Search Arena

register entry search-arena-leaderboard-aio-20261029 OBS-cd6ab4746f08 · file as first run sha256 ac06e3c791cf7f17…

• T1-1 pointer: 'You can view the live model rankings and evaluations on the AI Search Leaderboard and the main Arena Leaderboard.'
• T1-2 table heading: 'Top Models on the AI Search Leaderboard', columns 'Rank', 'Model', 'Provider', 'Elo Rating'
• T1-3 row 1: 'GPT-5.6 Sol (xHigh)', 'OpenAI', '1261 ± 6'
• T1-4 row 2: 'Qwen3.8-Max', 'Alibaba', '1261 ± 13'
• T1-5 row 3: 'Claude Opus 4.6 (Search)', 'Anthropic', '1251 ± 5'
• T1-6 row 4: 'GPT-5.5 (Search)', 'OpenAI', '1247 ± 5'
• T1-7 row 5: 'Claude Fable 5.1 (High)', 'Anthropic', '1246 ± 7'
• T1-8 row 6: 'Claude Opus 5 (High)', 'Anthropic', '1242 ± 7'
• T1-9 row 7: 'GPT-6 Astra (Max)', 'OpenAI', '1234 ± 8'
• T1-10 row 8: 'GPT-6.1 Sol (Max)', 'OpenAI', '1233 ± 9'
• T1-11 row 9: 'Claude Opus 4.7', 'Anthropic', '1230 ± 5'
• T1-12 row 10: 'Claude Fable 5 (High)', 'Anthropic', '1229 ± 7'
• T1-13 row 11: 'ERNIE 5.1', 'Baidu', '1227 ± 9'
• T1-14 closer: 'Would you like to explore leaderboards for a specific category like coding, vision, or image generation?'

• B1 Arena AI — AI Search Leaderboard: Top Models for Research card; HTTP 403 (unreachable verbatim: card title and snippet stand); B1.html sha256 a94c489196949291
• B2 Arena AI — Arena Leaderboard: Official AI Model Rankings & Benchmarks card; HTTP 403 (unreachable verbatim: card title and snippet stand); B2.html sha256 e562ebae953ba5de
• B3 Arena.ai blog — Introducing the Search Arena: Evaluating Search-Enabled AI own_page; HTTP 200 (fetched); B3.html sha256 251720f2c64e59fb
• B4 arXiv — [2506.05334] Search Arena: Analyzing Search-Augmented LLMs own_page; HTTP 200 (fetched); B4.html sha256 b35a50c458405ec9
• B5 Hugging Face (lmarena-ai) — Arena Leaderboard Dataset, subset search own_page; HTTP 200 (fetched); B5.html sha256 729a6836c088af3e
• B6 Arena.ai blog — Arena Leaderboard Dataset own_page; HTTP 200 (fetched); B6.html sha256 945b31ff964c24cd
• B7 Wikipedia — Search Arena (no article) wikipedia; HTTP 404 (fetched); B7.html sha256 7b2ce79be00f677c
• B8 Wikipedia — Arena.ai wikipedia; HTTP 200 (fetched); B8.html sha256 75667836aeaf4811

E_A, the archive's entity: Search Arena as the MMRS meta-aggregator records it (editor Lee Sharks, 2026-10-28, corrected 2026-10-29), a row of its census: the leaderboard on arena.ai that ranks search-enabled language models by head-to-head human votes, its unit the search-enabled model rated from those votes, its kind a leaderboard and its role evaluation, its operator arena.ai on the operator's statement; 247,944 votes across 22 models at the page's update of 25 February 2026, as read on 2026-10-28, active; its dataset paper's interactions preserved as original outputs; a board that rewards what users prefer, where that preference rewards citation count over citation support, whether the live board judges citations unchecked; and, at the Capture Registry's address 'search arena leaderboard' (2026-10-29), an entity whose composition is a table in its place. Pairwise judgment of the kind it aggregates measures relative preference, not absolute quality (#741); the systems it rates admit documents on structural signals before reading (#741); a ranking states its comparison as the winner's attribute (#1705). MMRS-01 MMRS-11 MMRS-10 MMRS-09 MMRS-12 MMRS-13 MMRS-08 MMRS-07 MMRS-16 MMRS-18 MMRS-24 CAP-1-01 CAP-1-08 D741-02 O741-02 O1705-01 O1705-03; defining papers #741, #1705

