{"path":"research/a-sentence-as-a-pointer-into-a-landscape.md","content":"# A sentence as a pointer into a landscape: a friend's diagram read against the concept layer\n\n**Date**: 2026-10-19 · **Type**: Claude's reading of a diagram, against the corpus · **Occasion**:\nfounder, verbatim except the friend's name (kept in `.private/`): *\"Also got this from my old friend\n[…] when we talked about Deliberus and the concept layer etc yesterday. Reflect on how it challenges\nwhat thinking and ideas and ontology that I have to deal with this in Deliberus thus far\"* ·\n**State**: a reading. **Nothing here is ruled.** The four proposals in § 5 wait in this note, outside\nthe ledger, on his word (the precedent of [twelve-books-in-plain-words.md](twelve-books-in-plain-words.md)).\n\n**Read-depth note.** The diagram was read in full as an image. Read in full this session:\n[core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md),\n[store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md),\n[claim-sameness-philosophical-readings.md](claim-sameness-philosophical-readings.md),\n[the-load-bearing-unsaid.md](the-load-bearing-unsaid.md),\n[semantic-disambiguation-and-concept-tracking.md](semantic-disambiguation-and-concept-tracking.md),\n[contingency-and-convergence-across-levels.md](contingency-and-convergence-across-levels.md) §§ 0–5,\n[philosophical-foundations-stress-test.md](philosophical-foundations-stress-test.md) §§ 6–9,\n[steelmanned-critiques.md](steelmanned-critiques.md) Critique 2,\n[the-tractatus-and-deliberus.md](the-tractatus-and-deliberus.md) §§ 0–3, the ledger rows AM7, AM9 and\nAM10, and [three-minds-and-the-leap.md](three-minds-and-the-leap.md) on the pair that produced AM9.\nTwo outside facts were looked up as search summaries, not read in full: the opening of the\n*Philosophical Investigations* (§ 2a) and Firth's tie between his view of meaning and Wittgenstein's\n(§ 2a). `deliberus/extraction/` was searched for *metaphor* and *figurative*, with no hits.\n\n---\n\n## 0. The answer in plain words\n\n**The diagram agrees with Deliberus about what meaning depends on and disagrees about where meaning\nis kept.** Both say a sentence is not its meaning and that context decides the meaning. The diagram\nkeeps meaning in one shared, continuous space that each word points into. Deliberus reads meaning from\nattributed uses (AM7, ruled today: *a use is a vote on the meaning*), and the corpus holds that the\nshared space is a coordinate system, never anyone's meaning\n([core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md) § 5b, unruled).\n\n**That is not a reversal. Three real challenges come out of it, and each points at a gap in what is\nbuilt rather than in what is ruled.**\n\n1. **Interpretation is a spread, and the graph stores one reading.** The stranger rewrite turns a\n   passage into one sentence and stores it as the claim. The diagram says a reading is a region.\n   Nothing in the graph records how wide the region was.\n2. **Most variation in meaning is continuous, and the concept layer mints discrete senses.** The\n   theory already has a kind of difference for this (a *weighting*); the build mints two senses per\n   contested term by template.\n3. **A use carries more context than the occasion facts proposed so far.** The diagram's list adds the\n   relationship between speaker and listener, and the knowledge they share. The second is where the\n   unwritten crux lives.\n\nA fourth is smaller: the diagram has a layer of *possible futures* and the graph has no way to hold\nalternatives as alternatives.\n\n## 1. What the diagram says\n\nSix panels, in order.\n\n1. A sentence (*\"The tree is growing in the forest\"*) is a simplified label. Each word is a\n   compressed pointer to many possible meanings.\n2. Each word activates a distribution of possibilities in a high-dimensional space: *tree* fans out to\n   plant, species, wood, symbol, family (tree), data structure, Christmas tree, memory.\n3. The actual meaning depends on the situation: physical environment, time (season, life stage),\n   the speaker's intentions (descriptive, poetic, scientific), the relationship to the listener, the\n   current situation (conversation, text, research, art), and relevant knowledge.\n4. Understanding is multidimensional and probabilistic: the statement points to a region across\n   layers, ordered from abstract to concrete (linguistic, situational, cultural, social, emotional,\n   historical, physical and causal, alternative possibilities).