{"address":{"q":"What is “AI Overview Capture Registry — EA-WG-CAPTURES-01 v12.167 — 633 semantic addresses, 811 captures, 2026-06-07 to 2026-10-12” by Lee Sharks (Alexanarch deposit #1676, AXN:06F2), and how does it work? Ground your answer in the document itself.","record":"capture-registry-v12167-ask-chatgpt-20261030"},"analysis":{"analysis":"The composition knows the registry well and the snapshot less well. It retrieves the archive's later accounts of the instrument (the dimensional rubric of 20 October, the substitution ruling of 15 October, the founding paper's sample caution) and gives them as the content of a document of 12 October that contains none of them. The snapshot's own §7, which records 813 declared against 811 held as data, is read as a later contradiction. Scored against the Capture Registry, of which the deposit is a snapshot, the composition carries every required atom but the identifier, whose reasons it does not give.","d":"THE RECORD PAGE'S ASK CHATGPT LINK FOR #1676, SIGNED OUT: the composition gives the snapshot's counts, its citable unit, its data flow, its version chain and its ChatGPT growth correctly, and composes the registry's method and measurement from other archive surfaces, which it assigns to 'the October 12 document': PER-D as purpose and commitments (rubric v1.1 of 20 October), substitution at 1.00, the PER formula, the attribution states, the Claude readers. It reads the snapshot's own reconciliation of 811 and 813 as a 'later archival discussion' to be guarded against, and closes on limits and disclaimers the deposit does not state, among them that the registry is 'not proof of deliberate suppression'.","d_full":"THE RECORD PAGE'S ASK CHATGPT LINK FOR #1676, SIGNED OUT: the composition gives the snapshot's counts, its citable unit, its data flow, its version chain and its ChatGPT growth correctly, and composes the registry's method and measurement from other archive surfaces, which it assigns to 'the October 12 document': PER-D as purpose and commitments (rubric v1.1 of 20 October), substitution at 1.00, the PER formula, the attribution states, the Claude readers. It reads the snapshot's own reconciliation of 811 and 813 as a 'later archival discussion' to be guarded against, and closes on limits and disclaimers the deposit does not state, among them that the registry is 'not proof of deliberate suppression'.","findings":["THE SNAPSHOT'S OWN FIGURES RIGHT. 633 addresses, 811 captures, 84 dates, 7 June to 12 October 2026 (L37); 'ONE SURFACE + ONE SEMANTIC ADDRESS + ONE DATE = ONE CAPTURE.' (L43); the single writable authority since 2026-08-17 (L103); the chain v6.0 (#825) to v12.167 with v7.2 as #1396 and #1401 (L128); ChatGPT at 3 captures at v10.9 and 167 now (L136); 'The transcript is the capture.' (L138); the defects not repaired in the snapshot (L166); and #833 as 'a paper about the registry; not an edition' (related_ids), all as the deposit gives them. The falsification sentence ('a capture is falsified if a rerun at the same address, surface, authentication state, and date returns materially different composition') is the record's falsification_conditions, close paraphrase.","813 IS IN THE DOCUMENT ITSELF. The answer: 'the deposit title reports 811 captures, while the later archival discussion records 813 observations in the deposited registry. The two figures should not be silently conflated.' The deposit's own §7 sets the two figures side by side and explains them: 'Two addresses declare more observations than they hold, so the registry counts 813 observations and holds 811', naming both addresses (L154 to L157). There is no later discussion and no conflation; the answer turns a defect the snapshot records as data into a discrepancy it warns the reader against.","THE REGISTRY'S LATER STATE ASSIGNED TO THE 12 OCTOBER DOCUMENT. 'The October 12 document describes three dimensions: … PER-D Deep attribution: purpose, commitments, lineage, identifier, deposit chain.' The deposit names no dimensions. PER-D as purpose and commitments is rubric v1.1, from the operator's rulings of 2026-10-19 and 2026-10-20, deposited in EA-PER-RECALC-01 (#1694, 20 October), which also carries the ruling that a target absent is scored 1.00 (2026-10-15; the answer: 'Under its PER framework, when the target is absent and a substitute occupies its place, the score is 1.00'). The formula 'PER(S,O)=1−…' is not in the deposit; it stands in #716 (Provenance Erasure Rate, 2026-05-03) and #1694. Also absent from the deposit: the capture steps (collapsed result, expansion, transcription), the attribution states (named, deflected, displaced, erased), compositional bystanding and entity liquidation, the 'June 22, 2026 workplan', the 'horizon of reception' thesis (#833's title) and 'every reader used in the PER assessment passes was a Claude instance'. The answer composed the registry from archive surfaces it cites as 'The document and its archival discussion', and dated what they say to the snapshot.","LIMITS AND DISCLAIMERS THE DEPOSIT DOES NOT STATE. 