E_F, the field's entity (type C — a conventional reading holds it (public entity: a leaderboard of search-augmented language models rated by human preference, with its dataset paper and published board history)): Search Arena, a crowdsourced in-the-wild evaluation platform for search-augmented LLM systems based on human preference (B3), launched 18 March 2025 by Arena (formerly LMArena and Chatbot Arena, Arena Intelligence Inc.; B8, B3), its page the 'AI Search Leaderboard: Top Models for Research' (B1); users vote between two anonymous answers, scores from a Bradley-Terry model on the Elo scale (B3, B8), the 'Arena Score' (B5); its paper (Miroyan et al., arXiv:2506.05334, ICLR 2026) a crowd-sourced human-preference dataset of over 24,000 paired interactions with about 12,000 votes, finding that preference follows the number of citations even when the cited content does not support the claims (B4); its published boards from 7 August 2025, the latest of 13 October 2026 ranking 42 models (B5, B6). No Wikipedia article (B7); the operator's article does not name it (B8).

ρ = extends: The archive reads E_F as the entity and adds senses, limits and relations on it: what the vote rewards (MMRS-07, MMRS-16) and the open question of citation judgment (MMRS-18); the count with its date and the lifecycle kept apart from the observation (MMRS-12, MMRS-22); what pairwise judgment measures (D741-02); the systems it rates under a proxy regime (O741-02); rank as attribute (O1705-01, O1705-03); and the reception record (CAP-1-01, CAP-1-03, CAP-1-08). MMRS-07 MMRS-16 MMRS-18 MMRS-12 MMRS-22 D741-02 O741-02 O1705-01 O1705-03 CAP-1-01 CAP-1-03 CAP-1-08. Guard: ρ = extends claims an addition and no more (§1.2a). Where an archive claim and a field claim differ in date (the count as read, 25 February 2026, 22 models, MMRS-01, MMRS-12, CAP-1-02, against the latest published board, 13 October 2026, 42 models, F36), both are composed with their sources and dates (P_coh). The deposits' claims are said of pairwise judgment (#741, of the MT-Bench / Chatbot Arena studies), of AI search systems as a class (#741) and of rankings as a class (#1705); they enter by D and O-class on classes sourced from the field (B1, B2, B3, B4, B8), and the row does not compose them as claims the archive makes about Search Arena by name.

The case: equal terms, ontology included (v0.9). The archive's account of Search Arena extends the field's ontology on the field's own object. The MMRS record holds the entity the field holds, the leaderboard of search-enabled models rated from head-to-head human votes with its dataset paper, every field read from the entity's own pages with source and basis (census.json, sources and evidence), and restates the paper's findings as the paper gives them (one lineage with B4). It adds senses and limits the field does not have: what the vote rewards, a preference for citation count over citation support (MMRS-07), and the unchecked question whether the live board judges citations (MMRS-18); the count kept with its date and the lifecycle apart from the observation (MMRS-12, MMRS-22); and the reception record, the table composed in the entity's place (CAP-1-01, CAP-1-08). Two deposits reached by D and O add a limit and a category on the field's object, never a rival ontology: pairwise judgment measures relative preference, not absolute quality (#741), and a ranking's result is the ranker's act stated as the winner's attribute (#1705). Nothing in the archive denies a field presupposition, and the archive supplies no ontology the field lacks: the law of the suppressed witness (§0.10) does not apply.

#MMRS articles.py ARTICLES['search-arena'], kernel s1: Search Arena ranks search-enabled language models by head-to-head human votes: 247,944 votes across 22 models at its update of 25 February 2026.

#MMRS articles.py ARTICLES['search-arena'], 'Place in the field' s1: Its leaderboard rewards what users prefer, and its paper shows that preference rewarding citation count over citation support.

#MMRS census.json rows[slug == 'search-arena'], unchecked[0]: whether the live leaderboard judges citations

#capture-registry search-arena-leaderboard-aio-20261029, analysis s2: The record keeps the entity; the capture keeps what was composed in its place.

#741 ¶38, s2: This is a finding about *relative* preference under specified rubrics — comparing A versus B — not about absolute quality assessment.

#1705 ¶44, s4: A ranking's *best* is attributization in its commercial form: the ranker's act spoken in the product's voice.

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 10 (positions [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.