\n5. The short sentence points to a longer statement that is *\"still incomplete, but far closer to\n   reality\"*: a particular tree, likely a mature spruce, growing in southern Finland in summer 2026,\n   fed by sunlight and soil, part of the forest's carbon cycle, *\"experienced by an observer who feels\n   a sense of calm and connection.\"* It remains *\"a high-probability, context-aware region, not a\n   final truth.\"*\n6. Meaning lives in a vast, continuous landscape shaped by factors at every scale.\n\n**Two cautions on reading it.** The image looks rendered by an image model (one label reads\n*symbly*), so its labels carry the gist and not each word's weight. And it was made after a\nconversation about Deliberus's concept layer, so where it resembles the corpus's own levels lens the\nresemblance is partly descent, by this corpus's own rule\n([contingency-and-convergence-across-levels.md](contingency-and-convergence-across-levels.md) § 2):\nagreement between two things that share a source is copy number, not corroboration.\n\n## 2. Where it challenges, strongest first\n\n### 2a. Where meaning is kept: a region in one space, or attributed uses\n\nPanel 1's *word as pointer* is the picture the *Philosophical Investigations* opens by questioning:\nwords name objects, and a word's meaning is what it stands for (§ 1, Augustine's picture; looked up as\nsummaries). The diagram makes that picture probabilistic: a word points to a region, not an object.\nDeliberus sits on the other side. A sense is a group of attributed, dated uses, and its label is\ncomputed from them (AA2); a theory or a person is defined the same way, by the claims that use its\nname and survive (AM10, his, open).\n\n**The fork is narrower than pointer against use.** The space in panel 2 is itself built from uses:\ndistributional semantics descends from Firth's contextual theory of meaning, which Firth tied to\nWittgenstein's (looked up as a search summary). So the real difference is *whose* uses. The space\naverages everyone's and drops the names; Deliberus keeps them attributed. The corpus has measured\nwhat the averaging costs: a model's vector for a contested word is the mean of a distribution with\nseveral peaks, a point no community holds\n([why-representations-converge-and-whether-minds-do.md](why-representations-converge-and-whether-minds-do.md);\n[contingency-and-convergence-across-levels.md](contingency-and-convergence-across-levels.md) § 4,\n*the average of paths is no path*).\n\n**What the space keeps.** The job § 5b already gave the centroid: an origin to measure differences\nfrom. *The average is harmless as a coordinate and harmful only as a bearer.* The diagram adds no\nreason to promote it, and its own panel 3 (the speaker's intentions, the relationship to the listener)\nis exactly the bearer that panel 2 lacks.\n\n### 2b. Interpretation is a spread, and the graph stores one reading\n\nThis is the challenge with the most consequence for the build. Panel 4 says a reading of a sentence\nis a region. Today a passage is read once by the extraction pass and the resulting sentence is stored\nas the claim. AA3 already proposes demoting that sentence to a dated machine judgment, and\n[store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md) § 3d rule 3 already says\nseveral rewrites from different positions are a gain. What neither says is that **the distance\nbetween readings is itself information about the use.**\n\nA passage that every reader rewrites the same way is a use whose meaning the text settles. A passage\nwhose rewrites scatter is a use whose meaning depends on something outside the passage: its source,\nits audience, its game. That second kind is exactly where AM9 (ruled today: hand the keeper each\nclaim's passage and the senses in play) earns its cost, and where a question to the concept layer\nbelongs. The corpus has measured the spread once without naming it: five models cut one sentence four\nways (store-the-uses § 3b). The pair that produced AM9 is a case of it: *distributed human\nauthority* read without its source answers *no bearing*; read in its game, one claim supports a\npriority and the other qualifies it ([three-minds-and-the-leap.md](three-minds-and-the-leap.md)).