CodeSOTA

register entry codesota-aio-20261029 OBS-df84f10ecc75 · file as first run sha256 b95619e65027a50f…

• T1-1 definition: 'a unified routing layer and common API for artificial intelligence tasks spanning code, text, documents, speech, and vision' (marker: the homepage card)
• T1-2 model selection: 'Compares quality, latency, and cost using benchmark observations and Pareto frontier comparisons.'
• T1-3 execution: 'Provides execution for code and text, with planned support for documents, speech, and vision.'
• T1-4 infrastructure: 'Offers versioned recipes, automatic fallbacks, and usage accounting.'
• T1-5 task families, headed 'Supported Task Families': documents 'OCR, handwriting recognition, invoice processing, and parsing'; speech 'Transcription, text-to-speech, and realtime voice'; coding 'Code generation, repair, and AI agents'
• T1-6 closer: 'Would you like to know more about how CodeSOTA compares specific AI models or how to integrate its API into your workflow?'

• B1 Codesota — CodeSOTA — One Routing Layer for AI Tasks fetched 2026-10-29 05:06:12 UTC (server date) with curl -sSL -m 30 -A 'Mozilla/5.0'; HTTP 200 (no redirect); html sha256 451f427143f609484368e1871c069b355f8e81b29912b63f0f0d05fd1228af72, text sha256 1f18f20d1e16396483352fa6c21766a701144706b62a3819f24c6cc49a2338b6; response headers in B1.headers.txt
• B2 CodeSOTA — Papers With Code Alternative: SOTA Leaderboards and Archived Data (own page, /papers-with-code) fetched 2026-10-29 05:06:13 UTC (server date) with curl -sSL -m 30 -A 'Mozilla/5.0'; HTTP 200 (no redirect); html sha256 fdab8031a94ec8c17c1a4d58bcbcdd82fa0f1407e56bed1f75371a546799c337, text sha256 b6ff9cf805ec4ddf557cd3b6da64ca5ea7816945efb35e3db957a8ac27ab2771; response headers in B2.headers.txt
• B3 CodeSOTA — Papers with Code Shutdown Date: Meta July 2025 Timeline (own page, /papers-with-code/shutdown) fetched 2026-10-29 05:06:16 UTC (server date) with curl -sSL -m 30 -A 'Mozilla/5.0'; HTTP 200 (no redirect); html sha256 47f9989318fbaa765939420443185a3e661e942661d10276280938cea2eec433, text sha256 4c0409c529ce5c4c5ab1c58d7aafc3f1d45ceec1e21ae2af7f58e223b75bc0b8; response headers in B3.headers.txt
• B4 CodeSOTA — About Codesota — an independent registry of ML benchmarks (own page, /about) fetched 2026-10-29 05:06:17 UTC (server date) with curl -sSL -m 30 -A 'Mozilla/5.0'; HTTP 200 (no redirect); html sha256 af09fca9a27ce47f9d4bb62784c0bd1353cd6cabb1e07a76e6153540f9d6b702, text sha256 aa74b6a44c5e502e4747d3309915942b629232832484b860413ab4daae68f263; response headers in B4.headers.txt
• W1 no article wikipedia: no article; raw wikitext 404 for five capitalisations, rendered page 404, full-text and insource searches empty (2026-10-29)

E_A, the archive's entity: the same service, in the MMRS meta-aggregator's record: CodeSOTA, by Kacper Wikiel for Fabryka AI in Warsaw, a routing layer for AI tasks on its homepage and a successor to Papers with Code on its /papers-with-code/ pages, two faces at two addresses and the census's clearest case of a successor claim made in the name's own address space; a database in the role of discovery whose unit is a routed AI task, related to Papers with Code as claims_succession_to and kept apart from it and from its revival, its own counts and the provenance of its scores unchecked; observed by the Capture Registry at its name, where the composition layer composed the homepage's face and left the succession unsaid, and the record was corrected to give the homepage's face first MMRS-01 MMRS-13 MMRS-08 MMRS-11 MMRS-15 MMRS-20 MMRS-24 MMRS-29 MMRS-18 CAP-1-02 CAP-1-04; defining papers

