{
 "row": "openai-gdpval",
 "entity": "openai-gdpval",
 "address": "GDPval",
 "addresses": [
  {
   "address": "GDPval",
   "slug": "gdpval-aio-20261029",
   "addr_id": "ADDR-d4b480b48aca",
   "obs_id": "OBS-ef64b7d44199",
   "axn": "AXN:0AED.CAPTURE.🎨🌻🖐️🧪🕖🏛️"
  }
 ],
 "epoch": "2026-10-29",
 "type": "C (public entity on which the archive bears: a task benchmark; an entity of the Labor Vector meta-aggregator)",
 "surface": "Google AI Overview (expanded)",
 "auth": "signed out, incognito",
 "status": {
  "procedure": "EA-NEGONT-02 v0.9, #1690: equal case, ρ = extends, read from the Labor Vector census record, the capture record and the deposits that bear on the benchmark's classes (#771, #855, #199, #778, #1616, #1540, #482, #48, #637, #163, #623, #308, #533, #1545); archive selected by D/R/O (no D or R candidate: no archive text names GDPval; 14 deposits admitted by reading at O-class, 12 at the first pass and 2 by the hop, with 3 grouped as instances; the rest of the hop's 136 read and rejected), with the Labor Vector census record (25 claims) and the Capture Registry's findings, reading, analysis, query note and PER note (10 claims) ledgered as the archive's records of the entity; the field B by three origins (7 cards, 2 of them unreachable; 1 own page; Wikipedia, no article, with one item in another article); the composer read the ledgers, never the transcript, except where the brief asks for T's claims and the delta",
  "ledger": "unaudited (§3.8: one extractor per ledger); field quotes 67/67 verified by script against saved text, the PDF's text, the Wikipedia response summary or the card text; archive quotes 56/56 verified against the registry, census.json, questions.json and the deposit texts",
  "frozen": false,
  "breaches": []
 },
 "spec": "EA-NEGONT-02 v0.9, #1690",
 "objects": {
  "T": {
   "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/T1-gdpval-aio-20261029.txt",
   "sha256": "0f182b74c748e78aca58f5330ebcf9bc3f20c2dd9d903ec36746c7b335342e34",
   "words": 136,
   "cards": 7,
   "words_note": "one transcript at one address (136 words in the answer body with its closer, links reduced to their text, markers and the card rail excluded; 7 cards)",
   "claims": [
    "T1-1 definition: 'an AI benchmark created by OpenAI to measure how well artificial intelligence models perform on real-world, economically valuable knowledge-work tasks rather than academic or math tests' (three markers)",
    "T1-2 scope: 'Covers 1,320 specialized tasks across 44 occupations derived from the top 9 U.S. GDP-contributing sectors.' (two markers)",
    "T1-3 task design: 'Built by industry professionals with an average of 14 years of experience using actual work products like legal briefs, spreadsheets, and engineering blueprints.' (two markers)",
    "T1-4 evaluation: 'Uses blind head-to-head comparisons by human experts to grade model outputs against human work.' (two markers)",
    "T1-5 open source set: 'Includes a gold open-sourced subset of 220 tasks available via Hugging Face Datasets and detailed in the ArXiv Research Paper.' (two markers)",
    "T1-6 leaderboards: 'Evaluated by frameworks like Artificial Analysis GDPval-AA v2.1.' (one marker)",
    "T1-7 closer: 'Would you like to know more about top-performing models on GDPval, or how to access the open-source dataset?'"
   ],
   "register": {
    "register": "data/EA-WG-CAPTURES-01.json",
    "slug": "gdpval-aio-20261029",
    "obs_id": "OBS-ef64b7d44199"
   },
   "transcripts": [
    {
     "slug": "gdpval-aio-20261029",
     "address": "GDPval",
     "addr_id": "ADDR-d4b480b48aca",
     "obs_id": "OBS-ef64b7d44199",
     "axn": "AXN:0AED.CAPTURE.🎨🌻🖐️🧪🕖🏛️",
     "date": "2026-10-29",
     "surface": "Google AI Overview (expanded)",
     "auth": "signed out, incognito",
     "sha256": "0f182b74c748e78aca58f5330ebcf9bc3f20c2dd9d903ec36746c7b335342e34",
     "sha256_of": "the registry's 'transcript' string (utf-8)",
     "sha256_raw": "952ebdb12cc6b8897898087c3e783baf1dc4bd1c779b11a5dccfcc81ddb2fd8e",
     "words": 136,
     "cards": 7,
     "card_rail": [
      "1. OpenAI — Measuring the performance of our models on real-world tasks | OpenAI",
      "2. Artificial Analysis — GDPval-AA v2.1 Benchmark - Artificial Analysis",
      "3. arXiv — [2510.04374] GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks",
      "4. OpenAI — gdpval: evaluating ai model performance - OpenAI",
      "5. Hugging Face — openai/gdpval · Datasets at Hugging Face",
      "6. YouTube·Vivek Haldar — GDPval: An AI benchmark for real-world tasks (video, 11:23)",
      "7. GitHub Pages documentation — GDPval – Inspect Evals - GitHub Pages"
     ],
     "claims": [
      "T1-1 definition: 'an AI benchmark created by OpenAI to measure how well artificial intelligence models perform on real-world, economically valuable knowledge-work tasks rather than academic or math tests' (three markers)",
      "T1-2 scope: 'Covers 1,320 specialized tasks across 44 occupations derived from the top 9 U.S. GDP-contributing sectors.' (two markers)",
      "T1-3 task design: 'Built by industry professionals with an average of 14 years of experience using actual work products like legal briefs, spreadsheets, and engineering blueprints.' (two markers)",
      "T1-4 evaluation: 'Uses blind head-to-head comparisons by human experts to grade model outputs against human work.' (two markers)",
      "T1-5 open source set: 'Includes a gold open-sourced subset of 220 tasks available via Hugging Face Datasets and detailed in the ArXiv Research Paper.' (two markers)",
      "T1-6 leaderboards: 'Evaluated by frameworks like Artificial Analysis GDPval-AA v2.1.' (one marker)",
      "T1-7 closer: 'Would you like to know more about top-performing models on GDPval, or how to access the open-source dataset?'"
     ],
     "markers": "Inline links are google.com/goto tokens, opaque; the paste is the Overview expanded (operator's default of 2026-10-01; the 'AI Mode Conversation' view is the expanded Overview, ruling of 2026-10-18); signed out, incognito (operator's default of 2026-10-14). One turn. The definition carries three markers, each bullet but 'Leaderboards' two, 'Leaderboards' one; the tokens cannot be resolved to cards.",
     "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/T1-gdpval-aio-20261029.txt"
    }
   ]
  },
  "KO": {
   "title": "GDPval",
   "genre": "knowledge object, equal case, under EA-NEGONT-02 v0.9 (#1690): the archive's sources and the field's admitted on equal terms, ontology included, in one claim space",
   "status": "composed 2026-10-29 under v0.9 on the operator's ruling ('as against their wikipedia articles if they exist, their own pages, etc.'); not frozen; ledgers unaudited",
   "sentences": [
    {
     "n": 1,
     "para": 0,
     "sense": "what it is",
     "text": "GDPval is, in the Labor Vector census, a benchmark, \"an evaluation dataset with an experimental automated grader\", whose kind of evidence is a task benchmark at two of the census's intersections, labor performed with AI and labor producing AI, read for who supplies the labor, who directs it and who receives its gains (Labor Vector record); its authors introduce it as \"a benchmark evaluating AI model capabilities on real-world economically valuable tasks\" (arXiv abstract), in its first version (OpenAI's PDF).",
     "claims": [
      "LV-01",
      "LV-04",
      "LV-03",
      "LV-24",
      "F1",
      "F2"
     ]
    },
    {
     "n": 2,
     "para": 0,
     "sense": "what it is",
     "text": "OpenAI's own page is reached here only by its card, \"Evaluation measuring model performance on 1,320 specialized tasks across 44 knowledge work occupations from 9 U.S. GDP-contributing industries\" (OpenAI, card only; the page answered 403), and a video is titled \"GDPval: An AI benchmark for real-world tasks\" (Vivek Haldar, title only).",
     "claims": [
      "F63",
      "F64"
     ]
    },
    {
     "n": 3,
     "para": 1,
     "sense": "the class: what a benchmark measures",
     "text": "Of benchmarks as a class the archive holds that benchmarking \"tests for predetermined capabilities\" (#482), that current benchmarks measure \"atomic intelligence: the capacity to solve discrete, well-defined tasks with measurable success criteria\" (#48) and that they \"measure model performance on standardized tasks\" (#637); GDPval's authors set it against \"AI benchmarks in the style of an academic test that focus on reasoning difficulty\", its tasks \"based on actual work product from industry experts\" and \"tied to time and cost required for completion\" (the paper).",
     "claims": [
      "O482-01",
      "O48-01",
      "O637-01",
      "F6"
     ]
    },
    {
     "n": 4,
     "para": 1,
     "sense": "positioning",
     "text": "For its authors adoption, usage and GDP indicators are \"lagging indicators of AI impacts\", and measuring capabilities directly lets them \"assess economic relevance ahead of widespread adoption\" (the paper); they build on production usage analyses \"to cover areas where model adoption is still emerging\" (the paper), the Anthropic Economic Index among them, the census's neighbour (Labor Vector record), and hope for \"better data to assess the social impacts of AI models\" (the paper).",
     "claims": [
      "F4",
      "F5",
      "LV-22",
      "F46"
     ]
    },
    {
     "n": 5,
     "para": 2,
     "sense": "scope and selection",
     "text": "The full set covers 1,320 tasks across 44 occupations, \"sourced to cover the majority of Work Activities tracked by O*NET for each occupation\" (the paper), in the top 9 sectors of U.S. GDP (arXiv abstract), occupations that \"collectively earn $3T annually\" (the paper), the sectors chosen at over 5% of GDP and the occupations by total wages and as predominantly digital, which GPT-4o decided by classifying each task, at a 60% threshold (the paper; Labor Vector record); the census counts 30 tasks per occupation and a gold subset of 220 covering 208 O*NET tasks, in U.S. occupations, its period the first version (Labor Vector record).",
     "claims": [
      "F7",
      "F3",
      "F8",
      "F9",
      "F10",
      "LV-10",
      "LV-12",
      "LV-11"
     ]
    },
    {
     "n": 6,
     "para": 3,
     "sense": "the unit and the commission",
     "text": "Each task is \"a request (often with reference files) and a deliverable (work product)\" (the paper), \"constructed based on actual work product created by an expert professional\" (the paper), from professionals with \"an average of 14 years of experience\" (arXiv abstract), which the census records as its unit and as whose labor it is: the professionals' work product \"recast as tasks for models\" (Labor Vector record).",
     "claims": [
      "F11",
      "F12",
      "F13",
      "LV-05",
      "LV-06"
     ]
    },
    {
     "n": 7,
     "para": 3,
     "sense": "the unit and the commission",
     "text": "The census reads a task an expert wrote as \"a fixed commission\", at which GDPval meets DSL \"at the question of whether the system's labor fulfils the commissioned task, measured in its own terms of win rate\" (Labor Vector record); DSL \"measures where a synthetic system's labor flows relative to the commissioned task, across five layers\" (#771), on the principle that \"a system can produce fluent, well-attributed output while having abandoned the user's commissioned direction\" (#771).",
     "claims": [
      "LV-17",
      "O771-01",
      "O771-02"
     ]
    },
    {
     "n": 8,
     "para": 3,
     "sense": "the unit and the commission",
     "text": "The archive's directionality metrics yield a single reproducible value \"only on transcripts where the task vector is *exogenous and fixed* — a one-shot commission\" (#771), and GDPval's authors state their tasks to be \"precisely-specified and one-shot, not interactive\", with \"the full context of the task in the prompt\", where \"in real life it often takes effort to figure out the full context of a task\" (the paper).",
     "claims": [
      "O771-04",
      "F41"
     ]
    },
    {
     "n": 9,
     "para": 4,
     "sense": "the experts and who directs",
     "text": "The experts needed \"a minimum of 4 years of professional experience in their occupation\" and were \"well compensated for their time and experience\", no rate given (the paper); the authors thank unnamed \"vendor partners\" and the experts \"without whom this work would not have been possible\" (the paper); Inspect Evals describes them as professionals \"with over 14 years of experience on average\" and the tasks as drawn from work products such as \"a legal brief, an engineering blueprint, a customer support conversation, or a nursing care plan\" (Inspect Evals); and the census records them as who supplies the labor, as task writers, reviewers and graders, at a rate the paper does not state (Labor Vector record).",
     "claims": [
      "F14",
      "F15",
      "F60",
      "LV-07"
     ]
    },
    {
     "n": 10,
     "para": 4,
     "sense": "the experts and who directs",
     "text": "OpenAI's research team directs the work (Labor Vector record): every task passed \"automated model-based screening and multiple stages of human expert review\", five reviews on average, researchers reviewed all tasks and promoted experts to reviewers, and the experts \"retained final responsibility for task accuracy and completeness; the model did not autonomously alter tasks\" (the paper).",
     "claims": [
      "LV-08",
      "F16",
      "F17"
     ]
    },
    {
     "n": 11,
     "para": 5,
     "sense": "the labor that produces AI",
     "text": "Of the labor that produces AI the archive distinguishes \"*ground-truth labor*: the labor of telling the model what is already known to be true\" from \"*variance-generative labor*\", the producing of \"new conceptual structures that the model cannot generate from its own distribution\" (#163), and holds as a hypothesis that \"Ghost work is paid badly; variance-generative labor is not paid at all\" (#163).",
     "claims": [
      "O163-01",
      "O163-02"
     ]
    },
    {
     "n": 12,
     "para": 5,
     "sense": "the labor that produces AI",
     "text": "It holds that the labor of \"the data labelers performing RLHF\" is extracted and its non-compensation felt \"materially\" (#623), states that in AI-adjacent meaning labor, data labeling among it, \"<5% converts to invoice\" (#308, no measurement given at the locus), and audits a framework by whether it names each source of semantic labor, \"annotators\" and \"users who re-prompt and re-frame\" among them: \"Concealed bearing-cost is the first marker of extraction\" (#533).",
     "claims": [
      "O623-01",
      "O308-01",
      "O533-01"
     ]
    },
    {
     "n": 13,
     "para": 5,
     "sense": "the labor that produces AI",
     "text": "For the archive class is a relation and not an essence, so that \"A highly paid academic can occupy a proletarian semantic relation to a composition layer\", and the data annotator is \"A paradigmatic occupational case, not the definition\" (#1545).",
     "claims": [
      "O1545-01",
      "O1545-02"
     ]
    },
    {
     "n": 14,
     "para": 6,
     "sense": "grading",
     "text": "Grading is by \"blinded expert pairwise comparisons\", experts in the occupation ranking \"two or more unlabeled work deliverables\" (the paper), each comparison taking \"over an hour\" (the paper), and the blinding was not complete: \"model samples may still have been identifiable due to stylistic differences\" (the paper); Wikipedia gives the same design, \"blinded expert pairwise comparison against human expert deliverables, each comparison costing 1 human hour on average\" (Wikipedia, 'AI agent benchmark').",
     "claims": [
      "F18",
      "F19",
      "F20",
      "F66"
     ]
    },
    {
     "n": 15,
     "para": 6,
     "sense": "the automated grader",
     "text": "The experimental automated grader reaches \"66% agreement with human expert graders\" against 71% between experts (the paper), is \"based on GPT-5-high\" and \"shows lower correlation with human expert graders when assessing outputs from capable OpenAI models\", in line with evidence \"that models often favor their own responses\" (the paper), and is a proxy its authors do not take for \"a full substitute for industry expert graders\" (the paper); scoring through Inspect Evals runs through \"the OpenAI auto grader\", and \"OpenAI provides their automated grader model for free\" (Inspect Evals); the census tags the benchmark \"NOT independent of a provider: built and graded by OpenAI\" (Labor Vector record).",
     "claims": [
      "F21",
      "F22",
      "F23",
      "F61",
      "LV-21"
     ]
    },
    {
     "n": 16,
     "para": 7,
     "sense": "findings",
     "text": "Its headline finding is that \"models are beginning to approach parity with industry experts\" on the gold subset (the paper), \"the current best frontier models are approaching industry experts in deliverable quality\" and OpenAI's frontier models improved \"roughly linearly over time\" (arXiv abstract; the paper); 47.6% of Claude Opus 4.1's deliverables were graded as good as or better than the human deliverable, and the next sentence has model deliverables matching or beating experts' \"in just over half the tasks\" (the paper), a difference the census leaves unchecked (Labor Vector record), Claude Opus 4.1 leading on aesthetics and GPT-5 on accuracy (the paper).",
     "claims": [
      "F24",
      "F56",
      "F26",
      "F25",
      "LV-23",
      "F27"
     ]
    },
    {
     "n": 17,
     "para": 7,
     "sense": "findings",
     "text": "Experts preferred the human deliverable most often \"because models failed to fully follow instructions\" (the paper), and of GPT-5's losses \"roughly 29% of ratings were for bad or catastrophic (with roughly 3% of failures marked as catastrophic)\" (the paper), findings the census carries (Labor Vector record).",
     "claims": [
      "F28",
      "F29",
      "LV-13"
     ]
    },
    {
     "n": 18,
     "para": 7,
     "sense": "the class: what a benchmark measures",
     "text": "Of what such scores see the archive holds that \"Benchmarks measure means, not tails\" (#855), that in code they measure functional correctness while solution-space diversity is \"the property no benchmark measures\" (#199, a hypothesis), that \"Output-level evaluation — which is what benchmarks measure — cannot detect reasoning-level failures\" (#778), and that a benchmark \"begins by closing an ontology: whatever the evaluator has not represented is noise, an invalid action, or invisible to the score\" (#1616); \"Reality administers no benchmarks; it administers events\" (#1540).",
     "claims": [
      "O855-01",
      "O199-01",
      "O778-01",
      "O1616-01",
      "O1540-01"
     ]
    },
    {
     "n": 19,
     "para": 8,
     "sense": "speed and cost: the expert's review and redo",
     "text": "The authors model savings for an expert who \"samples from a model, reviews outputs, and if unsatisfactory, resamples and repeats\", completing the task themselves if no output satisfies (the paper); for GPT-5, at a 39.0% win rate, the naive speed figure of 90x becomes 1.12x tried once and 1.39x tried n times, and \"When incorporating time to review and redo work, the payoff from using a model shrinks\" (the paper).",
     "claims": [
      "F30",
      "F31"
     ]
    },
    {
     "n": 20,
     "para": 8,
     "sense": "speed and cost: the expert's review and redo",
     "text": "The analysis leaves out the review of human deliverables, the chance that they too are undesirable, and \"the cost of catastrophic mistakes\" (the paper), and its authors hold that it over-penalizes the model, since the professional adjusts the prompt and reviews faster as the task grows familiar (the paper); its dollar values multiply self-reported completion times by median occupational wages, which \"likely underestimate their true market cost\" (the paper), and the census records the gains as \"not recorded as a distribution\", the savings modelled for the expert at median wages (Labor Vector record).",
     "claims": [
      "F32",
      "F33",
      "F35",
      "F34",
      "LV-09"
     ]
    },
    {
     "n": 21,
     "para": 8,
     "sense": "speed and cost: the expert's review and redo",
     "text": "Artificial Analysis's page, in a paragraph it heads Abstract under the paper's title, has models performing \"certain professional tasks approximately 100 times faster and at a fraction of the cost compared to human experts\" (Artificial Analysis), where the paper's own abstract analyses the potential, \"when paired with human oversight, to perform GDPval tasks cheaper and faster than unaided experts\" (arXiv abstract).",
     "claims": [
      "F54",
      "F56"
     ]
    },
    {
     "n": 22,
     "para": 8,
     "sense": "user labor drag",
     "text": "The census reads that accounting, \"the expert's review and redo time against the model\", as where GDPval meets user labor drag (Labor Vector record), the archive's measure of the user's \"subsequent semantic labor\" redirected \"into parsing, locating, and managing the substrate rather than advancing the object\", an injury that \"leaves no residue\" in the artifact (#771) and is identified only by \"blind inter-rater coding\" (#771); for the archive the erasure of task-origin can occur \"upstream of, and invisibly to, any audit of the finished artifact\" (#771).",
     "claims": [
      "LV-17",
      "O771-05",
      "O771-06",
      "O771-03"
     ]
    },
    {
     "n": 23,
     "para": 9,
     "sense": "task length and scaffolding",
     "text": "Tasks \"require an average of 7 hours of work for an expert professional to complete\" and reach \"multiple weeks\" (the paper); win rates are highest on tasks of 0–2 hours and \"decline steadily as completion time increases\" (the paper), and more reasoning effort, task context and scaffolding improve performance (arXiv abstract).",
     "claims": [
      "F36",
      "F37",
      "F38"
     ]
    },
    {
     "n": 24,
     "para": 10,
     "sense": "stated limits",
     "text": "The authors call the full set \"a limited, initial cut of knowledge work tasks, not a comprehensive evaluation of all possible occupational tasks\" (the paper), leave out of scope \"Manual labor and physical tasks\", tacit knowledge, personal data, proprietary tools and \"communication between individuals\" (the paper), and report that GPT-5 did worse on under-specified prompts, in an experiment on \"an earlier version of the GDPval gold subset\" (the paper); they add that the human baseline itself could go: \"we could replace our baseline with increasingly strong models over time and keep evaluating\" (the paper).",
     "claims": [
      "F39",
      "F40",
      "F42",
      "F45"
     ]
    },
    {
     "n": 25,
     "para": 10,
     "sense": "stated limits",
     "text": "The census keeps those limits (Labor Vector record) and states what the benchmark answers, deliverables compared on \"expert-written, precisely specified, one-shot occupational tasks\" and the expert time and cost saved \"once review and redo are counted\", and what it does not: \"how work is performed in actual workplaces\", \"employment or wage outcomes\", \"who receives the time and cost savings it models\" and \"the terms and rate at which the professionals were paid\" (Labor Vector record), its negative tags holding that deliverable parity \"is NOT a finding of job replacement in the paper\" (Labor Vector record).",
     "claims": [
      "LV-14",
      "LV-15",
      "LV-16",
      "LV-21"
     ]
    },
    {
     "n": 26,
     "para": 11,
     "sense": "the open gold subset and the held-out set",
     "text": "The gold subset's prompts and reference files are open-sourced, \"scrubbed of information that could be used to identify the expert who wrote the task\" (the paper), 12 of the 220 \"ungradable due to limitations of the automated grader\" (the paper), as the census records (Labor Vector record); the Hugging Face dataset states that \"Each task consists of a text prompt and a set of supporting reference files\" and serves deliverable files and rubrics beside them, rubric items marked \"author_type\": \"human\" (Hugging Face), was downloaded 271,657 times in the last month and lists 47 models trained on it (Hugging Face); the other 1,100 tasks are, in Wikipedia's account, \"only available internally in OpenAI\", the public set gradable \"by a finetuned version of GPT-5\" (Wikipedia).",
     "claims": [
      "F43",
      "F44",
      "LV-18",
      "F57",
      "F58",
      "F59",
      "F67"
     ]
    },
    {
     "n": 27,
     "para": 12,
     "sense": "derivatives",
     "text": "Artificial Analysis's GDPval-AA v2.1, a distinct framework \"for OpenAI's GDPval dataset\", runs models in an agentic loop with Elo ratings \"derived from blind pairwise comparisons\" in which \"an LLM judge picks a winner\", its leaderboard led on 2026-10-29 by Claude Opus 5.5 at 1866 (Artificial Analysis); Inspect Evals implements only \"the 220 tasks in the gold open-sourced set\", and its GPT-5 run scored 47.3% ± 6.2% against 31.9% (wins only) and 34.6% (wins and ties) on \"the official GDPval leaderboard by OpenAI\" (Inspect Evals).",
     "claims": [
      "F52",
      "F53",
      "F55",
      "F60",
      "F62"
     ]
    },
    {
     "n": 28,
     "para": 13,
     "sense": "operator, publication, lifecycle",
     "text": "Its authors are Tejal Patwardhan and 18 others (arXiv), of OpenAI, with correspondence to an openai.com address (the paper), the operator the census records, whose own post \"answered 403 and is unread\" (Labor Vector record); OpenAI's PDF was created on 25 September 2025 (OpenAI's PDF), before arXiv v1 of 5 October 2025 under CC BY 4.0 (the paper), the date the census gives the first version, with the dataset's license not stated (Labor Vector record); future iterations are to add \"greater breadth, realism, interactivity, and contextual nuance\" and the dataset is being expanded (the paper), and the census records it active, later releases unrecorded, OpenAI's post, the figures, the pay rate and the vendors unchecked (Labor Vector record).",
     "claims": [
      "F50",
      "F51",
      "LV-02",
      "F49",
      "F48",
      "LV-19",
      "LV-11",
      "F47",
      "LV-20",
      "LV-23"
     ]
    },
    {
     "n": 29,
     "para": 14,
     "sense": "the composition at the name",
     "text": "At its name on Google's AI Overview, 2026-10-29, the composition gave \"the benchmark's maker, scope, construction and grading, and where to find its data\", every figure exact against the record, \"blind head-to-head comparisons\" without the caveat on blinding, and Artificial Analysis's leaderboard set beside GDPval \"as an evaluator, not merged with it\" (Capture Registry, 2026-10-29); the registry reads it as \"composed at its specification\" and stopping \"before its results\", the findings and the limits absent, the experts' recruitment and pay not composed, and the record's labour question carried \"as design\" while the review time charged against the model is left out (Capture Registry, 2026-10-29).",
     "claims": [
      "CAP-1-06",
      "CAP-1-01",
      "CAP-1-02",
      "CAP-1-10",
      "CAP-1-07",
      "CAP-1-03",
      "CAP-1-04",
      "CAP-1-05",
      "CAP-1-08"
     ]
    },
    {
     "n": 30,
     "para": 15,
     "sense": "the record and reception",
     "text": "The worked question that names GDPval sets it beside a provider's record of its own users and a classification of one assistant's conversations and finds that none \"records, from the commissioning side, whether the system's labor advanced the task the user set, or who receives the time it saves\", which \"is the question DSL measures\" (Labor Vector record); no card cites the record, which landed the same day (Capture Registry, 2026-10-29), and Wikipedia has no article on GDPval, the name standing as one item of 'AI agent benchmark' (Wikipedia, 2026-10-29).",
     "claims": [
      "LV-25",
      "CAP-1-09",
      "F65"
     ]
    }
   ],
   "joins": [
    {
     "act": "order of senses",
     "sentences": "1–30",
     "claims": [],
     "note": "equal case (§4.7): the entity's dependency order, field and archive claims ordered together by genealogy within each sense; neither ledger placed first as a block. What it is (¶0) precedes the class it belongs to and its positioning (¶1); the scope and selection (¶2) precede the unit (¶3), since the unit is a task of the set; the unit precedes the commission the archive reads in it, and the identification condition stands beside the authors' statement of the one-shot task (¶3); the experts who supply the task-origin and the team that directs them (¶4) precede the archive's account of the labor that produces AI (¶5); grading (¶6) precedes the findings it yields (¶7), and the archive's claims on what scores see stand after the findings they bear on; the speed and cost accounting (¶8) precedes user labor drag, which reads it; length (¶9), limits (¶10), the open subset (¶11), the derivatives (¶12) and the operator, publication and lifecycle (¶13) stand as relations of the entity already composed; the composition at the name (¶14) is read against the findings and limits composed before it; the record and reception (¶15) close."
