[ChatGPT (chatgpt.com), signed out, incognito (operator: 'all opened in incognito'; the paste shows the unauthenticated interface: 'Log in', 'Chat with ChatGPT', 'Ask ChatGPT'). Received 2026-11-07 01:57 EST. The operator's prompt is not in the paste; its wording is the text of the record page's 'Ask ChatGPT' link for #1739 (operator: 'see the link for prompt text'), read here off the record page. Source chips ('Alexanarch', '+1'), image-strip captions and equations broken one glyph to a line kept as pasted; the closing interface furniture cut.] [QUERENT] What is “AI Overview Capture Registry — EA-WG-CAPTURES-01 v12.559 — 1001 semantic addresses, 1193 captures, 2026-06-07 to 2026-11-04” by Orin Trace (Alexanarch deposit #1739, AXN:0C01), and how does it work? Ground your answer in the document itself. [SYSTEM] What it is The AI Overview Capture Registry — EA-WG-CAPTURES-01 v12.559 is a versioned empirical research dataset that records how AI-powered search and conversational systems represent named entities when prompted with particular queries. It is not simply a database of AI search results, nor is it a conventional tracker of whether an entity appears in Google's AI Overviews. Its purpose is to preserve and compare what different AI systems say about the same subject, at different times, under identifiable query and authentication conditions. The document describes its central purpose as: A longitudinal instrument for recording dated machine-mediated reception events: what composition layers make of a named entity when asked for it at a specific semantic address, on a specific date, in a specific authentication state. The work is attributed to Dr. Orin Trace, dated November 4, 2026, and deposited as Alexanarch deposit #1739, identifier AXN:0C01. It is a snapshot of registry version 12.559, superseding version 12.167 (deposit #1676). The document reports: 1,001 Semantic addresses Distinct registered subjects or query positions 1,193 Captures Recorded observations of AI-generated responses 104 Dates Distinct observation dates 5 months Observation period June 7 – November 4, 2026 In other words, it attempts to make AI-generated representations of people, works, institutions, concepts, and other entities into dated, inspectable, citable evidence. 1. The fundamental idea: a capture The most important sentence in the document is: ONE SURFACE + ONE SEMANTIC ADDRESS + ONE DATE = ONE CAPTURE. This is the registry's basic unit of evidence. The three components mean: Surface: The AI system or interface being observed, such as Google AI Overview, ChatGPT, Grok, or Perplexity. Semantic address: The registered subject or query position being investigated. Date: When the observation was made. For example, imagine investigating how ChatGPT describes a particular scholarly work. A capture would preserve: The identity of the work being investigated. The specific semantic address assigned to it. The platform on which the answer appeared. The date of the observation. The authentication state (signed in, signed out, incognito, etc.). The resulting response, ideally as a verbatim transcript. A screenshot or frame, when available. An accompanying analysis and provenance measurements. This distinction matters because the registry treats an individual AI response as evidence of what a particular system produced under particular conditions—not as definitive evidence about the underlying subject. 2. Semantic addresses: what is being investigated? A semantic address is the registry's persistent identity for a subject being investigated. It is not simply a URL or a search-engine result. It is a stable position in the registry associated with a named entity or subject. The document reports addresses distributed across subject categories: Section Addresses Machine Reception 263 Works 141 Archive 91 Frameworks 86 Heteronyms 83 Semantic Economy 49 Coinages 48 Revelation & Theology 32 Classics & Philology 29 Provenance & Erasure 25 Institutions 24 Sites 19 People 17 Valuation 17 There are also smaller categories, including Architecture, Platform Erosion, Reference & General Knowledge, Projects, Operative Semiotics, and Concepts. Two important rules govern addresses: The address's slug is inherited from its first capture and is never renamed. An address can accumulate captures across multiple platforms and dates. This allows researchers to ask two different questions: What did a particular AI system say about this entity on a particular occasion? How has the representation of this entity changed across systems and over time? The document explicitly distinguishes these citation functions: captures/{slug}/ identifies an individual observation. addresses/{slug}/ identifies the persistent subject and its accumulated history. That distinction is fundamental to the dataset's research design. 3. Which AI systems does it monitor? The 1,193 captures are distributed across multiple surfaces. Surface Captures Google AI Overview 745 ChatGPT (all recorded labels) 278 Google AI Mode 21 Grok 9 Perplexity 9 Bing Copilot 6 Claude 3 Qwen 3 Bing AI Overview 2 Google Knowledge Panel 2 Other single-capture surfaces 9 Undetermined 56 Unresolved 50 The document makes an important distinction between an AI model and the interface through which its output is encountered. For example, ChatGPT's 278 captures are subdivided into different recorded labels: 230 under the main ChatGPT grouping. 27 under Luna 5.6. 12 under Luna. 8 under chatgpt.com. 1 web, logged out. This is useful because the same underlying service can produce different experiences depending on its interface, configuration, and access conditions. The registry therefore does not treat every observation from a particular AI provider as interchangeable. All figures above are transcribed from section 2 of the deposited document. 