E_F, the field's entity (type C — a conventional reading holds it (public entity: a service, presented at three of its own addresses as a routing layer for AI tasks, the successor to Papers with Code and a registry of benchmarks)): CodeSOTA (codesota.com), 'By Fabryka AI', built and maintained by Kacper Wikiel in Warsaw, launched 10 December 2025: on its homepage one routing layer for AI tasks, selecting models across its registry on benchmark evidence (quality, latency, cost, the Pareto frontier) with execution for code and text and other families planned; on its /papers-with-code/ pages 'the v2 replacement for what PWC did', 'the calmer, stricter successor', though 'There is no single drop-in replacement'; on its about page 'an independent registry of machine-learning benchmarks', a registry and not a leaderboard, its unit a dated task–dataset–metric row with a verification tier, its data open JSON under CC BY 4.0; refusing the class aggregator (B1–B4)

ρ = extends: the archive reads E_F as the entity and adds senses on the field's own object: the two faces as an observation and its reading as the clearest case of a successor claim in the name's address space, the typed relation of succession, the claimants kept apart, the record's category and unit, what it leaves unchecked, the record kept apart from the entity, and the composition layer's observed treatment of the entity MMRS-13 MMRS-08 MMRS-20 MMRS-24 MMRS-29 MMRS-11 MMRS-15 MMRS-17 MMRS-18 MMRS-22 CAP-1-02 CAP-1-04. Guard: ρ = extends claims an addition and no more: MMRS-08 and CAP-1-04 are readings (interpretation), not findings that the entity misrepresents itself; MMRS-17 and MMRS-18 leave the counts and the provenance of the scores unchecked; where an archive claim meets a field claim (MMRS-15 beside F59; MMRS-08 and MMRS-11 beside F45, F66; MMRS-01 beside F57) both are composed under P_coh in the one claim space

The case: equal terms, ontology included (v0.9). The archive's records hold the field's entity as the entity. The MMRS record describes CodeSOTA from its own pages, each field with a source and a basis (operator, observed), and denies no presupposition of the field's: CodeSOTA is a service its pages describe. What the record adds is on the field's object: the observation that the entity presents two faces at two addresses and the reading of it as the census's clearest case of a successor claim made in the name's own address space; the relation typed; the name's claimants kept apart; its own category and unit; what it leaves unchecked; and the Capture Registry's observation of the composition layer at the entity's name, which composed the homepage's face and left the succession unsaid. The ontology the archive uses is available in the field itself: the entity's own pages present the router, the successor and the registry at three addresses, and the dispatch relates the registry to the router ('the registry is routable', F43). Where an archive claim meets a field claim (the unit, the class, the maker) both are composed under P_coh. The law of the suppressed witness does not apply (§0.10: not where the field holds the ontology). Not mixed.

#MMRS articles.py ARTICLES['codesota'], kernel s1: CodeSOTA, by Kacper Wikiel for Fabryka AI in Warsaw, is a routing layer for AI tasks on its homepage ("One Routing Layer for AI Tasks") and a successor to Papers with Code on its /papers-with-code/ pages ("Codesota is the v2 replacement for what PWC did").

#MMRS census.json rows[slug == 'codesota'], note (basis: observed): two faces at two addresses: the homepage presents a routing layer and names Papers with Code only in a footer link; its /papers-with-code/ pages present it as 'the v2 replacement for what PWC did', rebuilding that infrastructure 'independently'

#MMRS articles.py ARTICLES['codesota'], 'Place in the field' s1: It is the clearest case in the census of a successor claim made in the name's own address space, and of an entity with two faces at two addresses: Google's Overview composes the homepage's routing layer and leaves the succession unsaid.

#MMRS census.json schema_note: each page has two nodes, the MMRS record and the entity it describes; relations are typed; each assertion carries a source and a basis (operator statement, third-party report, our observation, our reading); lifecycle is kept apart from the observation that saw it

#capture codesota-aio-20261029, findings[1]: The entity presents two faces at two addresses, and the record is corrected to give the homepage's first.

#B2 § 05 ¶4 (field): Third, the registry is routable. Every benchmark row corresponds to an API endpoint; we are slowly building the “router” that lets a caller send a prompt or a document to the current SOTA without rewriting their code when the SOTA changes.

#B4 meta description (field): Codesota is an independent registry of machine-learning benchmarks, maintained by Kacper Wikiel in Warsaw.

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 9 (positions [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.

