{"path":"research/claim-or-system.md","content":"# Claim, record, reading or form: where a rule under a link belongs\n\n**Date**: 2026-10-10 (evening) · **Status**: a reading of what the corpus already holds on one\nquestion, a correction to what was said earlier the same day, and two proposals (ledger AG12,\nAG13). **AG12 was ruled and built the same evening** (§ 4 for his words, § 4a for what was built\nand what was not). AG13, the four questions of § 3, is open.\n\nThe founder's question, on being told that the rules a machine writes under a link are not in the\ngraph:\n\n> *\"Ok well it makes sense to store them as claims, right? Or what might belong as a claim and what\n> might belong as a feature of the ontology and software itself? This is an ever-present question in\n> Deliberus. Let's mine the corpus graph for everything that's relevant to this question.\"*\n\n**The short answer.** Yes, as claims, and the corpus had mostly said so already in pieces that had\nnot been put side by side. What a machine writes under a link has four kinds of thing in it, and\nonly one of them is a claim. The piece nobody had ruled when this was written is the standing of a\nclaim that a machine wrote and nobody asserts. He ruled it that evening: it is a claim from the\nmoment it is written.\n\n---\n\n## 1. A correction first: four things sit under a link today, not one\n\nEarlier the same day the founder was told, and three notes were made to say, that the one thing the\ngraph keeps under a link is a standing question. That was the answer to *what is kept as a node*.\nRead again from the code and the live graph, with the question put the right way:\n\n| What sits under a link | How it is kept | On the live graph | Can a person dispute it? |\n|---|---|---|---|\n| The linking step's own reason for drawing the link, a sentence or three | a text field on the arrow | 1,481 of the 1,636 support and attack links between claims from texts or people | no. A field has no address |\n| A machine's reading of how much the first claim bears on the second | a number on the arrow | 190 links | no |\n| A standing question about the link | a node that holds both claims' ids | 1,225 | yes, by taking up one of its two answer claims |\n| **A claim about the link** (how much the first bears on the second, argued) | an ordinary claim, whose id the arrow holds, joined by a link that passes no strength on | **none.** Built 2026-09-28 with the ruling that a link's effect is its supporter's strength times its relevance; only a signed-in person can make one | yes. It is attacked and supported like any claim, and the link's relevance follows its strength |\n| A claim that a part is required | the same, on a part-of link | none. Built 2026-09-17 | yes |\n\nSo two statements made earlier that evening were too strong:\n\n- *\"Nothing can be attached to a link.\"* Nothing can be attached to the arrow itself. **The arrow\n  can hold the id of a claim about it, and that is built.** It has never been used.\n- *\"A machine's sentence under a link is stored nowhere.\"* The written-out steps of the two trials\n  are not in the graph. They are in the record of machine judgments, which has kept every model\n  call since 6 October with the texts each was shown. And a different machine sentence, the\n  linking step's own reason, is under nine links in ten as a field.\n\nThat last row is the one that matters for the question. **The graph already holds a machine's\nsentence under nearly every link, in the one form nobody can point at.**\n\n*This section is the state before the ruling of § 4. Since it, the written-out steps can be in the\ngraph as claims, and the slot for a claim about a link can be filled by a machine (§ 4a). The\nlinking step's own reasons are still fields.*\n\n---\n\n## 2. What the corpus graph returned\n\nThe index of what notes say about each other was asked with the question as written, and then\nabout thirteen notes and ledger rows it named. What bears, with where each stands:\n\n| Source | What it says | Standing |\n|---|---|---|\n| The self-similar decomposition principle (`self-similar-decomposition-and-claim-ontology.md`) | Nothing is permanently atomic, **and this applies to the system's own representations**. A field holding free prose (a claim's text, a reason) is the part not yet turned into structure | decided, April 2026 |\n| The sorting test (`what-belongs-in-the-ontology.md`) | A commitment belongs in the graph if the system can hold two people disagreeing about it and keep working, and in the ontology if it cannot. Three answers when the test finds one baked in: make it a recorded setting, give it a way to say *does not fit*, or publish it as a claim | the corpus's working test since 2026-08-22. The third answer is under his hold of 2026-08-30 on putting the project's own beliefs into the graph |\n| Machine work is judged by addressability | Does