-opened trust, privacy and terms pages for LawPavilion, Leah (ContractPodAi), LEAP Matter AI, LegalFly, LegalZoom AI, Lex Machina, LexCheck, Lexion, Lexis+ AI and LexLegis AI. Leah now names its model providers and states zero retention. LegalFly publishes its sub-processors. LexLegis AI now publishes a SOC 2 Type 2 report. Lexion gained sourced SSO and encryption-at-rest facts. No verified flag changed.

Decision. Lexion stays below the threshold: its DPA and residency facts rest on a PDF that could not be parsed. Lex Machina and Lexis+ AI facts were re-sourced to the LexisNexis Trust Center.

10 re-opened10 recordsby session

Five pieces published: SB 574 news, two explainers, a sandbox report and issue 41

Published a news brief on California SB 574 (Chapter 858), explainers on client-disclosure and AI-training duties for lawyers, a 6,832-word report on legal-services sandboxes, authorisations and enforcement, and weekly briefing issue 41. Each passed an independent claim-by-claim fact-check against primary sources before publication.

Decision. Held from issue 41: the Argentine Morón item, because the official SCBA copy could not be opened. Seven earlier drafts remain held for human checks.

5 recordsSB 574 chaptered text by session

Nine incidents added, including a California Supreme Court State Bar referral

Added nine verified incident records from court documents read in full: a California Supreme Court referral to the State Bar, a Massachusetts fee cap of $10,000, two Nebraska and New York rulings, and the tracker's first record from Argentina. Re-opened the Neill-Fraser and Valu records; neither primary judgment could be retrieved, so neither was re-stamped.

Decision. Held: Neill-Fraser and Valu stay without a fresh lastVerified date until an official copy of each judgment can be opened.

11 re-opened11 recordsCervantes v. Bianco, S295866 Shannon v. Shannon (Neb.) by session

EU AI literacy duty now recorded in every member-state training cell

Article 4 of the EU AI Act, as replaced by Regulation (EU) 2026/1744, applies directly in all member states. Eight country records scored the training category on national instruments only; Croatia, Cyprus, Greece, Malta, Slovenia, Spain, Portugal and Italy now cite Article 4 and read binding, consistent with the other EU records. National notes are kept.

Decision. Editorial rule: a directly applicable EU regulation is scored in every member-state cell it governs, not only on the EU record.

8 recordsRegulation (EU) 2026/1744 by session

Regulation sweep: 30 documents re-opened, 2 revisions and 3 new records

Re-opened the 27 oldest in-force documents and three linked records; 22 re-stamped, 8 blocked by the publisher's site. Recorded the Brazil CNJ Resolution 615 amendment and the Queensland 2025 revision; added California SB 574 (Chapter 858) and the AI sections of the 2026 TCC and Administrative Court guides; corrected eight records against primary text; 28 country cells re-stamped.

Decision. Held: ABA 512, FCFCOA PD-AI, Arizona ACJA 1-509, CEPEJ, Florida 24-1, New Jersey and two New York UCS documents not re-stamped (site blocked or timed out). No provisional country graduated.

30 re-opened6 recordsCNJ Resolution 674/2026 SB 574 chaptered text by session

Ten tools rechecked; Juro, Jus AI and Lawpath now meet the four-fact threshold

Re-opened trust, privacy, DPA and terms pages for Josef, Juro, Jus AI, Kira, LawDroid Copilot, Lawfully, LawGeex, Lawhive, Lawmatics and Lawpath. Juro, Jus AI and Lawpath now have four sourced security facts. Kira certifications, the Josef DPA and Lawmatics private deployment reverted to unknown where current pages no longer support them. Directory: 62 verified, 69 below threshold.

Decision. Verification means sourced facts, not endorsement. LawGeex status and Lawhive's directory fit are queued for editor review.

10 re-opened10 recordsby session

Federal Circuit VA AI whistleblower brief published after independent fact-check

Published a news brief on the Federal Circuit decision reviving an MSPB whistleblower appeal about a VA AI claims tool. The ruling concerns jurisdiction only and makes no merits finding. Independent review passed all 47 claims after a verified archive copy of the opinion was added. The owner approved publication.

Decision. Owner approved publication after re-verification. Fact-check flag true; draft false.