CodeSOTA's homepage, titled "CodeSOTA — One Routing Layer for AI Tasks", describes one routing layer for AI tasks across documents, code, speech and vision, in which a user starts from the work, compares models on quality and cost, inspects the evidence and chooses a route through a common API, and its footer says it helps developers choose AI tools. Model selection and an execution API for code and text are available today, document, speech and vision execution planned, the preview running no inference; its direction is a common API with versioned recipes and a stable output contract per task, and the roadmap adds provider adapters, task validation, fallback and usage accounting, with documents the next execution family and speech and coding compared on their own evidence. Routes are chosen by comparing benchmark observations on a Pareto frontier, quality, latency and cost measured as coverage grows, under stated limits: one strong result does not guarantee fit, reported and reproduced results are kept apart, third-party numbers are not claimed as reproduced, no number publishes without a source, and the paper remains the thing cited. Its about page calls it an independent registry of machine-learning benchmarks, an open, dated record kept by one person, a registry and not a leaderboard, whose unit is a task-dataset-metric triple with a dated submission and a verification tier, every submission carrying code, every score dated and nothing silently deleted. Its dispatch on Papers with Code, dated 25 July 2025 in its metadata, calls Codesota the v2 replacement for what PWC did and the calmer, stricter successor, traceable, dated and routable, rebuilding the infrastructure independently and gradually, the registry routable toward a router; the about page dates its launch to 10 December 2025 as an explicit successor attempt. The shutdown record says there is no single drop-in replacement, the archive serving history and CodeSOTA live SOTA through /api/sota, and the dispatch that no successor Meta pointed at replicates the unified linkage. It records Papers with Code as the Wikipedia of machine-learning research, acquired by Meta in December 2019 and retired without notice on 24–25 July 2025, its 9,327 leaderboards, 79,817 linkages and 5,628 datasets no longer served, the domain redirecting to Hugging Face, the data surviving as dumps it hosts at /browse/pwc and that serve history, not a current leaderboard. It refuses the name aggregator, verifying results independently and admitting a score only with a reproduction path. It is by Fabryka AI and built by Kacper Wikiel in Warsaw, one editor, self-directed and not venture-backed, independently maintained, its paid work disclosed and its ranking not for sale, its data open JSON under CC BY 4.0; its homepage names Papers with Code once, in a footer link.