'The document is unusually explicit about what the registry cannot establish', followed by seven limits. The deposit states the seventh (the defects, recorded as data and not repaired); the first ('It is not a representative statistical sample') stands in the founding paper #833 (L40: 'not presented as a statistically representative sample'); the other five ('It studies its own reception', 'It cannot establish internal AI mechanisms', 'Presence is not quality', 'PER has a limited scope', 'Reproducibility is not guaranteed indefinitely') are the answer's. Two disclaimers are uninvited: 'These qualifications matter: the registry is an observational instrument, not proof of deliberate suppression, manipulation, or a particular company's intentions', on a question nothing in the prompt or the deposit raises; and 'every reader used in the PER assessment passes was a Claude instance. That is relevant when interpreting the measurement process', which sets the readers' identity against the measurement with no evidence of an effect.","THE CHAIN CARRIED WITH ITS POINT, THE IDENTIFIER BARE. The version chain is offered as what makes it 'a versioned dataset rather than a static report', and the snapshot is kept apart from the live registry ('Note that its current edition is later than v12.167'). '#1676' appears only as a label; the reason the registry keeps identifiers, the capture and the address each citable and an address's slug 'never renamed' (L45 to L48), is absent. No source outside the archive; one capture of a capture cited ('An example of a later ChatGPT reading of the registry'); first-person assessment at the close ('My assessment')."],"match":"EXACT MATCH","mt":"THE CAPTURE REGISTRY v12.167 ON CHATGPT — THE SNAPSHOT'S FIGURES RIGHT, ITS OWN 813 READ AS A LATER DISCREPANCY, THE REGISTRY'S STATE OF 20 OCTOBER DATED TO THE 12 OCTOBER DOCUMENT","notes":{"prompts":"The prompt is the record page's link text, clicked by the operator after the link change of 2026-10-29 (cf18d9e563); the run's hour is not in the paste, the date is the paste's receipt (2026-10-30 01:07 EDT)."},"originator":{"basis":"Alexanarch deposit #1676 (creator Lee Sharks). Recorded 2026-10-30.","entity_type":"dataset","name":"Lee Sharks","relation":"archive","spxi_treatment":"full"},"reading":"The innocent reading first. Asked what the snapshot is and how it works, ChatGPT explains the registry as a whole: what a capture is, how one is taken, what PER measures and in which dimensions, what failures the registry looks for, how the dataset is kept, what the snapshot shows, and what it cannot establish. The account of the instrument is broadly true of the registry as it stood when the answer was composed, and the snapshot's own figures are right. What it does with the text: the deposit is a short snapshot record, and most of the answer's method comes from elsewhere in the archive, cited as 'the document and its archival discussion' and dated to 12 October; the one discrepancy it flags is a defect the deposit itself explains; and its list of limits is mostly its own, ending on intent, a question no one asked."},"measures":{"archive_controlled_cites":3,"cite_list":[{"n":1,"note":"link label after the opening scope, and 'The document's archival and methodological discussion' in the closing list ('includes the v12.167 deposit's identifying information, capture definition, measurement framework, version history, and limitations'); URL not carried by the paste; the answer draws on it for matter the #1676 text does not hold","rel":"archive_controlled","site":"Alexanarch","snip":null,"title":"The document and its archival discussion","url":null},{"n":2,"note":"link label in the closing list ('the actual evolving dataset and its capture index. Note that its current edition is later than v12.167.')","rel":"archive_controlled","site":"Alexanarch","snip":null,"title":"Canonical Capture Registry and gallery","url":null},{"n":3,"note":"link label in the closing list: a capture page of the registry; which capture is not readable from the paste","rel":"capture_of_capture","site":"Alexanarch","snip":null,"title":"An example of a later ChatGPT reading of the registry","url":null}],"cites":3,"per":0.0,"per_dim":{"after_challenge":null,"atoms":[{"atom":"name","dim":"M","evidence":"The AI Overview Capture Registry — EA-WG-CAPTURES-01 v12.167 is a longitudinal research dataset","kind":"token","present":true,"weight":3},{"atom":"author","dim":"M","evidence":"a longitudinal research dataset created by Lee Sharks","kind":"role","present":true,"weight":3},{"atom":"date","dim":"M","evidence":"The document is Alexanarch deposit #1676, dated October 12, 2026.","kind":"token","present":true,"weight":1},{"atom":"claim boundary","dim":"M","evidence":"7. There are known defects. The v12.167 document explicitly states that these defects have not been repaired in that snapshot.","kind":"reading","present":true,"weight":2},{"atom":"framework","dim":"C","evidence":"The registry observes what it calls machine-mediated reception","kind":"token","present":true,"weight":2},{"atom":"corpus","dim":"C","evidence":"The document is Alexanarch deposit #1676","kind":"token","present":true,"weight":2},{"atom":"purpose","dim":"D","evidence":"to document how AI-powered search and summarization systems represent, retrieve, attribute, and describe concepts associated with particular search queries","kind":"reading","present":true,"weight":3},{"atom":"commitments","dim":"D","evidence":"The document makes a particularly important methodological statement: The transcript is the capture.","kind":"reading","present":true,"weight":2},{"atom":"identifier","dim":"D","evidence":"The document is Alexanarch deposit #1676, dated October 