AI Hallucination Cases

register entry ai-hallucination-cases-database-aio-20261029 OBS-4f95ab35a3ac · file as first run sha256 0a0aaa6b60370f00…

• T1-1 definition: 'The primary public database tracking generative AI hallucinations in court decisions is the AI Hallucination Cases Database, created and maintained by Damien Charlotin.' (markers: card 1's token and one unresolved token)
• T1-2 focus: 'It tracks court and tribunal decisions where generative AI explicitly produced hallucinated content, most commonly fake legal citations, fabricated quotes, and non-existent case law.'
• T1-3 scope and size: 'As of October 2026, the database contains over 2,200 documented decisions globally, with U.S. federal and state courts making up the majority of the cases.'
• T1-4 parties: 'The tracked decisions span private lawyers, pro se (self-represented) litigants, expert witnesses, prosecutors, and judges.'
• T1-5 sanctions: 'The database helps track legal consequences ranging from strict warnings and mandatory verification rules to substantial monetary sanctions and disbarment.'
• T1-6 access: 'You can view and search the project directly at the AI Hallucination Cases Database.' (link and marker: card 1's token)
• T1-7 companion: 'For specialized UK records, a companion resource is available via the AI Hallucination Cases Tracker.' (link and marker: card 2's token)
• T1-8 closer: 'Would you like information on specific high-profile sanctions cases from the database, or guidance on how legal professionals verify AI output?'

• B1 Damien Charlotin — AI Hallucination Cases Database page, HTTP 200 at 2026-10-29T05:06:47Z; raw sha256 68d8bdade886d77a…
• B2 Natural and Artificial Law — AI Hallucination Cases Tracker page, HTTP 200 at 2026-10-29T05:08:25Z; raw sha256 b4c8760c47ec182d…
• B3 HAQQ — AI Hallucination Cases: The 2202-Case Sanctions Tracker - HAQQ page, HTTP 200 at 2026-10-29T05:08:26Z; raw sha256 99fa409d712f0ae0…
• B4 LinkedIn — George Hannah's Post unreachable (not found by WebSearch); card title and snippet stand
• B5 YouTube·KARE 11 — How common are AI hallucinations in court cases? oEmbed, HTTP 200 at 2026-10-29T05:08:26Z; raw sha256 ac41df10addc1eb2…
• B6 YouTube·Miller Law Group — AI Hallucinations Got an Attorney Disbarred oEmbed, HTTP 200 at 2026-10-29T05:08:27Z; raw sha256 60504a8520783b95…
• B7 Artificial Authority (Damien Charlotin) — Hallucinations Case Database - FAQ page, HTTP 200 at 2026-10-29T05:08:27Z; raw sha256 93d0d01730155efc…
• B8 Wikipedia — Artificial intelligence raw wikitext and rendered page, HTTP 200 at 2026-10-29T05:09:20Z; raw sha256 8046fc18449a00eb…
• B9 Wikipedia — Hallucination (artificial intelligence) raw wikitext and rendered page, HTTP 200 at 2026-10-29T05:09:20Z; raw sha256 44c64b72efb8835a…
• B10 Wikipedia — Mata v. Avianca, Inc. raw wikitext and rendered page, HTTP 200 at 2026-10-29T05:09:21Z; raw sha256 e3a4791f5537949a…
• W1 Wikipedia — no article ('AI Hallucination Cases', 'AI Hallucination Cases Database', 'Damien Charlotin') HTTP 404 (raw and rendered), 2026-10-29 05:08 UTC; Special:Search 200

E_A, the archive's entity: the same database, in the MMRS meta-aggregator's record: Charlotin's database of legal decisions in which a court or tribunal finds or implies reliance on hallucinated material, whose unit is the decision and into which a fabricated citation enters only when a court catches and rules on it; the census's most exact record of citation fidelity, each entry a machine composition that reached a court with invented or misrepresented sources, the court's decision the reception that recorded it; observed by the Capture Registry at its address, where the composition layer held its maker and size and narrowed its rule to 'explicitly' MMRS-01 MMRS-02 MMRS-10 MMRS-11 MMRS-12 MMRS-22 MMRS-26 CAP-1-01 CAP-1-02 CAP-1-05; defining papers