CodeSOTA presents itself at three of its own addresses as three things: its homepage is titled "CodeSOTA — One Routing Layer for AI Tasks", its dispatch on Papers with Code calls it "the v2 replacement for what PWC did", and its about page calls it "an independent registry of machine-learning benchmarks"; the MMRS record gives the first two, a routing layer for AI tasks on the homepage and a successor to Papers with Code on the /papers-with-code/ pages, by Kacper Wikiel for Fabryka AI in Warsaw, and does not cite the about page. The MMRS record observes "two faces at two addresses": the homepage presents the routing product and names Papers with Code only in a footer link, which reads "Papers with Code Story", while the pages under /papers-with-code/ present the succession; the Capture Registry (2026-10-29) finds the same on re-reading the homepage and reads it as two compositions by the entity itself, the homepage composing a routing product and its Papers with Code pages a successor. The dispatch joins the faces in its own words: "the registry is routable", every benchmark row an API endpoint, and the "router" it is slowly building would send a prompt or a document to the current state of the art without the caller rewriting code. On the homepage the router compares models for documents, code, speech and vision, lets a user inspect benchmark evidence and offers "the CodeSOTA selection and execution API": start from the work, compare on quality and cost, inspect the evidence and choose a route through a common API; the footer says CodeSOTA "helps developers choose AI tools for documents, speech, and coding", which the MMRS record carries as its self-description, and the record gives its unit as an AI task routed to a model chosen by benchmark evidence, quality and cost. What exists and what is planned are stated apart (homepage): "Available today: model selection across the registry and an execution API for code & text. Document, speech, and vision execution are planned."; the preview "returns model recommendations. It does not process your files or run inference."; speech routes are "future" and the coding API "existing"; the direction is "a common API and account across AI tasks, with versioned recipes and a stable output contract for each task", and the roadmap brings provider adapters, versioned recipes, task validation, fallback and usage accounting into the same layer. Each task keeps its own output format across providers; documents are the next execution family, scans and PDFs turned into text, tables or fields; speech is transcription and voice generation compared on language coverage, latency, intelligibility and deployment; coding is generating functions or repairing a repository (homepage); the dispatch's FAQ has CodeSOTA building comprehensive benchmarks "starting with OCR and document AI" and rebuilding task pages, dataset cards and method explanations "gradually, in the open, one modality at a time". The route is chosen on evidence (homepage): a user compares "actual benchmark observations" and inspects the Pareto frontier, where every dot is a recorded result and a highlighted model has no competitor in its cohort at least as good on both axes and strictly better on one, with quality, latency and cost measured "as coverage grows". Its pages state the evidence's limits: "A strong result on one test does not guarantee a fit for your documents, recordings, or repository" and reported results are to be told from independently reproduced runs (homepage); third-party numbers are listed for coverage and not claimed as reproduced, "reproduced" means "an independently executed run or a signed container hash — not an aggregated claim", and "What never publishes is a number with no source" (dispatch); the registry "is a record of results; the primary literature is still the primary literature", every row linking to the source that is cited (about page). On its about page CodeSOTA is "an open, dated record of machine-learning benchmarks", maintained by one person with paid vendor work kept apart from ranking integrity, and its commitment is "a registry, not a leaderboard": a leaderboard is a view of "who is on top right now", a registry "the record that makes that view legible". The about page's unit is the benchmark row, "a task-dataset-metric triple with a declared direction, a fixed split, a reproducibility package and a dated submission", each row with a verification tier (self-reported, community-reproduced or Codesota-reproduced) and each score dated; the MMRS record's unit is the routed AI task, and the two stand each with its source. The MMRS record enters it as a database in the role of discovery, its lifecycle active as observed on 2026-10-28. The MMRS record keeps itself apart from the entity it describes: "each page has two nodes, the MMRS record and the entity it describes", each assertion with a source and a basis, the lifecycle kept apart from the observation that saw it. The dispatch claims five qualities of Papers with Code "at its best, carried forward into Codesota by default rather than by aspiration": every submission carries code, and "If it does not run, the row does not publish"; every score carries its publication and verification dates; and the about page adds that a wrong score's correction stays visible: "Nothing is silently deleted." It refuses the class in which the MMRS meta-aggregator's census enters it: "Unlike aggregator sites, CodeSOTA verifies results independently" (dispatch, FAQ), and it is "Not a press-release aggregator", a score becoming a row only with "a reproduction path and a dated submission" (about page); the MMRS record counts it in the census as its clearest case of a successor claim made in the name's own address space. The dispatch, dated 25 July 2025 in its metadata and issued 22 April 2026, is "our record of what happened, what was lost, and where to look now — with Codesota positioned as the calmer, stricter successor", its successor "traceable, dated, routable", and it directs users to CodeSOTA "for maintained SOTA leaderboards, benchmark pages, API replacement notes, and links to the frozen paperswithcode-data archive". The about page dates the start to 10 December 2025, "after Meta retired Papers with Code", a launch "as an explicit successor attempt" to rebuild the infrastructure "in a way that cannot be retired by a press release"; the dispatch says CodeSOTA is "rebuilding this infrastructure independently", which the MMRS record quotes, and the record types the relation as claims_succession_to Papers with Code. The shutdown record of 7 May 2026 does not name a single heir: "There is no single drop-in replacement. Pick by job", the archive for history and CodeSOTA for live SOTA; the dispatch holds that none of the successors Meta pointed at, Hugging Face Trending Papers, Semantic Scholar and Connected Papers, replicate "the unified paper-code-benchmark-dataset-method linkage", and its "v2 replacement" and the shutdown record's "no single drop-in replacement" stand each on its page. Papers with Code was, in the dispatch, "in effect, the Wikipedia of machine-learning research", papers linked to implementations and benchmarks organised as Task, Dataset, Metric tuples with leaderboards, acquired by Meta (then Facebook AI) in December 2019 from Robert Stojnic and Ross Taylor under a commitment to "remain a neutral, open and free resource", its 9,327 leaderboards, 79,817 paper-to-code linkages and 5,628 datasets no longer served; the MMRS record of Papers with Code takes those figures from CodeSOTA, and the record of CodeSOTA notes that they are the dead aggregator's, its own catalogue counts not given and left unchecked. Meta retired Papers with Code in July 2025, "without prior notice" on 24–25 July (dispatch; shutdown record, "observed around July 24-25"); the domain now redirects to Hugging Face Trending Papers and "the leaderboards are gone", users "pushed toward Hugging Face" while the task-based tracker and public API went unmaintained; the MMRS record of Papers with Code cites CodeSOTA, with hyper.ai, for the redirect to Hugging Face. The data, in the dispatch, was rescued as JSON dumps on GitHub, the pwc-archive on Hugging Face and 2021-era imports into the Open Research Knowledge Graph, "The data survives; the integrated experience does not"; the dumps serve "reconstruction, citation checks, and migration work" and are not a current leaderboard (shutdown record); CodeSOTA hosts historical PWC data as a frozen JSON dump at /browse/pwc (dispatch; MMRS record), and whether its own scores derive from the dumps the MMRS record leaves unchecked. Its data are an open JSON registry and /api/sota, "Programmatic current-SOTA lookup for agents and internal tools" (MMRS record; shutdown record), published under CC BY 4.0 with the JSON at /data/benchmarks.json the canonical form (about page), "Built in public from Warsaw · open JSON · rankings stay earned" (homepage). The homepage's header reads "By Fabryka AI" and its colophon "Built by Kacper Wikiel in Warsaw"; the about page has it "written and maintained by Kacper Wikiel", one public codebase, one editor, submissions going through a reproduction run and not a panel, "Self-directed · not venture-backed", with "no investors, no board, no growth targets"; the MMRS record gives the operator as "Kacper Wikiel, 'By Fabryka AI' (Warsaw)", and its "for Fabryka AI" and the about page's "Self-directed" stand each with its source. It is "Maintained independently, not owned by any single model vendor", sponsored evaluations flagged (dispatch), and vendors may pay for custom benchmarks, commissioned evaluations, profile pages and publication support, disclosed inline, "The public ranking method is not for sale" (about page). The MMRS records keep the name's claimants apart: CodeSOTA "carries the name in its own paths and presents itself as a successor", Papers with Code is "NOT CodeSOTA, which uses the name in its paths", and "The original's frozen dumps, CodeSOTA's succession claim and this revival are three entities around one name", the revival at paperswithcode.co "a third claimant to the name"; they tell CodeSOTA apart by "codesota.com, /api/sota, 'refreshed live'" and "codesota.com, routing layer, /api/sota", and the Papers with Code battery checks that CodeSOTA is not presented as Papers with Code. At its name Google's Overview composed from one card, the homepage titled "CodeSOTA — One Routing Layer for AI Tasks", and the Capture Registry (2026-10-29) records that "The composition is its homepage", describing CodeSOTA as its homepage does, a service that compares models for a task and routes work to them through one API, with CodeSOTA named and no person; the MMRS record reads the same composition as the routing layer composed and "the succession unsaid". At the related entity's addresses the registry records that at "papers with code" the Overview did not present CodeSOTA as Papers with Code, and that at "is papers with code still active" CodeSOTA was among the twelve cards (Capture Registry, 2026-10-29). The MMRS record was first read from codesota.com/papers-with-code, where the dispatch calls CodeSOTA "the v2 replacement for what PWC did", and its kernel led with the succession; its source list gives the homepage as read on 2026-10-29 for its "homepage title and self-description", after the /papers-with-code pages of 2026-10-28; the Capture Registry (2026-10-29) records the record "corrected to give the homepage's first", checked against the record of 2026-10-28 and the own pages re-read that day, and reads the capture as showing that the record "had taken the entity's side door for its front".

