{"path":"research/store-the-uses-read-the-meaning.md","content":"# Store the uses, read the meaning\n\n**Status (2026-10-06)**: Claude's reading of a third mail from the friend whose project is read in\n[what-a-record-linkage-project-teaches-deliberus.md](what-a-record-linkage-project-teaches-deliberus.md),\nmade at the founder's request: *\"Read his last email also and think through deeply.\"* **Proposals\nonly**, ledger Arc AA ([session26-open-rulings.md](session26-open-rulings.md)). Nothing is built.\nThe friend and his project are not named, and his own words are paraphrased: the correspondence is\nprivate.\n\n**What the mail holds, and how far each part was read** (all of it in full, about 390 lines):\n\n| Part | What it is |\n|---|---|\n| Ten lines in the friend's own words | A doubt about storing sentences at all |\n| An answer from his coding assistant on our concept layer | It says it read two of our documents in full |\n| A design for a graph of real things and their relations, computed from our claims | Written by the same assistant on a request the founder put to it. It lists ten of our documents read in full, and Graphiti's source and paper |\n\n**What was checked here.** Every claim the mail makes about this corpus that was looked up held:\nthe 49 of 90 concepts reading *bifurcated*, the five kinds of difference between senses, the line\nabout a coordinate and a bearer, the April sentence, five documents that mention entity linking.\nIts claims about Graphiti were checked against the copy of Graphiti's source on the keeper's\nmachine (§ 5). Nothing was run and nothing was sent.\n\n---\n\n## In plain language\n\n**The friend's doubt is right, and Deliberus has already ruled the answer.** A stored word does not\nkeep its meaning, because meaning is how a word is used and use moves. So what can be stored is the\nuse: who said what, where, when. Meaning is read from the uses. We ruled exactly this for one thing\non 2026-09-15 (*a horizon kind is read, never set*) and have applied it to four more since. The\nmail shows it is the storage law of the whole system.\n\n**Two things in the graph still break that law.** A word's senses are stored as a model's\nparaphrase. And a claim's wording is a machine's rendering of a passage, stored as if it were the\npassage.\n\n**Deliberus cannot throw the sentences away, and does not need to.** Disagreement is its content,\nso what was said has to be kept, with who said it. That costs almost nothing: all the claim wording\nin the graph is under 2 MB. What gets compressed is what is *shown* (a map of ten boxes) and what\nis *held together* (claims that say the same thing, uses of a word in one sense, mentions of one\nreal thing).\n\n**The graph of real things is the founder's own April idea, and this is its first design.** It is\nworth having, as a layer computed from the claims and never written by hand. It stands on the\nrecord of machine judgments, so nothing new is owed before that record exists. Four conditions are\nadded here (§ 4d).\n\n---\n\n## 1. The friend's doubt, taken as four claims\n\nParaphrased: things cannot be stored unchanged. If a database holds the word *wine*, that word\nkeeps becoming something else. That is language games. He does not want to store sentences. He\nwants determinism, and something compressed, where a store of propositions would be one enormous\ntangle. He distrusts vectors and sees no other way. And he wonders whether he has understood what\nDeliberus is about.\n\n**(a) Meaning is use, and use moves.** True, and it is this corpus's own position on sameness: two\nclaims are the same when they are used the same way\n([claim-sameness-philosophical-readings.md](claim-sameness-philosophical-readings.md)). What\nfollows is the rule in this document's title. A use does not change: this person wrote these words\nin this source on this date. What is computed from the uses can be computed again tomorrow, from\nmore of them.\n\n**(b) A store of sentences is a tangle.** Two different things could be stored, and they want\nopposite treatment:\n\n- **What was said.** This is the ledger. It has to be kept whole and attributed, because the\n  minority position and the unstated crux live in what a summary drops\n  ([saturation-and-the-long-tail.md](saturation-and-the-long-tail.md)). It is cheap: measured\n  today, all claim wording in the graph is 863,000 characters.\n- **What is held to be the case.** This is the chart. It should be compressed and computed.\n\nSo the tangle is a problem of reading, and reading is where Deliberus compresses: a claim page\nopens on about ten boxes. Storage keeps everything once.\n\n**(c) Determinism.