4. How the registry measures provenance loss One of the most distinctive features is its measurement system called PER: Provenance Erasure Rate. The underlying concern is that AI-generated summaries can retain some information about a subject while losing the information needed to identify, contextualize, or verify it. The original PER system uses four units: Author. Institution. Identifier. Source. These represent different dimensions of provenance. For example, an AI-generated summary might accurately describe a book but: Omit its author. Misidentify the institution responsible for it. Leave out its persistent identifier. Fail to identify the source from which its information originated. The registry attempts to measure these losses rather than merely judging whether an answer sounds plausible. The evolution of PER The document describes a significant methodological change. Originally, PER was represented by a scalar score derived from a four-unit vector. In October 2026, the registry introduced a dimensional version, recorded as per_dim, under rubric v1.1. This expanded the assessment to include: PER-M: The measurement associated with material provenance. PER-C: The contextual dimension. PER-D: The deeper dimension, including purpose and commitments. The newer rubric also incorporates inversion and standing substitution. The exact operational definitions and atom sets are documented in the associated methodological work, EA-PER-RECALC-01 (deposit #1694). The v12.559 document reports: 907 Historical scalar PER Observations with the original measurement 0.429 Mean scalar PER Across those 907 observations 1,171 Dimensional PER Observations carrying the newer assessment 20 Pending assessments With recorded reasons for being pending The historical mean had been 0.506 in version 12.167 and 0.603 in version 10.9. These figures should not automatically be interpreted as evidence that AI systems improved or deteriorated. They are measurements across a changing collection of observations, and the measurement instrument itself evolved. The document's important methodological decision is that the original scalar is now derived automatically from its four-unit vector at the time of seating, rather than manually entered. This reduces the possibility of inconsistent hand-entered scores. 5. How the data actually gets produced The document describes a specific, ordered processing pipeline. Capture Registry workflow Observation and intake The operator collects the AI-generated response, associated material, and observation metadata. Capture intake scripts/capture_intake.py --seat This script admits, routes, normalizes, validates, and inserts the observation. It also derives the scalar PER from its four-unit vector. Gallery generation scripts/build_capture_gallery.py Builds the capture gallery from the registered observations. Postflight processing scripts/seat_capture_postflight.py Links and bakes records, runs gates, builds the semantic-address layer and address pages, allocates AXNs, synchronizes the dataset projection, and checks citability. Successful seating A capture is considered seated when the postflight script exits with status 0. There is an important provenance safeguard before records enter the registry. Each intake batch is preserved in a dated directory: rebuild/capture-registry/intake-YYYYMMDD-*/ These directories preserve: The original pasted material. The operator's messages. The script responsible for authoring each capture. The document additionally states that every quotation on which a finding rests is asserted against its source before anything is written. This is intended to prevent an analytical claim from silently becoming detached from the evidence supporting it. What happens when an existing capture needs correction? The document distinguishes between initial intake and repair. A seated capture is revised using: scripts/revise_capture_entry.py This revises the existing record while retaining its identity fields. The distinction is significant: the registry does not simply replace problematic observations with cleaner versions and conceal their history. It attempts to preserve the identity and provenance of the original observation. 6. AXNs: how individual observations are identified Another important feature is the AXN identification system. Since October 27, 2026, each semantic address holds one position in a shared allocation ledger: AXN:<HEX>.CAPTURE Each individual observation also receives an AXN whose glyphs derive from the SHA-256 hash of the observation sealed together with its analysis and measurements. This provides a connection between: The registered subject. Its individual observations. The sealed evidence and analysis. The associated deposit. The deposit's own identifier is AXN:0C01, while #1739 identifies its position in the Alexanarch deposit sequence. These are not interchangeable identifiers. The document also supplies SHA-256 hashes for its data files and its complete archive bundle. This makes it possible to check whether a downloaded file matches the deposited version. 7. What is actually preserved? The deposit is more than a spreadsheet or a collection of transcripts. It contains a substantial archival structure. File Function EA-WG-CAPTURES-01-v12.559.json Canonical registry data EA-WG-CAPTURES-01.schema.json Formal structure and validation rules semantic-addresses-v12.559.json Semantic-address index capture-deposit-links-v12.559.json Links between captures and deposits captures-axn-v12.559.json Capture AXN index manifest.json File inventory, checksums, sizes and roles .tar.gz bundle Complete packaged archive The complete bundle contains 2,992 members. The manifest lists 492 files, including: 485 capture frames. The principal data files. The packaged archive. Other supporting materials. The 485 frames account for approximately 186.5 MB. This matters because a transcript alone does not always establish what an interface displayed. Screenshots can preserve visual evidence of the interface and its generated composition. The registry therefore attempts to preserve both textual and visual evidence. 8. Known defects and limitations One of the more interesting aspects of this deposit is that it explicitly documents its own defects. It does not claim that all 1,193 captures are complete, perfectly attributed, or fully verifiable. The document reports the following defects: Defect Count Null citations 243 Analysis without finding 151 Truncated by interface 145 Surface unresolved 50 Citations unextracted 47 Unsupported citations 10 Date unresolved 3 Stale state 2 Distinction collapse 1 There are also three specifically discussed limitations: 1. Stale summary statistics The top-level _defect_counts field still contains the figures from v12.167. The document therefore calculates the current defect table from the individual records instead. 