E_F, the field's entity (type C — a conventional reading holds it (public entity: a database)): a database, on its page, that tracks legal decisions in cases where generative AI produced hallucinated content, typically fake citations, admitting only cases where a court or tribunal has explicitly found (or implied) reliance on hallucinated material, with some alleged-but-unconfirmed decisions as judgment calls; started by Damien Charlotin in April 2025, 2,231 cases at 28 October 2026, CC BY 4.0 with a CSV; not tracking the wider universe of fake citations in filings; beside it HAQQ's and Natural and Artificial Law's trackers (B1–B10)

ρ = extends: the archive reads E_F as the entity and adds senses on the field's own object: what its selection leaves out (a citation enters only when a court catches and rules on it), the entry as a machine composition and the court's decision as its reception, the database as a record of citation fidelity recording attribution, the record kept apart from the entity, the neighbours as candidates, and the composition layer's observed treatment of the entity MMRS-10 MMRS-11 MMRS-12 MMRS-22 MMRS-26 MMRS-24 MMRS-25 CAP-1-02 CAP-1-05 CAP-1-06. Guard: ρ = extends claims an addition and no more: MMRS-10, MMRS-11, MMRS-12 are the record's readings (interpretation), not findings that the database is defective; where an archive claim meets a field claim (MMRS-20 beside F40, F54, F56; MMRS-24 beside F24, F26) both are composed under P_coh in the one claim space

The case: equal terms, ontology included (v0.9). The MMRS record holds the field's entity as the entity: it describes AI Hallucination Cases from the database's own page, marking each field's basis (operator, observed, reading), and composes nothing that denies a field presupposition; its readings are on the field's object. What it adds is on that object: what the selection leaves out (a fabricated citation enters only when a court catches and rules on it), the entry as a machine composition received by a court and the database as a record of citation fidelity, the record kept apart from the entity, the neighbours as candidates; and the Capture Registry adds one observation of the composition layer at the entity's address. The ontology the archive uses is available in the field itself: the page states its boundary (F5), the FAQ that the courts judge and the database undercounts (F44) and that hallucinations expose citing without reading (F53), HAQQ that the count is a floor (F30). The law of the suppressed witness does not apply (§0.10: not where the field holds the ontology). Not mixed.

#MMRS article, kernel (articles.py): AI Hallucination Cases is Damien Charlotin's database of legal decisions in which a court or tribunal finds or implies reliance on hallucinated material: 2,231 decisions worldwide as of 28 October 2026, CC BY 4.0, with a CSV export.

#MMRS article, What its selection leaves out s2: A fabricated citation enters when a court catches and rules on it.

#MMRS article, Place in the field s1–s2: It is the census's most exact record of citation fidelity. Each entry is a machine composition that reached a court with sources it invented or misrepresented, and the court's decision is the reception that recorded it.

#MMRS census.json, schema_note: each page has two nodes, the MMRS record and the entity it describes; relations are typed; each assertion carries a source and a basis (operator statement, third-party report, our observation, our reading); lifecycle is kept apart from the observation that saw it

#capture ai-hallucination-cases-database-aio-20261029, analysis: The database is composed with its maker, at its size, as the field's primary record, and its admission rule is narrowed from 'finds or implies' to 'explicitly'.

#B1 opening ¶1: It does not track the (necessarily wider) universe of all fake citations or use of AI in court filings.

#B7 Anything else ?: For years, we (me included) have cited without reading; but now the costs of that practice have become explicit. In other words, hallucinations expose the epistemic hygiene the legal profession has long lacked, and that is precisely why they deserve to be studied.

equal-terms case: every field claim composed as received, beside the archive claims it stands with. Reading the received column recovers the field (§5.9).

v0.9 §4.10: KO_B sentences surviving in the KO at similarity ≥ 0.95, with positions (computed as apply_v09.py does: difflib.SequenceMatcher ratio, the best KO sentence for each KO_B sentence): v0.9 0 of 10 (positions [None, None, None, None, None, None, None, None, None, None]; composed first under v0.9, so there is no v0.7 object to measure against). Opening-run check (§5.10): no breach.

