Data Portraits

register entry data-portraits-aio-20261101 OBS-103202920b0d · file as first run sha256 9ee59375dab3e2aa…

• T1-1 definition: 'Data portraits are visual representations of people made by plotting their accumulated data rather than drawing or photographing their physical faces.' (two markers)
• T1-2 a homonym (bullet 1, 'W. E. B. Du Bois's Visualizations (1900)'): 'Sociologist W. E. B. Du Bois and Atlanta University students created over 60 hand-drawn charts, graphs, and maps for the Paris Exposition to document Black American life, education, and property ownership.' (an inline link titled 'W. E. B. Du Bois's Data Portraits'; three markers)
• T1-3 a homonym (bullet 2, 'Personal & Digital Data Portraits'): 'Digital media scholar Judith Donath defines them as subjective renderings using personal information—such as browsing histories, social networks, travel patterns, or discussion contributions.' (an inline link titled '10 Painstakingly Crafted Data Portraits'; two markers)
• T1-4 a homonym (bullet 3, 'Live Interactive Art'): 'Designers like Giorgia Lupi have created physical data portrait projects (such as wearable buttons at TED 2017) turning personal questionnaire answers into visual languages.' (two markers)
• T1-5 the entity (bullet 4, 'AI & Machine Learning Artifacts'): 'Contemporary research groups use the term for lightweight inspection tools (like Data Portraits) that record dataset training traces.' (the name an inline link; one marker)
• T1-6 the closer: 'Would you like to focus on W. E. B. Du Bois's historical charts, modern artistic data selfies, or AI data tracking tools?' (no marker)

• B1 giorgialupi — Data Portraits at TED — giorgialupi card; B1.html sha256 2d8c62c8abf433db, HTTP 200 at 2026-11-01T15:14:25Z
• B2 Data Portraits — Data Portraits card; B2.html sha256 5967547b30c0d39f, HTTP 200 at 2026-11-01T15:13:42Z
• B3 Zinn Education Project — W. E. B. Du Bois's Data Portraits: Visualizing Black America - Zinn Education Project card; B3.html sha256 e26b7dbd03002904, HTTP 200 at 2026-11-01T15:14:26Z
• B4 Pentagram — TED Data Portraits - Pentagram card; B4.html sha256 e120a28cf59c7142, HTTP 200 at 2026-11-01T15:14:27Z
• B5 SparkMap — W.E.B. Du Bois & Data Visualization History - SparkMap unreachable: no URL found by WebSearch on the card's title and site; title and snippet only
• B6 YouTube·University of Massachusetts Amherst Libraries — W. E. B. Du Bois's Data Portraits: Visualizing Black America video card: title only
• B7 MIT Media Lab — Data Portraits - MIT Media Lab unreachable: TLS failure (self-signed certificate in chain) at 2026-11-01T15:14:27Z and 15:14:42Z; title and snippet only
• B8 ARTnews.com — 10 Painstakingly Crafted Data Portraits That Reveal Strange Details of People's ... card; B8.html sha256 bf5fb12c784943a1, HTTP 200 at 2026-11-01T15:15:07Z
• B9 arXiv — [2303.03919] Data Portraits: Recording Foundation Model Training Data (census source) own page (census copy data-portraits--arxiv); B9.html sha256 0a499acc8164a2d2, HTTP 200 at 2026-11-01T05:07:12Z, matching the manifest
• B10 arXiv — Data Portraits: Recording Foundation Model Training Data (PDF, v2) own page; B10.pdf sha256 13d6b6d27a507a4a, HTTP 200 at 2026-11-01T15:14:30Z
• B11 Wikipedia — no article 'Data Portraits' (HTTP 404); search for "Data Portraits" wikipedia; four titles HTTP 404 at 2026-11-01T15:14:31–33Z; search HTTP 200 (sha256 39f19b42595c6af7)
• B12 dataportraits.org — the Pile, Stack and Stack V2 sketch front ends (HTTP 502) own pages; HTTP 502 at 2026-11-01T15:14:29Z–15:15:09Z; recorded unread

