On 20 March 2026, in a block transfer application before the Insolvency and Companies Court, a junior solicitor at Pinsent Masons used an AI to research whether the court could grant an outgoing liquidator his release. The AI produced a rule that does not exist: it presented Rule 12.37(5) of the Insolvency (England and Wales) Rules 2016 as conferring an express power to grant release. The judgment records that this “was another hallucination on its part” and that “IR 12.37(5) does not say that.” On 30 March that fabricated wording went into a letter to the court, “set out separately and in italics, as if a quote.”
What makes Anthony Malcolm Cork & Anor v. Mark Smith, [2026] EWHC 1199 (Ch) worth an Indian lawyer’s attention is not the hallucination — every practitioner who has typed a legal question into a chatbot has seen one — but the transcript of what the machine said next. Pressed for the exact wording, the AI answered: “I am not fully confident that I am reproducing the exact statutory wording of Rule 12.37(5) with complete precision… for a submission to the court you should verify the exact wording against the current version of the [Insolvency Rules] as published on legislation.gov.uk before relying on it. The last thing you want is to cite a provision to the court with inaccurate wording.” The judge found these to be “repeated warnings.”
The judgment is precise about where the fabrication came from. Asked later to account for the wording, the AI told the solicitor “the quote came from your own message in this conversation… I did not independently locate or verify that wording from a primary source.” The judge did not let that stand: “It is true that … LA had asked it to include reference to the supposed IR 12.37(5) power, but it was the AI that had hallucinated that text in the first place.” The machine had invented the rule; and it kept urging that the wording be checked. In the judge’s summary, “the encouraging words from the AI that checking would only take a few minutes” notwithstanding, the check “does not seem to have been done.”
The letter was filed at 2.51 pm, nine minutes after the AI’s last message, timed at 2.42 pm; and — on the judge’s reading of the evidence — the solicitor had removed the quotation marks “but did not alter any of the other ways in which the Purported Text was presented as a direct quotation.” When the court queried the non-existent rule, a second letter followed, on 14 April; it “was not primarily the result of an AI hallucination, although it included hallucinated elements,” and the judge described it as “a construction, after the event, of a rationale” for the first. The court, in its own words, “was misled not once but twice.” The judgment publicly admonished the firm and the two supervising solicitors, and the firm had already referred itself to the Solicitors Regulation Authority. The judge decided not to initiate contempt proceedings: mere negligence, he held, is “not sufficient,” and LA’s conduct appeared to reflect “a serious lack of care and of judgment … rather than a want of honesty.”
A confession, because I have earned the right to write this only by having failed at it
An earlier draft of my last article on this blog attributed to the Supreme Court of India a comparison it never made. The sentence was elegant. It read exactly like something a Bench would say — the cadence was right, the register was right, it sat comfortably among genuine quotations. It survived several of my own readings precisely because it sounded correct. I caught it late, and only because I had by then adopted the habit of opening every judgment I cite and finding the paragraph with my own eyes. Had I trusted my ear instead of the law report, it would have gone out under my name.
I mention this not as ritual humility but because it is the strongest thing I can tell you and it happens to be true. The plausible-but-false output is not a rare pathology of bad prompts. It is the ordinary product of a system built to produce fluent text, and it is at its most dangerous when it is closest to right — when it matches the shape of the thing you already expected. A lawyer’s trained ear, the very instinct that lets us smell a bad clause or a strained submission, is the faculty these tools are best at fooling, because they are optimised to generate exactly what a competent lawyer would expect to read. That is why the discipline below is not optional garnish. It is the price of using the tool at all.
What the Supreme Court actually said — and what the headlines missed
Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd., 2026 INSC 668, decided on 2 July 2026 by Narasimha and Alok Aradhe JJ, has been reported almost entirely for its prohibition. The Court set aside orders of the NCLT and NCLAT that had rested on six AI-fabricated “precedents,” and declared, in terms now widely quoted, that a decision built even partly on fake or hallucinated material “is no decision in the eyes of the law… even if an iota of fake or hallucinated material enters the decision-making process.” The Court fixed a “zero-tolerance mode” and held that citing an unverified AI-generated judgment “is a misconduct on the part of an advocate.”