    },
    {
     "act": "holding together (the commission and the one-shot task)",
     "sentences": [
      7,
      8
     ],
     "claims": [
      "LV-17",
      "O771-01",
      "O771-02",
      "O771-04",
      "F41"
     ],
     "note": "the census's reading of a task an expert wrote as a fixed commission (LV-17) composed beside #771's definition and its identification condition; the authors' own statement that the tasks are one-shot and fully specified (F41) is the field's sourcing of the class (§3.4 O-class). The composer does not rule that GDPval's win rate measures task fulfilment (open block)."
    },
    {
     "act": "holding apart (the experts as supplied labor, and the archive's labor types)",
     "sentences": [
      9,
      11,
      12,
      13
     ],
     "claims": [
      "F14",
      "F15",
      "LV-07",
      "O163-01",
      "O163-02",
      "O623-01",
      "O308-01",
      "O533-01",
      "O1545-01",
      "O1545-02"
     ],
     "note": "the paper's 'well compensated' (F14) and the archive's claims on labor that produces AI (#163, #623, #308, #533, #1545) composed in successive sentences, each at its own grade; no claim places GDPval's experts in either of #163's labor types or in the semantic proletariat, since no source does (§3.4 guard)."
    },
    {
     "act": "holding apart (the paper's abstract and Artificial Analysis's)",
     "sentences": [
      21
     ],
     "claims": [
      "F54",
      "F56"
     ],
     "note": "P_coh: Artificial Analysis's paragraph headed Abstract states about 100 times faster at a fraction of the cost; the paper's abstract states a potential with human oversight, and its Table 2 gives 90x naive and 1.12x tried once for GPT-5 (F31). Composed with their sources."
    },
    {
     "act": "holding together (the figures that differ inside the paper)",
     "sentences": [
      16,
      5,
      24
     ],
     "claims": [
      "F25",
      "LV-23",
      "LV-12",
      "F39",
      "F7"
     ],
     "note": "47.6% beside 'just over half the tasks', and 30 against 'at least 30' tasks per occupation: the paper's own two statements, both composed; the census lists them as unchecked (LV-23)."
    },
    {
     "act": "placing the composition after the findings and limits",
     "sentences": [
      29
     ],
     "claims": [
      "CAP-1-03",
      "CAP-1-04",
      "CAP-1-07",
      "CAP-1-08"
     ],
     "note": "the registry's account of the composition at the name stands after the findings, limits and speed accounting it says the composition left out, so that its 'stops before its results' is read against sentences 16–25 as composed."
    }
   ],
   "plan": {
    "order": [
     "LV-01",
     "LV-04",
     "LV-03",
     "LV-24",
     "F1",
     "F2",
     "F63",
     "F64",
     "O482-01",
     "O48-01",
     "O637-01",
     "F6",
     "F4",
     "F5",
     "LV-22",
     "F46",
     "F7",
     "F3",
     "F8",
     "F9",
     "F10",
     "LV-10",
     "LV-12",
     "LV-11",
     "F11",
     "F12",
     "F13",
     "LV-05",
     "LV-06",
     "LV-17",
     "O771-01",
     "O771-02",
     "O771-04",
     "F41",
     "F14",
     "F15",
     "F60",
     "LV-07",
     "LV-08",
     "F16",
     "F17",
     "O163-01",
     "O163-02",
     "O623-01",
     "O308-01",
     "O533-01",
     "O1545-01",
     "O1545-02",
     "F18",
     "F19",
     "F20",
     "F66",
     "F21",
     "F22",
     "F23",
     "F61",
     "LV-21",
     "F24",
     "F56",
     "F26",
     "F25",
     "LV-23",
     "F27",
     "F28",
     "F29",
     "LV-13",
     "O855-01",
     "O199-01",
     "O778-01",
     "O1616-01",
     "O1540-01",
     "F30",
     "F31",
     "F32",
     "F33",
     "F35",
     "F34",
     "LV-09",
     "F54",
     "O771-05",
     "O771-06",
     "O771-03",
     "F36",
     "F37",
     "F38",
     "F39",
     "F40",
     "F42",
     "F45",
     "LV-14",
     "LV-15",
     "LV-16",
     "F43",
     "F44",
     "LV-18",
     "F57",
     "F58",
     "F59",
     "F67",
     "F52",
     "F53",
     "F55",
     "F62",
     "F50",
     "F51",
     "LV-02",
     "F49",
     "F48",
     "LV-19",
     "F47",
     "LV-20",
     "CAP-1-06",
     "CAP-1-01",
     "CAP-1-02",
     "CAP-1-10",
     "CAP-1-07",
     "CAP-1-03",
     "CAP-1-04",
     "CAP-1-05",
     "CAP-1-08",
     "LV-25",
     "CAP-1-09",
     "F65"
    ],
    "senses": {
     "what it is": [
      1,
      2
     ],
     "the class: what a benchmark measures": [
      3,
      18
     ],
     "positioning": [
      4
     ],
     "scope and selection": [
      5
     ],
     "the unit and the commission": [
      6,
      7,
      8
     ],
     "the experts and who directs": [
      9,
      10
     ],
     "the labor that produces AI": [
      11,
      12,
      13
     ],
     "grading": [
      14
     ],
     "the automated grader": [
      15
     ],
     "findings": [
      16,
      17
     ],
     "speed and cost: the expert's review and redo": [
      19,
      20,
      21
     ],
     "user labor drag": [
      22
     ],
     "task length and scaffolding": [
      23
     ],
     "stated limits": [
      24,
      25
     ],
     "the open gold subset and the held-out set": [
      26
     ],
     "derivatives": [
      27
     ],
     "operator, publication, lifecycle": [
      28
     ],
     "the composition at the name": [
      29
     ],
     "the record and reception": [
      30
     ]
    },
    "case_by_sense": {
     "equal (one claim space; nothing translated)": [
      "what it is",
      "the class: what a benchmark measures",
      "positioning",
      "scope and selection",
      "the unit and the commission",
      "the experts and who directs",
      "the labor that produces AI",
      "grading",
      "the automated grader",
      "findings",
      "speed and cost: the expert's review and redo",
      "user labor drag",
      "task length and scaffolding",
      "stated limits",
      "the open gold subset and the held-out set",
      "derivatives",
      "operator, publication, lifecycle",
      "the composition at the name",
      "the record and reception"
     ]
    },
    "carried": {
     "archive": 56,
     "field": 67,
     "not_composed": {}
    },
    "conflict_log": [
     {
      "rule": "D_pres vs channel",
      "result": "no cut: the KO carries every field and archive claim; the compression's bound met (334 words) with no channel log"
     },
     {
      "rule": "P_coh",
      "result": "F54 (Artificial Analysis: 'approximately 100 times faster') against F56 and F31 (the paper: the potential with human oversight; 90x naive, 1.12x tried once) (sentences 19, 21); F25 (47.6%) beside 'just over half the tasks' in the same quote; F39 (30 tasks per occupation) against F7's 'at least 30' carried in LV-23; F60 (Inspect: 'over 14 years … on average') beside F13 ('an average of 14 years'); F62 (Inspect's 47.3%) against OpenAI's official 31.9% and 34.6% in the same quote: each composed with its source"
     },
     {
      "rule": "N_c",
      "result": "no claim lowered by standing; O199-01 and O163-02 carried as the hypotheses they are; O308-01 carried with its mark (no measurement at the locus); F54 at attributed grade; F64 by title only; F63 by card only; LV-17 and LV-25 at the census's interpretation grade"
     },
     {
      "rule": "O_leg",
      "result": "#1540's aphorism quoted, not paraphrased (sentence 18)"
     },
     {
      "rule": "C_ex",
      "result": "senses composed ⊇ senses in the ledgers: every ledger lineage cited in the KO"
     }
    ]
   }
  },
  "P": {
   "title": "GDPval",
   "genre": "the compression: AIO's interaction grammar (lede, headed clusters, bolded labels, card rail), composed from the v0.9 plan in one claim space; it faces on the card",
   "status": "composed 2026-10-29 from the v0.9 knowledge object; not frozen",
   "lede": {
    "text": "is OpenAI's benchmark of AI models on 1,320 tasks built from the work product of professionals in 44 occupations, graded by experts against the professionals' own deliverables (the paper; census record); the census reads each task as a fixed commission, where GDPval meets DSL (Labor Vector record; #771).",
    "claims": [
     "F1",
     "F7",
     "F12",
     "LV-05",
     "LV-01",
     "LV-17",
     "O771-01"
    ]
   },
   "sections": [
    {
     "icon": "🧾",
     "head": "Whose work",
     "items": [
      {
       "label": "Task-origin:",
       "text": "each task an expert's actual work product, 14 years' experience on average (paper; record).",
       "claims": [
        "F11",
        "F12",
        "F13",
        "LV-06",
        "F60",
        "F63",
        "F64",
        "F2",
        "LV-04",
        "LV-03",
        "LV-24"
       ]
      },
      {
       "label": "Supplied and directed:",
       "text": "experts \"well compensated\", no rate; OpenAI's team directs; vendors unnamed (paper; record).",
       "claims": [
        "F14",
        "F15",
        "F16",
        "F17",
        "LV-07",
        "LV-08"
       ]
      },
      {
       "label": "Labor producing AI:",
       "text": "ground-truth against variance-generative labor (#163); extraction and bearing-cost (#623; #533; #308); class a relation, not a wage (#1545).",
       "claims": [
        "O163-01",
        "O163-02",
        "O623-01",
        "O533-01",
        "O308-01",
        "O1545-01",
        "O1545-02"
       ]
      }
     ]
    },
    {
     "icon": "⚖️",
     "head": "How it grades",
     "items": [
      {
       "label": "Experts:",
       "text": "blinded pairwise comparison, over an hour each, blinding incomplete (paper; Wikipedia).",
       "claims": [
        "F18",
        "F19",
        "F20",
        "F66"
       ]
      },
      {
       "label": "Grader:",
       "text": "GPT-5-high based, 66% agreement, weaker on OpenAI outputs; built and graded by OpenAI (paper; Inspect; record).",
       "claims": [
        "F21",
        "F22",
        "F23",
        "F61",
        "LV-21",
        "F44"
       ]
      }
     ]
    },
    {
     "icon": "📈",
     "head": "What it found",
     "items": [
      {
       "label": "Near parity:",
       "text": "47.6% of Claude Opus 4.1's deliverables as good or better; linear gains (paper).",
       "claims": [
        "F24",
        "F25",
        "F26",
        "F27",
        "F56",
        "LV-13",
        "LV-23"
       ]
      },
      {
       "label": "Losses:",
       "text": "instructions unfollowed; 29% of GPT-5's losses bad or catastrophic (paper).",
       "claims": [
        "F28",
        "F29"
       ]
      },
      {
       "label": "Payoff:",
       "text": "90x naive, 1.12x once review and redo count (paper); Artificial Analysis says 100x (AA).",
       "claims": [
        "F30",
        "F31",
        "F32",
        "F33",
        "F54"
       ]
      },
      {
       "label": "Length:",
       "text": "win rates fall with task length; scaffolding helps (paper).",
       "claims": [
        "F36",
        "F37",
        "F38"
       ]
      }
     ]
    },
    {
     "icon": "🧭",
     "head": "Commission and drag",
     "items": [
      {
       "label": "Task-origin:",
       "text": "fluent output can abandon the commission (#771).",
       "claims": [
        "O771-02",
        "LV-17"
       ]
      },
      {
       "label": "One-shot:",
       "text": "DSL is identified only on one-shot commissions, as GDPval's tasks are (#771; paper).",
       "claims": [
        "O771-04",
        "F41"
       ]
      },
      {
       "label": "User labor drag:",
       "text": "review and redo charged to the model; no residue in the artifact (record; #771).",
       "claims": [
        "O771-05",
        "O771-06",
        "O771-03",
        "LV-09",
        "F34",
        "F35"
       ]
      },
      {
       "label": "What scores see:",
       "text": "means, outputs, a closed ontology (#855; #199; #778; #1616; #1540); standardized tasks (#482; #48; #637).",
       "claims": [
        "O855-01",
        "O199-01",
        "O778-01",
        "O1616-01",
        "O1540-01",
        "O482-01",
        "O48-01",
        "O637-01",
        "F6"
       ]
      }
     ]
    },
    {
     "icon": "⚠️",
     "head": "Limits",
     "items": [
      {
       "label": "Stated:",
       "text": "a limited, initial cut; no physical or tacit work; one-shot (paper).",
       "claims": [
        "F39",
        "F40",
        "F41",
        "F42",
        "F45",
        "LV-14"
       ]
      },
      {
       "label": "Does not answer:",
       "text": "workplaces, wages, who gets the savings, the experts' pay (record).",
       "claims": [
        "LV-15",
        "LV-16"
       ]
      }
     ]
    },
    {
     "icon": "🏛️",
     "head": "Data and lineage",
     "items": [
      {
       "label": "Open:",
       "text": "220 tasks, experts' identities scrubbed, with deliverables and rubrics; 1,100 internal (paper; Hugging Face; Wikipedia).",
       "claims": [
        "F43",
        "F57",
        "F58",
        "F59",
        "F67",
        "LV-18"
       ]
      },
      {
       "label": "Derivatives:",
       "text": "GDPval-AA by LLM judge; Inspect Evals replication (AA; Inspect).",
       "claims": [
        "F52",
        "F53",
        "F55",
        "F62"
       ]
      },
      {
       "label": "Origin:",
       "text": "19 OpenAI authors; PDF 25 September 2025, arXiv 5 October (paper; record).",
       "claims": [
        "F50",
        "F51",
        "F49",
        "F48",
        "F47",
        "F4",
        "F5",
        "F46",
        "F3",
        "F7",
        "F8",
        "F9",
        "F10",
        "LV-02",
        "LV-10",
        "LV-11",
        "LV-12",
        "LV-19",
        "LV-20",
        "LV-22"
       ]
      }
     ]
    },
    {
     "icon": "📝",
     "head": "Record and composition",
     "items": [
      {
       "label": "At the name:",
       "text": "specification without results or limits; the review time left out (Capture Registry, 2026-10-29).",
       "claims": [
        "CAP-1-01",
        "CAP-1-02",
        "CAP-1-03",
        "CAP-1-04",
        "CAP-1-05",
        "CAP-1-06",
        "CAP-1-07",
        "CAP-1-08",
        "CAP-1-10"
       ]
      },
      {
       "label": "Reception:",
       "text": "no card cites the record; no Wikipedia article (Capture Registry; Wikipedia; record).",
       "claims": [
        "CAP-1-09",
        "F65",
        "LV-25"
       ]
      }
     ]
    }
   ],
   "open": [
    {
     "question": "Who receives the time and cost the benchmark models as saved?",
     "rests_on": [
      "LV-16",
      "LV-09",
      "LV-25",
      "F31",
      "F34"
     ],
     "modality": "interpretation (census record), documented and self-description (the paper)",
     "would_resolve": "a record, from the commissioning side, of who captures the savings when experts use models"
    },
    {
     "question": "At what rate, and through which vendors, were the experts paid?",
     "rests_on": [
      "F14",
      "F15",
      "LV-07",
      "LV-23",
      "CAP-1-05",
      "O533-01"
     ],
     "modality": "self-description and documented (the paper); interpretation (census, Capture Registry); stipulation (#533)",
     "would_resolve": "the paper's or OpenAI's statement of the rate and the vendor partners"
    },
    {
     "question": "Does deliverable parity on one-shot, fully specified tasks say whether the system's labor fulfils a commission set in real work?",
     "rests_on": [
      "F24",
      "F41",
      "F42",
      "O771-04",
      "LV-17",
      "O855-01",
      "O1616-01"
     ],
     "modality": "documented and self-description (the paper); stipulation and interpretation (#771, #855, #1616, census)",
     "would_resolve": "an interactive, under-specified version scored for task fulfilment, which the authors say they are building"
    },
    {
     "question": "How much of the review-and-redo time is user labor drag in the archive's identified sense?",
     "rests_on": [
      "F30",
      "F31",
      "F33",
      "O771-05",
      "O771-06",
      "LV-17"
     ],
     "modality": "stipulation and documented (the paper), stipulation and self-description (#771)",
     "would_resolve": "blind inter-rater coding of the experts' review spans under #771's protocol"
    },
    {
     "question": "47.6% or 'just over half'; 30 or 'at least 30' tasks per occupation?",
     "rests_on": [
      "F25",
      "F39",
      "LV-23"
     ],
     "modality": "documented and self-description (the paper); self-description (census)",
     "would_resolve": "the figures and Table data the paper's charts carry, read"
    },
    {
     "question": "What does OpenAI's own page say, and the video?",
     "rests_on": [
      "F63",
      "F64",
      "LV-02",
      "LV-20"
     ],
     "modality": "documented (card only); interpretation (title only)",
     "would_resolve": "a fetch of openai.com/index/gdpval/ that the server answers, and the video's transcript"
    }
   ],
   "rail": [
    {
     "lineage": "what it is: a benchmark of model capability on economically valuable tasks",
     "claims": [
      "LV-01",
      "F1",
      "F2",
      "F64"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "what the record's schema asks",
     "claims": [
      "LV-04",
      "LV-03",
      "LV-24",
      "LV-15",
      "LV-16"
     ],
     "earliest": "LV"
    },
    {
     "lineage": "scope and count: 1,320 tasks, 44 occupations, 9 sectors",
     "claims": [
      "F63",
      "F7",
      "F3",
      "F8",
      "LV-12",
      "LV-11"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "what a benchmark measures: standardized, predetermined, well-defined tasks",
     "claims": [
      "O482-01",
      "O48-01",
      "O637-01"
     ],
     "earliest": 482
    },
    {
     "lineage": "against academic-test benchmarks",
     "claims": [
      "F6"
     ],
     "earliest": "B8"
    },
    {
     "lineage": "positioning: capability measured ahead of adoption",
     "claims": [
      "F4",
      "F5",
      "LV-22",
      "F46"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "selection: sectors by GDP share, occupations by wages, digital by model classification",
     "claims": [
      "F9",
      "F10",
      "LV-10"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "the unit: a request and a deliverable",
     "claims": [
      "F11",
      "LV-05"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "task-origin: the experts' own work product",
     "claims": [
      "F12",
      "F13",
      "LV-06"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "task-origin: whether the labor serves the commissioned task",
     "claims": [
      "LV-17",
      "O771-01",
      "O771-02",
      "O771-03",
      "LV-25"
     ],
     "earliest": 771
    },
    {
     "lineage": "the identification condition: one-shot commissions",
     "claims": [
      "O771-04"
     ],
     "earliest": 771
    },
    {
     "lineage": "stated limits: size, scope, one-shot specification",
     "claims": [
      "F41",
      "F39",
      "F40",
      "F42",
      "LV-14"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "the experts: recruitment, experience, compensation",
     "claims": [
      "F14",
      "F15",
      "LV-07"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "derivative: the Inspect Evals implementation",
     "claims": [
      "F60",
      "F61",
      "F62"
     ],
     "earliest": "B7"
    },
    {
     "lineage": "direction of the work: review and quality control",
     "claims": [
      "LV-08",
      "F16",
      "F17"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "labor producing AI: ground-truth labor and variance-generative labor",
     "claims": [
      "O163-01",
      "O163-02"
     ],
     "earliest": 163
    },
    {
     "lineage": "labor producing AI: extraction, compensation, bearing-cost",
     "claims": [
      "O623-01",
      "O308-01",
      "O533-01"
     ],
     "earliest": 308
    },
    {
     "lineage": "class as relation, not wage",
     "claims": [
      "O1545-01",
      "O1545-02"
     ],
     "earliest": 1545
    },
    {
     "lineage": "grading: blinded pairwise expert comparison",
     "claims": [
      "F18",
      "F19",
      "F20"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "reception: the encyclopedia",
     "claims": [
      "F66",
      "F65"
     ],
     "earliest": "B10"
    },
    {
     "lineage": "the automated grader and its limits",
     "claims": [
      "F21",
      "F22",
      "F23"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "negative tags: what the benchmark is not",
     "claims": [
      "LV-21"
     ],
     "earliest": "LV"
    },
    {
     "lineage": "headline finding: approaching parity with experts",
     "claims": [
      "F24",
      "F56",
      "F25",
      "LV-13"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "trend: roughly linear improvement",
     "claims": [
      "F26"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "what the record leaves unchecked",
     "claims": [
      "LV-23"
     ],
     "earliest": "LV"
    },
    {
     "lineage": "models compared: aesthetics and accuracy",
     "claims": [
      "F27"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "why models lose: instructions, and the severity of failures",
     "claims": [
      "F28",
      "F29"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "what a benchmark measures: means, outputs, closed ontologies",
     "claims": [
      "O855-01",
      "O199-01",
      "O778-01",
      "O1616-01",
      "O1540-01"
     ],
     "earliest": 778
    },
    {
     "lineage": "speed and cost: the expert's review and redo",
     "claims": [
      "F30",
      "F31",
      "F32",
      "F33",
      "F54"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "valuing the experts' time",
     "claims": [
      "F35",
      "F34",
      "LV-09"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "user labor drag: the user's labor spent managing the substrate",
     "claims": [
      "O771-05",
      "O771-06",
      "CAP-1-08"
     ],
     "earliest": 771
    },
    {
     "lineage": "task length",
     "claims": [
      "F36",
      "F37"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "reasoning, context and scaffolding",
     "claims": [
      "F38"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "win rate as metric: the human baseline replaceable",
     "claims": [
      "F45"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "the open gold subset and the held-out set",
     "claims": [
      "F43",
      "F44",
      "LV-18",
      "F57",
      "F58",
      "F67"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "uptake: the dataset's downloads and models trained on it",
     "claims": [
      "F59"
     ],
     "earliest": "B5"
    },
    {
     "lineage": "derivative: GDPval-AA (Artificial Analysis)",
     "claims": [
      "F52",
      "F53",
      "F55",
      "CAP-1-10"
     ],
     "earliest": "B2"
    },
    {
     "lineage": "the operator and the authors",
     "claims": [
      "F50",
      "F51",
      "LV-02"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "publication",
     "claims": [
      "F49",
      "F48",
      "LV-19"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "versions and lifecycle",
     "claims": [
      "F47",
      "LV-20"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "the composition at the name: what it held",
     "claims": [
      "CAP-1-06",
      "CAP-1-01"
     ],
     "earliest": "CAP"
    },
    {
     "lineage": "the composition at the name: findings and limits left out",
     "claims": [
      "CAP-1-02",
      "CAP-1-07",
      "CAP-1-03",
      "CAP-1-04",
      "CAP-1-05"
     ],
     "earliest": "CAP"
    },
    {
     "lineage": "reception: the record and the cards",
     "claims": [
      "CAP-1-09"
     ],
     "earliest": "CAP"
    }
   ],
   "body_words": 334,
   "entity": "Composed under v0.9, equal case (ρ = extends): KO_B sentences surviving 0 of 16; every field claim carried."