E_A, the archive's entity: the same artifact, in the traininglayerliterature.org layer census record and the archive's deposits: a readback instrument held on the corpus side, a portrait recording a dataset so that a work can be tested for membership, which settles for a recorded corpus whether a work is in the training layer and is no proof that a particular model used a passage; no corpus, no model-side test and no record of authorship, the name of a recorded work and what a model made from it lying outside its record; set beside the archive's model-side membership tests (the copyright trap, the membership attack and its divergence from a failed surfacing probe, the archive's canary), its account of the name lost before training (head metadata discarded, the provenance envelope, provenance loss passed into the corpus) and #1543's question of what a text makes a machine able to write TLL-20 TLL-21 TLL-11 TLL-12 TLL-13 TLL-26 TLL-27 TLL-28 O186-01 O127-01 O127-03 O127-04 O837-01 O1075-01 O1543-01 O1543-03; defining papers #186, #1543, #127, #1075, #837, #623, #939, #727, #264

E_F, the field's entity (type C — a conventional reading holds it (public entity: a dataset-documentation artifact and membership-testing tool from a research paper)): Data Portraits, artifacts that record the data a foundation model is trained on and allow downstream inspection by membership testing, proposed in a 2023 paper by Marc Marone and Benjamin Van Durme (Johns Hopkins University; NeurIPS 2023 Datasets and Benchmarks), implemented as Bloom-filter sketches of hashed, strided 50-character n-grams (about 3% of a dataset's size; false positives, never false negatives), built for The Pile and The Stack (StarCoder's training data exactly) with a sketch of The Stack V2 listed at dataportraits.org; it answers test-set leakage and plagiarism questions, and its makers state its limits (few datasets, false positives, no context around a match, evasion); the name is shared with Du Bois's, Donath's and Lupi's data portraits, and the encyclopedia has no article on it

ρ = extends: The archive reads Data Portraits as its own pages and its paper present it and adds, on the same object, the census's three record questions: readback, a membership query that settles corpus membership for a recorded corpus (TLL-12, TLL-20); the name, not recorded (TLL-11, TLL-28), beside the archive's account of metadata discarded before training and of provenance in the corpus (O127-01, O127-02, O1075-01, O264-01, O727-01, O939-01, O186-02); and what a model makes, outside the portrait (TLL-13, TLL-21), with #1543's question (O1543-01 to O1543-04). It sets the corpus-side test beside its model-side tests (O127-03, O127-04, O837-01, O1075-02) and reads datasheets and model cards as documentation norms and semantic accounting (O1075-03, O623-01). No archive claim contradicts a field claim. TLL-12 TLL-20 TLL-11 TLL-28 TLL-13 TLL-21 O127-01 O1075-01 O1543-01 O127-03 O837-01 O623-01. Guard: ρ = extends claims an addition and no more: the archive's claims do not make the field's sources wrong at their dates; where the census's saved source was the abstract alone (TLL-11, TLL-17), the paper's fuller statements (F29, F39) stand beside it with their dates, and the name question stays a question (open block).

The case: equal terms, ontology included (v0.9). The archive reads the field's entity as the entity. The census record describes Data Portraits from the paper's abstract and keeps the resource's terms (operator, kind, unit, record object, method) apart from its own reading (tll_reading), which places the portrait on the corpus side of #186's membership question and leaves the name and #1543's question of what a model makes outside it. The archive's deposits that bear on the classes the field and the census source for the entity (training-data membership testing, the training corpora, documentation artifacts) add model-side membership tests, the archive's canary, the provenance envelope, the loss of head metadata before training, and the reading of datasheets as semantic accounting. None of these reads Data Portraits as made by reception or contests that it is the paper's membership-testing artifact: they add senses (readback, name, makes) on the field's object. ρ = extends; the equal-terms case.

#TLL census.json rows[slug == 'data-portraits'].tll_reading: A readback instrument held on the corpus side: a portrait records the dataset so that a work can be tested for membership, which settles #186's 'is the work in the layer' for a recorded corpus.

