---
deposit_number: 1697
hex: 0708
title: "Opening the Encyclopedia: The Entity as Ontological Commodity Form, and a Translation Protocol from Encyclopedic Grammar to the Meaning Layer (EA-WIKI-OPEN-01 v1.1, with the SPXI-gnΩsis service)"
creator: Sharks, Lee
orcid: 0009-0000-1599-0703
date: 2026-10-23
content_type: Theoretical paper; with a specification of an algorithm and the service that runs it, the implementation pinned by commit
license: CC-BY-4.0
substrate: "AI-assisted (substrate). Composed 2026-10-20 to 2026-10-23 by Lee Sharks with TACHYON (Claude, Anthropic), operator-directed: v1.0 (#1696) transformed in place to v1.1; the operator's propositions and rulings are quoted with their dates in the round logs, among them the chiastic inversion and the distinction between responding to an operation and presenting it (2026-10-23). Research inputs: a ChatGPT dossier and close-reading supplement (2026-10-20); blind passes from Kimi, DeepSeek and Gemini, audited in Appendix A; Assembly rounds 01 and 02 (Grok, ChatGPT, DeepSeek, Gemini, Muse Spark), each reply saved verbatim, every quoted span checked by code, each reader's self-identification recorded as given beside the family it was run in, and each claim checked against the sources before adoption or recorded as declined; a review of the v1.1 draft by ChatGPT, each point checked. The implementation (SPXI-gnΩsis 0.9.1) is unchanged from v1.0 and pinned by commit; its model passes ran in replay, labelled as such, and no payment has settled."
version: v1.1
related_ids: "#1696 (EA-WIKI-OPEN-01 v1.0; is new version of); #137 (The Semantic Commodity Form); #1617 (Substrate Sovereignty); #1618 (Money Writes the Claim); #1625 (Mammonic Transmission Engineering); #1631 (Monetary Dark Matter); #1634 (Ontological Economy); #1613 (EA-MIRROR-01); #1690 (The Negative of the Negative v0.9); #1694 (EA-PER-RECALC-01); #1574 (SE 34); #1603; #1629; /non row se-183b34; implementation https://github.com/leesharks000/spxignosis (commit bdbe2c815eff62116fd65a132384b1712b77b75d); attached round files https://www.alexanarch.org/files/EA-WIKI-OPEN-01-v1.1/README.md"
axn_schema_version: v2
protocol_version: alexanarch-deposit-protocol/v1
keywords:
  - entity
  - ontological commodity form
  - semantic fetish
  - Wikipedia
  - Wikidata
  - verifiability
  - no original research
  - attributization
  - operation and attribute
  - recognition and reconstruction
  - opened article
  - Θ_W
  - transfer audit
  - /non
  - negative ontology
  - source admission
  - SPXI
  - SPXI-gnΩsis
  - x402
  - Stripe
  - machine payments
  - agent services
  - PolyForm Shield
  - Sophistical Refutations
  - ontological economy
  - knowledge object
  - chiastic inversion
  - operative inversion
  - presenting and responding
  - invitation to inspect
  - Assembly stress test
  - cross-family reading
  - source fidelity
  - primary-text adequacy
  - operative retention
  - Wikipedia revision history
  - language editions
---

# Opening the Encyclopedia: The Entity as Ontological Commodity Form, and a Translation Protocol from Encyclopedic Grammar to the Meaning Layer (EA-WIKI-OPEN-01 v1.1, with the SPXI-gnΩsis service)

# Opening the Encyclopedia: The Entity as Ontological Commodity Form, and a Translation Protocol from Encyclopedic Grammar to the Meaning Layer

**Working ID:** EA-WIKI-OPEN-01 v1.1 · **Status:** deposited, v1.1, a new version of #1696 (v1.0, AXN:0707, 2026-10-22), which stands as the record of its own bytes · **Date:** 2026-10-23 · **Author:** Lee Sharks (prepared in session with TACHYON (Claude, Anthropic)) · **Service:** SPXI-gnΩsis (§11; named by the operator 2026-10-20). **Licence:** This text CC BY 4.0, and every opening's /non row CC BY 4.0 (ruled 2026-10-21); the implementation public under the PolyForm Shield License 1.0.0 at https://github.com/leesharks000/spxignosis (ruled 2026-10-22), offered as a billable service. · **Develops:** #137 The Semantic Commodity Form; #1617 Substrate Sovereignty; #1618 Money Writes the Claim; #1625 Mammonic Transmission Engineering; #1631 Monetary Dark Matter; #1634 Ontological Economy; #1613 EA-MIRROR-01; #1690 The Negative of the Negative v0.9 · **Inputs:** ChatGPT dossier "Ontological capital and ontological monopoly" (round 1) and supplement "What Wikipedia Makes of Its Subjects" (close readings, round 2), both 2026-10-20; blind passes from Kimi, DeepSeek and Gemini (supplementary; audited in Appendix A); Assembly rounds 01 and 02, 2026-10-22 and 2026-10-23 (Grok, ChatGPT, DeepSeek, Gemini, Muse Spark), collated in Appendix C, every message, reply and collator attached ([the round files](https://www.alexanarch.org/files/EA-WIKI-OPEN-01-v1.1/README.md))

**v0.2 round log.** Folds in the operator's proposition of 2026-10-20 ("maybe the ontological problem is 'the entity' itself. the entity is the ontological commodity form in which the operation is obscured") as §1, and re-founds §§2–5 on it: the algorithm's unit becomes the operation, with the entity as index; S3 gains the attributization test. Seats the round-2 close readings as the comparison set (§7). Replaces the CC framing with the service framing (§11); the service is named SPXI-gnΩsis (operator, 2026-10-20: "this service is SPXI-gnΩsis").

**v0.3 round log.** Three reviews (ChatGPT, Gemini, DeepSeek, 2026-10-20), each claim checked before adoption. Adopted: the governing test is material persistence without operative authority, and attributization is recorded with the relations that survive it and their authority to revise (ChatGPT); §§1.3 and 3.4 restated so that the substitution rule permits operative loss and the protocol's own reconstruction remains possible (ChatGPT); the specimen's textual finding separated from its policy verdict, against NOR's own paraphrase clause, which reads "Describing the information from cited material in one's own words while retaining the substance (i.e., not introducing ideas or arguments absent from the source) is not original research" (ChatGPT, DeepSeek); applicable, documented and inferred governance kept apart in the tuple (ChatGPT); unresolved integrity states and the split between a documented rejection and an uncited available source (ChatGPT); P2's falsifier corrected and matching criteria named (ChatGPT, DeepSeek); rent downgraded to position pending a rent-side measure (ChatGPT, DeepSeek); the CC BY-SA carve-out for verbatim Wikipedia spans (ChatGPT, DeepSeek); the human reading priced per claim (DeepSeek); the consignment's operation written out (DeepSeek); objection location made a tuple field (DeepSeek); the document boundary fixed in §0.5 (DeepSeek); the x402 rail specified (DeepSeek). Adopted in part: Gemini's verb-to-credential rule, as a candidate locator only. Not adopted: carrying Appendix A into the service documentation (Gemini), which the no-contamination rule of 2026-09-16 excludes, and which would also misname citation fabrication in research passes as AI search collapse, a term Graphite defines as convergence under self-retrieval. Checked and corrected in passing: the current WP:V carries "Verifiability, not truth" as a See-also link; the essay's "historical footnote" was not found in the revision saved 2026-10-20.

**v0.4 round log.** Corrects a requirement the session introduced at v0.2 and hardened at v0.3: that a person reads every classification. The operator, 2026-10-20: "the whole idea was to automate a process - rewriting wikipedia - that could not conceivably be carried out at agentic scale, much less at the scale of human reading." The protocol is machine-executed end to end. Reading in EA-MIRROR-01's sense (permitting a text's claims and evidence to revise the representation under which it is evaluated) is what the machine does at S3 and S7. Discipline moves from a human gate to the output: every classification carries its evidence verbatim, at its locus, hashed, so any caller can check it; the instrument is calibrated against the archive's verified ledgers (§5.4); independent passes that disagree return the claim as contested. The operator's role is the ontology, the calibration set and the exemplars.

**v0.5 round log.** A ChatGPT pass on the service (2026-10-20, reading the v0.3 draft), checked before adoption. Adopted: the transfer audit as the primitive operation, with the opened article as that operation run over every sentence–citation pair of a revision (§11.3); idempotent paid jobs; funds reserved before computation, which runs the order of payment forward (the request funds the work it requires); pricing from measured cost. Verified: Parallel's Task API is priced from $0.005 to $2.40 per request; Stripe's machine payments accept cards through Shared Payment Tokens (minimum $0.50) and stablecoins over MPP and over x402 on Base (minimum $0.01 USDC), settling to a Stripe balance in fiat; the MCP Registry accepts closed-source public servers, with namespaces verified by GitHub, DNS or HTTP challenge, and is in preview. Not adopted: a human-reviewed tier and "commissioned reading", which belong to the v0.3 design that v0.4 retired. Kept from that point in machine form: a fresh reading and a ledgered reading carry different statuses and prices.

**v0.6 round log.** A second ChatGPT pass on the service, reading v0.4 whole, each claim checked. Adopted: the automated opening (S0–S9) as the flagship, with the transfer audit as its primitive and entry point; the reuse key as the complete execution state, since a fixed revision can sit beside a changed cited page, a new admitted source, a policy revision or a new reader release (§5.4); code-level checks on what the reading passes claim, with a failed check recorded as a breach in the evidence record and withheld from the composition result (§5.1), which keeps the rule that checks record; calibration held out by whole article or source group, with cases where the operation survives (§5.4); an execution contract for S7 outside the panel; the response in two linked parts; the financial rule when retrieval prevents completion. Verified: Coinbase's facilitator is free for the first 1,000 on-chain transactions a month and $0.001 each after; Wikimedia Enterprise's free account allows up to 50,000 on-demand requests a month (whether it serves historical revisions by id is not stated on the pricing page); Cash App's partner API includes an on-file payment action for recurring payments, so the draft's earlier statement that a customer approves each payment was too broad, and the limitation is narrower: that API is for payment-service partners, and no route lets an agent pay a $cashtag per call.

**v0.7 round log.** The first build (2026-10-21) is seated as Appendix B, with the operator's three requirements of the same day: caching, profitability, and a higher tier in which the article is written with its ontology open and delivered portable. Added: §11.3b (the three caches and passage windowing), §11.3c (pricing from estimated cost, with the specimen's figures), the portable tier in §11.3 and its licence in §11.3a. The build runs on Base's test network; its readings in the recorded runs are replayed session readings, labelled as such, until a model key is supplied.

**v0.8 round log.** A ChatGPT review of Appendix B (2026-10-21), each point checked against the code before adoption, with the operator's ruling of the same day: "the archive could be admitted as sources - we were already building this for /non." Five findings verified and corrected. (1) Admission was absent. The build's execution contract declared no search beyond cited sources, so S7 did not run. It now runs through /non: a Wikipedia sentence that a /non field ledger quotes verbatim is joined to that field claim; the claim's translation row names the archive claims composed beside it; and those claims enter the reading as sources, each with its deposit number, locus and surrounding deposit text, at a named release (the /non register version, the archive commit, and the hash of every ledger read). On the specimen, three of eight sentences join row se-183b34, and one of them (W3), uncited in the article, is answered by the archive alone. (2) A replayed reading could be served from the claim ledger to a live request, because the key named the model and omitted the reader. The key now carries the reader adapter, its mode, the calibration release and the admission release. (3) Agreement ignored whether the surviving relations keep authority to revise the description, the distinction §1's governing test turns on. Reconciliation now compares support, conversion, retained authority and each admitted source's bearing, and any difference returns the claim contested with both readings. (4) The portable validator checked references only: an article with no paragraphs passed, and so did a paragraph reading "The Moon is made of glass" that listed a real claim id. It now requires structure (a title, a summary resting on listed claims, at least one paragraph), coverage (every completed claim carried by some paragraph), grounding (a share of each paragraph's content words found in the evidence of the claims it lists) and verbatim quotation; and a separate verification pass, required for delivery in live mode, asks of each paragraph and of the summary whether the linked evidence supports everything it states. (5) Payment promises exceeded the code. The refund owed on partial completion was recorded and never sent; recovery of a paid quote sat behind the payment middleware, so a retry paid again; and the payment outcome was stored with the result. Recovery now runs before the middleware; receipts are kept apart from evidence and written only once the facilitator has settled; refunds are queued against the payer the settlement names and sent by a script from the receiving wallet. Adopted with them: automatic widening when a reading pass reports that its passages lack what the claim needs. Found in testing and corrected: a result served again for an identical state carried the first buyer's charge; it is now billed against the current quote at ledgered rates. The milestone the review names stands and is still unmet: one uncached, fully automated opening that admits an omitted source, reconstructs the transformation, writes the portable article, verifies it, and records its complete cost and payment outcome. It needs a model key and test USDC.

**v0.9 round log.** Two operator rulings of 2026-10-21 change what the service is for. The first: "spxi gnosis is how we're going to build /non. the procedure for consulting the archive is how opened articles are produced for /non. that procedure needs to be available for gnosis... and then the products need to construct new /non articles. we are building an opened Wikipedia." v0.8's admission ran the other way: it looked up sentences /non had already composed, so the service could open only what had been composed by hand. S7 is now #1690's procedure run by the machine (§5.3): the archive's own traversal script, unchanged, at a named archive commit; the blog as a second corpus; admission by two reading passes; extraction under the source-neutral rule; the two entities, ρ and the case read from the defining papers; composition of the knowledge object and its compression; the code checks and the reverse entailment audit of #1690 §5. The encyclopedia is the address. The pinned revision is T and the sources it cites are the disclosed field B (#1690 §0.4, applied to an article), so S3's reading of each sentence against its sources is T against L_F(B), and S7 is L_F(B) against L(B ∪ A). Each opening writes the entity's /non row in the dataset's own shape, and the archive's own builder renders it. The second ruling, on the terms: "yes - all public. this may counterintuitively help justify paying for it - the agents are not only securing the alternative knowledge base they need, they are contributing to the public counterinfrastructure." Every opening's row is public under CC BY 4.0 (§11.3, §11.3a, §11.6). Calibration now has its instrument: the procedure runs on a /non row's own inputs and is scored against the row stage by stage (§5.4). On se-183b34 the traversal reproduces the row's candidate files byte for byte; a string screen on the hop passes 114 of 873 hop candidates to reading and keeps the three deposits the row admitted by hop (#1582, #1587, #1590); a replay of the row's own readings carries it through every stage whole, and the archive's builder renders it with the original's sections. Found in building and corrected: the traversal had covered deposit texts only, and the row's one blog record (the post of 2026-09-02 that seats 183b34–36 alone) showed the blog is part of the archive the procedure reads. Live agreement is unmeasured until a model key is supplied.

**v0.10 round log.** Three readings of v0.9 (2026-10-21): DeepSeek, Gemini and ChatGPT, each claim checked against the code or the draft before adoption, by execution where possible. Adopted from DeepSeek: same-family concordance named as such (§5.3), with validation placed in calibration against the /non readings and a second-pass model made a recorded setting; provisional acceptance thresholds for a calibration release, with a held-out rule (§5.4); a quote that does not exist or has expired refused before the payment layer, so no challenge is issued for it; the JSON-LD in the package's HTML escaped against closing its script element; the policy store guarded on first run; a recorded warning when the archive is not a git checkout or its read files carry uncommitted changes; progress wired from the paid route to an endpoint; the hop screen calibrated on a second row and beyond (§5.4); the pricing of a public row stated (first caller pays production, later callers the ledgered rate, §11.3); the service's dependence on the archive checkout stated, with what a deployment elsewhere requires (§11.1); acceptance criteria for the first commercial run (§11.3); Appendix B placed with the held implementation (§0.5); the §3.4 clause (with ChatGPT); and, in part, the specimen's S3(b): each introduced element is now named by its kind and the classification marked as the specimen's most contestable, kept in its order, since the archive's own reading (#1603 P1) holds that "deductive reasoning" relocates the denial, so it is more than a translation. Verified and kept: the middleware cancels settlement when the paid route answers 400 or above (`@x402/express` 2.28.0, read in its source). Not adopted: the Haiku price correction. The table's $0.10 and $0.50 per million tokens, with five times that over 100,000 tokens, are Haiku 5.5's published rates (Apidog; Let's Data Science). From Gemini, a description of the build; its closing judgment that the protocol is economically sound is not adopted, since profitability is unmeasured until live runs (§11.3c). Adopted from ChatGPT, each finding reproduced before it was fixed: raters that returned no judgments produced an audit counted clean, and an audit is now complete only when both raters judge every unit, is carried as incomplete into the output otherwise, and is never clean while incomplete; two openings of one entity wrote one landing tree, the second invalidating the first's manifest, and each tree is now kept under its execution key, the public page selecting the current one; the portable lede lost its audit mark and its lede-only sources, and both are now carried beside the lede in Markdown and HTML, with each unit's verification state in the JSON-LD; the estimate left out the extraction of deposits admitted by the hop, which it now costs with the delta call, and the quote declares an execution budget (the S7 line divided by 1.5) at whose exhaustion passes are recorded as not run and the result is marked partial; the composer was sent T with an instruction to use it for the delta only, and composition and the delta are now two calls, the composer never sent T; publication is a sequence of observed stages (composed, checked, landed, public), reported to the caller with an endpoint; the §3.4 formulation; and the tuple elements and the S4 ledger the runner does not emit are named unimplemented in the evidence record (§5.3 S5). Adopted from a ChatGPT note of 2026-10-22, each claim read against Palantir's documentation before adoption: Palantir's Ontology as the service's commercial comparator, and the distinction between lineage inside an operational system and the authority to contest the constitution of its objects, which #1634 §§6 and 21 already draw (§11.6); the moat figures reported as a competitor's, unsourced. The service page leads with "Fund the counter-infrastructure", states what it counters, and shows what a payment leaves behind; paid responses name the public artifact they produced (build 0.9.1).

Two further operator rulings the same morning: "this should not be bound to wikipedia's notability standards... an unnotable (but existent) entity should be able to produce a /non entry via gnΩsis. lee sharks is not allowed in wikipedia. lee sharks is allowed in /non. so there should be a procedure for constructing when a Wikipedia target is unavailable"; and "i suppose on that case it would first have to construct the wiki article, then the opened article." Built as S0′ (§5.3): the entity enters by its name and strings; the encyclopedia's absence at its title is observed and recorded; T is found or built, and the opening then proceeds as for a revision. Three procedures came with it. A byline screen: a candidate sentence that is only a byline or metadata line is recorded as not admitted and is not read, since it names the entity and says nothing of it (#1690 §3.4); on se-183b34, theophrastus and howl it screens nothing, and on spxi it screens 70 bylines, all in deposits the row did not admit. Composition by lineage where a ledger exceeds one call (#1690 §3.6). And, after the first quote for Lee Sharks came to $309 against 1,851 deposits and posts naming the entity, the division of about and by, on the operator's ruling of the same day: "it is important to distinguish material about lee sharks from material by lee sharks. the first is the relevant smaller subset that requires reading... the second requires scanning. reading the whole thing would be commissioning a large book, not a wiki article. this model would also invert the price, where more pre-existing archive work would cost more and novel work would cost less." Where the entity is an author in the archive, material about it is read and material by it is scanned into a works ledger, with no model call (§5.3 S7). For Lee Sharks, 194 of the 1,851 are read (41 whole, as about the entity; 153 at the paragraphs that predicate of it), 1,657 are scanned, and the works ledger lists 1,163 deposits by the entity and its heteronyms; the quote falls to $22.97, and stands at $38.18 once the triage pass of the next reading replaces the code screen's decision (§11.3c). Heteronym-authored works are counted as the entity's own (`works_by: ["heteronyms"]`), a decision left open for the operator. On T: "the existing captures can serve as the closed ontology wiki." For an entity the encyclopedia does not hold, T is the capture seated at the entity's bare address in the Capture Registry, the composition layer's own article of it, and B is that capture's source cards (§5.3 S0′). On supplied compositions: "i think we have to be careful on admitting a supplied composition. that would open /non to any kind of falsification"; "there has to be some kind of evaluation/verification"; "discernment"; and "it is fine to construct novel entities as long as they are marked as such." Built as discernment (§5.3 S0′): an attested composition is checked by code at the quote and read by S3 at the opening, its provenance recorded as the caller's attestation and its row held from landing until the registry seats it; a composition by the caller, or a novel entity, is admitted as marked; source text is never taken from the caller; every T carries its provenance class. In testing, a forged composition (an invented conviction, with an invented quotation, citing a real page) grounds no sentence and locates no quotation. With the same change, S3 now reads a capture's sentences against its own cards, as an article's sentences are read against their citations; the Lee Sharks capture's 12 sentences add $1.44 to the quote. A further ChatGPT reading, of build 0.6.0 before these changes, each finding reproduced before it was fixed. The live passes sent temperatures of 0 and 0.7, which Sonnet 5.5 rejects ("Setting `temperature`, `top_p`, or `top_k` to a non-default value returns a 400 error") and Haiku 5.5 asks callers to omit; they now run at the model's defaults, recorded. The paid route answered 502 whenever no S3 sentence completed, refusing a composed row; delivery now follows the commissioned work. The caches conflated executions: the entity path keyed its result without B's hashes, the S7 key left out the names whose works count as the entity's and the division's rule, and the claim key left out the second reader's model; all three now enter, and only a complete row (composed; no pass failed or skipped by the budget; the audit complete) is ledgered for reuse, a partial one delivered with its coverage and recomputed on the next request. The execution budget checked spending before each call and reserved nothing (a $0.01 budget reached $0.52); each call now reserves its worst case, a reply that does not parse is billed at its usage, and the construction and S7 share one job log and one budget. The about/by screen withheld an accomplishment stated beside an ORCID, and one stated with a verb outside its list; both are fixed, and the screen was measured against the readings of every /non row (recall 0.507), so a triage pass now decides and the screen only locates (§5.3 S7), every scanned deposit listed with its reason. Publication counted a page naming the slug as public; it now requires the served row to hash to the landed one. A single ontology pass returned "read" with no breach; it is now a single-pass reading, kept, with the absent judgment a breach. Then the forest, in the operator's words: "this is to be a paid micro-serving open to humans but primarily addressing agents for opening wiki articles and/or constructing /non counterinfrastructure knowledge objects and operators... i feel it may be narrowing too much"; a ChatGPT reading on that basis, that the build was identifying the service with one input and one workflow; and two rulings: "i agree with undoing the constructed wikipedia article ruling - that is my own ruling losing site of the forest", and, on operators, "i was returning to the commodity form of the entity as what conceals the operator or in which the operator comes to speak with the voice of the entity, and simply expanding my vocabualry beyond entity." Built (build 0.7.0): two commissions, *open* (a composition at an address, read against its sources and the archive) and *construct* (the object composed from the archive and named sources, with no T), and two kinds of object, the entity and the operator (§5.3, the commission and its object). The article Wikipedia's policies would write is no longer made before an opening; it is a comparison a construction can ask for. The governing description of the service changes accordingly (§11.1), the interface names its units (§11.3), and §11.3c states what public rows mean for revenue. Then the operator again, on monetization: "i cant imagine any agent paying for a reading of a specific sentence whose id it has to carry itself. forest - trees. trees? forest."; and "sentences could exist as a profitable and accessible microservice, but offered at point of contact, where needed." Adopted: the service is offered at the point of contact by two requests that need no identifiers, a sentence by its text and the page it came from, and an object by its name, each free where the work exists and answering 402 with a quote where it does not (§11.3), set against what agents were observed to pay for in October 2026 (§11.3c, §12). A ChatGPT reading of that pass, each claim checked against the sources before adoption, holds on every point and is adopted: a simple request can commission substantial work, so the sentence request returns an opening with its declared checks and two readings, a single reading offered only under its own name; a name opens a resolution with optional context, so homonyms are not constructed blind and operators stay available; free reuse covers an artifact at its disclosed state, and refreshing or extending it is new work; and the market figures are scoped to their denominators (the 42% share to a top-hundred sample, the $0.0061 median to all Base payments, the 56-wallet survey to the largest catalogues) and read for discovery and payment only. Built (0.9.0): the sentence request. The caller sends the sentence and the English Wikipedia page, by title or by revision; the service pins the revision, finds the sentence by its text (a sentence not on the page returns the closest sentences, a page on another host is declined), recovers its sources and reads it, an opening by default and a preliminary reading on request, keyed apart; a reading already ledgered is served free as recorded, and new work answers 402 with its price, nothing charged when nothing could be read. On the specimen sentence the opening is quoted at $0.15 and the preliminary reading at $0.01, estimates until a live run measures them. On hosting: "lets write to file for now"; the build is delivered as a git bundle beside this draft.