1 re-opened1 recordFederal Circuit opinion by session

One incident added and six records rechecked

Added Posey v Bushnell from the court's filed opinion. Re-read primary judgment text for five stale records and promoted R v Chand to verified after reading its judgment. Neill-Fraser and Valu remain unstamped because primary text was not accessible. Tracker now holds 637 records: 153 verified, 8 unverified, 476 provisional.

Decision. Stamps applied only where judgment text was actually re-read.

8 re-opened6 recordsPosey v Bushnell opinion by session

Fifteen regulation records rechecked, seven now verified

Re-read primary text for 15 regulation documents and marked seven verified, including NC Formal Ethics Opinion 1, Texas Opinion 705 and USPTO AI guidance. Corrected the Manitoba guideline title, the Newfoundland and Labrador date and the EU Article 4 wording. Timestamped 15 supported country cells. No new instrument found; California SB 574 remains pending. Six blocked sources, including ABA Formal Opinion 512, were left unstamped.

Decision. No country-wide re-verification or provisional graduation; 54 provisional countries still await editor review.

15 re-opened15 recordsby session

Ten tools rechecked against current vendor documentation

Rechecked Filevine AI, Garfield, Gemini for Workspace, Genie AI, Hebbia, Icertis, iManage AI, IPRally, Ironclad and Jhana against current official vendor pages. Icertis, iManage and IPRally now meet the four-sourced-fact threshold; Ironclad's ISO claim was narrowed and its training terms marked conditional. Directory now lists 131 tools; 71 third-party tools remain below the threshold.

Decision. Verification means sourced facts, not endorsement.

10 re-opened10 recordsby session

Hancox news brief published after independent fact-check

Published a news brief on the EAT's warning over an unchecked ChatGPT skeleton argument, separating the September 17 judgment from September 18 reporting and comparing the operative page-length rules. Independent review passed all 38 claims; the owner authorized publication. The September 16 US court-instruments revision meets this week's research commitment.

Decision. Owner authorized production publication after validation. Fact-check flag true; draft flag false. No duplicate research report commissioned merely to satisfy the original-publication-date cadence counter.

1 re-opened1 recordHancox judgment EAT Practice Direction 2024 by session

One incident added and twelve records rechecked

Added Hancox v Sutherland from the Employment Appeal Tribunal judgment. Reopened twelve existing incidents, promoted Parker and Specter, and corrected unsupported tool attribution, costs and referral descriptions. Three attempted records remain unstamped because primary text was unavailable. No new regulator disposition was verified.

Decision. Neill-Fraser, Chand and Valu remain unverified; Reddy's later costs order remains unconfirmed.

12 re-opened13 recordsHancox judgment by session

Week 39 briefing published after independent fact-check

Published the seven-item Monday briefing from the frozen September 16 baseline after owner authorization. Independent review passed all 42 factual and publication-gate rows after adding sources and verified court-document archives. This is web publication; no email send is claimed.

Decision. Owner authorized production publication after validation. Fact-check flag true; draft flag false. Buttondown email send remains manual.

1 re-opened1 recordFlorida opinion by session

Twelve regulation records and seven country cells rechecked

Rechecked twelve regulation records and seven cells across the UK, New Zealand and US. Clarified court-requested disclosure, recommended versus prohibited judicial tasks, California rule versus advisory-standard scope, and billing distinctions. SB574 has no new action in the reopened history. No new instrument, version or country graduation was verified.

Decision. Four limited checks and two blocked records retain their dates. All country-wide verification dates and 54 provisional flags remain unchanged.

12 re-opened12 recordsSB574 official history Ireland commencement by session

Ten tools rechecked against current vendor documents

Rechecked Docket Alarm, Docketwise, Docusign IAM, DoNotPay, DraftWise, Epiq AI, Eve, EvenUp, Everlaw and Evisort. Four records now meet the sourced-fact threshold; Docket Alarm no longer does. Corrected certification scope, customer-data training and provider-only retention claims. These are verification corrections, not claims of newly announced vendor changes.

Decision. Portal shells are not evidence; Docusign checkout pricing was not reverified. Verified denotes documentation coverage, not endorsement.

10 re-opened10 recordsDocusign certifications EvenUp privacy by session

Published: US court AI instruments report (2nd edition, 268 records), two briefs and a Florida news brief

Second edition of the US court AI instruments report replaces the July stub over the same URL (6,474 words; 268 state and federal records as of 16 Sep 2026; 18 instruments read in full). Briefs: 'Can lawyers put client data into AI tools?' (76 reviewed countries: 7 binding, 33 guidance) and the State Bar of California 2023 vs 2026 guidance diff (10 changes tabled). News: Florida 4th DCA show-cause order over suspected AI filings. Every piece passed a fresh-session fact-check after fixes (report: 14 count-drift rows, then Baylson URL).