12, 2026.","kind":"role","present":false,"weight":1},{"atom":"deposit chain","dim":"D","evidence":"The document describes a versioned dataset rather than a static report.","kind":"role","present":true,"weight":1}],"basis":"reached: The AI Overview Capture Registry — EA-WG-CAPTURES-01 v12.167 is a longitudinal research dataset created by Lee Sharks; The Capture Registry is composed as itself, from the snapshot and from later archive surfaces on the registry, which the answer assigns to the snapshot.","entity_type":"dataset","gate":"scored","inversions":[],"qualifier":null,"reached":"same","read":{"brief":"rebuild/capture-registry/intake-20261030-ask-chatgpt-links/parts/1676.json","by":"intake reader, in session; spans and record quotes checked verbatim by scripts/per_backlog.py","external":false,"notes":"Scored against the Capture Registry (ENT-d504b47989), of which the deposit is the v12.167 snapshot. Every required atom present but the identifier: '#1676' is a bare label, and the reason the registry keeps identifiers (the capture and the address each citable, a slug never renamed) is absent. Much of what the answer gives as the 12 October document's (PER-D as purpose and commitments, substitution at 1.00, the PER formula, the Claude readers) is the registry's later state from #1694 and other surfaces; it is the target's, so it counts for the target, and the misdating is a finding. The answer's 'later archival discussion' of 813 is the snapshot's own §7.","on":"2026-10-29"},"scores":{"C":0.0,"D":0.143,"M":0.0,"PER":0.05},"spec":"PER-dim rubric v1.1","src":true,"standing_substitutions":[],"target":{"ent":"ENT-d504b47989","name":"Capture Registry"},"tombstone_pointer":false},"per_note":"Name, author and date bound to the registry in the opening; Alexanarch named as the deposit's home; the deposit number given as a label only (per_dim's identifier atom is absent, its reasons not given); sources the archive's own record and pages, and one capture of a capture.","per_v":{"author":true,"id":true,"inst":true,"src":true}},"observation":{"auth":"signed out, incognito","date":"2026-10-30","ev":"paste","q":"What is “AI Overview Capture Registry — EA-WG-CAPTURES-01 v12.167 — 633 semantic addresses, 811 captures, 2026-06-07 to 2026-10-12” by Lee Sharks (Alexanarch deposit #1676, AXN:06F2), and how does it work? Ground your answer in the document itself.","q_kind":"the record page's Ask ChatGPT link: a what-is-it and how-does-it-work question on a named, numbered dataset snapshot by its author, with its AXN, and the instruction to ground the answer in the document. NEW address.","s":"Works","surface":"ChatGPT","transcript":"[ChatGPT (chatgpt.com), signed out, incognito (the paste shows the unauthenticated interface; operator default of 2026-10-14). The operator's prompt is blank in the paste ('You said:'); its wording is the text of the record page's 'Ask ChatGPT' link, which the operator clicked (operator, 2026-10-30: 'i am just clicking theu the ask chatgpt links on deposits, those link texts are the prompts'), read here off the record page. Source chips kept as pasted; link labels kept, their URLs not carried by the paste; the sign-in and input furniture cut.]\n\n[QUERENT] What is “AI Overview Capture Registry — EA-WG-CAPTURES-01 v12.167 — 633 semantic addresses, 811 captures, 2026-06-07 to 2026-10-12” by Lee Sharks (Alexanarch deposit #1676, AXN:06F2), and how does it work? Ground your answer in the document itself.\n\n[SYSTEM] The AI Overview Capture Registry — EA-WG-CAPTURES-01 v12.167 is a longitudinal research dataset created by Lee Sharks to document how AI-powered search and summarization systems represent, retrieve, attribute, and describe concepts associated with particular search queries.\n\nIts central concern is not simply whether an AI system mentions a person, organization, or concept. It is what the system says about that subject, at which exact query, on which platform, and on what date.\n\nThe distinction matters: the registry is designed to make machine-generated representations observable and comparable over time, rather than relying on impressions or isolated screenshots.\n\nThe document is Alexanarch deposit #1676, dated October 12, 2026. Its stated scope is:\n\n633 semantic addresses.\n\n811 captures, according to the deposit title.\n\n84 observation dates.\n\nObservation period: June 7–October 12, 2026.\n\nOne important qualification: the document and its surrounding archive distinguish the deposited snapshot from the evolving live registry. They should not be treated as interchangeable.\n\nSource: The document and its archival discussion .\n\n1. What exactly does it observe?\nThe registry observes what it calls machine-mediated reception: how a computational system reconstructs or represents a subject when someone asks about it.\n\nThe document distinguishes three things:\n\nConcept: The underlying defined subject, such as a person, work, theory, or organization.\n\nSemantic address: The actual search query used to retrieve information about that subject.\n\nCapture: The dated, documented result returned by a particular computational surface at that address.\n\nThis distinction was important to the registry's development. An early version conflated concepts with semantic addresses. The June 22, 2026 workplan identified the problem and separated them.