GDPval

register entry gdpval-aio-20261029 OBS-ef64b7d44199 · file as first run sha256 0f182b74c748e78a…

• 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?'

• B1 OpenAI — Measuring the performance of our models on real-world tasks | OpenAI unreachable: HTTP 403 at 2026-10-29T22:31:14Z; HTTP 403 at 2026-10-29T21:49:14Z; card title and snippet stand
• B2 Artificial Analysis — GDPval-AA v2.1 Benchmark - Artificial Analysis card; B2.html sha256 e320651e4d74e7af, HTTP 200 at 2026-10-29T22:31:16Z
• B3 arXiv — [2510.04374] GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks card; B3.html sha256 904393337a8b91ee, HTTP 200 at 2026-10-29T22:31:17Z
• B4 OpenAI — gdpval: evaluating ai model performance - OpenAI card; B4.pdf sha256 37c7604bc9da68cd, HTTP 200 at 2026-10-29T22:31:21Z
• B5 Hugging Face — openai/gdpval · Datasets at Hugging Face card; B5.html sha256 1de40771dafd9ab6, HTTP 200 at 2026-10-29T22:31:21Z
• B6 YouTube·Vivek Haldar — GDPval: An AI benchmark for real-world tasks (video, 11:23) unreachable: HTTP 200 at 2026-10-29T22:32:00Z; HTTP 429 at 2026-10-29T22:32:11Z; title only
• B7 GitHub Pages documentation — GDPval – Inspect Evals - GitHub Pages card; B7.html sha256 db6622b4af380504, HTTP 200 at 2026-10-29T22:31:48Z
• B8 arXiv — GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks (HTML, v1) own page (census copy, read 2026-10-29); B8.html sha256 1d4c4f2494000318, HTTP 200 at 2026-10-29T21:56:25Z
• B9 Wikipedia — GDPval (no article) wikipedia: HTTP 404; HTTP 404; HTTP 404; HTTP 429; HTTP 200; B9.txt sha256 13e34e1adcf88d98
• B10 Wikipedia — AI agent benchmark (the GDPval item) wikipedia: HTTP 200; HTTP 200; B10.txt sha256 ec6893fc1c953a16

E_A, the archive's entity: 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) 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

E_F, the field's entity (type C — a conventional reading holds it (public entity: a task benchmark, its dataset and the paper that presents it)): 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)

ρ = extends: 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. 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.

The case: equal terms, ontology included (v0.9). 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.

#771 Abstract s2: Both are silent on a third dimension: the *direction* in which a synthetic system's labor flows relative to the task its user commissioned.

#771 §9 'Falsification, negative cases, and method', 'Identification condition': These metrics yield a single reproducible value only on transcripts where the task vector is *exogenous and fixed* — a one-shot commission.

#855 § 'Every Factor Removes Friction', ¶ 'Benchmarks measure means, not tails', s1–3: Benchmarks measure means, not tails.

#1545 §2 'Formal class-position test', ¶ after the formula, s2–4: This is deliberately a **relation**, not an essence.

#LV census.json rows[slug == 'openai-gdpval'].lv_intersection: 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.

#LV questions.json questions[id == 'working-with-ai'].gap: 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.

#capture-registry gdpval-aio-20261029, analysis s1: A capability benchmark is composed at its specification and stops before its results.

equal-terms case: every field claim composed as received, beside the archive claims it stands with. Reading the received column recovers the field (§5.9).

v0.9 §4.10: KO_B sentences surviving in the KO at similarity ≥ 0.95, with positions (computed as apply_v09.py does: difflib.SequenceMatcher ratio, the best KO sentence for each KO_B sentence): v0.9 0 of 16 (positions [None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None]; composed first under v0.9, so there is no v0.7 object to measure against). Opening-run check (§5.10): no breach.

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).

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).

T against L_F(B) (AIO's ontology at work on its own cards): 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).

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.

L_F(B) against L_A(B ∪ A) (the intervention): 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)"}