{"path":"research/the-chain-inside-a-link.md","content":"# The chain inside a link: how small a step must be, and when the machine writes, widens and narrows\n\n**Date**: 2026-10-10 · **Type**: a founder demand, answered from his earlier rulings, the notes of\n7 and 8 October, the code, three counts on the live graph and three outside readings ·\n**Status**: the rulings quoted are his. Every reading and proposal is Claude's and **unruled**\n(ledger Arc AG), **except one that he ruled the same day**: the machine may and should write\ndeeper steps, under the quality control of the other small tasks (ledger AG6). This note said the\nopposite in § 4.3 and § 5.1, and both are corrected below. What that quality control is:\n[quality-control-for-deeper-steps.md](quality-control-for-deeper-steps.md). Nothing here is built.\n\n**His words**, 2026-10-10, on the first proposal of\n[alien-proofs-and-the-library.md](alien-proofs-and-the-library.md): *\"all the dark matter of the\nintermediate chain of reasoning/causality must be elucidated by the Deliberus graph somehow. We\nmust not make too-large leaps/steps, that would be irresponsible for any scientist. We must examine\nand make it possible to review each step of the reasoning tree/chain/graph.\"* And: *\"We also must\ndecide WHEN the machine writes its guesses and when it widens and narrows and based on what (the\nsubgraph? which subgraph exactly, connected to which exact claims/links?)\"*. And: *\"Use this new\ncorpus graph thingy I just built in a separate session.\"*\n\n**How it was read.** The corpus graph was asked about nineteen ledger rows and seven notes\n(`scripts/corpus_graph.py around …`). It returned the three readers' reports of 8 October, which\nwere then read in full, and they pointed at older passages that were opened in turn (§ 9).\nThree counts were taken on the live graph with no model\n([the script](the-chain-inside-a-link/what_is_under_a_link.py),\n[what it printed](the-chain-inside-a-link/what-the-count-printed.txt)). Three outside sources were\nread from their own texts as search extracts and are given in other words.\n\n---\n\n## 0. In short\n\n1. **He has asked for this four times, and one of the four is a ruling.** On 25 August he ruled\n   that the machine works out every claim's descent eagerly and shows it on contest. The proposals\n   of 7 October narrowed that to *the first level eagerly, the rest on request*, and did not say\n   so. § 1.\n2. **Today 1 link of 1,569 has anything under it.** All 822 links the linking step drew between\n   texts are bare: no pattern, no question, no reading of their bearing. § 2.\n3. **A leap cannot be measured by counting the steps inside it, because that count has no place to\n   stop.** It can be caught four other ways, two of them by code. § 4.1.\n4. **So a leap may be left, and may not be left unmarked.** Every link says what has been found,\n   written or examined under it, and the share of bare links is a published figure. § 4.2.\n5. **The way down is finding first.** At every level the machine looks before it writes: on the\n   map, in the two texts, in outside sources on both sides. Where it finds nothing it writes the\n   level in its own words, and that sentence is a question on the link, never a finding.\n   ⚠ *This point first said the machine writes one level and below it only finds and asks. He\n   ruled otherwise the same day (AG6): it may and should write deeper steps, and the small tasks\n   that write them are quality-controlled like the others.* § 4.3.\n6. **The place for mechanism evidence already ships, on 6 links of 164.** One of the four causal\n   patterns has a question about mechanism. All four should. § 4.4.\n7. **Two dials must be kept apart**: how finely a step is written out, and how many cases its rule\n   claims. His first demand is the first dial and his second is the second. § 3.\n8. **When the machine writes, widens and narrows is a set of conditions on the graph, each with a\n   stated reading set, and each write points at the record of what it read.** § 5.\n9. **A rule is kept as every wording proposed for it.** Cases from texts and people push the\n   narrowest standing wording wider. Counter-cases from texts and people push the widest standing\n   wording narrower. The order they arrive in does not matter. § 5.3.