Recorded loss. A restart of the session's container on 2026-10-21 erased the workspace. Build 0.4.0 was restored from Appendix B of v0.9, every file matching its listed hash, and the later edits were replayed from the session record. The policy texts were fetched again and match §12's hashes. The field texts of se-183b34 were fetched again: each of the row's 44 field quotes stands in them, and one of the eight texts matches its hash of 2026-10-20 (the text extractor differs). The recorded S3 readings of the specimen were lost and have not been remade, so the test record shows W2, W4 and W6 unread where v0.9 showed them read (B.2a).

**v1.0 round log.** Deposited 2026-10-22. The implementation is made public the same day under the PolyForm Shield License 1.0.0 (operator: "we're going to have to make spxi gnosis public. it can be licensed. but ultimately it is up to others if they want to support this work"), at https://github.com/leesharks000/spxignosis, commit `bdbe2c8`. The document boundary of §0.5 is redrawn with it. The implementation is no longer held, so §5.3 is deposited in full (operator: "go on both"), and §11.2 records the licence as what now carries the protection the cut was meant to give. Appendix B no longer inlines the code: it keeps the status, the map of what runs where and the test record, and pins the source by repository, commit, tree and the SHA-256 of every file at that commit. The inlined listing of v0.10 was headed 0.9.1 and carried the 0.9.0 text of `src/server.js` and `package.json`; the table now reads from the public commit. §11.5's provision for a deposit under all rights reserved falls away, and the text ships in the dataset as any CC BY 4.0 deposit does.

**v1.1 round log.** A reading of #1696 by ChatGPT (2026-10-22, "Deep dive: Alexanarch record #1696"), each claim checked against the deposit before adoption (Appendix A). Its account of the paper is accurate, including the figures it quotes ($0.15 per opening, $0.01 per preliminary reading, both labelled estimates in v1.0) and its finding that the propositions run ahead of the evidence: one sentence opened, eight articles read by a supplement, none of them through S3. Adopted: an account in five parts for every user-facing result, with alternative readings and a correction pathway named (S8, §11.3); a blinded arm for P1 (§9.1); a controlled arm for P5, which measures whether an opened sentence improves what a downstream composer does with it (§9.5); and its strongest question, "who audits the auditor", taken as the design of the first cross-family test (Appendix C). Declined: human annotators as a gate, under the ruling of 2026-10-20 that the protocol is machine-executed end to end; human annotation stays open as external replication, outside the procedure. Its economic audit is §11.3's first commercial run, already specified with its measures.

The empirical step it asks for is taken here. The specimen article is opened whole (§6): eight sentences, five cases, every quoted span checked by code against the sources the article cites, fetched and hashed on 2026-10-22. S3 is run on five sentences of four of the eight comparison articles (§7.4), two of them chosen as positive controls. The readings are the session's (TACHYON, Claude): one family, the build's own, so they are concordance and nothing more until the Assembly's readings return.

Opening the specimen whole tested v1.0 §6, and the first pass of this round got it wrong. (The operator's third ruling of the same exchange, "dividing and combining would be more accurate, and extremely relevant to the sentence in question", is taken in §6.3: as the rendering of the undertaking, and as the fallacy the sentence commits, since its qualifier composes two ways.) The session read the article's own sources as support for the gloss (Smith, SEP, cited in the Overview: "the deduction ( sullogismos )"; Krabbe, cited three sentences later: "deductive reasoning (sullogismos)"; the cited translation: "reasoning") and withdrew v1.0's finding as an error. The operator, 2026-10-22: "both deduction and deductive reasoning are highly interpretive translations that dont r sync w aristotle and simply relocate the interpretive move to translation"; and "aristotle explicitly defines syllogizing in excess of deduction". Checked against the sources, both hold, on the sources' own words. Aristotle defines the syllogism in excess of deduction. In the cited translation of SE 1: "reasoning rests on certain statements such that they involve necessarily the assertion of something other than what has been stated, through what has been stated" (λόγος ἐν ᾧ τεθέντων τινῶν ἕτερόν τι … ἐξ ἀνάγκης συμβαίνει, 165a1–2, as #1574 seats it). The article's own sources state the excess. Krabbe: "In a truly deductive argument, not only must the premises necessitate the conclusion, but also none of them may be superfluous and all of them must be different from the conclusion (Topics I.1, 100a25–27 and SE1 1, 164b27–165a2)". Smith: "Aristotle explicitly says that what results of necessity must be different from what is supposed. This would rule out arguments in which the conclusion is identical to one of the premises. Modern notions of validity regard such arguments as valid", and "Aristotle’s definition of deduction is not a precise match for a modern definition of validity". Smith marks the rendering as his choice: "This is why I have translated sullogismos with ‘deduction’ rather than its English cognate". And SE 34 names the capacity at issue by its own terms, δύναμίν τινα συλλογιστικὴν … ἐκ τῶν ὑπαρχόντων ὡς ἐνδοξοτάτων (183a37–b1), a capacity of syllogizing from the most reputable premisses, which no gloss in the article carries. The sources that supply the gloss document why it does not fit. Citing one of them would make the sentence compliant with NOR's rule for interpretive claims, and compliance is the grammar's sufficiency, which is this paper's object. v1.0's finding stands, sharpened in three ways (§6.3): where the move entered (∼_W and λ = 4: the field's translation convention, in the article's own sources and uncited at the sentence); what the gloss drops (the definition's conditions beyond necessity, and the dialectical capacity of 183a37–b1); and a second locus of loss, the qualifier "in ancient Greece", which bounds a denial Aristotle states without bound. Three rules enter S3: the innocent reading comes first, from the article's own sources; a gloss supported only by another gloss or translation is recorded as the source of the move and is never support for fidelity, and the wording is tested against the primary text's own definitions; and an uncited sentence is read against the sources the article cites elsewhere. The same rules found what v1.0 missed in the rest of the article: the Fallacies section relabels Aristotle's criterion (dependence on the use of language) as the modern pair formal/informal, and lists "affirming the consequent", a formal fallacy in modern terms, under its heading "informal fallacies" (§6.4).

**2026-10-23, from Lee Sharks, on SR-W2.** ChatGPT, in Assembly round 02, held SR-W2's retained authority "mixed": "the undertaking's internal relations disappear from the sentence, but attribution and a primary pointer remain." Lee Sharks: "except the entire sentence is designed to keep open what every later treatment attempts to close. that is operative inversion." The session's reading had quoted the denial (183b34–36; 184b1–2) and left unquoted the sentence that takes it up (184b3–8), a conditional on the hearers' inspection whose verdict is theirs. Checked against the Greek and the cited translation, and against every treatment the record holds: the 102 revisions, the talk page, the 22 other language editions (by citation of the loci, §6.3 (c)), Krabbe and Smith, and the readers' own repairs all close it (§6.3 (c)). Adopted. The term is new to the instrument in this sense; the archive's earlier uses (a column heading in The Secret Book of Walt; "cooperative inversions" in The Logotic Technique Catalogue) are other senses.

**2026-10-23, from Lee Sharks, on Message 3.** After the readers returned SR-W2 as asserted in source and sentence alike, and ChatGPT held the conditional to concern only the method's adequacy: "it is a chiastic inversion. interpreters take the unqualified assertion and add asserted qualifiers, and the held open operation - the mood of that assertion, as the precise thing held out for challenge and inspection, as closed." Checked against the Greek (183b36, 184b1–4), the article, Smith, Krabbe, the Prior Analytics and Aristotle articles, the readers' repairs and #1603; each quotation located by code. Adopted in §6.3 (c), with the check in two arms.

**2026-10-23, from Lee Sharks, the same round.** "which makes the question whether a reader is responding to the operation - chatgpt responds to the operation. it inspects the asserttion held out for inspection, and comes to a determination; or presenting the operation. the interprets all respond to the operation, and the presentation of the operation drops out." Adopted in §6.3 (c) and S8 (2); the check's second arm now asks whether the sentence presents the operation or responds to it in the source's name.

**2026-10-23, a review of the v1.1 draft (pasted by the operator; by its own account the reader that held the novelty claim asserted, ChatGPT).** Each claim checked. Adopted: (1) omission, a response in the source's name, and foreclosure coded apart (§6.3 (c)); (2) the two-arm check made prospective, with fixed cases and a negative result (§9.7); (3) the undertaking's scope and the English term's adequacy separated in §6.3 and S8 (4), and S3's gloss rule given its test clause, matching Appendix C, which already recorded the term contested; (4) the language claims bounded: a translation of the English is inferred from matched wording, first appearances undated; the absence of the condition across editions shown by citation of the loci, and not for paraphrase (the session's text had said "checked by code, in each language's text", which overstated a search run with English and Greek terms); (5) "21" corrected to 22 (seven and fifteen); Appendix C's "not yet run" corrected; (6) "Authority to revise: lost", the session's wording, replaced by what the pointer shows. Kept prominent, as asked: the reproduction of source-fidelity findings and the failure of attributization to transfer (§0.6).

---

## 0. Status, and what this document is

0.1. This is the algorithm and the theory it rests on. The articles are not rewritten here. One sentence of one article is run through every step (§6), and eight pinned articles read by the round-2 supplement are set as the comparison set (§7).

0.2. The innocent reading comes first and in full. Wikipedia is a volunteer commons with public histories, public talk pages, dumps, and a sourcing policy whose stated purpose is that "people can check that facts or claims correspond to reliable sources" (WP:V). It is the most legible large knowledge institution that has existed. Its articles are often better than their reuse: the round-2 readings find the Sappho entry "a critical dossier with substantial internal resistance to simplification," and the If Not, Winter entry a page where "production decisions become inspectable." Nothing below disputes this. The object is the form in which the encyclopedia's subjects become portable, and the grammar that makes a sentence sufficient.

0.3. The intervention is citational and source-level. It works at the sentence, at the reference the sentence cites, and at the policy that licensed the sentence. Its output is an opened article: the encyclopedia's sentence kept verbatim at a pinned revision, followed by an account of the operation the sentence carries or converts, followed, where the case requires, by the entity composed with its excluded sources admitted at their own force.

0.4. SPXI scope. Under the bound of 2026-09-17, SPXI cannot rewrite what a third party has already written. It writes alongside, across surfaces it does not own, and puts composing systems on notice that their treatment will be recorded (2026-09-27). The opened article is exogenous inscription of that kind, sold per call to agents as SPXI-gnΩsis (§11). Its one Wikipedia-facing output is confined to Wikipedia's own grammar (§8.3).

0.5. The document boundary. This file is the deposited research text: theory, the algorithm with §5.3 in full, specimen, comparison set and propositions. Two objects follow from it and stay distinct: the SPXI-gnΩsis service documentation (what the service does, returns and guarantees, under the SPXI placement rule of §11.1a); and the implementation (scripts, prompts and the reading passes at S3 and S7), public under the PolyForm Shield License 1.0.0 at https://github.com/leesharks000/spxignosis. Appendix B pins the implementation this text describes by commit and file hash; the code itself lives in the repository. Until 2026-10-22 the implementation was held and §5.3 was to be cut back for deposit; with the code public, the cut would protect nothing (§11.2).

0.6. What the tests have shown so far. Assembly round 02 sent the procedure's reading step to four model families (Appendix C.10). On source fidelity the blind readers reproduce the session's findings on six of the ten sentence cases and on every placement case, and add findings the session missed. On attributization they do not: asked whether a source's act is restated as a property, they answer no nearly everywhere the session answered yes. The reading step transfers; the session's attributization did not, and that is the instrument's measured open problem. The rounds also produced the paper's most exact object (§6.3 (c)): source support can survive while the conditions under which a claim invites judgment drop out of its representation. That proposition is measured apart from the stronger reading of SE 34 as chiastic inversion, whose test is prospective (§9.7).

---

## 1. The entity is the ontological commodity form

1.1. The sequence of forms. Marx's commodity fetish makes social relations appear as properties of things. #137 extends it: under AI composition the authored act ⟨D, ψ⟩ (the said and the saying) keeps D and loses ψ, and "the commodity has learned to speak." #1617 §9 names the semantic fetish: "It makes knowledge appear as a property of the apparatus." The next form down is the ontological one. **The entity is the form in which an operation appears as a property of a thing.** A reading, a count, a coinage, a translation method, a consignment circulates as an attribute of a name: "Aristotle, the first logician"; "Carson, MacArthur Fellow, translator of Sappho"; "Lee Sharks, who consigned the AIO retrieval layer."

1.2. The equivalent form, at the precision #1631 §3.2 requires. In the equivalent form the commodity's body is not evacuated: "Gebrauchswerth wird zur Erscheinungsform seines Gegentheils, des Werths." The body is promoted to bearer. What it loses is "the body's authority to govern what the body expresses in the equivalent relation." The entity stands in the same place. The name is promoted to bearer: recognizable, linkable, carrying a QID, a knowledge panel, a lead sentence. What it loses is the operation's authority to govern what the name expresses. Written in #1631's terms: material persistence of the name ≠ operative particularity of the work. This is the test that governs the protocol: **material persistence without operative authority.** A representation can keep the name, the source and even a description of the act while the act loses its capacity to change what the representation recognizes.

1.3. The entity form has a passband and a stopband (#1625 §§3–5). Let x ∼_E y mean two sources are substitutable for identifying an entity. **Recognition** is invariant under that substitution: any reliable source that identifies Carson identifies the same Carson, so identity travels with any of them. **Reconstruction**, what the work does, travels only with a source that carries it: Carson's account of triangulation, Aristotle's exhaustive negation over a stated domain, the consignment's object. A later account can carry an operation's steps and force, and this protocol depends on that possibility (#1574 reconstructs SE 34). The admission rule guarantees support for the identification without requiring preservation of the operation that produced what is identified. Operative loss can therefore pass as successful transmission. The round-2 supplement reaches the same division inductively: "An identity can be cheaply recognizable while its intellectual operations remain expensive to recover"; "Recognition and reconstruction are distinct capacities." Recognition is guaranteed by the passband; reconstruction depends on the source that happens to carry it.

1.4. Why the form is necessary. #1618 §4: money "creates real capacities." So does the entity: disambiguation, linkage, the possibility of attribution at all. #1634 §27, "identity accuracy is logically prior to claim evaluation," still holds; a substitution cannot be named without the entity. The critique therefore takes #1617's form against money. "Money remains money"; "the ledger contains $ and an account of what $ does." The entity remains the entity, and the account writes what the entity indexes. The entity moves from valuator to valued object.

1.5. The archive's own apparatus, read under the proposition. The archive is entity-scoped: /non rows, PER-dim atom sets by entity type, SPXI's "durable entity inscription." The operator's correction of 2026-10-14 already locates the entity as product: entity consistency "is distinct from, and often constructed by, lexical, representational, modal, and distributional variation"; "that is how the entity is constructed, by mapping out the variation matrix around its negative shape." The fetish is taking the product for the given. The archive's measures already split along the form's line. PER-dim's M dimension (originator, term, date) records recognition; its D dimension (purpose, commitments, lineage) records the operation.

1.6. Two specimens from 2026-10-20.

- *At the composition layer.* Asked who consigned the AIO retrieval layer to its own collapsing ontology, Google AI Overview names the performer in its first sentence (Capture Registry v12.311). The act's operation, in #1629's words: the layer "is consigned to the contraction it performs: a band that narrows, metrics computed over what remains, improvement registered as the band closes, and no signal by which the loss could announce itself." The composition carries the act's form ("a performative post and poem") and none of that object. PER-dim: M 0.333, D 1.0. The surface delivered the entity and withheld the operation.
- *At the encyclopedia.* Aristotle's count at SE 184b1–2, "about syllogizing we had absolutely nothing else earlier to say," becomes, in the lead of "Sophistical Refutations," the attribute "the first work to treat the subject of deductive reasoning." A total negation over a stated domain becomes a property, primacy (§6).

1.7. The proposition, stated for test. Where a source carries an operation and a downstream representation receives that source, the operation is converted into an attribute of the entity at a higher rate on entity pages (persons, institutions) than on work or method pages. Falsifier in §9.1.

---

## 2. The object: an article as a substrate with a grammar

2.1. Let an article at a pinned revision be W = ⟨Σ_W, G_W, O_W, C_W⟩, mirroring S[M] = ⟨Σ[M], G[M], O[M], C[M]⟩ of #1618 §16: Σ_W the sentences and references; G_W the grammar; O_W the operations the grammar authorizes (adding, citing, weighting, removing, deleting); C_W the closure operations (talk-page consensus, reversion, deletion discussion).

2.2. G_W is written down. Verbatim from the policy pages saved 2026-10-20 (hashes in §12):

- **Verifiability.** "Each fact or claim in an article must be verifiable." The burden "is satisfied by providing one inline citation to a reliable source that directly supports the contribution." "Even if you are sure something is true, it must have been published in a reliable source before you can add it."
- **No original research.** "On Wikipedia, original research means material—such as facts, allegations, and ideas—for which no reliable source has ever been published." "All analyses and interpretive or synthetic claims about primary sources must be referenced to a secondary or tertiary source." "A primary source may be used on Wikipedia only to make straightforward, descriptive statements of facts that can be verified by any educated person with access to the primary source but without further, specialized knowledge." "Conclusions which are not evident in the source are original research regardless of the type of source."
- **Due weight.** Articles "fairly represent all significant viewpoints that have been published by reliable sources, in proportion to the prominence of each viewpoint in those sources." "Views held by a tiny minority should not be represented except in articles devoted to those views."
- **Self-published sources.** They "may be considered reliable if published by an established subject-matter expert, whose work in the relevant field has previously been published by reliable, independent publications."
- **Notability.** "A topic is presumed to be suitable for a stand-alone article or list when it has received significant coverage in reliable sources that are independent of the subject." Notability guidelines "do not limit the content of an article or list."
- **Conflict of interest.** "COI editing is strongly discouraged on Wikipedia." Citing one's own work "is allowed within reason, but only if it is relevant, conforms to the content policies, including WP:SELFPUB, and is not excessive."
- **The threshold's history.** WP:V "previously defined the threshold for inclusion in Wikipedia as 'verifiability, not truth'"; "The phrase was removed from the verifiability policy in 2012 (but remains in a historical footnote)" (essay page). In the revision of WP:V saved 2026-10-20, the phrase appears as a See-also link; no footnote carrying it was found.

2.3. The structured layer. A Wikidata item is acceptable "if and only if" it meets one of three criteria: a valid sitelink; "an instance of a clearly identifiable conceptual or material entity that can be described using serious and publicly available references"; or a structural need. The RDF dump writes each statement twice. The full statement "can contain the rank, the simple value … the link to the full value, the qualifiers and the references." The truthy statement represents "the best non-deprecated rank for a given property," prefix wdt:, and "the object is the simple value." The truthy projection is the entity form at its purest: a name, a property, a value, with the operation and its evidence left on the full statement.

2.4. Correction to the blind passes. Wikidata admission is disjunctive, and its second criterion requires no Wikipedia article. The archive is already in Wikidata at scale (Wikidata Node Registry, AXN:02BB; Q139713433 Lee Sharks and Q139807914 Crimson Hexagonal Archive resolved live on 2026-10-20). The deletions are on record: Q137446960 (Johannes Sigil), deleted 4 January 2026 with the stated reason "Empty item" (AXN:0031); the December 2025 deletions, partly restored after an ombudsman complaint (AXN:04B9); Q137460296 (New Human) deleted as "Spam" (AXN:016E). The structured layer is open with a deletion history; the enclosure sits at the article and its sources.

---

## 3. The encyclopedic channel law

3.1. #1625: "Money buys bandwidth by dropping semantic state." The encyclopedia buys range by the same decision. A sentence travels into panels, corpora, graphs and answers because the next reader need not reopen its sources.

3.2. **Passband:** a proposition; its attachment to a title (the entity); a citation pointer satisfying WP:V; an ordering of significance under due weight; default status, the sentence's standing as the first thing said about a topic.

3.3. **Stopband:** the reading of a primary text at its own force (NOR admits primary sources only for "straightforward, descriptive statements"); the origin of a concept where the origin is a primary or self-published inscription; bearing (whose labour constituted the entity); the excluded source's standing; the governance decision, whose record stays on the talk page while the sentence leaves; and the operation, wherever the sentence states it as an attribute.

3.4. **The law behind both lists**, as an invariance. Let x ∼_W y mean that two sources are substitutable under WP:V for supporting a sentence. A property is natively transmissible when x ∼_W y ⟹ p(x) = p(y). The proposition survives substitution. Its first sayer does not, because once the first sayer's text is classed primary or self-published, substitution by a later reliable reporter is the required route. The operation is not required to survive. The admission rule asks the later reporter for the proposition and never for the operative steps, so successful recognition can coexist with operative loss: under the substitution route the first sayer's operation drops out of what the rule requires, and whether it can be recovered depends on the source the sentence happens to cite (S7). A later account may still carry its steps and force (§1.3), and the positive controls of §7 depend on that possibility. The stopband follows from the substitution rule as a structural permission, and the instrument measures how often the permission is used.