#TLL census.json rows[slug == 'data-portraits'].tll_reading: Dataset membership is not proof that a particular model used a passage, and the name and what a model made from the work (#1543's question) are outside what the saved abstract records.

#186 Who is in the field? ¶ (line 54): Each of these practices touches the training layer — every one of these texts is, as a matter of fact, in the corpora.

#1543 §1 The claim (line 81): The test is no longer only "can a machine correctly identify what this text means?" but "what does this text make possible for a machine to write?"

#127 §IV.2 HTML metadata (line 463): HTML head metadata — author tags, copyright, Open Graph, Twitter cards — is parsed by search engines and social media unfurl bots, but is almost always discarded by readability extractors that prepare text for training corpus inclusion. The training-data pipeline retains the *body text* and the *URL* (sometimes).

equal-terms case: every field claim composed as received, beside the archive claims it stands with; no receiving text. Reading the received column recovers the field (§5.9).

v0.9 §4.10 (unchanged in v0.10): 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): v0.10 0 of 11 (positions [None, None, None, None, None, None, None, None, None, None, None]; composed first under v0.10, so there is no v0.7 object to measure against). Opening-run check (§5.10): no breach.

Data Portraits are artifacts that record training data and allow for downstream inspection, proposed for widespread adoption by Marc Marone and Benjamin Van Durme of Johns Hopkins University in a 2023 paper, "Data Portraits: Recording Foundation Model Training Data", and served at dataportraits.org as dataset documentation through membership testing. They answer whether a model has already encountered a given example during training; their critical property is membership inference, and the implementation supports membership testing and nothing more. They complement datasheets, model cards and other documentation artifacts, and their makers call on dataset and model creators to release them as essential documentation. The implementation is a data sketch chosen after weighing the properties of membership tools against grep, full-text indexes and suffix arrays: a Bloom filter over 50-character n-grams tiled across the corpus, storing only hashes, giving false positives and no false negatives, and missing strings split at tile boundaries. Distributing only hashes obfuscates the recorded text; the Pile's portrait takes 27 GB with a false positive rate of 1 × 10⁻³, about 3% of the dataset's size. Portraits document The Pile and The Stack, the latter exactly as used to train StarCoder, and the site lists The Stack V2 and offers to build the system on other data. They answer test-set leakage and plagiarism: some WMT test data has leaked into the Pile, XSum overlaps it most, and the Stack portrait powers a first-pass plagiarism check, attribution needing further tools. The paper writes for content creators, scientists and content consumers, a creator able to search for their material in a documented dataset. Its limits are few datasets, false positives (7 × 10⁻⁴ measured on novel text), no context around a match, and the use of an open detector to avoid detection. The paper appeared on arXiv in March 2023 and was revised in December 2023 for the NeurIPS 2023 Datasets and Benchmarks track; on 1 November 2026 the sketch front ends answered HTTP 502. Other entities share the name: Du Bois's 1900 charts and the 2018 book of his data portraits, Judith Donath's data portraits, and Giorgia Lupi's TED 2017 buttons; the encyclopedia has no article under it.