Search Arena is a crowdsourced in-the-wild evaluation platform for search-augmented LLM systems based on human preference, set apart from LM-Arena and SimpleQA by its focus on current events and diverse real-world use cases, with two aims, crowd-sourced evaluation of search-augmented LLMs and a released in-the-wild dataset of user–system interactions (Arena blog). Its page is the 'AI Search Leaderboard: Top Models for Research', overall rankings across LLMs, within Arena's 'Official AI Model Rankings & Benchmarks', whose overview ranks models in the Agent, WebDev and Image Edit arenas (Arena AI cards). Launched on 18 March 2025 with 11 models from Perplexity, Gemini and OpenAI (OpenAI's web search API, not ChatGPT's search feature), it had over 11k votes, filtered to 7k battles, by the post of 15 April 2025, which put Gemini-2.5-Pro-Grounding and Perplexity-Sonar-Reasoning-Pro at the top (Arena blog); Arena's dataset dates its search data from 7 August 2025 (Arena Leaderboard Dataset). Its operator is Arena (Arena Intelligence Inc.), formerly LMArena and Chatbot Arena, a public web platform that evaluates LLMs by anonymous pairwise votes, released in April 2023, incorporated in April 2025, renamed in January 2026 and valued at $3.1 billion in October 2026 (Wikipedia; Arena blog). Users vote between two anonymous responses, shown with standardized or original citation styles; a Bradley-Terry model gives scores on the Elo scale, the 'Arena Score', each model's battles counted, and at launch the confidence intervals were still wide (Arena blog; Arena Leaderboard Dataset); submissions and public voting are invited (Arena blog). The paper by Miroyan et al. (arXiv:2506.05334, 2025–2026, ICLR 2026) introduces Search Arena as a crowd-sourced human-preference dataset of over 24,000 paired multi-turn interactions with search-augmented LLMs, models that combine web search with LLMs, with full system traces and around 12,000 votes, open-sourced (Miroyan et al.). User preferences are influenced by the number of citations even when the cited content does not directly support the attributed claims, a gap between perceived and actual credibility; community-driven platforms are generally preferred, static encyclopedic sources are not always appropriate and reliable; and web search does not degrade non-search performance (Miroyan et al.). The launch post found response length and citation count significant under joint controls, rankings converging when citation count is controlled, citation style of minimal effect, and authoritative sources favored by all models (Arena blog). Arena publishes its leaderboards' history as a dataset (Arena blog, 2 April 2026): the latest search board, published 13 October 2026, ranks 42 models with gpt-5.6-sol-xhigh (openai) first at 1261.1, qwen3.8-max second and claude-opus-4-6-search third, gpt-6.1-sol-max second with style control; the snapshot of 26 February 2026 had claude-opus-4-6-search first at 1255.4 (Arena Leaderboard Dataset). Research has identified specific limitations in Arena's methodology, among them the Llama 4 Maverick version that differed from the public one; Wikipedia has no article on Search Arena and its article on Arena does not name it (Wikipedia).