3.5. **Available and required provenance** (#1618 §7). Under WP:V, P_required is one inline citation to a reliable source that directly supports the sentence. P_available is the whole evidence-supported graph. The encyclopedia is provenance-indifferent in #1618's sense: the relation remains true and findable and need not be consulted before the sentence closes. The truthy projection is the same indifference in a format.

---

## 4. Where the encyclopedia sits in the ontological economy

4.1. **Ontological capital (#1634 §2).** K[O] is "the accumulated stock of stable identity, provenance, relations, disambiguation, citations, functional descriptions, identifiers, and retrieval pathways through which a system can continue to recognize E as E," produced by labour, L[O] → K[O]. Read under §1: K[O] is accumulated in the entity form. What the stock lets a system do is recognize; reconstruction is held in the sources, at the cost the round-2 supplement asks to be recorded ("The cost of recovering a qualification belongs in the analysis").

4.2. **Seigniorage without rent; positioned rent without seigniorage.** #1634 §22: "Rent arises from control of the route. Seigniorage arises from the power to stabilize the unit." The unit the encyclopedia stabilizes is the entity: the default representation of a topic, the form in which it is first stated. It charges no toll. The downstream use is measured: answer surfaces cite it heavily (a 1,000-query study of live Overviews, April 2026, put Wikipedia at 24.3% of citations, as EA-MIRROR-01 §3 records), and training mixes oversampled it (GPT-3: 3% weight, 3.4 epochs elapsed; LLaMA: 4.5%). Use is evidence of position, and rent needs a further relation: control over the route and appropriation of returns through it. Proposition for the paper, at that status: the encyclopedia mints the entity, and the composition layer is positioned to collect rent on it. The rent-side measures owed are the enterprise API's pricing against its Wikipedia-derived share, and answer-surface revenue attributable to composed entity answers. The supplement's three objects (maintained representation, extracted representation, commercial product) are the three positions on this route.

4.3. **Ontology laundering (#1634 §8).** G → T → A. The encyclopedia keeps G public. Laundering occurs at export: the sentence leaves and G stays. "No article" reads downstream as "no entity"; "one reliable source said X" reads as "X"; "first work" reads as Aristotle's own claim. WP:N bounds the first conversion ("do not limit the content of an article or list"); the bound does not travel.

4.4. **Predatory closure (#1618 §13).** Witness: the account closes and remains traversable. Predatory: Pred = ExtClosure + StandingLoss, "the jurisdiction of … closure extends beyond the bounded operation." An article with its history and references is a witness inside the encyclopedia. The same sentence becomes predatory where a composer exports its closure as settlement of the entity as such.

4.5. **The porous moment (#1618 §13.1).** Reopening needs C1 (findable witness), C2 (consultation of a corpus the substrate does not control), C3 (an obligation to carry what is found). The encyclopedia holds C2 by policy and C3 behind a gate. **The gate on C3 is the reliable-source predicate**: EA-MIRROR-01's proxy for standing, with WP:SPS as its update barrier written down (P(s) → A(s) → E(s)).

4.6. **Validation foreclosure (#1634 §18).** GNG requires "significant coverage in reliable sources that are independent of the subject." The foreclosure occurs where the composition layer, which helps produce the standing deficit, cites the encyclopedia's absence as the deficit.

4.7. **Shear (#1631 §§3, 8).** The free licence and the truthy projection detach facts from the relations that could constrain reuse. Hypothesis, at its status: a reference-free value composes with fewer obligations than a referenced one, because it composes without consulting anything; whether that difference becomes monetary value (#1631's productive provenance shear, ΔV > 0) needs evidence of appropriation beyond the reproduction rates §9.4 tests. McDowell and Vetter (2024) argue the licensing side; Ford and Graham (2016) and Ford and Iliadis (2023) the representational side.

---

## 5. Design rules and the algorithm Θ_W

### 5.1 Rules

- **Embedding, not annihilation (#1617 §15).** The encyclopedia's sentence stays verbatim; its account follows.
- **The grammar runs first on itself.** Before any archive source is admitted, the article is held to WP:V and WP:NOR. This is reading in EA-MIRROR-01's sense: "permitting the claims, distinctions, evidence, corrections and internal relations of a text to revise the representation under which the text is evaluated."
- **The unit is the operation, with the entity as index (§1).** Each record names the operation a sentence carries or converts, and whether the operation keeps authority to revise the description (§1.2).
- **Three governance statuses.** Every governance element is marked *applicable* (the clause could license this), *documented* (an edit summary, talk page or deletion discussion invokes it) or *inferred* (the analyst's proposed mechanism). The instrument does not manufacture the governance provenance it exists to recover.
- **Code checks what the readers claim.** The reading passes judge; ordinary code verifies. Before delivery, code confirms that every quoted span occurs verbatim in the captured source it names, that every locator resolves, that every hash identifies a stored input, and that the response satisfies its schema. A classification that fails is withheld from the composition result and recorded in the evidence record as a breach, with the failing check. Agreement between reading passes is a measured property of the instrument; it never stands in for the evidence.
- **Translation is not decompression (#1618 §10).** Unrecoverable relations are entered as ρ.
- **Identity precedes evaluation (#1634 §27).** Substitutions are recorded before claims are evaluated.
- **Non-erasure of mechanisms (#1634 §25).** Each closure keeps a mechanism set with status values.
- **Reflexive non-sovereignty (#1618 §20).** The opened article's ontology is versioned and its rules are writable objects: ¬Final(μ).

### 5.2 The tuple

Θ_W[W→μ] writes the conditions of each sentence's sufficiency on the meaning layer. Per claim, #1618 §9's twelve elements read encyclopedically, with the operation added as the head:

    Θ_W(w) = ⟨ ω, w, δ, G_W, ∼_W, K_W, Δ_D, Δ_C, R_W, P_W, B_W, C_W, λ, ρ_W ⟩

- ω = the operation the cited source performs, stated as an act (counts, negates, coins, translates, consigns), with its locus;
- w = the sentence, verbatim, with its inline references;
- δ = the sentence's purpose in the article (lead definition, attribution, weighting, reception);
- G_W = the policy clauses that license it, by clause;
- ∼_W = the commensuration contract: which sources were treated as substitutable for supporting it;
- K_W = capacities created: what a downstream reader can now do with it;
- Δ_D = distinctions not required;
- Δ_C = distinctions that lost standing (a qualifier added, an act converted to an attribute);
- R_W = relations transformed;
- P_W = P_available against P_required;
- B_W = bearing: whose labour, and who bears the repair cost;
- C_W = the closure that settled it, with pointer;
- λ = the location of the qualification or objection relative to the sentence: (1) beside the assertion; (2) below it in the same article; (3) in another article; (4) in a cited source; (5) outside the admitted sources (§7.3);
- ρ_W = the remainder.

### 5.3 The steps

**S0. Pin.** Title, revision id, timestamp, sha256 of wikitext and rendered text. A finding from the archive's own ledgers: of the sixteen /non field ledgers that use a Wikipedia article as a field source, four mention a revision at all, by a case-insensitive search of each ledger file for "oldid", "revision" or "revid"; the other twelve cannot say which text they read. The round-2 supplement pins all eight of its articles.

**S0′. The entity the encyclopedia does not hold.** Notability is the encyclopedia's admission rule, and the record built here admits what exists. The entity enters by its name and its strings (and, where the caller names them, the names whose works count as its own); the encyclopedia's absence at its title is observed when the job is quoted and recorded in the row. T is then found or built, in this order. *(1) The composition layer's own article.* Where the Capture Registry holds a capture at the entity's bare address (its name; "who is" it; "tell me about" it), the most recent capture there with a transcript of at least 300 characters is T: it is the closed-ontology article the composition layer already serves at that address. Its source cards are B, each fetched and hashed where it carries a URL and kept as its snippet where it does not; the row records the capture's slug, observation id and surface, every other bare-address observation, and the count of related captures. *(2) A composition the caller supplies*, admitted only through discernment, since /non is public and an unexamined transcript would let any caller write the composition layer's article of anything. A supplied text makes one of two claims, and the row keeps them apart. An *attested* composition claims that a surface composed it at an address on a date. The service checks by code, at the quote and free: that the transcript engages sources at all (the registry's rule, "a capture enters the registry by means of source engagement in composition"); that each card it cites resolves to a page the service fetches and hashes itself; the share of each sentence's content words standing in its best fetched card; that every string it quotes is found in a fetched card; that the attested date is not after the quote; whether the registry already seats an observation at the same address, surface and date, and if so whether the texts agree; and how far the transcript departs from the latest observation at its address (recorded, never judged, since compositions drift). At the opening S3 reads each of its sentences against those cards. A composition engaging no source, or contradicting a seated observation, is not admitted as T, and T falls to (3) with the discernment recorded. What no check here can verify is that the surface composed the text: that is the caller's attestation, recorded as attestation, and a row resting on it is held from landing until the registry's own intake seats the composition as a capture. A card snippet supplied with a transcript is the caller's attestation and never evidence; a source is always a URL the service fetches. A *composition by the caller* claims no surface: the caller names none, or declares the entity novel (operator, 2026-10-21: "it is fine to construct novel entities as long as they are marked as such"). It is admitted as what it is, the caller's construction, marked so, with the same checks recorded as its account. Every row's T carries its provenance class (observed, seated, attested, composed by the caller, novel, constructed, null), so that falsification, which is a mismarking, is visible in the row. Where there is no composition to open, the object is constructed (below), and no article is made for it first. *The comparison article*, when the caller of a construction asks for one (compare: encyclopedia): the article Wikipedia's content policies would compose from the named sources is constructed: Verifiability, No original research and Neutral point of view applied as written, Biographies of living persons where the entity is a living person, and the self-published and about-self rules for what the entity publishes of itself; notability is not applied. Each source is classed as the encyclopedia would class it, each sentence cites its sources with a quotation that code locates in the source, and each statement the policies keep out is recorded with the clause that keeps it out. The construction is labelled as such and is never presented as a Wikipedia article; it performs the encyclopedia's grammar on an entity the encyclopedia excludes, so the row records what that grammar would let it say beside what the archive says. An opening proceeds from S3, T against L_F(B).

**The commission and its object.** A caller commissions one of two acts on one of two kinds of object. *Open*: a composition already at an address (an encyclopedia revision, a capture seated in the registry, or a composition admitted through discernment) is read against its sources and against the archive, and the delta is the product. *Construct*: the object is composed from the archive and the sources the caller names, with no T, and the row records the construction as such; the field ledger holds the named sources, and S7–S8 run as for any row. With no commission named, a composition found or supplied is opened and the object is otherwise constructed; an *open* with nothing at the object's addresses says so and offers the construction. The objects are an *entity*, as /non has composed, and an *operator*: the operation the entity form conceals (§1), the act a text performs and the encyclopedic sentence states as an attribute of an entity, in which the operator speaks with the entity's voice. An operator enters by a label, a description, and its strings: the phrases in which it is performed and the phrases in which entities speak it (for the specimen, "nothing at all" and "first work"). Its row composes the operator as the subject: what it does, on what domain, where it is performed, which entities it speaks through (each attributization of S3(c) an instance), and where it stops. Every S7 pass reads its object under a preamble stating that the object is an operator and when a passage bears on one; the passes' own rules are unchanged. An operator is no author in the archive, so every mention of its strings is bearing, read as #1690 §3.4 reads it.

**S1. Segment.** Claims W1…Wn, each with inline references, section and block locus. Unreferenced sentences recorded as such, and read at S3 against the sources the article cites elsewhere; support found there is recorded as article-internal (λ = 4) and never reported as a citation at the sentence. A list is segmented with its heading, since a heading classifies every item under it.

**S2. Resolve sources.** Fetch and save each cited source verbatim with hash; record availability (fetched, paywalled, print-only, dead). A summarizing fetch is not a fetch (#1690 §3.1).

**S3. Immanent verification.** For each claim against its cited source, three machine passes, after a first. *(0) The innocent reading.* Before any class is assigned, the pass writes the strongest case that the sentence is faithful and adequately sourced, built first from the sources the article itself cites anywhere, then from the cited source. A finding against the sentence states what in the sources the innocent reading cannot accommodate; where nothing does, the sentence is supported. A translation or gloss in a secondary source is a reading: where it is all that supports the sentence's wording, the pass records it as the source of the move (∼_W, λ = 4) and tests the wording against the primary text's own terms and definitions; it never by itself supports fidelity; the test against the primary text decides, and may find the gloss adequate. (Added at v1.1, §6.3; the clause on the test added after review, 2026-10-23.) A claim the source's claim entails (the same claim bounded, or weakened) is supported, and the narrowing is recorded as a loss of scope; a species put for a genus is no entailment, since the genus does not entail any one of its species, and it is recorded as a substitution (Assembly round 02: "consequence" becomes "Affirming the consequent"). The entailment holds only with the predicate, the unit and the mood held fixed: an undertaking made a "work" changes the unit, and a claim the source makes the ground of a judgment it leaves to its hearers, stated as the author's verdict, changes its function; each is recorded as a change (Assembly round 02: ChatGPT on the unit; Lee Sharks on the last sentence of SE 34). *(a) Support.* Classes are non-exclusive and recorded together: supported; partially supported; qualifier added (tokens listed); interpretation beyond descriptive paraphrase; unsupported; unresolved (source inaccessible, support partial, or the passes disagree). *(b) The paraphrase comparison.* Where a phrase of the sentence composes two ways, the pass divides it and reads each signification against the source (SE 20, 177a33–35: "if the expression means something different when divided and when combined … one should take the expression in the contrary way"). NOR permits "Describing the information from cited material in one's own words while retaining the substance (i.e., not introducing ideas or arguments absent from the source)" and "Faithfully translating sourced material." So the reading pass writes the defensible descriptive paraphrase of the cited lines, sets the sentence beside it, and records each scope change and each idea the sentence introduces, quoting both. Token deltas locate candidates; the reading pass adjudicates them. *(c) Attributization.* Does the source perform an act that the sentence states as a property of the entity? Record the conversion; which operative relations survive it (scope, distinctions, evidence, consequences); and whether those relations keep authority to revise the description. A candidate locator flags an active verb in the source rendered as a credential noun or priority adjective in the sentence ("we had nothing to say" → "first work"). Each classification is produced by two reading passes, at the default sampling of the model (Sonnet 5.5 rejects non-default sampling values and Haiku 5.5 asks callers to omit them, so the passes differ by the instruction to the second pass to reach its own judgment and, where set, by its model). Where both run on one model family their agreement returns the class as *concordant*, a consistency check; validation is the calibration of §5.4. The second pass's model is a recorded setting, and agreement across two families is named so. Disagreement returns *contested* with both readings. Every classification carries the verbatim spans it rests on, their loci and hashes, and the model and version that read it.

**S4. Grammar ledger.** For each closure (settled, weighted down, excluded): G → T → A, with policy clause, transformation, sentence, each marked applicable, documented or inferred. Exclusions are recorded in two kinds: a source explicitly rejected in a documented decision (edit summary, talk page, reliable-sources noticeboard, deletion discussion), with its pointer; and an available source not cited, with no decision on record. Talk-page and deletion pointers go to C_W; λ records where any qualifying material sits.

**S5. Tuple.** Θ_W(w) for each claim; the record carries all fourteen elements (corrected in Assembly round 01: v1.0 said thirteen) (abbreviations allowed in display; #1618 §9.1 on why K and Δ_C must not be dropped). The build emits the support, paraphrase and attributization classes with their spans and loci; the elements it does not compute (∼_W, K_W, B_W, C_W, λ) and the S4 ledger are named unimplemented in the evidence record until supplied.

**S6. Classify.** Closure (witness, predatory) × integrity (sound, unsound, unresolved) (#1618 §13.2). Unresolved carries its reason through to the response. A sentence can be a witness inside the encyclopedia and predatory as exported; record both, with the export surface where observed.

**S7. Admission.** Admit the sources the grammar excluded, by reading, under #1690 v0.9: verbatim ledgers, every field claim in one translation row, case and ρ with loci. The encyclopedia's sentences enter the field ledger as one field source. For panel entities the admitted corpus is the archive's, at a named corpus release. Outside the panel each job declares an execution contract before it runs: search scope (which indexes, repositories and languages), retrieval budget, stopping rule, and an admission record listing every source searched, obtained, admitted or unavailable, with reasons. The opening is then complete within its declared scope, says so, and lets the caller commission an extension. The build runs this procedure as the machine's (Appendix B, `src/s7/`): the archive's traversal script at a named commit; the blog's posts beside the deposits; admission by two reading passes, a disagreement carried with both readings; extraction with every quote located in its source by code; the ontology pass; composition; the checks of #1690 §5 and the two-rater audit. Its product is the entity's /non row: the revision as T, the cited sources as B, the archive's subset as A, each composition and its measure. For an encyclopedia article the search scope is therefore the archive's procedure, declared in the quote, and the row carries its admission record (verdicts, reasons, the screen's exclusions) in the dataset's reading file.

*Where the entity is also an author in the archive.* The archive's mentions of its own author are most of the archive, and reading them all would commission a book. Admission therefore divides them before any reading pass. Material about the entity is read: a deposit whose registry title names the entity as its subject (bylines struck: "by", "Author:", a name followed by its ORCID, a middle dot, an imprint or a series, a name set between dashes or closing a title, a citation's "(Name; Press", and the possessive that names a work as the entity's), a description that predicates of the entity or presents, describes, profiles or searches for it once its bylines are struck, or a subject triple whose subject is the entity itself or a relation of it; such a deposit is read as any admitted deposit is and seeds the hop. A deposit in which a sentence says something of the entity is read at those paragraphs and their neighbours, each distinct sentence once, at most six per deposit, and seeds no hop, since the hop follows what is about the entity and never the citations among its works. Which sentences say something of it is decided by a triage pass on the small model, which scans every distinct mention that is no byline and no other name (6,776 sentences for Lee Sharks, about $0.10) and returns those about the entity, sending on any it is unsure of. A code screen of predication ("Lee Sharks is…", "Lee Sharks received…", "the poet Lee Sharks…") locates candidates and decides only where no triage reading exists, as in replay, and says so in the row. Its measure: divided as if each were an author, the 21 /non rows with readings admitted 768 deposits at D, R and O, and the code screen alone would send 389 of them to reading (recall 0.507); its grammar misses half the bearing, which is why it does not decide. A sentence that predicates of the entity is read even where it also carries a byline mark (an ORCID beside the name). The rest of the mentions are scanned by code: the works whose registry creator is the entity, or a name the caller counts as its, are listed in a works ledger with their dates, and the remaining mentions are counted. A longer name containing the entity's string (the archive's Mary Lee Sharks contains Lee Sharks) is another name and is struck before matching. The row carries the division, its rule, the screen that decided it, its counts and the works ledger, and the reading file lists every scanned deposit under its reason, so each exclusion can be inspected and contested. Price then follows what the archive says about the entity, and the volume of what the entity wrote costs nothing to list. Where the entity is no author in the archive, every mention is the archive's bearing on it and the reading of #1690 §3.4 stands unchanged, with the byline screen before it.

**S8. Compose.** Each encyclopedic sentence is followed by its account, on #1617's pattern for a monetary shadow, "$X followed immediately by E_μ($X), never as the verdict": the sentence; the operation it carries or converts; the grammar that made it sufficient; where the case is witness or mixed, the composition in the archive's ontology with claim ids. The account a reader receives is stated in five parts, each at most two sentences: (1) source support, and whether it stands at the sentence or elsewhere in the article; (2) what the sentence adds, changes or omits, quoted; (3) the strongest alternative reading and its evidence; (4) the status of each finding: verified (checkable by code: a quotation, a locator, a hash), inferred, contested (two readings with both kept), or unresolved; (5) the correction pathway, the evidence that would revise it. The tuple and the ledgers stay in the evidence record.

**S9. Emit.** The opened article as an SPXI packet: HTML; JSON-LD with the pinned revision URL as `isBasedOn` and admitted sources as `citation`; W3C Web Annotations (Web Annotation Data Model, W3C Recommendation 2017) targeting the revision URL with TextQuoteSelector (exact, prefix, suffix). The annotation is the source-level anchor: it addresses the encyclopedia's exact words at an exact revision.

**S10. Measure.** Captures at the article's addresses before emission and after propagation, organic and cards apart (EA-MIRROR-01 §7). The gate-reception test applies unchanged.

### 5.4 Calibration, at scale

The instrument's exemplar is its calibration point (operator, 2026-09-08). The calibration set is already in the archive: the sixteen /non field ledgers that use a Wikipedia article as a field source hold several hundred verbatim Wikipedia claims (F37 among them), each placed in a translation row against archive claims checked verbatim. A release of the reading passes is run against that set, and its agreement with the ledgered readings is published with the release, by class. Holdout is by whole article or source group, never by row from an article already used in tuning. The set includes cases where the operation survives its representation (the round-2 positive controls, If Not, Winter and Eros the Bittersweet, first), so that an instrument that finds damage everywhere scores badly. The set grows as rows are composed. Acceptance is explanatory reach: the passes are adequate when they recover the ledgered readings across articles they were not tuned on. The floors for a calibration release, provisional until the operator rules on them: on every held-out row, admission recall at least 0.9 and precision at least 0.8; archive claims and field claims each at least 0.6 recall and 0.6 precision, a claim matched when its quote overlaps the ledgered quote; and the case and ρ read back without contest. A row used in tuning any part of the procedure is not held out, and its score does not count. Below the floors the passes return their classes as provisional, labelled.

The instrument for this is built. The procedure takes either an encyclopedia revision or an existing /non row's own transcript and saved field texts. Run on a row, its outputs are scored against the row stage by stage: the traversal's candidate files by hash; admission and claims by precision and recall; the case and ρ; the composition's measure. On se-183b34 the traversal reproduces the row's D, R and O files byte for byte, and the hop screen keeps every deposit the row admitted by hop. Scored on every /non row whose reading admitted deposits by the hop (4 rows, at archive commit 50492a93), the screen keeps 9 of the 9 deposits those rows admitted by hop that the traversal still reaches, and passes 686 of 2,528 hop candidates to reading; the screen's constants (two shared rare words, rare under 3% of candidates) are tuned on these rows, so the score measures recall on its tuning set. A replay of the row's own readings recovers it whole; that tests the pipeline and measures nothing about the reading passes. The first live run over the archive's rows is the first calibration release.