2. Missing physical frame files Two referenced screenshots are not held at their expected repository paths: screengrab-20260910-105838.png screengrab-20260910-105901.png 3. Pending dimensional assessments Twenty observations have pending dimensional PER records, each accompanied by a stated reason. The document explicitly says these defects have not been repaired in this snapshot. That is an important archival principle: defects are retained as documented properties of the deposited dataset rather than silently corrected during snapshot creation. 9. Why version 12.559 matters The registry has undergone substantial expansion. Between v12.167 and v12.559: Semantic addresses increased from 633 to 1,001. Captures increased from 811 to 1,193. Dimensional PER became the standard assessment at intake. Capture-specific AXNs were introduced. Surface identification was expanded. The registry's postflight process was strengthened. The new version is therefore not merely a larger dataset. It represents an evolution of the underlying research instrument. 10. What is the broader intellectual significance? I think the most important idea in this document is that it treats AI-generated knowledge as something that can itself become an object of empirical research. Traditionally, researchers study: Books and publications. Authors and institutions. Search-engine rankings. Citations and bibliographies. The transmission and reception of ideas. This registry adds another object of study: The way machine-mediated systems receive, reconstruct, summarize, attribute, and sometimes distort existing knowledge. Consider a hypothetical example. A researcher publishes a book with a clearly identified author, institution, publisher, and identifier. Several months later, three AI systems are asked about that book. System A correctly identifies the author but omits the institution. System B describes the book accurately but attributes it to another researcher. System C gives a plausible description but cites unrelated sources. A conventional search-results archive might simply save the three responses. The Capture Registry instead provides a framework for recording each response as a distinct observation, associating it with a persistent semantic address, preserving its evidence, and assessing its provenance. Over time, the researcher can investigate whether these errors are: Persistent or temporary. Specific to one platform. Shared across multiple platforms. Associated with particular types of entities. Changing as the underlying AI systems change. That is the difference between collecting AI answers and building an empirical instrument for studying AI-mediated reception. 11. How should we interpret the document critically? There is an important distinction between what the deposit establishes and what it claims its methodology can establish. What it establishes internally: A defined dataset with 1,001 semantic addresses and 1,193 observations. A stated collection period and platform distribution. A documented intake and processing pipeline. A formalized citation and identification structure. A version history. File-level checksums. Explicitly recorded defects and pending assessments. What it does not independently establish: That every observation accurately represents the complete experience of an ordinary user. That the selected semantic addresses constitute a representative sample of all AI-generated knowledge. That its PER methodology is an independently validated universal measure of provenance loss. That the observed changes in PER necessarily indicate changes in overall AI quality. That the registry's own interpretation of every captured response is correct. The distinction is especially important because the document identifies a human–machine production process: the snapshot was assembled and written by TACHYON (Claude, Anthropic), under MANUS (Lee Sharks) direction, while the underlying captures were taken by the operator. This is a provenance statement about the production of the dataset itself. It should not be confused with independent verification of all the conclusions derived from the dataset. My overall assessment The AI Overview Capture Registry is best understood as a longitudinal archival and measurement infrastructure for studying machine-mediated knowledge reception. Its central innovation is not simply collecting AI-generated answers. It is combining: Persistent semantic identities for the subjects being investigated. Dated, platform-specific observations. Verbatim transcripts and visual evidence. Provenance-loss measurements. Cryptographic identification and integrity checks. A versioned, reproducible intake and correction process. Explicit documentation of defects. The particularly important conceptual move is making an AI-generated representation into a citable historical event. Instead of saying, "ChatGPT says X about this book," a researcher can potentially identify precisely which ChatGPT observation, concerning which semantic address, on which date, under which access conditions, with which transcript and associated evidence. That makes the registry useful for studying not only AI search visibility, but also attribution, knowledge transmission, entity resolution, historical distortion, and the relationship between source material and machine-generated accounts. In one sentence: The registry attempts to turn the changing ways AI systems represent the world into a structured, measurable, historically traceable body of evidence. Primary document The account above is grounded in the deposited document itself, rather than treating its descriptions as independently verified research findings. Author: Orin Trace Title: AI Overview Capture Registry — EA-WG-CAPTURES-01 v12.559 Date: November 4, 2026 Deposit: #1739 Identifier: AXN:0C01 License: CC BY 4.0 The document is available through the Alexanarch deposit record, which includes the complete Markdown text, JSON dataset, schema, individual indexes, manifest, and downloadable archive. Source: Goodreads