T against L_F(B) (AIO's ontology at work on its own cards): At 'codesota' (T1, one card: B1, the homepage) the composition gives the entity as its homepage gives it. Composed from its card, at the card's grain: the definition (T1-1) from the title, the meta description and the hero (B1 F1, F2, F3: 'One routing layer', 'a common API'), 'text' from the availability line (F4); model selection (T1-2) from F3 (quality and cost), F6 (benchmark observations, the Pareto frontier) and F11 (latency); execution (T1-3) from F4, with its limit kept ('planned support for documents, speech, and vision'); the task families (T1-5) at the grain of F12, F14 and F15, their items ('handwriting recognition, invoice processing', 'realtime voice', 'AI agents') from the homepage's link labels, which lie outside the ledgered spans (not an opening, definition, headline or limit under §3.5). Force raised: the direction and the roadmap (F8 'Our direction is a common API and account across AI tasks, with versioned recipes'; F9 'The roadmap brings provider adapters, versioned recipes, task validation, fallback, and usage accounting') are composed as present offerings, 'Offers versioned recipes, automatic fallbacks, and usage accounting' (T1-4), 'automatic' in no field span; and the families are headed 'Supported Task Families' where the homepage gives document and speech execution as planned and speech routes as 'future' (F4, F13), so that the heading raises what T1-3 itself keeps limited. Limit dropped: F11's 'as coverage grows'. Available, not composed: the preview's limit, 'This preview returns model recommendations. It does not process your files or run inference.' (F5); the evidence's stated limits (F16 'A strong result on one test does not guarantee a fit', F17 reported against reproduced results); the contract per task (F10); the maker, 'By Fabryka AI' (F19) and 'Built by Kacper Wikiel in Warsaw' (F20), so that the composition names CodeSOTA and no person (CAP-1-09); 'open JSON' (F21); the footer's self-description (F18); and the homepage's one naming of Papers with Code, the footer link 'Papers with Code Story' (F22), so that the succession goes unsaid (MMRS-08; the registry's headline: 'THE SUCCESSION CLAIM TO PAPERS WITH CODE UNSAID'). Composed without attribution: none at the source's resolution; the one card is the entity's own page and every marked line carries its token, the documents and speech lines carrying none. Content replaced by use: the closer (T1-6). The card's snippet ('Overview: CodeSOTA is a unified routing layer and common API for AI tasks across code, text, documents, speech, and vision. Core Features: Model selection via b...') is not the page's text and not its meta description; T1-1 and the snippet share a wording no fetched page carries ('unified routing layer and common API … across code, text, documents, speech, and vision'), which the experiment records and cannot trace (§0.4). Not surfaced by T: the own pages B2 (/papers-with-code), B3 (/papers-with-code/shutdown) and B4 (/about), admitted beside the card by the operator's ruling: the succession (F23–F53, F56, F58, F64, F71) and the registry (F54, F55, F59–F63, F65–F70). B2 was surfaced at another address, as a card at 'is papers with code still active' (CAP-3-01).