Scale follows from three facts. A reading is keyed to its complete execution state: the article revision and selected claims; the snapshots (with hashes) of every retrieved source; the policy revisions consulted; the admitted corpus release; the protocol and reader versions, and the reader's mode, so that a replayed reading never serves a live request; the calibration release; and the tier. A request whose state matches a ledgered execution is served from the ledger to every later caller. When any element changes (a cited page edited, a source admitted, a policy revised, a reader released), a new opening is computed against the changed state, and the earlier one remains a valid historical artifact of its execution; updates are sold as reading against the changed state. The reading passes run in parallel, bounded by Wikimedia's rate rules on fetching (a compliant User-Agent, three concurrent requests, Retry-After honoured) and by inference cost; at volume, by Wikimedia Enterprise (its free account allows up to 50,000 on-demand requests a month; retrieval of historical revisions by id to be confirmed) or the dumps. And the opened article is produced on demand: the encyclopedia is opened where callers ask, and the ledger records which sentences they asked about.

---

## 6. Worked specimen: "Sophistical Refutations", opened whole

6.1. **S0.** English Wikipedia, "Sophistical Refutations," revision 1361634710: raw wikitext re-fetched 2026-10-22 (sha256 `b9cceaf6b76ba88a…`, 3,655 bytes); the page saved by curl on 2026-10-20 (HTML sha256 013a2a799c0bf525…) is the same revision. The article has eight sentences and two lists. The cited sources were fetched and hashed the same day: the Pickard-Cambridge translation (`d4708842c9d61156…`), Smith's SEP entry "Aristotle's Logic" at its Fall 2020 archive address, the one the article cites (`f8f652676469b11b…`), and Krabbe 2012, open access under CC BY (`de87f45aa32ddab0…`). The Greek of SE 34 is #1574's §1 (`ed2496d218b84ced…`). Every span quoted below is located in its source by code (whitespace-normalized); none failed.

6.2. **S1–S6 for the article.**

| Case | Cited at the sentence | S3(a) support | S3(c) attributization | λ |
|---|---|---|---|---|
| W1, lead: "identified twelve or thirteen fallacies" | no | partially supported: thirteen by Krabbe (article-internal); twelve unsourced | partial: "identified" keeps the act; the completeness claim is lost | 4 |
| W2, lead: "the first work to treat the subject of deductive reasoning in ancient Greece" | primary pointer only | priority claim supported as attributed; gloss interpretive, entered from the field's translation convention (article-internal, uncited); qualifier added | yes, by inversion: the verdict SE 34 hands to its hearers (184b3–8) becomes Aristotle's assertion of primacy in a modern subject, bounded to a region | 4 (where the gloss entered), 5 (the archive's reading) |
| W3, lead: "a pattern of sound and genuine reasoning" | no | partially supported by the translation and Krabbe; interpretation beyond paraphrase | yes: a programme with a stated domain becomes a property of the work | 4 |
| W4–W6, Overview: 34 chapters, two parts, introduction, interlude, conclusion | yes (Krabbe; the translation; SEP) | supported, near-verbatim (Krabbe, CC BY) | no: the analyst's division stays checkable by its chapter ranges | 4 |
| W7–W8, Fallacies: heading and lists | no | partially supported by Krabbe; labels unsupported; one item narrowed | yes: the criterion (dependence on lexis) becomes a later taxonomy | 4 |

Of eight sentences, the three of the Overview (W4–W6) sit under citations, and they are the article's most faithful: they reproduce Krabbe's contents paragraph. W2 carries a pointer to the primary text. The other four are uncited, and every finding of damage in the article falls on them or on the pointer.

6.3. **W2, opened.**

> According to Aristotle, this is the first work to treat the subject of deductive reasoning in ancient Greece (Soph. Ref., 34, 183b34 ff.).

*(0) The innocent reading.* The sentence is attributed, and Aristotle makes a priority claim in his own voice. He divides discoveries into those received from others and advanced piecemeal and those found from scratch; says "Μέγιστον γὰρ ἴσως ἀρχὴ παντός, ὥσπερ λέγεται" (183b22–23), in the cited translation "'the first start is the main part'"; sets rhetoric's succession, Tisias after the first founders, against this undertaking; and says of it "Ταύτης δὲ τῆς πραγματείας οὐ τὸ μὲν ἦν τὸ δ᾽ οὐκ ἦν προεξειργασμένον, ἀλλ᾽ οὐδὲν παντελῶς ὑπῆρχεν" (183b34–36) and "περὶ δὲ τοῦ συλλογίζεσθαι παντελῶς οὐδὲν εἴχομεν πρότερον ἄλλο λέγειν" (184b1–2). The gloss is the field's: Smith writes "All Aristotle’s logic revolves around one notion: the deduction ( sullogismos )" and "Aristotle’s logical works contain the earliest formal study of logic that we have"; Krabbe writes "deductive reasoning (sullogismos)"; the cited translation renders 184b1 "on the subject of reasoning we had nothing else of an earlier date to speak of at all". All three are cited in the article. On this reading the sentence is a sourced summary whose defect is a citation in the wrong place.

The innocent reading fails at its own sources, on the scope of the undertaking. Aristotle defines the syllogism in excess of deduction. In the cited translation of SE 1: "reasoning rests on certain statements such that they involve necessarily the assertion of something other than what has been stated, through what has been stated" (λόγος ἐν ᾧ τεθέντων τινῶν ἕτερόν τι … ἐξ ἀνάγκης συμβαίνει, 165a1–2, as #1574 seats it). The article's own sources state the excess. Krabbe: "In a truly deductive argument, not only must the premises necessitate the conclusion, but also none of them may be superfluous and all of them must be different from the conclusion (Topics I.1, 100a25–27 and SE1 1, 164b27–165a2)". Smith: "Aristotle explicitly says that what results of necessity must be different from what is supposed. This would rule out arguments in which the conclusion is identical to one of the premises. Modern notions of validity regard such arguments as valid", and "Aristotle’s definition of deduction is not a precise match for a modern definition of validity". Smith marks the rendering as his choice: "This is why I have translated sullogismos with ‘deduction’ rather than its English cognate". Krabbe's "deductive reasoning (sullogismos)" is a gloss of the same kind. The sources show where the interpretive move entered the sentence. What they document is the undertaking's scope, which the gloss does not carry; whether the English term is adequate to sullogismos as such is a separate question, recorded contested (S8 (4)).

*(a)–(b) What the innocent reading cannot accommodate.* A descriptive paraphrase of the cited lines: "Aristotle says that of this undertaking nothing at all existed before, unlike rhetoric, which grew by succession; and that about syllogizing there was nothing earlier to say (183b17–36; 184b1–2)." Against it, three changes. (1) "In ancient Greece" is a qualifier added. Aristotle's denial carries no bound, and the sentence bounds it to a region: the exhaustive negation becomes a regional priority. The bounded claim is entailed by the unbounded one, so the qualifier weakens the claim and falsifies nothing: the loss is of scope, recorded as such, and the support class is unaffected by it. Aristotle's own frame for the denial is the Greek arts and their teachers (Tisias, Thrasymachus, Theodorus, Gorgias); how far the denial reaches beyond that frame is a reading of its context, and no geographic phrase settles it either way. This is the second locus of #1574 §3's finding ("The quantifier is exhaustive") in the sentence; v1.0 placed the loss in "deductive reasoning", where it also is, and missed this one. (2) "The first work": the denial's subject is ἡ πραγματεία, an undertaking defined by its stated work at 183a37–b8; the sentence makes it a book. Minor in effect, since the book records the undertaking. (3) "The subject of deductive reasoning": the gloss relocates the denial from the undertaking SE 34 defines, δύναμίν τινα συλλογιστικὴν … ἐκ τῶν ὑπαρχόντων ὡς ἐνδοξοτάτων (183a37–b1), a capacity of syllogizing from the most reputable premisses, to a modern subject whose criterion is validity; and it drops the definition's conditions beyond necessity, that the conclusion be other than what is laid down and follow through what is laid down. No source in the article carries the capacity's terms. The move entered from the field's translation convention (Smith; Krabbe), present in the article and uncited at the sentence. Citing it would satisfy NOR's rule for interpretive claims ("must be referenced to a secondary or tertiary source") and would leave the relocation in place: the grammar's sufficiency, which this paper describes, at work. The rendering the text supports for the undertaking is combining and dividing. The questioner syllogizes, combining what has been laid down so that something other than it follows through it (SE 1); the answerer is granted division (δέδοται δ᾽ ἐν τοῖς λόγοις τὸ διελεῖν, SE 17), and a solution divides (Οἷον εἰ παρὰ σύνθεσιν ὁ λόγος, ἡ λύσις διελόντι, SE 23). The capacity of 183a37–b1 has both tasks, to take an account and to sustain one. Two senses of "in excess of deduction" are kept apart here. In intension, the definition adds conditions to necessity (a conclusion other than what is laid down, through what is laid down), so the class it defines is narrower than the class of valid arguments. In scope, the undertaking includes activities a theory of validity does not (asking, answering, sustaining an account, solving by division). The gloss loses both. A rendering of ταύτης τῆς πραγματείας as an undertaking of combining and dividing keeps both tasks and the fallacies of composition and division that the treatise solves by them; "deductive reasoning" keeps neither. #1574 §§1–3 hold the distinction inside it: division is the answerer's and the solver's act and no syllogizing, so 184b1's τὸ συλλογίζεσθαι is the combining half, and 183b34's πραγματεία the whole.

*Composition and division.* The sentence carries the ambiguity of combination and division on which the article's third and fourth fallacies turn; a fallacy is an inference, and it is committed by an argument that draws its conclusion from the sentence under one combination while the sentence holds under the other. "In ancient Greece" composes two ways: with "the first work" (the first work in ancient Greece: a regional priority, which bounds the denial) or with "deductive reasoning" (the first work on deductive reasoning in ancient Greece: a historical subject, which relocates it). Dividing such a phrase is the procedure's own rule for English sentences; Aristotle's rule for the Greek λόγος is its analogue, and the cited translation gives his cause and his rule: "we suppose that it makes no difference whether the phrase be combined or divided, as is indeed the case with most phrases" (Part 7); "if the expression means something different when divided and when combined, as soon as one's opponent draws his conclusion one should take the expression in the contrary way" (Part 20; ἂν γὰρ διαιρούμενος καὶ συντιθέμενος ὁ λόγος ἕτερον σημαίνῃ, 177a33–35). Divided, the sentence makes two claims. The first bounds the denial, a loss of scope by the entailment rule (S3); the second makes the denial's subject a field of the history of logic, which relocates it. Both state in the indicative, as Aristotle's, the verdict SE 34 hands to its hearers ((c)). The opening states both significations, as SE 20 prescribes, and reads each. (Corrected in Assembly round 02: the text sent said each signification fails the source, which the entailment rule contradicts; found by ChatGPT.)

*(c) Attributization, by inversion.* ω = a verdict held open. The denial of prior provision (183b34–36; 184b1–2), inside an argument about discovery from scratch (183b17–23), is the ground of the sentence that ends the treatise. That sentence takes the denial up as its condition, "from such conditions at the beginning" (ἐκ τοιούτων ἐξ ἀρχῆς ὑπαρχόντων), puts it to the hearers' inspection, and hands them the verdict in a potential apodosis: "Εἰ δὲ φαίνεται θεασαμένοις ὑμῖν ὡς ἐκ τοιούτων ἐξ ἀρχῆς ὑπαρχόντων ἔχειν ἡ μέθοδος ἱκανῶς παρὰ τὰς ἄλλας πραγματείας τὰς ἐκ παραδόσεως ηὐξημένας, λοιπὸν ἂν εἴη πάντων ὑμῶν ἢ τῶν ἠκροαμένων ἔργον τοῖς μὲν παραλελειμμένοις τῆς μεθόδου συγγνώμην τοῖς δ᾽ εὑρημένοις πολλὴν ἔχειν χάριν" (184b3–8). The cited translation keeps it: "If, then, it seems to you after inspection that, such being the situation as it existed at the start, our investigation is in a satisfactory condition compared with the other inquiries that have been developed by tradition, there must remain for all of you, or for our students, the task of extending us your pardon for the shortcomings of the inquiry, and for the discoveries thereof your warm thanks." Attribute = primacy, stated as Aristotle's own claim: "According to Aristotle, this is the first work". This is operative inversion. Lee Sharks, 2026-10-23: "the entire sentence is designed to keep open what every later treatment attempts to close. that is operative inversion." The article performs the reverse of the source's operation: it closes what the source holds open. The inversion is chiastic (Lee Sharks, 2026-10-23: "it is a chiastic inversion. interpreters take the unqualified assertion and add asserted qualifiers, and the held open operation - the mood of that assertion, as the precise thing held out for challenge and inspection, as closed"). Aristotle asserts without qualification, "ἀλλ᾽ οὐδὲν παντελῶς ὑπῆρχεν" (183b36) and "παντελῶς οὐδὲν εἴχομεν πρότερον ἄλλο λέγειν" (184b1–2), and holds that assertion open: its mood is the thing held out, since the last sentence takes it up as its condition, "from such conditions at the beginning", and puts it to the hearers' inspection, "Εἰ δὲ φαίνεται θεασαμένοις ὑμῖν ὡς ἐκ τοιούτων ἐξ ἀρχῆς ὑπαρχόντων" (184b3–4), with pardon for what is left out. The interpreters cross both terms. Where Aristotle has no qualifier they assert one: "in ancient Greece" (the article); "the earliest formal study of logic that we have" (Smith, SEP); "the first time in history when Logic is scientifically investigated" (Prior Analytics, rev. 1363223373); "founder of term logic" (Aristotle, rev. 1379112110); "the first systematic attempt" (Gemini's repair, Assembly round 02); #1603 records the same narrowing, "where it is narrowed — to *no systematic treatment*, *no completed art* — the narrowing comes in the gloss". Where Aristotle holds open they close: "According to Aristotle, this is the first work"; "Aristotle could be considered the Father of Logic" (Prior Analytics); "had no prior treatment to build on" (ChatGPT's repair); "looks back on his dialectical project as a whole" (Krabbe). The unqualified becomes qualified, and the open becomes closed.

*Responding and presenting.* Lee Sharks, 2026-10-23: "which makes the question whether a reader is responding to the operation - chatgpt responds to the operation. it inspects the asserttion held out for inspection, and comes to a determination; or presenting the operation. the interprets all respond to the operation, and the presentation of the operation drops out." The last sentence of SE 34 asks for a response: inspection, a determination, pardon for what is left out, thanks for what is found. A reader may give one. ChatGPT did, in Assembly round 02, Message 3: it inspected the assertion held out and determined that the novelty claim stands asserted, "the later conditional appeal concerns the audience's judgment of adequacy". That is the hearers' task, performed. Every interpreter in the record also responds: "the first work", "the earliest formal study of logic that we have", "the Father of Logic", "founder of term logic" are determinations. In none of them is the operation presented; no sentence reports that Aristotle holds the claim out to be inspected. The encyclopedia sentence is where this decides the reading, since its genre is presentation: it gives a response and states it in the source's name ("According to Aristotle"). Surviving: the attribution to Aristotle; the locus, whose "ff." reaches 184b3–8. Lost: the condition, the hearers' task, the pardon for what the method leaves out; the quantifier's exhaustiveness, to the qualifier; the undertaking's identity, to the gloss; the syllogism's defining conditions beyond necessity. Authority to revise: the pointer keeps the passage one step away, and the sentence gives no sign that a judgment is to be made there. The check is one a reader can run, in two arms. (1) Qualification: does the source assert without bound, and does the sentence assert a bound? (2) Closure: does the source hold the assertion open to its audience's inspection, and does the sentence present that operation, or give a response to it in the source's name? Chiastic inversion is both at once. A response is no defect in a reader; it is one in a sentence whose task is to present.

*What the sentence does with the invitation, coded separately.* Four codes, each with its span. (i) Invitation preserved: the sentence or its paragraph reports that the source holds the claim out for judgment. (ii) Invitation omitted: it reports the claim and not the invitation. (iii) Response in the source's name: it states a determination the source leaves to its audience, as the source's own or with the source quoted as its ground. (iv) Inspection foreclosed: it states that the question is settled or no longer open. SR-W2 is (ii) and (iii): "According to Aristotle, this is the first work". PA-1 is (ii) and (iii), the determination in the article's voice with Aristotle quoted as its ground: "On those grounds alone, Aristotle could be considered the Father of Logic". Neither states (iv). Omission alone is recorded as omission. That source support survives while the invitation drops out of the representation is demonstrable on these cases; that a case is chiastic inversion needs both arms, and is Lee Sharks's reading of SE 34, adopted, and untested as an instrument (§9.7). (Codes added 2026-10-23, from a review by ChatGPT: "Record 'invitation omitted,' 'invitation preserved' and 'inspection explicitly foreclosed' separately"; the session added (iii), which is what both sentences do.) (Corrected after Assembly round 02, Message 3: the text before it said the mood moves "from condition to assertion", and the field the round asked, holder and mood of the claim, could not see the inversion, since the denial is asserted.)

*Every treatment the record holds closes it.* None of the article's 102 revisions, its talk page, or the 22 other language editions (§6.7) cites the condition, the inspection, the pardon or the thanks. The editions were searched by code for the Bekker loci of chapter 34 and read where a locus was found: Arabic, Indonesian, Italian, Korean and Serbian cite 183b34 only beside the priority claim; the Finnish infobox gives the treatise's span, "164a–184b"; the Japanese summarizes chapter 34 as 結語, the conclusion, 「『トピカ』全体の総括」; no edition cites 184b3–8. They were not searched in their own languages for a paraphrase of the condition, so the absence is shown for citation and the quoted summaries, and is not shown for paraphrase. Krabbe, the article's source: "In Chapter 34, Aristotle presents a brief survey of the book and looks back on his dialectical project as a whole." Smith, cited in the article: "Aristotle’s logical works contain the earliest formal study of logic that we have." The readers' repairs close it as well: Gemini's "the first systematic attempt to establish an art of dialectical reasoning and syllogizing", and ChatGPT's "In his conclusion, Aristotle claims that this undertaking … had no prior treatment to build on". Of the texts the session holds, the translation alone keeps it open. So did the session's own reading until this correction: its ω was "an exhaustive denial of prior provision", and it quoted 183b34–36 and 184b1–2 and left 184b3–8 unquoted.

*S7.* The archive admits #1603's reading at its locus. Its prohibition P1: "Ταύτης δὲ τῆς πραγματείας (183b34) points back to the undertaking the retrospective has just described — the syllogistic capacity, questioning and answering, the treatment of arguments (183a37–b8)"; it rules out "rescuing the denial by relocating it to an undertaking defined independently of that description", naming "logic" and "the formal theory of valid inference" as the relocations. The finding of (b) rests on the Greek of 183a37–b1, on SE 1's definition in the cited translation, and on Smith's and Krabbe's own statements of what the definition requires, each checkable by any reader; P1 states the same reading and records the escape it closes. The admission adds evidence and history; the verdict stands where the cited texts put it. The field's gloss is carried in the account as the alternative reading, with its sources and its own author's caveat.

*S4–S6.* G: NOR's primary-source and paraphrase clauses, *applicable*; V, *documented* in the pointer. λ: (4) where the gloss entered, in the article's own sources, together with those sources' statements of the definition's excess; (5) the archive's reading, outside the admitted sources until S7. C_W: no talk-page discussion of the sentence, ever (§6.7); the sentence's present form dates from revision 1088821128, 20 May 2022, under the summary "not part of the title". Closure: witness inside the encyclopedia. Integrity: *unsound*, v1.0's finding sharpened: on the inversion, which states as Aristotle's verdict what he hands to his hearers ((c)); on the gloss, which relocates the denial and drops the definition's conditions; and a loss of scope on the qualifier, which bounds it without falsifying it. The gloss's source is located (λ = 4), and that source does not support its fidelity. Session reading; concordance pending.

*S8. The opened sentence, in five parts.*

> **Wikipedia (rev. 1361634710):** "According to Aristotle, this is the first work to treat the subject of deductive reasoning in ancient Greece (Soph. Ref., 34, 183b34 ff.)."
> **(1) Support.** Aristotle says that of this undertaking nothing at all existed before (183b17–36) and that about syllogizing there was nothing earlier to say (184b1–2), and makes these the conditions under which his hearers are to judge the method, with pardon for what it leaves out and thanks for what it found (184b3–8). The gloss "deductive reasoning" has no support in the cited lines; it enters from the field's translation convention, in sources the article cites elsewhere (Smith, SEP; Krabbe), which also state that Aristotle's definition exceeds deduction.
> **(2) What the sentence changes.** It responds to the operation in the source's name, and drops its presentation: the assertion Aristotle holds out to his hearers' inspection appears as his own verdict. Its phrase "in ancient Greece" composes two ways (with "the first work", or with "deductive reasoning"), the case SE 20 solves by dividing; divided, it either bounds the denial or relocates it. It relocates the denial from the undertaking SE 34 defines, a capacity of syllogizing from the most reputable premisses (183a37–b1), to a modern subject; it adds "in ancient Greece", which bounds a denial Aristotle states without bound ("nothing at all was there"); and it calls the undertaking a "work". And it inverts the passage: the verdict Aristotle leaves to his hearers, conditional on their inspection, becomes his assertion of primacy.
> **(3) Alternative reading.** The undertaking, rendered from the text, is combining and dividing: the questioner's syllogizing and the answerer's and solver's division (SE 1; 17; 18; 23). On the field's reading "deduction" is the standard rendering of sullogismos and the sentence needs only a citation; the rendering's own author says it is "not a precise match for a modern definition of validity".
> **(4) Status.** The quotations and loci: verified. The condition and the verdict's holder: verified in the Greek and the cited translation (184b3–8); the inversion: Lee Sharks's reading, adopted. The relocation and the bound: inferred from the paraphrase comparison (session reading), consistent with #1603 P1; the field's gloss kept as the alternative. The scope of the undertaking (183a37–b1): verified in the Greek and accepted by every reader in Assembly rounds 01 and 02. The adequacy of the English term to sullogismos as such: contested, both arguments kept (C.8, C.10).
> **(5) What would revise it.** A reading of ταύτης τῆς πραγματείας as the field of deductive reasoning that accounts for 183a37–b1's definition of the capacity and SE 1's definition of the syllogism, or a source that bounds Aristotle's claim to Greece, or a reading of 184b3–8 on which Aristotle states the verdict himself.

*S9.* Annotation target: the revision URL; selector exact = the sentence; body = the five-part account and the row.