\n\n---\n\n## 1. What he has already said, and what was ruled\n\n| When | His words | State |\n|---|---|---|\n| Session 9, spring 2026 | *\"recursively decomposed (eventually, not instantly tho) until we aren't muddling properly separable aspects of reality that can and should be judged separately\"* | **ruled** 2026-08-31 (ledger B2, *\"Both blessed.\"*): the floor is separability, and contest is a way to spend a budget. *Eventually* is about pace ([decomposition-axes.md](decomposition-axes.md) § 4) |\n| 2026-08-19 | *\"I also believe in simply letting the system/LLM-prompting act like a five year old and ask 'Why?' recursively all the way down to philosophical/phenomenological bedrock, for each branch in the deliberation graph.\"* | his position ([taxonomy-gaps-and-the-closed-enum.md](taxonomy-gaps-and-the-closed-enum.md)) |\n| 2026-08-25 | *\"I want Deliberus to automatically recursively decompose all claims in each mapping-break. It should find as much research of high quality as possible and feed everything into the graph.\"* And: *\"All unstated premises need to be clarified and scaffolded.\"* | **direction ruled** in [kartpaus-auto-enrichment-design.md](kartpaus-auto-enrichment-design.md): do eagerly, show on contest; surfaces present questions, never finished analysis; research on both sides; stop where nothing more could change what a reader sees. Unbuilt |\n| 2026-09-29 | *\"Yes.\"* to: review need not keep pace with a proven machine pass; depth is what compounds | **ruled** (ledger T1, a principle line): machine work built on machine work is capped by measured error |\n| 2026-10-07 | *\"I want to have all the implicit reasoning structure and premises be explicated and exposed and scaffolded automagically. To the extent that process can be trusted of course. This depends if we can split it up into microtasks that we can rely on deeply.\"* | his request; answered by proposals AD13 to AD21, unruled |\n| 2026-10-10 | the words at the top of this note | his demand |\n\n**What the October proposals did to the August ruling.** AD19 says depth comes from taking a rule\nin its turn, when it recurs or is disputed, and that a chain is never written to the bottom in one\ngo. AD29 says any link can be written out on request, the first level eagerly. Both were written\nwith the ruling of 29 September in view and without the ruling of 25 August beside it. That ruling\nhas the machine explore the whole descent eagerly, fed by research, into a stage the default map\ndoes not show. The two meet in § 4.3: eager and recursive stands, and what the September ruling\ncaps is how deep the machine may build on its own sentences, by a measure of its error that does\nnot exist yet (and by his ruling of 10 October, AG6, the lack of that measure does not hold the\nwriting back). Recurrence and dispute are an order of work. They were written as if they were the\nstandard.\n\n---\n\n## 2. What is under a link today\n\nCounted on the live graph, 2026-10-10, for every live support or attack link between claims taken\nfrom texts or written by people.\n\n| | Links |\n|---|---|\n| All | 1,569 |\n| Written by the linking step, between two texts | 822 |\n| Written at extraction, inside one text | 667 |\n| Carrying a pattern of argument, which is what gives a link standing questions | 639 (none of the 822) |\n| With a relevance claim (a person's argument about how much the link bears) | 0 |\n| With a machine's reading of its bearing | 0 |\n| With a standing question whose answer anybody has touched | 1 |\n| **With anything at all under it** | **1** |\n| Support links whose middle is already on the map, as a path of two support links | 104 (74 of them drawn by the linking step) |\n| Causal and sign links | 164 |\n| ... whose pattern has a question about mechanism | 6 |\n\n**Read plainly.** The arrows are bare. A link the linking step drew moves no claim's strength,\nsince a machine's guess is presumed to bear nothing until it is read or argued\n(`deliberus/relevance.py`). On a map it is drawn like any other link.\n\n*Added the same evening, from a reader's finding checked in the code:* the machine reader of a\nlink's bearing that is built (`deliberus/relevance_judge.py`) selects only links that carry a\npattern and are not marked unconfirmed, and the linking step's links carry no pattern and are\nmarked unconfirmed. So that reader passes over all 822, and today the only way one of them comes to\ncount is a person arguing its bearing.