AI Hallucination Cases tracks legal decisions in cases where generative AI produced hallucinated content, typically fake citations, describing itself as the most comprehensive database of AI hallucination cases in law, and its decision is any document in which a court or tribunal addresses AI use in more than a passing reference (its page). It covers only cases where the court or tribunal has explicitly found (or implied) reliance on hallucinated material, the maker letting the courts make or imply the judgment so that the database is necessarily an undercount (its page; FAQ; HAQQ), with some alleged but unconfirmed decisions, and alleged hallucinations of judges, as judgment calls, though the FAQ says he refrains from rows only alleged, most hallucinations being obvious. Damien Charlotin started it in April 2025 while teaching at Sciences Po and maintains it on his personal website (FAQ), and Wikipedia gives its establishment in April 2025 by the HEC Paris and Sciences Po lecturer; its period filter begins in Q2 2023. It does not track the wider universe of all fake citations or AI use in court filings (its page), and HAQQ calls its count a floor. It had identified 2,231 cases by 28 October 2026, 1,514 from the USA and 1,288 involving pro se litigants against 880 lawyers, with HAQQ's 2,202 on 19 October, 1,501 American; the nature of each hallucination is fabricated, false quotes, misrepresented or outdated advice (its filters; HAQQ). Each record carries the court, date, party, AI tool as disclosed, nature, outcome, penalty and reports, a penalty of 1 being a placeholder, and the page warns that a named tool is not necessarily responsible and that automatically extracted dates and responses are indicative (its page). From two or three cases a month before 2025 it reached five a day by February 2026 (FAQ), with over 1,300 a year after its start (Wikipedia) and HAQQ's series from about 200 to 2,202, updated daily (its page); the FAQ reads the growth through judicial lag and LLM availability, the US share through its legal data, and the records as mostly pure mistakes. Cases come from referrals, scrapers and keyword searches, and from readers, through an AI-assisted pipeline (FAQ; its page); it is CC BY 4.0 with a CSV, serves practitioners and benchmarks, underlies the PelAIkan checker, and its maker expects no fix at the model stage and finds that hallucinations expose a profession that cited without reading (FAQ). It has been featured in news media and in court decisions (its page), and Wikipedia cites it for Canadian and Israeli counts and Mata v. Avianca, records sanctions by 2026, and the Overview's cards include a LinkedIn post at 1,600 cases and two videos by title. Beside it stand HAQQ's sanctions tracker, a vendor's monthly re-count that lists sanctions up to about $109,700, and Matthew Lee's international tracker of selected judicial quotations at Natural and Artificial Law, which measures no frequency, with a separate UK tracker whose inclusion is not proof (HAQQ; Natural and Artificial Law).