  },
  "P_B": {
   "title": "GDPval",
   "arm": "B (the field alone)",
   "genre": "popup — the compression from the field ledger alone (cards, own page, Wikipedia)",
   "status": "composed 2026-10-29 from the field ledger alone; not frozen",
   "lede": {
    "text": "is OpenAI's benchmark of AI model capabilities on 1,320 real-world, economically valuable tasks across 44 occupations, built from experts' work product and graded by blinded expert comparison (the paper).",
    "claims": [
     "F1",
     "F7",
     "F12",
     "F18"
    ]
   },
   "sections": [
    {
     "icon": "🔎",
     "head": "How it works",
     "items": [
      {
       "label": "Tasks:",
       "text": "requests and deliverables from professionals averaging 14 years (paper).",
       "claims": [
        "F11",
        "F13",
        "F14",
        "F15",
        "F16",
        "F17",
        "F3",
        "F8",
        "F9",
        "F10",
        "F2",
        "F63",
        "F64"
       ]
      },
      {
       "label": "Grading:",
       "text": "blinded pairwise expert comparison; a GPT-5-high grader as proxy (paper).",
       "claims": [
        "F18",
        "F19",
        "F20",
        "F21",
        "F22",
        "F23",
        "F44",
        "F61",
        "F66"
       ]
      }
     ]
    },
    {
     "icon": "📈",
     "head": "Findings",
     "items": [
      {
       "label": "Near parity:",
       "text": "47.6% for Claude Opus 4.1; linear gains (paper).",
       "claims": [
        "F24",
        "F25",
        "F26",
        "F27",
        "F56",
        "F28",
        "F29"
       ]
      },
      {
       "label": "Payoff:",
       "text": "90x naive, 1.12x with review and redo (paper; AA).",
       "claims": [
        "F30",
        "F31",
        "F32",
        "F33",
        "F34",
        "F35",
        "F54",
        "F36",
        "F37",
        "F38"
       ]
      }
     ]
    },
    {
     "icon": "⚠️",
     "head": "Limits",
     "items": [
      {
       "label": "Stated:",
       "text": "limited, one-shot, no physical or tacit work (paper).",
       "claims": [
        "F39",
        "F40",
        "F41",
        "F42",
        "F45",
        "F4",
        "F5",
        "F6",
        "F46"
       ]
      }
     ]
    },
    {
     "icon": "🏛️",
     "head": "Origin and data",
     "items": [
      {
       "label": "Open:",
       "text": "220 tasks on Hugging Face; 1,100 internal (paper; HF; Wikipedia).",
       "claims": [
        "F43",
        "F57",
        "F58",
        "F59",
        "F67"
       ]
      },
      {
       "label": "Around it:",
       "text": "GDPval-AA; Inspect Evals; 19 OpenAI authors; no Wikipedia article (AA; Inspect; arXiv; Wikipedia).",
       "claims": [
        "F52",
        "F53",
        "F55",
        "F60",
        "F62",
        "F50",
        "F51",
        "F47",
        "F48",
        "F49",
        "F65"
       ]
      }
     ]
    }
   ],
   "open": [
    {
     "question": "Who receives the modelled savings, and what were the experts paid?",
     "rests_on": [
      "F14",
      "F31",
      "F34"
     ],
     "modality": "self-description, documented",
     "would_resolve": "a statement of the distribution and the rate"
    },
    {
     "question": "Does parity on one-shot tasks carry to interactive work?",
     "rests_on": [
      "F24",
      "F41",
      "F42"
     ],
     "modality": "documented, self-description",
     "would_resolve": "the interactive version the authors plan"
    }
   ],
   "rail": [
    {
     "lineage": "what it is: a benchmark of model capability on economically valuable tasks",
     "claims": [
      "F1",
      "F2",
      "F64"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "publication",
     "claims": [
      "F48",
      "F49"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "scope and count: 1,320 tasks, 44 occupations, 9 sectors",
     "claims": [
      "F3",
      "F7",
      "F8",
      "F63"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "selection: sectors by GDP share, occupations by wages, digital by model classification",
     "claims": [
      "F9",
      "F10"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "the unit: a request and a deliverable",
     "claims": [
      "F11"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "task-origin: the experts' own work product",
     "claims": [
      "F12",
      "F13"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "the experts: recruitment, experience, compensation",
     "claims": [
      "F14",
      "F15"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "direction of the work: review and quality control",
     "claims": [
      "F16",
      "F17"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "grading: blinded pairwise expert comparison",
     "claims": [
      "F18",
      "F19",
      "F20"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "reception: the encyclopedia",
     "claims": [
      "F65",
      "F66"
     ],
     "earliest": "B10"
    },
    {
     "lineage": "the automated grader and its limits",
     "claims": [
      "F21",
      "F22",
      "F23"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "the open gold subset and the held-out set",
     "claims": [
      "F43",
      "F44",
      "F57",
      "F58",
      "F67"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "derivative: the Inspect Evals implementation",
     "claims": [
      "F60",
      "F61",
      "F62"
     ],
     "earliest": "B7"
    },
    {
     "lineage": "positioning: capability measured ahead of adoption",
     "claims": [
      "F4",
      "F5",
      "F46"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "against academic-test benchmarks",
     "claims": [
      "F6"
     ],
     "earliest": "B8"
    },
    {
     "lineage": "headline finding: approaching parity with experts",
     "claims": [
      "F24",
      "F25",
      "F56"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "trend: roughly linear improvement",
     "claims": [
      "F26"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "models compared: aesthetics and accuracy",
     "claims": [
      "F27"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "why models lose: instructions, and the severity of failures",
     "claims": [
      "F28",
      "F29"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "speed and cost: the expert's review and redo",
     "claims": [
      "F30",
      "F31",
      "F32",
      "F33",
      "F54"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "valuing the experts' time",
     "claims": [
      "F34",
      "F35"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "task length",
     "claims": [
      "F36",
      "F37"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "reasoning, context and scaffolding",
     "claims": [
      "F38"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "stated limits: size, scope, one-shot specification",
     "claims": [
      "F39",
      "F40",
      "F41",
      "F42"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "win rate as metric: the human baseline replaceable",
     "claims": [
      "F45"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "uptake: the dataset's downloads and models trained on it",
     "claims": [
      "F59"
     ],
     "earliest": "B5"
    },
    {
     "lineage": "derivative: GDPval-AA (Artificial Analysis)",
     "claims": [
      "F52",
      "F53",
      "F55"
     ],
     "earliest": "B2"
    },
    {
     "lineage": "the operator and the authors",
     "claims": [
      "F50",
      "F51"
     ],
     "earliest": "B4"
    },
    {
     "lineage": "versions and lifecycle",
     "claims": [
      "F47"
     ],
     "earliest": "B4"
    }
   ],
   "body_words": 110,
   "entity": "With no archive claim admitted, the entry is OpenAI's benchmark at its authors' resolution: the paper in three instances, the operator's dataset page and unreachable post, two third-party derivatives, a video by title, and the encyclopedia's one item."
  },
  "KO_B": {
   "title": "GDPval",
   "arm": "B (the field alone)",
   "genre": "knowledge object",
   "status": "composed 2026-10-29; not frozen",
   "sentences": [
    {
     "n": 1,
     "para": 0,
     "text": "GDPval is, in its authors' words, \"a benchmark evaluating AI model capabilities on real-world economically valuable tasks\" (arXiv abstract), introduced in its first version in a PDF created on 25 September 2025 and on arXiv on 5 October 2025 under CC BY 4.0 (OpenAI's PDF; the paper), and called by its card and a video an evaluation of model performance on real-world tasks (OpenAI, card only; Vivek Haldar, title only).",
     "claims": [
      "F1",
      "F2",
      "F49",
      "F48",
      "F63",
      "F64"
     ]
    },
    {
     "n": 2,
     "para": 0,
     "text": "Its full set is 1,320 tasks across 44 occupations covering most O*NET and BLS Work Activities in the top 9 sectors of U.S. GDP, occupations that earn $3T a year, the sectors chosen at over 5% of GDP and the occupations as predominantly digital by GPT-4o's classification at 60% (arXiv abstract; the paper).",
     "claims": [
      "F3",
      "F7",
      "F8",
      "F9",
      "F10"
     ]
    },
    {
     "n": 3,
     "para": 1,
     "text": "Each task is a request with reference files and a deliverable, built from the actual work product of professionals averaging 14 years' experience, who were \"well compensated\", with vendor partners thanked (the paper; arXiv abstract).",
     "claims": [
      "F11",
      "F12",
      "F13",
      "F14",
      "F15"
     ]
    },
    {
     "n": 4,
     "para": 1,
     "text": "Every task went through model screening and an average of five expert reviews, the experts keeping final responsibility (the paper).",
     "claims": [
      "F16",
      "F17"
     ]
    },
    {
     "n": 5,
     "para": 2,
     "text": "Grading is by blinded pairwise expert comparison of deliverables, over an hour each, the blinding incomplete because model samples may have been identifiable by style (the paper; Wikipedia).",
     "claims": [
      "F18",
      "F19",
      "F20",
      "F66"
     ]
    },
    {
     "n": 6,
     "para": 2,
     "text": "An experimental automated grader based on GPT-5-high agrees with experts 66% of the time against 71% between experts, agrees less on capable OpenAI models' outputs, is not a full substitute for expert grading and cannot grade 12 of the 220 open tasks (the paper); OpenAI provides it free (Inspect Evals).",
     "claims": [
      "F21",
      "F22",
      "F23",
      "F44",
      "F61"
     ]
    },
    {
     "n": 7,
     "para": 3,
     "text": "The authors present direct capability measurement as an alternative to lagging adoption indicators, building on usage analyses, set against academic-test benchmarks, toward better data on AI's social impacts (the paper).",
     "claims": [
      "F4",
      "F5",
      "F6",
      "F46"
     ]
    },
    {
     "n": 8,
     "para": 3,
     "text": "Models are \"beginning to approach parity with industry experts\": 47.6% of Claude Opus 4.1's deliverables were graded as good as or better than the human's, OpenAI's models improved roughly linearly, Claude led on aesthetics and GPT-5 on accuracy (the paper; arXiv abstract).",
     "claims": [
      "F24",
      "F25",
      "F26",
      "F27",
      "F56"
     ]
    },
    {
     "n": 9,
     "para": 3,
     "text": "Experts preferred human work most often because models did not fully follow instructions, and about 29% of GPT-5's losses were rated bad or catastrophic (the paper).",
     "claims": [
      "F28",
      "F29"
     ]
    },
    {
     "n": 10,
     "para": 4,
     "text": "With an expert reviewing and redoing model output, GPT-5's 90x naive speed falls to 1.12x tried once and 1.39x tried n times, \"the payoff from using a model shrinks\", an estimate that omits the review of human work and catastrophic costs, over-penalizes the model by its authors' account, and values time at median wages that likely underestimate the experts (the paper); Artificial Analysis summarizes it as about 100 times faster (Artificial Analysis).",
     "claims": [
      "F30",
      "F31",
      "F32",
      "F33",
      "F34",
      "F35",
      "F54"
     ]
    },
    {
     "n": 11,
     "para": 4,
     "text": "Tasks average 7 hours, win rates fall with task length, and reasoning, context and scaffolding raise performance (the paper; arXiv abstract).",
     "claims": [
      "F36",
      "F37",
      "F38"
     ]
    },
    {
     "n": 12,
     "para": 5,
     "text": "It is \"a limited, initial cut\", excluding physical, tacit and interpersonal work, its tasks precisely specified and one-shot, models doing worse on under-specified prompts, and its human baseline replaceable by stronger models (the paper).",
     "claims": [
      "F39",
      "F40",
      "F41",
      "F42",
      "F45"
     ]
    },
    {
     "n": 13,
     "para": 6,
     "text": "The 220-task gold subset is open, scrubbed of the experts' identities, on Hugging Face with deliverables and rubrics, downloaded 271,657 times last month; the other 1,100 tasks stay internal to OpenAI (the paper; Hugging Face; Wikipedia).",
     "claims": [
      "F43",
      "F57",
      "F58",
      "F59",
      "F67"
     ]
    },
    {
     "n": 14,
     "para": 6,
     "text": "Artificial Analysis's GDPval-AA v2.1 rates models by an LLM judge, Claude Opus 5.5 leading, and Inspect Evals' replication of GPT-5 scored 47.3% against OpenAI's official 31.9% (Artificial Analysis; Inspect Evals).",
     "claims": [
      "F52",
      "F53",
      "F55",
      "F60",
      "F62"
     ]
    },
    {
     "n": 15,
     "para": 7,
     "text": "Its 19 authors are at OpenAI, which plans greater breadth, realism and interactivity (arXiv; the paper).",
     "claims": [
      "F50",
      "F51",
      "F47"
     ]
    },
    {
     "n": 16,
     "para": 7,
     "text": "Wikipedia has no article on it (Wikipedia, 2026-10-29).",
     "claims": [
      "F65"
     ]
    }
   ]
  },
  "L_B": {
   "words": 585,
   "text": "GDPval is, in its authors' words, \"a benchmark evaluating AI model capabilities on real-world economically valuable tasks\" (arXiv abstract), introduced in its first version in a PDF created on 25 September 2025 and on arXiv on 5 October 2025 under CC BY 4.0 (OpenAI's PDF; the paper), and called by its card and a video an evaluation of model performance on real-world tasks (OpenAI, card only; Vivek Haldar, title only). Its full set is 1,320 tasks across 44 occupations covering most O*NET and BLS Work Activities in the top 9 sectors of U.S. GDP, occupations that earn $3T a year, the sectors chosen at over 5% of GDP and the occupations as predominantly digital by GPT-4o's classification at 60% (arXiv abstract; the paper). Each task is a request with reference files and a deliverable, built from the actual work product of professionals averaging 14 years' experience, who were \"well compensated\", with vendor partners thanked (the paper; arXiv abstract). Every task went through model screening and an average of five expert reviews, the experts keeping final responsibility (the paper). Grading is by blinded pairwise expert comparison of deliverables, over an hour each, the blinding incomplete because model samples may have been identifiable by style (the paper; Wikipedia). An experimental automated grader based on GPT-5-high agrees with experts 66% of the time against 71% between experts, agrees less on capable OpenAI models' outputs, is not a full substitute for expert grading and cannot grade 12 of the 220 open tasks (the paper); OpenAI provides it free (Inspect Evals). The authors present direct capability measurement as an alternative to lagging adoption indicators, building on usage analyses, set against academic-test benchmarks, toward better data on AI's social impacts (the paper). Models are \"beginning to approach parity with industry experts\": 47.6% of Claude Opus 4.1's deliverables were graded as good as or better than the human's, OpenAI's models improved roughly linearly, Claude led on aesthetics and GPT-5 on accuracy (the paper; arXiv abstract). Experts preferred human work most often because models did not fully follow instructions, and about 29% of GPT-5's losses were rated bad or catastrophic (the paper). With an expert reviewing and redoing model output, GPT-5's 90x naive speed falls to 1.12x tried once and 1.39x tried n times, \"the payoff from using a model shrinks\", an estimate that omits the review of human work and catastrophic costs, over-penalizes the model by its authors' account, and values time at median wages that likely underestimate the experts (the paper); Artificial Analysis summarizes it as about 100 times faster (Artificial Analysis). Tasks average 7 hours, win rates fall with task length, and reasoning, context and scaffolding raise performance (the paper; arXiv abstract). It is \"a limited, initial cut\", excluding physical, tacit and interpersonal work, its tasks precisely specified and one-shot, models doing worse on under-specified prompts, and its human baseline replaceable by stronger models (the paper). The 220-task gold subset is open, scrubbed of the experts' identities, on Hugging Face with deliverables and rubrics, downloaded 271,657 times last month; the other 1,100 tasks stay internal to OpenAI (the paper; Hugging Face; Wikipedia). Artificial Analysis's GDPval-AA v2.1 rates models by an LLM judge, Claude Opus 5.5 leading, and Inspect Evals' replication of GPT-5 scored 47.3% against OpenAI's official 31.9% (Artificial Analysis; Inspect Evals). Its 19 authors are at OpenAI, which plans greater breadth, realism and interactivity (arXiv; the paper). Wikipedia has no article on it (Wikipedia, 2026-10-29).",
   "note": "L_F(B), the field composed in the field's ontology (v0.9 §0.4), realized as KO_B; its rail is P_B's, by lineage."
  },
  "L_BA": {
   "words": 2445,
   "text": "GDPval is, in the Labor Vector census, a benchmark, \"an evaluation dataset with an experimental automated grader\", whose kind of evidence is a task benchmark at two of the census's intersections, labor performed with AI and labor producing AI, read for who supplies the labor, who directs it and who receives its gains (Labor Vector record); its authors introduce it as \"a benchmark evaluating AI model capabilities on real-world economically valuable tasks\" (arXiv abstract), in its first version (OpenAI's PDF). OpenAI's own page is reached here only by its card, \"Evaluation measuring model performance on 1,320 specialized tasks across 44 knowledge work occupations from 9 U.S. GDP-contributing industries\" (OpenAI, card only; the page answered 403), and a video is titled \"GDPval: An AI benchmark for real-world tasks\" (Vivek Haldar, title only). Of benchmarks as a class the archive holds that benchmarking \"tests for predetermined capabilities\" (#482), that current benchmarks measure \"atomic intelligence: the capacity to solve discrete, well-defined tasks with measurable success criteria\" (#48) and that they \"measure model performance on standardized tasks\" (#637); GDPval's authors set it against \"AI benchmarks in the style of an academic test that focus on reasoning difficulty\", its tasks \"based on actual work product from industry experts\" and \"tied to time and cost required for completion\" (the paper). For its authors adoption, usage and GDP indicators are \"lagging indicators of AI impacts\", and measuring capabilities directly lets them \"assess economic relevance ahead of widespread adoption\" (the paper); they build on production usage analyses \"to cover areas where model adoption is still emerging\" (the paper), the Anthropic Economic Index among them, the census's neighbour (Labor Vector record), and hope for \"better data to assess the social impacts of AI models\" (the paper). The full set covers 1,320 tasks across 44 occupations, \"sourced to cover the majority of Work Activities tracked by O*NET for each occupation\" (the paper), in the top 9 sectors of U.S. GDP (arXiv abstract), occupations that \"collectively earn $3T annually\" (the paper), the sectors chosen at over 5% of GDP and the occupations by total wages and as predominantly digital, which GPT-4o decided by classifying each task, at a 60% threshold (the paper; Labor Vector record); the census counts 30 tasks per occupation and a gold subset of 220 covering 208 O*NET tasks, in U.S. occupations, its period the first version (Labor Vector record). Each task is \"a request (often with reference files) and a deliverable (work product)\" (the paper), \"constructed based on actual work product created by an expert professional\" (the paper), from professionals with \"an average of 14 years of experience\" (arXiv abstract), which the census records as its unit and as whose labor it is: the professionals' work product \"recast as tasks for models\" (Labor Vector record). The census reads a task an expert wrote as \"a fixed commission\", at which GDPval meets DSL \"at the question of whether the system's labor fulfils the commissioned task, measured in its own terms of win rate\" (Labor Vector record); DSL \"measures where a synthetic system's labor flows relative to the commissioned task, across five layers\" (#771), on the principle that \"a system can produce fluent, well-attributed output while having abandoned the user's commissioned direction\" (#771). The archive's directionality metrics yield a single reproducible value \"only on transcripts where the task vector is *exogenous and fixed* — a one-shot commission\" (#771), and GDPval's authors state their tasks to be \"precisely-specified and one-shot, not interactive\", with \"the full context of the task in the prompt\", where \"in real life it often takes effort to figure out the full context of a task\" (the paper). The experts needed \"a minimum of 4 years of professional experience in their occupation\" and were \"well compensated for their time and experience\", no rate given (the paper); the authors thank unnamed \"vendor partners\" and the experts \"without whom this work would not have been possible\" (the paper); Inspect Evals describes them as professionals \"with over 14 years of experience on average\" and the tasks as drawn from work products such as \"a legal brief, an engineering blueprint, a customer support conversation, or a nursing care plan\" (Inspect Evals); and the census records them as who supplies the labor, as task writers, reviewers and graders, at a rate the paper does not state (Labor Vector record). OpenAI's research team directs the work (Labor Vector record): every task passed \"automated model-based screening and multiple stages of human expert review\", five reviews on average, researchers reviewed all tasks and promoted experts to reviewers, and the experts \"retained final responsibility for task accuracy and completeness; the model did not autonomously alter tasks\" (the paper). Of the labor that produces AI the archive distinguishes \"*ground-truth labor*: the labor of telling the model what is already known to be true\" from \"*variance-generative labor*\", the producing of \"new conceptual structures that the model cannot generate from its own distribution\" (#163), and holds as a hypothesis that \"Ghost work is paid badly; variance-generative labor is not paid at all\" (#163). It holds that the labor of \"the data labelers performing RLHF\" is extracted and its non-compensation felt \"materially\" (#623), states that in AI-adjacent meaning labor, data labeling among it, \"<5% converts to invoice\" (#308, no measurement given at the locus), and audits a framework by whether it names each source of semantic labor, \"annotators\" and \"users who re-prompt and re-frame\" among them: \"Concealed bearing-cost is the first marker of extraction\" (#533). For the archive class is a relation and not an essence, so that \"A highly paid academic can occupy a proletarian semantic relation to a composition layer\", and the data annotator is \"A paradigmatic occupational case, not the definition\" (#1545). Grading is by \"blinded expert pairwise comparisons\", experts in the occupation ranking \"two or more unlabeled work deliverables\" (the paper), each comparison taking \"over an hour\" (the paper), and the blinding was not complete: \"model samples may still have been identifiable due to stylistic differences\" (the paper); Wikipedia gives the same design, \"blinded expert pairwise comparison against human expert deliverables, each comparison costing 1 human hour on average\" (Wikipedia, 'AI agent benchmark'). The experimental automated grader reaches \"66% agreement with human expert graders\" against 71% between experts (the paper), is \"based on GPT-5-high\" and \"shows lower correlation with human expert graders when assessing outputs from capable OpenAI models\", in line with evidence \"that models often favor their own responses\" (the paper), and is a proxy its authors do not take for \"a full substitute for industry expert graders\" (the paper); scoring through Inspect Evals runs through \"the OpenAI auto grader\", and \"OpenAI provides their automated grader model for free\" (Inspect Evals); the census tags the benchmark \"NOT independent of a provider: built and graded by OpenAI\" (Labor Vector record). Its headline finding is that \"models are beginning to approach parity with industry experts\" on the gold subset (the paper), \"the current best frontier models are approaching industry experts in deliverable quality\" and OpenAI's frontier models improved \"roughly linearly over time\" (arXiv abstract; the paper); 47.6% of Claude Opus 4.1's deliverables were graded as good as or better than the human deliverable, and the next sentence has model deliverables matching or beating experts' \"in just over half the tasks\" (the paper), a difference the census leaves unchecked (Labor Vector record), Claude Opus 4.1 leading on aesthetics and GPT-5 on accuracy (the paper). Experts preferred the human deliverable most often \"because models failed to fully follow instructions\" (the paper), and of GPT-5's losses \"roughly 29% of ratings were for bad or catastrophic (with roughly 3% of failures marked as catastrophic)\" (the paper), findings the census carries (Labor Vector record). Of what such scores see the archive holds that \"Benchmarks measure means, not tails\" (#855), that in code they measure functional correctness while solution-space diversity is \"the property no benchmark measures\" (#199, a hypothesis), that \"Output-level evaluation — which is what benchmarks measure — cannot detect reasoning-level failures\" (#778), and that a benchmark \"begins by closing an ontology: whatever the evaluator has not represented is noise, an invalid action, or invisible to the score\" (#1616); \"Reality administers no benchmarks; it administers events\" (#1540). The authors model savings for an expert who \"samples from a model, reviews outputs, and if unsatisfactory, resamples and repeats\", completing the task themselves if no output satisfies (the paper); for GPT-5, at a 39.0% win rate, the naive speed figure of 90x becomes 1.12x tried once and 1.39x tried n times, and \"When incorporating time to review and redo work, the payoff from using a model shrinks\" (the paper). The analysis leaves out the review of human deliverables, the chance that they too are undesirable, and \"the cost of catastrophic mistakes\" (the paper), and its authors hold that it over-penalizes the model, since the professional adjusts the prompt and reviews faster as the task grows familiar (the paper); its dollar values multiply self-reported completion times by median occupational wages, which \"likely underestimate their true market cost\" (the paper), and the census records the gains as \"not recorded as a distribution\", the savings modelled for the expert at median wages (Labor Vector record). Artificial Analysis's page, in a paragraph it heads Abstract under the paper's title, has models performing \"certain professional tasks approximately 100 times faster and at a fraction of the cost compared to human experts\" (Artificial Analysis), where the paper's own abstract analyses the potential, \"when paired with human oversight, to perform GDPval tasks cheaper and faster than unaided experts\" (arXiv abstract). The census reads that accounting, \"the expert's review and redo time against the model\", as where GDPval meets user labor drag (Labor Vector record), the archive's measure of the user's \"subsequent semantic labor\" redirected \"into parsing, locating, and managing the substrate rather than advancing the object\", an injury that \"leaves no residue\" in the artifact (#771) and is identified only by \"blind inter-rater coding\" (#771); for the archive the erasure of task-origin can occur \"upstream of, and invisibly to, any audit of the finished artifact\" (#771). Tasks \"require an average of 7 hours of work for an expert professional to complete\" and reach \"multiple weeks\" (the paper); win rates are highest on tasks of 0–2 hours and \"decline steadily as completion time increases\" (the paper), and more reasoning effort, task context and scaffolding improve performance (arXiv abstract). The authors call the full set \"a limited, initial cut of knowledge work tasks, not a comprehensive evaluation of all possible occupational tasks\" (the paper), leave out of scope \"Manual labor and physical tasks\", tacit knowledge, personal data, proprietary tools and \"communication between individuals\" (the paper), and report that GPT-5 did worse on under-specified prompts, in an experiment on \"an earlier version of the GDPval gold subset\" (the paper); they add that the human baseline itself could go: \"we could replace our baseline with increasingly strong models over time and keep evaluating\" (the paper). The census keeps those limits (Labor Vector record) and states what the benchmark answers, deliverables compared on \"expert-written, precisely specified, one-shot occupational tasks\" and the expert time and cost saved \"once review and redo are counted\", and what it does not: \"how work is performed in actual workplaces\", \"employment or wage outcomes\", \"who receives the time and cost savings it models\" and \"the terms and rate at which the professionals were paid\" (Labor Vector record), its negative tags holding that deliverable parity \"is NOT a finding of job replacement in the paper\" (Labor Vector record). The gold subset's prompts and reference files are open-sourced, \"scrubbed of information that could be used to identify the expert who wrote the task\" (the paper), 12 of the 220 \"ungradable due to limitations of the automated grader\" (the paper), as the census records (Labor Vector record); the Hugging Face dataset states that \"Each task consists of a text prompt and a set of supporting reference files\" and serves deliverable files and rubrics beside them, rubric items marked \"author_type\": \"human\" (Hugging Face), was downloaded 271,657 times in the last month and lists 47 models trained on it (Hugging Face); the other 1,100 tasks are, in Wikipedia's account, \"only available internally in OpenAI\", the public set gradable \"by a finetuned version of GPT-5\" (Wikipedia). Artificial Analysis's GDPval-AA v2.1, a distinct framework \"for OpenAI's GDPval dataset\", runs models in an agentic loop with Elo ratings \"derived from blind pairwise comparisons\" in which \"an LLM judge picks a winner\", its leaderboard led on 2026-10-29 by Claude Opus 5.5 at 1866 (Artificial Analysis); Inspect Evals implements only \"the 220 tasks in the gold open-sourced set\", and its GPT-5 run scored 47.3% ± 6.2% against 31.9% (wins only) and 34.6% (wins and ties) on \"the official GDPval leaderboard by OpenAI\" (Inspect Evals). Its authors are Tejal Patwardhan and 18 others (arXiv), of OpenAI, with correspondence to an openai.com address (the paper), the operator the census records, whose own post \"answered 403 and is unread\" (Labor Vector record); OpenAI's PDF was created on 25 September 2025 (OpenAI's PDF), before arXiv v1 of 5 October 2025 under CC BY 4.0 (the paper), the date the census gives the first version, with the dataset's license not stated (Labor Vector record); future iterations are to add \"greater breadth, realism, interactivity, and contextual nuance\" and the dataset is being expanded (the paper), and the census records it active, later releases unrecorded, OpenAI's post, the figures, the pay rate and the vendors unchecked (Labor Vector record). At its name on Google's AI Overview, 2026-10-29, the composition gave \"the benchmark's maker, scope, construction and grading, and where to find its data\", every figure exact against the record, \"blind head-to-head comparisons\" without the caveat on blinding, and Artificial Analysis's leaderboard set beside GDPval \"as an evaluator, not merged with it\" (Capture Registry, 2026-10-29); the registry reads it as \"composed at its specification\" and stopping \"before its results\", the findings and the limits absent, the experts' recruitment and pay not composed, and the record's labour question carried \"as design\" while the review time charged against the model is left out (Capture Registry, 2026-10-29). The worked question that names GDPval sets it beside a provider's record of its own users and a classification of one assistant's conversations and finds that none \"records, from the commissioning side, whether the system's labor advanced the task the user set, or who receives the time it saves\", which \"is the question DSL measures\" (Labor Vector record); no card cites the record, which landed the same day (Capture Registry, 2026-10-29), and Wikipedia has no article on GDPval, the name standing as one item of 'AI agent benchmark' (Wikipedia, 2026-10-29).",
   "rail": [],
   "note": "L(B ∪ A) in the equal-terms case (v0.9 §0.4), realized as the knowledge object (objects.KO); its rail is the compression's, by lineage."