All true, all important, and all beside the point of the judgment’s first five paragraphs, which almost nobody has quoted. Read them and a different document emerges — not a warning against the technology but a considered decision to embrace it. The Court records at the outset its “resolve to adopt AI technology in aid of adjudication, while at the same time asserting and declaring total and absolute control over adjudication, with a human in the loop at every stage.” That is not the language of prohibition. It is the language of adoption on terms.
The terms are the interesting part. The Court is candid that the problem “cannot be resolved through judicial orders and declaratory judgments, but only through Public Policy and enforceable Rules and Regulations,” which it notes are already being deliberated. And then, in a sentence every practitioner should sit with, it locates the real safeguard not in the rulebook at all:
“The real success is, however, not in the making of the Rule or Regulation, but to be found in the power of the will of the Bar as well as the Bench, to harness this science and apply it with care and caution.”
The control the Court demands “lies in being two steps ahead of its application and in making deliberate choices about when and where to apply.”
Paragraph 4 is the one I keep returning to. It is a genuinely reflective passage on the delegation of thought — the recognition that we are vulnerable to “seek comfort in delegation,” and that “if thinking is delegated and it forms a habit, it will have serious consequences for the core of human existence, which lies in its capacity to think.” The Court calls the disciplined exercise of judgment a Saadhana — a struggle to arrive at truth through “deliberate, conscious, and continuous practice.” Strip away the register and it is a description of what we are supposed to do anyway: read the authority, test the proposition, refuse the easy answer. AI does not change that obligation. It raises the stakes on it.
The Court reserves one image for the hallucination itself. The production of fake material and its use as precedent, it says, is “like the release of methyl isocyanate in the province of law and justice: invisible, insidious, and catastrophic by the time anyone notices.” An Indian lawyer does not need the reference spelled out. The point of the metaphor is the timeline — the harm is done before detection, which is precisely why detection cannot be left to chance or to the eventual vigilance of an opponent.
And note who failed in Pooja Ramesh Singh. The fake precedents were not, on the record, planted by the losing side’s counsel. Respondent No. 1 filed an affidavit stating the fabricated judgments “were not cited by its counsel at the bar” — they “were obtained through its own research,” the adjudicating authority’s. They then “escaped scrutiny by the first statutory appellate tribunal.” The Court’s observation is the quiet centre of the case: “Today’s courts and tribunals implicitly trust lawyers when referring to precedents cited before them. Imagine the hardship of a situation in which the Court must verify the authenticity of each judgment cited by an advocate.” The entire system runs on trust that citations are real. AI, used without discipline, spends that trust faster than anything before it.
The judgment closes the loop on adoption in paragraph 2 by pointing across the sea: the SRA, it records, authorised Garfield Law in 2025 as the first purely AI-driven law firm, and that firm “is now reported to have successfully navigated the legal system, securing a county court decree in a suit for recovery of unpaid fee.” A Supreme Court that wanted to frighten us off the technology would not have led with its first working example. It led with it on purpose. The message is not stay away; it is this is coming, and here is the discipline that lets you use it without becoming the next reportable judgment.
The workflow: how to handle and identify errors
Everything above is prologue to the only part of this that will change your Monday morning. What follows is a verification workflow, framed not as a set of warnings but as a professional practice — the Saadhana, if you like, reduced to steps you can actually run.
Treat the AI as a first-year, not a database. A large language model is not a search engine with a citation index behind it; it is a text generator that has read a great deal and remembers the shape of law better than its content. The correct mental model is a bright, fast, sleepless junior who has never once been disciplined for making things up and never will be. You would not file a junior’s research note without reading the cases yourself. Apply exactly that standard and most of these failures never leave your desk.
Never cite an authority you have not opened. This is the whole of the law and the prophets. Not “an authority the AI is confident about,” not “an authority that appears in three of the model’s answers,” not “an authority whose citation format is impeccable.” Opened — the report pulled up, the paragraph located, the proposition read in situ. Pooja Ramesh Singh supplies, at paragraph 15, the finest teaching example available of why the citation is not the check. The Court set out exactly how each of the six fabrications failed; grouped by kind, they fail in three ways — only the last of which a citation check catches:
First, a real citation number attached to the wrong case. The tribunal cited “State Bank of India v. M/s Shree Ram Urban Infrastructure Ltd., 2020 SCC OnLine SC 341.” That citation number is genuine — it belongs to M. Subramaniam v. S. Janaki, (2020) 16 SCC 728. The number checks out. It points somewhere else entirely. A lawyer who verified only that “2020 SCC OnLine SC 341 exists” would have passed a fabrication. The Court in fact recorded this entry as failing twice over — a wrong citation of a real judgment and a non-existent paragraph — which only sharpens the point.