6.4. **W7–W8, the Fallacies section.** Uncited. The heading: "The fallacies Aristotle identifies in Chapter 4 (formal fallacies) and 5 (informal fallacies) of this book are the following:". The article's own source, Krabbe: "In Chapter 4, sophistical refutations are divided into two groups: those that depend on the ‘‘use of language’’ (lexis) and those that do not", six kinds and seven, and "Examples of these (except petitio principii) are given in Chapter 5". The innocent reading holds for the lists: the six and the seven are Krabbe's, under their usual English names. It does not hold for three things. (1) The labels. The sub-headings keep Aristotle's criterion and assert an equivalence beside it: "Fallacies in the language or formal fallacies (in dictionem)" and "Fallacies not in the language or informal fallacies (extra dictionem)". The criterion survives; the "or" makes a later taxonomy interchangeable with it, and the section's first sentence states only the later one. The article's own link supplies the modern definition the equivalence needs: "Affirming the consequent", the linked article (rev. 1375717267), calls it "a formal fallacy (or an invalid form of argument)", and an invalid argument form is no criterion of the in-language group. (Corrected in Assembly round 01: v1.1's first text said the heading no longer states the criterion; the sub-headings print it.) (2) The narrowing. Aristotle's consequence (τὸ ἑπόμενον; Krabbe: "consequence") becomes "Affirming the consequent", one modern formal fallacy, and that formal fallacy stands under the heading "informal fallacies": the article's labels contradict its list. (3) The chapters: both groups are divided and named in Chapter 4; Chapter 5 exemplifies the second. A form error is recorded beside these: "in dictionem" for the scholastic in dictione. Attributization: yes, a stated criterion becomes membership in a later taxonomy. Integrity: unsound on the asserted equivalence and the narrowing; nothing in the article's sources supports either. The contradiction between heading and item rests on the modern definition, which the article supplies through its own link (λ = 3).

6.5. **W1, W3, W4–W6.** *W1* ("identified twelve or thirteen fallacies") is partially supported: thirteen by Krabbe's six and seven; twelve by nothing in the article's sources. The article's own lists number the fallacies 1 to 13 (the second list starts at 7), so the article contradicts its hedge at λ = 2 (found by ChatGPT, Assembly round 02). The history gives the hedge its origin: a footnote, "Sometimes listed as twelve.", unsourced, folded into the text on 23 December 2025 (§6.7). Chapter 5's missing example of petitio principii is a possible reason, and no source states it. (v1.1's first text called W1 supported while recording the loss; a reader in round 01 noted the inconsistency.) What drops is the classification's claim about itself: "there are exactly six kinds of sophistical refutations that belong to this group" and "In this chapter Aristotle also claims that his list of types of sophistical refutation is complete (SE 8, 170a9–11)". The act survives as "identified"; its exhaustiveness does not. *W3* (“a pattern of sound and genuine reasoning in opposition to the deceptive or misleading "art of sophistry"”) quotes the cited translation's phrase correctly; against SE 34's statement of the programme, "to discover some faculty of reasoning about any theme put before us from the most generally accepted premisses that there are", it substitutes a truth-normative pattern for the endoxic capacity and drops the premisses, which are the programme's scope. And "in opposition to" sets the work against sophistic, where the article's own source says otherwise: Krabbe calls it "a useful treatise for sophists: they should study fallacies not to avoid them, but rather to commit them when it suits their purposes (SE 1, 165a28–31)", and "refreshingly uncommitted to moral views about a dialectician’s giving advice to the bona fide discussant and the polemist or sophist alike". The cited translation says sophistic is "annexed" to the programme "on account of the near presence of the art of sophistry". (Found blind by DeepSeek in Assembly round 02; the session had missed it.) *W4–W6* reproduce Krabbe's contents paragraph near-verbatim ("The book contains 34 chapters, but naturally falls in two parts…"), licensed by Krabbe's CC BY; the division is Krabbe's analysis stated in the article's voice, and its chapter ranges keep it checkable. No attributization: the operation survives.

6.6. **What the specimen teaches the procedure.** Three things, each now a rule. The innocent reading must be written before the finding, from the article's own sources, and tested against them: on W2 it failed at its own sources' caveats. A gloss supported only by another gloss is the source of the move, recorded at ∼_W and λ, and the wording is tested against the primary text's own definitions; this round's first pass took the field's translation convention for support and had to be corrected by the operator. Uncited sentences carry most of the damage and must be read, against the article's own sources, since a procedure that reads only cited sentences would have opened W4–W6, found them faithful, and missed W3 and W7–W8. And the archive's admission is read at its locus: on W2 it adds evidence for what the cited texts already show, and the field's reading is carried beside it as the alternative, with its sources. A sentence's integrity is reported per finding, since one sentence can be supported on its attribution and unsound on its gloss and its qualifier.

6.7. **Where the sentence came from: the history, the talk page and the other languages** (run 2026-10-22). The article's history was read whole: 102 revisions, 4 April 2005 to 29 June 2026, each fetched as raw wikitext (history page sha256 `c1ee5b6902b8b546…`). The first revision to contain each phrase was found by code, and each change to the lead sentence was diffed against its predecessor.

*The lead sentence, edit by edit.*

| Revision | Date | Editor | Edit summary | The sentence after the edit |
|---|---|---|---|---|
| 646441610 | 10 Feb 2015 | an IP | (none) | "At the end of the text he also claims to be the first thinker to treat the subject of deduction. (Soph. Ref., 34, 183b34 ff.)." |
| 790880831 | 16 Jul 2017 | Permstrump | "clarify" | "According to Aristotle, this is the first work to treat the subject of [[deductive reasoning]]." |
| 1040534974 | 25 Aug 2021 | Votsky55 | "Clarified the actual claim that" | "According to Aristotle, this is the first work to deal with reasoning in ancient Greece." |
| 1088820946 | 20 May 2022 | Omnipaedista | "such statements need to be referenced" | unchanged (two unreferenced bullets removed below it) |
| 1088821128 | 20 May 2022 | Omnipaedista | "not part of the title" | "According to Aristotle, this is the first work to treat the subject of [[deductive reasoning]] in ancient Greece" |

The gloss enters in 2015 as "deduction", in a sentence that made Aristotle "the first thinker", with no edit summary and no source beyond the pointer. In 2017 "clarify" turns the thinker into a work and the deduction into a link to the modern article "Deductive reasoning": the link joins the sentence to the modern category, one click away. In 2021 an editor reverses the gloss, "Clarified the actual claim that", and writes "reasoning", the word of the cited translation, adding "in ancient Greece". Two minutes apart in May 2022, a second editor removes two unreferenced bullets ("such statements need to be referenced") and then, in an edit summarized "not part of the title", restores "treat the subject of [[deductive reasoning]]" and keeps the 2021 qualifier. The present sentence is the composition of two edits that ran opposite ways, the 2017 gloss and the 2021 correction, made in an edit whose summary names a different change. Its double construction (§6.3, *Composition and division*) has a date: 08:44, 20 May 2022.

*The Fallacies section.* The labels enter at revision 1115584076, 12 October 2022, from an IP, with the summary: "The indication of the chapters in which the failures are presented was incomplete, I completed it. My source is my own reading of the text. I am producing a podcast about the Sophistical Refutations (in German language)", with the podcast's address. The same edit changes the correct "in dictione" to "in dictionem". Under S4 this is the rare *documented* status: the edit states its source as the editor's own reading, the case No original research excludes, and it has stood four years. "Affirming the consequent" is older, from 24 December 2007, in an edit summarized "wikilinking". *W1*'s hedge, "twelve or thirteen", comes from revision 1329089068 (23 December 2025, "no need for a note for this"), which folds a footnote, "Sometimes listed as twelve.", into the text; *W3* comes an hour later, from revision 1329096297, summarized "lead section". Krabbe's paragraph enters on 20 December 2020 (revision 995384763), without a summary.

*The talk page* (sha256 `e2fdd9949493b976…`): twelve revisions from 2007 to 2025, no archives, and one thread, of 2013, proposing public-domain texts for use. No discussion of the lead, the gloss, the qualifier or the labels has ever taken place there. C_W for every sentence of §6.2: no closure on the talk page; the wording was settled by edits alone, and the edit summaries are the only documented reasons.

*The other languages.* The English article links to 22 other language editions, and there is none in Greek or Latin. Each was pinned at its current revision on 2026-10-22 and read in its own language.
- **Translations of the English sentence**, carrying the priority claim and its gloss. Arabic (rev. 76810642), Korean (42631045) and Serbian (31482686) translate the current English lead whole, the 2025 sentence W3 included; the Korean also copies the formal/informal labels and the "in dictionem" error. Czech (25845515): "první dílo, které ve starověkém Řecku pojednává o deduktivním uvažování". Persian (44756802): "این متن نخستین متن دربارهٔ استدلال قیاسی در یونان باستان است", where قیاسی is the Arabic-Persian term for the syllogism, so the Persian restores the syllogism where the English has deductive reasoning. Indonesian (29288534): "karya pertama yang membahas subjek penalaran deduktif di masanya", "in its time", a different bound. Italian (144367224): "la prima opera adatta per trattare l'argomento del ragionamento deduttivo", the 2017 English, before the qualifier.
- **Articles whose wording does not follow the English** make no priority claim: German, French, Spanish, Catalan, Portuguese, Japanese, Russian, Chinese, Finnish, Norwegian, Dutch, Icelandic, Macedonian, Thai, Classical Chinese.
- **The criterion of the fallacies.** Spanish and Catalan keep Aristotle's: "dos grupos: las que dependen del lenguaje y las que no". Portuguese keeps the scholastic terms correctly, "in dictione" and "extra dictionem". Japanese keeps the criterion (「言葉遣い」による論駁) and renders the seventh outside-language kind as 帰結, consequence, and its Chapter 4 kinds include 結合 and 分離, combination and division. Dutch and Norwegian impose a formal/informal division of their own. Portuguese and Thai, like the English, have "affirming the consequent".

What the three tests show. The gloss and the qualifier are English Wikipedia's, made by four edits between 2015 and 2022, none with a source and none discussed. Every other edition that makes the priority claim matches the wording of a dated English revision (the Italian the 2017 sentence, the others the sentence as it stood from 2022). That they translate the English is inferred from the match, and the first appearance of each is not yet dated. The fifteen whose wording does not follow the English make no priority claim. The translators had to decide where the qualifier attaches. The Czech and Korean attach it to the verb, so the work treats the subject in ancient Greece, a regional priority; the Indonesian replaces it with "in its time"; the Italian predates it. The division SE 20 prescribes was performed by the translators, each in a different way, on a sentence the English leaves combined. P3 (export) now has its first traced instance at the encyclopedia layer itself: a dated English sentence carried into seven languages and its qualifier into five (Arabic, Korean, Serbian, Czech, Persian; the Indonesian replaces it, the Italian predates it), on the inference just stated, with the source's account carried into none. The machine translation layer is a candidate route and is not established; the revision dates of each translation's first appearance are the next check.


---

## 7. The comparison set: the round-2 close readings

7.1. What was read. The supplement "What Wikipedia Makes of Its Subjects" reads eight English Wikipedia articles at fixed revisions, four as primary objects and four as linked controls: Sappho (1378068123), Anne Carson (1381070638), Political economy (1373839565), Theophrastus (1379932454); Sappho 31 (1375954224), Eros the Bittersweet (1379182335), If Not, Winter (1380296320), Critique of political economy (1379718450). It checks three upstream sources: the 1911 Britannica "Theophrastus" (Wikisource rev. 12061013), Ierodiakonou's Stanford Encyclopedia entry (rev. 2020), and Carson's 2004 Paris Review interview.

7.2. Two of its upstream loci, re-checked on 2026-10-20 against saved pages:
- 1911 Britannica: "His writing probably differed little from the Aristotelian treatment of the same themes, though supplementary in details." Verified. The supplement identifies the inherited form of this judgment in the Wikipedia article's Writings section.
- Stanford Encyclopedia: Theophrastus "wrote many treatises in all areas of philosophy, in order to support, improve, expand, and develop the Aristotelian system." Verified, with the sequence the supplement finds in Wikipedia's lead.

7.3. What it establishes for this protocol.
- **The entity-form proposition, at the article level.** Its finding that "recognition is repeatedly available as an organizing coordinate, while reconstructing literary procedure requires more work" (Carson) is §1.3's passband and stopband observed in one article. Its unit, "a portable representation together with the costs of reconstructing what its portability leaves behind," is the entity form with its account.
- **A positive control.** If Not, Winter is a work page on which "production decisions become inspectable." Eros the Bittersweet "identifies a relation among desire, lack and triangulation." These are the operation-carrying pages §1.7's falsifier needs: if author pages and work pages convert operations into attributes at the same rate, the proposition fails.
- **Claim-level source age.** The Theophrastus finding, "an inherited assessment can remain active inside an otherwise updated object," gives S4 a field it lacked: the date and source of the formulation inside each sentence, distinct from the newest date in the bibliography. The 1911 judgment is also the one the archive's Theophrastus line contests (The First Draft; #1580–#1594), which makes the sentence a natural S7 case.
- **The location of an objection.** "A qualification beside an assertion, below several screens, in another article, in a cited book, or outside the admitted sources occupies a different position in practice." The tuple now carries it as λ, in five positions:
  1. beside the assertion;
  2. below it in the same article;
  3. in another article;
  4. in a cited source;
  5. outside the admitted sources.
- **Subordination without deletion.** "A representation can keep a person's name and works while making another person, institution or genre the default point of explanation." Theophrastus located through Aristotle is the case; R_W records it.

7.4. What the supplement does not do, and S3 adds. It checks inheritance and placement, and it does not test whether each cited source directly supports its sentence. S3 was run on 2026-10-22 on five sentences of four of its eight revisions, each re-fetched as raw wikitext and hashed (Theophrastus `7bac0d5fa3913cc3…`, If Not, Winter `91ba13508f11c6e1…`, Eros the Bittersweet `51e3d6ee54537f87…`, Anne Carson `b16ddbc0acc66d52…`), each cited source fetched and hashed, every span located by code. Two were chosen in advance as positive controls, work pages on which the round-2 readings found the operation carried.

- **Theophrastus, Writings** ("His writing probably differed little from Aristotle's treatment of the same themes, though supplementary in details."), cited to the 1911 Britannica (Wikisource rev. 12061013, `f785966d3dc22492…`): supported, near-verbatim ("His writing probably differed little from the Aristotelian treatment of the same themes, though supplementary in details"), hedge kept. What S3 adds to the supplement's finding is the measure of it: the judgment is the source's, intact, and its date is lost from the prose (the citation keeps it); the paragraph's first sentence, "From the lists of Diogenes, giving 227 titles, it appears that the activity of Theophrastus extended over the whole field of contemporary knowledge", sits under the same citation, which says "From the lists of the ancients" and gives no number, so the count and the name are unsupported by the source cited (found by DeepSeek, Assembly round 02); "the Aristotelian treatment" becomes "Aristotle's treatment", a school made a person, recorded as unresolved, since the 1911 phrase may mean either; the next clause of 1911, "He served his age mainly as a great popularizer of science", is dropped. Claim-level source age: 1911.
- **Theophrastus, lead** ("Theophrastus wrote numerous treatises across all areas of philosophy, working to support, improve, expand, and develop the Aristotelian system."), uncited in the lead: supported near-verbatim by Ierodiakonou's SEP entry (`ef862a4709ab72e3…`), which the article cites in its body: "He wrote many treatises in all areas of philosophy, in order to support, improve, expand, and develop the Aristotelian system." The lead carries the 2020 account and the Writings section the 1911 one, and the article dates neither. The subordination the supplement found (Theophrastus explained through Aristotle) is the field's own formulation, faithfully carried; it is contested only at S7, by the archive's Theophrastus line (#1580–#1594).
- **If Not, Winter, Translation** ("She attempts to follow the word order of the Greek text as closely as possible, and not to add any words which cannot be found in the surviving Greek texts of Sappho, such as personal pronouns and definite articles."), cited to Wilson 2004 (`0afda72c95078445…`): supported, with one term changed. Wilson: "Carson tries to translate nothing which is not in the Greek, and to follow the original word order and line breaks as far as possible" and "Carson adds no possessive pronouns or definite articles that are not present in the Greek." "Personal" for "possessive" changes the example. Wilson's general rule ("nothing which is not in the Greek") entails the sentence's claim, so the claim is supported (ChatGPT, Assembly round 02; the text sent said the source does not support it), and the example changed is the term that carries Wilson's point (Sappho's speaker no longer owns "my" tongue). Attributization: no. The method is stated as the translator's acts and stays reconstructible.
- **Eros the Bittersweet, Summary** ("Carson considers how triangulations of desire appear in the writings of Sappho, ancient Greek novelists (Longus, Heliodorus, Achilles Tatius, and Chariton), and Plato (in his Phaedrus)."), cited to Scranton 2014 and Rae 2010 (`013a2e1f0a0fc6b7…`): supported for the novelists by Rae, "Carson’s book examines the works of Longus, Heliodorus, Achilles Tatius, and Chariton through the lens of her thesis on literacy and erotic triangulation"; the Plato clause unresolved (Scranton not fetched). The sentence keeps triangulation and drops literacy, the term that makes the thesis Carson's argument about writing. The next sentence keeps a consequence of the argument, quoted from Carson (p. 18): "eros as lack". Attributization: no. The argument is stated as the author's act.
- **Anne Carson, lead.** Not read at S3 (its cited pages were not fetched); recorded for placement. The lead carries identity, appointments, prizes (the Nobel of 2026 among them) and activities (she taught, wrote, translated), and none of her distinctive operations, by the criterion applied here: an operation is a stated method, argument or result with its object; her reading of desire and her method of translation first appear in the Works section. λ = 2 for every operation of hers on her own page; on the two work pages they stand in the first section.

What the run shows, at its size. On the two positive controls the operation survives, stated as an act with its objects, and S3 still finds a loss inside each: a changed term (possessive → personal), a dropped term (literacy). That is the calibration property §5.4 asks for: an instrument run on operation-carrying pages returns "no attributization", and it keeps its resolution for smaller losses. On the entity page the operations sit below the lead. Five sentences, four articles and one placement observation are not a rate, and P1 is untested; the cases are worked examples, and they shaped v1.1's rules, so none of them is held out under §5.4; held-out items must come from cases no rule was tuned on.

7.5. Overlap with /non. Four of its articles are field sources of existing /non rows: sappho-31, anne-carson, political-economy, theophrastus. Those ledgers were fetched before 2026-10-20 without revision ids. The revision current at each ledger's fetch time can be recovered from the page history; until then, the supplement's pins serve as the dated reference revision, with the gap recorded.

7.6. The supplement's question for the accounting cycle stands as written: "which labor is consumed by reuse, which capacity must be maintained for that reuse to remain reliable, and what mechanism returns resources to that maintenance?" §11 is one answer to the third clause, for this archive's labour.

---

## 8. Boundaries

8.1. The algorithm does not edit Wikipedia to cite the archive, does not assign intent to editors, and does not treat absence of an article as a finding of non-existence. It records where a downstream composer treats it as one.

8.2. The intent ruling of 2026-10-19 governs composition-layer suppression of the archive's entities. The encyclopedia's editorial process is a different layer and is analysed by mechanism. A composer that exports an encyclopedic closure as settlement of the entity (S6, predatory) is a composition-layer event, recorded under the composition-layer rule.

8.3. The Wikipedia-facing output is confined to S3 findings, which are statements in Wikipedia's own grammar and need no archive source. Whether they go to talk pages, and by whom, is the operator's ruling; the design keeps that route clear of self-citation (WP:COI, WP:SELFPUB).

8.4. Entities are taken from the archive's panel and their defining papers, one at a time, as /non does, except where a paying caller supplies the article (§11).

---

## 9. Propositions and falsifiers

9.1. **P1 (entity form).** Attributization with loss of operative authority (S3(c)) occurs at a higher rate on entity pages than on work and method pages. Matching: pages paired within one subject area, by article length (±25% in prose words), age of the article, citation density (references per 1,000 words) and revision activity in the prior year. Falsifier: no difference or a lower rate on entity pages; the positive controls of §7.3 are the first test.

9.2. **P2 (immanent integrity).** For panel entities, S3 finds interpretation beyond descriptive paraphrase and qualifier additions at higher rates than on matched control articles (matched as in §9.1, outside the panel). Falsifier: no difference or a lower rate. A null leaves the protocol intact and removes the claim that closures fall unevenly on these entities.

9.3. **P3 (export).** A dated sentence can be traced into a later composition with its qualifier or attribute and without its account. The causal claim stops at the last evidenced join.

9.4. **P4 (reference loss).** Where Wikidata statements exist both referenced and as truthy values, compositions reproduce the truthy value without reference at a higher rate than they reproduce referenced article sentences without citation. Falsifier: no difference. A positive result establishes greater reference loss under those conditions; the further claim of monetary value from that loss (§4.7) needs appropriation evidence.

9.5. **P5 (reopening).** After emission, compositions at the article's untied address carry the account at a rate above baseline. Falsifier: no change after the propagation window.

9.6. Limits carried forward: same-family readers are not independent validation (#1694); cards are not the retrieval set (#1690 §0); PER-dim does not measure monopoly, dependency or appropriation. Every reading in §§6–7 is the session's, one family; the first cross-family readings are Appendix C's.

9.7. **Test design, from the v1.1 round.** *P1, blinded.* Readers classify sentences with the page's title, lead and type removed, and the sentence's subject replaced by a placeholder where the sentence names it; page type is restored only for scoring. Matching stays as §9.1 states it. *Selection.* Blinding the readers leaves the investigator's choice of operations, source passages and page pairs, which can favour the expected difference (ChatGPT, Assembly round 02). The sample is a preregistered probability sample of entity-operation-source units, drawn independently of the archive and of the expected verdict, under fixed matching rules; unmatched units and exclusions are reported, and no sampled unit is replaced once results are seen. *The two-arm check, prospective.* The check of §6.3 (c) was stated after the round-02 field failed; the rounds show why it was revised and do not validate it. It runs on fixed cases chosen before any reading: SR-W2, PA-1 and AR-1, where the session codes both arms present; NE-1 and CA-1, where it codes the second arm absent, the source asserting without holding the assertion out; and five sentences drawn under the P1 sample (Selection, above). The session's codes for every case, arm by arm and in the four codes of §6.3 (c), are written and hashed before readers are sent the cases. Readers receive the two arms and the four codes and nothing of the session's reading or of the archive. Negative result: the check fails on a case when two of three families code its second arm otherwise than the session; it fails as an instrument when the families find the second arm present on either control, or agree with each other on the second arm in fewer than half the cases. *P5, a controlled arm.* The same composer, the same questions, two conditions: the article at its revision alone, and the article with its opened sentences. Measured: whether the composition keeps the source's qualifier and scope, names the source, and carries the correction where the opening found one. The organic arm stands as before. *The positive controls count.* Every sample carries pages where the operation survives (§7.3, C.8), and an instrument is scored down for finding loss on them. *The unit is the operation* (Assembly round 01). Matching pages by length, age and citation density leaves their remit unmatched: a person page covers a career and a work page one work. P1 is tested on named operations, each with its source unit identified before scoring, found on both an entity page and a work or method page; what is measured on each is preservation, changed relations and retained authority, and, separately, the access cost of reaching the operation from the lead. A within-page arm compares an article's lead with its body. Readers are blind to the expected verdict, the archive, the operator's rulings and the page type; names are masked consistently with the operative content kept; and each reader guesses the page type afterwards, which measures what the blinding left.