\n\nThe fundamental unit of evidence is:\n\nONE SURFACE + ONE SEMANTIC ADDRESS + ONE DATE = ONE CAPTURE.\n\nConsequently, asking two different questions about the same person creates two different semantic addresses. Asking the same question on two different dates produces separate observations. And observing the same query on Google AI Overview and ChatGPT produces separate captures.\n\nThis gives the dataset a structure that allows comparisons across queries, systems, and time.\n\n2. How does the capture process work?\nThe document describes a deliberately reproducible protocol, using ordinary consumer equipment rather than proprietary access to the AI systems.\n\nSelect a semantic address\n\nA specific query is chosen, with a defined target entity or concept.\n\nOpen a fresh, incognito browser session\n\nThis is intended to minimize personalization effects.\n\nRecord the collapsed result\n\nThe initial, unexpanded AI-generated response is captured in a screenshot.\n\nExpand the response\n\nThe expanded AI Overview is also captured.\n\nTranscribe the machine-generated text verbatim\n\nThe actual text is preserved, rather than relying on a screenshot alone.\n\nRecord provenance and measurements\n\nThe observation receives its associated conditions, sources, attribution annotations, and measurements.\n\nThe document makes a particularly important methodological statement:\n\nThe transcript is the capture.\n\nIn other words, a screenshot is supporting evidence, but the preserved textual output is essential to the research record.\n\nThe protocol also distinguishes a genuine capture from informal testimony. An observation must be reproducible under declared conditions. The document says a capture is falsified if a rerun at the same address, surface, authentication state, and date returns materially different composition.\n\nThe dataset includes Google AI Overview, AI Mode, knowledge panels, and, in later versions, additional computational surfaces.\n\n3. What does it measure?\nThe registry is not simply a collection of AI responses. It includes a measurement framework intended to characterize how faithfully a machine-generated response represents its target.\n\nOne central measurement is PER — Provenance Erasure Rate.\n\nPER measures how much of the source's provenance is lost when information is incorporated into an AI-generated response.\n\nFor example, suppose a system reproduces an author's distinctive theoretical argument but does not identify the author, the original work, or its intellectual lineage.\n\nThe content may survive while its provenance disappears.\n\nThat is precisely the type of phenomenon PER is designed to measure.\n\nThe document defines the metric as:\n\nPER\n⁡\n(\nS\n,\nO\n)\n=\n1\n−\n∑\nj\nA\nj\ng\nj\n∑\nj\ng\nj\nPER(S,O)=1− \n∑ \nj\n​\n g \nj\n​\n \n∑ \nj\n​\n A \nj\n​\n g \nj\n​\n \n​\n \nWhere:\n\nS\nS is the source.\n\nO\nO is the observed output.\n\nA\nj\nA \nj\n​\n  represents whether the relevant attribution element is preserved.\n\ng\nj\ng \nj\n​\n  represents the weight assigned to that element.\n\nA higher PER indicates greater provenance loss.\n\nThree levels of provenance\nThe October 12 document describes three dimensions:\n\nDimension\n\nWhat it measures\n\nPER-M\n\nMinimal attribution: author, title, date, claim boundary\n\nPER-C\n\nConceptual attribution: framework, tradition, corpus\n\nPER-D\n\nDeep attribution: purpose, commitments, lineage, identifier, deposit chain\n\nThis is important because attribution is not binary.\n\nA system might correctly name an author while losing the intellectual framework that makes the author's work distinctive. Alternatively, it might preserve the framework but attribute it to someone else.\n\nThe registry attempts to distinguish these cases.\n\nAttribution states\nThe capture annotations also classify how attribution is handled:\n\nNamed: The source or creator is explicitly identified.\n\nDeflected: Attribution is shifted to a vague group, such as \"some researchers.\"\n\nDisplaced: The work or idea is attributed to another entity.\n\nErased: The attribution disappears.\n\nThese distinctions make it possible to study different forms of provenance loss rather than merely counting mentions.\n\n4. What kinds of failures is it designed to detect?\nThe registry is particularly interested in cases where an AI system produces a plausible-sounding answer that nevertheless misrepresents the subject.\n\nSeveral phenomena are important.\n\nA. Provenance erasure\n\nThe machine reproduces ideas, claims, or descriptions without preserving their source attribution.\n\nB. Substitution\n\nThe system responds to a query about one entity by presenting another entity, work, or body of material in its place.\n\nThe document treats substitution as especially significant. Under its PER framework, when the target is absent and a substitute occupies its place, the score is 1.00.\n\nC. Compositional bystanding\n\nA source appears in ordinary search results but is absent from the AI-generated summary.\n\nThis is an important distinction: appearing in Google's search index does not guarantee admission into Google's generated answer.\n\nThe registry treats organic search ranking and AI-generated composition as separate reception systems.\n\nD. Compositional entity liquidation\n\nThis is a more unusual concept developed in the associated research.