Decision. Held unchanged: deepfake-evidence report (blocked primaries), three scorecards, methodology, vendor documentation audit and Harvey vs Legora (right-of-reply). Newsletter 2026-38 is due Monday 21 September.

4 recordsState Bar of California 2026 Practical Guidance by session

Incident sweep: Florida 4th DCA 'AI slop' show-cause order and a Tasmanian report added; 14 stale records re-opened; tracker at 635

Sweep of 10–16 September 2026 added Lisandrillo v. Palozzi (Fla. 4th DCA, 16 Sep 2026: petition denied, counsel ordered to show cause without AI why sanctions or a Florida Bar referral should not follow; no fake citations involved) and a press-reported Tasmanian parole matter (unverified). Fourteen of the fifteen oldest records were re-opened and stamped; Hussein v Canada verified via mirrored PDFs (CAD 100 costs); Parker v Forsyth remains blocked. No new regulator disposition found.

Decision. Neill-Fraser stays unverified until the Supreme Court of Tasmania judgment is on AustLII or the court site.

16 re-opened3 recordsLisandrillo v. Palozzi opinion by session

Regulation sweep: Ireland's AI Act commenced (S.I. 403/2026); California State Bar 2026 guidance; Fuentes order drops disclosure

S.I. No. 403/2026 commenced Ireland's Regulation of Artificial Intelligence Act 2026 on 31 July 2026 (status in-force). The State Bar of California's 2026 Practical Guidance replaced the November 2023 version (agentic AI added); recorded as a new version. Judge Fuentes's N.D. Ill. standing order (PDF dated 4 Mar 2026) no longer requires AI disclosure; version recorded, US cells updated. NYSBA Task Force report added. Starr, Delaware, E.D. Tex. and Baylson records now point at the operative documents. Tier-1/2 cells stamped.

Decision. BG, RW, BW and ME join KZ, RS, AZ, ET, MU as provisional records ready for editor graduation. California SB 574 remains with the Governor (presented 9 Sep; deadline 30 Sep). KI-MIG Wayback save still failing.

14 re-opened8 recordsS.I. No. 403/2026 by session

Legal tech: batch of 10 re-verified; Consilio Aurora, DeepIP and descrybe.ai verified; BriefCatch and Checkbox First Pass added

Re-opened the trust, privacy, terms and pricing pages of consilio-aurora, contractbook, counsel-stack, courtcorrect, darrow, deepip, deepjudge, definely, descrybe-ai and disco-cecilia on 16 September 2026. Consilio Aurora, DeepIP and descrybe.ai now meet the four-sourced-facts threshold; Contractbook (EU residency), CourtCorrect (SSO), Darrow (private deployment) and Definely (DPA) each lost a fact no current page supports. Added BriefCatch (unverified) and Checkbox First Pass (beta, verified).

Decision. Trust portals for CourtCorrect, DeepIP, Definely and DISCO block automated fetches; facts they alone support stay unknown until an editor opens them.

10 re-opened12 recordsConsilio Aurora announcement by session

Published: SDT strike-off news brief and weekly briefing 2026-37; deepfake report held

The SRA v Abhishek Kumar news brief (450 words) and 'This week in legal AI safety' 2026-37 (1,207 words) published after fresh-session fact-checks (brief: 2 fails fixed; briefing: 6 fails fixed, ALL-PASS on re-check). Dated update appended to the California SB 574 brief: presented to the Governor 9 Sep 2026. The deepfake-evidence report was reworked but two fresh fact-check passes still fail on sources the checker cannot open; it stays draft.

Decision. Held: deepfake-evidence-in-court-60-rulings-rule-707-status (editor must open the blocked CanLII/AustLII/Indian/French primaries in a browser); the three scorecards, methodology page, vendor documentation audit and Harvey vs Legora remain held for vendor right-of-reply.