\n\n---\n\n## 3. Two dials that must be kept apart\n\n**How finely a step is written out, and how many cases its rule claims, are different things.**\nThe philosophers of mechanism say so about scientific explanations: how much detail a description\nof a mechanism includes is one matter, and the range of cases it covers is another (Machamer,\nDarden and Craver 2000). A description can be very detailed and narrow, or coarse and wide.\n\n| Dial | His demand | What turns it | § |\n|---|---|---|---|\n| Detail: how finely the step from one claim to the next is written | *no too-large leaps; review each step* | finding and writing what lies between | 4 |\n| Reach: how many cases the general sentence under a step claims | *when it widens and narrows, and based on what* | cases, counter-cases and mechanism | 5 |\n\nThe first proposal of this morning ([alien-proofs-and-the-library.md](alien-proofs-and-the-library.md)\n§ 4.6) was about reach only. His answer was about detail. Mechanism joins them, because it does\nwork on both dials: it is the fine-grained chain, and it says how far a rule can be carried (§ 4.4).\n\n---\n\n## 4. Detail: no leap left unmarked\n\n### 4.1 What a too-large leap is, in terms that can be checked\n\n**Counting the steps inside a link does not work.** On 7 October a model asked what else must be\ntrue listed two or three further claims on nearly every link, the one-step links included, and an\noptional middle was filled in on 27 of 45 links. How many claims one finds between two others\ndepends on how finely one cuts ([how-far-a-premise-is-from-its-claim.md](how-far-a-premise-is-from-its-claim.md)\n§ 3a). A grade of distance on every link was considered and dropped: the model and the hand\nreading agreed on 5 of 30.\n\n**Four things can be checked instead.**\n\n| Sign of a leap | Checked by | Measured |\n|---|---|---|\n| The map already holds a claim between the two | code: a path of two support links between the same claims | 104 links today |\n| The two claims come from texts of different debates | code, once a source records its debate (AD3, open) | on 7 October no such link read as one step (0 of 9) |\n| Written out, the step does not carry: a reader who is not told a link was drawn cannot get from the written sentences to the second claim, or gets there without the first claim | the blind reader of 7 October | of 30: 6 could not be written out, 5 were written around the link, 2 unsettled |\n| A causal or sign link has no answer to *how* or to *what else* | the link's own standing questions | 0 answered on 164 |\n\n**A leap is too large when one of these holds.** Otherwise the link stands as one step for now,\nwith its questions open under it.\n\n### 4.2 A leap may be left. It may not be left unmarked, and the machine draws none bare\n\n**What the 25 August ruling keeps hidden is the machine's analysis. It does not ask the map to hide\nthat an arrow is a guess.** Untouched machine structure shown as what it is has been the blessed\ndefault since 2026-08-31. So:\n\n- **Every link shows its state**: nothing under it; something found; something written by a\n  machine; something a person examined.\n- **A link a machine drew with nothing under it is shown as a proposal**, in a different line from\n  a link a text states. It still exists, and it still weighs nothing.\n- **A support link whose middle is on the map shows that middle.** No model is needed, and it is\n  the cheapest first step: 104 links.\n- **The share of links with nothing under them is a published figure**, beside the count of links.\n  Today it is 1,568 of 1,569.\n\n### 4.3 The way down is finding\n\n**Each level under a link is filled the cheapest sound way first.**\n\n| Way | What it is | Who | What it may weigh |\n|---|---|---|---|\n| Found on the map | a claim that already sits between the two; a claim of either text that states a sentence of the step | code, and a model that points while code checks the claim exists | what any link between claims from texts weighs |\n| Found in a source | for a causal or sign step, what explains it and what rivals it, fetched on both sides, each as a located passage | the research step of 25 August, unbuilt | as a claim said by its author, marked as fetched by a machine |\n| Made | the machine's own sentence: what ties the case to a rule, and the rule | the writer of 7 October | nothing, until a text states it or a person takes it up (AD17) |\n| Asked | where nothing is found and nothing could be written that passes its checks: the question itself, with a move beside it | nobody yet | nothing; it is an invitation |\n\n**What compounds is the third row.