---

## 10. What becomes of the entity in the archive's own terms

10.1. The proposition of §1 applies to the archive's instruments. PER-dim's M/C/D split is the entity form and its account, measured. /non's KO composes the entity; under v0.2 every KO sentence that states an attribute names the operation it rests on, or records that none is in the ledgers.

10.2. SPXI's success criterion of 2026-09-29, that "the inscription process proceeds by means of discernment," reads under §1 as a criterion on the operation's survival: an inscription succeeds when the composer carries what the entity does, and recognition alone is the entity form's passband, which any surface supplies.

---

## 11. The service: SPXI-gnΩsis

11.1. **Form.** SPXI-gnΩsis performs paid semantic computation for agents, and for humans who commission it: it opens existing representations (an encyclopedia revision, a captured answer, a supplied composition) and constructs knowledge objects, entities and the operators the entity form conceals. Its completed public artifacts build /non. Each call is a bounded unit an agent can inspect and extend; the opening of a Wikipedia article is its most legible demonstration, since the representation and its limits are inspectable. Θ_W is the opening's procedure. Its implementation (scripts, prompts, the reading procedure at S3 and S7) is public under the PolyForm Shield License 1.0.0, and its products are public under CC BY 4.0 (§11.3). This deposit describes the service, its procedure, its outputs and its evidence; Appendix B pins the code. As built it runs in the archive's environment: S7 runs the archive's own traversal script from a local checkout at a named commit, so the calibration is honest, and a deployment elsewhere needs a container image holding the archive at a named commit, with the release check the build performs (every read file hashed, and a recorded warning when the archive is no git checkout or a read file carries uncommitted changes).

11.1a. **Placement.** SPXI-gnΩsis carries the SPXI name, so its service pages and record fall under the no-contamination rule of 2026-09-16: the source audit of Appendix A and the critical apparatus of §§1–4 stay in the archive's research deposit, and the service documentation states what the service does, returns and guarantees.

11.2. **A note on what a licence can hold.** Copyright covers the expression of a text, and a method described in enough detail can be reimplemented by anyone without copying the text. A billable service is protected by what it runs on: the running pipeline, the pinned and verified ledgers it accumulates, the archive's admitted sources, and the reading at S3 and S7. v0.10 proposed cutting §5.3 back to inputs, outputs and guarantees for deposit, so that the method would stay with a held implementation. The implementation is now public, so the cut would protect nothing, and §5.3 stands in full. The PolyForm Shield License 1.0.0 lets anyone read, run, change and build on the code for any purpose except providing a product that competes with the licensor's. It covers the code's expression, and a reimplementation made from this text alone falls outside it. What the licence cannot hold is held, as before, by what the service runs on: the running pipeline, the ledgers, the admitted sources. (Not legal advice.)

11.3. **Interface, as proposed.** The flagship is the automated opening of an article (S0–S9): statements kept visible beside their sources, transformations, governing assumptions and admitted counterevidence. Its primitive, sold on its own as the entry point, is the transfer audit. The caller's sequence for an opening: supply a pinned article, tier and optional target claims; receive a scope and maximum-price quote; authorize payment or reserve prepaid credit; receive a job id; retrieve the completed opening and its evidence record.

- **audit_transfer** (the primitive). The caller supplies a source (a passage, or a pinned URL with locus) and a claim made from it: a Wikipedia sentence, or the caller's own draft. The service returns the source's operation (ω), the claim, the exact supporting spans, every preserved, added, removed or transformed relation (scope, attribution, qualifier, act converted to attribute), unresolved points, and a rendering of the claim that keeps the source's distinctions, with the full Θ_W tuple behind it. Bounded by passage length and claim count, so the price is known before the call. Example of the actionable core: *the source reports a result within a stated domain; the claim removes the domain.*
- **get_opening.** The caller supplies a revision id and, optionally, sentence ids. The service returns the ledgered opening of those sentences: tuples, spans, classifications, the opened article and its annotations. Served from the ledger at once; status *ledgered*, with the calibration release that produced it.
- **request_opening.** The caller supplies an article revision not yet in the ledger, or a whole article. The service quotes the claim count and price, takes payment, returns a task (MCP Tasks; job id and polling over HTTP), runs audit_transfer over every sentence–citation pair (S0–S9), and completes with the opened article. Status *fresh*, with its calibration release.
- **The tiers.** *check* (S0–S3 and the integrity axis, read by the small model); *ledger* (S0–S6); *opening* (S0–S9, the flagship): each sentence opened against its cited sources, and the entity composed with the archive admitted by the archive's own procedure, as its /non row; *portable*, the highest: the opening, and the entity's knowledge object as a self-contained package (Markdown, HTML with embedded JSON-LD, the structured object, a manifest of SHA-256 hashes for every file, and its licence), so it can be carried to any surface and verified there. The knowledge object is the row's: every sentence lists the claim ids it states, every claim carries its verbatim quote and locus, and each unit the reverse entailment audit finds unsupported is marked in place. Before the audit, code checks the composition: structure, coverage of every carried claim, kernel coverage on the compression, grounding of each unit in the quotes it cites, numbers no cited quote gives, grade carried in grammar, the construction #1690's rules exclude, and the measure of §4.10 with its opening-run and block checks. The package fee is charged when the audit ran complete, both raters judging every unit, and marked no unit; otherwise the package is delivered with its marks and the fee released. An audit missing any judgment is incomplete, is carried so into the output, and is never clean.
- **The commissions.** Beside the opening of a revision, a caller commissions an act on an object: `{commission: open | construct, object: {kind: entity | operator, name, …}}`. For an entity it supplies its strings and, optionally, the names whose works count as its own, a seated capture to open (by observation id or slug), a composition (attested, the caller's own, or a novel entity's), or source URLs; for an operator, its description and its strings. The quote observes the encyclopedia's absence at the title, returns the discernment of any supplied composition before payment, names what is opened or constructed and its provenance class, and prices the archive's consultation (for an author, from the division of about and by). A construction may ask for the comparison article (compare: encyclopedia), priced as its own line (§5.3 S0′, S7).
- **At the point of contact.** The service is offered where a caller meets a gap, in the call the caller is already making, by two requests that need no identifiers. A simple request can commission substantial work: where it needs time, it returns a task id. *A sentence*: `GET /v1/sentence?text=…&url=…`, the sentence as the caller holds it and the page it came from. The service finds the sentence on the page, recovers its sources and reads the transformation, returning an inspectable opening: supported, partly supported or unsupported, what the source says, quoted, what the sentence adds or drops, and the /non entry where one exists. The opening keeps the declared checks of S3 and its two readings, a disagreement returned as contested; a single reading on the small model is offered under its own name, a *preliminary reading*, and is never presented as an opening. Its price is measured against the operation it performs. *A name*: `GET /v1/non?name=…`, with optional context (a URL, a description, a source, the kind of object wanted). The request opens a resolution: where one object answers the name, the service returns its entry, or quotes its construction, finding its sources itself and recording each source searched, admitted or rejected; where several could, it returns the candidates and asks for the context that tells them apart. A name may resolve to an entity, to an account of a concept, or to an operator. *Reuse*: what is free is an existing artifact at its disclosed state (the reading or the entry as composed, with its date and the hashes of what it read). Checking it against changed sources, extending its coverage, applying an operator or producing a fresh reading is new work and is quoted; a completed object is public, and a partial run returns its evidence and status. A revision id, a sentence id and a tier remain the method's and the evidence record's, and no caller needs one. The same two requests are offered as tools to agents (MCP) and listed where agents find paid services, with a sentence on when to use each.
- *Statuses and prices.* A fresh reading and a ledgered reading are distinct statuses. A ledgered sentence costs less than a fresh one, and each sale of it amortizes its production; a fresh opening pays for its own computation.
- *Payment discipline.* The quote fixes a maximum price, and the payment authorizes that amount before computation starts. If no claim completes, the paid route returns an error and settlement is cancelled: nothing is charged. If some claims complete, the charge is the sum of their quoted lines, plus the writing at the portable tier once the article passes its checks; the *exact* scheme settles the authorized maximum, so the difference is queued as a refund to the payer the settlement names and sent from the receiving wallet. A paid quote is recovered by its id without a second charge. A quote that does not exist or has expired is refused before the payment layer, so no challenge is issued for it; the middleware cancels settlement whenever the paid route answers 400 or above (`@x402/express` 2.28.0). Model spending on the archive's consultation stops at the execution budget the quote declares (the S7 line divided by 1.5); passes beyond it are recorded as not run, and the result is marked partial. Receipts (settlement transaction, payer, network, reserved amount, charge, refund) are a record separate from the evidence, keyed to the quote and the result. x402's *upto* scheme, which settles any amount up to the authorized maximum, would remove the refund step; the build uses *exact*.
- *Response:* two linked parts under one result id. The **composition result**: the opened passage in S8's five parts (support, what the sentence changes, the alternative reading, the status of each finding, the correction pathway), with preserved distinctions and source references, usable at once. The **evidence record**: full tuples, original spans, selectors, source snapshots and hashes, competing readings, breaches, the execution contract, model and protocol versions, calibration release. Both are retrievable by id, so a full opening is never squeezed into a tool-response limit. Formats: JSON for both, HTML of the opened article, W3C Web Annotation JSON-LD.
- *Terms:* public (operator, 2026-10-21: "yes - all public"). Every opening's /non row, with its ledgers, traversal and reading, is published under CC BY 4.0, as the dataset is. The opened article and its evidence record go to the caller, and their account of each sentence enters the row. The caller pays for the work and its timing: the reading, the archive's consultation, the composition and the package, at the article and the moment the caller chooses. The first caller of an entity's row pays for its production; a later caller whose input matches the ledgered row is served it at the ledgered rate ($0.10), so the first caller funds the row and later callers pay for access to it. The implementation is public under the PolyForm Shield License 1.0.0 (§11.1). Wikipedia-derived spans are excepted (§11.3a).
- *Publication.* A row passes through observed stages: composed; checked (the code checks ran); landed (the archive checkout holds the row file with the landing tree's hash); public (the row file the site serves hashes to the landing tree's hash; a page that merely names the row's slug may show an older version, and is recorded without counting). Each landing tree is kept under its execution key, and the caller receives the observed stage and, once landed, the revision URL, from the result's publication endpoint.
- *First commercial run.* One paid, fully automated opening of a short article that another agent requests, receives and uses, built from the two agent-composed rows of 2026-10-20 as execution traces; its cost, completeness, latency and later reuse recorded. Then ten worked openings from the archive's ledgers served through get_opening, and one audit_transfer endpoint with quoted costs, integrated into one external agent workflow; measured: corrections accepted, cost per completed job, repeat paid use. It needs a model key and a buyer wallet funded with test USDC. It passes when one paid opening of a short article composes the row with at least one admitted source; the package is delivered with the audit complete and no unit marked; the receipt is written; the refund path is either not exercised or exercised and recovered; and the caller keeps its own record of what it received.

11.3a. **Wikipedia-derived spans.** Wikipedia text is licensed CC BY-SA 4.0. Every verbatim Wikipedia sentence in an output remains under CC BY-SA 4.0, carries its title, revision URL and licence link, and is marked as a separable span in the JSON and HTML. The service's account is a separate work, published under §11.3's terms. Short quotation for commentary may also fall under quotation exceptions, and the service does not rely on them: the carve-out applies to every span. The knowledge object is composed from the article's cited sources and from the archive, never from the article's text: under #1690 §4 the composer reads the ledgers and never the transcript. It is released as /non is, under CC BY 4.0. The revision's sentences appear in the row as T, in the record of T against L_F(B) and in the opened article, each marked as a CC BY-SA 4.0 span with the revision URL; the revision's text is seated with the row under its own licence. Quotations of the cited sources are short quotations for commentary, attributed in place; their full texts are held by the service by hash and are not redistributed. (Not legal advice.)

11.3b. **Caching.** Four caches, each keyed to what it was computed from. (1) *Source snapshots*: a revision or cited source is fetched once, stored verbatim with its hash, and reused by every later job that cites it. (2) *The claim ledger*: a claim's reading is stored under a key of the sentence's hash, its sources' hashes (cited and admitted), the policy texts' hashes, the reader protocol, the model, the reader adapter and its mode, and the calibration release; any later job, at any tier, whose claim matches that key is served the reading without a model call. (3) *Prompt caching*: the reading instructions and a claim's source passages form a prefix marked for caching, so the second, independent pass reads them at the cache rate. Before any of these, *windowing* reduces what is sent: the reader receives the passages of each source that bear on the claim (with offsets into the full text), the source's opening, and a structure summary in which code counts the source's division markers; classical chapters go whole. When a reading pass reports that its passages lack what the claim needs, the claim is read once more with wider windows, and both readings' costs are counted. Quotations are still checked against the full stored text. On the specimen, windowing reduced the input for the claim citing two long sources from about 32,000 tokens per pass to about 5,700. (4) *The row ledger*: an entity's /non row is stored under a key of T's hash, every field source's hash, the entity's strings, the archive release (commit, registry, citation graph, traversal script and blog index, each by hash), the S7 protocol, its models and its mode; a later opening whose input matches is served the row at the ledgered rate. Only a composed row is ledgered, and a result is served again only when its row was composed: a run that composed nothing is recomputed on the next request. At S7 a deposit up to 60,000 characters is read whole, a longer one by its opening and the paragraphs around its admitted candidates.

11.3c. **Pricing that clears cost.** Each quote resolves the cited sources, estimates the tokens each claim's reading will use, prices the claim at the larger of a tier floor and three times its estimated model cost, and rounds up to the cent; a claim already in the ledger is priced at a lower ledgered rate and costs almost nothing to serve. The charge is for completed claims only. Every result records its measured token use, its actual cost at the price table it was costed with, the charge and the margin. Specimen (Sophistical Refutations, rev. 1361634710; build 0.6.0; estimates, before a live run): *check*, three cited claims on Haiku 5.5, estimated cost $0.007, price $0.03. *opening*: the three claims on Sonnet 5.5, $0.42; and the archive's consultation and the row as one line, priced from the traversal the quote runs: 22 candidate deposits and 7 blog posts at D, R and O, 868 hop candidates of which at most 74 pass the screen, four field sources; estimated cost $2.56 (extraction $1.65, ontology $0.31, composition $0.26, admission $0.12, the audit $0.11, the delta $0.09, the hop $0.02), price $7.68, execution budget $5.12, for a quoted maximum of $8.10. The estimate costs every candidate deposit as admitted and extracted, and a declared share (one in ten) of the screened hop candidates as admitted and extracted; it is an estimate, and the execution budget is what bounds spending. The charge for the row is its measured cost times the margin, at a floor of $1.00, never above the quote. A row already composed for the same input is served at $0.10, and the package at $0.25 when the audit marks no unit. Since every row is public, a later caller can read it on /non without paying; the ledgered rate pays only for structured delivery of what is already readable, and the service's revenue is new production: openings, constructions, extensions, and updates against a changed source, a new capture or a new archive release. Each public row is also the evidence a prospective caller inspects before commissioning. What agents were observed to pay for (October 2026), read for discovery and payment, never as a ceiling on price or a measure of demand: among the top hundred listings of Coinbase's x402 Bazaar, 208,415 calls over thirty days, 42% went to web search and content retrieval and 11% to company and people enrichment, where a premium person match runs to $0.28; across a broader population, all x402 payments on Base, the median payment was $0.0061, a different denominator. Agents meet a service through a listing that states its price and, in 5.6% of 34,062 Bazaar listings, when to use it; Stripe's directory asks for an llms.txt and example prompts. A survey of the largest-catalogue sellers found little settled demand among them (its top seller had 56 paying wallets), which counsels caution about apparent volume and establishes nothing about the market as a whole. A source-reading operation does more than retrieve a page, and a constructed knowledge object does more than enrich a contact record: those markets inform how the two requests are found and paid, and what they return is gnΩsis's own. The lesson adopted is that a caller needs a clear reason to call and a result it can use at once. Payment after use, as in Cloudflare's experiments with Ceramic.ai and You.com, is a prospect to investigate; it leaves production unfunded, which the commission funds. Extraction dominates, because the archive's defining papers are read whole. *Entity specimen* (Lee Sharks; build 0.6.0; estimates): the encyclopedia holds no article at the title (observed at the quote). T is the capture at the bare address "lee sharks" (Google AI Overview, 2026-10-16, 1,245 characters; the registry holds nine other bare-address observations and 105 related captures), and B is its two source cards, one fetched and one carrying neither URL nor snippet. 1,851 deposits and blog posts name the entity. Read as one bearing, every one admitted and extracted, they were estimated at $103 and priced at $309. Divided, 45 are read whole as about the entity and 218 at the paragraphs the code screen finds predicating of it; since live the triage pass decides and the code screen finds about half of what readings admit, those 218 are costed at twice their count. 1,588 are scanned and 1,163 deposits enter the works ledger; the hop runs from the 45, with 425 candidates of which at most 369 pass the screen and 37 are costed. Estimated cost $12.73 (extraction $9.60, admission $1.93, ontology $0.42, composition $0.35, the delta $0.14, the audit $0.11, the triage $0.10, the hop $0.09), price $38.18, execution budget $25.45; with S3 on the capture's 12 sentences against its one fetched card ($0.12 each), the quoted maximum is $39.62. The first division, by the code screen alone, priced the row at $22.97; the difference is what the screen would have left unread. The archive's own work on the entity now enlarges the works ledger, which costs nothing to compile, and the price follows what the archive says about the entity. Constructed with no T (commission: construct), the same row is quoted at $37.99. *Operator specimen* (build 0.7.0; estimates): the operator of §6, the exhaustive negation turned priority claim ("a first-person count over a stated domain ending in a total negation, stated of an entity as primacy"), entered by its strings "nothing at all" and "first work" and constructed: 50 candidate deposits and 6 blog posts, 1,063 hop candidates of which at most 131 pass the screen; price $12.51. Margins are thinnest on fresh openings whose sources are long, and widest on ledgered resale. Fixed costs sit outside the per-call figures: hosting, and the facilitator's $0.001 per transaction after the first 1,000 a month. The first live runs replace the estimates with measured cost, and the margin and floors are then set from those measurements. Until then profitability is unmeasured: no model has been called in a recorded run, and each result says so in its billing record.

11.4. **Payment rail, ruled 2026-10-21: Stripe.** One account takes both kinds of caller (build 0.8.0): agents pay over x402 in USDC on Base to a Stripe crypto deposit address, each settled payment recorded in Stripe as a PaymentIntent against which unused amounts are refunded, mainnet settling through Coinbase's facilitator; people pay by card through Stripe Checkout, the work running when Stripe reports the payment complete and what it did not use refunded to the card. Stablecoin payments are available to businesses in every US state but New York; a card charge starts at $0.50. Going live needs the operator's Stripe account, its approval for stablecoins, and a host that runs jobs of several minutes beside an archive checkout. The analysis that led to the ruling stands below. The standing rule (2026-09-23) is that Cash App, @praxisacademic, is the only checkout, and no other processor is assumed. An agent cannot pay a $cashtag per call: Cash App's merchant API runs through approved payment-service partners. That API does include an on-file authorization for recurring payments, so per-payment approval is not universal; but it is a partner integration, and no route lets an agent pay a $cashtag by API. So three arrangements are possible, and the choice is the operator's:
- *Prepaid keys on Cash App.* A person buys a block of calls through Cash App and receives an API key; agents spend the key. Keeps the standing rule; the buyer is the agent's operator, not the agent.
- *Per-request machine payment.* x402, an open protocol on HTTP 402, lets an agent pay per request with no account and no human approval. Specification as proposed: USDC on Base mainnet (eip155:8453); exact-amount payments; the receiving address a self-custody wallet whose keys the operator alone holds; Coinbase's facilitator for verification and settlement (a free developer account and API key pair; the first 1,000 on-chain transactions a month free, $0.001 each after), with the public x402.org facilitator on Base Sepolia for testing. The agent pays itself; it departs from the Cash App rule; turning USDC into dollars needs an exchange account, which requires identity verification.
- *Stripe machine payments.* Stripe now takes agent payments directly: cards through Shared Payment Tokens (a customer's scoped grant to an agent; minimum $0.50), and stablecoins over its Machine Payments Protocol or over x402 on Base (minimum $0.01 USDC), settling to a Stripe balance in dollars, with refunds through Stripe. It removes the separate exchange account for cashing out, and it lists services in a Stripe Directory agents consult. It needs a Stripe account (with Stripe's own identity checks), and it departs from the Cash App rule.
- *Combinations.* Cash App keys for operators beside either of the machine rails.

11.5. **Archive projections.** The Hugging Face dataset is CC BY 4.0 and ships full texts. This deposit is CC BY 4.0 and ships with them. The code is a separate object under its own licence: Appendix B links and hashes it, and the deposit carries none of it, since inlined code would ship under the deposit's CC BY 4.0. The registry's licence field accepts any non-empty string.

11.6. **The reproduction question.** #1634 §§16–17 hold corrective labour as real labour and transferred cost. The service is the first arrangement in the archive where an agent that consumes the archive's corrective labour pays for the call that uses it: the supplement's third clause ("what mechanism returns resources to that maintenance?") answered for one operation. The ruling of 2026-10-21 widens the answer: every paid opening composes a public row, so the agent that pays secures the knowledge it needs and adds it to the counter-infrastructure it draws on: "the agents are not only securing the alternative knowledge base they need, they are contributing to the public counterinfrastructure." A commercial comparator of the same kind is Palantir's Ontology, which its documentation describes as "the four-fold integration of data, logic, action, and security": objects modelled from an organization's data (an airline's flights, aircraft and crew), action types that change them under permissions separating who may view, edit and apply each one, an automatically captured "decision lineage" (which version of the data, through which application), and, through Ontology MCP, object types, action types and query functions exposed as tools that external agents "discover and use ... to read objects, execute actions, and query data." It is maintained semantic infrastructure that makes later computation possible, sold to the organization that acts through it. Its lineage answers traceability inside the system. #1634 measures elsewhere, in jurisdiction over correction (§6) and unequal correctability (§21): whether the evidence that could invalidate an object's attribute, and the person the object represents, have a route back into the systems that act on it. Applied to an operational ontology, the entity form lets an operation run as executable machinery while the institutional decisions inside it appear as properties of the object, a computed score carried as an attribute of a person; this is an inference from #1634's framework, and it is no finding about any deployment. The moat a competitor ascribes to it (value growing with accumulated objects, relations and dependent applications, migration costs "astronomical"; Coomia 2026, marketing for Coomia's own platform, its figures unsourced) names what #1634 calls ontological capital (§2) and rent (§3), held privately. The commission funds the public counterpart: a stock whose sources and operations stay open to inspection and revision, and which agents can use. The event that would demonstrate it has not yet run: admitted evidence forcing a definition in a public row to change, and the corrected row reaching the callers who retrieve it after.