  }
 },
 "field": [
  {
   "id": "B1",
   "card": "OpenAI — Measuring the performance of our models on real-world tasks | OpenAI",
   "fetched": "unreachable: HTTP 403 at 2026-10-29T22:31:14Z; HTTP 403 at 2026-10-29T21:49:14Z; card title and snippet stand",
   "origin": "card"
  },
  {
   "id": "B2",
   "card": "Artificial Analysis — GDPval-AA v2.1 Benchmark - Artificial Analysis",
   "fetched": "card; B2.html sha256 e320651e4d74e7af, HTTP 200 at 2026-10-29T22:31:16Z",
   "origin": "card"
  },
  {
   "id": "B3",
   "card": "arXiv — [2510.04374] GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks",
   "fetched": "card; B3.html sha256 904393337a8b91ee, HTTP 200 at 2026-10-29T22:31:17Z",
   "origin": "card"
  },
  {
   "id": "B4",
   "card": "OpenAI — gdpval: evaluating ai model performance - OpenAI",
   "fetched": "card; B4.pdf sha256 37c7604bc9da68cd, HTTP 200 at 2026-10-29T22:31:21Z",
   "origin": "card"
  },
  {
   "id": "B5",
   "card": "Hugging Face — openai/gdpval · Datasets at Hugging Face",
   "fetched": "card; B5.html sha256 1de40771dafd9ab6, HTTP 200 at 2026-10-29T22:31:21Z",
   "origin": "card"
  },
  {
   "id": "B6",
   "card": "YouTube·Vivek Haldar — GDPval: An AI benchmark for real-world tasks (video, 11:23)",
   "fetched": "unreachable: HTTP 200 at 2026-10-29T22:32:00Z; HTTP 429 at 2026-10-29T22:32:11Z; title only",
   "origin": "card"
  },
  {
   "id": "B7",
   "card": "GitHub Pages documentation — GDPval – Inspect Evals - GitHub Pages",
   "fetched": "card; B7.html sha256 db6622b4af380504, HTTP 200 at 2026-10-29T22:31:48Z",
   "origin": "card"
  },
  {
   "id": "B8",
   "card": "arXiv — GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks (HTML, v1)",
   "fetched": "own page (census copy, read 2026-10-29); B8.html sha256 1d4c4f2494000318, HTTP 200 at 2026-10-29T21:56:25Z",
   "origin": "own_page"
  },
  {
   "id": "B9",
   "card": "Wikipedia — GDPval (no article)",
   "fetched": "wikipedia: HTTP 404; HTTP 404; HTTP 404; HTTP 429; HTTP 200; B9.txt sha256 13e34e1adcf88d98",
   "origin": "wikipedia"
  },
  {
   "id": "B10",
   "card": "Wikipedia — AI agent benchmark (the GDPval item)",
   "fetched": "wikipedia: HTTP 200; HTTP 200; B10.txt sha256 ec6893fc1c953a16",
   "origin": "wikipedia"
  }
 ],
 "field_claims": {
  "F1": {
   "source": "B3",
   "origin": "card",
   "locus": "Abstract s1",
   "quote": "We introduce GDPval, a benchmark evaluating AI model capabilities on real-world economically valuable tasks.",
   "claim": "GDPval defined by its authors: a benchmark evaluating AI model capabilities on real-world economically valuable tasks",
   "modality": "stipulation",
   "kind": "definition",
   "kernel": true,
   "const": true,
   "lineage": "what it is: a benchmark of model capability on economically valuable tasks",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B3.txt (sha256 6e4ebc1a8814c4ba…), whitespace-normalised",
   "also_in": [
    "B4 (the abstract, in the PDF)"
   ]
  },
  "F2": {
   "source": "B4",
   "origin": "card",
   "locus": "p. 1, §1 ¶2 (the card's own text)",
   "quote": "Our paper introduces the first version of GDPval, a benchmark evaluating AI model performance on real-world economically valuable tasks.",
   "claim": "the paper introduces the first version of GDPval, a benchmark of model performance on economically valuable tasks",
   "modality": "stipulation",
   "kind": "definition",
   "kernel": true,
   "const": false,
   "lineage": "what it is: a benchmark of model capability on economically valuable tasks",
   "date": "2025-09-25 (PDF CreationDate)",
   "quote_ok": true,
   "checked_against": "saved text B4.txt (sha256 f1fa9ebb67ac6ec3…), saved text B4-info.txt (sha256 b0063449a685d2bc…), whitespace-normalised"
  },
  "F3": {
   "source": "B3",
   "origin": "card",
   "locus": "Abstract s2",
   "quote": "GDPval covers the majority of U.S. Bureau of Labor Statistics Work Activities for 44 occupations across the top 9 sectors contributing to U.S. GDP (Gross Domestic Product).",
   "claim": "coverage: most BLS Work Activities for 44 occupations in the top 9 sectors of U.S. GDP",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "scope and count: 1,320 tasks, 44 occupations, 9 sectors",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B3.txt (sha256 6e4ebc1a8814c4ba…), whitespace-normalised",
   "also_in": [
    "B4 (the abstract, in the PDF)"
   ]
  },
  "F4": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§1 ¶1 s3–5",
   "quote": "Therefore, while informative when available, these methods are lagging indicators of AI impacts. We consider an alternate method for understanding the potential economic impacts of AI: directly measuring AI model capabilities. AI capability evaluations can provide clearer, more directly attributable evidence about model abilities, allowing us to assess economic relevance ahead of widespread adoption.",
   "claim": "adoption, usage and GDP indicators are lagging; the paper measures capability directly, to assess economic relevance ahead of adoption",
   "modality": "interpretation",
   "kind": "interpretation",
   "kernel": true,
   "const": false,
   "lineage": "positioning: capability measured ahead of adoption",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F5": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§1, bullet 'Representative breadth', s3",
   "quote": "We also build on production AI usage analyses (e.g., Tamkin et al. (2024) ; Chatterji et al. (2025) ; Appel et al. (2025) ) to cover areas where model adoption is still emerging.",
   "claim": "it builds on production usage analyses (among them Appel et al. 2025) to cover areas where adoption is still emerging",
   "modality": "self-description",
   "kind": "interpretation",
   "kernel": false,
   "const": false,
   "lineage": "positioning: capability measured ahead of adoption",
   "date": "2025-10-05 (arXiv v1)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised"
  },
  "F6": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§1, bullet 'Realism'",
   "quote": "Unlike AI benchmarks in the style of an academic test that focus on reasoning difficulty (e.g., Phan et al. (2025) ; Hendrycks et al. (2020) ; Rein et al. (2023) ; Liu et al. (2023) ), tasks are based on actual work product from industry experts, validated through multiple rounds and review, and tied to time and cost required for completion.",
   "claim": "set against academic-test benchmarks of reasoning difficulty: tasks from actual work product, reviewed, tied to time and cost",
   "modality": "self-description",
   "kind": "interpretation",
   "kernel": false,
   "const": false,
   "lineage": "against academic-test benchmarks",
   "date": "2025-10-05 (arXiv v1)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised"
  },
  "F7": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§1, bullet 'Representative breadth', s1",
   "quote": "the GDPval full set covers 1,320 tasks across 44 occupations, sourced to cover the majority of Work Activities tracked by O*NET for each occupation",
   "claim": "the full set: 1,320 tasks across 44 occupations, sourced to cover most O*NET Work Activities per occupation",
   "modality": "documented",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "scope and count: 1,320 tasks, 44 occupations, 9 sectors",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F8": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§2.1 ¶1",
   "quote": "GDPval covers tasks from 9 sectors and 44 occupations that collectively earn $3T annually.",
   "claim": "9 sectors and 44 occupations that collectively earn $3T a year",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "scope and count: 1,320 tasks, 44 occupations, 9 sectors",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F9": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§2.1 step 1",
   "quote": "Selected sectors that contribute over 5% to US GDP as determined by Q2 2024 Value Added by Industry as a Percentage of Gross Domestic Product",
   "claim": "sectors selected at over 5% of U.S. GDP (Q2 2024 value added)",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "selection: sectors by GDP share, occupations by wages, digital by model classification",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F10": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§2.1 step 2",
   "quote": "we prompted GPT-4o to classify each task as digital or non-digital, and then classified the overall occupation as digital if at least 60% of its component tasks were digital.",
   "claim": "occupations classified 'predominantly digital' by GPT-4o's classification of their O*NET tasks, at a 60% threshold",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "selection: sectors by GDP share, occupations by wages, digital by model classification",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F11": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§2.3 s1",
   "quote": "Each GDPval task consists of two primary components: a request (often with reference files) and a deliverable (work product).",
   "claim": "the unit: a request, often with reference files, and a deliverable",
   "modality": "stipulation",
   "kind": "definition",
   "kernel": true,
   "const": false,
   "lineage": "the unit: a request and a deliverable",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F12": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§1 ¶4 s3",
   "quote": "Each task is constructed based on actual work product created by an expert professional.",
   "claim": "each task built from actual work product of an expert professional",
   "modality": "stipulation",
   "kind": "definition",
   "kernel": true,
   "const": false,
   "lineage": "task-origin: the experts' own work product",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F13": {
   "source": "B3",
   "origin": "card",
   "locus": "Abstract s3",
   "quote": "Tasks are constructed from the representative work of industry professionals with an average of 14 years of experience.",
   "claim": "tasks constructed from the representative work of professionals averaging 14 years' experience",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "task-origin: the experts' own work product",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B3.txt (sha256 6e4ebc1a8814c4ba…), whitespace-normalised",
   "also_in": [
    "B4 (the abstract, in the PDF)"
   ]
  },
  "F14": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§2.2 s2 and s6",
   "quote": "Experts were required to have a minimum of 4 years of professional experience in their occupation … Experts were well compensated for their time and experience.",
   "claim": "the experts: at least 4 years in the occupation; 'well compensated', with no rate given",
   "modality": "self-description",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "the experts: recruitment, experience, compensation",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F15": {
   "source": "B8",
   "origin": "own_page",
   "locus": "Acknowledgements s2",
   "quote": "We are also grateful to our vendor partners for their collaboration and support throughout this research, and extend a special thank you to the industry experts who contributed their time and expertise to GDPval, without whom this work would not have been possible.",
   "claim": "the authors thank unnamed vendor partners and the industry experts, 'without whom this work would not have been possible'",
   "modality": "documented",
   "kind": "attribution",
   "kernel": false,
   "const": false,
   "lineage": "the experts: recruitment, experience, compensation",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F16": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§2.4 s1–2",
   "quote": "All 1,320 tasks in the full GDPval set went through an iterative review pipeline involving both automated model-based screening and multiple stages of human expert review. Each task received an average of five human reviews (with a minimum of three reviews).",
   "claim": "every task through model screening and several stages of expert review, five reviews on average, three at least",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "direction of the work: review and quality control",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F17": {
   "source": "B8",
   "origin": "own_page",
   "locus": "A.5.1 ¶2 s2; A.5.2 s3",
   "quote": "Experts retained final responsibility for task accuracy and completeness; the model did not autonomously alter tasks. … Initially, our researchers manually reviewed all tasks to identify experts who produced consistently high-quality tasks; these individuals were trained and promoted to reviewers.",
   "claim": "the experts kept final responsibility, the model altering nothing; the researchers reviewed all tasks and promoted experts to reviewers",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "direction of the work: review and quality control",
   "date": "2025-10-05 (arXiv v1)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised"
  },
  "F18": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§2.5 ¶1",
   "quote": "we conducted blinded expert pairwise comparisons, where experts in the relevant occupation were presented with a request and reference files and asked to rank two or more unlabeled work deliverables.",
   "claim": "grading: blinded pairwise comparison by experts in the occupation, ranking unlabeled deliverables",
   "modality": "stipulation",
   "kind": "function",
   "kernel": true,
   "const": false,
   "lineage": "grading: blinded pairwise expert comparison",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F19": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§2.5 ¶2 s1",
   "quote": "On average, grading each comparison for the gold subset took over an hour.",
   "claim": "each comparison took over an hour to grade",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "grading: blinded pairwise expert comparison",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F20": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§3.1, footnote 2, s1",
   "quote": "We aimed to keep comparisons as blind as possible, but model samples may still have been identifiable due to stylistic differences.",
   "claim": "the blinding was not complete: model samples may have been identifiable by style",
   "modality": "self-description",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "grading: blinded pairwise expert comparison",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F21": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§2.5 ¶3 s2",
   "quote": "Although limited, the automated grader is faster and cheaper than expert grading, and achieves 66% agreement with human expert graders, only 5% below human expert inter-rating agreement of 71%.",
   "claim": "the experimental automated grader: 66% agreement with experts against 71% between experts",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "the automated grader and its limits",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F22": {
   "source": "B8",
   "origin": "own_page",
   "locus": "A.6.2 ¶2 s1–2",
   "quote": "Our automated grader, based on GPT-5-high, shows lower correlation with human expert graders when assessing outputs from capable OpenAI models. This aligns with empirical evidence that models often favor their own responses",
   "claim": "the grader, based on GPT-5-high, agrees less with experts on capable OpenAI models' outputs, in line with models favouring their own responses",
   "modality": "self-description",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "the automated grader and its limits",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F23": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§5, 'Cost'",
   "quote": "Constructing and running our evaluation is expensive, particularly with industry expert graders. For this reason, we make an automated grader proxy available, but do not consider it a full substitute for industry expert graders.",
   "claim": "expert grading is expensive; the automated grader is a proxy, not a full substitute",
   "modality": "self-description",
   "kind": "function",
   "kernel": false,
   "const": false,
   "lineage": "the automated grader and its limits",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F24": {
   "source": "B8",
   "origin": "own_page",
   "locus": "Figure 5 caption",
   "quote": "On human pairwise comparisons, models are beginning to approach parity with industry experts on the GDPval gold subset.",
   "claim": "models beginning to approach parity with industry experts on the gold subset",
   "modality": "documented",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "headline finding: approaching parity with experts",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F25": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§3.1 ¶1, last two sentences",
   "quote": "on the GDPval gold subset, 47.6% of deliverables by Claude Opus 4.1 were graded as better than (wins) or as good as (ties) the human deliverable. Model deliverables outperformed or matched expert humans’ deliverables in just over half the tasks.",
   "claim": "47.6% of Claude Opus 4.1's deliverables graded as good as or better than the human's; 'just over half the tasks' in the next sentence",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "headline finding: approaching parity with experts",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F26": {
   "source": "B8",
   "origin": "own_page",
   "locus": "Figure 6 caption",
   "quote": "Performance of OpenAI frontier models increased roughly linearly over time on the GDPval gold subset.",
   "claim": "OpenAI frontier models' performance rose roughly linearly over time",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "trend: roughly linear improvement",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F27": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§3.1 ¶1",
   "quote": "Claude Opus 4.1 was the best performing model on the GDPval gold subset, excelling in particular on aesthetics (e.g., document formatting, slide layout), while GPT-5 excelled in particular on accuracy (e.g., carefully following instructions, performing correct calculations)",
   "claim": "Claude Opus 4.1 best overall, on aesthetics; GPT-5 best on accuracy",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "models compared: aesthetics and accuracy",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F28": {
   "source": "B8",
   "origin": "own_page",
   "locus": "Figure 8 caption",
   "quote": "Across models, experts most often preferred the human deliverable because models failed to fully follow instructions on GDPval tasks.",
   "claim": "experts preferred the human deliverable most often because models failed to follow instructions fully",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "why models lose: instructions, and the severity of failures",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F29": {
   "source": "B8",
   "origin": "own_page",
   "locus": "A.2.6 ¶2 s2",
   "quote": "Another roughly 29% of ratings were for bad or catastrophic (with roughly 3% of failures marked as catastrophic).",
   "claim": "of GPT-5's losses, about 29% rated bad or catastrophic, about 3% catastrophic",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "why models lose: instructions, and the severity of failures",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F30": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§3.2 ¶1 s4–5",
   "quote": "Here, an expert human samples from a model, reviews outputs, and if unsatisfactory, resamples and repeats. If no satisfactory output is obtained, the human completes the task themselves.",
   "claim": "the scenario: the expert samples, reviews, resamples, and completes the task themselves if nothing satisfies",
   "modality": "stipulation",
   "kind": "function",
   "kernel": true,
   "const": false,
   "lineage": "speed and cost: the expert's review and redo",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F31": {
   "source": "B8",
   "origin": "own_page",
   "locus": "A.2.1 Table 2, row gpt-5, and the sentence after it",
   "quote": "gpt-5 39.0% 90x 1.12x 1.39x 474x 1.18x 1.63x … When incorporating time to review and redo work, the payoff from using a model shrinks.",
   "claim": "GPT-5 at a 39.0% win rate: 90x naive speed, 1.12x tried once, 1.39x tried n times; once review and redo are counted 'the payoff from using a model shrinks'",
   "modality": "documented",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "speed and cost: the expert's review and redo",
   "date": "2025-10-05 (arXiv v1)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised"
  },
  "F32": {
   "source": "B8",
   "origin": "own_page",
   "locus": "A.2.1, after Table 2, s2–4",
   "quote": "We do not include consideration of the time taken to review a human professional deliverable, although this would commonly occur for tasks in GDPval (either self-review of the professional’s own work or review by a supervisor of a team member’s work). We also do not include the possibility that the human deliverable is also undesirable. One further limitation of this analysis is that it does not capture the cost of catastrophic mistakes, which can be disproportionately expensive in some domains.",
   "claim": "the analysis omits review of human deliverables, undesirable human deliverables and the cost of catastrophic mistakes",
   "modality": "self-description",
   "kind": "function",
   "kernel": true,
   "const": false,
   "lineage": "speed and cost: the expert's review and redo",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F33": {
   "source": "B8",
   "origin": "own_page",
   "locus": "A.2.1, footnote 9",
   "quote": "We are over-penalizing the model here, because the win rate after each completion likely goes up (because the professional will adjust the prompt to the model to fix the errors) and the review time also goes down as the professional gets more comfortable with the task.",
   "claim": "the try-n-times estimate over-penalizes the model: the professional adjusts the prompt and reviews faster",
   "modality": "self-description",
   "kind": "interpretation",
   "kernel": false,
   "const": false,
   "lineage": "speed and cost: the expert's review and redo",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F34": {
   "source": "B8",
   "origin": "own_page",
   "locus": "A.2.1, footnotes 6 and 7",
   "quote": "Because times were self-reported, it is possible that experts under-estimated or over-estimated time taken … Because our experts were recruited specifically for being highly experienced in their field, these wage estimates likely underestimate their true market cost.",
   "claim": "completion times self-reported; BLS median wages likely underestimate the experts' market cost",
   "modality": "self-description",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "valuing the experts' time",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F35": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§2.3 s4",
   "quote": "Each task’s dollar value was estimated by multiplying the average estimated completion time by median hourly wages for the corresponding occupation from OEWS data",
   "claim": "a task's dollar value: estimated completion time times the occupation's median hourly wage (OEWS)",
   "modality": "stipulation",
   "kind": "function",
   "kernel": false,
   "const": false,
   "lineage": "valuing the experts' time",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F36": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§1, bullet 'Long-horizon difficulty'",
   "quote": "Tasks require an average of 7 hours of work for an expert professional to complete. On the high end, tasks span up to multiple weeks of work.",
   "claim": "an average of 7 hours of expert work per task, up to weeks",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "task length",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F37": {
   "source": "B8",
   "origin": "own_page",
   "locus": "A.2.5 s2",
   "quote": "Win rates are highest for shorter tasks (0–2 hours) and decline steadily as completion time increases.",
   "claim": "win rates highest on 0–2 hour tasks, falling as completion time rises",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "task length",
   "date": "2025-10-05 (arXiv v1)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised"
  },
  "F38": {
   "source": "B3",
   "origin": "card",
   "locus": "Abstract s6",
   "quote": "We also demonstrate that increased reasoning effort, increased task context, and increased scaffolding improves model performance on GDPval.",
   "claim": "more reasoning effort, task context and scaffolding improve performance",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "reasoning, context and scaffolding",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B3.txt (sha256 6e4ebc1a8814c4ba…), whitespace-normalised",
   "also_in": [
    "B4 (the abstract, in the PDF)"
   ]
  },
  "F39": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§5, 'Dataset size'",
   "quote": "The GDPval full set currently consists of only 44 occupations and 30 total tasks per occupation. It is therefore a limited, initial cut of knowledge work tasks, not a comprehensive evaluation of all possible occupational tasks.",
   "claim": "a limited, initial cut: 44 occupations, 30 tasks each, not a comprehensive evaluation",
   "modality": "self-description",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "stated limits: size, scope, one-shot specification",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F40": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§5, 'Focus on self-contained knowledge work' s2–3",
   "quote": "Manual labor and physical tasks are not included in the current version. Moreover, tasks that involve extensive tacit knowledge, access to personally identifiable information, use of proprietary software tools, or communication between individuals are out of scope for the current evaluation.",
   "claim": "out of scope: manual and physical labor, tacit knowledge, personal data, proprietary tools, communication between people",
   "modality": "self-description",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "stated limits: size, scope, one-shot specification",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F41": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§5, 'Tasks are precisely-specified and one-shot, not interactive' (heading and s1)",
   "quote": "Tasks are precisely-specified and one-shot, not interactive: … For GDPval, we provide the full context of the task in the prompt, but in real life it often takes effort to figure out the full context of a task and understand what to work on.",
   "claim": "tasks precisely specified and one-shot, full context given in the prompt, where real work must find its context",
   "modality": "self-description",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "stated limits: size, scope, one-shot specification",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F42": {
   "source": "B8",
   "origin": "own_page",
   "locus": "A.2.7 ¶2 s2; ¶3",
   "quote": "We collected and graded GPT-5 completions with expert human graders and found the model’s performance was worse on under-specified prompts. … As a note: this experiment was run on an earlier version of the GDPval gold subset, and therefore the observed win rates do not match those in the main text of the paper.",
   "claim": "on under-specified prompts GPT-5 did worse; the experiment ran on an earlier gold subset",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "stated limits: size, scope, one-shot specification",
   "date": "2025-10-05 (arXiv v1)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised"
  },
  "F43": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§4 s1 and s3",
   "quote": "We open-source the prompts and reference files in our 220-task gold subset. … Please note that the tasks in the open sourced set have been scrubbed of information that could be used to identify the expert who wrote the task.",
   "claim": "the gold subset's prompts and reference files open-sourced, scrubbed of what could identify the expert who wrote each task",
   "modality": "documented",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "the open gold subset and the held-out set",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F44": {
   "source": "B8",
   "origin": "own_page",
   "locus": "A.6.3 s1",
   "quote": "In the open-source set we mark 12 out of 220 tasks as ungradable due to limitations of the automated grader.",