** A model's reading of a contested sentence is not deterministic and cannot be\nmade so by decree. Everything after it can be. Once a model's answer is recorded, it is a fixed\nfact, and every reading computed from recorded uses and recorded answers gives the same result\neach time. That is what the record of machine judgments buys\n([storing-the-record-of-machine-judgments.md](storing-the-record-of-machine-judgments.md)): the one\nplace where chance enters is written down and frozen.\n\n**(d) Vectors.** Our position already matches his distrust: an embedding may find candidates and\nmay serve as the point a difference is measured from; it is never anyone's meaning and never\ndecides ([core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md) § 5b).\nThere is another way for half of the problem. Real things have registers (Wikidata, DOIs, ORCID),\nso two mentions of one thing can be joined by an identifier from outside. For contested words\nthere is no register, and the other way is attributed uses, stated differences, and whether\nconclusions diverge downstream.\n\n**What his own system does not have to do.** It digs one answer out of a finished past: who owns\nwhat. Here the disagreement is the content. He can compress many records to one identity. We have\nto compress without choosing a side, so a group keeps its members and its dissent.\n\n## 2. One law, and where Deliberus already follows it\n\nEvery stored thing is one of three:\n\n- a **use**: something somebody said or did. Kept, never changed;\n- a **machine judgment**: what a model answered about a use. Recorded once, never edited; a later\n  answer is added beside it;\n- a **reading**: something computed from the first two. Never stored as if it were either. Cached\n  at most.\n\nSorted that way, what the system holds today:\n\n| Thing | What it is | Stored as | State |\n|---|---|---|---|\n| A source's text, a quoted passage, a person's own words, a vote | use | itself | kept |\n| A horizon kind | reading | computed, dated | ruled 2026-09-15, built |\n| A claim's strength | reading | computed when read | built |\n| A critical question nobody has answered | reading | rendered when read | built 2026-09-21 |\n| An answer to a question put to the graph | reading | computed at the time of asking | built |\n| A map | reading | computed | built |\n| A link's strength | reading | still a stored rating from extraction | ruled 2026-09-28, build pending |\n| The `confirmed` flag, the trust tier | reading: both are fixed by who wrote the link | stored on 4,600 links | ruled to retire, build pending |\n| A claim's confidence | reading: looked up from two other fields | stored on 4,903 claims | not raised before |\n| An embedding | reading: computed from the wording and the embedding model | stored on 3,895 claims, about three quarters of the graph's memory | proposed, ledger Z13 |\n| Which model answered what | machine judgment | nowhere, or as a name on some items | proposed, ledger Z1 and Z11 |\n| A word's senses | machine judgment | a model's paraphrase, stored as the sense | § 3 |\n| A claim's wording, where a machine wrote it | machine judgment | stored as the claim itself | § 3b |\n\nSo the law has been ruled here five times, one thing at a time, and never stated once. The friend's\nassistant says the same in one line: the philosophy is right and the storage is being fitted to it\nafterwards.\n\n**The test this gives for any new field**: is it a use, a judgment or a reading? A reading is not\nstored as a fact. A judgment is not stored without its record.\n\n## 3. Identity is a group of members\n\nThe second thing the mail shows is one rule under three layers.\n\n| Layer | The group | Its members | What must never stand in for it |\n|---|---|---|---|\n| Claims that say the same | a same-claim group | the claims, each with its own wording and links | a canonical wording, a summary |\n| A word's sense | a sense | attributed uses: this person used the word this way in this claim | a label, a definition a model wrote |\n| A real thing | an entity | mentions | a name |\n\nIn each: membership can be undone, a member is admitted on evidence, and a candidate is compared\nwith the members and never with a label or an average. For claims this is ledger Z5. The mail\nstates it for senses and for entities.\n\n### 3a. Senses\n\nIn the friend's terms a sense with no uses behind it cannot exist. Ours can: the detector writes a\nparaphrase and the paraphrase is the sense. That is the cause of three things measured on\n2026-09-06: 49 of 90 concepts read *bifurcated* because the detector writes two senses per\ncontested word; who holds which sense is in the data model and never written; and what a reader\nsees is the model's paraphrase.\n\nOne storage rule replaces three open proposals of\n[core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md) § 6 (a sense\nleads with its attributed claims; a split counts only when it divides along sources or\ncommunities; attribution is stored): **a sense is a group of attributed uses, and its label is\ncomputed from them.