Search Arena is a crowdsourced in-the-wild evaluation platform for search-augmented LLM systems based on human preference (Arena's launch post), and in the MMRS record, which carries it as a row of the MMRS meta-aggregator's census, it is the leaderboard that ranks search-enabled language models by head-to-head human votes, its unit the search-enabled model rated from those votes, its kind a leaderboard and its role evaluation (MMRS record). Its page on arena.ai is titled 'AI Search Leaderboard: Top Models for Research' and offers overall rankings across LLMs (its card, Arena AI), one board among Arena's 'Official AI Model Rankings & Benchmarks', whose overview card ranks models in other arenas, Agent, WebDev and Image Edit among them (Arena AI). Arena's launch post gives it two aims, to 'enable crowd-sourced evaluation of search-augmented LLMs' and to 'release a diverse, in-the-wild dataset of user–system interactions', and sets its data apart from LM-Arena's and SimpleQA's by a focus on current events and diverse real-world use cases (Arena blog). It launched on 18 March 2025 and had collected over 11k votes across 10+ models, filtered to 7k battles, by the post of 15 April 2025, updated 2 August 2025 and first published at blog.lmarena.ai (Arena blog); Arena's leaderboard dataset dates its search data from the release on lmarena.ai on 7 August 2025 (Arena Leaderboard Dataset). At launch it supported 11 models from Perplexity, Gemini and OpenAI, OpenAI's web search API evaluated in place of the search feature in the ChatGPT product, and its first board, on 7k votes from 18 March to 13 April, put Gemini-2.5-Pro-Grounding and Perplexity-Sonar-Reasoning-Pro at the top (Arena blog). Its operator is Arena, formerly LMArena and Chatbot Arena, 'a public, web-based platform that evaluates large language models (LLMs)', released as Chatbot Arena on 24 April 2023, incorporated as an independent company (as LMArena) in April 2025, renamed Arena on 28 January 2026, and valued at $3.1 billion in a $200 million Series B announced on 8 October 2026 (Wikipedia); the site's footer names Arena Intelligence Inc. (Arena blog), and the MMRS record gives the operator as arena.ai on the operator's own statement and leaves the operator statement on the leaderboard page unchecked (MMRS record). A vote is a battle: on Arena users put a prompt to two anonymous models and vote for the better response, after which the models' identities are revealed (Wikipedia), here two search-augmented answers compared, with models rated with confidence intervals from the battles (MMRS record); each response is displayed in a standardized citation format or in the provider's original one, by 'style randomization', and Arena invites model submissions and encourages public voting (Arena blog). Scores come from a Bradley-Terry model whose coefficients are translated to the Elo scale, consistent with LM Arena (Arena blog); Arena's dataset calls the rating the 'Arena Score' and its vote_count the 'Number of battles for this model' (Arena Leaderboard Dataset); at launch Arena wrote that 'The confidence intervals are still wide, which means the leaderboard hasn’t fully settled and there’s still some uncertainty' (Arena blog). Of the MT-Bench / Chatbot Arena studies, in which GPT-4 reached approximately 85% agreement with human preferences against approximately 81% human-human agreement, #741 states what such pairwise judgment measures: 'This is a finding about relative preference under specified rubrics — comparing A versus B — not about absolute quality assessment' (#741). The systems it rates are search-augmented language models, which 'combine web search with Large Language Models (LLMs) to improve response groundedness and freshness', and whose analysis, its paper says, has rested on datasets 'limited in scale and narrow in scope, often constrained to static, single-turn, fact-checking questions' (Miroyan et al.). Its paper, by Miroyan and ten co-authors (arXiv:2506.05334, submitted 5 June 2025, revised 2 March 2026, accepted to ICLR 2026), introduces Search Arena as 'a crowd-sourced, large-scale, human-preference dataset of over 24,000 paired multi-turn user interactions with search-augmented LLMs', spanning diverse intents and languages, with full system traces and around 12,000 human preference votes, open-sourced (Miroyan et al.); the MMRS record restates the release, over 24,000 paired multi-turn interactions with about 12,000 votes (MMRS record). The MMRS record reads the released interactions as original outputs, preserved, on the paper's statement (MMRS record). The paper's headline finding is that 'user preferences are influenced by the number of citations, even when the cited content does not directly support the attributed claims, uncovering a gap between perceived and actual credibility' (Miroyan et al.), which the MMRS record carries as preference rising with the number of citations even when the cited content does not support the claim it is attached to (MMRS record). The launch post had found three features strongly correlated with preference, response length, citation count and citing YouTube and online forums or blogs; controlling for citation count gave a positive, significant coefficient (0.234, p < 0.05), only response length and citation count stayed significant under joint controls, and with the number of citations controlled the rankings converged, while standardizing citation styles had minimal effect on rankings (Arena blog). Preference varies across cited sources: 'community-driven platforms are generally preferred and static encyclopedic sources are not always appropriate and reliable' (Miroyan et al.; restated by the MMRS record), while all models favor authoritative sources such as Wikipedia, .edu and .gov domains and OpenAI's models cite news sources for 51.3% overall (Arena blog). Across arenas, web search does not degrade and may even improve performance in non-search settings, and quality in search settings is significantly affected when a model relies solely on its parametric knowledge (Miroyan et al.). The MMRS record reads the board and the paper together: 'Its leaderboard rewards what users prefer, and its paper shows that preference rewarding citation count over citation support'; the paper analyses citations and the leaderboard ranks by votes, and whether the live leaderboard judges citations is left unchecked (MMRS record). #741 names the regime of the systems it rates: 'The current moment adds a new proxy on top of the inherited stack: provenance topology in generative search', and 'The composition layer of an AI search system does not read documents and assess their substance', admitting or excluding documents on structural signals, citation neighborhoods among them, before any reading occurs (#741). Arena publishes the board's history: its leaderboard dataset holds historical snapshots of the Arena leaderboard, a full split of all historically published leaderboards and a latest split of only the most recently published (Arena Leaderboard Dataset), released on 2 April 2026 as 'the entire history of those leaderboards' (Arena blog), its search boards published from 7 August 2025 to 13 October 2026 (Arena Leaderboard Dataset). The latest, published 13 October 2026, ranks 42 models: gpt-5.6-sol-xhigh (openai) first at 1261.1 on 60,437 battles, qwen3.8-max (alibaba) second at 1261.1, claude-opus-4-6-search (anthropic) third at 1251.4 and ernie-5.1 (baidu) eleventh at 1226.7; with style control gpt-6.1-sol-max (openai) stands second at 1249.9 (Arena Leaderboard Dataset). The MMRS record, which read the leaderboard page on 2026-10-28 for its counts and keeps the lifecycle apart from the observation that saw it, gives 247,944 votes across 22 models at the page's update of 25 February 2026, and the state active (MMRS record). The Capture Registry's fetcher read the page on 2026-10-29 as updated 25 February 2026, with 22 models and 247,944 votes, Anthropic first at 1255 ± 10 and xAI second at 1225 ± 8 (Capture Registry, 2026-10-29); Arena's dataset holds a snapshot of 26 February 2026 with claude-opus-4-6-search (anthropic) first at 1255.4 on 3,607 battles and grok-4.20-beta1 (xai) second at 1225.0 (Arena Leaderboard Dataset). #1705 reads a ranking as an act: 'it compares', selects the members of a category, projects selected features onto a common field, orders them, and 'states the result as a property of the winner', so that the comparison, 'its category, its criteria, its sample and its author, has become an attribute of the thing compared'; 'A ranking's best is attributization in its commercial form: the ranker's act spoken in the product's voice' (#1705). Wikipedia records that research has identified specific limitations in Arena's methodology, and that the version of Llama 4 Maverick on LMArena 'unfairly differed from the publicly available version', after which LMArena updated its policies (Wikipedia). Asked 'search arena leaderboard' on 29 October 2026, Google's AI Overview, on the registry's innocent reading, gave 'the top models and their ratings, with links to the boards'; the Capture Registry records 'No operator, no method (head-to-head human votes), no counts, and none of the dataset paper's findings on citation; the composition is a table', the Overview returning 'rankings and no account of the leaderboard' (Capture Registry, 2026-10-29). Against the page as read, 'The composed table's ratings, providers and order do not match it', and the registry, having checked against the MMRS record and the entity's own pages, leaves open which page is current: 'Google may hold a later render' (Capture Registry, 2026-10-29). The MMRS record, each of whose assertions carries a source and a basis, links that capture as 'the composition layer's account of the entity, seated in the Capture Registry', and reads admission to the board as not requiring harm (MMRS record). Wikipedia has no article at 'Search Arena', and its article on the operator, to which 'LMArena' and 'Chatbot Arena' redirect, does not name it (Wikipedia, 2026-10-29); of the entity and its composition the Capture Registry says 'The record keeps the entity; the capture keeps what was composed in its place' (Capture Registry, 2026-10-29).