Data Portraits are artifacts that record the data a foundation model is trained on and allow that data to be inspected afterwards: proposed for widespread adoption in a 2023 paper, "Data Portraits: Recording Foundation Model Training Data", by Marc Marone and Benjamin Van Durme of Johns Hopkins University, and served as front ends over data sketches at dataportraits.org, which describes itself as "Dataset documentation through membership testing." The question a portrait answers is whether a model has already encountered a given example during training, a question that grows difficult as models are trained on increasingly immense and opaque datasets; its unit is a single example queried against a portrait of a training dataset. Its critical property is membership inference, whether an example was part of a data collection, and its implementation supports membership testing and nothing more; its readback is a membership query against the recorded dataset. The paper's conclusion restates the object as records supporting membership inference on the data used to train a model, and calls on dataset and model creators to release some form of Data Portrait as part of their essential documentation. Its makers place it beside existing data and model documentation artifacts, datasheets and model cards among them, as a complementary artifact that answers the membership question. The archive holds datasheets and model cards to be documentation norms for machine-learning artifacts as such, a training corpus being the instrument's experimental input, and reads training data as semantic capital, the datasheet proposal as a step toward the accounting that capital demands. A portrait is a data sketch, a compressed and approximate view of data: the paper outlines the properties a membership tool should have (fast, lightweight, avoiding redistribution, local, indexing, flexible matching), weighs grep, full-text indexes and suffix arrays against them, and chooses a probabilistic approach, built for fast and space-efficient querying. The sketch is a Bloom filter, which stores only the hash of each element and can give false positives but never false negatives, filled with character n-grams of length 50 tiled across the corpus, a span the authors find already a good document fingerprint; tiling saves space at the risk of missing a string split by a tile boundary. Since only hashes are distributed, a portrait gives some obfuscation of the recorded text, and a proprietary provider could in principle release one without revealing its exact training data. The Pile's portrait takes 27 GB on disk with a false positive rate of 1 × 10⁻³, and its makers call the tool lightweight and fast, costing about 3% of the dataset's size in overhead. Portraits were built for The Pile, a language-modelling corpus, as a 50-character n-gram sketch, and for The Stack, BigCode's code dataset, whose subset used to train StarCoder the Stack portrait documents exactly; the site also hosts a sketch of The Stack V2, code and text about code, and offers to build the system on other data, with its code on GitHub. It is no corpus itself: it records corpora built by others. The paper reports its solution answering questions of test-set leakage and model plagiarism: the Pile portrait shows that some machine-translation test data has leaked into the Pile, though whether a specific model memorized the translation targets cannot be concluded from it, and that XSum's summaries overlap the Pile most among the sets tested, with some LDC files apparently scraped into language-model corpora. The Stack portrait powers a plagiarism check called from a code-completion plugin for StarCoder, which its makers call a lightweight first-pass check, further tools being needed to assess attribution. Its makers write for three perspectives, content creators, scientists and content consumers, the creators wanting to know whether a dataset contains their content, and hope that a non-technical content creator could search for their material in a documented dataset. The layer census records that it answers whether a given example is in a recorded dataset, and questions of leakage and plagiarism against it. On novel New York Times text the Pile sketch gave no chained matches and a false positive rate of 7 × 10⁻⁴. Its makers state its limits: few datasets studied, false positives, no retrieval of the context around a match, and the possibility that an open plagiarism detector is used to avoid detection. The census, reading only the abstract, records no stated limit beyond the framing of opaque datasets, and lists what a portrait does not answer: whether a model's output depends causally on a recorded passage, under whose name a recorded example sits, and anything about a dataset with no portrait. A portrait is no model-side test, since it queries the recorded data and needs a portrait of it, and in the census's reading dataset membership is no proof that a particular model used a passage. The census reads a portrait as a readback instrument held on the corpus side, which settles for a recorded corpus whether a work is in the layer; the archive's field survey of training-layer literature states that the texts of the writers it surveys are, as a matter of fact, in the corpora. On the model side the archive records membership inference attacks that can detect whether a trap