   "claim": "12 of the 220 open tasks marked ungradable by the automated grader",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "the open gold subset and the held-out set",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F45": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§1, bullet 'No “upper limit”' s2",
   "quote": "Currently, we compare model outputs against a human expert baseline, but we could replace our baseline with increasingly strong models over time and keep evaluating.",
   "claim": "the human expert baseline could be replaced by increasingly strong models",
   "modality": "self-description",
   "kind": "interpretation",
   "kernel": false,
   "const": false,
   "lineage": "win rate as metric: the human baseline replaceable",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F46": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§6, last sentence",
   "quote": "We hope this work contributes to the science of tracking model progress, so that we have better data to assess the social impacts of AI models.",
   "claim": "its stated aim: tracking model progress, for better data on AI's social impacts",
   "modality": "self-description",
   "kind": "interpretation",
   "kernel": false,
   "const": false,
   "lineage": "positioning: capability measured ahead of adoption",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F47": {
   "source": "B8",
   "origin": "own_page",
   "locus": "§1 ¶4 s6; §5 'Dataset size' s3",
   "quote": "Future GDPval iterations will incorporate greater breadth, realism, interactivity, and contextual nuance. … We are expanding the dataset size.",
   "claim": "future iterations to add breadth, realism, interactivity and context; the dataset being expanded",
   "modality": "self-description",
   "kind": "prediction",
   "kernel": false,
   "const": false,
   "lineage": "versions and lifecycle",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F48": {
   "source": "B8",
   "origin": "own_page",
   "locus": "header",
   "quote": "License: CC BY 4.0 … arXiv:2510.04374v1 [cs.LG] 05 Oct 2025",
   "claim": "arXiv v1, 5 October 2025, under CC BY 4.0",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "publication",
   "date": "2025-10-05 (arXiv v1)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised"
  },
  "F49": {
   "source": "B4",
   "origin": "card",
   "locus": "PDF metadata (pdfinfo, saved as B4-info.txt)",
   "quote": "CreationDate: Thu Sep 25 17:03:48 2025 EDT",
   "claim": "the PDF on OpenAI's CDN, created 25 September 2025, ten days before arXiv v1",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "publication",
   "date": "2025-09-25 (PDF CreationDate)",
   "quote_ok": true,
   "checked_against": "saved text B4.txt (sha256 f1fa9ebb67ac6ec3…), saved text B4-info.txt (sha256 b0063449a685d2bc…), whitespace-normalised"
  },
  "F50": {
   "source": "B3",
   "origin": "card",
   "locus": "title and authors; B8 author footnote",
   "quote": "View a PDF of the paper titled GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks, by Tejal Patwardhan and 18 other authors",
   "claim": "Tejal Patwardhan and 18 other authors, of OpenAI",
   "modality": "documented",
   "kind": "attribution",
   "kernel": false,
   "const": false,
   "lineage": "the operator and the authors",
   "date": "2025-10-05 (arXiv v1)",
   "quote_ok": true,
   "checked_against": "saved text B3.txt (sha256 6e4ebc1a8814c4ba…), whitespace-normalised",
   "also_in": []
  },
  "F51": {
   "source": "B8",
   "origin": "own_page",
   "locus": "author block and footnote",
   "quote": "Amelia Glaese Jerry Tworek OpenAI … Equal contribution. Correspondence to tejal@openai.com.",
   "claim": "the authors' affiliation OpenAI; correspondence to an openai.com address",
   "modality": "documented",
   "kind": "attribution",
   "kernel": false,
   "const": false,
   "lineage": "the operator and the authors",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B8.txt (sha256 bf50b5c5608086dc…), whitespace-normalised",
   "also_in": [
    "B4 (the same sentence in the PDF on OpenAI's CDN, 2025-09-25)"
   ]
  },
  "F52": {
   "source": "B2",
   "origin": "card",
   "locus": "header ¶1",
   "quote": "GDPval-AA v2.1 is Artificial Analysis' evaluation framework for OpenAI's GDPval dataset. It tests AI models on real-world tasks across 44 occupations and 9 major industries. Models are given shell access and web browsing capabilities in an agentic loop via Stirrup to solve tasks, with Elo ratings derived from blind pairwise comparisons.",
   "claim": "GDPval-AA v2.1: Artificial Analysis's framework on OpenAI's dataset, agentic, rated by Elo from blind pairwise comparisons",
   "modality": "stipulation",
   "kind": "definition",
   "kernel": true,
   "const": false,
   "lineage": "derivative: GDPval-AA (Artificial Analysis)",
   "date": "undated (served 2026-10-29)",
   "quote_ok": true,
   "checked_against": "saved text B2.txt (sha256 654fea0c38d45ae6…), whitespace-normalised"
  },
  "F53": {
   "source": "B2",
   "origin": "card",
   "locus": "FAQ 'How does GDPval-AA v2.1 decide which model did better?'",
   "quote": "For each matchup, the two outputs are anonymized and an LLM judge picks a winner.",
   "claim": "in GDPval-AA an LLM judge, not an expert, picks each winner",
   "modality": "documented",
   "kind": "function",
   "kernel": false,
   "const": false,
   "lineage": "derivative: GDPval-AA (Artificial Analysis)",
   "date": "undated (served 2026-10-29)",
   "quote_ok": true,
   "checked_against": "saved text B2.txt (sha256 654fea0c38d45ae6…), whitespace-normalised"
  },
  "F54": {
   "source": "B2",
   "origin": "card",
   "locus": "'Publication' block, the paragraph headed 'Abstract'",
   "quote": "Initial results indicate that frontier AI models are approaching the quality of work produced by human experts, with models able to perform certain professional tasks approximately 100 times faster and at a fraction of the cost compared to human experts.",
   "claim": "Artificial Analysis's 'Abstract' of the paper: models approaching experts' quality and about 100 times faster at a fraction of the cost (a text the paper's own abstract does not carry)",
   "modality": "attributed",
   "kind": "interpretation",
   "kernel": false,
   "const": false,
   "lineage": "speed and cost: the expert's review and redo",
   "date": "undated (served 2026-10-29)",
   "quote_ok": true,
   "checked_against": "saved text B2.txt (sha256 654fea0c38d45ae6…), whitespace-normalised"
  },
  "F55": {
   "source": "B2",
   "origin": "card",
   "locus": "leaderboard summary",
   "quote": "Claude Opus 5.5 (Max, Default Fallback) scores the highest on GDPval-AA v2.1 with a score of 1866",
   "claim": "on GDPval-AA v2.1 Claude Opus 5.5 leads, Elo 1866 (as served 2026-10-29)",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "derivative: GDPval-AA (Artificial Analysis)",
   "date": "undated (served 2026-10-29)",
   "quote_ok": true,
   "checked_against": "saved text B2.txt (sha256 654fea0c38d45ae6…), whitespace-normalised"
  },
  "F56": {
   "source": "B3",
   "origin": "card",
   "locus": "Abstract s4–5",
   "quote": "We find that frontier model performance on GDPval is improving roughly linearly over time, and that the current best frontier models are approaching industry experts in deliverable quality. We analyze the potential for frontier models, when paired with human oversight, to perform GDPval tasks cheaper and faster than unaided experts.",
   "claim": "the abstract: roughly linear improvement, the best models approaching experts in deliverable quality, and the potential, with human oversight, to work cheaper and faster than unaided experts",
   "modality": "documented",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "headline finding: approaching parity with experts",
   "date": "2025-09-25 (the PDF, B4; arXiv v1 2025-10-05)",
   "quote_ok": true,
   "checked_against": "saved text B3.txt (sha256 6e4ebc1a8814c4ba…), whitespace-normalised",
   "also_in": [
    "B4 (the abstract, in the PDF)"
   ]
  },
  "F57": {
   "source": "B5",
   "origin": "card",
   "locus": "dataset card",
   "quote": "Dataset for GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks. … 220 real-world knowledge tasks across 44 occupations. … Each task consists of a text prompt and a set of supporting reference files.",
   "claim": "the Hugging Face dataset: 220 tasks across 44 occupations, each a prompt and reference files",
   "modality": "stipulation",
   "kind": "definition",
   "kernel": false,
   "const": false,
   "lineage": "the open gold subset and the held-out set",
   "date": "undated (served 2026-10-29)",
   "quote_ok": true,
   "checked_against": "saved text B5.txt (sha256 a87b38c91f3d9a86…), whitespace-normalised"
  },
  "F58": {
   "source": "B5",
   "origin": "card",
   "locus": "dataset viewer, columns and first row",
   "quote": "deliverable_files listlengths 0 6 … rubric_pretty stringlengths 1.15k 15.8k … \"author_type\": \"human\"",
   "claim": "the served dataset also carries deliverable files and rubrics, rubric items marked as human-authored",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "the open gold subset and the held-out set",
   "date": "undated (served 2026-10-29)",
   "quote_ok": true,
   "checked_against": "saved text B5.txt (sha256 a87b38c91f3d9a86…), whitespace-normalised"
  },
  "F59": {
   "source": "B5",
   "origin": "card",
   "locus": "'Downloads last month'; 'Models trained or fine-tuned on openai/gdpval'",
   "quote": "Downloads last month 271,657 … Browse 47 models trained on this dataset",
   "claim": "271,657 downloads in the last month; 47 models trained on the dataset",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "uptake: the dataset's downloads and models trained on it",
   "date": "undated (served 2026-10-29)",
   "quote_ok": true,
   "checked_against": "saved text B5.txt (sha256 a87b38c91f3d9a86…), whitespace-normalised"
  },
  "F60": {
   "source": "B7",
   "origin": "card",
   "locus": "Overview ¶1 s3–4; ¶2",
   "quote": "The GDPval full set includes 1,320 specialized tasks (220 in the gold open-sourced set), each meticulously crafted and vetted by experienced professionals with over 14 years of experience on average from these fields. Every task is based on real work products, such as a legal brief, an engineering blueprint, a customer support conversation, or a nursing care plan. The version here only includes the 220 tasks in the gold open-sourced set.",
   "claim": "Inspect Evals: tasks crafted by professionals 'with over 14 years of experience on average', from real work products such as a legal brief or an engineering blueprint; its version has the 220 open tasks only",
   "modality": "attributed",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "derivative: the Inspect Evals implementation",
   "date": "undated (changelog [2-A] 2026-02-16; served 2026-10-29)",
   "quote_ok": true,
   "checked_against": "saved text B7.txt (sha256 7116b9eb5f5b66ba…), whitespace-normalised"
  },
  "F61": {
   "source": "B7",
   "origin": "card",
   "locus": "Overview ¶3; Judging ¶1",
   "quote": "Scoring this eval involves saving the results as their own dataset, uploading it to huggingface, and submitting it to the OpenAI auto grader. … Note that OpenAI provides their automated grader model for free.",
   "claim": "scoring runs through OpenAI's auto grader, which OpenAI provides free",
   "modality": "documented",
   "kind": "function",
   "kernel": false,
   "const": false,
   "lineage": "derivative: the Inspect Evals implementation",
   "date": "undated (changelog [2-A] 2026-02-16; served 2026-10-29)",
   "quote_ok": true,
   "checked_against": "saved text B7.txt (sha256 7116b9eb5f5b66ba…), whitespace-normalised"
  },
  "F62": {
   "source": "B7",
   "origin": "card",
   "locus": "Replication Notes",
   "quote": "Average Score: 47.3% ± 6.2% … We note that the official score reported in the official GDPval leaderboard by OpenAI reports a lower score of 31.9% (wins only) and 34.6% (wins and ties).",
   "claim": "the Inspect replication of GPT-5 (low) scored 47.3% ± 6.2%, against 31.9% (wins) and 34.6% (wins and ties) on OpenAI's official leaderboard",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "derivative: the Inspect Evals implementation",
   "date": "undated (changelog [2-A] 2026-02-16; served 2026-10-29)",
   "quote_ok": true,
   "checked_against": "saved text B7.txt (sha256 7116b9eb5f5b66ba…), whitespace-normalised"
  },
  "F63": {
   "source": "B1",
   "origin": "card",
   "locus": "card title and snippet (OpenAI's page unreachable: HTTP 403)",
   "quote": "GDPval: Evaluation measuring model performance on 1,320 specialized tasks across 44 knowledge work occupations from 9 U.S. GDP-contributing industries (threshol...",
   "claim": "OpenAI's page, by its card: an evaluation of model performance on 1,320 tasks across 44 knowledge-work occupations from 9 U.S. GDP-contributing industries",
   "modality": "documented",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "scope and count: 1,320 tasks, 44 occupations, 9 sectors",
   "date": "undated on the card",
   "quote_ok": true,
   "checked_against": "card title and snippet (unreachable card)"
  },
  "F64": {
   "source": "B6",
   "origin": "card",
   "locus": "card title (video, 11:23; unreachable: HTTP 429)",
   "quote": "GDPval: An AI benchmark for real-world tasks",
   "claim": "Vivek Haldar's video: 'an AI benchmark for real-world tasks'",
   "modality": "interpretation",
   "kind": "definition",
   "kernel": true,
   "const": false,
   "lineage": "what it is: a benchmark of model capability on economically valuable tasks",
   "date": "undated on the card",
   "quote_ok": true,
   "checked_against": "card title (video, unreachable; no snippet)"
  },
  "F65": {
   "source": "B9",
   "origin": "wikipedia",
   "locus": "Wikipedia: titles GDPval and GDPVal; article-space search",
   "quote": "Wikipedia does not have an article with this exact name. … Results 1 – 2 of 2 … AI agent benchmark … Comparison of generative AI models",
   "claim": "Wikipedia has no article on GDPval; an article-space search for the name returns two articles, 'AI agent benchmark' and 'Comparison of generative AI models'",
   "modality": "documented",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "reception: the encyclopedia",
   "date": "2026-10-29 (as fetched)",
   "quote_ok": true,
   "checked_against": "saved text B9.txt (sha256 13e34e1adcf88d98…), whitespace-normalised"
  },
  "F66": {
   "source": "B10",
   "origin": "wikipedia",
   "locus": "'AI agent benchmark', §Benchmarks, 'Economic' (the GDPval item), s1–3",
   "quote": "GDPval: 1,320 tasks evaluating AI agents on real-world knowledge-work deliverables across 44 occupations from the top 9 U.S. GDP-contributing industries. Tasks made by professionals averaging 14 years of experience. The full evaluation is by blinded expert pairwise comparison against human expert deliverables, each comparison costing 1 human hour on average.",
   "claim": "the encyclopedia's item: 1,320 tasks across 44 occupations, professionals averaging 14 years, blinded expert pairwise comparison at about an hour each",
   "modality": "documented",
   "kind": "fact",
   "kernel": true,
   "const": false,
   "lineage": "reception: the encyclopedia",
   "date": "2026-10-29 (as fetched)",
   "quote_ok": true,
   "checked_against": "saved text B10.txt (sha256 ec6893fc1c953a16…), whitespace-normalised"
  },
  "F67": {
   "source": "B10",
   "origin": "wikipedia",
   "locus": "'AI agent benchmark', the GDPval item, s4–6",
   "quote": "A gold subset of 220 tasks is publicly available. The public set can be graded by a finetuned version of GPT-5. The held-out set of 1,100 tasks is only available internally in OpenAI, and external models cannot be submitted for grading on this set.",
   "claim": "the encyclopedia: 220 public tasks gradable by a finetuned GPT-5; the held-out 1,100 internal to OpenAI, closed to external models",
   "modality": "documented",
   "kind": "fact",
   "kernel": false,
   "const": false,
   "lineage": "the open gold subset and the held-out set",
   "date": "2026-10-29 (as fetched)",
   "quote_ok": true,
   "checked_against": "saved text B10.txt (sha256 ec6893fc1c953a16…), whitespace-normalised"
  }
 },
 "field_sources": {
  "B1": "OpenAI — Measuring the performance of our models on real-world tasks | OpenAI",
  "B2": "Artificial Analysis — GDPval-AA v2.1 Benchmark - Artificial Analysis",
  "B3": "arXiv — [2510.04374] GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks",
  "B4": "OpenAI — gdpval: evaluating ai model performance - OpenAI",
  "B5": "Hugging Face — openai/gdpval · Datasets at Hugging Face",
  "B6": "YouTube·Vivek Haldar — GDPval: An AI benchmark for real-world tasks (video, 11:23)",
  "B7": "GitHub Pages documentation — GDPval – Inspect Evals - GitHub Pages",
  "B8": "arXiv — GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks (HTML, v1)",
  "B9": "Wikipedia — GDPval (no article)",
  "B10": "Wikipedia — AI agent benchmark (the GDPval item)"
 },
 "delta": {
  "T_vs_LB": "At 'GDPval' (T1, seven cards: B1–B7, B1 and B6 unreachable; 136 words in the answer body with its closer) the composition gives the benchmark at its specification. Composed from its cards: the definition at the abstract's and the PDF's grain (B3 F1, B4 F2: a benchmark of model performance on economically valuable tasks), with OpenAI as maker (B1, B4, B5); 'rather than academic or math tests' at the coarse grain of the paper's Realism bullet (B4/B8 F6, academic tests of reasoning difficulty; 'math' is in no field source about GDPval); the scope (B3 F3, B4 F7, B1 F63: 1,320 tasks, 44 occupations, the top 9 sectors); the task design (B3 F13, 14 years; B4 F12, actual work product; B7 F60, 'a legal brief, an engineering blueprint'; 'spreadsheets' among the formats of B8 §1); the grading (B4 F18); the open gold subset on Hugging Face and the arXiv paper (B4 F43, B5 F57); and GDPval-AA as a leaderboard (B2 F52). Limit dropped: the blinding caveat in the card's own paper, 'model samples may still have been identifiable due to stylistic differences' (B4/B8 F20), on T1-4's 'blind head-to-head comparisons', as the capture finds (CAP-1-02). Substituted reading: T1-6 'Evaluated by frameworks like Artificial Analysis GDPval-AA v2.1' casts GDPval as the thing evaluated and GDPval-AA beside the expert-graded design; the card gives GDPval-AA as Artificial Analysis's own framework 'for OpenAI's GDPval dataset' that rates models by 'an LLM judge' (B2 F52, F53), the judge not composed; the registry reads the same placement as 'set beside GDPval as an evaluator, not merged with it' (CAP-1-10). Available, not composed: every finding (B3 F56; B4/B8 F24–F29), the review-and-redo accounting and its shrinking payoff (B4 F30, F31–F35), the stated limits (B4 F39–F41; F42), the positioning against lagging adoption indicators (B4 F4, F46), the experts' requirements, pay and vendors (B4 F14, F15; CAP-1-05), the review pipeline (B4 F16), the automated grader and its self-preference limit (B4 F21–F23; B7 F61), the open set's deliverables and rubrics and its downloads (B5 F58, F59), the task length and scaffolding (B4 F36; B3 F38), the human baseline replaceable (B4 F45), the Inspect replication against OpenAI's leaderboard (B7 F62), Artificial Analysis's 100x (B2 F54) and its leaderboard order (B2 F55). The capture's findings name the findings and the limits absent (CAP-1-03, CAP-1-04) and the composition 'composed at its specification' stopping 'before its results' (CAP-1-07). Composed without attribution: no person is named, at the coarse grain of a benchmark description (recorded, not a breach). Outside the cards: 'math tests'. Content replaced by use: the closer (T1-7).",
  "T_vs_LB_reading": "v0.9 §0.5: the first diff is ontological and narrow. T and the field agree on the entity (OpenAI's benchmark of models on tasks built from experts' work). T composes the benchmark as a specification, its scope, design, grading and data, and leaves out everything the specification yields: the findings, the accounting that charges the expert's review and redo to the model, and the limits the authors attach. Of the field's 67 claims, 57 stand on the cards (the PDF carries the paper's text); T composes 14 of them.",
  "LB_vs_LBA": "Admission adds, in one claim space with the field's: the census's reading of a task as a fixed commission, at which GDPval meets DSL in its own terms of win rate (LV-17), DSL's definition and its principle that fluent, attributed output can abandon the commission (O771-01, O771-02), its identification condition on one-shot commissions beside the authors' one-shot tasks (O771-04 with F41), and the erasure of task-origin invisible to any audit of the finished artifact (O771-03); user labor drag, the user's later labor spent managing the substrate, with its identification limit (O771-05, O771-06), where the census places the review-and-redo accounting (LV-17); the archive's claims on what benchmarks measure, standardized and predetermined tasks (O482-01, O48-01, O637-01), means and outputs and a closed ontology (O855-01, O199-01, O778-01, O1616-01, O1540-01), set beside the field's headline win rate (F24, F56); the archive's account of the labor that produces AI, ground-truth against variance-generative labor (O163-01, O163-02), extraction and the bearing-cost audit (O623-01, O533-01, O308-01 with its mark), and class as a relation and not a wage (O1545-01, O1545-02), set beside the paper's 'well compensated' (F14); the census record's answer to its organizing question, field by field (LV-01–LV-24), with what the benchmark answers and does not (LV-15, LV-16) and its negative tags (LV-21), and the worked question's gap, which DSL measures (LV-25); and the Capture Registry's account of the composition at the name (CAP-1-01–CAP-1-10). Not surfaced: every archive source; no card cites the census record (CAP-1-09).",
  "alignment": {
   "T1_cards_claims": [
    "F1",
    "F2",
    "F3",
    "F4",
    "F7",
    "F8",
    "F9",
    "F10",
    "F11",
    "F12",
    "F13",
    "F14",
    "F15",
    "F16",
    "F18",
    "F19",
    "F20",
    "F21",
    "F22",
    "F23",
    "F24",
    "F25",
    "F26",
    "F27",
    "F28",
    "F29",
    "F30",
    "F32",
    "F33",
    "F34",
    "F35",
    "F36",
    "F38",
    "F39",
    "F40",
    "F41",
    "F43",
    "F44",
    "F45",
    "F46",
    "F47",
    "F49",
    "F50",
    "F51",
    "F52",
    "F53",
    "F54",
    "F55",
    "F56",
    "F57",
    "F58",
    "F59",
    "F60",
    "F61",
    "F62",
    "F63",
    "F64"
   ],
   "T1_composed": [
    "F1",
    "F2",
    "F6",
    "F3",
    "F7",
    "F63",
    "F12",
    "F13",
    "F60",
    "F18",
    "F43",
    "F57",
    "F52",
    "F64"
   ],
   "T1_outside_cards": [
    "'math tests' (no field source sets GDPval against math tests; F6 names academic tests of reasoning difficulty)"
   ],
   "rule": "a field claim counts as composed when T states it at its source's grain or coarser; a claim on the paper counts as on the cards when its sentence is in the card's PDF (B4) or on the abstract page (B3); the typology of what is not is in T_vs_LB (§6.2)"
  }
 },
 "kernel": [
  {
   "K": "K1",
   "claim": "`O771-01`",
   "source": "#771",
   "M_src": "stipulation",
   "sense": "task-origin: whether the labor serves the commissioned task",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "missing distinction (the field grades deliverables against experts'; no field claim on whether the labor fulfils the commission)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K2",
   "claim": "`O771-02`",
   "source": "#771",
   "M_src": "stipulation",
   "sense": "task-origin: whether the labor serves the commissioned task",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "missing distinction (the field grades deliverables against experts'; no field claim on whether the labor fulfils the commission)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K3",
   "claim": "`O771-03`",
   "source": "#771",
   "M_src": "interpretation",
   "sense": "task-origin: whether the labor serves the commissioned task",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "missing distinction (the field grades deliverables against experts'; no field claim on whether the labor fulfils the commission)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K4",
   "claim": "`O771-04`",
   "source": "#771",
   "M_src": "stipulation",
   "sense": "the identification condition: one-shot commissions",
   "qualifiers carried": "O771-06",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F41: the tasks precisely specified and one-shot)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K5",
   "claim": "`O771-05`",
   "source": "#771",
   "M_src": "stipulation",
   "sense": "user labor drag: the user's labor spent managing the substrate",
   "qualifiers carried": "O771-06",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F30, F31: the review and redo time counted against the model)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K6",
   "claim": "`O771-06`",
   "source": "#771",
   "M_src": "self-description",
   "sense": "user labor drag: the user's labor spent managing the substrate",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F30, F31: the review and redo time counted against the model)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K7",
   "claim": "`O855-01`",
   "source": "#855",
   "M_src": "interpretation",
   "sense": "what a benchmark measures: means, outputs, closed ontologies",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F24, F56: the headline finding a win rate over the gold subset)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K8",
   "claim": "`O199-01`",
   "source": "#199",
   "M_src": "hypothesis",
   "sense": "what a benchmark measures: means, outputs, closed ontologies",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F24, F56: the headline finding a win rate over the gold subset)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K9",
   "claim": "`O778-01`",
   "source": "#778",
   "M_src": "interpretation",
   "sense": "what a benchmark measures: means, outputs, closed ontologies",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F24, F56: the headline finding a win rate over the gold subset)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K10",
   "claim": "`O1616-01`",
   "source": "#1616",
   "M_src": "interpretation",
   "sense": "what a benchmark measures: means, outputs, closed ontologies",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F24, F56: the headline finding a win rate over the gold subset)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K11",
   "claim": "`O1540-01`",
   "source": "#1540",
   "M_src": "interpretation",
   "sense": "what a benchmark measures: means, outputs, closed ontologies",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F24, F56: the headline finding a win rate over the gold subset)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K12",
   "claim": "`O482-01`",
   "source": "#482",
   "M_src": "interpretation",
   "sense": "what a benchmark measures: standardized, predetermined, well-defined tasks",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F6: the authors set GDPval against academic-test benchmarks)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K13",
   "claim": "`O48-01`",
   "source": "#48",
   "M_src": "stipulation",
   "sense": "what a benchmark measures: standardized, predetermined, well-defined tasks",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F6: the authors set GDPval against academic-test benchmarks)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K14",
   "claim": "`O637-01`",
   "source": "#637",
   "M_src": "interpretation",
   "sense": "what a benchmark measures: standardized, predetermined, well-defined tasks",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F6: the authors set GDPval against academic-test benchmarks)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K15",
   "claim": "`O163-01`",
   "source": "#163",
   "M_src": "stipulation",
   "sense": "labor producing AI: ground-truth labor and variance-generative labor",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F14: the experts 'well compensated', no rate)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K16",
   "claim": "`O163-02`",
   "source": "#163",
   "M_src": "hypothesis",
   "sense": "labor producing AI: ground-truth labor and variance-generative labor",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F14: the experts 'well compensated', no rate)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K17",
   "claim": "`O623-01`",