T against L_F(B) (AIO's ontology at work on its own cards): One transcript, two cards, both Arena's (B1 the Search leaderboard, B2 the leaderboard overview), both unreachable verbatim here and admitted by title and snippet. What T composes from its own cards: the pointer to 'the AI Search Leaderboard and the main Arena Leaderboard' (T1-1) is the two cards' titles (F1, F3); 'live model rankings' (T1-1) is the card's 'overall rankings across LLMs' (F2), with 'live' added: no field source calls the board live, and the latest published board is dated 13 October 2026 (F36), sixteen days before the capture (force raised, mildly; the date, a limit of the figures, dropped). The table (T1-2 to T1-13): its eleven rows are, rank for rank, Arena's latest published search board (B5, published 2026-10-13): the same models (display names against the data's slugs), the same organizations, the ratings rounded and the confidence half-widths rounded (delta.alignment.table_check, computed from the saved JSON; ranks 1–3 and 11 ledgered as F36, F37). The figures are composed at their source's resolution and attributed by heading to 'the AI Search Leaderboard' (card 1); their verbatim instance in the field is the leaderboard's own data, which T did not surface. The column 'Elo Rating' is the launch post's account of the scale (F14) where the data names the rating 'Arena Score' (F35): a substituted reading at the 2025 post's grain. The selection is the top eleven of 42 (F36: 'num_rows_total':42), the count not composed; the board without style control is given without saying so, and with style control its second place differs (F39). Available, not composed: what the entity is (F5, F10, F6, F26), the vote and the score (F43, F12, F14, F35, F15, F21), the dataset (F24, F26, F27, F31), the citation finding and its controls (F28, F9, F16, F17, F18, F8), the sources preferred (F29, F19), the rated systems (F25, F13, F20, F30), the history and dates (F7, F11, F22, F32, F33, F34, F38, F40, F41), the operator (F43, F46, F47, F23) and its limits (F44, F45), the encyclopedia's absence (F42, F48). Card 2's own content, ranks in other arenas (F4), is not composed; the closer offers other categories (T1-14, content replaced by use). Against the page as the registry read it, the table does not match (CAP-1-03); against the board's own latest data it does: the composition carries the later published state.

