Aeneas

register entry aeneas-deepmind-aio-20261031 OBS-ed654d93f896 · file as first run sha256 50e1962dff6f15cd…

• T1-1 definition: 'an artificial intelligence model developed by Google DeepMind (introduced in July 2025) designed to contextualize, restore, date, and locate ancient, damaged, or fragmentary Latin inscriptions.' (two markers)
• T1-2 training corpus: 'Trained on the Latin Epigraphic Dataset (LED), containing over 176,000 inscriptions spanning from 700 BCE to 800 CE across 62 Roman provinces.' (two markers)
• T1-3 text and image input: 'A multimodal model that processes both weathered stone images and text to identify geographical provenance and shared syntax or formulas.' (two markers)
• T1-4 gap restoration: 'Restores missing or broken text segments of arbitrary length, achieving high top-20 accuracy on character sequences.' (two markers)
• T1-5 chronological and geographic attribution: 'Predicts when and where an inscription was made, typically dating texts within roughly 13 years of expert historian estimates.' (two markers)
• T1-6 collaborative research: 'In testing with historians, Aeneas provided useful contextual parallels 90% of the time and boosted researcher confidence by 44%.' (two markers)
• T1-7 access: 'You can explore or test the tool directly through Predicting the Past.' (one marker)
• T1-8 closing offer: 'Would you like to know more about how Aeneas compares to earlier DeepMind historical AI tools like Ithaca, or how researchers use it for specific archaeological finds?' (no marker)

• B1 Google DeepMind — Aeneas transforms how historians connect the past — Google DeepMind card; census own page aeneas--blog (byte-identical, read 2026-10-31, HTTP 200); B1.html sha256 c53e5cf5b5336f47
• B2 blog.google — Aeneas: Google DeepMind AI connects ancient inscriptions card; B2.html sha256 bfea5da464bccc15, HTTP 200 at 2026-10-31T04:39:17Z
• B3 University of Oxford — Aeneas: an AI model for restoring and contextualizing Latin inscriptions | Faculty of Classics card; B3.html sha256 772454851fa204ca, HTTP 200 at 2026-10-31T04:39:18Z
• B4 GitHub — GitHub - google-deepmind/predictingthepast · GitHub unreachable: https://github.com/google-deepmind/predictingthepast HTTP 403 from the build environment's proxy; card title and snippet only
• B5 Nature — Contextualizing ancient texts with generative neural networks | Nature card; B5.html sha256 07c431681107fad5, HTTP 200 at 2026-10-31T04:39:20Z
• B6 Predicting the Past — Predicting the Past | Contextualising, restoring, and attributing ancient texts card; census own page aeneas--site (byte-identical, read 2026-10-31, HTTP 200); B6.html sha256 08a11cf87711fe87
• B7 MIT Technology Review — Google DeepMind’s new AI can help historians understand ancient Latin inscriptions | MIT Technology Review card; B7.html sha256 1526eaa5bd4dd61d, HTTP 200 at 2026-10-31T04:39:21Z
• B8 Wikipedia — Aeneas (the model) (no article) no article: raw HTTP 404 for five title variants; disambiguation, hero and Google DeepMind articles HTTP 200; API search HTTP 429, index.php search no article; B8.txt sha256 dab1678fd4253e41

E_A, the archive's entity: the same model, in the Revelation First census record and the archive's deposits: an AI model that retrieves textual and contextual parallels for ancient inscriptions, at the intersection of AI and ancient texts, whose unit is an inscription with its retrieved parallels and whose texts are Latin inscriptions, adaptation stated; evidence that can settle which inscriptions share wording, syntax, formulas or provenance and cannot settle the direction of dependence between parallel texts, nor anything of the New Testament; read by the archive as retrieving the kind of evidence the midrashim transform argues from, the direction of which the Revelation First thesis states as its falsifier (#202) and requires be shown against shared ancestry (#832); a parallel kept apart from dependence (#828, #1458, #1474); a model of the class whose gain, on Ithaca, lies in ranked hypotheses rather than a verdict (#1443); composed at its name as 'a better Ithaca for Latin', its reach to papyri left out (Capture Registry, 2026-10-31) RF-01 RF-05 RF-07 RF-08 RF-14 RF-15 RF-16 RF-17 O202-01 O832-01 O828-01 O1443-01 O1443-02 CAP-1-06 CAP-1-07; defining papers #202, #832, #828, #1458, #1474, #1443