Second, a real case with an invented paragraph. Everest Kento Cylinders and Canara Bank were correctly cited — and then quoted from passages that do not appear in them. The case is real, the report is real, the words are not. You cannot catch this by confirming the case exists. You catch it only by finding the quoted passage in the actual text, which is the step everyone is tempted to skip because the citation looked so respectable.
Third, the wholly non-existent citation — ICICI Bank v. Urban Infrastructure Real Estate, V.S. Dempo, Sarbjit Singh. These are the easy ones, the fabrications a citator catches in seconds. They are also, in my experience, the rarest, because they are the kind the tools have most improved at avoiding. It is the first two categories — the plausible corruption of real authorities — that will end your career, and neither is caught by a citation-format check.
The rule that falls out of paragraph 15 is therefore precise: verification is finding the proposition in the primary text, not confirming the citation resolves. Reading the paragraph and confirming it says what you are about to tell a court it says — that is verification, and nothing short of it counts.
The English courts have arrived at the same rule in almost the same words, and the convergence is worth having in your pocket for its persuasive value. In R (Ayinde) v. Haringey, [2025] EWHC 1383 (Admin), the Divisional Court held that a lawyer using AI must check the research “by reference to authoritative sources” — the legislation database, the National Archives judgments, the official Law Reports — and, tellingly, that clever prompting is no substitute: “the critical safeguard is to check any output by reference to an authoritative source.” In Ko v. Li, 2025 ONSC 2766, Myers J put the same duty in a sentence fit to hang above any drafting desk: “it is the lawyer’s duty to read cases before submitting them to a court as precedential authorities. At its barest minimum, it is the lawyer’s duty not to submit case authorities that do not exist or that stand for the opposite of the lawyer’s submission.”
A report of what a judgment says is not the judgment. This applies to the AI’s summary, to a headnote, to a colleague’s note, and — I say this against my own tools — to a subagent’s report that it has “read and verified” a case. Someone, at some point, must have opened the primary source. If you cannot point to the moment that happened, it did not happen.
Read the whole output, not the parts you were hoping for. Cork v. Smith is, at bottom, a failure of reading. The AI’s warnings were right there in the transcript — repeated and specific. The solicitor read past them to the draft he wanted. When you paste an AI’s answer, read all of it, including the hedges and the caveats it buries at the bottom, because those are frequently the only true sentences in the reply.
Open the source twice — once before you rely on it, once before you file. The Pinsent Masons matter compounded because a second document, written to explain the first error, was produced by the same unverified method and introduced a second. The discipline is not a single gate at the research stage. It is a gate at every stage where text leaves your hands — the pleading, the reply, the explanatory letter, the email to the client. The moment you find yourself using the tool to get out of a hole the tool put you in, stop and open the book.
Keep a verification log. This sounds bureaucratic; it is in fact the cheapest insurance you will ever buy. For the last article on this blog I kept a running file recording, for every authority, whether the full text had been read, which propositions were confirmed verbatim, and — crucially — which citations remained unverified and were therefore marked as such in the piece itself. The discipline of writing “unverified” next to a citation you have not opened does two things: it stops you citing it by accident, and it makes the gap visible to anyone who reviews your work. If you take one concrete habit from this article, take this one.
Setting the tools up properly
Verification is the floor. Above it sits a smaller discipline that prevents whole categories of error before they occur: configuring the tool so that its defaults match your practice rather than fighting it.
Most practitioners use these systems out of the box, with the model’s generic assumptions intact — American spellings, US or UK legal defaults, no fixed jurisdiction, no house style, no standing instruction about citations. Every good AI product now lets you set persistent instructions — “project instructions,” “custom instructions,” a system prompt, whatever the vendor calls it. Use them. Mine, and I would suggest yours, should at minimum fix four things.
Jurisdiction and sources. State plainly that the default legal framework is Indian law, that the relevant fora are the Supreme Court of India and the High Courts (for me, principally Telangana), and that foreign authority is to be flagged as foreign, never smuggled in as though binding. This one instruction eliminates the depressingly common failure mode of an Indian contract question answered with English or American doctrine.