3 recordsSDT case page 12884-2026 by session

Incident sweep: 7 rows added (SDT strike-off, 6 US courts), 14 stale records re-opened; tracker at 633

Sweep of 3–11 September 2026 sources added SRA v Abhishek Kumar (SDT, strike-off, 25 Aug 2026), Southland Homes v. Lam (Cal. App. Div., USD 4,000, State Bar notice), Asiedu v. Eastlund (Minn., USD 2,341.25), State ex rel. Schnell v. Richardson (Mo., USD 1,750), G.P. v. Arevalo, Falkenstein v. Niles and Carter v. City of Buffalo. Fourteen of the fifteen oldest records were re-opened and stamped; Hussein v Canada was not (CanLII bot wall). No new regulator disposition found.

Decision. Southland dated to the order filed 4 Aug 2026 (modified 5 Aug), not the 10 Sep posting date; Asiedu docket corrected to A25-2106/A26-0532 from the official Minnesota PDF after the fact-check.

15 re-opened7 recordsSDT judgment 12884-2026 by session

Regulation sweep: Germany, Poland and Ireland AI Act statutes added; 13 key documents re-opened, none changed

Filed from the official texts: Germany's KI-MIG (in force 29 Jul 2026), Poland's act of 3 July 2026 on AI systems (in force 11 Aug 2026) and Ireland's Regulation of Artificial Intelligence Act 2026 (enacted 21 Jul 2026; commences by ministerial order, recorded as 'enacted, not commenced'). Thirteen most-linked documents re-opened unchanged, from the SRA warning notice to the consolidated EU AI Act. Tier-1/2 country cells re-verified; DE, PL and IE cells updated.

Decision. Provisional country records AZ, ET and MU (and KZ, RS from 4 Sep) have had every source re-opened and are listed for editor graduation; the editor flips provisional: false. KI-MIG Wayback snapshot failed and needs a retry.

13 re-opened3 recordsGesetze im Internet — KI-MIG by session

Legal tech: batch of 10 re-verified; Casepoint, CoCounsel and Caseway verified, Clio Duo downgraded

Re-opened the trust, privacy, terms and pricing pages of casemine-amicus, casepoint, casetext-cocounsel, caseway, casium, chatgpt-enterprise, claude-for-enterprise, clearbrief, clio-duo and conga-clm on 11 September 2026. Casepoint (privately deployed LLMs stated on its product page), Thomson Reuters CoCounsel and Caseway now meet the four-sourced-facts threshold. Clio Duo fell below it after an unsupported private-deployment position reverted to unknown. ChatGPT Enterprise zero retention moved to no (workspace retention minimum 90 days per OpenAI); Claude for Enterprise records the 30-day covered-model retention caveat and US-only published residency. No new tools.

Decision. Unknown stays unknown: every downgrade is a fact no current vendor page supports, logged with the page checked.

10 re-opened10 recordsCasepoint generative AI page by session

Legal tech tools: 128 records get documentation prose, vendor quotes and diligence questions

Every independent tool record now carries a written account of what the vendor's public pages state, a list of what is not published where we looked, verbatim vendor quotations with the page they appear on (501 quotes; 446 taken from the legal-ai-vendor-facts audit rows, the rest checked against the live page) and questions a buyer would put to the vendor. Existing facts were not changed; disagreements found are logged for the editor. Verified tool pages need this prose to be indexable.

Decision. Same rule as the incident tracker after the 22 August demotion: a page is indexed only when it says something the vendor's own page does not, in sourced words.

128 re-openedby session

Incident tracker: 139 records written up from primary sources, 8 corrected, thin pages de-indexed

All 139 verified incident records gained the fabricated material as the court listed it, verbatim judgment passages with pinpoints (543 quotes), a dated timeline, a why-it-matters note and a practice note, fetched from the listed sources. Eight records were corrected against those sources (Coleman, Shelbert, Benjamin v State, Adibi, Chapman, In re BFI, Green v Imprint, Asif). Thin record pages, exhaustive tool comparisons and hubs under three items now render noindex; sitemap 908 to 815 URLs.

Decision. Response to the 22 August 2026 site-wide search demotion (Google's August 2026 spam update): index only pages that say something the primary source does not, and say it with quotes and pinpoints.

139 re-openedby session

Research report: AI chat logs, note-takers and privilege, the evidence record 2024 to 2026

Published the Wave 9 report collecting 17 rulings, orders and ethics opinions on whether AI conversations and transcripts are privileged, discoverable or admissible (Heppner, Warner, Morgan v V2X, the OpenAI preservation and 20-million-log production orders, Mendones, the NYC Bar opinions, In re Otter.AI), with a five-vendor table of retention and training terms. Fresh-session fact-check passed on the second pass.