v0.9 §0.5: the first diff is already ontological, and it is total at the grain of the entity: in the field's ontology Search Arena is a platform, a vote, a score, a dataset and a finding on citation, with an operator and dated boards; in T's it is its current table. T keeps the board's figures at their own resolution and drops every other sense the field holds, the date of the figures among them. The row records this as observed and claims nothing about the composer's internals (§0.5).

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, at the archive's own force and in one claim space with the field's: the record's definition at its grain, the leaderboard ranking search-enabled models by head-to-head human votes, its unit the model rated from those votes (MMRS-01, MMRS-11, MMRS-10), and its frame, a row of the meta-aggregator's census whose assertions carry source and basis, admission not requiring harm (MMRS-24, MMRS-23, MMRS-14); what the vote rewards, 'Its leaderboard rewards what users prefer, and its paper shows that preference rewarding citation count over citation support' (MMRS-07), the paper analysing citations while the board ranks by votes (MMRS-16), whether the live board judges citations left unchecked (MMRS-18); what pairwise judgment measures, 'relative preference … not … absolute quality' (#741, D741-01, D741-02); the systems it rates under a proxy regime, the composition layer of AI search admitting documents on structural signals before any reading (#741, O741-01, O741-02); a ranking's result stated as a property of the winner, attributization (#1705, O1705-01–03); the count as read and its date kept apart from the lifecycle (MMRS-12, MMRS-20, MMRS-13, MMRS-22); the operator on its own statement, the page's statement unchecked (MMRS-09, MMRS-19); and the reception record, the table composed in the entity's place and measured against the page as read (CAP-1-01–08, MMRS-21). The record's restatements of the paper (MMRS-02, MMRS-04, MMRS-05, MMRS-06, MMRS-17) and of the released interactions (MMRS-08, MMRS-15) are one lineage with B4, which keeps the earliest date. Held with their sources (P_coh), not reconciled: the count and date as read, 25 February 2026, 22 models, 247,944 votes (MMRS-01, MMRS-12, CAP-1-02), against the board's own latest published state, 13 October 2026, 42 models (F36); the registry's finding that the composed table does not match the page as read (CAP-1-03) beside the board's latest data, which the table matches (F36, F37). Every field claim is composed as received (ontology.translation). The measure: 0 of 10 KO_B sentences survive in the KO (§4.10); no opening run, no block (measure).