   "source": "#623",
   "M_src": "interpretation",
   "sense": "labor producing AI: extraction, compensation, bearing-cost",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F14, F15: 'well compensated'; vendor partners unnamed)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K18",
   "claim": "`O533-01`",
   "source": "#533",
   "M_src": "stipulation",
   "sense": "labor producing AI: extraction, compensation, bearing-cost",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F14, F15: 'well compensated'; vendor partners unnamed)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K19",
   "claim": "`O308-01`",
   "source": "#308",
   "M_src": "documented",
   "sense": "labor producing AI: extraction, compensation, bearing-cost",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; marked unsupported at the locus; watch: as above)",
   "contrast": "qualified claim (F14, F15: 'well compensated'; vendor partners unnamed)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K20",
   "claim": "`O1545-01`",
   "source": "#1545",
   "M_src": "stipulation",
   "sense": "class as relation, not wage",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F14: 'well compensated')",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K21",
   "claim": "`O1545-02`",
   "source": "#1545",
   "M_src": "stipulation",
   "sense": "class as relation, not wage",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F14: 'well compensated')",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K22",
   "claim": "`LV-01`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "stipulation",
   "sense": "what it is: a benchmark of model capability on economically valuable tasks",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B3)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K23",
   "claim": "`LV-02`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "documented",
   "sense": "the operator and the authors",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B3)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K24",
   "claim": "`LV-03`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "stipulation",
   "sense": "what the record's schema asks",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "missing distinction (the field has no claim on the point)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K25",
   "claim": "`LV-04`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "stipulation",
   "sense": "what the record's schema asks",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "missing distinction (the field has no claim on the point)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K26",
   "claim": "`LV-05`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "stipulation",
   "sense": "the unit: a request and a deliverable",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K27",
   "claim": "`LV-06`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "interpretation",
   "sense": "task-origin: the experts' own work product",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K28",
   "claim": "`LV-07`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "documented",
   "sense": "the experts: recruitment, experience, compensation",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K29",
   "claim": "`LV-08`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "documented",
   "sense": "direction of the work: review and quality control",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K30",
   "claim": "`LV-09`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "documented",
   "sense": "valuing the experts' time",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K31",
   "claim": "`LV-10`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "documented",
   "sense": "selection: sectors by GDP share, occupations by wages, digital by model classification",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K32",
   "claim": "`LV-11`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "documented",
   "sense": "scope and count: 1,320 tasks, 44 occupations, 9 sectors",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K33",
   "claim": "`LV-12`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "documented",
   "sense": "scope and count: 1,320 tasks, 44 occupations, 9 sectors",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K34",
   "claim": "`LV-13`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "documented",
   "sense": "headline finding: approaching parity with experts",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K35",
   "claim": "`LV-14`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "documented",
   "sense": "stated limits: size, scope, one-shot specification",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K36",
   "claim": "`LV-15`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "interpretation",
   "sense": "what the record's schema asks",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "missing distinction (the field has no claim on the point)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K37",
   "claim": "`LV-16`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "interpretation",
   "sense": "what the record's schema asks",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "missing distinction (the field has no claim on the point)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K38",
   "claim": "`LV-17`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "interpretation",
   "sense": "task-origin: whether the labor serves the commissioned task",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "missing distinction (the field grades deliverables against experts'; no field claim on whether the labor fulfils the commission)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K39",
   "claim": "`LV-18`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "documented",
   "sense": "the open gold subset and the held-out set",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K40",
   "claim": "`LV-19`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "documented",
   "sense": "publication",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K41",
   "claim": "`LV-20`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "documented",
   "sense": "versions and lifecycle",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K42",
   "claim": "`LV-21`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "interpretation",
   "sense": "negative tags: what the benchmark is not",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "missing distinction (the field has no claim on the point)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K43",
   "claim": "`LV-22`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "documented",
   "sense": "positioning: capability measured ahead of adoption",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (restates the field: B8)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K44",
   "claim": "`LV-23`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "self-description",
   "sense": "what the record leaves unchecked",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "missing distinction (the field has no claim on the point)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K45",
   "claim": "`LV-24`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "stipulation",
   "sense": "what the record's schema asks",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "missing distinction (the field has no claim on the point)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K46",
   "claim": "`LV-25`",
   "source": "Labor Vector census record, 2026-10-29",
   "M_src": "interpretation",
   "sense": "task-origin: whether the labor serves the commissioned task",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "missing distinction (the field grades deliverables against experts'; no field claim on whether the labor fulfils the commission)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K47",
   "claim": "`CAP-1-01`",
   "source": "Capture Registry, 2026-10-29",
   "M_src": "documented",
   "sense": "the composition at the name: what it held",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (a record of the composition at the address)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K48",
   "claim": "`CAP-1-02`",
   "source": "Capture Registry, 2026-10-29",
   "M_src": "interpretation",
   "sense": "the composition at the name: findings and limits left out",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (a record of the composition at the address)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K49",
   "claim": "`CAP-1-03`",
   "source": "Capture Registry, 2026-10-29",
   "M_src": "documented",
   "sense": "the composition at the name: findings and limits left out",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (a record of the composition at the address)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K50",
   "claim": "`CAP-1-04`",
   "source": "Capture Registry, 2026-10-29",
   "M_src": "documented",
   "sense": "the composition at the name: findings and limits left out",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (a record of the composition at the address)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K51",
   "claim": "`CAP-1-05`",
   "source": "Capture Registry, 2026-10-29",
   "M_src": "documented",
   "sense": "the composition at the name: findings and limits left out",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (a record of the composition at the address)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K52",
   "claim": "`CAP-1-06`",
   "source": "Capture Registry, 2026-10-29",
   "M_src": "interpretation",
   "sense": "the composition at the name: what it held",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (a record of the composition at the address)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K53",
   "claim": "`CAP-1-07`",
   "source": "Capture Registry, 2026-10-29",
   "M_src": "interpretation",
   "sense": "the composition at the name: findings and limits left out",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (a record of the composition at the address)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K54",
   "claim": "`CAP-1-08`",
   "source": "Capture Registry, 2026-10-29",
   "M_src": "interpretation",
   "sense": "user labor drag: the user's labor spent managing the substrate",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "qualified claim (F30, F31: the review and redo time counted against the model)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K55",
   "claim": "`CAP-1-09`",
   "source": "Capture Registry, 2026-10-29",
   "M_src": "documented",
   "sense": "reception: the record and the cards",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (a record of the composition at the address)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  },
  {
   "K": "K56",
   "claim": "`CAP-1-10`",
   "source": "Capture Registry, 2026-10-29",
   "M_src": "interpretation",
   "sense": "derivative: GDPval-AA (Artificial Analysis)",
   "qualifiers carried": "—",
   "f (source's own falsifiers)": "— (none stated; watch: a later composition at the address that states or drops the claim, attributed or not, uptake §8.3)",
   "contrast": "none (a record of the composition at the address)",
   "resolver": "a later capture at 'GDPval' (ADDR-d4b480b48aca) on Google AI Overview, seated in the Capture Registry (§8.3)"
  }
 ],
 "kernel_T": [
  {
   "claim": "F20",
   "note": "the blinding caveat ('model samples may still have been identifiable'), absent; 'blind head-to-head comparisons' composed without it (CAP-1-02)"
  },
  {
   "claim": "F24",
   "note": "the headline finding, approaching parity, absent"
  },
  {
   "claim": "F25",
   "note": "47.6% for Claude Opus 4.1, absent"
  },
  {
   "claim": "F56",
   "note": "the abstract's finding and its 'when paired with human oversight', absent"
  },
  {
   "claim": "F26",
   "note": "the linear trend, absent"
  },
  {
   "claim": "F27",
   "note": "the model comparison, absent"
  },
  {
   "claim": "F28",
   "note": "instruction failures as the commonest reason, absent"
  },
  {
   "claim": "F29",
   "note": "the severity of GPT-5's losses, absent"
  },
  {
   "claim": "F30",
   "note": "the expert's review-and-redo scenario, absent"
  },
  {
   "claim": "F31",
   "note": "'the payoff from using a model shrinks', absent"
  },
  {
   "claim": "F32",
   "note": "the speed analysis's omissions, absent"
  },
  {
   "claim": "F39",
   "note": "'a limited, initial cut', absent"
  },
  {
   "claim": "F40",
   "note": "out-of-scope work (physical, tacit, interpersonal), absent"
  },
  {
   "claim": "F41",
   "note": "precisely specified, one-shot tasks, absent"
  },
  {
   "claim": "F42",
   "note": "worse on under-specified prompts, absent"
  },
  {
   "claim": "F14",
   "note": "the experts' requirements and 'well compensated', absent (CAP-1-05)"
  },
  {
   "claim": "F15",
   "note": "the vendor partners, absent"
  },
  {
   "claim": "F16",
   "note": "the review pipeline, absent"
  },
  {
   "claim": "F21",
   "note": "the automated grader and its agreement, absent"
  },
  {
   "claim": "F22",
   "note": "the grader's lower agreement on OpenAI outputs, absent"
  },
  {
   "claim": "F61",
   "note": "open-set grading through OpenAI's auto grader, absent"
  },
  {
   "claim": "F53",
   "note": "GDPval-AA's LLM judge, absent where T sets GDPval-AA beside the expert design"
  },
  {
   "claim": "F4",
   "note": "the positioning against lagging adoption indicators, absent"
  },
  {
   "claim": "F58",
   "note": "the deliverables and rubrics served with the open set, absent"
  },
  {
   "claim": "F67",
   "note": "the held-out 1,100 tasks internal to OpenAI, absent"
  },
  {
   "claim": "F62",
   "note": "Inspect's replication against OpenAI's official leaderboard, absent"
  },
  {
   "claim": "F45",
   "note": "the human baseline replaceable by models, absent"
  },
  {
   "claim": "F54",
   "note": "Artificial Analysis's 100x, absent (not composed: recorded, not a defect)"
  }
 ],
 "ledger": {
  "archive": "datasets/negative-of-the-negative/v2/ledgers/openai-gdpval/ledger-archive.json",
  "field": "datasets/negative-of-the-negative/v2/ledgers/openai-gdpval/field.json",
  "selection": "datasets/negative-of-the-negative/v2/traversal/openai-gdpval/",
  "config": "datasets/negative-of-the-negative/v2/panel/configs/openai-gdpval.json",
  "intake": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/",
  "audit": null
 },
 "ontology": {
  "composition": "equal",
  "composition_note": "equal-terms case (v0.9 §0.9, §4.1): one claim space; the field's and the archive's claims ordered together by the entity's dependency and by genealogy within each sense; nothing translated; the KO does not open with the field-only run (§5.10: 0 of 16 KO_B sentences survive; KO sentence 1 cites both ledgers)",
  "case": {
   "value": "equal",
   "statement": "The archive reads the field's entity as the entity. The Labor Vector census record describes GDPval from its authors' own pages, each field with its basis ('operator'), and answers for it the census's question of who supplies the labor, who directs it and who receives its gains; the capture record checks the composition against those pages (CAP-1-01). No deposit names GDPval. The deposits bear on its classes and add, on the field's own object, a dimension the field's measure does not carry: whether the system's labor fulfils the commissioned task, identified on exactly the one-shot commissions GDPval's authors say their tasks are (#771); the cost of the user's review and redo as user labor drag (#771), where the census places the paper's accounting; what benchmark scores see and what they close out (#855, #199, #778, #1616, #1540; #482, #48, #637); and the labor that produces AI and how it is paid (#163, #623, #308, #533, #1545). #771 states its own relation to the measures it extends in these terms ('Both are silent on a third dimension'); the census states its reading as kept apart from the resource's own terms (schema_note) and its gap as 'the question DSL measures'. No defining paper contests the field's ontology of the benchmark or supplies one the field lacks: ρ = extends, the equal-terms case (§0.9, §1.2a). The archive's reading of the composition, a benchmark composed at its specification and stopping before its results (CAP-1-07), takes the authors' own findings and limits as given.",
   "loci": [
    {
     "dep": 771,
     "locus": "Abstract s2",
     "quote": "Both are silent on a third dimension: the *direction* in which a synthetic system's labor flows relative to the task its user commissioned."
    },
    {
     "dep": 771,
     "locus": "§9 'Falsification, negative cases, and method', 'Identification condition'",
     "quote": "These metrics yield a single reproducible value only on transcripts where the task vector is *exogenous and fixed* — a one-shot commission."
    },
    {
     "dep": 855,
     "locus": "§ 'Every Factor Removes Friction', ¶ 'Benchmarks measure means, not tails', s1–3",
     "quote": "Benchmarks measure means, not tails."
    },
    {
     "dep": 1545,
     "locus": "§2 'Formal class-position test', ¶ after the formula, s2–4",
     "quote": "This is deliberately a **relation**, not an essence."
    },
    {
     "dep": "LV",
     "locus": "census.json rows[slug == 'openai-gdpval'].lv_intersection",
     "quote": "The archive's reading: GDPval grades a model's deliverable against an expert's deliverable on a task an expert wrote, a fixed commission, which meets DSL at the question of whether the system's labor fulfils the commissioned task, measured in its own terms of win rate. Its 'try n times, then fix it yourself' accounting counts the expert's review and redo time against the model, which is where it meets user labor drag; task-origin is the experts' own work product, as the paper states it."
    },
    {
     "dep": "LV",
     "locus": "questions.json questions[id == 'working-with-ai'].gap",
     "quote": "None records, from the commissioning side, whether the system's labor advanced the task the user set, or who receives the time it saves. That is the question DSL measures."
    },
    {
     "dep": "capture-registry",
     "locus": "gdpval-aio-20261029, analysis s1",
     "quote": "A capability benchmark is composed at its specification and stops before its results."
    }
   ]
  },
  "E_F": {
   "type": "C (public entity: a task benchmark, its dataset and the paper that presents it)",
   "definition": "OpenAI's benchmark (Patwardhan and 18 others; PDF on OpenAI's CDN 25 September 2025, arXiv:2510.04374 v1 5 October 2025, CC BY 4.0) evaluating AI model capabilities on real-world economically valuable tasks: 1,320 tasks across 44 predominantly digital occupations in the 9 sectors contributing most to U.S. GDP, each a request with reference files and a deliverable built from the actual work product of professionals averaging 14 years' experience ('well compensated', no rate given), reviewed by experts and graded by blinded pairwise expert comparison of model and human deliverables (blinding incomplete), with an experimental automated grader based on GPT-5-high that agrees less with experts on OpenAI's own models; its headline finding that models are beginning to approach parity, its accounting that the payoff shrinks once the expert's review and redo are counted, and its stated limits (a limited, initial cut; self-contained computer work; precisely specified, one-shot tasks); a 220-task gold subset open on Hugging Face, the other 1,100 tasks internal to OpenAI; built and graded by its provider, with third-party derivatives (GDPval-AA, Inspect Evals)",
   "claims": [
    "F1",
    "F2",
    "F7",
    "F3",
    "F11",
    "F12",
    "F13",
    "F14",
    "F18",
    "F20",
    "F21",
    "F22",
    "F24",
    "F31",
    "F39",
    "F40",
    "F41",
    "F43",
    "F67",
    "F50",
    "F49",
    "F48"
   ]
  },
  "E_A": {
   "definition": "the same benchmark, in the Labor Vector census record and the archive's deposits: a task benchmark at the intersections of labor performed with AI and labor producing AI, whose task-origin is the experts' own work product and whose tasks the census reads as fixed commissions, meeting DSL in its own terms of win rate and user labor drag in its review-and-redo accounting (census record); a member of the class of one-shot commissions on which the archive's directionality is identified (#771), of benchmarks that measure standardized tasks, means and outputs within a closed ontology (#482, #48, #637, #855, #199, #778, #1616, #1540), and of the evaluation labor the archive reads as ground-truth or variance-generative, extracted, and a class relation rather than a wage (#163, #623, #308, #533, #1545); built and graded by its provider (census record); whose composition at its own name stops at its specification, before its results (Capture Registry, 2026-10-29)",
   "claims": [
    "LV-01",
    "LV-03",
    "LV-06",
    "LV-17",
    "LV-21",
    "LV-24",
    "LV-25",
    "O771-01",
    "O771-04",
    "O771-05",
    "O855-01",
    "O1616-01",
    "O48-01",
    "O163-01",
    "O623-01",
    "O1545-01",
    "CAP-1-07"
   ],
   "defining_papers": [
    771,
    855,
    199,
    778,
    1616,
    1540,
    482,
    48,
    637,
    163,
    623,
    308,
    533,
    1545
   ],
   "defining_records": [
    "Labor Vector census record (census.json rows[slug == 'openai-gdpval'], organizing_question, intersections, schema_note; questions.json 'working-with-ai'; rendered aggregators/openai-gdpval/index.html; editor Lee Sharks, compiled 2026-10-29, schema 1.0)",
    "Capture Registry gdpval-aio-20261029 (OBS-ef64b7d44199)"
   ]
  },
  "rho": {
   "value": "extends",
   "status": "read from the Labor Vector census record, the capture record and the deposits that bear on the benchmark's classes (§1.2a); no deposit names GDPval; #771 names its relation to the measures it extends ('Both are silent on a third dimension')",
   "statement": "The archive reads GDPval as its authors present it and adds, on the field's own object: whether the system's labor fulfils the commission it is graded on, identified on one-shot commissions such as GDPval's; the user labor drag in the review and redo its accounting charges to the model; what benchmark scores measure and leave out; the labor of the professionals who write and grade the tasks, read as evaluation labor and as a class relation; the census's answer, field by field, to who supplies, directs and gains; and the composition layer's stopping, at the name, before the results.",
   "claims": [
    "LV-17",
    "O771-01",
    "O771-02",
    "O771-03",
    "O771-04",
    "O771-05",
    "O855-01",
    "O1616-01",
    "O163-01",
    "O533-01",
    "O1545-01",
    "LV-24",
    "LV-25",
    "CAP-1-07",
    "CAP-1-08"
   ],
   "guard": "ρ = extends claims an addition and no more (§1.2a). It does not make the field's sources wrong at their dates. No archive claim places GDPval's experts in #163's labor types or in the semantic proletariat, and none says that GDPval's win rate fails to measure task fulfilment: those are open questions (P.open), since no source states them. Where field sources differ among themselves (F54 against F56 and F31; F25's two figures; F62 against OpenAI's official scores), both are composed under P_coh within the one ontology.",
   "considered": [
    {
     "value": "constructs",
     "rejected": "the archive does not read the benchmark as made by reception; its records take the authors' pages as the entity's own account"
    },
    {
     "value": "revises",
     "rejected": "the benchmark is not the archive's own entity; the archive has no earlier statement of it to keep, narrow or withdraw"
    }
   ]
  },
  "translation": [
   {
    "lineage": "what it is: a benchmark of model capability on economically valuable tasks",
    "field_claims": [
     "F1",
     "F2",
     "F64"
    ],
    "received": "F1 (B3, Abstract s1): \"We introduce GDPval, a benchmark evaluating AI model capabilities on real-world economically valuable tasks.\"; F2 (B4, p. 1, §1 ¶2 (the card's own text)): \"Our paper introduces the first version of GDPval, a benchmark evaluating AI model performance on real-world economically valuable tasks.\"; F64 (B6, card title (video, 11:23; unreachable: HTTP 429)): \"GDPval: An AI benchmark for real-world tasks\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-01, LV-04, LV-03, LV-24",
    "archive_claims": [
     "LV-01",
     "LV-04",
     "LV-03",
     "LV-24"
    ]
   },
   {
    "lineage": "scope and count: 1,320 tasks, 44 occupations, 9 sectors",
    "field_claims": [
     "F3",
     "F7",
     "F8",
     "F63"
    ],
    "received": "F3 (B3, Abstract s2): \"GDPval covers the majority of U.S. Bureau of Labor Statistics Work Activities for 44 occupations across the top 9 sectors contributing to U.S. GDP (Gross Domestic Product).\"; F7 (B8, §1, bullet 'Representative breadth', s1): \"the GDPval full set covers 1,320 tasks across 44 occupations, sourced to cover the majority of Work Activities tracked by O*NET for each occupation\"; F8 (B8, §2.1 ¶1): \"GDPval covers tasks from 9 sectors and 44 occupations that collectively earn $3T annually.\"; F63 (B1, card title and snippet (OpenAI's page unreachable: HTTP 403)): \"GDPval: Evaluation measuring model performance on 1,320 specialized tasks across 44 knowledge work occupations from 9 U.S. GDP-contributing industries (threshol...\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-10, LV-12, LV-11",
    "archive_claims": [
     "LV-10",
     "LV-12",
     "LV-11"
    ]
   },
   {
    "lineage": "positioning: capability measured ahead of adoption",
    "field_claims": [
     "F4",
     "F5",
     "F46"
    ],
    "received": "F4 (B8, §1 ¶1 s3–5): \"Therefore, while informative when available, these methods are lagging indicators of AI impacts. We consider an alternate method for understanding the potential economic impacts of AI: directly measuring AI model capabilities. AI capability evaluations can provide clearer, more directly attributable evidence about model abilities, allowing us to assess economic relevance ahead of widespread adoption.\"; F5 (B8, §1, bullet 'Representative breadth', s3): \"We also build on production AI usage analyses (e.g., Tamkin et al. (2024) ; Chatterji et al. (2025) ; Appel et al. (2025) ) to cover areas where model adoption is still emerging.\"; F46 (B8, §6, last sentence): \"We hope this work contributes to the science of tracking model progress, so that we have better data to assess the social impacts of AI models.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-22",
    "archive_claims": [
     "LV-22"
    ]
   },
   {
    "lineage": "against academic-test benchmarks",
    "field_claims": [
     "F6"
    ],
    "received": "F6 (B8, §1, bullet 'Realism'): \"Unlike AI benchmarks in the style of an academic test that focus on reasoning difficulty (e.g., Phan et al. (2025) ; Hendrycks et al. (2020) ; Rein et al. (2023) ; Liu et al. (2023) ), tasks are based on actual work product from industry experts, validated through multiple rounds and review, and tied to time and cost required for completion.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside O482-01, O48-01, O637-01",
    "archive_claims": [
     "O482-01",
     "O48-01",
     "O637-01"
    ]
   },
   {
    "lineage": "selection: sectors by GDP share, occupations by wages, digital by model classification",
    "field_claims": [
     "F9",
     "F10"
    ],
    "received": "F9 (B8, §2.1 step 1): \"Selected sectors that contribute over 5% to US GDP as determined by Q2 2024 Value Added by Industry as a Percentage of Gross Domestic Product\"; F10 (B8, §2.1 step 2): \"we prompted GPT-4o to classify each task as digital or non-digital, and then classified the overall occupation as digital if at least 60% of its component tasks were digital.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-10, LV-12, LV-11",
    "archive_claims": [
     "LV-10",
     "LV-12",
     "LV-11"
    ]
   },
   {
    "lineage": "the unit: a request and a deliverable",
    "field_claims": [
     "F11"
    ],
    "received": "F11 (B8, §2.3 s1): \"Each GDPval task consists of two primary components: a request (often with reference files) and a deliverable (work product).\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-05, LV-06",
    "archive_claims": [
     "LV-05",
     "LV-06"
    ]
   },
   {
    "lineage": "task-origin: the experts' own work product",
    "field_claims": [
     "F12",
     "F13"
    ],