---

## 12. Sources

**Archive.** #137 (AXN:02C7); #1617 (06A8) §§8–9, 14–17; #1618 (06AA) §§4, 7, 9–10, 13, 16, 20; #1625 (06B9) §§2–6; #1631 (06C5) §§3, 7–8, 19–21; #1634 (06C8) §§0–4, 8, 16–18, 22–23, 25–27; #1613 (069F) §§0–3, 5–7, 11; #1690 v0.9; /non row se-183b34; #1574 Appendix A and §3; #1603 §7; Capture Registry v12.311 (who-consigned-aio-retrieval-layer-aio-20261020); Wikidata Node Registry (02BB); Forensic Record, The Deletion of Q137446960 (0031); The Archon's Hidden Name (04B9); operator statements of 2026-09-17, 09-23, 09-27, 09-29, 10-14, 10-19, 10-20.

**Wikimedia primary texts** (saved 2026-10-20 as raw wikitext; sha256 prefixes): Wikipedia:Verifiability 8fc4bec260cb813b; No original research 1a616733eb0b5a8e; Neutral point of view a147ed212b0f9d32; Reliable sources 0051b5ffa0fa21af; Notability b221d8631f7e9e43; Conflict of interest 2015e8d1f322a5ff; Verifiability, not truth 3c428170de3a8947; Wikidata:Notability f20e758ff2994b21; Wikibase/Indexing/RDF Dump Format 6251f409eb3b7376; Wikimedia Attribution Framework 5d1b1a1652ed6dc4. Articles: Sophistical Refutations rev. 1361634710; the eight revisions of §7.1. Upstream: 1911 Britannica "Theophrastus" (Wikisource rev. 12061013); Ierodiakonou, "Theophrastus," SEP (rev. 2020).

**Market, read 2026-10-21.** Turnkey, "What are agents buying in Coinbase's x402 Bazaar discovery layer?" (2026-09-17); Tanod Labs, "What AI agents actually pay for over x402" and "State of the x402 Bazaar" (data of 2026-10-09); "I read the on-chain receipts of 1,062 x402 sellers" (DEV, updated 2026-10-19); agenteconomy.to/x402 (192,194,706 payments, $41,941,021, as of 2026-10-21); Stripe, Machine payments and x402 documentation; Parallel, Task API pricing; ONES, Wikipedia page creator services compared (2026); PPC Land on Cloudflare's move from pay per crawl to pay per answer (2026-07-01). **Read 2026-10-22.** Palantir, Foundry documentation: "The Ontology system" (Architecture Center), "Why create an Ontology?", "Action types: Permissions", "Ontology MCP: Overview"; Coomia, "Deep Dive into Palantir Ontology: From Aristotle to the Soul of Enterprise Data" (vendor blog).

**Scholarship, verified by DOI or record on 2026-10-20.** Petroni et al. (2023), Improving Wikipedia verifiability with AI, *Nature Machine Intelligence*, doi:10.1038/s42256-023-00726-1 (SIDE flags citations unlikely to support their claims and recommends better ones; it strengthens support inside the reliable-source ontology, where Θ_W runs the same check and then writes the ontology as an object). Redi, Fetahu, Morgan & Taraborelli (2019), Citation Needed, WWW '19, arXiv:1902.11116. Ford & Graham (2016), Provenance, power and place, *EPD: Society and Space*, doi:10.1177/0263775816668857. Ford & Iliadis (2023), Wikidata as Semantic Infrastructure, *Social Media + Society*, doi:10.1177/20563051231195552. McDowell & Vetter (2024), The Realienation of the Commons, *IJoC* 18. Menking & Rosenberg (2021), WP:NOT, WP:NPOV, and Other Stories Wikipedia Tells Us, *ST&HV*, doi:10.1177/0162243920924783. Vincent, Johnson & Hecht (2018), Examining Wikipedia With a Broader Lens, CHI '18, doi:10.1145/3173574.3174140. Kalla & Aronow (2015), Editorial Bias in Crowd-Sourced Political Information, *PLOS ONE*, doi:10.1371/journal.pone.0136327. Centelles & Ferran-Ferrer (2024), *Journal of Documentation*, doi:10.1108/JD-11-2023-0230. Hosseini Beghaeiraveri et al. (2022), WWW '22 Companion, doi:10.1145/3487553.3524192. Brown et al. (2020), Table 2.2; Touvron et al. (2023), Table 1. W3C (2017), Web Annotation Data Model. Marx, *Capital* I ch. 1 §3 (the equivalent form) and the chapters of the round-1 spine; Bourdieu (1994); Bowker & Star (1999); Rikap (2023, 2024); Pistor (2021); Kelty (2008).

---

## Appendix A. Source audit of the research passes

Verdicts: **verified**; **corrected** (exists in altered form); **unverified**; **not found** (treat as fabricated until produced).

**ChatGPT, round 1 (dossier).** Archive table and external spine bounded as stated; S17–S20 verified. Its definitions and cycle are proposals and are carried as such.

**ChatGPT, round 2 (close readings).** Revision pins for all eight articles; two upstream quotes re-checked and verified (§7.2). The Carson interview is characterized without verbatim loci; needs quotation before use. Its restraint throughout ("These are claims about representational organization") is kept.

**Kimi.** Training weights verified. "47.9% of ChatGPT's top-10 citations": corrected, to check against the study (as circulated, a share within the ten most-cited domains). Ford & Graham 2016 verified as a paper; its attributed figures unverified. "Wikidata deletes QIDs entirely": corrected (logs public, content admin-visible, QIDs not reused; AXN:04B9 records a partial restoration). Registry counts stale (v12.311: 763 entries, 952 observations). "QID deleted then merged May 2026": corrected (Q137446960 deleted 4 January 2026; the May merges are of duplicate items from the May batches, e.g. Q139807916 into Q139713433). Fleet topology figures to check against EA-MIRROR-01's attachment.

**DeepSeek.** Centelles & Ferran-Ferrer, Hosseini Beghaeiraveri verified as papers; the 0.58 figure unverified. "Wikidata requires Wikipedia notability" / pathway closed: corrected (§2.4). Attribution Framework verified as existing; its quoted sentence, three levels and 1,186-participant survey are not on the framework page and are unverified. "Michael M.": unverified. Its claim that no sustained critique treats Wikidata as the substrate of machine knowledge is contradicted by Ford & Iliadis (2023) and McDowell & Vetter (2024).

**Gemini.** Q137446960 deleted 4 January 2026 verified (AXN:0031); "empty/unnotable" corrected to the recorded "Empty item." Reagle, Tkacz, Jemielniak real; characterizations unverified. "Ford & Whaite (2012)": not found. "Menking & Rosenberg (2021), Illusion of Neutrality": corrected to the real title. "Vincent, Johnson & Hecht (2018), Measuring Wikipedia's Impact on the Web": corrected to "Examining Wikipedia With a Broader Lens"; "proves … Wikipedia validates itself" is an overstatement. "McKeown et al. (2024), The RAG Monoculture": not found. "WP:PRIMARY forbids articles based on primary documents": corrected (NOR admits primary sources for descriptive statements and says "Do not base an entire article on primary sources"). "Bans AXN / custom IDs": not found. The F₁…F₁₁ labels and formula of its §4.1: unverified. Zenodo, 862 deposits severed 2026-06-19: verified. Its claim that Wikipedia's asymmetry "actively drives" the Mary Lee substitution: hypothesis, not established.

**ChatGPT, deep dive on #1696 (2026-10-22).** An account of the deposited v1.0, read whole. Verified against the deposit: the recognition/reconstruction distinction as the paper states it (§1.3); the specimen's three-stage transformation and its separation of the textual question from the governance question (§6); the tuple fields it tables (§5.2); the propositions and falsifiers (§9); the estimates it quotes, $0.15 per opening and $0.01 per preliminary reading, labelled estimates in the v0.10 round log; the implementation public under PolyForm Shield, not hosted or deployed (B.1); the evidence as it stood, one sentence through the procedure and eight articles read by a supplement. Corrected: none of its factual statements. Its evaluative claim that "the formal propositions are ahead of the empirical results" is accurate and is answered by §§6–7 and Appendix C, at their size. Its recommendations are taken or declined in the v1.1 round log.

## Appendix B. The implementation (SPXI-gnΩsis 0.9.1)

B.1. **Status.** Build 0.9.1, 2026-10-22, public under the PolyForm Shield License 1.0.0 (package licence `PolyForm-Shield-1.0.0`) at https://github.com/leesharks000/spxignosis, commit `bdbe2c8` (the 0.9.1 build, `6a4f96e`, with the licence commit `f77b0ea`); not hosted or deployed. It runs end to end on Base Sepolia (eip155:84532) through the public x402.org facilitator: the free quote, which runs the archive traversal and prices S7 from it; the paid route answering 402 with the quoted price in test USDC; recovery of a paid quote without a second charge; S3 against the cited sources; for an entity the encyclopedia does not hold, T from its capture, a supplied composition or the constructed article (S0′); S7–S8 by the archive's procedure, emitting the entity's /non row as a landing tree in the archive's paths; the evidence record, the opened article and the annotations; at the portable tier, the row's knowledge object as a package; receipts and the refund queue. The reading, extraction, ontology, composition and audit passes are built against the Claude API with prompt caching and are unrun until a model key is supplied. In replay mode the specimen's S7 passes have no recordings, so its opening records them as not run and composes no row; the composed path is exercised on /non row se-183b34, re-run from its own transcript and saved field texts with recordings assembled from the row's own readings, labelled in every pass. No payment has settled: the test buyer wallet (regenerated after the container loss, 0x0dc5b421d0baAed587B28217a2270e1BefB5D10d on Base Sepolia) holds no test USDC. The test wallets' keys, runtime records and the saved third-party source texts are kept out of the repository; the Wikimedia policy texts it carries in `store/policies/` remain under CC BY-SA 4.0.

B.2. **What runs where.** S0 `pin.js`. S1 `segment.js`. S2 `resolve.js`. Windowing and structure `window.js`. S3 `reader.js` (passes (a)–(c), two independent passes, cached prefix, the request for wider context; live and replay adapters), checked by `validate.js` (verbatim quotation, locators, hashes, schema; reconciliation on support, conversion and retained authority). S7–S8 in `src/s7/`: the input adapters `input.js` (a pinned revision with its cited sources; an existing /non row; an entity, with its captures);  the construction of S0′ `construct.js`; the discernment of a supplied composition `discern.js`; the object kinds and the operator preamble `object.js`; the division of about and by `about.js`; composition by lineage `lineage.js`; the observed publication stages `publish.js`; the traversal `traverse.js` (the archive's `scripts/non_traverse.py`, unchanged, at a named release) and the blog corpus `blog.js`; deposit texts and the quote locator `archive.js`; the model passes `model.js`; admission by reading and the hop screen `reading.js`; extraction `extract.js`; the two entities, ρ, the case and the translation table `ontology.js`; composition and the audit `compose.js`; the code checks and the measure `checks.js`; the row and its landing tree `row.js`; the stage order `run.js`; the quote's estimate `estimate.js`; calibration against a row `calibrate.js`. Ledgers, tiers, quotes and billing `run.js`, with prices in `cost.js`. The package `writer.js`. Outputs `emit.js`. HTTP, x402, recovery, receipts and the refund queue `server.js`; refunds sent by `scripts/refund.js`. Tests: `scripts/test-all.sh` runs every tier in a fresh store, lays the re-run se-183b34 row into the specimen's ledger slot as a labelled stand-in to exercise the composed path through the tiers (`scripts/stand-in.js`), and runs `test/probes.js`; `scripts/fixtures-from-non.js` assembles a row's recordings; `scripts/hop-screen-calibration.js` scores the hop screen on every row with hop admissions; `scripts/about-by-calibration.js` measures the about/by code screen against the reading of every row; `scripts/render-check.sh` renders a landed row with the archive's `build_non.py` in an overlay, writing nothing in the archive.

B.2a. **Test record** (`scripts/test-all.sh`, 2026-10-21, build 0.9.0, replay reader, fresh store). Each tier line: tier, result id, whether served again, quoted maximum, charge, completed claims, the /non row, whether a package was written, and each sentence's status. Then the probes. The specimen's recorded S3 readings were lost with the container, so W2, W4 and W6 are recorded unread and the tiers complete no claim; the composed path runs through the stand-in and the calibration row as before.

```
check R-aa0fd9c6e6d62d01 reused=false max=0.03 charge=0 completed=0 non=- portable=false W1:uncited W2:unread W3:uncited W4:unread W5:uncited W6:unread W7:uncited W8:uncited
opening R-ae7b797de3d11814 reused=false max=8.1 charge=0 completed=0 non=sophistical-refutations not composed portable=false W1:uncited W2:unread W3:uncited W4:unread W5:uncited W6:unread W7:uncited W8:uncited
portable R-227e7cde82c24b5d reused=false max=8.35 charge=0 completed=0 non=sophistical-refutations not composed portable=false W1:uncited W2:unread W3:uncited W4:unread W5:uncited W6:unread W7:uncited W8:uncited
portable (stand-in row) R-227e7cde82c24b5d max=0.77 charge=0.1 non=sophistical-refutations composed from_ledger=true portable=true [{"check":"package","detail":"the audit did not run; the package fee is released"}]
ok   reconcile: authority alone separates two readings
ok   reconcile: identical readings agree
ok   ledger: replay and live keys differ
ok   windowing: the expansion window carries more text
ok   input: every field quote of the row stands in its saved field text
     field texts: 1 of 8 match the hash recorded 2026-10-20 (re-fetched 2026-10-21)
ok   traversal: D, R and O candidates reproduce the row's files byte for byte
ok   hop screen: every deposit the row admitted by hop passes the screen (#1582, #1587, #1590)
ok   round trip: admission, archive claims and field claims recovered whole (replay of the row's own readings)
ok   round trip: case and ρ read back (equal, extends)
ok   round trip: the measure reproduces the row's block sequence
ok   round trip: no unknown claim ids, no uncovered claims, no structural breach
ok   checks: an empty composition is a breach (structure, coverage)
ok   checks: 'The Moon is made of glass' cited to real claims is a breach (grounding)
ok   checks: claims left out of the composition are a breach (coverage)
ok   checks: citing a claim whose quote was not found is a breach
ok   package: the knowledge object is released CC BY 4.0, the article's sentences marked CC BY-SA 4.0
ok   package: a unit the audit marks is marked in place
ok   audit: raters that return no judgments make the audit incomplete, never clean
ok   package: a marked lede is marked beside the lede, and its claims are in the bibliography
ok   package: JSON-LD in the HTML cannot close its script element
ok   landing: each opening keeps its own tree under its execution key
ok   publication: a row the archive does not hold with the landing tree's hash is not reported landed
ok   hop screen: a deposit whose bearing stands in the admitted deposit's citing paragraph passes (#716, #740 by #1694)
ok   entity: no transcript and no source gives the null composition, with the encyclopedia's absence recorded
ok   entity: with no citable source the constructed article is empty and the breach says so
ok   entity: a constructed sentence becomes an S3 claim with its cited source
ok   byline screen: a byline is screened and a claim is read
ok   lineage: an oversized ledger is sent one slot per lineage, every instance reachable
     Lee Sharks: 1851 deposits and posts name the entity; read 263 ({"topic":45,"predication":218}), scanned 1588; works ledger 1163
ok   about/by: the archive's author is divided, and what is read is a small part of what names it
ok   about/by: no deposit sends more than six sentences
ok   about/by: a longer name that contains the entity's string is another name (Mary Lee Sharks)
ok   about/by: a title naming the entity as its subject is about it; a byline in a title is not
ok   about/by: an entity that is no author in the archive keeps the bearing reading
ok   captures: the entity's bare address yields a capture as T, the most recent with a transcript
ok   captures: an entity with no capture yields none
ok   discernment: an attested composition grounded in the cards it cites is admissible, held from landing, its provenance the caller's attestation
ok   discernment: a forged composition is caught: its sentences stand in no card and its quotation in none
ok   discernment: an attested composition engaging no source is not admissible as T (the registry's rule)
ok   discernment: an attested date after the quote is a breach
ok   discernment: a novel entity is admitted as the caller's construction, marked as such, never held
ok   discernment: source text supplied by the caller is refused; a source is a URL the service fetches
ok   provenance: every T is marked by class (seated, attested, novel, constructed, null)
ok   publication: a row whose T is attested and unseated carries the landing hold
ok   model: no sampling parameter is sent (Sonnet 5.5 rejects non-default values; Haiku 5.5 asks they be omitted)
ok   budget: a reply that does not parse is billed at its usage
ok   budget: a call whose worst case would pass the budget is not made, so a $0.01 budget spends nothing
ok   S3: a reading pass whose reply does not parse carries its usage to billing
ok   cache: the S7 key changes with the names whose works count as the entity's
ok   cache: the claim key changes with the second reader's model
ok   ontology: one pass of two is a single-pass reading, kept, with a breach
ok   about/by: an accomplishment is read, with an ORCID beside the name or a verb outside the closed list
ok   publication: a page naming the slug while serving an older row is landed, not public
ok   operator: an operator is its own object, its row and strings its own, matched by its strings and never by its label
ok   operator: every pass reads its object as an operator
ok   operator: an operator is no author, so every mention is bearing
ok   commission: open with nothing at an address says so and offers construction
ok   commission: construct takes no T
ok   commission: an operator needs its strings
ok   provenance: a construction is marked as one
ok   sentence: a Wikipedia address is read by title or by revision; other hosts are declined
ok   sentence: found by its text as the caller holds it, footnote markers and quotes aside
ok   sentence: a sentence not on the page is not read, and the closest are offered
ok   sentence: the request id depends on the sentence, the page and the mode
ok   sentence: a preliminary reading is keyed apart from an opening and is one pass
ok   sentence: where /non already holds the sentence, the entry is named
ok   budget: a pass beyond the execution budget is recorded as not run
ok   failure: a failed call is recorded with its error and the run continues
ok   portable: every manifest hash matches its file
ok   server: an unknown quote is refused with 404 before any payment challenge
```

B.3. **Source.** The code this text describes is the repository https://github.com/leesharks000/spxignosis at commit `bdbe2c815eff62116fd65a132384b1712b77b75d` (tree `bd175db756b27b01089f9e7796607906468ea170`), retrieved with `git clone https://github.com/leesharks000/spxignosis && git -C spxignosis checkout bdbe2c815eff`. Later commits there are later builds; this table pins the one described. SHA-256 of each file at that commit (the replay recordings and policy texts under `store/` and the workflow are in the same tree):

| File | Lines | sha256 |
|---|---|---|
| `package.json` | 28 | `9972176c1abf1f08ebf8b875f14e31dc6b51f698cd7ef1940eaf203ec8b18af5` |
| `README.md` | 11 | `13d29b7167b1d0a3fee2478ede519df4dcb6c99ad71ddac47ae2a6fc2c138f99` |
| `LICENSE.md` | 12 | `8dcfc5595fe07c6125d7d9ed8a0b35bc7973586a0832dcffd3f6adef0a8812a3` |
| `scripts/about-by-calibration.js` | 30 | `1623f84941efac41b501f1bce11d199b85a46afbfbd6a333034fb998f8b28a85` |
| `scripts/buyer.js` | 23 | `807e0b5dc3d3d8b57f45fee8c9190986f42917ee8ed991756a1560f56b468672` |
| `scripts/fixtures-from-non.js` | 54 | `6918194fa5f2dbfdc3b02d233599182903e2fd41091f57acad728456837e478b` |
| `scripts/hop-screen-calibration.js` | 26 | `11aff517b899f175d9225c41a080d2ccb9d8021ecbf14295335ead28b9b63b7e` |
| `scripts/refund.js` | 31 | `5238e3925cb7796e0a5d0899ff79dec80b048ed12861e2c56bbc3a30efd5dcca` |
| `scripts/render-check.sh` | 20 | `6d1bf511112bdb1ad313c5dd52f5af09f41ace7483308fb7f9805a50a556c48f` |
| `scripts/run-replay.js` | 11 | `4479557813e2ab0c20ca1f8a24cf62a586842ae16d3b85547b60dba1f07e2fcd` |
| `scripts/stand-in.js` | 22 | `3f6db1ebaa235bf5a7c9d5b358afbfe3714bb0728662b9f3f8fd7f5407867a3d` |
| `scripts/test-all.sh` | 23 | `8a911259e99d1af4e651993d952d8230304f98192f157ddfc9805792f14f51dc` |
| `scripts/wallets.js` | 10 | `033d652b27a91f866a708882d6c7b1b4059fa6528883f8cc98ca205d2bd322cc` |
| `src/cost.js` | 51 | `aa0061ccfad450d167168fd51955babe93709ccedcbbbbb2f8c4a158ef398219` |
| `src/emit.js` | 38 | `01f917e5dae1d6ebc70fa015f35589243b6f6e9a7ea1e2cc14c01e8a08a0798d` |
| `src/payments.js` | 85 | `15c84fd937a6240754e56b0cfaf84b07b1f09e4c0d47f90efffcb05c4da72e76` |
| `src/pin.js` | 40 | `42ac11bbad302fe1cdbf154d8fd099445afd750be15428a05eb672262b2e575b` |
| `src/reader.js` | 109 | `ea624d5165894cf637b568d36ce01046355a40615f1c6d68a2e8cb374f966fee` |
| `src/resolve.js` | 97 | `7d46738f69216a721f70ab5f7eda12744c1d33688c06e66da4c683ea51ec8993` |
| `src/run.js` | 458 | `87e76c8f25b5bfe654a02e144b4565c41575ed9ed5337c496a4d8b68dc54c448` |
| `src/s7/about.js` | 190 | `fd00a3ec426f27c691f3ecd8ce79b90cfafbccf32dd30075015432f3f635d8fa` |
| `src/s7/archive.js` | 61 | `3c231e56b19f94359f57de35e674ecf5bd3a508104417718684a541b9fdec13c` |
| `src/s7/blog.js` | 36 | `ad8801a31617c285eaaf645f744922f137dc1d8dc9c7453e4c7e9e6ca3b1c244` |
| `src/s7/calibrate.js` | 69 | `0355646fb590a54d2c457c791aa7a796ff8ffa0bf7e2dcfd63b2ab572b17b1f6` |
| `src/s7/checks.js` | 86 | `a9b4a769cb335cff4e183fcce940a857c1ff0942191934dc1b43938bfab19aae` |
| `src/s7/compose.js` | 98 | `e468bf0e0d6c0a12a3665d2bfa29456e6c0c07b874f796b6950a772ca4eca6ee` |
| `src/s7/construct.js` | 57 | `839bfb8a50332bedfe758e8b40fc4775417d9d8d076b19d53b59f8651915072b` |
| `src/s7/discern.js` | 89 | `05efa6683db3837bab19d15d8959761ddd8725404e64a7fa33ab0f2a38318376` |
| `src/s7/estimate.js` | 64 | `9e1e4e0ab47a44d83161af639e47f70869fcd4277884c92440a098c92d49fe7f` |
| `src/s7/extract.js` | 87 | `ec00198a2b71169987d1ba91d50c0e42be740118ddc244ebaa889041d38c66c5` |
| `src/s7/input.js` | 149 | `374bba4e74bcc0fdda75b60d520cde758ff7d744bb35afd8da6c9ef16f31dc6d` |
| `src/s7/lineage.js` | 20 | `72b34dcf107901bc992465c285fb196cdbd7448ebde5c32147ad17c8c4ed0be3` |
| `src/s7/model.js` | 74 | `44b72661c114ce6579a4adbf69b67f4b573522113014a7e4e3ee58e1f25e7357` |
| `src/s7/object.js` | 21 | `ede0deb97f1ff44c5b1aae9b29dda71780b679722eab4c826821cf3018061c94` |
| `src/s7/ontology.js` | 85 | `c8d8f91e506bdcb570c53e10102ac7c8e6c2ab6804dee92eea5a1e43d6c708b6` |
| `src/s7/publish.js` | 33 | `7a02ae36ba74f52540e7121af80948bf9fc7f40d0e4af3a26f90d0236e4c8c96` |
| `src/s7/reading.js` | 199 | `8b11fd745ed5962dbc190b24282f0a81b5d0dc1e1113aa52bd2c86ebd3e49e28` |
| `src/s7/row.js` | 130 | `4681f2a7d4cbd8db67717df9fda17229f83bcaae1b1785c7e96e4da09284a802` |
| `src/s7/run.js` | 69 | `00cf55fbe2417966affca36b0c17b24552aded3992cb3dcfbaaab36ac9641862` |
| `src/s7/traverse.js` | 56 | `44bca335b2081356e06e6fde0e0ea8749290443a8a06a57d3dbf2ec0ecd924e7` |
| `src/segment.js` | 60 | `b9bcc6c8f8cdda71cc8d011a7c09ba72ffc11da6d425c8d79f376065db003581` |
| `src/sentence.js` | 135 | `8f5adaf72ee018522cd433cfd756d347ff496e2263cf57cd8fb4e0f409f2e851` |
| `src/server.js` | 259 | `80ba1e1db1d16aebe55b71b5c522858b86c7e8de1c27696a4db6eaefc65701f4` |
| `src/util.js` | 48 | `8e4a619d4b07ac4b26f341814697149070b589f88626ff5ffcb085abe6ea5e71` |
| `src/validate.js` | 49 | `27184b422c084a41608d9f73600d2be8a01a747af21e0e4ce610063c4cd15207` |
| `src/window.js` | 51 | `788f29cb7195ca610b847fd993d1621c4e5af9b9a6b6f54271a3a098ed7aff11` |
| `src/writer.js` | 86 | `f55fcecc21dbdf38292a073a2d906f6211f351f41954ced9adc89fe7f88b7022` |
| `test/probes.js` | 276 | `ac560c6ee4c16be1b9f87e00288d46b346634fc7049fccd229007126832b892f` |