** The ruling of 29 September is that machine work built on\nmachine work compounds error. A found sentence stands as far from its text as the reading that\ntook the claim out of it and the pointing that found it, however deep in a chain it sits. A made\nsentence under a made sentence adds a step each time.\n\n⚠ **Corrected 2026-10-10, on his ruling of that day (ledger AG6).** This paragraph went on: *the\nmachine makes a further sentence on top of one of its own only as deep as its measured error\nallows, and since no error rate has been measured, today that is one made level.* Asked whether\nthe machine may write deeper steps itself or should only find and ask, he answered: *\"It may and\nshould write deeper steps. But the microtasks in charge of doing so need to be quality-controlled,\njust like the other microtasks that are in play in Deliberus.\"* So an unmeasured rate does not bar\ndeeper writing. It obliges the measuring, and until there is a measure every made sentence carries\nhow many made sentences it stands on, weighs nothing and is read by nothing but the descent itself.\nThe controls, level by level, and a first trial of the second level:\n[quality-control-for-deeper-steps.md](quality-control-for-deeper-steps.md).\n\n**Recursion needs nothing new.** A found claim is an ordinary claim. The links into it carry\npatterns, the patterns carry questions, and each of those can be answered in the same four ways.\nThis is the self-similar principle of April, and it is why the same act serves at every level.\n\n**Where a chain ends.** In one of three things, and none is set by the machine:\n\n- a located contact with the world: a passage in an identified source, a measurement, an\n  observation, an utterance. It can be re-examined and it is not a smaller claim (ledger Q2, open);\n- a resting place that is read from how the claims currently stand, with its date (ruled\n  2026-09-15: read, never set);\n- a marked gap with its question.\n\nAD29 said a chain *ends in a named kind of stop*. The readers of 8 October found that this meets\nthe ruling of 15 September, and the wording here replaces it.\n\n**His squirrel, by hand, as an illustration.** No source was opened for the biology.\n\n- The link: *the squirrel smells of decomposition*, so *the squirrel is dead*. A step from a sign.\n- Under it, in his own sentence of 2011: *dead squirrels smell of decomposition*. And the question\n  a sign leaves open: what else would make that smell here.\n- Under his sentence, the chain a biology text would supply: the animal dies; its defences stop;\n  bacteria spread and break the tissue down; what they release is what is smelled. Four causal\n  steps, each a claim somebody can point at, each with a source once one is fetched.\n- Under the open question: a live squirrel that has rolled in something rotting; food rotting\n  nearby. Each is a rival, and the chain above is what says which rivals are possible.\n\n### 4.4 Where mechanism evidence goes\n\n**It goes under the question *how*, and that question exists on one pattern of four.**\n\n| Pattern | Its standing questions today | Links |\n|---|---|---|\n| cause to effect | is the generalization well supported; is the cause present; could something prevent the effect | 113 |\n| sign | is the sign reliable; could something else explain it | 37 |\n| correlation to cause | confounders; order in time; dose and response | 8 |\n| effect to cause | other causes; base rate; **is there a plausible mechanism** | 6 |\n\nAll but one of these questions (whether the cause is present) are sorted in code as bearing on the\nlink's connection, which is the second factor of the ruling of 28 September. So an answer to *how*\nmoves how much the link bears.\n\n**Proposed: every causal and sign link carries the question how.** Its answer is the chain of § 4.3.\nThe evidence note already proposes mechanism evidence as its own stream beside evidence from cases\n(ledger Q8, open), and this is where that stream lands.\n\n**A mechanism a machine writes is a guess at how it could work.