the output give a reader more places to disagree, or fewer | approved 2026-09-21, a standing principle |\n| Always-mint for reported speech | A position a source reports is made a claim of its own, asserted by nobody here, joined by a link that passes no strength on, and every count says which claims are asserted | ruled 2026-08-15 |\n| A link's relevance is a claim (ledger Q33) | the slot in § 1 | ruled 2026-09-28 |\n| The toll bridge (ledger AD22, `philosophy-of-the-unwritten-step.md` § 3) | A rule is a licence. Written in as one more premise it needs another rule, without end. So what a step needs is drawn as the grounds of the link's relevance claim, never as further arrows into the conclusion | proposed |\n| The squirrel run on the strength code (`the-squirrel-and-the-strength-calculus.md`) | Those grounds have to be required parts of the relevance claim, since separate supports add, and writing more would raise a link | proposed |\n| A question is a potential, not a node (`latent-scaffolding-and-the-type-token-split.md`) | A standing question can be rendered from a template when read, so it need not be stored | endorsed, built 2026-09-21 |\n| A reader's finding on that note, 2026-10-08 | **A rule is not the same shape.** It is varying prose, so it cannot be rendered on read; it has to be a stored judgment first. And no question may be shown without both of its claims, which bars listing a rule apart from its links | a found connection |\n| Store the uses, read the meaning (ledger AA1) | Every stored thing is a use (somebody said it), a machine judgment (recorded once) or a reading (computed). A reading is never stored as a fact | proposed |\n| A reader's finding on that note, 2026-10-08 | For a rule: *the wording is a recorded judgment, one pointer on the link*. A rule's count across links, and what a step commits its author to, are readings | a found connection |\n| The machine's sentence is nobody's claim (ledger AD17, AG7) | Shown as a question with a move beside it, weighing nothing and read by no pass but the descent, until a text states it or a person takes it up | proposed (AG7 was Claude's, the same day) |\n| A rule is kept as every wording proposed for it (ledger AG4, AD25) | Rules in the plural on one link; a rule's count across links is its record | proposed |\n| The machine may and should write deeper steps (ledger AG6) | under quality control like the other small tasks | **ruled 2026-10-10** |\n| Depth is what compounds (ledger T1) | a measured error rate caps how deep a machine may build on its own work | ruled 2026-09-29 |\n| The splitter's rule | write what was found, hold what was added; a written part caps its whole as a person's does | ruled 2026-09-21 |\n| One graph, staged maturity | no separate universe of drafts. Rough material is told apart by stage and by who made it, not by living somewhere else | a standing product directive |\n| The chain written in scores lower (`the-chain-inside-a-link.md` § 4.6) | Run on the strength code, a link with its middle claim written in gives its conclusion 0.500 where the bare link gives 0.519 | measured 2026-10-10 |\n\nTwo of these had not been set beside the question before, and they decide most of it.\n\n**The ruling of the same day needs a node.** A deeper step is written under a sentence: under the\nrule, why the rule holds. A field cannot have anything under it. Neither can a row in the record.\nSo once the machine writes more than one level, the first level's sentences have to be things\nother things can hang from, and in this graph the only such thing is a claim. A second kind of\nnode for *the machine's sentences that are not claims* would be the separate universe of drafts\nthe product directive rules out.\n\n**Only a node can be shared.** The proposals want a rule to recur across links, with the count as\nits record. A field on one arrow cannot be the same field on another arrow. A claim can be pointed\nat from many.\n\n---\n\n## 3. The sort\n\nTake one written-out step: *the squirrel smells of decomposition, so it is dead*. Under it a\nmachine writes a fact about the case, a tie between the words of the two claims, and a rule. Each\npiece, and each thing recorded about the step, goes to one of four places.