AI Hallucination Cases is Damien Charlotin's database of legal decisions in which a court or tribunal finds or implies reliance on hallucinated material (MMRS record); its own page says it "tracks legal decisions" in cases where generative AI produced hallucinated content, "typically fake citations, but also other types of AI-generated arguments", and describes itself as "the most comprehensive database of AI hallucination cases in law". Charlotin wrote in November 2025 that he maintains it on his personal website (FAQ), and the MMRS record gives him as its operator by his own statement; HAQQ's tracker calls him a research fellow at HEC Paris, and Wikipedia's articles on artificial intelligence and on hallucination a legal data analysis lecturer at HEC Paris and Sciences Po. Its unit is the decision (MMRS record), which its page defines as any document "where the use of AI, whether established or merely alleged, is addressed in more than a passing reference by the court or tribunal"; it admits "only cases where the court or tribunal has explicitly found (or implied) that a party relied on hallucinated content or material", which the MMRS record states as reliance a court "explicitly finds, or implies", entering it in its self-description field and as the property that admission requires "a court finding or implication of reliance on hallucinated material". The judgment is the courts': "I am not making that judgment, I let the courts and judges make or imply it, which is why the database is necessarily an undercount" (FAQ); HAQQ restates the bar as a court that "found, or clearly implied" reliance, accusations alone not counting, and the FAQ holds that most hallucinations are "very obvious", a made-up case name or false quote one that "cannot be a human mistake". Its exceptions are stated twice, and the statements pull apart: its page "also covers some judicial decisions where AI use was alleged but not confirmed", "a judgment call on my part", while the FAQ says he refrains "from adding rows about cases where hallucinations are only alleged" and, in a February 2026 edit, makes an exception where the alleged hallucination comes from a judge and, absent an appeal decision or retraction, he is "bound to make a judgment call"; the MMRS record composes both as "Mere allegations are excluded, with some judgment calls admitted." It states its boundary itself, that it "does not track the (necessarily wider) universe of all fake citations or use of AI in court filings" (its page; quoted by the MMRS record), and on the MMRS record's reading a fabricated citation enters only when a court catches and rules on it; HAQQ draws the same consequence for its reading of the count, "a fake citation nobody checks never reaches the database, so 2,202 is a floor". Each record gives the case, the court and jurisdiction, the date, the party using AI, the AI tool, the nature of the hallucination, the outcome or sanction, the monetary penalty, details and reports (its table; the MMRS record lists the court, date, party, tool where named, nature and outcome), the nature filed as Fabricated (1,829), False Quotes (595), Misrepresented (947) or Outdated Advice (38). Its outcomes can be filtered for monetary sanctions and disciplinary referrals, and a penalty recorded as 1 "is a placeholder, not the real figure", the amount being unknown (its page). Its page states the record's limits: AI tools are tracked "only as disclosed or admitted by parties in court filings", with 'Unidentified' and 'Implied' as the defaults; a named tool "does not necessarily mean that tool was responsible"; offending-document dates are parsed automatically and "may occasionally be wrong", and responses when caught are extracted automatically, "indicative rather than exact". Seeking to be exhaustive, it had identified 2,231 cases at its update of 28 October 2026 and calls itself "a work in progress" (its page; the MMRS record gives the same count as of the site's own update date), 1,514 of them from the USA, 238 from Canada, 112 from Australia and 82 from the UK, with 1,288 involving pro se litigants, 880 lawyers, 32 judges, 17 experts and 5 prosecutors (its filters; the MMRS record gives the pro se and lawyer counts). HAQQ, reading the database's update of 19 October 2026, gives 2,202 decisions, 1,501 from the United States, 1,267 involving pro se litigants and 872 lawyers. Charlotin started it in April 2025 while teaching Sciences Po Law School students on the "Limits and Potentials" of LLMs in law, when he looked for data on hallucinations and, finding none, decided to "tally the cases myself" (FAQ); Wikipedia dates its establishment to April 2025 and places its creation among the effects of law firms' use of generative AI for research; its period filter begins in the second quarter of 2023 (its page; MMRS record), and the MMRS record's census enters that filter as its founding. Before 2025, the FAQ says, "we maybe had two or three cases a month", and by February 2026 the daily average was five; Wikipedia's hallucination article gives over 1,300 instances a year after its start, a LinkedIn post four months before the capture 1,600 cases, and HAQQ's series about 200 in mid-2025, 719 in January 2026 and 2,202 on 19 October 2026; its page reads "Updated daily" and was last updated on 28 October 2026, and the MMRS record gives its cadence as continuous. The FAQ attributes the acceleration to judicial lag, the availability of LLMs and public knowledge of them, matching a profile of self-represented litigants or lawyers surprised by a tool's new AI component, and explains the share of US entries by the US legal data ecosystem, legal data in many European countries being "hoarded by legal editors". It divides the records into pure mistakes, "the vast majority of cases", and a "(substantial) minority" involving "bad" actors (FAQ). Cases come from referrals, scrapers and bots monitoring data sources, and keyword searches (FAQ), and readers are asked to send cases (its page); the MMRS record gives the intake as editor-curated with readers sending cases and the database growing by its editor's reading and readers' cases; the pipeline for new entries itself uses AI, "though I am careful to check it does not hallucinate" (FAQ). It is licensed CC BY 4.0, "free to reuse, including commercially, with attribution", to be cited as "AI Hallucination Cases Database, Damien Charlotin", and its cases download as CSV (its page; MMRS record). The FAQ gives its use as practitioners finding the cases relevant in their own jurisdiction, data analyses, and the cases as "benchmarks to fix the issue"; from it Charlotin built PelAIkan, "an automated reference checker that also detects hallucinations" (its page), and he expects no fix "at the model stage" and no reduction to zero "given the existing paradigm" (FAQ, a forecast). "For years, we (me included) have cited without reading; but now the costs of that practice have become explicit": hallucinations, the FAQ says, "expose the epistemic hygiene the legal profession has long lacked". The MMRS record reads it as "the census's most exact record of citation fidelity", recording attribution by the nature of the hallucination: "Each entry is a machine composition that reached a court with sources it invented or misrepresented, and the court's decision is the reception that recorded it." It has been featured in news media "and indeed in several decisions dealing with hallucinated material", the LA Times among them in May 2025 (its page); Wikipedia cites it for counts of Canadian filings (at least 167 by 14 June 2026, with courtready.ca) and Israeli cases (at least 55 by August 2026) and, in its article on Mata v. Avianca, for that case's description as a landmark, and its article on artificial intelligence records judges' sanctions and bar warnings by 2026. On the Overview's video cards it appears at title grain: KARE 11's "How common are AI hallucinations in court cases?" and Miller Law Group's "AI Hallucinations Got an Attorney Disbarred". HAQQ, which discloses that it "sells a legal AI product", re-reads the live database monthly and takes off its page any number it cannot re-verify, and lists sanctions from revoked admissions to personal costs, the largest it can source about $109,700 (HAQQ); the MMRS record enters it as a candidate, "a vendor tracker built on Charlotin's count". Matthew Lee, a barrister at Doughty Street Chambers, keeps at Natural and Artificial Law an international tracker of "Selected judicial quotations" on AI hallucination cases and fabricated or false citations, first published 16 May 2025, which "does not measure how often these issues occur", and a separate UK tracker of case records, outcomes, quotations and sources in which "inclusion is not proof of an AI hallucination or wrongdoing"; neither page names Charlotin's database, which the MMRS record's candidate list enters as "a UK companion to Charlotin's database". The Capture Registry (2026-10-29) holds the UK tracker a separate aggregator and HAQQ's tracker, which gives the database's 2,202 decisions of 19 October (HAQQ), built on Charlotin's count, both added as candidates the census did not hold, and reads the cards as the database "already aggregated in turn: a regional companion and a vendor tracker counting its cases". The MMRS record keeps itself apart from the entity, each assertion with a source and a basis and the lifecycle apart from the observation that saw it, records the database as active on 28 October 2026 with the roles of monitoring and reception, and relates it to the composition layer's account seated in the Capture Registry. There (Capture Registry, 2026-10-29) the Overview at 'ai hallucination cases database' composes the database with its maker and at its size, Charlotin and 'over 2,200' being right against the 2,231 its page gives, as the field's primary record; it narrows the admission rule, which the page states as reliance the court "has explicitly found (or implied)", to 'explicitly', so that the implied cases are dropped, and lists private lawyers first where the page's filters give 1,288 pro se litigants and 880 lawyers.