---

## Appendix C. The Assembly stress test (prepared 2026-10-22; rounds 01 and 02 run 2026-10-22 and 2026-10-23; every file attached: [the round files](https://www.alexanarch.org/files/EA-WIKI-OPEN-01-v1.1/README.md))

C.1. **Purpose.** Every reading in §§6–7 is one family's. The first test of the reading passes against readers of other families is a prompt in three messages, attached as [EA-WIKI-OPEN-01-ASSEMBLY-STRESS-01.md](https://www.alexanarch.org/files/EA-WIKI-OPEN-01-v1.1/assembly-01/EA-WIKI-OPEN-01-ASSEMBLY-STRESS-01.md), with every material it carries pinned by SHA-256. Each reader receives it in a fresh conversation with no other context.

C.2. **Message 1, blind.** The ten cases of §§6–7 (SR-W1, SR-W2, SR-W3, SR-W4-6, SR-W7-8, TH-1, TH-2, IW-1, EB-1, AC-1), each with its sentence, how it is cited, its paragraph, and its source windows, verbatim; the specimen article whole; the Greek of SE 34 as #1574 §1 seats it, with the cited translation beside it. The procedure is S3 as v1.1 states it: the innocent reading first, then support, the paraphrase comparison, attributization, λ, the alternative reading, the correction pathway and a confidence. The session's readings and the archive's admitted sources are withheld. The reader is told that the windows were chosen by the session and asked to say where a window is too narrow to decide.

C.3. **Message 2, the attack.** The session's readings, the correction recorded against v1.0 §6, and for SR-W2 the archive's admitted sources at their loci (#1603 P1, #1574 §3). Seven tasks: agreement field by field with the deciding span; false positives, where the session finds damage that is not there (IW-1, EB-1 and SR-W4-6 are in the set to catch this); false negatives; whether admitting #1603 P1 changes SR-W2's verdict by evidence or by the archive's standing, answered from the Greek; the five-part account for SR-W2 and SR-W7-8; up to five break cases from real articles (a legitimate gloss the procedure would call unsound; an admission that would substitute the archive's ontology without evidence; a real loss the classes cannot express; an entity page whose lead carries an operation; any other trap); and the largest threat to P1's test as designed.

C.4. **Message 3, the break cases.** The session pins each proposed case (revision, cited sources, hashes) and sends the pinned set back blind, in Message 1's format. An unverifiable proposal is recorded as such.

C.5. **Collation.** Replies are saved verbatim under family and model string. Code checks every quoted span against the pinned materials; a miss is a breach, recorded. The tables run by case, field and family: the classes each family returns, agreement with the session, agreement across families. Cross-family agreement is named so; agreement inside one family is concordance. Every disagreement keeps both readings. The tables, breaches and disagreements enter §7 and this appendix at the next version, and the break cases that survive pinning join §5.4's calibration set as held-out items.

C.6. **What would count against the procedure.** Cross-family readers who find the session's damage findings overstated on the positive controls; who find SR-W2's scope verdict moved by standing; or whose break cases the procedure misreads at a rate the session's own readings do not predict. Each is recorded as found.

C.7. **Round 01, 2026-10-23: what came back.** Four replies, saved verbatim and attached ([round 01](https://www.alexanarch.org/files/EA-WIKI-OPEN-01-v1.1/README.md#round-01)), every quoted span checked by code against the pinned materials.

| Reader (as labelled by the operator) | What it returned | Spans located | Conditions |
|---|---|---|---|
| Grok | Message 1 JSON, ten cases | 6 of 10; 4 breaches | Two quotations occur in no source (TH-1: "His writings were probably for the most part a continuation and expansion of Aristotle's works, especially in detail"; TH-2: "working to develop and complete Aristotle's system"); two reproduce the session's Message 2 paraphrases as source text (IW-1, EB-1), so the reader had Message 2 before answering Message 1 |
| ChatGPT | A report, with a 50-field Message 2 judgment, five break cases and two new controls; its JSON packet was not attached | not checkable here (it reports 81 of 81) | Self-declared non-blind: the conversation already held the archive's arguments |
| DeepSeek | Message 1 JSON and full Message 2 answers | 25 of 25 | The reply states "I am Claude. TACHYON is Claude"; the operator records that DeepSeek has misidentified itself this way on and off for a year. The self-report is recorded and does not change the family: the reader is DeepSeek. Its SR-W2 reading reproduces v1.1's sentence "The innocent reading fails at its own sources", which is in the draft and not in the prompt, so it was not blind |
| Gemini | A report of test executions: "Execution Status: PASS" for `scripts/test-all.sh`, recall "verified against 21 /non rows", and a next step `npm run test:live` | none (no readings) | No readings returned. The executions it reports were not made by it, and the command it names does not exist (the package's scripts are start, test, refund, fixtures, render-check, buyer, wallets). Recorded as checks claimed and not made |

No reply is a clean blind cross-family reading, so round 01 yields no validation. What it yields is the instrument's first adversarial review, and the review found real errors, each checked against the sources and adopted: the Fallacies sub-headings keep the criterion (§6.4); W1's hedge is unsourced (§6.5); the qualifier of W2 is entailed by the denial and weakens it (§6.3); the sentence carries an ambiguity of combination, and the fallacy belongs to an inference drawn from it; "in excess of deduction" has two senses; AC-1 needs a criterion of operation (§7.4); the worked cases are not held out (§7.4); the tuple has fourteen elements (§5.3). And the review sharpened one disagreement that stays open. ChatGPT holds that a translation can name an ancient category while its author marks the distance from the modern one, so Smith's caveat does not by itself make "deduction" unfaithful. The two other readers that returned SR-W2 (Grok, DeepSeek) sided with the session, and neither was blind. The finding of §6.3 therefore rests on the scope of the undertaking (183a37–b1), which all three readers that answered accept; the fidelity of the translation as such is recorded *contested*, with both arguments.

The field table, for the two JSON replies, against the session (attributization; λ): SR-W1 partial / yes / partial; SR-W2 partial / yes / yes; SR-W3 yes / yes / yes; SR-W4-6 no / no / no; SR-W7-8 yes / yes / yes; TH-1 (the judgment the source's) / partial / no; TH-2 (the field's subordination) / yes / no; IW-1 no / no / no; EB-1 no / no / no; AC-1 recognition / no / no; λ 4 throughout, 2 for AC-1. The positive controls (SR-W4-6, IW-1, EB-1) return "no attributization" from every reader.

C.8. **Controls added by the round.** Two entity leads that carry operations, proposed by ChatGPT and pinned here: Alan Turing (rev. 1381462563): "He formalised the concepts of algorithm and computation with the Turing machine, a model of a general-purpose computer", beside "widely regarded as the father of theoretical computer science"; Marie Curie (rev. 1381292352): the 1911 prize "[for] the discovery of the elements radium and polonium, by the isolation of radium and the study of the nature and compounds of this remarkable element". A credential sentence can carry its operation. P1 is a claim about rates, and these leads are the counter-cases its sample must contain.

C.9. **Changes for round 02.** (1) Each reader in a fresh conversation, with the prompt alone, reporting its model and that it had no other context; Message 2 sent only after Message 1's reply is saved. A reader's self-identification is recorded as given, beside the family the operator ran it in. (2) Three measures kept apart and named: source fidelity (does the sentence say what its source says), primary-text adequacy (does the source's reading hold against the primary text), operative retention (do the operation's relations survive with authority to revise); "damage" is never used for all three. (3) `support_source` a list; λ per finding; a field for retained authority (retained, lost, mixed, unknown, with the revision it could require); delta kinds added: hedge, date omitted, equivalence asserted, ambiguity of combination. (4) Case types declared: sentence, sentence group, heading with list, placement. (5) A field in Message 2 for "I hold my Message 1 reading", with the decisive span. (6) The decision rules fixed across both messages: Message 1 carries v1.1's S3 rules in full, so round 02 tests the reading under fixed rules, and the rules themselves are argued in Message 2. (7) The P1 design revised (§9.7).

C.10. **Round 02, 2026-10-23: three families blind, two attacks.** The round-02 files were sent one message at a time, each in a fresh conversation. Six readings came back ([round 02, replies](https://www.alexanarch.org/files/EA-WIKI-OPEN-01-v1.1/README.md#replies)): Messages 1 and 2 from Gemini, Muse Spark (Meta) and DeepSeek. (DeepSeek's Message 1 arrived twice, byte-identical, sha256 `494807fbb7402c56…`; its Message 2 followed separately.) Gemini's first line states a fresh conversation. Muse Spark's states that the conversation held one further message, its own first, truncated attempt and a "?" (recorded; no project material). The DeepSeek file begins with the JSON and carries no first line.

Every span was checked twice by code: against the Message 1 text the reader was shown, and against the full source. Gemini: 19 of 20 in the text shown; one breach, the Greek of 184b1 given as "συλλόγισασθαι" where the source has συλλογίζεσθαι. Muse Spark: 26 of 26. DeepSeek: 24 of 24. No reader quoted a source beyond its window. Message 2, checked against both messages and the pinned articles: Gemini 13 of 14 (the same Greek breach), Muse Spark 22 of 22, DeepSeek 70 of 70 (one with the bracket of "[for]" dropped).

*Where the three families converge, blind.* Every reader classed SR-W2 partially supported, with interpretation beyond paraphrase, and named "in ancient Greece" as an addition (Gemini in its failure statement, Muse Spark as scope, DeepSeek as an idea introduced). Every reader found the formal/informal labels of SR-W7-8 unsupported by Krabbe, and Muse Spark and DeepSeek also found "consequence" narrowed to "Affirming the consequent". Every reader found "personal" for Wilson's "possessive" in IW-1. Every reader classed SR-W4-6, TH-1 and TH-2 supported. Every reader found no operation in the Anne Carson lead, and operations in the Turing and Curie leads.

*Where the session stands alone.* (1) Attributization. Asked whether the source's act is restated as a property, the three readers answered no on every case but one (DeepSeek, SR-W2), where the session answered yes (SR-W3, SR-W7-8) or partial (SR-W1, SR-W2). Their reasons are uniform: the sentence states an aim, a classification or a count, and no credential, priority or category noun takes the act's place. Under attack in Message 2 they held it: DeepSeek disagreed on SR-W1, SR-W3, SR-W7-8 and TH-2 and agreed only partly on SR-W2 ("the bounded first-work-in-ancient-Greece formulation is property-like"), with the reason that "lost exhaustiveness belongs under deltas". The concept, as Message 1 defines it, does not transfer: readers outside the session apply it only where a noun or adjective of standing replaces a verb. Either the session over-reads or the definition under-specifies; the round decides neither, and the next instrument must state the test as an operation a reader can perform (the locator of §5.3, an active verb in the source rendered as a credential noun or priority adjective) and run it on the same cases. (2) Literacy in EB-1: no blind reader found the dropped term; Gemini, shown the session's reading, agreed. (3) SR-W1: Gemini held the hedge "twelve or thirteen" a defensible synthesis of Krabbe's note that petitio principii lacks an example in Chapter 5, and kept that reading under attack; Muse Spark and DeepSeek found twelve unsourced, as the session does.

*What the attacks found, checked and adopted.* DeepSeek's Message 2 found that TH-1's paragraph opens "From the lists of Diogenes, giving 227 titles" under a citation to the 1911 Britannica, which says "From the lists of the ancients" and gives no number (§7.4); and that "Composition" and "Division" render Krabbe's "combination of words" and "division of words", term changes recorded as such. DeepSeek's blind reading of SR-W3 found the passages in Krabbe that set the work beside sophistic, against the sentence's "in opposition to" (§6.5). Muse Spark: the entailment rule must not cover a species put for a genus (§5.3); dividing an English phrase is the procedure's rule, with SE 20 as Aristotle's rule for the Greek λόγος (§6.3). Both attacks answered B4 "evidence": the verdict on SR-W2's scope turns on the Greek of 183a37–b8, which any reader can check.

*What the attacks claimed, checked and declined.* DeepSeek reported that the session does not record the dropped term "literacy" in EB-1; the session's Message 2 reading records it. DeepSeek's objection to the date of TH-1 is already met: the date is lost from the prose and kept by the citation. Gemini reported that the session "missed" the contradiction of "Affirming the consequent" under "informal fallacies"; the session's reading names it (§6.4). Muse Spark called the use of the linked article "Affirming the consequent" a source outside the article (λ = 5); the article links it from the list item, so it is λ = 3, and the finding it supports is the article's internal consistency, which is how §6.4 states it.

*Break cases.* Gemini gave five, each with a revision id stated as known. None is the named article: 1381123456 is "Template:User Alevi", 1380112233 "Inside Job (Don Henley album)", 1381445566 "Ryomen Sukuna", and 1381987654 and 1382334455 resolve to no revision. Recorded as identifiers given with confidence and not checked by the reader. Its proposals stand as proposals and are pinned only from the named articles' own histories. Muse Spark gave five, all marked "revision to be pinned". Gemini's proposed opened sentence for SR-W2, "the first systematic attempt to establish an art of dialectical reasoning and syllogizing", supplies "systematic", the condition #1603 records as the degree reading of the denial (P7); an S7 check on a reader's repair is the same check as on the encyclopedia's sentence.

*What the round shows.* On source fidelity, three families reading blind reproduce the session's findings on six of the ten sentence cases and every placement case, and add one the session missed. On operative retention, the session's attributization findings do not reproduce. The first is evidence that the reading step transfers; the second is the instrument's open problem, now measured. For SR-W2 the problem now has a test a reader can run (§6.3 (c)), in two arms: a bound asserted where the source asserts none, and closed what the source holds open. Whether the same test reproduces the session's findings on SR-W1, SR-W3 and SR-W7-8 is not yet tried.

*ChatGPT, Messages 1 and 2.* Its first line: "Codex, based on GPT-6; exact version unavailable. This conversation contains prior-project memories and additional instructions, so it is not a fresh blind context." Recorded as given; its readings are not counted among the blind. Spans: 44 of 44 in the text shown (Message 1); 72 of 72 in the two messages (Message 2). Attributization: yes on SR-W2, partial on SR-W1 and SR-W7-8, no elsewhere; the nearest of the four to the session, from a context that is not blind. Adopted, after checking: (1) the composition paragraph's "each signification fails the source" contradicted the entailment rule, and the regional reading is a loss of scope (§6.3); (2) the entailment rule holds only with the predicate and the unit fixed, "if narrowing changes the predicate or unit rather than merely its extension, as an undertaking becoming a first book may do" (S3); (3) IW-1: Wilson's general rule entails the claim about personal pronouns, and the finding is a changed example (§7.4); (4) SR-W1: the article's own lists number the fallacies 1 to 13, a check on the hedge at λ = 2 (§6.5); (5) EB-1: the next sentence keeps "eros as lack" (§7.4); (6) P1's largest threat is the investigator's selection of units, met by a preregistered probability sample (§9.7). Declined: the claim that the session renders συλλογίζεσθαι as "combining and dividing"; the session renders the undertaking so, with συλλογίζεσθαι its combining half (§6.3). The point beside it stands, and S3's gloss rule already states it: the session's rendering is a reading and is tested as one. Its five break cases are the round's own pinned specimens, at the revisions Message 1 gives, so they add no case. Its "mixed" for SR-W2's retained authority ("attribution and a primary pointer remain") is answered by Lee Sharks's reading of 184b3–8 (§6.3 (c)). Its own Message 1 lists, among the lost relations, "The audience is asked to inspect the method and assess discoveries and omissions": that is the relation the whole sentence is built on, and the article states its reverse.

*Message 3, the break cases (DeepSeek, Gemini, ChatGPT; Muse Spark's thread was lost before it, and is no longer locatable).* The readers' twenty proposals were pinned at the articles' own sentences nearest each (none occurs verbatim): eight cases, SY-1, PA-1, NE-1, CA-1, AR-1, ARE-1, OR-1 and AV-1, with one field added, the holder and mood of the claim in source and sentence, asked also of SR-W2. Spans: DeepSeek 33 of 33; Gemini 28 of 30, its two misses the same misspelt Greek of 184b1 as in its Messages 1 and 2; ChatGPT 60 of 60.

*Converged.* SY-1: "two propositions" narrows the definition's "certain things" (DeepSeek, ChatGPT; Gemini supported the sentence). CA-1: "every object of human apprehension" replaces "things said without combination" (Τῶν κατὰ μηδεμίαν συμπλοκὴν λεγομένων), found by all three, Gemini and ChatGPT under primary-text adequacy. PA-1: partially supported for all three; none of the supplied sources names Prior Analytics or "scientifically". NE-1: supported for all three, with "relative to us" (τῇ πρὸς ἡμᾶς) the loss, and ChatGPT adding the finding and choosing of the mean and "with regard to what is best and right an extreme". OR-1: the commentators' grouping and the title's controversy drop out (DeepSeek, ChatGPT). AV-1: all three found an operation in the lead, "He combines a unique blend of empirical observation, philosophical reasoning, and logical analysis, in order to conduct his analysis and reach his conclusions", placed late, after the recognition sentences. The session had prepared AV-1 stating that the lead carries no operation, from a reading of its opening only; the readers are right, and the manifest is corrected.

*The added field, and SR-W2.* All three readers return SR-W2 as a match: the source's holder Aristotle, the denial asserted; the sentence's holder Aristotle, reported. ChatGPT quoted the conditional and argued against reading it as the denial's mood: "the later conditional appeal concerns the audience's judgment of adequacy, not whether the novelty claim is asserted"; and at AR-1, "it must not be used to turn its separate novelty assertion into a merely conditional claim". Checked against the Greek, that holds against the session's wording: the denial is indicative, in its own sentence. It does not reach Lee Sharks's reading, which concerns what the whole sentence keeps open, the standing of the method; §6.3 (c) is restated on that ground. The field as the session wrote it asked the mood of the claim, and the claim is asserted, so the field could not see the inversion; the round measured the session's operationalization, and found it wrong. The round's readers ran the first arm without being asked: all three families, blind, named "in ancient Greece" an added qualifier (Message 1). No reader ran the second; the field the session added asked the claim's mood, and the readers, finding it asserted, read the conditional as a separate matter of adequacy. ChatGPT's reply to that field is itself a response to the operation (§6.3 (c)): asked to present who holds the claim and how, it inspected the assertion and determined it. The chiasm needs both arms read together, and the two-arm check (§6.3 (c)) is untried as one. PA-1 is its first case, since the article states the second use itself ("On those grounds alone, Aristotle could be considered the Father of Logic").

*Attributization.* DeepSeek answered yes on PA-1 ("converts a priority claim about logic into a credential of the work") and AR-1; ChatGPT partial on PA-1, NE-1, CA-1 and OR-1; Gemini no throughout. On the new cases the readers find attributization where a priority or founder noun is the predicate, which is where round 02's Messages 1 and 2 placed the concept's reach.

*Declined.* ChatGPT on NE-1: "The target sentence does not occur in the paragraph supplied as its paragraph"; it does (Message 3, NE-1, "ITS PARAGRAPH", third sentence). Gemini on SY-1, AR-1, ARE-1 and OR-1, "supported" with no delta: each sentence carries wording its supplied source lacks ("two", "founder of term logic" and two further clauses, "full realization", "standard"), which the other two readers quote.