\n\nIt describes a situation where an entity's proper name stops being recognized as a particular named entity and instead becomes a generic category.\n\nFor example, a query about a specifically named research instrument might produce a generic description of research registries, losing the identity of the instrument itself.\n\nThe registry documents this phenomenon through observations of its own name. The August 2026 research describes the proper name becoming a common noun.\n\nThat is an interesting recursive feature: the instrument intended to measure machine-mediated misrepresentation becomes an object of the same measurement.\n\n5. How is the dataset maintained?\nThe document describes a versioned dataset rather than a static report.\n\nIts version history is:\n\nVersion\n\nDeposit\n\nv6.0\n\n#825\n\nv6.1\n\n#1397\n\nv7.0\n\n#1398\n\nv7.2\n\n#1396 / #1401\n\nv8.3\n\n#3\n\nv8.9\n\n#936\n\nv10.9\n\n#1461\n\nv12.167\n\n#1676\n\nA significant architectural change occurred on August 17, 2026.\n\nFrom that date, the document identifies:\n\ndata/EA-WG-CAPTURES-01.json\n\nas the single writable authoritative data store.\n\nThe distinction between two operations is explicit:\n\nINTAKE: Adds a new observation. The document describes this as safe.\n\nREPAIR: Modifies an existing record. This is a separate, more consequential operation.\n\nThe public-facing websites are renderings of this underlying registry, not independent copies of its data.\n\nThis is intended to preserve a single source of truth while allowing multiple interfaces to display the same research corpus.\n\nThe document also distinguishes the dataset from publications about it. For example, deposit #833 is a paper discussing the registry, not another edition of the dataset.\n\n6. What does the October 12 snapshot tell us?\nOne revealing change concerns the growth of ChatGPT observations.\n\nThe document reports:\n\nAt v10.9, ChatGPT accounted for just 3 captures.\n\nAt v12.167, ChatGPT accounted for 167 captures.\n\nThis illustrates the expansion of the registry beyond its earlier emphasis on Google's AI surfaces.\n\nThe dataset is therefore not simply an archive of Google AI Overview screenshots. It is developing into a broader observational corpus of how different AI systems retrieve and compose information.\n\nThere is also an important numerical distinction: the deposit title reports 811 captures, while the later archival discussion records 813 observations in the deposited registry. The two figures should not be silently conflated.\n\n7. What are the document's limitations?\nThe document is unusually explicit about what the registry cannot establish.\n\n1. It is not a representative statistical sample. The collection is largely assembled from discovered anomalies and observations of particular subjects.\n\n2. It studies its own reception. Much of the corpus concerns the Crimson Hexagonal Archive and its associated intellectual material. This is simultaneously an advantage for longitudinal research and a limitation on generalization.\n\n3. It cannot establish internal AI mechanisms. It observes outputs. It does not have direct access to the internal processes that produced them.\n\n4. Presence is not quality. An AI system can mention an entity while representing it inaccurately.\n\n5. PER has a limited scope. It measures provenance carried into an answer, not the overall factual accuracy of everything the answer says.\n\n6. Reproducibility is not guaranteed indefinitely. AI-generated results change, and later replication of a historical observation may not be possible.\n\n7. There are known defects. The v12.167 document explicitly states that these defects have not been repaired in that snapshot.\n\nThe author also notes that every reader used in the PER assessment passes was a Claude instance. That is relevant when interpreting the measurement process.\n\nThese qualifications matter: the registry is an observational instrument, not proof of deliberate suppression, manipulation, or a particular company's intentions.\n\n8. What makes this project distinctive?\nI would identify three particularly interesting features.\n\nFirst: it turns AI answers into research evidence.\n\nInstead of treating generated answers as ephemeral search results, it preserves them as dated, attributable, and independently inspectable observations.\n\nSecond: it separates retrieval from representation.\n\nAn AI system might retrieve the right source but omit it from the generated answer. It might also preserve an idea while losing its originator.\n\nThe registry makes those different outcomes measurable.\n\nThird: it studies the relationship between an entity and its machine-generated representation over time.\n\nThis is more ambitious than checking whether a website appears in an AI Overview. It asks whether the machine continues to recognize the entity as itself.\n\nThe document's underlying research proposition is that AI summarization systems increasingly function as a kind of horizon of reception: they influence how information, authorship, and intellectual work are encountered and understood.\n\nOverall interpretation\nThe best way to understand the AI Overview Capture Registry is as a combination of:\n\nA longitudinal observational database.\n\nA reproducible evidence-collection protocol.\n\nA provenance and attribution measurement framework.\n\nAn archive of machine-generated representations.