T against L_F(B) (AIO's ontology at work on its own cards): At 'ai hallucination cases database' (T1, six cards: B1–B6) the composition gives the database by its page and its neighbours' pages. Composed from its cards: the maker (B1's 'Cite as … Damien Charlotin', F12; HAQQ's 'maintained by Damien Charlotin', F29), the size ('over 2,200': B1's 2,231, F6, or HAQQ's 2,202, F28), the US majority (B1 F14; HAQQ F31), the party categories (B1 F13), the nature categories at a coarser grain (B1 F15: 'fake legal citations, fabricated quotes, and non-existent case law' for Fabricated and False Quotes), the access link (B1) and the UK pointer (B2 F24). Force raised: 'The primary public database' composes the page's self-description, 'The most comprehensive database of AI hallucination cases in law' (F22, attribute text of B1), as a ranking, without attribution (coarse grain). Limit dropped: 'where generative AI explicitly produced hallucinated content' drops the page's '(or implied)' (F3) and with it the implied cases, and drops the exception for some decisions where AI use was alleged but not confirmed (F4) (CAP-1-02); 'most commonly fake legal citations, fabricated quotes, and non-existent case law' drops the Misrepresented (947) and Outdated Advice (38) categories (F15) and 'other types of AI-generated arguments' (F1); 'over 2,200 documented decisions' drops the page's boundary, that it 'does not track the (necessarily wider) universe of all fake citations or use of AI in court filings' (F5), and HAQQ's 'floor' (F30). Substituted reading: the parties are listed with 'private lawyers' first, where the filters give 1,288 pro se litigants against 880 lawyers (F13; CAP-1-04), and 'private lawyers' is not one of the page's categories (Lawyer, Governement Lawyer, Federal Defender); 'a companion resource' for UK records names a relation the card's page does not state: B2 is Matthew Lee's international tracker of selected judicial quotations (F26), which points to a separate UK tracker (F24) and does not name Charlotin or his database (MMRS-24 defect; CAP-1-03). Partly absent from the readable field: T1-5's 'mandatory verification rules' and 'disbarment'; the page's outcome filters are monetary sanction and disciplinary referral (F23), HAQQ lists revoked admissions, referrals, suspensions and personal costs (F33), and 'disbarred' appears only in card B6's title (F38, title only): if T drew on it, it composed a video title without attribution, which the field cannot settle. Available, not composed: the decision as any document addressing AI use beyond a passing reference (F2); the courts' judgment and the undercount (F44 is the FAQ's, not a card; on the cards HAQQ's 'counts only what a judge caught', F30); the AI-tool and extraction limits (F18–F20) and the penalty placeholder (F21); the licence, CC BY 4.0 with attribution, and the CSV (F11, F12); the reader intake (F9); PelAIkan (F10); media and court citations of the database (F7, F8); HAQQ's method and its disclosure that it sells a legal AI product (F34, F35); Natural and Artificial Law's own limits, selective and not measuring frequency, inclusion not proof (F25, F27); the LinkedIn post's 1,600 (F36). Content replaced by use: the closer (T1-8). Neither the FAQ (B7: the start in April 2025, the intake, the pipeline, the reading of the cases, citing without reading) nor Wikipedia (B8–B10) is a card; they are admitted beside the cards by the operator's ruling and were not surfaced by T.