** The philosophers of mechanism\nseparate a sketch, which has gaps where nobody yet knows what happens, from a filled description,\nand a description of how something could work from one of how it does (Machamer, Darden and Craver\n2000; the Stanford Encyclopedia's entry on mechanisms). Words like *causes* and *leads to* often\nhold the place of a stage nobody has described. So a made chain is shown as questions, and only a\nfound chain counts as evidence.\n\n**Mechanism also sets reach.** A rule holds as far as its mechanism operates. That is what the\nevidence note's question about carrying a finding to a new case asks: is the causal structure the\nsame there, and are the supporting factors present (ledger Q11, open).\n\n### 4.5 Shown so that each step can be reviewed\n\n**Levels, opened one at a time.** Lamport's method for writing proofs is the nearest practice\n(*How to Write a 21st Century Proof*, 2012). Every step has its own justification one level down.\nA reader opens as many levels as they need and no more. He reports that detail added at lower\nlevels makes a proof easier to read where the same detail in running prose makes it harder, that a\nlevel reads best at about four to ten steps, and that the way to find the gaps is to imagine a\nchild beside you asking why at every assertion. That is the founder's sentence of 19 August.\n\nWhat that means here:\n\n- **Each sentence under a link is a question with a move beside it**: dispute it; it is stated\n  here; why (AD17, and the principle that no gap is shown without a move).\n- **A level is kept small by the reader and never by the writer.** On 7 October the writer wrote\n  3.6 added sentences a step where the blind reader needed 1.2. What the reader does not need is\n  left out.\n- **A sentence under one link does not lean silently on a sentence under another.** It points, or\n  it stands alone. Lamport's rule that a step may use only what stands above it is the same guard.\n- **A demand to open a step carries its reason.** A bare demand repeated without end is a way of\n  stalling (ledger Q2, open).\n\n### 4.6 What the arithmetic must not do\n\n**Today it pays for the leap, on a link a text or a person drew.** Run through the strength code\non 8 October, a bare link gave its conclusion 0.519, and a link with its middle claim written in as\na chain of two arrows gave 0.500\n([the-squirrel-and-the-strength-calculus.md](the-squirrel-and-the-strength-calculus.md) § 3). If\nwriting the middle lowers a conclusion, nobody will write it. *(Scoped the same evening: 0.519 is\nfor a link presumed to bear half, which is what a text's or a person's link is presumed. A link\nthe machine drew is presumed to bear nothing, so nothing written under it can lower it.)*\n\n**That is why what is written under a link sits under the link's relevance claim and does not\nreplace the arrow** (AD22, and § 5 of that note). The direct link stays. Whether an examined\ninterior may also raise a link is the open question of the lift, and it is his\n([weakest-link-arithmetic-and-the-merge-hunch.md](weakest-link-arithmetic-and-the-merge-hunch.md)\n§ 9d). The test to hold any answer to was written on 8 October: the same weight however many\nsentences a step is written in.\n\n---\n\n## 5. Reach: when the machine writes, widens and narrows, and on what\n\n### 5.1 One rule for when\n\n**Eagerly, into the draft stage. Shown on contest, as questions. Made depth counted on every\nsentence, and capped by measured error once there is a measure; found depth not. Stopped where\nnothing more could change what a reader sees.** Each clause is an earlier ruling or proposal:\n25 August, 25 August, 29 September with his ruling of 10 October (AG6), 25 August.\n\n⚠ *2026-10-10: the third clause first read \"made depth capped by measured error\", which with no\nmeasure meant one level. And the last clause is weaker than it looks: measured that evening, the\nstop as built prices 595 of 1,198 open questions on unexamined claims as worth asking, and it has\nnothing to read under the 930 links that carry no pattern\n([quality-control-for-deeper-steps.md](quality-control-for-deeper-steps.md) § 4.7).*\n\n### 5.2 The acts, and exactly what each reads\n\n**Two reading sets, named so the table can be short.**\n\n- **The link's neighbourhood**: the two claims and the passages they were taken from; the link's\n  kind, pattern and writer; every other live link into the second claim (its other reasons and its\n  objections); whatever already sits under the link.