E_F, the field's entity (type C — a conventional reading holds it (public entity: an AI model for contextualizing ancient inscriptions)): the generative neural network Google DeepMind introduced in Nature on 23 July 2025 as the first model for contextualizing ancient inscriptions (DeepMind; Nature, Assael, Sommerschield et al.), co-developed with the University of Nottingham and researchers at Warwick, Oxford and AUEB, which retrieves textual and contextual parallels (texts sharing wording, syntax, formulas or provenance) by historically enriched embeddings from the Latin Epigraphic Dataset of 176,861 inscriptions, determines provenance from text and image, restores gaps of unknown length and dates within about 13 years; used by twenty-three historians as a source of starting points (90% and +44% with its predictions, 75% and +23% from parallels alone); named after the wandering hero and built upon Ithaca, which it now powers; trained for Latin and stated adaptable to papyri, manuscripts and coinage; limited by a small, survival-biased corpus; free as an interface, its code and weights open; with no Wikipedia article

ρ = extends: The archive reads Aeneas as its maker and its paper present it and adds, on the field's own object: what its evidence, parallels, can and cannot settle, the direction of dependence among what it cannot; the Revelation First thesis's own test of direction and its confound of shared ancestry, which set what a retrieved parallel would have to show; the distinction of parallel from dependence in the archive's philology; for the class, the location of the machine's contribution in ranked hypotheses rather than a verdict; the record's answer, field by field, to the census's question; and the composition layer's omission, at its name, of the reach that would bring it to the papyri. RF-14 RF-15 RF-17 O202-01 O832-01 O828-01 O1458-01 O1474-01 O1443-01 O1443-02 CAP-1-07. Guard: ρ = extends claims an addition and no more (§1.2a). It does not make the field's sources wrong at their dates. Where an archive claim stands against a field claim (RF-15 and RF-17, that retrieval does not settle direction, beside F31, the paper's parallel read as direct influence; RF-11's 'thousands' beside F11 and F28; RF-19's licence 'not read' beside F36; CAP-1-04's 'the saved sources say thousands' beside F11), both are composed under P_coh within the one ontology. The archive's class claims are claims on the classes the field sources (parallels: F3, F25; AI models for ancient texts: F1, F8), never on the entity by the composer's judgment.

The case: equal terms, ontology included (v0.9). The archive reads the field's entity as the entity. The Revelation First census record describes Aeneas from its maker's post, each evidence quote with its basis ('operator'), and asks of it the census's question of what each resource holds on the composition of the New Testament writings and how machines read them; its one reading of its own, rf_intersection, takes the model's parallels as the kind of evidence the midrashim transform argues from and holds that retrieval does not settle the direction of dependence (RF-17), which the record also states as what the evidence cannot answer (RF-15). The capture record checks the composition against those pages. No deposit names the model (every D candidate is the hero). The deposits bear on the classes the field sources for it and add senses and limits on the field's own object: the thesis states the direction of parallels as its falsifier (#202) and requires that shared ancestry be excluded before a parallel counts as inheritance (#832); the philological notes keep parallel apart from dependence (#828, #1458, #1474); the field survey reads Ithaca, the model Aeneas builds on, as a gain from ranked hypotheses rather than a verdict, the hazard of machine reading located in single-verdict interfaces (#1443). None of these contests what the field says the model is, and none supplies an ontology the field lacks: ρ = extends, the equal-terms case (§0.9, §1.2a). Where the paper reads a parallel as direct influence (F31) and the record holds that retrieval does not settle direction (RF-15, RF-17), both are composed under P_coh within the one ontology.

#202 the falsification conditions (line 33): The thesis is falsified if: (a) a document in the NT can be securely dated earlier than any plausible date for Revelation; (b) the literary-structural parallels between Revelation and the later NT books can be shown to run the other direction;

#832 WS-4 verification condition (line 307): to demonstrate direct inheritance rather than parallel derivation.

#828 § III. Method (p17): It does not claim direct literary dependence between the five case studies.

#1443 §2.3 Ithaca (2022): Ithaca is empirical evidence that **the hazard is a property of single-verdict interfaces, not of machine reading as such**, and that ranked-alternative presentation is the mitigation.