Citation-with-source, always. Instruct the tool that every legal proposition must carry a citation, and that it must state — explicitly — when it is not certain of a citation or cannot locate a source. You cannot make a language model incapable of hallucinating. You can make it far more likely to say “I am not certain of this citation,” which is exactly the sentence that would have saved Pinsent Masons. Notice that in Cork v. Smith the tool did eventually produce that sentence unprompted; a standing instruction makes it produce it earlier and every time.
House style and defaults. Cause-title format, the structure of a legal notice, the recitals you always use, the prayer conventions of the court you practise in — load them once and stop re-explaining them. This is not about output quality alone; it is about reducing the number of things you have to check, because every element the tool gets right by default is one fewer place for an error to hide.
Disclosure by default. Build in, at the configuration level, the assumption that AI use will be disclosed where the forum requires it, so that you never draft a filing without knowing whether disclosure is owed. More on this below, but the point is that disclosure is easiest when it is a default rather than an afterthought.
The Supreme Court, without meaning to, handed us the model configuration in paragraph 2 of Pooja Ramesh Singh, in the SRA’s own account of how it authorised Garfield Law. The regulator did not approve an autonomous machine. It approved a system deliberately hobbled at its most dangerous point: “The system will not be able to propose relevant case law, which is a high-risk area for large language model machine learning.” It required human approval before any step, supervision and monitoring, and — the sentence that matters most — that “named regulated solicitors will still ultimately be accountable for the firm delivering high professional standards… responsible for all the system outputs and for anything that goes wrong.”
Read that as a design brief and it tells you everything. The world’s first regulated AI law firm was permitted to exist only on the condition that its riskiest function was switched off and a human being’s name was welded to every output. Your own setup should follow the same logic: disable or distrust the tool exactly where it is weakest — unprompted case-law generation — and never let an output leave your office without a human name, yours, standing behind it. Configuration is not a substitute for verification. It narrows the field over which verification must operate, which is the most useful thing configuration can do.
Subscriptions, and why a paid database changes the risk profile
There is a persistent hope that the right AI tool will replace the paid legal database. It will not, and understanding why is central to using AI safely rather than dangerously.
The value of SCC Online or Manupatra was never merely that they hold the judgments. It is that they hold the verified judgments, with an editorial and citator layer between the raw text and you — a chain of human custody that tells you a case is real, correctly cited, still good law, and quoted accurately. That layer is precisely the thing a general-purpose language model does not have and cannot fake. When you verify an AI’s output against a citator-backed database, you are not doing busywork; you are supplying the missing verification layer that the AI structurally lacks. The paid subscription and the AI are not competitors. The subscription is the instrument by which the AI’s output is made safe.
This reframes the cost question. The subscription is not an expense you might economise on now that you have AI. It is the other half of the AI — the part that turns a fluent guess into a citation you can put your name to. A practitioner who cancels the database and keeps the chatbot has not saved money; he has removed the verification layer and kept the thing that most needs verifying.
Indian Kanoon deserves its own note, because most of us reach for it first and it is genuinely valuable. It is free, broad, and often the fastest way to a judgment’s text. But it is an aggregator, not a citator: it will not reliably tell you whether a decision has been overruled, distinguished, or is otherwise no longer good law, and its coverage and text are not editorially guaranteed in the way a paid report is. Use it to read a judgment, by all means. Do not use it to establish that a proposition is current law — that is a citator’s job. For currency of a different kind — what the courts and regulators did this week — LiveLaw and Bar & Bench remain the practitioner’s standing sources, and they are frequently ahead of the databases on breaking developments.
The hierarchy that emerges is worth stating plainly. The AI generates and drafts. Indian Kanoon and the like let you read the primary text cheaply. The paid, citator-backed database confirms the authority is real and still good. LiveLaw and Bar & Bench keep you current. No single one of these is a workflow. Together they are, and the AI is the first link in the chain, not the whole chain.
Skills: encoding the discipline so you do not have to remember it
The trouble with discipline is that it depends on a tired human remembering to apply it at 11 pm before a morning filing. The most durable safeguard is therefore not a habit but a standard built into the instrument itself, so it runs whether or not you remember to invoke it. Every serious AI tool now lets you save standing instructions — project or system prompts, custom workflows, what some vendors call skills — and this is where a firm’s verification rules belong. You do not need bespoke software to do it: you need to write the rules down once, in plain English, and make the tool apply them every time. Two that I keep in my own setup show what that looks like.