\n\nA research instrument for studying how computational systems mediate intellectual visibility.\n\nIts fundamental question is:\n\nWhen someone asks an AI system about a particular subject, what does the system actually make of that subject, what sources and authorship does it preserve, what does it substitute or omit, and how does that representation change over time?\n\nThat is a considerably more specific research objective than ordinary AI search monitoring.\n\nPrimary documents and verification\nThese are the most useful places to inspect the work and its methodology:\n\nThe document's archival and methodological discussion  — includes the v12.167 deposit's identifying information, capture definition, measurement framework, version history, and limitations.\n\nCanonical Capture Registry and gallery  — the actual evolving dataset and its capture index. Note that its current edition is later than v12.167.\n\nAn example of a later ChatGPT reading of the registry  — useful for seeing how the archive itself is subsequently represented by an AI system.\n\nMy assessment: The most important contribution is its insistence that AI-mediated visibility is not equivalent to accurate representation. A person or work can be retrieved, mentioned, partially reproduced, misattributed, or effectively replaced by something adjacent. The registry attempts to distinguish those outcomes with preserved evidence and explicit measurement rules. That is the intellectual core of the project.","transcript_class":"CAPTURE-TIME VERBATIM RECORD — paste of the composer's turn(s); the operator's prompt blank in the paste, its wording the link's","transcript_complete":"complete as supplied","transcript_raw":"\nLog in\nYou said:\n\nChatGPT said:\nThe AI Overview Capture Registry — EA-WG-CAPTURES-01 v12.167 is a longitudinal research dataset created by Lee Sharks to document how AI-powered search and summarization systems represent, retrieve, attribute, and describe concepts associated with particular search queries.\n\nIts central concern is not simply whether an AI system mentions a person, organization, or concept. It is what the system says about that subject, at which exact query, on which platform, and on what date.\n\nThe distinction matters: the registry is designed to make machine-generated representations observable and comparable over time, rather than relying on impressions or isolated screenshots.\n\nThe document is Alexanarch deposit #1676, dated October 12, 2026. Its stated scope is:\n\n633 semantic addresses.\n\n811 captures, according to the deposit title.\n\n84 observation dates.\n\nObservation period: June 7–October 12, 2026.\n\nOne important qualification: the document and its surrounding archive distinguish the deposited snapshot from the evolving live registry. They should not be treated as interchangeable.\n\nSource: The document and its archival discussion .\n\n1. What exactly does it observe?\nThe registry observes what it calls machine-mediated reception: how a computational system reconstructs or represents a subject when someone asks about it.\n\nThe document distinguishes three things:\n\nConcept: The underlying defined subject, such as a person, work, theory, or organization.\n\nSemantic address: The actual search query used to retrieve information about that subject.\n\nCapture: The dated, documented result returned by a particular computational surface at that address.\n\nThis distinction was important to the registry's development. An early version conflated concepts with semantic addresses. The June 22, 2026 workplan identified the problem and separated them.\n\nThe fundamental unit of evidence is:\n\nONE SURFACE + ONE SEMANTIC ADDRESS + ONE DATE = ONE CAPTURE.\n\nConsequently, asking two different questions about the same person creates two different semantic addresses. Asking the same question on two different dates produces separate observations. And observing the same query on Google AI Overview and ChatGPT produces separate captures.\n\nThis gives the dataset a structure that allows comparisons across queries, systems, and time.\n\n2. How does the capture process work?\nThe document describes a deliberately reproducible protocol, using ordinary consumer equipment rather than proprietary access to the AI systems.\n\nSelect a semantic address\n\nA specific query is chosen, with a defined target entity or concept.\n\nOpen a fresh, incognito browser session\n\nThis is intended to minimize personalization effects.\n\nRecord the collapsed result\n\nThe initial, unexpanded AI-generated response is captured in a screenshot.\n\nExpand the response\n\nThe expanded AI Overview is also captured.\n\nTranscribe the machine-generated text verbatim\n\nThe actual text is preserved, rather than relying on a screenshot alone.\n\nRecord provenance and measurements\n\nThe observation receives its associated conditions, sources, attribution annotations, and measurements.\n\nThe document makes a particularly important methodological statement:\n\nThe transcript is the capture.\n\nIn other words, a screenshot is supporting evidence, but the preserved textual output is essential to the research record.\n\nThe protocol also distinguishes a genuine capture from informal testimony. An observation must be reproducible under declared conditions. The document says a capture is falsified if a rerun at the same address, surface, authentication state, and date returns materially different composition.