#RF census.json rows[slug == 'aeneas'].rf_intersection: Parallels as evidence: the model retrieves the kind of evidence the midrashim transform argues from (for example Rev 2:10 // Jas 1:12). Retrieval finds parallels; the direction of dependence, which the thesis argues, is not something it settles.

#RF census.json rows[slug == 'aeneas'].does_not_answer[0]: the direction of dependence between parallel texts

#capture-registry aeneas-deepmind-aio-20261031, analysis s1: Aeneas is composed as a better Ithaca for Latin.

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 14 (positions [None, None, None, None, 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.

Aeneas is "the first model for contextualizing ancient inscriptions, designed to help historians better interpret, attribute and restore fragmentary texts" (DeepMind), introduced in Nature on 23 July 2025 as a generative neural network for contextualizing ancient texts (Nature; DeepMind). It was co-led by Yannis Assael and Thea Sommerschield and co-developed by Google DeepMind with the University of Nottingham and researchers at Warwick, Oxford and AUEB, Jonathan Prag of Oxford among them; the interface card adds Durham University (DeepMind; Oxford; Predicting the Past card). Parallels are texts sharing wording, syntax, standardized formulas or provenance, or phrasing, function or cultural setting (DeepMind; Nature); Aeneas retrieves them across thousands of Latin inscriptions in seconds by embeddings, beyond fuzzy string matching (DeepMind; Google's blog; Nature). On a Mainz altar a retrieved parallel supports, for the authors, the hypothesis of direct influence of an earlier inscription on a later one (Nature). It determines provenance from text and image, images for place only, restores gaps of unknown length, and scores 73% top-20 restoration, 72% among 62 provinces and dating within 13 years (DeepMind; Nature; MIT Technology Review). Its Latin Epigraphic Dataset holds over 176,000, or 176,861, inscriptions from the seventh century BCE to the eighth century CE, images for 5%, almost 150,000 in MIT Technology Review's account (DeepMind; Nature; MIT Technology Review). Twenty-three historians found its parallels useful starting points in 90% of cases with confidence up 44%, figures that are 75% and 23% for parallels alone (Nature; DeepMind; MIT Technology Review). It is designed to fit historians' workflows as a source of starting points, not to automate epigraphy, a leap beyond Ithaca in Prag's words (DeepMind; MIT Technology Review; Oxford). On the Res Gestae it gave two dating peaks matching both prevailing hypotheses (DeepMind; Nature). Its limits are a small, survival-biased corpus and weaker performance where data is thin; it does not interpret meaning, and its long-term use is unclear (Nature; MIT Technology Review). Trained for Latin, it is stated to be adaptable to other languages and media, from papyri to coinage (Google's blog; Nature; DeepMind). Named for the wandering hero, it builds on Ithaca, which now runs on it (DeepMind; Nature). Its interface is free at predictingthepast.com, its code Apache 2.0 and weights CC BY-SA 4.0 (DeepMind; Predicting the Past; Nature; GitHub card). Wikipedia has no article on it; its Google DeepMind article carries Ithaca, and the name's disambiguation opens with the hero (Wikipedia).