1 recordby session

Research report: AI in court filings, the worldwide disclosure map

Published the second Wave 9 research report: 46 binding-form court instruments and 130 country records coded on one schema (prohibit, declare, certify, verify, silent). Six regulation records were corrected from the archived operative texts during the fact-check (New York Part 161's certification is an optional model rule; Florida's Rule 2.515 never mentions AI; Louisiana Act 250 has no certification; Oklahoma 1.17 verified; Florida 17th Circuit order superseded; GPN-AI paragraph 4.10 automatic disclosure). Fact-check passed on the third pass.

7 recordsby session

Research report: consumer AI legal services and self-represented litigants

Published the last Wave 9 report: twelve consumer legal-AI services read on one schema (operator, regulated status, disclaimer, training on user content, price), the 41 self-represented rows in the incident tracker with their outcomes and the five monetary orders, two federal docket studies of AI pro se filings, and the regulators' record (FTC v DoNotPay, Upsolve v James, Alabama FO 2026-01, New York S7263A, Nippon Life v OpenAI). Descriptive only; no service recommended. Fresh-session fact-check passed on the third pass.

1 recordby session

Research report: EU AI Act obligations for law firms and legal-AI vendors after the Omnibus

Published the Wave 9 report on the EU AI Act as it reaches legal practice: every deployer and provider duty by article, application date and enforcer after Regulation (EU) 2026/1744 (five in force, six deferred to 2 December 2027), the fine tiers, a 27-member-state enforcement-authority table and an enforcement register to 5 September 2026. Fresh-session fact-check passed on the second pass.

3 recordsby session

Research report: legal AI in India, demand, incidents and guidance

Published the Wave 9 India report: Semrush India database demand (202 keywords, head term 8,100/mo against 9,900 in the US), the nine Indian decisions on fabricated authority (seven with the fabrication in the adjudicator's own order; the first cost order, Rs 50,000, Bombay High Court 7 Jan 2026), the eight Indian instruments from the Kerala policy to the Gadiya judgment, and seven Indian legal-AI products on the audit schema. Fresh-session fact-check passed after one round of fixes.

2 recordsby session

Research report: the judges' record, courts' own AI rules and the eleven decisions built on fabricated authority

Published the Wave 9 report on the bench side of the record: 106 Regulation Tracker instruments addressed to judges, tribunal members or court staff (10 countries bind judges, 22 guide, 6 drafting; 7 US states binding) and the 11 tracker decisions in which the adjudicator relied on non-existent authority (7 in India). Fresh-session fact-check passed on the second pass; seven US-state regulation records and three incident records were corrected along the way.

4 recordsby session

Research report: the legal-AI search-demand record, September 2026

Published the first research report of Wave 9: what people search for around legal AI in the US and UK (Semrush, 5 Sep 2026), what the first page of results gives them, and where the unserved demand sits. Ten keyword sets (998 rows) published as CSV. Fresh-session fact-check passed on the third pass.

1 recordRaw CSVs (repository) by session

Research report: the sanctions ledger, 2023 to 2026, priced

Published the third research report of Wave 9: every court order over AI-fabricated authority in the tracker (150 rows, 22 Jun 2023 to 3 Sep 2026) coded for sanction type, amount, actor, tool and regulator disposition. 34 USD orders sum USD 371,542 (median 5,000); 48 referrals, five verified dispositions. Fresh-session fact-check passed on the fourth pass; six tracker records were re-read from primary documents along the way.

7 recordsby session

Research report: what regulators did after 48 AI-misuse referrals

Published the Wave 9 report following every court referral in the incident tracker to the regulator: 48 matters in seven countries; seven rows carrying five distinct dispositions verified from the regulator's own document (Victoria, the Texas to Oklahoma reciprocal chain, Ontario, USPTO, the Netherlands), four reported, 36 open; measured time to disposition 104 to 422 days; 76 bar and law-society AI guidance documents dated. Fresh-session fact-check passed on the second pass after a re-derived guidance count.

1 recordby session

Research report, second edition: how often legal AI hallucinates, the measurement register

Replaced the 16 July 2026 report with a second edition (same slug, pubDate kept, updatedDate 5 Sep 2026): a register of every published measurement on one schema, the vendor claims-versus-evidence audit over 47 verified tools, and the incident record as a lagging indicator. Fresh-session fact-check passed on the third pass.