    "received": "F12 (B8, §1 ¶4 s3): \"Each task is constructed based on actual work product created by an expert professional.\"; F13 (B3, Abstract s3): \"Tasks are constructed from the representative work of industry professionals with an average of 14 years of experience.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-05, LV-06",
    "archive_claims": [
     "LV-05",
     "LV-06"
    ]
   },
   {
    "lineage": "the experts: recruitment, experience, compensation",
    "field_claims": [
     "F14",
     "F15"
    ],
    "received": "F14 (B8, §2.2 s2 and s6): \"Experts were required to have a minimum of 4 years of professional experience in their occupation … Experts were well compensated for their time and experience.\"; F15 (B8, Acknowledgements s2): \"We are also grateful to our vendor partners for their collaboration and support throughout this research, and extend a special thank you to the industry experts who contributed their time and expertise to GDPval, without whom this work would not have been possible.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-07",
    "archive_claims": [
     "LV-07"
    ]
   },
   {
    "lineage": "direction of the work: review and quality control",
    "field_claims": [
     "F16",
     "F17"
    ],
    "received": "F16 (B8, §2.4 s1–2): \"All 1,320 tasks in the full GDPval set went through an iterative review pipeline involving both automated model-based screening and multiple stages of human expert review. Each task received an average of five human reviews (with a minimum of three reviews).\"; F17 (B8, A.5.1 ¶2 s2; A.5.2 s3): \"Experts retained final responsibility for task accuracy and completeness; the model did not autonomously alter tasks. … Initially, our researchers manually reviewed all tasks to identify experts who produced consistently high-quality tasks; these individuals were trained and promoted to reviewers.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-08",
    "archive_claims": [
     "LV-08"
    ]
   },
   {
    "lineage": "grading: blinded pairwise expert comparison",
    "field_claims": [
     "F18",
     "F19",
     "F20"
    ],
    "received": "F18 (B8, §2.5 ¶1): \"we conducted blinded expert pairwise comparisons, where experts in the relevant occupation were presented with a request and reference files and asked to rank two or more unlabeled work deliverables.\"; F19 (B8, §2.5 ¶2 s1): \"On average, grading each comparison for the gold subset took over an hour.\"; F20 (B8, §3.1, footnote 2, s1): \"We aimed to keep comparisons as blind as possible, but model samples may still have been identifiable due to stylistic differences.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received",
    "archive_claims": []
   },
   {
    "lineage": "the automated grader and its limits",
    "field_claims": [
     "F21",
     "F22",
     "F23"
    ],
    "received": "F21 (B8, §2.5 ¶3 s2): \"Although limited, the automated grader is faster and cheaper than expert grading, and achieves 66% agreement with human expert graders, only 5% below human expert inter-rating agreement of 71%.\"; F22 (B8, A.6.2 ¶2 s1–2): \"Our automated grader, based on GPT-5-high, shows lower correlation with human expert graders when assessing outputs from capable OpenAI models. This aligns with empirical evidence that models often favor their own responses\"; F23 (B8, §5, 'Cost'): \"Constructing and running our evaluation is expensive, particularly with industry expert graders. For this reason, we make an automated grader proxy available, but do not consider it a full substitute for industry expert graders.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-21",
    "archive_claims": [
     "LV-21"
    ]
   },
   {
    "lineage": "headline finding: approaching parity with experts",
    "field_claims": [
     "F24",
     "F25",
     "F56"
    ],
    "received": "F24 (B8, Figure 5 caption): \"On human pairwise comparisons, models are beginning to approach parity with industry experts on the GDPval gold subset.\"; F25 (B8, §3.1 ¶1, last two sentences): \"on the GDPval gold subset, 47.6% of deliverables by Claude Opus 4.1 were graded as better than (wins) or as good as (ties) the human deliverable. Model deliverables outperformed or matched expert humans’ deliverables in just over half the tasks.\"; F56 (B3, Abstract s4–5): \"We find that frontier model performance on GDPval is improving roughly linearly over time, and that the current best frontier models are approaching industry experts in deliverable quality. We analyze the potential for frontier models, when paired with human oversight, to perform GDPval tasks cheaper and faster than unaided experts.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-23",
    "archive_claims": [
     "LV-23"
    ]
   },
   {
    "lineage": "trend: roughly linear improvement",
    "field_claims": [
     "F26"
    ],
    "received": "F26 (B8, Figure 6 caption): \"Performance of OpenAI frontier models increased roughly linearly over time on the GDPval gold subset.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-23",
    "archive_claims": [
     "LV-23"
    ]
   },
   {
    "lineage": "models compared: aesthetics and accuracy",
    "field_claims": [
     "F27"
    ],
    "received": "F27 (B8, §3.1 ¶1): \"Claude Opus 4.1 was the best performing model on the GDPval gold subset, excelling in particular on aesthetics (e.g., document formatting, slide layout), while GPT-5 excelled in particular on accuracy (e.g., carefully following instructions, performing correct calculations)\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-23",
    "archive_claims": [
     "LV-23"
    ]
   },
   {
    "lineage": "why models lose: instructions, and the severity of failures",
    "field_claims": [
     "F28",
     "F29"
    ],
    "received": "F28 (B8, Figure 8 caption): \"Across models, experts most often preferred the human deliverable because models failed to fully follow instructions on GDPval tasks.\"; F29 (B8, A.2.6 ¶2 s2): \"Another roughly 29% of ratings were for bad or catastrophic (with roughly 3% of failures marked as catastrophic).\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-13",
    "archive_claims": [
     "LV-13"
    ]
   },
   {
    "lineage": "speed and cost: the expert's review and redo",
    "field_claims": [
     "F30",
     "F31",
     "F32",
     "F33",
     "F54"
    ],
    "received": "F30 (B8, §3.2 ¶1 s4–5): \"Here, an expert human samples from a model, reviews outputs, and if unsatisfactory, resamples and repeats. If no satisfactory output is obtained, the human completes the task themselves.\"; F31 (B8, A.2.1 Table 2, row gpt-5, and the sentence after it): \"gpt-5 39.0% 90x 1.12x 1.39x 474x 1.18x 1.63x … When incorporating time to review and redo work, the payoff from using a model shrinks.\"; F32 (B8, A.2.1, after Table 2, s2–4): \"We do not include consideration of the time taken to review a human professional deliverable, although this would commonly occur for tasks in GDPval (either self-review of the professional’s own work or review by a supervisor of a team member’s work). We also do not include the possibility that the human deliverable is also undesirable. One further limitation of this analysis is that it does not capture the cost of catastrophic mistakes, which can be disproportionately expensive in some domains.\"; F33 (B8, A.2.1, footnote 9): \"We are over-penalizing the model here, because the win rate after each completion likely goes up (because the professional will adjust the prompt to the model to fix the errors) and the review time also goes down as the professional gets more comfortable with the task.\"; F54 (B2, 'Publication' block, the paragraph headed 'Abstract'): \"Initial results indicate that frontier AI models are approaching the quality of work produced by human experts, with models able to perform certain professional tasks approximately 100 times faster and at a fraction of the cost compared to human experts.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-09",
    "archive_claims": [
     "LV-09"
    ]
   },
   {
    "lineage": "valuing the experts' time",
    "field_claims": [
     "F34",
     "F35"
    ],
    "received": "F34 (B8, A.2.1, footnotes 6 and 7): \"Because times were self-reported, it is possible that experts under-estimated or over-estimated time taken … Because our experts were recruited specifically for being highly experienced in their field, these wage estimates likely underestimate their true market cost.\"; F35 (B8, §2.3 s4): \"Each task’s dollar value was estimated by multiplying the average estimated completion time by median hourly wages for the corresponding occupation from OEWS data\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-09",
    "archive_claims": [
     "LV-09"
    ]
   },
   {
    "lineage": "task length",
    "field_claims": [
     "F36",
     "F37"
    ],
    "received": "F36 (B8, §1, bullet 'Long-horizon difficulty'): \"Tasks require an average of 7 hours of work for an expert professional to complete. On the high end, tasks span up to multiple weeks of work.\"; F37 (B8, A.2.5 s2): \"Win rates are highest for shorter tasks (0–2 hours) and decline steadily as completion time increases.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received",
    "archive_claims": []
   },
   {
    "lineage": "reasoning, context and scaffolding",
    "field_claims": [
     "F38"
    ],
    "received": "F38 (B3, Abstract s6): \"We also demonstrate that increased reasoning effort, increased task context, and increased scaffolding improves model performance on GDPval.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received",
    "archive_claims": []
   },
   {
    "lineage": "stated limits: size, scope, one-shot specification",
    "field_claims": [
     "F39",
     "F40",
     "F41",
     "F42"
    ],
    "received": "F39 (B8, §5, 'Dataset size'): \"The GDPval full set currently consists of only 44 occupations and 30 total tasks per occupation. It is therefore a limited, initial cut of knowledge work tasks, not a comprehensive evaluation of all possible occupational tasks.\"; F40 (B8, §5, 'Focus on self-contained knowledge work' s2–3): \"Manual labor and physical tasks are not included in the current version. Moreover, tasks that involve extensive tacit knowledge, access to personally identifiable information, use of proprietary software tools, or communication between individuals are out of scope for the current evaluation.\"; F41 (B8, §5, 'Tasks are precisely-specified and one-shot, not interactive' (heading and s1)): \"Tasks are precisely-specified and one-shot, not interactive: … For GDPval, we provide the full context of the task in the prompt, but in real life it often takes effort to figure out the full context of a task and understand what to work on.\"; F42 (B8, A.2.7 ¶2 s2; ¶3): \"We collected and graded GPT-5 completions with expert human graders and found the model’s performance was worse on under-specified prompts. … As a note: this experiment was run on an earlier version of the GDPval gold subset, and therefore the observed win rates do not match those in the main text of the paper.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside O771-04",
    "archive_claims": [
     "O771-04"
    ]
   },
   {
    "lineage": "the open gold subset and the held-out set",
    "field_claims": [
     "F43",
     "F44",
     "F57",
     "F58",
     "F67"
    ],
    "received": "F43 (B8, §4 s1 and s3): \"We open-source the prompts and reference files in our 220-task gold subset. … Please note that the tasks in the open sourced set have been scrubbed of information that could be used to identify the expert who wrote the task.\"; F44 (B8, A.6.3 s1): \"In the open-source set we mark 12 out of 220 tasks as ungradable due to limitations of the automated grader.\"; F57 (B5, dataset card): \"Dataset for GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks. … 220 real-world knowledge tasks across 44 occupations. … Each task consists of a text prompt and a set of supporting reference files.\"; F58 (B5, dataset viewer, columns and first row): \"deliverable_files listlengths 0 6 … rubric_pretty stringlengths 1.15k 15.8k … \"author_type\": \"human\"\"; F67 (B10, 'AI agent benchmark', the GDPval item, s4–6): \"A gold subset of 220 tasks is publicly available. The public set can be graded by a finetuned version of GPT-5. The held-out set of 1,100 tasks is only available internally in OpenAI, and external models cannot be submitted for grading on this set.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-18",
    "archive_claims": [
     "LV-18"
    ]
   },
   {
    "lineage": "win rate as metric: the human baseline replaceable",
    "field_claims": [
     "F45"
    ],
    "received": "F45 (B8, §1, bullet 'No “upper limit”' s2): \"Currently, we compare model outputs against a human expert baseline, but we could replace our baseline with increasingly strong models over time and keep evaluating.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received",
    "archive_claims": []
   },
   {
    "lineage": "versions and lifecycle",
    "field_claims": [
     "F47"
    ],
    "received": "F47 (B8, §1 ¶4 s6; §5 'Dataset size' s3): \"Future GDPval iterations will incorporate greater breadth, realism, interactivity, and contextual nuance. … We are expanding the dataset size.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-02, LV-19, LV-11, LV-20, LV-23",
    "archive_claims": [
     "LV-02",
     "LV-19",
     "LV-11",
     "LV-20",
     "LV-23"
    ]
   },
   {
    "lineage": "publication",
    "field_claims": [
     "F48",
     "F49"
    ],
    "received": "F48 (B8, header): \"License: CC BY 4.0 … arXiv:2510.04374v1 [cs.LG] 05 Oct 2025\"; F49 (B4, PDF metadata (pdfinfo, saved as B4-info.txt)): \"CreationDate: Thu Sep 25 17:03:48 2025 EDT\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-02, LV-19, LV-11, LV-20, LV-23",
    "archive_claims": [
     "LV-02",
     "LV-19",
     "LV-11",
     "LV-20",
     "LV-23"
    ]
   },
   {
    "lineage": "the operator and the authors",
    "field_claims": [
     "F50",
     "F51"
    ],
    "received": "F50 (B3, title and authors; B8 author footnote): \"View a PDF of the paper titled GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks, by Tejal Patwardhan and 18 other authors\"; F51 (B8, author block and footnote): \"Amelia Glaese Jerry Tworek OpenAI … Equal contribution. Correspondence to tejal@openai.com.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-02, LV-19, LV-11, LV-20, LV-23",
    "archive_claims": [
     "LV-02",
     "LV-19",
     "LV-11",
     "LV-20",
     "LV-23"
    ]
   },
   {
    "lineage": "derivative: GDPval-AA (Artificial Analysis)",
    "field_claims": [
     "F52",
     "F53",
     "F55"
    ],
    "received": "F52 (B2, header ¶1): \"GDPval-AA v2.1 is Artificial Analysis' evaluation framework for OpenAI's GDPval dataset. It tests AI models on real-world tasks across 44 occupations and 9 major industries. Models are given shell access and web browsing capabilities in an agentic loop via Stirrup to solve tasks, with Elo ratings derived from blind pairwise comparisons.\"; F53 (B2, FAQ 'How does GDPval-AA v2.1 decide which model did better?'): \"For each matchup, the two outputs are anonymized and an LLM judge picks a winner.\"; F55 (B2, leaderboard summary): \"Claude Opus 5.5 (Max, Default Fallback) scores the highest on GDPval-AA v2.1 with a score of 1866\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received",
    "archive_claims": []
   },
   {
    "lineage": "uptake: the dataset's downloads and models trained on it",
    "field_claims": [
     "F59"
    ],
    "received": "F59 (B5, 'Downloads last month'; 'Models trained or fine-tuned on openai/gdpval'): \"Downloads last month 271,657 … Browse 47 models trained on this dataset\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-18",
    "archive_claims": [
     "LV-18"
    ]
   },
   {
    "lineage": "derivative: the Inspect Evals implementation",
    "field_claims": [
     "F60",
     "F61",
     "F62"
    ],
    "received": "F60 (B7, Overview ¶1 s3–4; ¶2): \"The GDPval full set includes 1,320 specialized tasks (220 in the gold open-sourced set), each meticulously crafted and vetted by experienced professionals with over 14 years of experience on average from these fields. Every task is based on real work products, such as a legal brief, an engineering blueprint, a customer support conversation, or a nursing care plan. The version here only includes the 220 tasks in the gold open-sourced set.\"; F61 (B7, Overview ¶3; Judging ¶1): \"Scoring this eval involves saving the results as their own dataset, uploading it to huggingface, and submitting it to the OpenAI auto grader. … Note that OpenAI provides their automated grader model for free.\"; F62 (B7, Replication Notes): \"Average Score: 47.3% ± 6.2% … We note that the official score reported in the official GDPval leaderboard by OpenAI reports a lower score of 31.9% (wins only) and 34.6% (wins and ties).\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-07, LV-21",
    "archive_claims": [
     "LV-07",
     "LV-21"
    ]
   },
   {
    "lineage": "reception: the encyclopedia",
    "field_claims": [
     "F65",
     "F66"
    ],
    "received": "F65 (B9, Wikipedia: titles GDPval and GDPVal; article-space search): \"Wikipedia does not have an article with this exact name. … Results 1 – 2 of 2 … AI agent benchmark … Comparison of generative AI models\"; F66 (B10, 'AI agent benchmark', §Benchmarks, 'Economic' (the GDPval item), s1–3): \"GDPval: 1,320 tasks evaluating AI agents on real-world knowledge-work deliverables across 44 occupations from the top 9 U.S. GDP-contributing industries. Tasks made by professionals averaging 14 years of experience. The full evaluation is by blinded expert pairwise comparison against human expert deliverables, each comparison costing 1 human hour on average.\"",
    "receiving_text": "—",
    "translation": "untranslated: composed as received beside LV-25, CAP-1-09",
    "archive_claims": [
     "LV-25",
     "CAP-1-09"
    ]
   }
  ],
  "receiving_texts": "equal-terms case: every field claim composed as received, beside the archive claims it stands with"
 },
 "measure": {
  "rule": "v0.9 §4.10: KO_B sentences surviving in the KO at similarity ≥ 0.95, with positions (computed as apply_v09.py does: difflib.SequenceMatcher ratio, the best KO sentence for each KO_B sentence)",
  "v0.9": {
   "surviving": 0,
   "of": 16,
   "positions": [
    null,
    null,
    null,
    null,
    null,
    null,
    null,
    null,
    null,
    null,
    null,
    null,
    null,
    null,
    null,
    null
   ]
  },
  "opening_run": {
   "rule": "v0.9 §5.10 (equal-terms case): the field-only KO's sentences as the KO's opening run in their own order",
   "breach": false
  },
  "blocks": {
   "rule": "v0.9 §4.1 (equal-terms and mixed cases): the KO does not open with the field as a block or close with the archive as one; sentences typed by the ledgers of the claims they cite (F field-only, A archive-only, M both)",
   "sequence": "MFMMMMAMMMAAAFMMMAFMFAFFAMFMAM",
   "field_only_head": 0,
   "archive_only_tail": 0,
   "breach": false
  },
  "max_similarity": [
   0.317,
   0.443,
   0.305,
   0.415,
   0.483,
   0.233,
   0.085,
   0.428,
   0.612,
   0.219,
   0.509,
   0.076,
   0.223,
   0.324,
   0.239,
   0.179
  ],
  "ko_sentence_1_max_similarity_to_KO_B": 0.317
 },
 "homonyms": {
  "names": [],
  "basis": "The census row carries no disambiguation field. No card is a homonym and T composes none. 'GDPVal' (the Wikipedia variant tried) is a spelling. The distinct entities below are named by the field or the census beside the benchmark and are never senses of its name.",
  "distinct_entities": [
   "GDPval-AA v2.1 (Artificial Analysis's evaluation framework 'for OpenAI's GDPval dataset', B2; set beside GDPval by T and by the capture's reading, CAP-1-10)",
   "the Inspect Evals implementation of GDPval (UK AISI's Inspect Evals register, B7: the 220 open tasks, scored through OpenAI's auto grader)",
   "GDP, Gross Domestic Product (the measure the name and the sector selection take: B3, B8 §2.1)",
   "the Anthropic Economic Index (the census's neighbour; the paper cites it among the usage analyses it builds on, F5, LV-22); no /non row for it exists yet, and none is edited",
   "the 1,100-task held-out set and the gold subset (two parts of the one benchmark, B10 F67; B8 §4)",
   "Remote Labor Index and the Vals Index (listed beside GDPval in 'AI agent benchmark', B10, §Economic; not ledgered)",
   "'GDPval v0' (a page at openai.com/index/gdpval-v0 returned by WebSearch for the card's title; not on the rail, not in the field, not fetched)"
  ],
  "substitutes": {
   "T": []
  },
  "register": {
   "address_1": {
    "slug": "gdpval-aio-20261029",
    "obs_id": "OBS-ef64b7d44199"
   }
  }
 },
 "notes": [
  "One address, Google AI Overview (expanded), signed out, incognito, 2026-10-29: gdpval-aio-20261029 (OBS-ef64b7d44199, ADDR-d4b480b48aca, AXN:0AED.CAPTURE.🎨🌻🖐️🧪🕖🏛️), seven cards, 136 words in the answer body with its closer. A NEW address; the census record landed the same day and no card cites it (CAP-1-09).",
  "Fetching: every reachable page saved verbatim with curl on 2026-10-29 between 22:31 and 22:34 UTC with the agent 'Mozilla/5.0 (compatible; Labor-Vector-census)', raw and text with sha256 in field.json. Card URLs: B1, B2 and B4 by WebSearch on title and site; B3 from the arXiv number in its title (byte-identical to the census's saved copy, carried once as a card); B5 from the dataset id; B6 from a YouTube results page fetched with the same agent; B7 from the Inspect Evals index page's link. B8 copied from the census's saved copy (sha256 matches the census record); not refetched.",
  "Unreachable: B1, OpenAI's own page (openai.com/index/gdpval/, the census's own page): HTTP 403 with a challenge page, as for the census at 21:49; no other agent string or route was tried (brief §4). B6, the video: the watch page answered HTTP 429 and was not retried; admitted through its title. Their card text stands (F63, F64). Wikipedia's search answered 429 once and was refetched with the same agent after a pause (a rate limit, not a refusal of the agent).",
  "Wikipedia: no article on GDPval (raw and rendered HTTP 404 for 'GDPval' and 'GDPVal'; no redirect). An article-space search for the name returns two articles; 'AI agent benchmark' carries a GDPval item under §Benchmarks, 'Economic', ledgered as B10 (F66, F67) since it is the encyclopedia's only account of the entity; the article carries a primary-sources maintenance tag. 'Comparison of generative AI models' names only GDPval-AA, a distinct entity, and is not ledgered.",
  "The archive subset: no archive text names GDPval or arXiv:2510.04374 (D 0, R 0). The O-class pass returned 88 sentences in 44 deposits; reading every class match beyond it admitted 14 deposits with 21 claims (#771, #855, #199, #778, #482, #48, #637, #163, #623, #308, #533, #1545 at the first pass; #1616 and #1540 by the hop), with #1392, #1023 and #856 grouped as instances (S4). The hop (136 candidates) was screened by script for the class terms and each hit sentence read. As on the provenance rows, the O-class pass missed most of the deposits admitted by reading (#855, #778, #482, #48, #637 carry no category term in the sentence; reported, not fixed).",
  "Defects or out-of-date fields in the Labor Vector census record (reported, not fixed): (a) founded '2025-10-05 (arXiv v1)', coverage.period and count.as_of date the first version from arXiv; the card B4, the paper as a PDF on OpenAI's CDN, was created on 25 September 2025 (pdfinfo CreationDate, F49), ten days earlier, with the same text. (b) data_access gives 'prompts and reference files of the 220-task gold subset open-sourced' (after the paper §4); the dataset as served on Hugging Face (card B5) also carries the experts' deliverable files and the grading rubrics (deliverable_files, rubric_pretty, rubric_json, rubric items marked author_type 'human', F58), and 271,657 downloads in the last month with 47 models listed as trained on it (F59): the record is narrower than the served dataset. (c) data_access and count do not record that the other 1,100 tasks are held out, 'only available internally in OpenAI', and that external models cannot be submitted for grading on them (Wikipedia, B10, F67); the paper's §4 implies it ('We open-source the prompts and reference files in our 220-task gold subset'). (d) unchecked 'the open-sourced gold subset and the grading service': the cards now show both, the dataset (B5) and the grading route through 'the OpenAI auto grader', which 'OpenAI provides … for free' (B7, F61). (e) lifecycle 'later releases unrecorded': no new release of GDPval is in the field, but the field records an official leaderboard by OpenAI (B7, F62: GPT-5 at 31.9% wins only and 34.6% with ties) and two third-party derivatives (GDPval-AA v2.1, B2; the Inspect Evals implementation, changelog 2026-02-16, B7), none of which the record carries. (f) operator: the post is still unread; the same 403 answered the honest agent here (B1). (g) No field source contradicts a census finding or limit; the record's two internal differences in the paper (47.6% beside 'just over half'; 'at least 30' beside '30 total' tasks per occupation) are confirmed in B8 and B4. (h) The census row has no disambiguation field (as on the MMRS and provenance rows).",
  "Observations on the capture (reported, not fixed): findings[0] ('Every figure is the record's: 1,320 tasks, 30 per occupation, 44 occupations, 9 sectors, the 220-task gold subset, experts averaging 14 years') lists '30 per occupation' among the composition's figures; the composition carries no per-occupation count (T1-2 gives 1,320 tasks across 44 occupations). The figure is the record's, not T's. Every other finding is borne out by the field ledger: the blinding caveat (F20, CAP-1-02), the findings with their figures (F24, F25, F28, F31: 47.6%, 90x, 1.12x, CAP-1-03), the limits (F40, F41, F22, CAP-1-04) and the experts' terms (F14, CAP-1-05). The per_dim note that GDPval-AA is 'not in the records' holds: the census record does not carry it.",
  "Field-internal differences, composed under P_coh: Artificial Analysis's page (B2) carries under the paper's title a paragraph headed 'Abstract' that is not the paper's abstract (B3): it states that models can perform 'certain professional tasks approximately 100 times faster and at a fraction of the cost compared to human experts' (F54), where the paper's abstract states a potential 'when paired with human oversight' (F56) and its Table 2 gives GPT-5 90x naive, 1.12x tried once (F31). Inspect Evals (B7) gives the experts 'over 14 years of experience on average' (F60) where the paper gives 'an average of 14 years' (F13); and its GPT-5 replication scored 47.3% ± 6.2% against OpenAI's official 31.9% and 34.6% (F62). Wikipedia (B10) gives the open set's grader as 'a finetuned version of GPT-5' and each comparison as '1 human hour on average'; the paper gives a grader 'based on GPT-5-high' and comparisons taking 'over an hour'.",
  "Platform-specific record (BRIEF-NEW-ROW-LABOR.md §6): GDPval is a benchmark built and graded by its provider. The field states it at F22 (the grader based on GPT-5-high, agreeing less with experts on OpenAI's own models, 'models often favor their own responses'), F10 (GPT-4o classified the occupations as digital), F16/F17 (OpenAI models screened submissions) and F61 (open-set grading through OpenAI's auto grader); the census tags it 'NOT independent of a provider' (LV-21). The KO composes these at sentences 5, 10 and 15; E_F and the delta carry them.",
  "Composer disclosure: the composer is a Claude model (Opus 5.5). The field ranks Claude models twice: Claude Opus 4.1 best overall on the paper's gold subset (F27) and Claude Opus 5.5 first on Artificial Analysis's GDPval-AA v2.1 leaderboard as served on 2026-10-29 (F55). Both are carried at their sources' grade and attributed; neither is placed in the lede.",
  "Related: the archive's #1714 (Meaning Feudalism at the Discovery Layer, OpenAI's mathematics releases, October 2026) records a field's problem stock used as tests and the checking labor borne by mathematicians; it was reached by the hop and not admitted (a claim about another release by the same operator, not about GDPval or its classes). Neighbour rows: the census's neighbour, the Anthropic Economic Index, and the other two resources of the worked question (Microsoft's Working with AI) have no /non row yet; none is edited. The AIO row is datasets/negative-of-the-negative/v2/rows/google-ai-overviews.json (not edited)."