\n\n| The piece | Where it belongs | The test that puts it there |\n|---|---|---|\n| The rule (*what smells so is dead*) | **a claim** | two people can disagree about it; it recurs under other links; cases and mechanism hang under it |\n| The fact about the case (*this smell is the smell of decomposition*) | **a claim** | it is the one piece that is a missing premise in the old sense |\n| The tie between two wordings (*this smell* and *that smell* are one thing) | **a claim**, and the concept layer | sameness of words is disputed like anything else |\n| *This link rests on that rule* | **the link's relevance claim**, with the three pieces above as its required parts | it is a claim about the link, and the slot for one is built |\n| Which shape the step has (rule and case, sign, like case, cases to a rule) | **form, recorded per step** | two readers can file one step differently, so it is a setting on the step and never a fixed property. The list of shapes is form too, and needs a *does not fit* |\n| Who wrote it, which model, shown which texts | **the record** of machine judgments | nobody can disagree about it; it is what happened |\n| What each check said of it | **the record** | a check's verdict is a machine judgment. A person who disagrees adds a claim; the verdict is not edited |\n| How strong it is, how often the rule recurs, which wording is the narrowest | **a reading** | each can be computed again from the claims and the record |\n| That a link may have a relevance claim, that its grounds are required parts, how required parts combine | **form** | the system cannot hold two people using different rules of combination on one link. Where it can be made a recorded setting it should be, as was done for how parts combine |\n| Whether a sentence nobody asserts may move a strength | **form, and the open question** (§ 4) | it is a rule of the software about standing, not a statement about the world |\n\nAs a rule of thumb, in the order to ask:\n\n1. **Could two people disagree about it?** Then it is a claim.\n2. **Does it say who did what, and when?** Then it is the record.\n3. **Can it be worked out again from the first two?** Then it is a reading, and is not stored as a\n   fact.\n4. **What is left is the form.** It should be as small as it can be, be a recorded setting wherever\n   two people could want it different, and have a way to say that something does not fit it.\n\nThe first is the sorting test of August. The second and third are the storage law proposed on\n6 October. The fourth is the three answers of the August note. Put in one list they give the same\nanswer to this question that each gives alone, which is some reason to trust the list.\n\n---\n\n## 4. The piece that was open, and his ruling: a claim that nobody asserts\n\n**RULED 2026-10-10.** Put the choice below as one question (is a machine's sentence a claim the\nmoment it is written, or only once somebody takes it up), the founder answered:\n\n> *\"The moment it's written. Go ahead and make it happen.\"*\n\nSo the second of the two answers below holds, and the first is withdrawn. The reasoning is kept as\nit was written, since it says what each answer costs.\n\nIf the rule is a claim, whose is it? The corpus holds two answers, and the second is partly the\nfounder's own.\n\n**It is not a claim until somebody takes it up** (AD17, AG7). The sentence is the machine's\nquestion. It is kept in the record with a pointer on the link, shown as a question, and becomes a\nclaim when a text states it or a person adopts it. What this protects: no count is inflated, no\nstrength moves on a sentence nobody holds, and no pass builds on it by accident.\n\n**It is a claim from the moment it is written, and it is inert** (always-mint, staged maturity,\nand his instinct in the question above). It gets an id and a page. A person can attack it\ndirectly. One rule can sit under many links. The next level can hang from it. It is marked as\nmachine-written and asserted by nobody, it is held at a draft stage, it weighs nothing and no pass\nreads it until somebody takes it up.\n\nBoth keep the same protections. They differ in what the sentence *is* while it waits, and that\ndecides three practical things:\n\n| | Not a claim until taken up | A claim at once, inert |\n|---|---|---|\n| Where the second level hangs | nowhere in the graph: a second structure has to be built for it | under the claim |\n| A person who thinks the rule is wrong | must first adopt it, then attack it | attacks it |\n| A rule under forty links | forty pointers to rows in the record, compared by a pass | one claim, forty links to it |\n| What a count of claims means | unchanged | grows by claims nobody asserts, as it already has for reported positions and for answers to standing questions, so the count must say so |\n| What can go wrong | the machine's sentences live in a place no page or agent can reach, and are forgotten | a pass that forgets the mark treats a machine's sentence as somebody's |\n\n**Proposed (AG12), and ruled as above: the second.** The ruling of the same day (AG6) presupposes it, the decided principle\nof April points at it, and the mechanism for holding a claim back from public surfaces by its\nstage exists. The cost is real and is the one the corpus has named before: a guard that holds only\nbecause every reader remembers a filter. The proposal therefore carries its own test, one that\nreads the code for a reader that forgets, as is done for replaced part links.