**\n\nIt also gives the concept layer time for nothing. Uses are dated. The sense of a word as it stood\non any date is read from the uses up to that date, which is the history of a concept that\n[fractal-scales-and-temporal-frame.md](fractal-scales-and-temporal-frame.md) lists as missing.\n\nTwo cautions. The five kinds of difference between senses came from one method, so by this\ncorpus's own rule they need a value for *none of these*. And *are these two uses the same sense*\ncan be read from what each use leads to only where something has been built on them. Today that is\nrare, so the usual answer is *cannot tell*.\n\n### 3b. Claim wording\n\nA claim a machine extracted is a rendering: the machine read a passage and wrote a sentence a\nstranger could follow. The passage is the use. The sentence is a judgment. The graph stores the\nsentence as the claim, under an identifier made up at storing time.\n\nThe cost shows when a source is read again. Its machine-written claims are deleted and new ones\nare written under new identifiers. Whatever people had attached to the old ones is cut loose. Two\nstanding rules pull against each other there: *no stale claims in the graph*, and *contributions\nare added, never overwritten*.\n\nIf a claim were anchored in its passage, with its wording as a dated rendering, reading a source\nagain would add renderings. Nothing would be deleted and what people attached would stay.\n\n**This is a question for design, with a real difficulty in it.** One sentence can yield several\nclaims, and five models cut one sentence four ways. So matching a new rendering to an old claim is\nitself a sameness judgment, though an easy kind: same source, same passage.\n\n### 3c. Can the use be stored without paraphrasing?\n\nThe founder's question on reading the above (2026-10-06): *\"Can we store the use without\nparaphrasing? What would that mean?\"*\n\n**Yes. The use needs no paraphrase at all.** It is the author's words in their place, so storing it\nmeans storing a pointer: this source, from this character to that one. The words are already in the\nkept source text. Nothing a machine wrote is part of it.\n\n**What it cannot do is replace the rewritten sentence.** Measured on 2026-10-06 over the 1,144\nextracted claims whose source text was kept, by code alone\n(`uv run python scripts/claim_wording_sources.py`):\n\n| What | Measured |\n|---|---|\n| The stored quote is in the source character for character, in exactly one place | 923 claims, 81%. For these a pointer can be computed today |\n| It is there once typography is evened out or an ellipsis is followed | 94% (`scripts/measure_quote_fidelity.py`) |\n| It occurs in several places | 1 claim |\n| Length of the quoted passage, and of the claim | median 14 words, and 23 |\n| The claim is its passage word for word | 4 claims |\n| Claims that share a passage with another claim | 169, 15% |\n| A claim's content words that come from its passage | 34% on average |\n| From elsewhere in the same source | 37% |\n| **Found nowhere in the source** | **29% on average, median 20%** |\n\nSo a passage is usually a fragment. One stored passage reads *but both eventually hit ceilings*.\nNobody can read that alone, attack it, or compare it with a claim from another source. The\nrewritten sentence does three jobs the passage cannot: it makes a fragment readable by a stranger,\nit cuts one passage into separate claims, and it gives claims from different sources a form that\ncan be compared.\n\n**What changes is which of the two is the claim.**\n\n- **The passage is the author's.** It is stored as a pointer, it never changes, and it is the\n  claim's anchor.\n- **The rewritten sentence is ours.** It is a machine judgment about the passage: recorded, dated,\n  open to being done again, and open to being refused by its author.\n\n**Five things follow.**\n\n1. **Code can check part of a rewrite.** Every content word of a claim is from its passage, from\n   elsewhere in the same source, or from nowhere in the source. The third pile is the machine's own\n   wording, and it is where weight gets added. One stored claim says self-evaluation *is a\n   significant technical hurdle*. Its author wrote *a second issue*. The word *hurdle* is nowhere\n   in that source. The two faults of this kind found so far were each found by a person\n   ([the-evals-loop-and-our-machine-passes.md](the-evals-loop-and-our-machine-passes.md) § 3a);\n   this count says where a reader should look. It is not a verdict: a word used in another sense\n   elsewhere in the source counts as the author's.\n2. **A rule for rewrites follows**: write in the author's words, borrow context only from the same\n   source, and count what is left over. The two extractions made by hand average 5 and 6% of words\n   from outside the source; the machine's range from 9 to 42%. The splitter already has this rule\n   one level down (*in the parent's own words only*), adopted after its first version was measured\n   adding content.