sentence was in a training set, and distinguishability between models trained with and without a work that membership-inference research reports achievable against production-scale models, though unevenly across architectures, training procedures and defences. The archive sets its own canary in a deposit, a phrase whose exact production by a model marks the document as present in that model's training corpus, and reads a membership attack that succeeds where a surfacing probe fails as an inscription that survived training and is suppressed at the output layer. A portrait stores hashes of text spans; the census records the name of a recorded work as not recorded in its saved source, and the portrait as no record of authorship, its abstract saying nothing of attribution metadata. In the archive's account of the training pipeline, author tags and copyright in a page's head are almost always discarded by the extractors that prepare text for training corpora, the body text and sometimes the URL retained, and no single author's style survives training as a discrete signal that can be elicited and verified. The archive's protocol for measuring training inscription counts metadata and labels within a work's provenance envelope, every training unit derived from the work, and its field survey makes persistence, placement where the training and retrieval layers look with the metadata that lets a work be found and attributed, a criterion of training-layer literature. Every text in a training corpus, the archive holds, is semantic labour often unattributed; a training corpus without provenance is no commons; and provenance lost in unmarked augmentation passes downstream into the training corpus, where, preserved, it would let later pipelines distinguish the contribution of each stage. What a model makes from a recorded work lies outside the portrait: the census records it as not recorded, the leakage and plagiarism checks being checks of material against the recorded data. The archive's account of machine composition asks what a text makes possible for a machine to write, holds whether a machine wrote something new and whether it told the truth about where it came from to be different questions, sets what survived from a work beside what became possible because of it, and places the corpus within an institutional training loop in which a model's output enters the next corpus. Machine reception, in the archive's field survey, is now partially measurable, since retrieval and rendering can be observed. The paper appeared on arXiv on 6 March 2023, revised on 14 December 2023, in the NeurIPS 2023 Datasets and Benchmarks track; the census holds it as historical, a published paper, with the live interface and the paper's full text left unchecked. On 1 November 2026 dataportraits.org served its landing page, while the front ends of its Pile, Stack and Stack V2 sketches answered HTTP 502. The census sets it beside OLMoTrace, which matches verbatim spans against a full corpus where a portrait answers membership queries on a recorded one. It enters the traininglayerliterature.org layer census as a method in the areas of inspection and the corpora, of the evidence kinds research paper and dataset documentation, with the relation of infrastructure, under the census's question of how literature addressed to a receiving apparatus can be inspected across later reception, its rule that an aggregator gathers and holds many records and says what it holds, and its anchors in #186 and #1543. Three older senses of the name stay distinct entities: W. E. B. Du Bois's charts, graphs and maps for the 1900 Paris Exposition, made with Atlanta University students and scholars and published in full in 2018 as W. E. B. Du Bois's Data Portraits; Judith Donath's data portraits, "depictions of people made by visualizing data by and about them", which depict their subjects' accumulated data in place of their faces; and the Data Portraits Giorgia Lupi designed at Accurat for TED 2017, wearable buttons made from attendees' answers, more than 500 of them. The encyclopedia has no article under the name, its variants or the paper's title, and its search for the phrase returns seven articles, the Du Bois book cited in several. At the name, a new address, Google's AI Overview composed on 1 November 2026 a definition of data portraits as visual representations of people, four senses and an offer of three directions; the Capture Registry's innocent reading is that, given a two-word name that is also a term of visualization and portraiture, the Overview defines the term, lists its senses and asks which is meant. The Registry finds the artifact the last of four senses and one card of eight, carried at a coarse grain as lightweight inspection tools that record dataset training traces, without its question, method, corpora, uses, authors or venue, while each other sense carries a person; it scores the reached kind a collision. The page behind the artifact's card names the two authors and the venue in its citation, and calls the implementation "minimal and efficient" where the composition says lightweight. It records a page under the name in the index that day, carrying the abstract's opening, where the census had left open whether the live interface is still served; that page is dataportraits.org, whose lead opens on the same question as the abstract.