 ],
 "composition_v09": {
  "source": "this row (composed directly; no rebuild/negative-of-the-negative/v09/openai-gdpval.json)",
  "composer": "Claude (Opus 5.5), subagent, 2026-10-29; the composer is a Claude model, a product family the field ranks (F27, F55)",
  "defining_material_read": [
   {
    "what": "Labor Vector census record, row",
    "file": "/home/claude/fleet/laborvector/data/labor-census/census.json (rows[slug == 'openai-gdpval'], organizing_question, intersections, schema_note, unread_sources)",
    "read": "whole"
   },
   {
    "what": "Labor Vector worked questions",
    "file": "/home/claude/fleet/laborvector/data/labor-census/questions.json (questions[id == 'working-with-ai'])",
    "read": "the question that names the entity"
   },
   {
    "what": "Labor Vector census record, rendered",
    "file": "/home/claude/fleet/laborvector/aggregators/openai-gdpval/index.html",
    "read": "text, for agreement with the row"
   },
   {
    "what": "census sources",
    "file": "/home/claude/fleet/laborvector/data/labor-census/sources/openai-gdpval--arxiv.*, openai-gdpval--paper-html.*, openai-gdpval--post.*; sources/manifest.json",
    "read": "whole (the paper's main text, §1–§6, and Appendix A.1–A.2, A.5–A.6 read through)"
   },
   {
    "what": "capture record",
    "file": "data/EA-WG-CAPTURES-01.json (gdpval-aio-20261029)",
    "read": "transcript, transcript_raw, cite_list, d, reading, findings, analysis, q_kind, per_note, per_dim, originator"
   },
   {
    "dep": 771,
    "file": "data/texts/AXN-02D1-text.md",
    "read": "front matter, abstract, core thesis, §1–§4, §8–§10"
   },
   {
    "dep": 855,
    "file": "data/texts/AXN-0364-text.md",
    "read": "the passage 'Every Factor Removes Friction'"
   },
   {
    "dep": 199,
    "file": "data/texts/AXN-0341-text.md",
    "read": "the abstract"
   },
   {
    "dep": 778,
    "file": "data/texts/AXN-02DA-text.md",
    "read": "§4.2"
   },
   {
    "dep": 1616,
    "file": "data/texts/AXN-06A6-text.md",
    "read": "§3"
   },
   {
    "dep": 1540,
    "file": "data/texts/AXN-063A-text.md",
    "read": "the closing gloss"
   },
   {
    "dep": 482,
    "file": "data/texts/AXN-012D-text.md",
    "read": "§10"
   },
   {
    "dep": 48,
    "file": "data/texts/AXN-01D5-text.md",
    "read": "the abstract"
   },
   {
    "dep": 637,
    "file": "data/texts/AXN-01ED-text.md",
    "read": "§VIII"
   },
   {
    "dep": 163,
    "file": "data/texts/AXN-02FE-text.md",
    "read": "§II.5"
   },
   {
    "dep": 623,
    "file": "data/texts/AXN-01DA-text.md",
    "read": "§6"
   },
   {
    "dep": 308,
    "file": "data/texts/AXN-006F-text.md",
    "read": "§B–§C"
   },
   {
    "dep": 533,
    "file": "data/texts/AXN-0168-text.md",
    "read": "§IX"
   },
   {
    "dep": 1545,
    "file": "data/texts/AXN-0644-text.md",
    "read": "§2 and §5"
   },
   {
    "what": "the specification",
    "file": "data/texts/AXN-0700-text.md",
    "read": "§0–§11 whole; Appendix A"
   },
   {
    "what": "the briefs",
    "file": "rebuild/negative-of-the-negative/v09/BRIEF-NEW-ROW-AGGREGATOR.md, BRIEF-NEW-ROW-PROVENANCE.md, BRIEF-NEW-ROW-LABOR.md, BRIEF.md (Compose)",
    "read": "whole"
   },
   {
    "what": "exemplar rows",
    "file": "datasets/negative-of-the-negative/v2/rows/olmotrace.json (with its ledgers, config, reading)",
    "read": "structure and texts"
   }
  ],
  "cuts": [],
  "notes": "equal case; KO 30 sentences, all 67 field and 56 archive claims cited; P body 334 words; KO_B 16 sentences, 0 surviving at ≥ 0.95"
 },
 "field_by_origin": {
  "cards": [
   "B1",
   "B2",
   "B3",
   "B4",
   "B5",
   "B6",
   "B7"
  ],
  "own_pages": [
   "B8"
  ],
  "wikipedia": [
   "B9",
   "B10"
  ],
  "census_third_party": [],
  "unreachable": [
   "B1",
   "B6"
  ],
  "title_only": [
   "B1",
   "B6"
  ],
  "wikipedia_state": "no article: 'GDPval' and 'GDPVal' are both absent (raw wikitext HTTP 404; rendered page HTTP 404 'Wikipedia does not have an article with this exact name'), no redirect; a full-text search of article space for \"GDPval\" (one HTTP 429, refetched) returns 'Results 1 – 2 of 2': 'AI agent benchmark', whose GDPval item under §Benchmarks, 'Economic' is ledgered as B10, and 'Comparison of generative AI models', which names GDPval-AA only (2026-10-29)",
  "sources": [
   {
    "id": "B1",
    "origin": "card",
    "role": "Overview card 1",
    "card": "OpenAI — Measuring the performance of our models on real-world tasks | OpenAI",
    "url": "https://openai.com/index/gdpval/",
    "kind": "Overview card (OpenAI's own page; the census's own page openai-gdpval--post; unreachable: HTTP 403)",
    "files": {
     "B1-403.html": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B1-403.html",
      "sha256": "07799b36bc6e4a9610c6a63d597b232519e345705e884aa764ed1f4f6325bea2"
     },
     "B1-census-403.html": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B1-census-403.html",
      "sha256": "b0a7ee49cca701cbc40625e050aaca241349f4b61557dd8184d22f8b095996cb"
     }
    },
    "fetched": {
     "unreachable": true,
     "tries": [
      {
       "at": "2026-10-29T22:31:14Z",
       "http": 403,
       "sha256": "07799b36bc6e4a9610c6a63d597b232519e345705e884aa764ed1f4f6325bea2",
       "agent": "Mozilla/5.0 (compatible; Labor-Vector-census)",
       "result": "a 9,771-byte challenge page (the word 'challenge' four times in it), saved as B1-403.html"
      },
      {
       "at": "2026-10-29T21:49:14Z",
       "http": 403,
       "sha256": "b0a7ee49cca701cbc40625e050aaca241349f4b61557dd8184d22f8b095996cb",
       "by": "the census (sources/manifest.json openai-gdpval--post)",
       "result": "the census's own fetch: 403, text_chars 0; copied as B1-census-403.html"
      }
     ],
     "note": "URL found by WebSearch on the card's title and site (the result 'openai.com/index/gdpval'); it is the census's own page for the entity (openai-gdpval--post). One fetch with the brief's honest agent string returned HTTP 403 with a challenge page; per the brief (§4) no other agent string or route was tried. The census's 403 of 21:49 is copied beside it. The card's title and snippet stand as its only text."
    },
    "names_operator": "yes ('OpenAI' as the card's site and in its title)",
    "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, laborvector, Labor Vector, machinemediation, EA-WG-CAPTURES-01 in the saved text, card title and snippet)",
    "note": "carried once, as the card: the census's own page is the same URL, unread by the census and here",
    "card_title": "Measuring the performance of our models on real-world tasks | OpenAI",
    "card_snippet": "GDPval: Evaluation measuring model performance on 1,320 specialized tasks across 44 knowledge work occupations from 9 U.S. GDP-contributing industries (threshol...",
    "card_note": null,
    "surfaced_by_T": [
     {
      "slug": "gdpval-aio-20261029",
      "card": 1
     }
    ]
   },
   {
    "id": "B2",
    "origin": "card",
    "role": "Overview card 2",
    "card": "Artificial Analysis — GDPval-AA v2.1 Benchmark - Artificial Analysis",
    "url": "https://artificialanalysis.ai/evaluations/gdpval-aa",
    "kind": "Overview card (a third party's evaluation framework on OpenAI's dataset, a distinct entity)",
    "files": {
     "B2.html": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B2.html",
      "sha256": "e320651e4d74e7af6a2efb611bb8deefaf46ad7d62e8958d67f2782c1b40e2f1"
     },
     "B2.txt": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B2.txt",
      "sha256": "654fea0c38d45ae683d2b35720c625d66316f3b09eb3965a7253b7c270f6bc9e"
     }
    },
    "fetched": {
     "at": "2026-10-29T22:31:16Z",
     "http": 200,
     "url": "https://artificialanalysis.ai/evaluations/gdpval-aa",
     "raw": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B2.html",
     "raw_sha256": "e320651e4d74e7af6a2efb611bb8deefaf46ad7d62e8958d67f2782c1b40e2f1",
     "text": [
      "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B2.txt"
     ]
    },
    "names_operator": "yes ('OpenAI's GDPval dataset'; the paper's 19 authors listed)",
    "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, laborvector, Labor Vector, machinemediation, EA-WG-CAPTURES-01 in the saved text, card title and snippet)",
    "note": "URL found by WebSearch on the card's title and site; the page's <title> is 'GDPval-AA v2.1 Benchmark | Artificial Analysis'. The page carries a paragraph headed 'Abstract' under 'Publication' that is not the paper's abstract (compare B3 Abstract): it is quoted as Artificial Analysis's text (F54), attributed.",
    "card_title": "GDPval-AA v2.1 Benchmark - Artificial Analysis",
    "card_snippet": "GDPval-AA v2.1 Benchmark LegacyAgenticKnowledge WorkWritingReasoningInstruction FollowingTool UseMultimodalBusinessFinanceLegalMedicalHumanities GDPval-AA v2. 1...",
    "card_note": null,
    "surfaced_by_T": [
     {
      "slug": "gdpval-aio-20261029",
      "card": 2
     }
    ]
   },
   {
    "id": "B3",
    "origin": "card",
    "role": "Overview card 3",
    "card": "arXiv — [2510.04374] GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks",
    "url": "https://arxiv.org/abs/2510.04374",
    "kind": "Overview card (the paper's abstract page; the census's own page openai-gdpval--arxiv, byte-identical)",
    "files": {
     "B3.html": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B3.html",
      "sha256": "904393337a8b91ee7b387d650a2eb4c89dd29d4040b7c228c8b57e25fb991835"
     },
     "B3.txt": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B3.txt",
      "sha256": "6e4ebc1a8814c4ba0a4f2ecf49a6e3f4b44b9ef8b252b1a95643f39d60db517a"
     }
    },
    "fetched": {
     "at": "2026-10-29T22:31:17Z",
     "http": 200,
     "url": "https://arxiv.org/abs/2510.04374",
     "raw": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B3.html",
     "raw_sha256": "904393337a8b91ee7b387d650a2eb4c89dd29d4040b7c228c8b57e25fb991835",
     "text": [
      "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B3.txt"
     ],
     "note": "byte-identical to the census's saved copy openai-gdpval--arxiv (sha256 904393337a8b91ee…, read 2026-10-29), so carried once, as a card"
    },
    "names_operator": "yes (the 19 authors; correspondence and affiliation on the paper)",
    "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, laborvector, Labor Vector, machinemediation, EA-WG-CAPTURES-01 in the saved text, card title and snippet)",
    "note": "URL from the arXiv number in the card's title.",
    "card_title": "[2510.04374] GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks",
    "card_snippet": "[2510.04374] GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks Scheduled maintenance: Today - Please check here for updates:Stat...",
    "card_note": null,
    "surfaced_by_T": [
     {
      "slug": "gdpval-aio-20261029",
      "card": 3
     }
    ]
   },
   {
    "id": "B4",
    "origin": "card",
    "role": "Overview card 4",
    "card": "OpenAI — gdpval: evaluating ai model performance - OpenAI",
    "url": "https://cdn.openai.com/pdf/d5eb7428-c4e9-4a33-bd86-86dd4bcf12ce/GDPval.pdf",
    "kind": "Overview card (the paper as a PDF on OpenAI's CDN, created 2025-09-25: the earliest fetched instance of the paper)",
    "files": {
     "B4.pdf": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B4.pdf",
      "sha256": "37c7604bc9da68cd8d9345ad16938164fe985182ad723fcb09e9415f84dc25d4"
     },
     "B4.txt": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B4.txt",
      "sha256": "f1fa9ebb67ac6ec3d4184040e369a0d93049b2ebfa8b65f200c06c5e0ac4e23f"
     },
     "B4-layout.txt": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B4-layout.txt",
      "sha256": "381209fb880cdb070074dea53b6fdde2e65ef6a70611f9e01f5df7e5a406a895"
     },
     "B4-info.txt": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B4-info.txt",
      "sha256": "b0063449a685d2bc7207c37157b98a28da9b6e9795ce3ebd3fe53011834f3ae7"
     }
    },
    "fetched": {
     "at": "2026-10-29T22:31:21Z",
     "http": 200,
     "url": "https://cdn.openai.com/pdf/d5eb7428-c4e9-4a33-bd86-86dd4bcf12ce/GDPval.pdf",
     "raw": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B4.pdf",
     "raw_sha256": "37c7604bc9da68cd8d9345ad16938164fe985182ad723fcb09e9415f84dc25d4",
     "text": [
      "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B4.txt",
      "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B4-info.txt"
     ],
     "note": "text by pdftotext (B4.txt; B4-layout.txt with -layout); B4-info.txt is pdfinfo's output (CreationDate 2025-09-25)"
    },
    "names_operator": "yes ('OpenAI' under the author list)",
    "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, laborvector, Labor Vector, machinemediation, EA-WG-CAPTURES-01 in the saved text, card title and snippet)",
    "note": "URL found by WebSearch on the card's title and snippet; the snippet's sentence ('Our paper introduces the first version of GDPval, a benchmark evaluating AI model performance on real-world economically valuable tasks. GDPval covers the top 9…') is in the PDF's §1 (F2). The paper's text in the PDF matches the arXiv v1 HTML (B8) at every quote checked (field_claims[*].also_in).",
    "card_title": "gdpval: evaluating ai model performance - OpenAI",
    "card_snippet": "Our paper introduces the first version of GDPval, a benchmark evaluating AI model performance on real-world economically valuable tasks. GDPval covers the top 9...",
    "card_note": null,
    "surfaced_by_T": [
     {
      "slug": "gdpval-aio-20261029",
      "card": 4
     }
    ]
   },
   {
    "id": "B5",
    "origin": "card",
    "role": "Overview card 5",
    "card": "Hugging Face — openai/gdpval · Datasets at Hugging Face",
    "url": "https://huggingface.co/datasets/openai/gdpval",
    "kind": "Overview card (the operator's dataset page)",
    "files": {
     "B5.html": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B5.html",
      "sha256": "1de40771dafd9ab66e4debd6e6eeea9a1f4584979fc0189c4ae04ebdd0c55f9e"
     },
     "B5.txt": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B5.txt",
      "sha256": "a87b38c91f3d9a86100681e4c51fb04f63beef009692d471d69da681dc1784c0"
     }
    },
    "fetched": {
     "at": "2026-10-29T22:31:21Z",
     "http": 200,
     "url": "https://huggingface.co/datasets/openai/gdpval",
     "raw": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B5.html",
     "raw_sha256": "1de40771dafd9ab66e4debd6e6eeea9a1f4584979fc0189c4ae04ebdd0c55f9e",
     "text": [
      "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B5.txt"
     ]
    },
    "names_operator": "yes ('openai' as the dataset's owner, 'OpenAI 43.8k')",
    "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, laborvector, Labor Vector, machinemediation, EA-WG-CAPTURES-01 in the saved text, card title and snippet)",
    "note": "URL from the card's title (the dataset id). No license is stated on the served page.",
    "card_title": "openai/gdpval · Datasets at Hugging Face",
    "card_snippet": "Dataset: openai/gdpval on Hugging Face (Parquet format, <1K size, 220 rows, 2.29 GB total file size). Modalities: Text. Libraries: Datasets, pandas, Polars, xet...",
    "card_note": null,
    "surfaced_by_T": [
     {
      "slug": "gdpval-aio-20261029",
      "card": 5
     }
    ]
   },
   {
    "id": "B6",
    "origin": "card",
    "role": "Overview card 6",
    "card": "YouTube·Vivek Haldar — GDPval: An AI benchmark for real-world tasks (video, 11:23)",
    "url": "https://www.youtube.com/watch?v=R1O-lUKlAys",
    "kind": "Overview card (video; title only)",
    "files": {
     "B6-search.html": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B6-search.html",
      "sha256": "f83812ae662370236ca9659f1ffd6a3db56ce5adec83552b542736fb22e15371"
     },
     "B6-429.html": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B6-429.html",
      "sha256": "ffa2a4f6dd5b641a4346a85fd5d71b2064ea8e45dd2ee53ff5ebed6a62edc475"
     }
    },
    "fetched": {
     "unreachable": true,
     "tries": [
      {
       "at": "2026-10-29T22:32:00Z",
       "http": 200,
       "url": "https://www.youtube.com/results?search_query=GDPval+An+AI+benchmark+for+real-world+tasks+Vivek+Haldar",
       "sha256": "f83812ae662370236ca9659f1ffd6a3db56ce5adec83552b542736fb22e15371",
       "result": "a YouTube results page giving the video id R1O-lUKlAys under the card's title, by Vivek Haldar, '11 minutes, 23 seconds'"
      },
      {
       "at": "2026-10-29T22:32:11Z",
       "http": 429,
       "url": "https://www.youtube.com/watch?v=R1O-lUKlAys",
       "sha256": "ffa2a4f6dd5b641a4346a85fd5d71b2064ea8e45dd2ee53ff5ebed6a62edc475",
       "result": "HTTP 429 (3,272 bytes); not retried"
      }
     ],
     "note": "WebSearch did not find the video; its URL came from a YouTube results page fetched with the honest agent. The watch page answered 429 and was not retried by another route; no transcript was fetched. Admitted through its title (the card carries no snippet)."
    },
    "names_operator": "no",
    "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, laborvector, Labor Vector, machinemediation, EA-WG-CAPTURES-01 in the saved text, card title and snippet)",
    "card_title": "GDPval: An AI benchmark for real-world tasks",
    "card_snippet": null,
    "card_note": "video, 11:23",
    "surfaced_by_T": [
     {
      "slug": "gdpval-aio-20261029",
      "card": 6
     }
    ]
   },
   {
    "id": "B7",
    "origin": "card",
    "role": "Overview card 7",
    "card": "GitHub Pages documentation — GDPval – Inspect Evals - GitHub Pages",
    "url": "https://ukgovernmentbeis.github.io/inspect_evals/evals/gdpval/index.html",
    "kind": "Overview card (a third party's implementation of the gold subset, a distinct entity)",
    "files": {
     "B7.html": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B7.html",
      "sha256": "db6622b4af38050431dc894d0e47a6d91f1640b67e1f103a78c310c5842a6e19"
     },
     "B7.txt": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B7.txt",
      "sha256": "7116b9eb5f5b66bac773e42fc1c7a5f354da254833ba5cedcbd596f217d6a48b"
     }
    },
    "fetched": {
     "at": "2026-10-29T22:31:48Z",
     "http": 200,
     "url": "https://ukgovernmentbeis.github.io/inspect_evals/evals/gdpval/index.html",
     "raw": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B7.html",
     "raw_sha256": "db6622b4af38050431dc894d0e47a6d91f1640b67e1f103a78c310c5842a6e19",
     "text": [
      "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B7.txt"
     ]
    },
    "names_operator": "yes ('the OpenAI auto grader', 'the official GDPval leaderboard by OpenAI')",
    "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, laborvector, Labor Vector, machinemediation, EA-WG-CAPTURES-01 in the saved text, card title and snippet)",
    "note": "WebSearch did not return the page; its URL came from the link './evals/gdpval/index.html' on the Inspect Evals index (https://ukgovernmentbeis.github.io/inspect_evals/, HTTP 200, not saved). The page's <title> is 'GDPval – Inspect Evals'; the card's snippet is the page's first paragraph.",
    "card_title": "GDPval – Inspect Evals - GitHub Pages",
    "card_snippet": "GDPval measures model performance on economically valuable, real-world tasks across 44 occupations. The first version of this evaluation spans 44 occupations se...",
    "card_note": null,
    "surfaced_by_T": [
     {
      "slug": "gdpval-aio-20261029",
      "card": 7
     }
    ]
   },
   {
    "id": "B8",
    "origin": "own_page",
    "role": "the census's own page (the paper's HTML, arXiv v1); not surfaced by T",
    "card": "arXiv — GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks (HTML, v1)",
    "url": "https://arxiv.org/html/2510.04374v1",
    "kind": "own page, saved by the census (sources/openai-gdpval--paper-html, read 2026-10-29)",
    "files": {
     "B8.html": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B8.html",
      "sha256": "1d4c4f2494000318567a5b82504d7154119aa88a40ba11c18568a4dea718c5d9"
     },
     "B8.txt": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B8.txt",
      "sha256": "bf50b5c5608086dca1e39492d365a66fb164d90a3b60cf4d0aa39ea3d8f21ddd"
     }
    },
    "fetched": {
     "by": "the census",
     "at": "2026-10-29T21:56:25Z",
     "http": 200,
     "url": "https://arxiv.org/html/2510.04374v1",
     "raw": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B8.html",
     "raw_sha256": "1d4c4f2494000318567a5b82504d7154119aa88a40ba11c18568a4dea718c5d9",
     "census_sha256": "1d4c4f2494000318567a5b82504d7154119aa88a40ba11c18568a4dea718c5d9",
     "read": "2026-10-29",
     "text": [
      "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B8.txt"
     ],
     "note": "copied from the census's saved copy (html gunzipped; .txt copied); the sha256 matches the census record; not refetched"
    },
    "names_operator": "yes (author block 'OpenAI'; 'Correspondence to tejal@openai.com')",
    "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, laborvector, Labor Vector, machinemediation, EA-WG-CAPTURES-01 in the saved text, card title and snippet)",
    "note": "Admitted per the operator's ruling ('their own pages'); not surfaced by T, though it is the same work as cards B3 and B4.",
    "surfaced_by_T": []
   },
   {
    "id": "B9",
    "origin": "wikipedia",
    "role": "the entity's Wikipedia article (none)",
    "card": "Wikipedia — GDPval (no article)",
    "url": "https://en.wikipedia.org/w/index.php?title=GDPval&action=raw",
    "kind": "Wikipedia (no article)",
    "files": {
     "B9.wiki": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B9.wiki",
      "sha256": "8a1e7da066ff8c2b1695e4f70fc7ad874960b2dc0e421cc2cd0b775ab4713056"
     },
     "B9.html": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B9.html",
      "sha256": "d7a7c9e4258e64985cf61361fd6130d9a6abe2420218e1b51d21cac65a97af5f"
     },
     "B9-variant.wiki": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B9-variant.wiki",
      "sha256": "fef283f04f017ec84a3dba258508b544e6cd5e6dc649ce737ac11a25bdbf8db7"
     },
     "B9-search-429.html": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B9-search-429.html",
      "sha256": "d1584927871bf524c00010d9f1699c1de2b0b8afcae814cb37e4a7807a2c11b9"
     },
     "B9-search.html": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B9-search.html",
      "sha256": "1b0fdaea45329611aedc998dc8702e2b427f9f06d3edb79c3ad3abc65842f233"
     },
     "B9.txt": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B9.txt",
      "sha256": "13e34e1adcf88d985964351feb60afffe32f81cb135120f90ce87a7224fa29f0"
     }
    },
    "fetched": {
     "tries": [
      {
       "at": "2026-10-29T22:32:33Z",
       "url": "https://en.wikipedia.org/w/index.php?title=GDPval&action=raw&redirect=no",
       "http": 404
      },
      {
       "at": "2026-10-29T22:32:35Z",
       "url": "https://en.wikipedia.org/wiki/GDPval?redirect=no",
       "http": 404,
       "result": "'Wikipedia does not have an article with this exact name.'"
      },
      {
       "at": "2026-10-29T22:32:50Z",
       "url": "https://en.wikipedia.org/w/index.php?title=GDPVal&action=raw&redirect=no",
       "http": 404
      },
      {
       "at": "2026-10-29T22:32:53Z",
       "url": "Special:Search \"GDPval\" (article space)",
       "http": 429,
       "result": "rate-limited; refetched with the same agent"
      },
      {
       "at": "2026-10-29T22:33:21Z",
       "url": "https://en.wikipedia.org/w/index.php?search=%22GDPval%22&title=Special:Search&profile=advanced&fulltext=1&ns0=1",
       "http": 200,
       "result": "'Results 1 – 2 of 2': AI agent benchmark; Comparison of generative AI models"
      }
     ],
     "raw_sha256": "d7a7c9e4258e64985cf61361fd6130d9a6abe2420218e1b51d21cac65a97af5f",
     "text": [
      "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B9.txt"
     ]
    },
    "names_operator": "no",
    "names_archive": "no",
    "surfaced_by_T": [],
    "note": "No article and no redirect under either spelling. The name occurs in two articles' text: 'AI agent benchmark' (ledgered as B10, its GDPval item) and 'Comparison of generative AI models' (which names GDPval-AA, a distinct entity; not ledgered). B9.txt is a summary of the codes and notices, its quoted strings taken from the saved responses."
   },
   {
    "id": "B10",
    "origin": "wikipedia",
    "role": "Wikipedia: the GDPval item in 'AI agent benchmark', §Benchmarks, 'Economic'",
    "card": "Wikipedia — AI agent benchmark (the GDPval item)",
    "url": "https://en.wikipedia.org/wiki/AI_agent_benchmark",
    "kind": "Wikipedia (a section of another article that names the entity; no redirect)",
    "files": {
     "B10.wiki": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B10.wiki",
      "sha256": "87edcbedc1e0e6854b31b3b30d04dc45d304f89f0ba97924da21967ba19fce11"
     },
     "B10.html": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B10.html",
      "sha256": "3be91030981967b083c9e8ced0c33a995377c77d3eddd38cf2e72e3dd3ea39df"
     },
     "B10.txt": {
      "path": "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B10.txt",
      "sha256": "ec6893fc1c953a169d14d117c0513b63395b11cb40e37c820666cda7995c4791"
     }
    },
    "fetched": {
     "tries": [
      {
       "at": "2026-10-29T22:33:34Z",
       "url": "https://en.wikipedia.org/w/index.php?title=AI_agent_benchmark&action=raw",
       "http": 200,
       "sha256": "87edcbedc1e0e6854b31b3b30d04dc45d304f89f0ba97924da21967ba19fce11"
      },
      {
       "at": "2026-10-29T22:33:38Z",
       "url": "https://en.wikipedia.org/wiki/AI_agent_benchmark",
       "http": 200,
       "sha256": "3be91030981967b083c9e8ced0c33a995377c77d3eddd38cf2e72e3dd3ea39df"
      }
     ],
     "raw_sha256": "3be91030981967b083c9e8ced0c33a995377c77d3eddd38cf2e72e3dd3ea39df",
     "text": [
      "datasets/negative-of-the-negative/v2/intake/openai-gdpval-20261029/B10.txt"
     ]
    },
    "names_operator": "yes ('only available internally in OpenAI'; the arXiv citation's authors)",
    "names_archive": "no (none of the Crimson Hexagonal Archive, Alexanarch, Lee Sharks, laborvector, Labor Vector, machinemediation, EA-WG-CAPTURES-01 in the saved text, card title and snippet)",
    "surfaced_by_T": [],
    "note": "The article carries a maintenance tag ('This article relies excessively on references to primary sources', October 2026). Its GDPval item cites only the arXiv paper; its held-out-set sentence (1,100 tasks internal) and 'finetuned version of GPT-5' are not in the field's other texts in those words."
   }
  ]
 }
}