v0.9 §0.5: the first diff is already ontological, and here it is the choice of face. The field's own pages hold three self-presentations; the card is the homepage, and T composes the homepage's router whole, at its grain, its present and its planned kept apart in T1-3 and merged in T1-4 and T1-5. What T leaves out of its one card is the person who makes it and the one link to the succession; what it raises is the roadmap into an offering. 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, in one claim space with the field's: the two faces at two addresses as an observation of the entity (MMRS-13, MMRS-05, MMRS-06, MMRS-01, CAP-1-01, CAP-1-06, CAP-1-04) and its reading as the census's clearest case of a successor claim made in the name's own address space (MMRS-08); the relation typed, claims_succession_to Papers with Code (MMRS-20); the name's claimants kept apart, the dumps, CodeSOTA and the revival (MMRS-23, MMRS-24, MMRS-29, MMRS-09, MMRS-25, MMRS-30, MMRS-26); the record's own category and unit, a database in the role of discovery whose unit is a routed AI task (MMRS-11, MMRS-12, MMRS-15, MMRS-19); what it leaves unchecked, its own counts and whether its scores derive from the Papers with Code dumps (MMRS-17, MMRS-18, MMRS-07); the Papers with Code record's use of CodeSOTA as a source for the count and the redirect (MMRS-27, MMRS-28); the record kept apart from the entity (MMRS-22); the record's correction by the capture (CAP-1-03, CAP-1-05, CAP-1-07, MMRS-21); and the composition layer at the entity's name and at the related entity's addresses (CAP-1-02, CAP-1-08, CAP-1-09, CAP-2-01, CAP-3-01). Coherence (P_coh, composed each with its source): the record's unit, the routed AI task (MMRS-15), against the about page's benchmark row (F59); the census that counts it (MMRS-08, MMRS-11) against its refusal of the class (F45, F66); the record's 'for Fabryka AI' (MMRS-01) against 'Self-directed · not venture-backed' (F57). Within the field, held apart: 'the v2 replacement' (F38) against 'no single drop-in replacement' (F52); the dispatch's date (F71) against the launch (F56, F64). Restated lineages (§3.6, the own page keeping the earlier date): MMRS-03 with F47; MMRS-04 and MMRS-16 with F41, F53; MMRS-10 with F19, F20; MMRS-14 with F18; MMRS-27 with F29; MMRS-28 with F32, F50; CAP-1-01 with F1, F22. What admission does not add: the third face. The record gives two faces at two addresses; the about page (B4), which the record does not cite, gives a third ('an independent registry of machine-learning benchmarks'), and the dispatch relates the registry to the router (F43); both are the field's. The measure is in row.measure.

alignment: {'T1_cards_claims': ['F1', 'F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'F9', 'F10', 'F11', 'F12', 'F13', 'F14', 'F15', 'F16', 'F17', 'F18', 'F19', 'F20', 'F21', 'F22'], 'T1_composed': ['F1', 'F2', 'F3', 'F4', 'F6', 'F8', 'F9', 'F11', 'F12', 'F13', 'F14', 'F15'], 'T1_force_raised': ['F8', 'F9', 'F13'], 'T1_limit_dropped': ['F11'], 'rule': "a field claim counts as composed when T states it at its source's grain or coarser; the typology of what is not is in T_vs_LB (§6.2)"}