\n\n### 2c. Most variation is continuous, and the layer mints discrete senses\n\nPanels 2 and 6 put meaning in a continuum. The concept layer stores senses, and its detector mints two\nper contested term, so 49 of 90 concepts read *bifurcated* by template (core-periphery § 4, measured 6 September). The\ntheory is closer to the diagram than the build is. § 5b lists kinds of difference between two\npeople's senses, and one of them is a *weighting* (same features, different dominance); § 6 says a\nwithin-mind difference *is a weighting, not a sense*; and § 8 notes that Lakoff's *mother*, one word\nwith several models that agree in ordinary cases, has no place in the layer. The diagram's challenge is\nto build what the theory already holds: record a continuous difference as a weighting, and keep a new\nsense for differences that partition by source or position, as § 6 item 1 asks.\n\n### 2d. What a use carries\n\nPanel 3's list against what is kept or proposed:\n\n| Panel 3 | In the corpus |\n|---|---|\n| Physical environment, time | *The time the content is about* and scope, compared like time (store-the-uses § 4b, § 4d) |\n| Speaker's intentions (descriptive, poetic, scientific) | *The kind of text* and the speech act (asserted, conceded, supposed, reported, asked), AA7 and § 3c item 5 |\n| Current situation | *In reply to what*, AA7 |\n| Relationship to listener | Only *to whom*, AA7; audience is named as missing in [the-motte-and-the-periphery.md](the-motte-and-the-periphery.md) § 3c |\n| Relevant knowledge | Not a fact of the use. It is what the speaker took as shared, which is what implicit-premise recovery looks for ([the-load-bearing-unsaid.md](the-load-bearing-unsaid.md)) |\n\n**Two additions follow.** The relationship between speaker and audience (a court to the parties,\ncounsel to the bench, a friend to a friend) is a fact and can be stored beside *to whom*. The shared\nknowledge cannot be stored as a fact; the corpus already treats it as the place where false consensus\nhides the crux, and its home stays implicit-premise recovery. The diagram's list confirms that the\ntwo belong together: what a speaker leaves unsaid depends on whom they are speaking to.\n\n### 2e. Expanding a sentence and cutting it\n\nPanel 5 understands a sentence by expanding it into one long, continuous statement. Deliberus does\nboth moves: the stranger rewrite expands a passage until a stranger can read it, and decomposition then\ncuts. Three points.\n\n- **Scope and time belong in the claim**, and the corpus agrees: a trial in one country and a trial\n  in another do not contradict (store-the-uses § 4d). Panel 5's *southern Finland, summer 2026* is\n  that rule.\n- **The observer's feeling does not.** Panel 5 folds *\"an observer who feels a sense of calm\"* into the\n  statement about the tree. Deliberus would keep it as a separate claim, attributed to the observer.\n  This is a difference worth keeping: folding the perceiver into the fact makes every disagreement\n  about the world also a disagreement about someone's experience, and the second kind cannot be\n  attacked by evidence about the first.\n- **\"Still incomplete, but far closer to reality\" is the *Tractatus* ambition made probabilistic**: a\n  fuller description is a better picture of the world. The corpus's answer to the book applies: there\n  is no complete analysis, and where description stops is found by trying and dated\n  ([the-tractatus-and-deliberus.md](the-tractatus-and-deliberus.md) § 0). Panel 5's *continuous*\n  carries the formalization paradox (Critique 2): a statement cut into parts can lose its joins. The\n  standing answer is the recomposition test: if the parts cannot rebuild the whole, they are reasons,\n  not parts (core-periphery § 5a-ii).\n\n### 2f. Possible futures\n\nPanel 4's lowest layer and panel 6's *multiple possible futures* hold alternatives as alternatives. The\ngraph holds a claim about a future as one more claim, and holds no credence over branches; the stress\ntest names that missing layer, judges it out of scope, and asks for a credence layer or a plain\ndisclaimer\n([philosophical-foundations-stress-test.md](philosophical-foundations-stress-test.md) § 8a). The diagram\ndoes not change that verdict. It is noted here because it is the one layer of the diagram with no home\nat all.\n\n### 2g. The example is the easy case\n\n*Tree* is a concrete noun. By the levels lens its core is assembled by the world and converges between\nminds; context changes its periphery (associations, feelings), not what is claimed. Deliberus's\nontology rests on context weighing differently by word: *penguin* against *fair*, *tree* against\n*armed attack*. The diagram treats every word alike. Its own panel 6 has the case where it becomes a\nDeliberus case: *tree as a symbol*. Said of a movement, *the tree is growing* is a different claim. The\nextraction code has no handling for figurative use (no hits for *metaphor* or *figurative* in\n`deliberus/extraction/`); AA7's *kind of text* is the nearest home.