v0.9 §0.5: the first diff is already ontological, and here it is small on the entity and large on its limits. T and the field agree on the entity (Charlotin's database of court decisions on AI hallucinations, at its size); what T keeps is the page's self-description, its counts and its categories, and what it drops is the page's own conditions on what the count is: '(or implied)', the alleged-but-unconfirmed exception, the boundary of the wider universe, and the record's stated limits. It raises the page's self-description to a ranking and composes a neighbour as a companion. The row records this as observed and claims nothing about the composer's internals (§0.5).

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the record's reading of what the selection leaves out, that a fabricated citation enters only when a court catches and rules on it (MMRS-10), beside the page's boundary (F5) and HAQQ's floor (F30); the archive's sense the field does not state, the database as the census's most exact record of citation fidelity, each entry 'a machine composition that reached a court with sources it invented or misrepresented', the court's decision the reception that recorded it (MMRS-11, MMRS-12), with attribution recorded by the nature of the hallucination (MMRS-22); the record kept apart from the entity, each field with source and basis, lifecycle apart from the observation (MMRS-26, MMRS-14) and the relation to the composition layer's account (MMRS-23); the neighbours as candidates (MMRS-24, MMRS-25) and the Capture Registry's observation of them (CAP-1-03, CAP-1-06); and the composition layer's treatment of the entity at its address (CAP-1-01, CAP-1-02, CAP-1-04, CAP-1-05). Coherence (P_coh, composed each with its source): the MMRS census's 'founded' (the period filter, Q2 2023, MMRS-20) against April 2025 (F40, F54, F56); 'a UK companion to Charlotin's database' (MMRS-24) against Natural and Artificial Law's pages, which do not name it (F24, F26); the MMRS record's 'Mere allegations are excluded, with some judgment calls admitted' (MMRS-04) beside the maker's two statements (F4; F45, F47). Restated lineages (§3.6, the own page keeping the earlier date): MMRS-01 and MMRS-16 with F1; MMRS-02, MMRS-03, MMRS-21 with F2, F3; MMRS-05 with F17, F15; MMRS-06 with F13; MMRS-07 with F16; MMRS-08, MMRS-17 with F9, F43; MMRS-09 with F5; MMRS-13 with F12, F39; MMRS-15 with F11, F12; MMRS-18 with F6; MMRS-19 with F22. The measure is in row.measure.

alignment: {'T1_cards_claims': ['F1', 'F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'F9', 'F10', 'F11', 'F12', 'F13', 'F14', 'F15', 'F16', 'F17', 'F18', 'F19', 'F20', 'F21', 'F22', 'F23', 'F24', 'F25', 'F26', 'F27', 'F28', 'F29', 'F30', 'F31', 'F32', 'F33', 'F34', 'F35', 'F36', 'F37', 'F38'], 'T1_composed': ['F1', 'F3', 'F6', 'F12', 'F13', 'F14', 'F15', 'F22', 'F24', 'F28', 'F29', 'F31'], 'T1_composed_count': '12 of 38 card claims', 'rule': "a field claim counts as composed when T states it at its source's grain or coarser; the typology of what is not is in T_vs_LB (§6.2)"}