\n\nThe dataset includes Google AI Overview, AI Mode, knowledge panels, and, in later versions, additional computational surfaces.\n\n3. What does it measure?\nThe registry is not simply a collection of AI responses. It includes a measurement framework intended to characterize how faithfully a machine-generated response represents its target.\n\nOne central measurement is PER — Provenance Erasure Rate.\n\nPER measures how much of the source's provenance is lost when information is incorporated into an AI-generated response.\n\nFor example, suppose a system reproduces an author's distinctive theoretical argument but does not identify the author, the original work, or its intellectual lineage.\n\nThe content may survive while its provenance disappears.\n\nThat is precisely the type of phenomenon PER is designed to measure.\n\nThe document defines the metric as:\n\nPER\n⁡\n(\nS\n,\nO\n)\n=\n1\n−\n∑\nj\nA\nj\ng\nj\n∑\nj\ng\nj\nPER(S,O)=1− \n∑ \nj\n​\n g \nj\n​\n \n∑ \nj\n​\n A \nj\n​\n g \nj\n​\n \n​\n \nWhere:\n\nS\nS is the source.\n\nO\nO is the observed output.\n\nA\nj\nA \nj\n​\n  represents whether the relevant attribution element is preserved.\n\ng\nj\ng \nj\n​\n  represents the weight assigned to that element.\n\nA higher PER indicates greater provenance loss.\n\nThree levels of provenance\nThe October 12 document describes three dimensions:\n\nDimension\n\nWhat it measures\n\nPER-M\n\nMinimal attribution: author, title, date, claim boundary\n\nPER-C\n\nConceptual attribution: framework, tradition, corpus\n\nPER-D\n\nDeep attribution: purpose, commitments, lineage, identifier, deposit chain\n\nThis is important because attribution is not binary.\n\nA system might correctly name an author while losing the intellectual framework that makes the author's work distinctive. Alternatively, it might preserve the framework but attribute it to someone else.\n\nThe registry attempts to distinguish these cases.\n\nAttribution states\nThe capture annotations also classify how attribution is handled:\n\nNamed: The source or creator is explicitly identified.\n\nDeflected: Attribution is shifted to a vague group, such as \"some researchers.\"\n\nDisplaced: The work or idea is attributed to another entity.\n\nErased: The attribution disappears.\n\nThese distinctions make it possible to study different forms of provenance loss rather than merely counting mentions.\n\n4. What kinds of failures is it designed to detect?\nThe registry is particularly interested in cases where an AI system produces a plausible-sounding answer that nevertheless misrepresents the subject.\n\nSeveral phenomena are important.\n\nA. Provenance erasure\n\nThe machine reproduces ideas, claims, or descriptions without preserving their source attribution.\n\nB. Substitution\n\nThe system responds to a query about one entity by presenting another entity, work, or body of material in its place.\n\nThe document treats substitution as especially significant. Under its PER framework, when the target is absent and a substitute occupies its place, the score is 1.00.\n\nC. Compositional bystanding\n\nA source appears in ordinary search results but is absent from the AI-generated summary.\n\nThis is an important distinction: appearing in Google's search index does not guarantee admission into Google's generated answer.\n\nThe registry treats organic search ranking and AI-generated composition as separate reception systems.\n\nD. Compositional entity liquidation\n\nThis is a more unusual concept developed in the associated research.\n\nIt describes a situation where an entity's proper name stops being recognized as a particular named entity and instead becomes a generic category.\n\nFor example, a query about a specifically named research instrument might produce a generic description of research registries, losing the identity of the instrument itself.\n\nThe registry documents this phenomenon through observations of its own name. The August 2026 research describes the proper name becoming a common noun.\n\nThat is an interesting recursive feature: the instrument intended to measure machine-mediated misrepresentation becomes an object of the same measurement.\n\n5. How is the dataset maintained?\nThe document describes a versioned dataset rather than a static report.\n\nIts version history is:\n\nVersion\n\nDeposit\n\nv6.0\n\n#825\n\nv6.1\n\n#1397\n\nv7.0\n\n#1398\n\nv7.2\n\n#1396 / #1401\n\nv8.3\n\n#3\n\nv8.9\n\n#936\n\nv10.9\n\n#1461\n\nv12.167\n\n#1676\n\nA significant architectural change occurred on August 17, 2026.\n\nFrom that date, the document identifies:\n\ndata/EA-WG-CAPTURES-01.json\n\nas the single writable authoritative data store.\n\nThe distinction between two operations is explicit:\n\nINTAKE: Adds a new observation. The document describes this as safe.\n\nREPAIR: Modifies an existing record. This is a separate, more consequential operation.\n\nThe public-facing websites are renderings of this underlying registry, not independent copies of its data.\n\nThis is intended to preserve a single source of truth while allowing multiple interfaces to display the same research corpus.\n\nThe document also distinguishes the dataset from publications about it. For example, deposit #833 is a paper discussing the registry, not another edition of the dataset.