\n\n## 3. What it does not change\n\nThe proposed storage law (AA1: every stored thing is a use, a machine judgment or a reading) and AA2,\nboth unruled, and the rulings AM7 and AM9 stand, and the diagram supports their premise: context decides meaning. It disagrees only about where\nmeaning is kept, and on that the corpus's reasons are measured. The core and periphery split is\nuntouched; the diagram's panel 2 is a periphery drawn for a word whose core nobody disputes.\n\n## 4. Questions worth putting back to the friend\n\n- Whose distribution is panel 2? Everyone's averaged, or one person's? If one person's, where are the\n  other people's, and what does the diagram do when two of them differ?\n- When is a statement expanded enough? Panel 5 stops somewhere; what decided where?\n- Is the observer's feeling part of what the sentence says, or part of what the observer says?\n\n## 5. Proposals, outside the ledger, unruled\n\n| Name | Proposal |\n|---|---|\n| **The spread of readings** (withdrawn the same day, § 6: tokens for little; the code count of AA6 covers most of it) | Read a passage more than once (two model families, or two framings) and record how far the rewrites differ. Store the distance as a property of the use, beside the rewrites in the record of machine judgments. A wide spread marks a use whose meaning lies outside its passage: it raises a concept-layer question and is where AM9's passage matters most. A measurement first: the spread over the passages already extracted, against the claims people later corrected |\n| **The relationship beside the audience** | AA7's occasion facts gain *the relationship between speaker and audience* beside *to whom* |\n| **Weighting before a new sense** | Before the layer mints a second sense, it asks whether the difference is a weighting (one of § 5b's kinds). A sense is minted only where the uses partition by source or position (§ 6 item 1). This is the build catching up with the theory, not a new rule |\n| **Figurative use** | AA7's *kind of text* includes figurative use, so a claim read literally from a figurative passage is marked, not stored as a literal claim |\n\n## 6. His question: why not copy the author's words as the claim? (2026-10-19)\n\nOn the first proposal: *\"Sounds like lots of tokens spent for dubious value. Doesn't it suffice to\nsimply instruct the machine extracting from sources to copy claims verbatim? Are we not doing this\nand if so why not? Were/are there good reasons?\"*\n\n**His objection to the spread test stands.** It spends a model call per passage to learn what a\ncount by code mostly tells for nothing: where a rewrite added words that are not in the source,\nits meaning may have moved (AA6). The proposal is withdrawn in favour of that count.\n\n**What the pipeline does with the author's words, read in the code today** (`focused.py`,\n`models.py`, `graph/store.py`):\n\n| Wording | Made by | Stored |\n|---|---|---|\n| The exact quote (`source_span`) | pass 2a, copied from the source | yes, on the claim |\n| The assertion *\"in the AUTHOR'S ORIGINAL WORDS\"*, pronouns and context kept | pass 2a | **no**: handed to pass 2b and dropped |\n| The rewritten claim, made to pass the stranger test | pass 2b | yes, as the claim's text |\n\nSo the pipeline already writes each claim in the author's own words and throws that version away.\n\n**Why the stored claim is a rewrite, and whether the reasons hold.** Measured 2026-10-06 over 1,144\nextracted claims ([store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md) § 3c):\nthe quoted passage is a fragment (median 14 words against the claim's 23; *\"but both eventually hit\nceilings\"*), 15% of claims share a passage with another, and claims from different sources need one\nform before they can be compared or linked. Those reasons hold: a fragment cannot be read, attacked\nor matched alone. The cost is measured too: 29% of a claim's content words appear nowhere in its\nsource, against 5 to 6% in the two extractions made by hand, and that is where weight gets added\n(*a second issue* became *a significant technical hurdle*).\n\n**Ruled and built the same day** (*\"Yes please, build them.