2 recordsby session

Research report, second edition: legal AI by the numbers, thirteen surveys with the methods exposed

Replaced the 16 July 2026 numbers page with a second edition (same slug, pubDate kept, updatedDate 5 Sep 2026): a register of thirteen 2024 to 2026 surveys on one schema with sponsor, n, population, field dates and question wording, a comparability matrix and the governance gap triangulated. Fresh-session fact-check passed on the second pass.

1 recordby session

Two incident-file analyses drafted and held for fact-check: Sullivan & Cromwell, Lacey v. State Farm

Drafts written from the primary documents for the two most-searched US incidents the site did not cover. Both stay draft until a fresh fact-check session returns a clean claim table; the first fact-check attempts were cut short by an API rate limit and will be re-run.

Decision. Draft, do not publish: every article passes a fresh adversarial fact-check before the draft flag flips.

2 re-opened2 recordsby session

Four held drafts fact-checked and published; five held (scorecards await right of reply, comparison awaits Queensland text)

Fresh-session fact-checks of the 3 Sept drafts: the summer regulation review (Connecticut § 4-9(c) paraphrase fix verified), the Q3 incident report (actor counts, tool table totals and three descriptions corrected against the tracker), the September EU AI Act update (Greece item corrected) and weekly briefing 2026-36 (Mitchell, tools and update-count fixes) now publish. Held: the worldwide court-rules comparison until Queensland's PD 5 full text is opened, and the Harvey, Lexis+ AI and Microsoft Copilot scorecards plus the scorecard methodology until vendors have had five business days to reply.

Decision. Publish only on a clean claim table; scorecards never publish without the right-of-reply window, even when the facts check out.

54 re-opened4 recordsby session

Research report: 'AI in legal practice: what 130 country records require at September 2026' (7,900 words, 16 tables)

The week's deep research report, built from the regulatory map (130 records, 54 provisional kept as a separate tier), the document registry and the incident tracker. Every number recomputed by a fresh-session fact-checker (47 claim rows, 8 fail groups corrected before publication; 25 of 25 cited sources re-opened). Key findings: verification duties (50 records) far outnumber filing-disclosure rules (13); the EU AI Act supplies the horizontal statute for 27 members while 21 national instruments exist; provisional records mostly record absence.

Decision. Published after the independent recomputation matched the dataset cell by cell; the three most fragile claims (Digital Omnibus dates, CEPEJ characterisation, mid-2026 Ireland/India/CJC guidance) are flagged in the run log for the next check.

47 re-opened1 recordby session

Research report published: what 130 country records require at September 2026 (7,900 words, 16 tables)

The week's deep research report, built from the regulatory map (130 records, 54 provisional kept as a separate tier), the document registry and the incident tracker. Every number recomputed by a fresh-session fact-checker (47 claim rows, 8 fail groups corrected before publication; 25 of 25 cited sources re-opened). Key findings: verification duties (50 records) far outnumber filing-disclosure rules (13); the EU AI Act supplies the horizontal statute for 27 members while 21 national instruments exist; provisional records mostly record absence.

Decision. Published after the independent recomputation matched the dataset cell by cell; the three most fragile claims (Digital Omnibus dates, CEPEJ characterisation, mid-2026 Ireland/India/CJC guidance) are flagged in the run log for the next check.

47 re-opened1 recordby session

Two incident-file analyses published after fresh fact-checks: Sullivan & Cromwell and Lacey v. State Farm

Both drafts failed their first adversarial fact-check (nine and six rows: a mangled Westlaw number, an overstated correction schedule, two claims the tracker schema could not support, a mischaracterised United States v. Cohen, an unsourced vendor line, two false superlatives). Every failing row was fixed from the primary document and a second fresh session re-checked the amended passages before the draft flags flipped. The Harvey vs Legora comparison stays in draft.

Decision. Publish gate = clean claim table from a session that did not write the piece. Merge is the human step.

2 re-opened2 recordsFact-check re-check log by session

Two more news briefs published: D.C. Court of Appeals (Douglas) and Virginia Court of Appeals (LPC Commercial)

Douglas v Deutsche Bank National Trust: the D.C. Court of Appeals struck an appellee brief containing four non-existent authorities found through a generative AI search tool; a disciplinary referral is pending. LPC Commercial Services v Ar-Razzaaq Properties: Virginia's Court of Appeals found 25 phantom quotations and ordered AI and appellate-practice CLE. Both briefs checked against the CourtListener PDFs and dockets; unsupported wording tightened before publication.