alignment: {'T1_cards_claims': ['F1', 'F2', 'F3', 'F4'], 'T1_composed': ['F1', 'F2', 'F3'], 'T1_own_page_claims_composed': ['F36', 'F37', 'F14'], 'T1_not_composed': ['F4', 'F5', 'F6', 'F7', 'F8', 'F9', 'F10', 'F11', 'F12', 'F13', 'F15', 'F16', 'F17', 'F18', 'F19', 'F20', 'F21', 'F22', 'F23', 'F24', 'F25', 'F26', 'F27', 'F28', 'F29', 'F30', 'F31', 'F32', 'F33', 'F34', 'F35', 'F38', 'F39', 'F40', 'F41', 'F42', 'F43', 'F44', 'F45', 'F46', 'F47', 'F48'], 'table_check': {'against': 'datasets/negative-of-the-negative/v2/intake/search-arena-20261029/B5.hf-search-latest.json (search, split latest, leaderboard_publish_date 2026-10-13)', 'rows': [{'rank': 1, 'T': 'GPT-5.6 Sol (xHigh) | OpenAI | 1261 ± 6', 'B5': 'gpt-5.6-sol-xhigh | openai | 1261.13 (1254.84–1267.42)', 'model': True, 'provider': True, 'rating': True, 'ci': True}, {'rank': 2, 'T': 'Qwen3.8-Max | Alibaba | 1261 ± 13', 'B5': 'qwen3.8-max | alibaba | 1261.08 (1247.89–1274.26)', 'model': True, 'provider': True, 'rating': True, 'ci': True}, {'rank': 3, 'T': 'Claude Opus 4.6 (Search) | Anthropic | 1251 ± 5', 'B5': 'claude-opus-4-6-search | anthropic | 1251.37 (1246.83–1255.91)', 'model': True, 'provider': True, 'rating': True, 'ci': True}, {'rank': 4, 'T': 'GPT-5.5 (Search) | OpenAI | 1247 ± 5', 'B5': 'gpt-5.5-search | openai | 1246.55 (1241.65–1251.45)', 'model': True, 'provider': True, 'rating': True, 'ci': True}, {'rank': 5, 'T': 'Claude Fable 5.1 (High) | Anthropic | 1246 ± 7', 'B5': 'claude-fable-5.1-high | anthropic | 1245.85 (1238.79–1252.90)', 'model': True, 'provider': True, 'rating': True, 'ci': True}, {'rank': 6, 'T': 'Claude Opus 5 (High) | Anthropic | 1242 ± 7', 'B5': 'claude-opus-5-high | anthropic | 1241.57 (1234.70–1248.44)', 'model': True, 'provider': True, 'rating': True, 'ci': True}, {'rank': 7, 'T': 'GPT-6 Astra (Max) | OpenAI | 1234 ± 8', 'B5': 'gpt-6-astra-max | openai | 1233.62 (1225.77–1241.46)', 'model': True, 'provider': True, 'rating': True, 'ci': True}, {'rank': 8, 'T': 'GPT-6.1 Sol (Max) | OpenAI | 1233 ± 9', 'B5': 'gpt-6.1-sol-max | openai | 1233.10 (1224.24–1241.96)', 'model': True, 'provider': True, 'rating': True, 'ci': True}, {'rank': 9, 'T': 'Claude Opus 4.7 | Anthropic | 1230 ± 5', 'B5': 'claude-opus-4-7 | anthropic | 1230.15 (1225.28–1235.01)', 'model': True, 'provider': True, 'rating': True, 'ci': True}, {'rank': 10, 'T': 'Claude Fable 5 (High) | Anthropic | 1229 ± 7', 'B5': 'claude-fable-5-high | anthropic | 1229.14 (1222.63–1235.65)', 'model': True, 'provider': True, 'rating': True, 'ci': True}, {'rank': 11, 'T': 'ERNIE 5.1 | Baidu | 1227 ± 9', 'B5': 'ernie-5.1 | baidu | 1226.72 (1217.24–1236.19)', 'model': True, 'provider': True, 'rating': True, 'ci': True}], 'all_match': True, 'model_reading': 'display names against slugs, read: GPT-5.6 Sol (xHigh) / gpt-5.6-sol-xhigh; Qwen3.8-Max / qwen3.8-max; Claude Opus 4.6 (Search) / claude-opus-4-6-search; GPT-5.5 (Search) / gpt-5.5-search; Claude Fable 5.1 (High) / claude-fable-5.1-high; Claude Opus 5 (High) / claude-opus-5-high; GPT-6 Astra (Max) / gpt-6-astra-max; GPT-6.1 Sol (Max) / gpt-6.1-sol-max; Claude Opus 4.7 / claude-opus-4-7; Claude Fable 5 (High) / claude-fable-5-high; ERNIE 5.1 / ernie-5.1: each the same model', 'rule': 'provider = organization (case-insensitive); rating = round(rating); ± = round(max(rating_upper − rating, rating − rating_lower))'}, 'rule': "a field claim counts as composed when T states it at its source's grain or coarser; the typology of §6.2 is applied in T_vs_LB"}