\n3. **Reading a source again adds a rewrite.** The passage stays, so the claim stays, and what\n   people attached to it stays (§ 3b).\n4. **The author rules on the rewrite and never on the claim.** Consent then has something to be\n   about: these are my words, and that is or is not what they say.\n5. **What the author did with the words belongs to the use where the author marked it.** Asserted,\n   conceded, supposed, reported, asked. Three measured faults are one fault here: parts that stated\n   as fact what their claim denied or conceded; a hypothesis stored as an observation, on 24 of 116\n   claims of one hand-made map whose author had typed his own labels; and reports stored without\n   the reporter's stance.\n\n**Four kinds of claim have no passage to point at.**\n\n| Kind | Live count | What it is |\n|---|---|---|\n| A person's own typed claim | 44 | Its own use. No rewrite, unless a machine rewords it |\n| An unstated premise | 34 | A judgment about a gap between two passages. Asserted by nobody until somebody says *yes, I hold that* |\n| The content of a report, minted as its own claim | 53 | Has a passage. Its asserter is the person reported, and the reporter's words are the use |\n| A critical question | 3,118 stored | Written from a template. Since 2026-09-21 a new one is rendered when read |\n\n**Two limits.** A pointer needs the source kept: 16 of the 33 stored sources have their text. And a\ntranslation is a rewrite from the first word to the last: the one Swedish source's claims are 97%\nwords from outside the source, which is correct, and its passages are still Swedish.\n\n**This corpus has said this three times and built a copy of the quote.** In March 2026 the first\npipeline run was read and the fix written down was that the first pass *should POINT, not\nPARAPHRASE*, giving *indices/anchors into the source text*\n([first-pipeline-run-analysis.md](first-pipeline-run-analysis.md)). What was built is a quoted\nstring on each claim, with no position. The field of argument mapping keeps what was said apart\nfrom what it was taken to mean, and joins the two by what the speaker did in saying it\n([argsbase-argument-web.md](argsbase-argument-web.md), where this corpus called it *critical for\nDeliberus*). And the group whose method the splitter follows keeps the pair, the part in the\nsource's words and its rewritten form ([the-jhu-decomposition-line.md](the-jhu-decomposition-line.md)).\n\n### 3d. What a use carries besides its words\n\nThe founder, the same hour: *\"In a way the original use is defined by the context of the source\ntext but also ofc of the author's mindscape, memescape and state of mind etc etc... So how do we\ntake into account this\"*\n\nHis three words are levels this corpus already has, from his own lens of 2026-09-06\n([contingency-and-convergence-across-levels.md](contingency-and-convergence-across-levels.md) § 3):\nthe community a person talks within (level 4), the one person's history (level 5), and the state\nof the moment, which re-weights the same parts each time a word is used (level 6).\n\n**What can be on record of each:**\n\n| What shaped the use | What can be kept | Kept today |\n|---|---|---|\n| The text around the passage | All of it | Yes, where the source is kept. A pointer leaves the passage inside its text; a copied quote cuts it out |\n| The occasion: when, where it was published, to whom, in reply to what, in what kind of text | Facts | When and where, yes. To whom, in reply to what, and the kind of text, no |\n| The community it was said from | Traces: whom the author cites, whose vocabulary they share, which senses of a word divide along which communities | Not as such. It is what a sense would be, as a group of attributed uses (§ 3a) |\n| The person's own history | Their other uses, dated, and what they have said about how their meaning differs | Their other claims are in the graph, joined to them by a name |\n| The state of mind at the time | Almost nothing. Tone and hesitation, where there is a recording. The author's own later words | A recording is kept for spoken sources only when someone keeps it |\n\n**Three rules follow from the table.**\n\n1. **Nothing is guessed into the use.** A machine's reading of what the author had in mind is a\n   judgment like any other: recorded, marked as a reading, open to refusal. This is the line\n   between a rewrite that resolves *both* to the two things named a sentence earlier, which the\n   text settles, and one that turns *a second issue* into *a significant technical hurdle*, which\n   reads a mind.\n2. **The author's mind enters the record only as more uses.