\n\nOne refinement the sort forces. AG7 said no pass but the descent reads a machine's sentence, the\nsameness check included. But a rule that recurs is found by asking whether two rules are the same.\n**So machine-written rules are compared with each other, and never merged with a claim a text or a\nperson made.** A match between a machine's rule and somebody's claim is exactly the event that\ngives the rule an asserter, and it is shown, not acted on.\n\n**One thing the proposal has to add to what is built.** A reader with none of this context, asked\nto compare the proposal with everything settled, found nothing reversed and could not tell one\nthing: whether this is the structure the two-factor ruling built or something beside it. It is\nthat structure. But today a relevance claim is made by a signed-in person and stored as a person's\nclaim. Where nobody has argued a link's relevance, which is every link today, the relevance claim\nthe machine's sentences hang under would itself be machine-made, and would carry the same mark.\n\n**What is not proposed.** That a machine's rule should weigh. On today's code, writing a step in\nmakes a link weaker (§ 2, last row), and whether a sound written-out step should earn anything is\nan open question of the strength rulings, not of storage. And nothing here asks for the form\nitself to be published as claims in the graph: that is under his hold.\n\n---\n\n## 4a. What was built on the ruling, and what was not\n\nBuilt the same evening, in `deliberus/machine_sentences.py`. Read the module for the exact fields;\nthis is the shape.\n\n**What a machine's sentence is in the graph.** An ordinary claim node with three marks: its kind\n(`machine_step`), its stage (`draft`, which every public listing holds back) and its author (an AI\ncontributor). It also says which kind of sentence it is (a rule, a fact about the case, a tie\nbetween two wordings, a wider sentence, a stage of a mechanism), which pass and model wrote it,\nand where its stored answer is kept.\n\n**Where it hangs.** Under a link, the sentences are parts of the link's relevance claim: the claim\nabout the link whose id the arrow holds (§ 1, fourth row). Where nobody has argued the link's\nrelevance, a relevance claim is made for the purpose, and it carries the same marks. Each part\nlink says it is a machine's (`provenance: machine_step`), which sentence kind it joins, and its\nplace in the step.\n\n```\n   Claim A ═══ SUPPORTS (relevance_claim: id of R) ═══► Claim B\n                              ┆\n   R  \"what A's bearing on B rests on\"     kind machine_step · stage draft\n   │                                       (or a person's claim, if one argued the link)\n   ├─ part ─► the names it ties together   ┐\n   ├─ part ─► the fact about the case      ├ each a claim of kind machine_step\n   └─ part ─► the rule                     ┘\n                  ┆\n                  └┄ a wider sentence, or the stages of how it comes about:\n                     claims that remember this rule (written_under), joined to it by no link\n```\n\n**One wording is one claim.** A sentence with the same words as an earlier machine sentence, with\ncase, spacing and punctuation set aside, is the same node, so one rule sits under every link it was\nwritten under and the count of those links can be read off it. This is sameness of wording only.\n\n**It moves no strength, and that is held by a mark, not by an accident.** The strength code leaves\na part link out when it is a machine's and nobody has argued that the part is required. Without\nthat clause, one reader's vote on a machine's part would have turned the relevance of a link a\nmachine had guessed from nothing into a half. A person who attacks the machine's relevance claim does lower the link, by\nthe attacker's own standing: that is a person's contribution and it counts.\n\n**A person's argument takes the place.** When a person argues a link's relevance and a machine's\nrelevance claim is in the slot, the person's claim takes the slot and the machine's parts hang\nunder the person's claim.\n\n**Readers of claims decide about the kind.** The passes that pick claims out of the graph (the\nsameness sitting, splitting, titles, embeddings, the map, the rarest-claim place, coherence, the\nhorizon and hinge readings, the listing and the sitemap) leave the kind out, most of them from one\nshared list. The linking step finds its candidates by embedding, and a machine's sentence has none. `tests/test_machine_sentence_readers.py` reads the code for a query that can return a\nclaim it did not name and has nothing that rules a machine's sentence out, and fails until the new\nreader either leaves the kind out or is listed with the reason it reads every claim.\n\n**Counts say so.** `/graph/stats` gives the claims by who stands behind them, with a machine's\nsentences under their own name. That closes, for that route, the old gap that the function which\nsplits a count had no caller.