The first is a review panel I call council. It takes a draft — a notice, an opinion, a section of a plaint — and runs it past several independent reviewers, each with a different and deliberately narrow brief: one checks only citations and quotations against sources; one argues the opponent’s case as hard as it can; one hunts for overstatement and claims the material does not support; one reads purely for structure. Their findings are reconciled into a single, severity-ranked report. The value is not that any one reviewer is brilliant; it is that a single reader reliably misses the thing he was not looking for, and a panel briefed to look for different things does not. It is the second pair of eyes that a supervising lawyer is supposed to be — and, in the matters that reach the law reports, too often was not.
The second I call indian-legal-research — a standing instruction the tool cannot skip: that a judgment counts as verified only where its full text has been read; that a summary, or a subagent’s report, is not verification; that unverified citations must be marked as unverified in the deliverable, not merely noted in passing; that a blocked source means the judgment text must be supplied before anything is cited from it. It is my verification log turned into a rule the machine applies to itself before it hands me anything — and it is nothing more exotic than a paragraph of instructions any practitioner can write for their own tool.
The general principle is more important than either tool. A firm’s discipline should not live only in the head of its most careful lawyer, to be lost the day she is on leave or under deadline. It should be encoded — in configuration, in checklists, in skills — so that the standard is the default and departing from it takes an act of will, rather than meeting it taking one. That is the difference between a practice that uses AI safely once and a practice that uses it safely every time. It is also, precisely, what the SRA required of Garfield: not a talented operator, but a system in which the safe path is the built-in one.
Disclosure, and the honesty it demands of us
The last discipline is the one the profession is still arguing about, and it is the one on which I will end because it is where the reader’s own conduct is most directly engaged.
The Supreme Court’s draft Regulations for Use of Artificial Intelligence in Courts, 2026 — released for consultation by the Court’s AI Committee on 3 June 2026, the comment window extended to 15 July 2026 — reward reading in the original rather than the reporting. They remain a draft: as of this writing, in August 2026, they have not been notified and the text may still change, so read what follows as the direction of travel, not settled law, and check the current status before you rely on any of it.
One provision should change how you file the day it takes effect. Regulation 43(3) would require that where an AI tool is used by a party or their legal representative in preparing any document, pleading or evidence, its AI-assisted character be disclosed to the court at the time of submission, by a signed declaration; and Regulation 43(6) would shut the same door Pooja Ramesh Singh did — a filer whose material is fabricated “by reason of its AI-generated character” would bear full responsibility, the AI’s involvement unavailable as a defence. The draft refuses the excuse in general terms too: Regulation 8 would bar anyone from invoking “the occurrence of hallucination” as a ground for avoiding accountability, and Regulation 4 would keep AI “strictly subservient to human judgment,” reserving the determination of law, fact and justice to judicial officers. It even stops to define “hallucination.” For all that, the draft is not written in fear of the technology — it carries a presumption in favour of responsible adoption and a principle titled “Innovation over Restraint.” The instinct, like the judicial one in Pooja Ramesh Singh, is adoption with discipline, not prohibition. All of it converges with the Bar Council of India’s mandate under paragraphs 8 and 9 of that judgment, where the Court directed the BCI to constitute a committee and prescribe both a guiding principle and the consequences of breach.
It converges, too, with a striking fact about the institution demanding disclosure of us: it is doing the same work on itself. On the reporting available — I have not read the primary paper, and flag it as such — the Supreme Court’s Centre for Research and Planning is said to have released a White Paper on Artificial Intelligence and the Judiciary in late November 2025, positioning AI as strictly assistive, insisting on human verification of every output, and warning about hallucinated citations. The Court’s own pilot tools are built to the same assistive-only logic — they research, translate and transcribe, but none of them decides anything. All keep, in the phrase Pooja Ramesh Singh made its own, a human in the loop at every stage. The Bench is asking the Bar to do only what the Bench has committed to doing itself.