Aeneas is, in its makers' words, "the first model for contextualizing ancient inscriptions, designed to help historians better interpret, attribute and restore fragmentary texts" (Google DeepMind, 23 July 2025), a generative neural network for contextualizing ancient texts set against digital methods "restricted to literal matches and narrow historical scopes" (Assael, Sommerschield et al., Nature, 2025), which the Revelation First record enters as "an AI model that retrieves textual and contextual parallels for ancient inscriptions, introduced in Nature (2025)" and tags "NOT a dating model alone: it retrieves parallels" (Revelation First record). DeepMind introduced it with the Nature paper as "the first artificial intelligence (AI) model for contextualizing ancient inscriptions" (DeepMind), the sentence the record cites for its kind and the date it keeps as its lifecycle, "published July 2025" (Revelation First record); the research was co-led by Yannis Assael and Thea Sommerschield (DeepMind). DeepMind co-developed it with the University of Nottingham and with researchers at Warwick, Oxford and the Athens University of Economics and Business (DeepMind), Jonathan Prag of Oxford's Faculty of Classics among the authors (Oxford Faculty of Classics), and the record carries that list as its operator (Revelation First record); the interface's card names "DeepMind, Durham University, Oxford, and partners" (Predicting the Past card). The Revelation First record asks of it the census's question, "What evidence does each resource hold on when and from what the New Testament writings were composed, and how do machines now read those texts?", places it at the intersection "AI and ancient texts: machine reading, restoration and dating" with machine reading, restoration or dating as its kind of evidence, and heads it with the census's correction, that Revelation First is the priority thesis and not the old pre-70 date (Revelation First record). Parallels are, in DeepMind's definition, texts that "share similarities in wording, syntax, standardized formulas or provenance", which historians find through "their expertise and specialized resources" (DeepMind), and in the paper's, "inscriptions with shared phrasing, function or cultural setting" by which historians contextualize, restore and attribute texts (Nature); the record keeps that work as its finding and takes "an inscription with its retrieved parallels" as its unit (Revelation First record). Aeneas "reasons across thousands of Latin inscriptions, retrieving textual and contextual parallels in seconds", searching "a vast collection of Latin inscriptions" with each text turned into "a kind of historical fingerprint" by embeddings (DeepMind; Google's blog), and ranks parallels from its dataset's training set by historically enriched embeddings that "surpass traditional fuzzy string-matching methods" (Nature), which is the record's method (Revelation First record). What this evidence can settle, the record answers, is "which inscriptions share wording, syntax, formulas or provenance with a given inscription", and what it cannot is "the direction of dependence between parallel texts"; its reading is that the model retrieves "the kind of evidence the midrashim transform argues from (for example Rev 2:10 // Jas 1:12)" and that the direction, which the thesis argues, retrieval does not settle (Revelation First record), the transform being "a formally specified structural operation that maps compressed forms in Revelation to their elaborated expressions in later NT texts" (#202). The thesis makes the direction its falsifier: it is falsified if "the literary-structural parallels between Revelation and the later NT books can be shown to run the other direction" (#202), and its work plan requires the transforms to be shown "not reducible to generic literary influence or shared Second Temple background (the common-substrate confound)", since transforms that run equally from Ezekiel and Daniel to both texts establish "only shared ancestry", not "direct inheritance rather than parallel derivation" (#832, with its versions #1211, #1216, #1217, #1400). The paper, on a votive altar from Mainz, reads a retrieved parallel's "rare textual formulas and an identical iconographical type" as "supporting the hypothesis that the later inscription was directly influenced by the earlier one" (Nature); the archive's philological notes keep parallel and dependence apart, one claiming no "direct literary dependence" between its cases (#828), one holding that terms found together are "a convergence, not a chain of custody" (#1458), and one that verbal parallels are "verifiable in the text" while the claim that a seam was composed is "interpretive and unfalsifiable as stated" (#1474). It is stated to be the first model to determine a text's geographical provenance from multimodal input and to restore gaps of unknown length (DeepMind), though only the geographical head takes images, restoration and dating being text-only (Nature); from a transcription and a scan it gives possible dates, places and fill-ins (MIT Technology Review), with top-20 restoration accuracy of 73% in gaps of up to ten characters (58% at unknown length), the province among 62 at 72%, and dates "within 13 years of the date ranges provided by historians" (DeepMind). Its corpus is the Latin Epigraphic Dataset, cleaned, harmonized and linked from the Epigraphic Database Roma, the Epigraphic Database Heidelberg and the Clauss–Slaby database, "over 176,000 Latin inscriptions" (DeepMind), in the paper 176,861 inscriptions of 16 million characters from the seventh century BCE to the eighth century CE, images for 5% (Nature), "almost 150,000" in MIT Technology Review's account, and in the record "thousands of Latin inscriptions" across the Roman world, its period "as its corpus (not stated in the saved post)" (Revelation First record); of Ithaca's corpus the archive holds that "The dataset construction is the hidden achievement" (#1443). In a study with twenty-three historians the best results came from the model's context and predictions together (DeepMind); the paper's abstract reports parallels useful as starting points "in 90% of cases, improving their confidence in key tasks by 44%", historian and model together outperforming either alone, and dating at a 13-year distance (Nature), while its results give the parts: parallels alone 75% and confidence up 23%, rising to 90% and a further 21% "when Aeneas’ predictions were also available" (Nature), MIT Technology Review reporting "90% of inscriptions". It is "designed to integrate within historians' existing research workflows", its suggestions "valuable starting points for historical inquiry" (DeepMind), not, in Yannis Assael's words, to automate epigraphy but "crafting a tool that will integrate with the workflow of a historian" (MIT Technology Review), and for Jonathan Prag, in a field where "the answers are not black and white, and the data is simultaneously rich and incomplete", a leap beyond Ithaca "based upon previous scholarship" (Oxford Faculty of Classics). Of Ithaca, the model Aeneas builds on, the archive holds that the historians' gain with the model "cannot come from accuracy" and comes from interface design, "ranked top-20 hypotheses with saliency, not a verdict", and that Ithaca is evidence that the hazard named by differential register prioritization (#1406) is "a property of single-verdict interfaces, not of machine reading as such", with ranked alternatives its mitigation (#1443). On the Res Gestae Divi Augusti it produced not a single date but "a detailed distribution of possible dates, showing two distinct peaks", which "captured both prevailing dating hypotheses in a quantitative way" (DeepMind; Nature). Its stated limits are the corpus, small beside the datasets of natural-language processing and subject to "inscription survival bias", with performance "weaker where data is limited" (Nature); Kathleen Coleman notes that it "doesn’t guess the meaning of texts", cannot interpret new finds on its own and was tested on one of the best-studied inscriptions (MIT Technology Review); and the archive notes that machine readers of ancient texts "work on authors who cannot be consulted" (#1443). It was trained for Latin, chosen because few machine-learning studies treat Latin inscriptions (Google's blog; Nature), and is stated to be adaptable "to other ancient languages, scripts and media, from papyri to coinage" (DeepMind), in the paper "to any ancient language and written medium (such as papyri, manuscripts or coinage)" (Nature); the record keeps Latin as its text and stated limit, the adaptation as stated, and tags it "NOT trained on Greek or on papyri in the published version" (Revelation First record). The record answers nothing of the New Testament with it, and its worked question on machine dating, premised that "Machine dating is calibrated on a corpus", names it for "parallels for Latin inscriptions, the kind of evidence dependence arguments use", with the gap that none of its four resources "dates a New Testament papyrus or a composition" (Revelation First record). Named, by both its post and its paper, after "the wandering hero of Graeco-Roman mythology", it builds upon Ithaca, DeepMind's model for ancient Greek inscriptions, which the release upgraded "to be powered by Aeneas" (DeepMind; Nature); the record names Ithaca as its neighbour, "DeepMind's earlier model for Greek inscriptions" (Revelation First record), and the encyclopedia's Google DeepMind article carries Ithaca and names no Aeneas (Wikipedia). An interactive version is free at predictingthepast.com, a page titled "Contextualising, restoring, and attributing ancient texts", and its code and dataset are open (DeepMind; Predicting the Past), the code under Apache License 2.0 and the weights under CC BY-SA 4.0 at the google-deepmind/predictingthepast repository (Nature; GitHub card); the record gives the interface as its address and alias, "an interactive version" as its access and its licence as "not read", and lists the Nature paper among what it left unchecked (Revelation First record). Wikipedia has no article on the model, and the name's disambiguation opens "Aeneas was a Trojan hero in Virgil's Aeneid and Homer's Iliad" and lists no model (Wikipedia). At 'Aeneas DeepMind' on Google's AI Overview, 31 October 2026, a new address at which no card cites the record built the day before, the operator's disambiguator selects the model from the hero, the composition is of the model throughout, contextualization first, and its Latin inscriptions, its date and its parallels are the record's (Capture Registry, 2026-10-31), on the innocent reading "the model, its corpus, its tasks and its tested usefulness, from its maker, its paper and its press" (Capture Registry, 2026-10-31). The registry reads it as "a better Ithaca for Latin": the retrieval of parallels is kept and the stated reach to papyri, "the claim that would bring it to the field", is left out, and Ithaca appears only in a closing offer to compare, not as the post's "builds upon" (Capture Registry, 2026-10-31; DeepMind). The body names Google DeepMind alone where DeepMind's post names its partners, offers the interface without the commitments of free access and open code and dataset, gives no person and no identifier, its one way to the work an alias, and scores a dimensional PER of 0.421; the registry finds its larger figures from the cards and does not charge those the saved post does not state (Capture Registry, 2026-10-31; DeepMind).