** What they wrote elsewhere, how they\n   use the word in other places, and what they answer when shown a rewrite. That answer is a new\n   use, dated now. It is not a window onto then: a person's account of what they meant earlier is\n   a reading too, and gets no extra weight\n   ([what-human-judgment-is-for.md](what-human-judgment-is-for.md)). What cannot be wrong is their\n   refusal, today, to stand behind a rewrite.\n3. **A rewrite is made from somewhere too.** The reader has a community, a history and a state as\n   well, and a model has its training. So a rewrite carries who made it, and several rewrites of\n   one passage from different positions are a gain, where one canonical rewrite would hide the\n   position it was made from. For a model this is the record of machine judgments.\n\n**The limit, stated.** The mind is not in the record and will not be. Two things stand where it\nwould: the author's standing right to answer, and the meeting itself. Two people at one table read\neach other's state directly, which is why that setting is the one this corpus keeps choosing for\nits first real tests ([islands-of-coherence.md](islands-of-coherence.md) § 5c), and why\nunderstanding in the full sense is said here to happen in the encounter and not before it\n([claim-sameness-philosophical-readings.md](claim-sameness-philosophical-readings.md), on Gadamer).\n\n**Three facts are proposed for a use**, since they are facts and are missing: to whom it was said,\nin reply to what, and the kind of text. The first is already named as missing\n([the-motte-and-the-periphery.md](the-motte-and-the-periphery.md) § 3c), and the third has been\nmeasured to decide what kind of horizon a disagreement ends at\n([frontier-extraction-experiment.md](frontier-extraction-experiment.md)).\n\n## 4. A graph of real things, computed from the claims\n\n### 4a. Whose idea\n\nThe founder's, on 2026-04-05: a continually updated graph that *\"outputs a graph representing the\nobject level reality as Deliberus 'understands it right now'\"*\n([truth-graph-evidence-system.md](truth-graph-evidence-system.md) § The founder's April 5–6\nmessages). What was built instead answers a question at the time of asking, from the argument\ngraph. No design for the object-level graph exists in this corpus, and the ledger arc that\nrecovered that April session has no row for it. The mail holds the first.\n\n### 4b. The design in the mail\n\nThree layers, of which one is computed.\n\n| Layer | What it holds | How it comes to be |\n|---|---|---|\n| What each claim is about | The things it mentions, and one proposition: subject, relation, object or value, qualifiers, the time the content is about | A model proposes it. Recorded as a machine judgment, never edited |\n| Which mentions are one thing | Entities as groups of mentions | A merge needs evidence: an outside identifier, two mentions in one source that refer to each other, or a matching pattern of relations. A similar name only proposes |\n| The world links | *X relates to Y, during T* | Computed from the propositions that fall on the same link. Each link carries the claims for it, the claims against it, who asserts it, and a state |\n\n**An example, made up.** A claim reads: *a trial in one country in 2017 found that a basic income\ndid not reduce how much people worked*. The first layer notes two things (basic income, hours\nworked), one relation (*does not reduce*), a place and a year. The second layer decides that *basic\nincome* here is the same thing as *UBI* in another source. The third layer shows one link, *basic\nincome and hours worked*, with every claim that speaks to it, for and against, each still attached\nto its source and its argument.\n\nOnly claims about what is the case feed it. A report (*X says Y*) gives a link about Y, with X as\nits asserter. Words like *freedom* stay in the concept layer. Recomputing from the same claims\ngives the same result.\n\nIt takes two things from Graphiti and refuses three. Taken: four timestamps on a fact, and the\nrule that two facts whose periods do not overlap are both true. Refused: settling a contradiction\nby letting the newer fact win, merging two things because their names match, and rewriting a\nsummary in place.\n\n### 4c. What holds\n\n- **It is computed and never written by hand.** That is the horizon ruling one layer up.\n- **A claim is not a fact.** Fed to a memory system as it stands, a claim would lose its support,\n  its attacks, its kind and its asserter.\n- **Two opposite assertions made at one time are the product.** A rule that lets the newer one win\n  is a machine giving a verdict.\n- **Its first layer is the record of machine judgments.** The design says each proposal is kept\n  with its pass, its model and its instructions, and that a later pass adds a proposal beside it.\n  That is ledger Z1 and Z11. So the record is the ground this layer stands on, which is a second\n  reason to build the record first.