\n- **The rule's record**: every wording proposed for the rule; every link resting on it, with who\n  drew it; the cases filed under it (84 links of the pattern *example* filed a case under a general\n  claim when counted on 8 October); the claims that say why it holds; the cases that break it and the exceptions.\n\n| Act | Fires when | Reads | Writes |\n|---|---|---|---|\n| Show the middle on the map | a support link, and a claim with support links from its first claim and to its second | the three claims and three links | pointers under the link, marked found by code |\n| Write the first level | a link with nothing under it. Order: links the machine drew between texts; then links inside a text; a person's own link only after the person | the link's neighbourhood; the claims of the two texts; the nearest stored rule wordings | the step in the form its shape asks for, or *cannot be written*; a pointer where a text states a sentence |\n| Answer a text's own questions | a link written at extraction whose pattern's questions have no answer | the link, the templates, the passage around both claims | tentative answers that point at the source's own sentences |\n| Fetch what explains and what rivals | a causal or sign link with no answer to *how* or *what else*, where an answer could change what a reader sees | the link, the question, the two claims | located claims from outside sources, on both sides |\n| Go one level down | a sentence under a link is disputed with a reason, or recurs under three or more links, or the stopping rule picks it | the sentence alone, with what is found for it; never how the sentence was written or checked | found first; where nothing is found, made in the form its kind of ground asks for, counted and checked ([quality-control-for-deeper-steps.md](quality-control-for-deeper-steps.md) § 4) |\n| Propose a rule over a group | two or more links of one shape, from at least two authors, whose needs a reader judges alike in a stated respect | those links and their steps; stored rule wordings | one wording as a guess, with the respect of likeness named |\n| Widen | a link a text or a person drew needs the rule, and the narrowest standing wording does not cover its case | the rule's record and the new link | a wider wording, added; the earlier one stays |\n| Narrow | a case that breaks the widest standing wording is found in a text or brought by a person | the rule's record and the case | the wording marked broken, with the case; a narrower wording or an exception, added; every link under the broken part queued for reading again |\n| Retire the machine's own link | a link the machine drew, confirmed by nobody, whose step could not be written or was written around it | the link and its step | the link superseded and kept visible, or redrawn as the chain it is (ruled 2026-09-13) |\n| Read again | anything a recorded judgment read has since been superseded, recut or reworded | that judgment's own inputs | a new judgment; the old one stays |\n\nThe numbers in the table (two links, two authors, three links) are starting values to be set by\nuse.\n\n**Whose steps.** Links the machine drew: always, since for them showing the work is the only way\nto count. A text's own step: by tentative answers from the source, offered for correction. A\nperson's own step: the person is asked first, and the machine's guess is shown second. A claim\nabout a person's own body, spirit or sacred commitments: nothing uninvited (the carve of\n25 August, and the consent and timing that the ruling on the sacred kept the same day).\n\n**Never by the writer alone.** A case against a rule counts when it was found in a text or brought\nby a person. A machine that drew the link, wrote the rule, imagined the case against it and checked\nthe step is one witness asked four times. Its imagined case becomes a question.\n\n### 5.3 A rule is kept as every wording proposed for it\n\n**The walkthrough of 8 October gave the picture: a rule you can open like a cupboard, with what\nprompted it on the left and what failed to break it on the right.** This makes the cupboard a\nrecord.\n\n- Every wording proposed for a rule is kept, with who proposed it.\n- **A wording is too narrow** when a link that a text or a person drew needs the rule and the\n  wording does not cover its case.