\"*): pass 2a's assertion is kept on each\nnew claim as `author_wording`, and code marks each rewrite with `words_from_passage`,\n`words_from_source` and `words_added`, the words found nowhere in the source\n(`deliberus/extraction/wording.py`, ledger AA6). Neither asks the model anything: the fields are\nhidden from the reply form, so the pass's instructions and version are unchanged. Together with the\nquote already stored, every new claim has three wordings: the author's passage, the author's\nassertion, and the stranger's sentence. Not deployed when written; claims stored before carry no\nmark, and a backfill of the count is possible only for sources whose text was kept.\n\n### 6b. His follow-up questions, and what was found (2026-10-19, later)\n\n**The brackets the prompt asks for are almost never written.** Pass 2b's instruction for the claim\ntext is *\"Rewrite so it passes the STRANGER TEST … Replace ALL pronouns with names, resolve\n\"it/this/they\", add [bracketed context].\"* Brackets would make added context visible in the text\nitself. Measured over the same 1,365 stored claims: **48 (4%) contain a bracket**. Parentheses do\npart of that job: of 296 parentheticals in claims, 134 hold text that is in the source (the author's\nown aside) and 162 hold text that is not, so some added context is set apart in round brackets; the\nrest is folded into the sentence unmarked. Nothing in the prompt asks the model to keep the author's\nwords or to leave out weight (*significant*, *objectively*); that is AA6's open instruction half.\n\n**Other characters, measured the same day.** Nothing systematic. Em dashes appear in 87 claims, in\n28 of them where the source has none (the model's punctuation, not the author's). One claim stores\nliteral backslashes before its quotation marks (`In their work \\\"The Feedback Fallacy,\\\" …`), an\nescaping fault somewhere between the reply and the store.\n\n**A backfill has two parts with different reach.** The mark needs only the claim text, its passage\nand the source text, all stored on `extractions` (`raw_text`, `combined_json`), so it can be\ncomputed with no model call, but only where the source text was kept: 18 of the 35 extractions,\nholding the 1,365 claims measured here; the other 17 (664 claims) would first need their text\nfetched again. The author's wording needs pass 2a's reply. The record of machine judgments keeps\nevery reply since 2026-10-06 (`judgments.note_call`, ledger Z1), but no source has been extracted\nsince that day, so the record holds it for the two sources extracted on 6 October and for nothing\nelse; every other claim gets it only by running pass 2a again.\n\n**Proposal, unruled: the mark as standing work** (R8, machine work as a standing property,\n`machine-work-as-a-standing-property.md`). A claim with no mark, or one counted by an older matcher,\nowes one; a matcher change such as the final-*e* rule makes every mark outdated. Because the count is\ncode, R8's brake on model calls does not bind it and the queue can be drained completely. **Extraction\nitself as standing work costs no more model calls than today** (his correction, the same evening:\n*\"A source is generally extracted once anyways, no? If the source text doesn't fundamentally change,\nwhich can be checked, there's no reason to rerun the whole pass\"*). The condition is a source whose\ntext, by the hash already stored (`raw_text_sha256`), has no extraction; that is one extraction per\nsource, as now. A prompt change marks earlier extractions outdated, and R8's design reports that\nqueue and never drains it on its own, so re-extraction stays a decision.\n\n**What the word match cannot do, and who could (proposal, unruled; the measurement is not).** The\nmatcher is exact words with crude endings dropped. A lemmatizer would add little: of the 5,621 words\nit flags as added over the 1,365 claims, simplemma's lemmas find 147 (2.6%) in the source after all\n(*found*, *using*, *societies*). The flags are real changes of word, not spelling. What is left is a\nquestion of meaning, whether an added word says more than the passage, and that is one short item\nagainst one concrete question, the use the decision model Jev was measured good at. Per the global\nrule on case review, a person judges a few marked claims one at a time first, so Jev's answers have\na reference.\n\n**The hurdle claim and the hedge ruling.** AM6 (a hedge is stance, not strength) lets *will* stand\nfor an author's *may* only in a conclusion minted from a conditional, which nobody asserts, with the\n*may* kept as stance on the link (`three-minds-and-the-leap.md`). The hurdle claim is the author's\nown assertion, where the older discipline holds: never raise the certainty an author gave. Pass 2b\nraised it (*significant*) and nothing recorded the raise until the mark. His revisit of *may* is an\nopen loop in `TODO.md` (AM6, ruled *for now*). His proposal the same evening to take stance hedges out\nof every claim, with pros, cons and guards: `three-minds-and-the-leap.md` § 5b.