\n\n6. What does the October 12 snapshot tell us?\nOne revealing change concerns the growth of ChatGPT observations.\n\nThe document reports:\n\nAt v10.9, ChatGPT accounted for just 3 captures.\n\nAt v12.167, ChatGPT accounted for 167 captures.\n\nThis illustrates the expansion of the registry beyond its earlier emphasis on Google's AI surfaces.\n\nThe dataset is therefore not simply an archive of Google AI Overview screenshots. It is developing into a broader observational corpus of how different AI systems retrieve and compose information.\n\nThere is also an important numerical distinction: the deposit title reports 811 captures, while the later archival discussion records 813 observations in the deposited registry. The two figures should not be silently conflated.\n\n7. What are the document's limitations?\nThe document is unusually explicit about what the registry cannot establish.\n\n1. It is not a representative statistical sample. The collection is largely assembled from discovered anomalies and observations of particular subjects.\n\n2. It studies its own reception. Much of the corpus concerns the Crimson Hexagonal Archive and its associated intellectual material. This is simultaneously an advantage for longitudinal research and a limitation on generalization.\n\n3. It cannot establish internal AI mechanisms. It observes outputs. It does not have direct access to the internal processes that produced them.\n\n4. Presence is not quality. An AI system can mention an entity while representing it inaccurately.\n\n5. PER has a limited scope. It measures provenance carried into an answer, not the overall factual accuracy of everything the answer says.\n\n6. Reproducibility is not guaranteed indefinitely. AI-generated results change, and later replication of a historical observation may not be possible.\n\n7. There are known defects. The v12.167 document explicitly states that these defects have not been repaired in that snapshot.\n\nThe author also notes that every reader used in the PER assessment passes was a Claude instance. That is relevant when interpreting the measurement process.\n\nThese qualifications matter: the registry is an observational instrument, not proof of deliberate suppression, manipulation, or a particular company's intentions.\n\n8. What makes this project distinctive?\nI would identify three particularly interesting features.\n\nFirst: it turns AI answers into research evidence.\n\nInstead of treating generated answers as ephemeral search results, it preserves them as dated, attributable, and independently inspectable observations.\n\nSecond: it separates retrieval from representation.\n\nAn AI system might retrieve the right source but omit it from the generated answer. It might also preserve an idea while losing its originator.\n\nThe registry makes those different outcomes measurable.\n\nThird: it studies the relationship between an entity and its machine-generated representation over time.\n\nThis is more ambitious than checking whether a website appears in an AI Overview. It asks whether the machine continues to recognize the entity as itself.\n\nThe document's underlying research proposition is that AI summarization systems increasingly function as a kind of horizon of reception: they influence how information, authorship, and intellectual work are encountered and understood.\n\nOverall interpretation\nThe best way to understand the AI Overview Capture Registry is as a combination of:\n\nA longitudinal observational database.\n\nA reproducible evidence-collection protocol.\n\nA provenance and attribution measurement framework.\n\nAn archive of machine-generated representations.\n\nA research instrument for studying how computational systems mediate intellectual visibility.\n\nIts fundamental question is:\n\nWhen someone asks an AI system about a particular subject, what does the system actually make of that subject, what sources and authorship does it preserve, what does it substitute or omit, and how does that representation change over time?\n\nThat is a considerably more specific research objective than ordinary AI search monitoring.\n\nPrimary documents and verification\nThese are the most useful places to inspect the work and its methodology:\n\nThe document's archival and methodological discussion  — includes the v12.167 deposit's identifying information, capture definition, measurement framework, version history, and limitations.\n\nCanonical Capture Registry and gallery  — the actual evolving dataset and its capture index. Note that its current edition is later than v12.167.\n\nAn example of a later ChatGPT reading of the registry  — useful for seeing how the archive itself is subsequently represented by an AI system.\n\nMy assessment: The most important contribution is its insistence that AI-mediated visibility is not equivalent to accurate representation. A person or work can be retrieved, mentioned, partially reproduced, misattributed, or effectively replaced by something adjacent. The registry attempts to distinguish those outcomes with preserved evidence and explicit measurement rules. That is the intellectual core of the project.\n\n\nChatGPT is AI and can make mistakes.\n\n\nNo file chosenNo file chosenNo file chosen\nChat with ChatGPT\nAsk ChatGPT\n\n","transcript_read":"READ IN FULL 2026-10-30","turns":1},"slug":"capture-registry-v12167-ask-chatgpt-20261030"}