\n\n**What exists to store today.** No pass in the pipeline writes steps yet. The only sentences are\nthe stored answers of the two trials, and `scripts/write_machine_sentences.py` makes claims of\nthem (it says what it would write unless told to apply, and it can take them out again). Run it\nagainst the live graph only after the site runs the code that knows the kind. How many are in a\ngraph: `uv run python scripts/graph_facts.py --shape`, under the kinds of claim.\n\n**Calls made in the build that he did not rule**, each Claude's and each reversible:\n\n- *The second level is joined to its rule by no link.* A sentence written under a machine's rule\n  remembers the rule in a field. Which kind of link belongs there is the open question of ledger\n  AG1 to AG5, and drawing a part-of or a support link now would decide it by accident.\n- *The linking step's own reasons stay fields.* The 1,481 reasons of § 1 were not made claims. They\n  are a different kind of sentence (§ 6), and nobody has read them as a set.\n- *A machine's sentence gets no embedding*, so no search by meaning returns it and the linking\n  step never meets it.\n- *The route that argues a link's relevance hands the slot to the person* (above).\n\n**Not built, and open.** Comparing machine-written rules with each other by meaning, and showing a\nmatch with a claim somebody made (the refinement above). A way for a person to take a rule up as\ntheir own. A page that shows a link's written-out step from the graph (the step review still reads\nits file). A pass that writes steps as part of the pipeline.\n\n**What has to be settled before such a pass runs: volume.** A reader of the notes found the piece\nthis note had left out (`drowning-in-claims.md` § 2): the graph has drowned itself once already,\ntwo thirds of its claims being possible answers nobody has given, and the flood there was machine\nminting. This ruling opens a second source of the same kind. By the trials' own figures a\nwritten-out step is 3.5 sentences and a relevance claim, and what is written under a rule is 2.7\nmore. Carried to all 1,569 live support and attack links that is about 7,000 claims at the first\nlevel and about 4,300 at the second, against 6,417 claims of every kind in the graph when counted,\nof which a reader sees 2,301. Three things hold that today: nothing writes steps but two stored\ntrials (about two hundred claims); one wording is one claim; and the kind is left out of every\nlisting and count of what somebody asserts. None of them is a limit. The August note names the\ngovernors that would be one: a gate on what is written, a spend gate, and a pause when too many\nproposals are refused. His ruling of 25 August has the machine work a descent out eagerly and show\nit on contest, with a stop rule; measured on 2026-10-10 that stop rule is a weak brake on fresh\nmaterial (`quality-control-for-deeper-steps.md`). Which of these a writing pass runs under is open, and\nhis.\n\n---\n\n## 5. What this changes in the day's other notes\n\n- `quality-control-for-deeper-steps.md` § 4.2 and `docs/architecture-reference.md` § Graph Database\n  said the standing question is the only thing kept under a link, and that a machine's sentence is\n  stored nowhere. Both now say what § 1 says.\n- Ledger AG7 keeps its protections and loses one sentence, now that AG12 is ruled: *it is never\n  filed as a claim*.\n- The three places a machine's sentence could be kept, listed in § 4.2 earlier that evening, were one\n  place seen three ways: a claim (§ 3), under the link's relevance claim (the slot), shared across\n  links (because it is a node).\n\n## 6. What was not checked\n\n- The route that makes a relevance claim has never been used by anyone, on the live site or in a\n  trial. It is tested. Whether its wording works for a reader is unknown.\n- The 1,481 reasons on arrows were counted and three were read. Nobody has read them as a set, and\n  they are a different kind of sentence from a written-out step: the linking step's account of\n  why it drew the link, in its own words, with no form asked of it.\n- The index returned what readers have found so far. Readers have been through a part of the notes,\n  so a note that bears and that no reader has opened would not have come back.\n\n## Cross-references\n\n`what-belongs-in-the-ontology.md` · `store-the-uses-read-the-meaning.md` ·\n`storing-the-record-of-machine-judgments.md` · `philosophy-of-the-unwritten-step.md` § 3 ·\n`the-squirrel-and-the-strength-calculus.md` · `latent-scaffolding-and-the-type-token-split.md` ·\n`self-similar-decomposition-and-claim-ontology.md` · `relevance-and-the-two-factors.md` ·\n`the-chain-inside-a-link.md` · `quality-control-for-deeper-steps.md` § 4.2 ·\n`docs/architecture-reference.md` § Graph Database · ledger AA1, AD17, AD22, AD25, AG4, AG6, AG7,\nAG12, AG13, Q33, T1 · `session49-what-goes-under-a-link.md`\n"}