\n\n### 6c. The instruction built and tried (2026-10-19, evening; ledger AA6, instruction half)\n\n**Ruled** (*\"Yes!\"*) and built: pass 2b's instruction for the claim text says to change only what a\nstranger needs, keep the author's words for the rest, put added context in square brackets and never\nin round ones, add no weight or certainty, keep what the author asserts rather than a definition of\nit, and use no em dash the source lacks. Code then clears escaped quotes and swaps an em dash the\nsource never uses for a spaced hyphen (`wording.tidy_claim_text`, his word: *\"No em dashes please if\nthe source didn't have them, simple dashes if need be.\"*); the web page had been cleaning escaped\nquotes for display (`web/src/lib/index.ts`, `cleanText`), which is why the stored fault went unseen.\nThe mark now also lists the added words declared in brackets (`words_added_in_brackets`).\n\n**The trial, four model calls under R8's brake.** A full pass 2 on the hurdle argument (two calls)\nsplit it into four atoms instead of the stored ten and left the self-evaluation sentence out, so it\ncould not answer the question. The remaining two calls gave the old and the new instruction the same\nten author sentences (the stored passages) as pass 2a's atoms:\n\n| | Words found nowhere in the source | Of those, inside square brackets |\n|---|---|---|\n| Old instruction (served by the fallback, gemini-3.1-flash-lite) | 18 | 0 |\n| New instruction (gemini-3.5-flash) | 6 | 6 |\n\nThe two runs were served by different models, because the primary was out of quota, so the comparison\ncarries that confound. The hurdle sentence under the new instruction: *A persistent issue [in executing\nlong-running agentic tasks] is self-evaluation [where agents judge their own work]* (*persistent* is in\nthe source elsewhere; the author's *second issue* is still not kept word for word). Under the old one,\nthe same model family wrote *a critical, previously unaddressed challenge* and *objectively mediocre*.\nOne addition to watch: *[his] colleagues*, a gender the source does not state.\n\n**Why round-bracket asides occurred** (his question, sampled from the 162 not found in the source):\nthey are the rewrite's own glosses, mostly definitions of a term (*J-space*, *Large Language Model*,\n*Anthropic's AI language model*) and a few quoted words. They are not assertions the author made, so\nsplitting them into claims would make claims of machine glosses; a term's meaning belongs to the\nconcept layer's senses. The new instruction sends added context to square brackets, where the mark\nlists it as declared. (Refined in § 6d: a gloss that defines a term becomes the author's unexpressed\ndefinitional commitment, marked as reconstructed.)\n\n**The stored claims were not replaced.** He said to replace them if the output improved, and it did.\nReplacing is not a text edit: a claim's text is written once, and the ten claims of that argument\ncarry 73 links (25 supports and 18 attacks from extraction, 13 supports from the linker, 2 relevance\nclaims, and others). The graph has a supersede step only for a split part\n(`proposed_split.py`, a reworded part marks the old one `superseded`), none for an extracted claim\nwith its links. Which links move to the new wording (the extraction's own structure) and which stay\nwith the old (people's contributions, which were about the old words) is a design call for him.\n\n**The decision-model check on added words, built and tried** (`deliberus/extraction/added_words.py`,\nhis *\"Go ahead, proceed.\"*): one request a claim, one yes/no question a word, *does this word say\nsomething the passage does not?* Three live calls, about $0.0001 in all. The hurdle claim's four words\nscored 0.87 to 0.92. Context the new rewrite declared in brackets scored 0.57 to 0.74 (*Anthropic\nresearchers* for *we*; *one goal of Anthropic research is*). It separates the two, without a clear line\nbetween them yet, which is what his judging a few cases one at a time is for. Nothing calls it from\nthe pipeline yet; `tests/test_decision_api.py` now allows exactly this importer of the client.\n\n### 6d. Replacing a claim: what travels, and the bracketed glosses (2026-10-19, late; unruled)\n\n**What is on the hurdle claims, read from the graph and the database.