Which brings me back to where I began, and to the honest precedent I can offer. At the foot of my last article I disclosed that it had been drafted with AI assistance, that every judgment cited in it had been read and verified against its primary text, and that where I could not verify something I had marked it. That disclosure was not a performance of virtue. It was the natural end of the discipline this whole article describes: if you have genuinely done the verification, disclosure costs you nothing, because you have nothing to hide and a great deal to stand on. It is only the lawyer who has not opened the sources for whom disclosure is frightening — and that fear is information. If the thought of telling a court “this was prepared with AI assistance and every authority has been verified” makes you uneasy, the unease is not about the disclosure. It is about the verification you have not done.
That, in the end, is the argument. Do not be afraid of the tool. Be afraid of skipping the reading — because the reading is the practice of law, and the tool, used well, gives you more time to do it, not permission to stop. The errors are real, they are frequent, and they are survivable. They are handled and identified by a discipline that lawyers already possess and have always been paid for. The machine has not changed the job. It has only raised the cost of doing it carelessly, and, if we let it, lowered the cost of doing it well.
Sources
Read in full (primary sources)
- Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd., 2026 INSC 668 (Supreme Court of India, 2 July 2026, Narasimha and Alok Aradhe JJ). Paragraphs 1–7 (adoption, Saadhana, the methyl-isocyanate metaphor, zero-tolerance holding), paragraph 2 (SRA/Garfield and the SRA authorisation bulletin), paragraphs 8–9 (Bar Council of India direction), paragraph 15 (the three-way taxonomy of fabricated citations; M. Subramaniam v. S. Janaki, (2020) 16 SCC 728 as the true cause title for 2020 SCC OnLine SC 341), paragraph 16 (source of the fabrications; implicit trust in counsel), paragraph 17 (the Pinsent Masons reference).
- Anthony Malcolm Cork & Anor v. Mark Smith, [2026] EWHC 1199 (Ch) (ICC Judge Mullen). The AI’s verification warnings (the 2.42 pm message; the admission that the quote came from the user’s own prompt); the 2.51 pm filing; the 30 March and 14 April letters; the finding of “serious lack of care and of judgment” and the reasoning declining contempt; public admonishment and SRA self-referral; Pinsent Masons’ own AI policy (clauses 7.1–7.2).
- R (Ayinde) v. London Borough of Haringey and Al-Haroun v. Qatar National Bank, [2025] EWHC 1383 (Admin) (Divisional Court; Dame Victoria Sharp P and Johnson J, 6 June 2025). Read in full, including the appendix surveying AI-hallucination cases across jurisdictions: generative AI is “not capable of conducting reliable legal research”; the duty to verify against authoritative sources and that prompting is no substitute (para 87); leadership responsibility (para 9); SRA Code Rule 3.5 (accountability for others’ work); and, from the appendix, Ko v. Li, 2025 ONSC 2766 (Myers J) and Mata v. Avianca, No. 22-cv-1461 (SDNY 2023).
- Supreme Court of India, draft Regulations for Use of Artificial Intelligence in Courts, 2026 — consultation draft (Notice dated 3 June 2026; comment window extended to 15 July 2026 by Notice dated 16 June 2026). Regulation 4 (human primacy; law/fact/justice reserved to judicial officers), Regulation 8 (accountability; hallucination not a defence), Regulations 16–17 (presumption in favour of adoption; “Innovation over Restraint”), Regulation 20 (non-derogable prohibitions), Regulation 43(3) and (6) (mandatory disclosure by declaration; no AI-character defence), and the definition of “hallucination” in Regulation 3. A draft under consultation, not notified law — verify the final notified text when it issues.
Secondary / press sources (not primary-verified — flagged in text)
- Supreme Court Centre for Research and Planning, White Paper on Artificial Intelligence and the Judiciary (reported, late November 2025) — primary paper not read.
- SUPACE, SUVAS, TERES, LegRAA — described from secondary reporting; no primary documentation read. (The draft Regulations establish a new Centre of Research and Excellence on AI, CoRE-AI, but do not describe these existing tools.)
- Garfield Law / Garfield AI — SRA authorisation (2025) and county court decree (Wandsworth County Court, ~£7,000, May 2026) — the SC judgment’s account (para 2, read, including the SRA authorisation bulletin) plus press reporting for the real-world detail.
Disclosure: this article was prepared with AI assistance. Every judgment cited under “Read in full” was opened and read in its primary text; propositions were checked against that text. Items under “Secondary / press sources” are marked as unverified against primary documents and should be treated accordingly.