T against L_F(B) (AIO's ontology at work on its own cards): At 'Aeneas DeepMind' (T1, seven cards: B1–B7, B4 unreachable; 151 words, a closing offer of 27) the composition gives the model, its corpus, its tasks and its tested usefulness. Composed from its cards: the definition at the grain of the post and the paper ('contextualize, restore, date, and locate', B1 F1, B5 F24, B6 F44), the date (B1 F2, 'July 2025'); the corpus (B1 F11 'over 176,000'; B5 F28 'seventh century bce to the eighth century ce', as '700 BCE to 800 CE'); multimodal input and gaps of arbitrary length (B1 F10; B4 F23); the top-20 and the 13 years (B1 F13; B5 F26); the 90% and 44% (B5 F26); parallels as 'shared syntax or formulas' (B1 F3) and 'useful contextual parallels' (B1 F4); the interface (B6 F44); Ithaca in the closing offer (B1 F8, as comparison only). Available, not composed: the partners (B1 F6, B3 F21) and the co-leads (B1 F18), the body naming Google DeepMind alone; the stated reach to 'papyri to coinage', 'any ancient language and written medium' (B1 F5, B2 F19, B5 F33); free access and open code and dataset, the licences (B1 F7, B5 F36); Ithaca's upgrade to run on Aeneas (B1 F17); the embeddings and the definition of parallels (B1 F12, B5 F25, F30); the Res Gestae (B1 F14, B5 F32); the Mainz altar's reading of direct influence (B5 F31); the corpus's source databases (B1 F11, its first part). Limit dropped: the paper's own qualification of its figures, that 90% and 44% hold only when the model's predictions are added to its parallels, parallels alone giving 75% and 23% (B5 F27): T1-6 gives the 90% and 44% to the parallels; the stated limits of the corpus, its size and survival bias, weaker where data is thin (B5 F34, F35); Coleman's limit, that it does not guess meaning and was tested on a famous stone (B7 F42). Force raised: T1-3 attaches the image input to 'shared syntax or formulas'; the paper confines images to geographical attribution, restoration and dating text-only (B5 F29). Substituted reading: T1-2 sets '62 Roman provinces' as the corpus's span; in the field the 62 provinces are the attribution task's classes (B1 F13, B5 F29's head). Composed without attribution: none at the name (the maker named; the figures carried with markers to the cards). Outside the cards: none of T's claims lacks a card; 'typically dating … of expert historian estimates' stands nearest B1 F13 'date ranges provided by historians'. Not surfaced, admitted by the operator's ruling: no own page stands outside the rail (the census's own pages are cards B1 and B6); Wikipedia has no article (B8 F46–F48). Of the cards' 45 readable claims, T composes 13 at their grain or coarser.