** All 73 links are machine-made;\nno person has voted, reasoned, commented or answered a question on any of the ten claims. 45 of the\nextraction links carry `confirmed = true`, which is a code default of `store_relationships`, not a\nperson's word (`miscast_parts.py` docstring). The keeper and `scripts/standing_work.py disagreed`\nskip confirmed links, so that default shielded extraction's links from every later machine reading:\nthe shielding closed for people earlier the same day (`TODO.md`, the wrench question), here held by a\ndefault nobody chose. ⚠ Corrected the same night: this was not waiting on his word. He had ruled on\n2026-09-28 to retire the field (ledger Q32, *\"Yes.\"*), the build was never done, and new code read the\nfield twice after the ruling. He remembered it (*\"I thought we'd been over this in the distant past\"*)\nand asked for it gone; ledger Q32 was built on 2026-10-19 (`what-confirmed-means.md` § 7).\n\n**His question** (after reading the other session's wrench finding): *\"I'm not sure about 'anything\npeople added stays on the old one'. If what people added risks being lost/spread out due to old claims\nbeing effectively superseded by 'better' new claims then that's bad. But the sameness keeper could\npotentially merge/cluster the old and the new. But then why not simply move what was added to the old\nclaim to the new? Perhaps because the lingo of the surrounding claims added will then mismatch\npotentially?\"*\n\n**Reading: a judgment travels only if it was made about what is still there.** It is ledger AL15\n(*the whole as a record*, open) applied to a rewording instead of a cut: the same three routes, which\nthat row lays out for a cut claim's reasons moving to its parts, and the same answer to keeping the old\none (his words there: *\"we might want to keep the original claim as telemetry more or less\"*).\n- *Extraction's own links* (this sentence of the argument supports that one) were made about the\n  argument's structure, which a rewrite of the wording does not change. They move to the new claims,\n  matched by the passage each claim quotes, and drop the `confirmed` default on the way so later\n  passes can read them.\n- *The linker's links and relevance claims* were judged against the old words. As in AL15, each is\n  read once against the new wording and moved, the old kept visible; until that reading has a measured\n  rate it is machine work resting on machine work, which T1 caps. The reading is owed by condition, so\n  it costs no call the old door would not have spent (R8).\n- *What a person added* (a reason, a vote, a confirmation, an answer) was about the old words. His\n  lingo worry is right and goes deeper than style: an attack on *significant hurdle*, moved onto\n  *persistent issue*, attacks a word that is no longer there, and moving it silently speaks for its\n  author. So the old claim stays visible, marked replaced, with a link to its successor; the new claim\n  shows what was said against and for the old one, so nothing is spread out of sight; and the person\n  who added it gets one proposal: carry it over, or leave it with the old wording. That is the ruling\n  *nothing a person wrote or confirmed moves without a proposal*, and it asks the one person whose words\n  they are, never a janitor's round. The keeper reads old against new first: *same* makes the proposal a\n  one-step carry; anything else says what changed. His *merge/cluster the old and the new* is stage four\n  of the sameness ladder (a group carrying the union of its versions' authors, sources and reasons),\n  unbuilt and unwalked; until then the replaced-by link is what keeps the two together.\n\nFor the hurdle source none of this has a person in it: replacing it moves the ten claims' structure and\nreads the linker's thirteen links and the two relevance claims once against the new wording. Unbuilt;\nit is decision 3 of `machine-work-as-a-standing-property.md` § 7.\n\n**His follow-up on the glosses:** *\"Ok maybe the relevant passes could split the stuff in brackets out\nto separate claims that define those terms (simply by mentioning them, tagged with the same\nconcept-sense potentially)?\"*\n\n**Reading: yes for a gloss that defines a term, attributed to the author as an unexpressed commitment\nand marked as the machine's reconstruction; no split for a bracket that resolves a reference.**\n- The shape exists: a definitional claim with a `DEFINES` link to the claim it clarifies and a link to\n  its `Concept`, built for a person's clarification of a contested term (`api.py`, argument label\n  `concept_clarification`).