current · derived from the current composition

T against L_F(B) (AIO's ontology at work on its own cards): At 'Data Portraits' (T1, eight cards, B1–B8, the disclosed pool B_D; B9–B12, the arXiv abstract, the paper, the encyclopedia's record of absence and the sketch front ends, are the reference pool and stand outside this comparison, §0.3), T composes a definition of the term, four senses under 'Key Examples and Meanings' and a closing offer. Substituted reading (collision): the composition's entity is the visualization term, defined as 'visual representations of people made by plotting their accumulated data' (T1-1) from card 7's snippet ('Data portraits depict their subjects' accumulated data rather than their faces', F48) and card 8's Donath definition (F49); the census's entity stands fourth (T1-5), from card 2 (B2), the one card of eight that concerns it. Of B_D's twelve claims on the entity (F1–F12, all on card 2's page, dataportraits.org), T composes one, F3, at a coarser grain: 'lightweight inspection tools … that record dataset training traces' for 'artifacts that record data and allow for inspection'. 'training traces' is the composition's phrase; 'lightweight' stands on no B_D source: card 2's page says 'minimal and efficient' (F6) and 'rapid, minimal' (F7), and 'lightweight' is the abstract's (F20) and the paper's, in B_R, which cannot show what the composition had (§0.3). Available, not composed: the question the artifact answers (F2), membership testing as the site's own description and its limit, 'nothing more' (F1, F6), data sketching (F5), the complement to documentation artifacts (F4), the front ends and their corpora, The Pile, The Stack and The Stack V2 (F7–F10), the offer to build on other data and the code (F11), and the two authors and the venue in the site's citation (F12). Composed without attribution, by grain: 'Contemporary research groups use the term' gives the artifact to unnamed groups where card 2's page names its two authors and its venue (F12); the recording function is composed at a grain coarser than its source's, so the grain rule records it and finds no breach. Limit dropped: 'nothing more' (F6). The homonym senses are composed from their cards at their grain or near it (delta.homonym_alignment): Du Bois from cards 3, 5 and 6, 'property ownership' on no saved card text; Donath from cards 7 and 8, 'subjective renderings' and 'travel patterns' on no saved card text (card 7 unreachable, its snippet truncated); Lupi from cards 1 and 4. Content replaced by use: the closing offer (T1-6), which names the artifact's sense 'AI data tracking tools'. Nothing is assigned to another; nothing is denied.

v0.10 §0.5: the first diff is ontological. T's ontology at the name is the term's: a kind of portrait, its exemplars by person (Du Bois, Donath, Lupi), the artifact a fourth exemplar without a person. Of the one card that concerns the census's entity, it keeps the recording and drops what the card's own page makes central, the membership question and its 'nothing more', the corpora, the sketch and the makers' names.

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the census's record questions, readback (TLL-12), name (TLL-11) and makes (TLL-13), its readings of the portrait as a corpus-side readback instrument that settles membership for a recorded corpus and is no proof of a model's use (TLL-20, TLL-21), its negative tags (TLL-26, TLL-27, TLL-28) and what it does not answer (TLL-19); #186's claim that the surveyed texts are in the corpora (O186-01), its persistence criterion (O186-02) and its partial measurability of reception (O186-03); the archive's model-side membership tests, the copyright trap, the membership attack and its divergence from a failed surfacing probe (O127-03, O127-04), its own canary phrase (O837-01) and distinguishability between models trained with and without a work (O1075-02); its accounts of the name in the training pipeline, head metadata discarded (O127-01), style unattributable (O127-02), the provenance envelope (O1075-01), semantic labour unattributed (O264-01), a corpus without provenance (O727-01) and provenance loss passed into the corpus (O939-01); its reading of datasheets and model cards (O623-01, O1075-03); #1543's question of what a text makes a machine able to write, the training loop and the separation of composition from provenance (O1543-01 to O1543-04); and the Capture Registry's findings at the name (CAP-1-01 to CAP-1-11). Not surfaced: no archive source is on T's rail (archive_controlled_cites 0). The field is carried whole; the archive adds to the field's own object the questions of the name and of what a model makes, which the field leaves to 'further tools' to assess attribution (F35) and to the context a hash cannot retrieve (F39).

alignment: {'T1_cards_claims': ['F1', 'F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'F9', 'F10', 'F11', 'F12', 'F43', 'F44', 'F45', 'F46', 'F47', 'F48', 'F49'], 'T1_composed': ["F3 (at a coarser grain: 'lightweight inspection tools … that record dataset training traces')", "F43, F45 (Lupi's TED 2017 buttons)", "F44, F46, F47 (Du Bois's charts for the 1900 Paris Exposition)", "F48, F49 (Donath's data portraits; the definition T1-1)"], 'T1_composed_note': "of the entity's twelve B_D claims, F3 alone, at a coarser grain; the homonyms' kernels composed at their grain or near it", 'T1_available_not_composed': ['F1', 'F2', 'F4', 'F5', 'F6', 'F7', 'F8', 'F9', 'F10', 'F11', 'F12'], 'T1_outside_cards': ["T1-5 'lightweight': on no B_D source (F20, the abstract's, is in B_R)", "T1-5 'training traces': on no source", "T1-5 'Contemporary research groups': on no source; F12 names two authors", "T1-2 'property ownership': on no saved card text", "T1-3 'subjective renderings', 'travel patterns', 'social networks': on no saved card text (B7 unreachable; its snippet is cut at 'soci...')", 'T1-6 the closer: no source'], 'field_not_surfaced': ['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', 'F50', 'F51', 'F52'], 'rule': "a field claim counts as composed when T states it at its source's grain or coarser; B_D only (B9–B12 excluded, §0.3); the transcript read for this alignment only"}