v0.9 §0.5: the first diff is ontological and narrow. T and the field agree on the entity (DeepMind's model for Latin inscriptions). T composes the maker's model at the grain of its tasks and figures and leaves out what the field says about its reach (papyri, any written medium), its makers beyond DeepMind, its openness, and the paper's own qualification of the figures it carries, the 90% and 44% that belong to parallels with predictions, not to parallels.

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the Revelation First record's account, field by field, with its basis (RF-01–RF-30): parallels as the evidence the midrashim transform argues from, and what that evidence can and cannot settle, 'the direction of dependence between parallel texts' among what it cannot (RF-14, RF-15, RF-17, RF-28), nothing of the New Testament (RF-16, RF-29); the thesis's own falsifier on the direction of parallels (O202-01) and its work plan's common-substrate confound (O832-01), which set what a retrieved parallel would have to show to bear on priority, beside the paper's reading of a parallel as direct influence (F31, held under P_coh); the archive's philological notes keeping parallel apart from dependence (O828-01, O1458-01, O1474-01); of Ithaca, the model the field says it builds on, the gain located in interface design and the hazard of single-verdict interfaces (O1443-01, O1443-02), the dataset as the hidden achievement (O1443-03), and authors who cannot be consulted (O1443-04); the record's own state where the field is fuller (its count 'thousands', RF-11, against F11 and F28; its licence 'not read', RF-19, against F36); and the Capture Registry's account of the composition at the name (CAP-1-01–CAP-1-15). Not surfaced: every archive source; no card cites the census record (CAP-1-14).

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', 'F39', 'F40', 'F41', 'F42', 'F43', 'F44', 'F45'], 'T1_composed': ['F1', 'F24', 'F2', 'F11', 'F28', 'F10', 'F13', 'F26', 'F4', 'F44', 'F23', 'F8', 'F3'], 'T1_composed_from_own_page': [], 'T1_outside_cards': [], '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"}