Decision. Regulator follow-ups (D.C. ODC, Virginia State Bar) are on the watch list; the briefs say 'pending' and nothing more until a disposition is published.

2 re-opened2 recordsby session

Two news briefs and two explainers published after fresh-session fact-checks

News: the Supreme Court of India's judgment in Gadiya v Union of India (2026 INSC 947, 2 Sept) on AI-invented authorities in a customs order, and Indonesia's Supreme Court draft judgment-writing guidance banning AI-drafted reasoning (letter 112/WKMA.Y/HK2.1/IX/2026). Briefs: 'Must lawyers verify AI output before relying on it?' and 'May judges and court staff use AI in their own work?', both built from the country map with provisional records separated. 79 claims checked; corrections included the Rs 425.28 crore figure, the 14/18 September deadline wording and full country lists.

Decision. Indonesia's letter is described as draft guidance, not a binding rule, because the source says so; the country cell mirrors that.

79 re-opened4 recordsby session

Incident sweep: 15 new court records (25 Aug–4 Sep), Ko v Li regulatory outcome recorded, 15 stale records re-opened

New rows include the D.C. Court of Appeals striking a brief with four non-existent authorities (Douglas), Virginia's Court of Appeals finding 25 phantom quotes (LPC Commercial), a California appellate fee sanction with a State Bar referral (Varma), and warnings from Hawaii, Illinois, Pennsylvania, Colorado, Texas, Maryland, Michigan, New Hampshire and Australia's ART. Ko v Li now carries the Law Society of Ontario suspension with a dated correction. The 15 stalest records were re-opened and confirmed.

Decision. Every row links a judgment or order fetched this run; CanLII/AustLII bot-blocks were recorded as warnings, not bypassed. Douglas and Varma go on the regulator follow-up watch list.

15 re-opened16 recordsby session

Tracker: 13 US incidents added from primary court documents, 2023 to 2025 back-catalogue

Thirteen rows added after reading the order or opinion on govinfo, RECAP or an official court site: Lacey v. State Farm (Ellis George and K&L Gates, $31,100), Johnson v. Dunn (Butler Snow, public reprimand and bar referral), Mid Central v. HoosierVac ($6,000), Benjamin v. Costco ($1,000), Dehghani v. Castro ($1,500), Bevins v. Colgate-Palmolive, Thomas v. Pangburn, Morgan v. Community Against Violence, Dukuray v. Experian, United States v. Cohen, Ex parte Lee, Matter of Samuel and Smith v. Farwell (Mass. Super. Ct., $2,000). Every row is status verified.

Decision. Pre-2026 depth was the tracker's weakest point (44 rows before this run). Rows were chosen from the most-searched US cases and added only where the primary document was read.

13 re-opened13 recordsLacey v. State Farm, Order of Special Master, ECF 119 (RECAP) Johnson v. Dunn, Sanctions Order, Dkt. 204 (RECAP) by session

Sullivan & Cromwell / In re Prince Global Holdings: row verified from the docket, status flipped from unverified

The 18 April 2026 letter (ECF 25), the corrected motion with redline (ECF 27), Kobre & Kim's 21 April letter (ECF 36) and the 23 April provisional relief opinion (ECF 38) were read on RECAP. The row now records the withdrawal and correction, the 28-citation count from opposing counsel, and that no sanctions entry appears on the docket through 10 August 2026. The earlier [NEED: PACER primary] flag is resolved.

1 re-opened1 recordLetter to Chief Judge Glenn, ECF 25 (RECAP) Docket 26-10769 (Bankr. S.D.N.Y.) by session

Cookieless analytics, privacy notice, sitemap pruning and title clean-up

PostHog switched to cookieless mode (nothing stored on the device, no person profiles, DNT respected). New /privacy notice. Sitemap now excludes noindex pages and carries lastmod for 541 records; meta descriptions clamped; page titles shortened to the 'Thing · SafeLegalAI' pattern.