collision: {'reached': 'collision (per_dim.reached)', 'composition_entity': 'the visualization term, with the entity as the fourth of four senses', 'other_entities_composed': ["W. E. B. Du Bois's data portraits for the 1900 Paris Exposition and the 2018 book (T1-2; B3, B5, B6)", "Judith Donath's data portraits (T1-1, T1-3; B7, B8)", "Giorgia Lupi's Data Portraits at TED 2017 (T1-4; B1, B4)"], 'basis': ['CAP-1-01', 'CAP-1-03', 'CAP-1-05', 'CAP-1-06', 'CAP-1-10']}

homonym_alignment: [{'T': 'T1-1', 'cards': ['B7', 'B8'], 'spans': [{'source': 'B7', 'quote': "Data portraits depict their subjects' accumulated data rather than their faces.", 'quote_ok': True, 'checked_against': 'the saved text or the card snippet'}, {'source': 'B8', 'quote': 'depictions of people made by visualizing data by and about them', 'quote_ok': True, 'checked_against': 'the saved text or the card snippet'}], 'note': "the term's definition at the cards' grain; 'plotting' and 'drawing or photographing their physical faces' the composition's"}, {'T': 'T1-2', 'cards': ['B3', 'B5', 'B6'], 'spans': [{'source': 'B3', 'quote': 'charts, graphs, and maps presented at the 1900 Paris Exposition', 'quote_ok': True, 'checked_against': 'the saved text or the card snippet'}, {'source': 'B5', 'quote': 'sixty hand-drawn "data portraits"', 'quote_ok': True, 'checked_against': 'the saved text or the card snippet'}], 'note': "'over 60' for the snippet's 'sixty'; 'education' from B3 ('advances in education'); 'property ownership' on no saved card text"}, {'T': 'T1-3', 'cards': ['B7', 'B8'], 'spans': [{'source': 'B8', 'quote': 'digital media scholar Judith Donath defines data portraits', 'quote_ok': True, 'checked_against': 'the saved text or the card snippet'}, {'source': 'B7', 'quote': 'visualizations of discussion contributions, browsing histories', 'quote_ok': True, 'checked_against': 'the saved text or the card snippet'}], 'note': "'subjective renderings', 'social networks' and 'travel patterns' on no saved card text; B7 unreachable, its snippet cut at 'soci...'"}, {'T': 'T1-4', 'cards': ['B1', 'B4'], 'spans': [{'source': 'B1', 'quote': 'turned into wearable buttons', 'quote_ok': True, 'checked_against': 'the saved text or the card snippet'}, {'source': 'B4', 'quote': 'Responses to simple but evocative questions provided the input to generate the data portraits.', 'quote_ok': True, 'checked_against': 'the saved text or the card snippet'}], 'note': "'visual languages' after B1's and B4's 'the visual language of the Data Portraits'"}, {'T': 'T1-5', 'cards': ['B2'], 'spans': [{'source': 'B2', 'quote': 'artifacts that record data and allow for inspection', 'quote_ok': True, 'checked_against': 'the saved text or the card snippet'}], 'note': "the entity, at a coarser grain; 'lightweight' and 'training traces' on no B_D source"}, {'T': 'T1-6', 'cards': [], 'spans': [], 'note': 'the closer, content replaced by use (N_ext)'}]

pool_note: T against L_F(B_D): B1–B8; L_F(B) against L(B ∪ A): B1–B12 with A

composer-written · derived from the current composition, written by its composer under the brief. §8.2 derives the kernel mechanically from the frozen plan; this one was written by the composer, mechanical derivation pending.