\n- **Its test is the right one**: a full rebuild must equal what was kept up step by step, link for\n  link.\n\n### 4d. Four conditions added here\n\n**1. A proposition read from a claim is not a second kind of evidence about that claim.** It is a\nmodel reading the same sentence. By ledger Z2 that is the wording again. What is new is the\nregister behind a real thing: an identifier from outside. So this layer makes part of sameness\ncheap to decide, the part about named things, and leaves the rest a judgment. It should not be\ndescribed as two claims meeting by structure alone.\n\n**2. Scope is to place and people what a period is to time.** Graphiti's sound rule is that two\nfacts whose periods do not overlap are both true. The same holds for scope: a finding about one\ncountry's trial and a finding about another's do not contradict. Without scope in the comparison\nthis layer would manufacture contradictions. In the example above, a trial in another country that found the opposite does not contradict the first, any more than a trial thirty years later would. That is the mirror of the fourteen false support\nlinks, which joined look-alike claims about different things. Hedges belong here too: *reduces*\nand *is associated with less* are different propositions.\n\n**3. A world link carries no number of its own.** Three claims for and two against could be\nboiled down to one strength, a link that reads *73% true*. That would be a new aggregate over claims that may share one origin,\nand shared origin is not represented here yet (ledger Q4). The link shows its claims and their\nseparate strengths. Its state says what the record shows: *asserted and never challenged*,\n*challenged, and the challenges answered*, *standing claims on both sides*, *lapsed*. The mail\ncalls one state *established*, on the page where it notes that *truth graph* is a word of finality.\n\n**4. The list of relations is a taxonomy we would be inducing.** It gets what the list of argument\nschemes got: an open, curated set, a value for *does not fit* with a required note, and the\nexpectation that a large share will not fit at first. A relation earns its place if it makes two\npropositions comparable that were not.\n\nOne more thing follows from *computed, never written*: **a correction goes to the claim layer.**\nSomeone who thinks a world link is wrong disputes a claim under it, or the proposition read from a\nclaim, or the grouping of two mentions. Nobody edits the link. The day somebody can, this is a\nsecond store, which *one graph, staged maturity* rules out.\n\n### 4e. What it would cost\n\n| What | Cost |\n|---|---|\n| A proposition for each claim about what is the case | One model call per claim, once: under 2,000 calls for today's corpus. About 1 KB each in the record, so about 2 MB on disk |\n| Entities | Rows in a table |\n| The world links | Nothing stored. Computed, or cached |\n| Memory in the graph | None, as long as the layer is read from the record |\n\n### 4f. What it would move that is stuck\n\n- **Links across debates.** One measured pair of opposing texts gained eight resemblance links,\n  all between its own two texts and none to the other 23 sources, because similarity of wording\n  measures vocabulary ([dogfood-run-6-israel-palestine-cross-domain.md](dogfood-run-6-israel-palestine-cross-domain.md)).\n  Two claims about the same thing and relation, with values that cannot both hold, can be found by\n  looking them up.\n- **Claims that go stale at an event.** *The sitting president supports this* ends at an election.\n  With real things and dated links, a change of office is something the system can see.\n- **A reader that asks about a thing.** *What does Deliberus hold about X, and how contested is\n  it* is the question an agent would ask first.\n- **The friend's own doubt.** A proposition is the compressed form he asked for, and comparing two\n  of them is a lookup.\n\n### 4g. The smallest honest experiment\n\nAfter the record exists: read a proposition for each claim about what is the case, write nothing\nto the graph, and list the pairs across sources whose propositions cannot both hold. Count how\nmany the linker never proposed. A person reads them.\n\n**What would show the layer is not worth building**: most claims fit no relation in the list; or\nthe pairs found are look-alikes with different scope; or the list holds nothing the linker did not\nalready have.\n\n## 5. Graphiti, checked against the local copy\n\nThe keeper's machine has a copy of Graphiti's source from August 2026. The mail read the current\nversion.\n\n| The mail says | Against the local copy |\n|---|---|\n| When two facts contradict and their periods overlap, the one that became true earlier is ended; when either has no date, nothing is ended | **Holds**, line for line (`resolve_edge_contradictions`) |\n| Entity summaries are rewritten in place | **Holds** (the summary prompt says to update it) |\n| A new mention is merged with an existing entity of the same name without asking a model | **Not in the local copy.