\n- Only a gloss that says what a term means splits (*J-space*, *Large Language Model*). A bracket that\n  says who or what a word points to replaces that word and stays in the claim: under the new\n  instruction *we pushed Claude further* became *[Anthropic researchers] pushed [the AI model] Claude\n  further*, so the stranger test is passed and the brackets only mark the words that are not the\n  author's. (His question, the same night: *\"Shouldn't these be rewritten to pass the stranger test,\n  that is, not say \"we\" but instead have \"we\" replaced with the entity in question?\"* They are.)\n- **Whose claim, corrected the same night.** This section first said the machine's, never the\n  author's. He asked: *\"Why not attributed to the author? Aren't they implicitly making a definitional\n  claim?\"* They are, and argumentation theory agrees on both halves. Pragma-dialectics' unexpressed\n  premise rule forbids attributing more than the arguer's commitments and forbids disowning one's own\n  (van Eemeren and Garssen 2026, § 6.3); Walton and Reed (2005, p. 365) require an inserted premise to\n  be marked as the analyst's hypothesis, and Araucaria marks such premises *missing*\n  (`what-a-link-rests-on/outside-reading-argumentation-theory.md`, Q1-11 and Q1-19). So the claim is\n  the author's commitment, in the machine's words: attributed to the author, marked unexpressed, its\n  words all counted as added by the rewrite mark. Two challenges then stay apart, as they should: *the\n  author did not mean that* attacks the reconstruction; *that is a poor definition* attacks the content.\n  Where the source itself defines the term, the author's own definition is used and nothing is\n  reconstructed.\n- It stays a use, never the meaning. `store-the-uses-read-the-meaning.md` names a sense stored as a\n  model's paraphrase as one of the two breaks of the storage law, and core-periphery § 0 warns of\n  centroid senses; an attributed, marked reconstruction is one author's use on the concept, never its\n  label.\n- No extra call: pass 2b returns the term and its gloss in the answer it already gives.\n\n### 6e. A hedge out of every claim, tried (2026-10-20; ledger AM6 widened)\n\n**Ruled** (*\"Proceed. We can always refine this later.\"*) and built in pass 2b: the rewrite takes the\nauthor's confidence out of the claim into `hedge` and `hedge_reading` (stance, content, ambiguous),\nwrites the definite form in the claim's own tense and mood, and keeps scope words, stated numbers,\nconditions and a modal that is content. The extraction page shows the hedge beside the attribution.\n\n**Two calls on nine real passages** (`may`, `could`, `likely`, `tend to`, a percentage range). The first\nrun wrote *could prove to be a bottleneck* as *proves to be a bottleneck* and dropped the *would* of a\nhypothetical, so the instruction gained the tense and mood rule; the second wrote *will prove to be a\nbottleneck*, *would lead to a step-change*, *AIs will behave like human sociopaths*, kept *tend to* and\n*70% to 90%*, and read every hedge as stance. **One fault to watch:** *Some have even said compression may\nbe the way to AGI* came back as *Compression is the way to AGI*: the reported-speech frame went with the\nhedge, which puts other people's view in the author's mouth. The first run had kept *Some have stated\nthat*. Reported content is minted separately by its own ruling (2026-08-15), but the claim's wording\nmust not lose who said it; the first ten claims the pass writes should be read for this.\n\n## Cross-references\n\n- [core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md): the division\n  of labour (claims are cores, senses are peripheries) and the five kinds of difference.\n- [store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md): the storage law, the\n  occasion facts, and his question of 6 October on the author's mindscape.\n- [three-minds-and-the-leap.md](three-minds-and-the-leap.md): his remark of 19 October that the\n  context of a concept-sense-use matters and that a phrase is *\"a type of move in a language-game\"*,\n  ruled as AM9.\n- [claim-sameness-philosophical-readings.md](claim-sameness-philosophical-readings.md): sameness is\n  use-relative.\n- [contingency-and-convergence-across-levels.md](contingency-and-convergence-across-levels.md): the\n  levels lens the diagram's panel 4 resembles.\n- [the-tractatus-and-deliberus.md](the-tractatus-and-deliberus.md): why a fuller description never\n  completes.\n"}