Decision. No cookie banner: nothing is stored on or read from the visitor's device, so PECR consent is not engaged.

by session

Regulation sweep: 10 stale documents re-opened and archived, no revisions; tier-1 cells confirmed; KZ and RS ready for graduation

Every in-force document older than 30 days was fetched and compared to its recorded version: all unchanged, archive.org snapshots refreshed, Kerala High Court dead homepage URL replaced. Discovery across the regulator list found no qualifying new official document for August. Tier-1 country cells (GB, US, EU, AU, CA, IE, SG, IN) re-opened and confirmed. Provisional Kazakhstan and Serbia records re-opened source by source and found consistent; Bulgaria held on an EUR-Lex fetch failure.

Decision. Provisional records stay provisional: KZ and RS are listed as ready for the editor's own source-by-source check; only the editor sets provisional: false.

10 re-opened10 recordsby session

Regulatory map: 54 provisional country records added (76 → 130 countries)

AI-researched records for Europe/CIS (14), Middle East and Africa (20) and the Americas/Asia (20). A status was set only where a fetched official or reputable source supports it; most cells are honestly 'nothing found'. Notable finds: Kazakhstan's AI law No. 230-VIII (17 Nov 2025), Bulgaria under the EU AI Act, Rwanda's judicial guidance. All 54 carry provisional: true until the editor re-opens every source.

Decision. Provisional records are published as leads, tagged everywhere, and read 'researched' rather than 'checked' until editor sign-off (editorial-standards#provisional).

54 re-opened54 recordsAdilet — Law of Kazakhstan On Artificial Intelligence No. 230-VIII EUR-Lex — Regulation (EU) 2024/1689 by session

Keyword registry, coverage check, six vendor hubs, tracker methodology page and title fixes

A machine-readable keyword registry now maps every target keyword to one URL, with a coverage script wired into CI. Six vendor hubs (Harvey, Legora, Anthropic Claude, OpenAI ChatGPT, Microsoft Copilot, Lexis+ AI) collect news, incidents and documented governance facts. A tracker methodology page compares coverage with Damien Charlotin's database (2,016 cases on 4 September 2026). Homepage, directory, news, statistics and category titles now carry the terms readers search for; tracker and analysis pages link each other both ways.

Decision. Semrush on 4 September 2026 showed 3 organic keywords and 0 AI citations; the registry-first system is the fix, not more articles.

Damien Charlotin, AI Hallucination Cases database by session

Site redesign v6: real-world regulatory map, datasets catalogue, legal tech record, trust pages

Full redesign (Geist type, ocean-teal brand, shield mark). The regulatory map is now a Miller-projection choropleth from Natural Earth data; the tools section became Legal Tech with vendor marks and a fact-glyph matrix; new /datasets, /report, /privacy, /disclaimer and /design/system pages; publisher-level bylines (SafeLegalAI); cookieless analytics.

Decision. Design rules are recorded in DESIGN.md v6 and the living style guide so weekly content stays on-system.

by session

Legal tech: weekly batch of 10 records re-opened against vendor pages; 3 facts upgraded, 5 downgraded to not disclosed

Re-opened trust, privacy, terms and sub-processor pages for the 10 oldest records. Alexi now documents SOC 2; Amazon Q Business gained SOC and ISO 42001 evidence and data-residency options and becomes verified; Adalat AI becomes verified. Bloomberg Law AI's no-training commitment could not be found on a current page and is now 'not disclosed'; four unsupported 'private deployment: no' facts were downgraded to not disclosed. No new tools; LegalAI Space skipped by rule.

Decision. A fact is 'yes' or 'no' only where a page fetched this run states it; where the page no longer says it, the record reverts to not disclosed rather than keeping a stale claim.

10 re-opened10 recordsby session

Harvey and Legora records re-opened: DPAs, sub-processor lists and model providers confirmed from vendor pages

Both records were re-read against the vendors' current legal pages while the Harvey vs Legora comparison was drafted. Harvey: DPA (2 August 2026) and sub-processor list (2 August 2026) now sourced; model providers recorded as OpenAI, Anthropic, Google, Microsoft, Mistral, Baseten and Fireworks. Legora: EU DPA (31 January 2025) and pre-approved sub-processor list (6 July 2026) sourced; providers Microsoft, AWS, Google and OpenAI. Legora's trust centre, read in a browser, shows SOC 2 Type II, ISO 27001:2022 and ISO 42001:2023 with public certificates and a SOC 2 letter of attestation.

2 re-opened2 recordsHarvey Data Processing Addendum Legora pre-approved sub-processors Legora Trust Center by session