** There a model decides, shown names and types. Newer than our copy, or wrong |\n| The extraction prompt forbids abstract concepts | Not found in the local copy |\n\nSo the refusal of *the newer one wins* rests on code that was read here. The refusal of merging by\nname rests on the mail alone.\n\n## 6. Proposals (ledger Arc AA, unruled)\n\n| Ledger | Proposal |\n|---|---|\n| AA1 | ***Read, never set* is the storage law of the whole system.** Every stored thing is a use, a machine judgment or a reading, and a reading is never stored as a fact. The test applies to every new field. Stored readings today: a claim's confidence, embeddings, the trust tier, the `confirmed` flag |\n| AA2 | **Identity is a group of members, and a name, a label or a summary is computed from the group and never compared against.** One rule for same-claim groups, word senses and real things. For senses: a sense is a group of attributed, dated uses, and one with no use behind it cannot exist |\n| AA3 | **Should a claim be anchored in its passage, with its wording as a dated rewrite?** The passage can be stored as a pointer with no paraphrase, today for 81% of extracted claims; the rewritten sentence stays, as a judgment its author can refuse. Reading a source again would then add rewrites and delete nothing. A question for design: 15% of claims share a passage (§ 3b, § 3c) |\n| AA4 | **The graph of real things is a reading over the claims, built on the record of machine judgments, and never a second store.** With four conditions: a proposition is not a second kind of evidence; scope is compared like time; a world link carries no number and no state named as final; the list of relations is open, with a value for *does not fit*. A correction goes to the claim layer |\n| AA5 | **Its first experiment, after the record exists**: propositions for the claims about what is the case, nothing written to the graph, and a person reads the contradictions it finds across sources that the linker never proposed |\n| AA6 | **A rewrite is held to its author's words, and code counts what it adds.** Words found nowhere in the source are 29% of a claim on average, 5 to 6% in the extractions made by hand; they are where weight gets added (§ 3c) |\n| AA7 | **A use carries its occasion as facts, and nothing about the author's mind is guessed into it.** To whom it was said, in reply to what, and the kind of text are stored with it. A reading of what the author had in mind is a recorded judgment. The author's mind enters only as more uses, their answer to a rewrite included (§ 3d) |\n\n## 7. What would show this reading wrong\n\n- **If the law costs more than it saves.** Computing a reading each time it is wanted can be slow.\n  A cache is allowed; the test is only that it can be thrown away and rebuilt to the same result.\n- **If people need to write to the chart.** A reader who sees a wrong world link may want to fix\n  it there and then. If sending them to the claim underneath loses them, *a correction goes to the\n  claim layer* is wrong as an interface, though it may still be right as storage.\n- **If anchoring a claim in its passage cannot be done.** Where two readings of a source cut it so\n  differently that old and new claims cannot be matched, § 3b fails and the two standing rules stay\n  in tension.\n- **If the friend's assistant misread its own source.** Two of the four claims about Graphiti\n  could not be found in the copy read here.\n\n## Cross-references\n\n- [storing-the-record-of-machine-judgments.md](storing-the-record-of-machine-judgments.md): the\n  record every layer here stands on.\n- [what-a-record-linkage-project-teaches-deliberus.md](what-a-record-linkage-project-teaches-deliberus.md):\n  the first reading of the same friend's project.\n- [what-belongs-in-the-ontology.md](what-belongs-in-the-ontology.md) § 6c: *read, never set*, as\n  ruled for horizon kinds.\n- [core-periphery-and-the-concept-layer.md](core-periphery-and-the-concept-layer.md): senses,\n  differences from a shown reference, and what was measured of the concept layer.\n- [truth-graph-evidence-system.md](truth-graph-evidence-system.md): the April request and what was\n  built from it.\n- [supersession-and-bitemporal-lifecycles.md](supersession-and-bitemporal-lifecycles.md): stale,\n  superseded and refuted as three states, and the earlier reading of Graphiti's timestamps.\n- [the-harness-conviction.md](the-harness-conviction.md): the chart and the ledger.\n- [what-agent-memory-research-says-about-agent-surfaces.md](what-agent-memory-research-says-about-agent-surfaces.md): what research on memory for AI agents says about the agent\n  surfaces, with a measured failure behind the refusal of summaries rewritten in place.\n"}