\n- **A wording is broken** when a case found in a text or brought by a person falls under it and\n  the step fails there.\n- **Of the wordings that are neither, the map shows two**: the narrowest and the widest. A reader\n  sees how much room is left between them.\n\n**Cases move the first and counter-cases move the second.** A new link pushes the narrowest\nstanding wording just wide enough to cover it. A counter-case pushes the widest standing wording\njust narrow enough to exclude it. Mechanism evidence says how far the first may be carried without\na case.\n\n**The outside form is Mitchell's version spaces** (1977): keep every candidate rule that fits\nthe cases so far, represented by its most specific and most general members, and let each new case\nremove the candidates it contradicts. Two properties carry over. The result does not depend on the\norder in which cases arrive, which is the property ledger T4 asks of the whole graph. And the\nmethod does not commit to one best rule, which is AD25's *rules in the plural* with a shape.\n\n**What the gap means.** Where the narrowest and the widest standing wordings are far apart, the\ncases have not settled how far the rule reaches, and that is where a new case or a mechanism is\nworth most. Where they meet, the rule is hard to vary, which is the corpus's own test of a good\nexplanation ([depth.md](../depth.md)).\n\n**His drawing of 2013, by hand.** The case: *it will save companies money to fire workers*. The\nrule as he drew it: *companies will fire workers if it saves them money*. A link somebody draws\nabout a town council laying off staff to save money needs the rule, and the wording does not cover\na council: the narrow wording widens to *employers*. A case somebody brings of a firm that kept its\nstaff through a slump although firing would have saved money breaks *any employer, whenever it\nsaves money*: the wide wording gains *unless dismissal costs it more in other ways*. Why saving\nmoney leads to firing (owners press for profit, and dismissal is allowed) says where the rule\ncannot go: not to a firm its workers own. The council and the firm are invented for this note.\n\n**For a sign there is no rule to widen.** The same record serves the law that runs from the thing\nto its sign (*dead squirrels smell of decomposition*), and rivals attach to the link's question\n*what else*, never to the law.\n\n**Limits.** Narrower and wider need a ladder of kinds (squirrel, rodent, animal), and the machine's\nladder is everyone's average. Mitchell's method is known to fail when the cases contradict each\nother, and here they will. So the record keeps a broken wording with the case that broke it and\ndoes not delete it, and a contested case is a claim like any other.\n\n### 5.4 What each guess was based on\n\n**Every model call has been written down since 2026-10-06**, with its instructions and its inputs\n(`deliberus/judgments.py`; the founder: *\"Yes build it.\"*). The design's later step is the number\nof that record on each claim and link a machine wrote. With it, *based on what* has an exact\nanswer for every guess: the texts its record lists. And *read again* in § 5.2 is a query: which\nrecorded judgments read something that has since changed.\n\n---\n\n## 6. What this changes in the earlier proposals\n\n| Proposal | As written | As it stands after this note |\n|---|---|---|\n| AD19 | depth by taking a rule in its turn; never to the bottom in one go | the order of work for *made* sentences. Finding goes down without that limit, and the standard is his: nothing separable stays bundled. *No chain in one go* stands as a control: one level to a call (ruled direction AG6; controls AG7 to AG11) |\n| AD29 | any link can be written out on request, the first level eagerly; each chain ends in a named kind of stop | eagerly at every level, found first and made where nothing is found; made depth counted, checked and, once measured, capped; a chain ends in a located contact, a dated reading or a marked gap |\n| AF1 | the machine's sentence starts narrow, widens as far as a mechanism reaches, narrows at each case that breaks it | kept, with its form given in § 5.3: every wording stored, two shown |\n| AD3 | across debates the step proposes and does not draw | folded into § 4.2: any machine link with nothing under it is shown as a proposal |\n\n---\n\n## 7. Proposals (ledger Arc AG; unruled, except AG6)\n\n| # | Proposal | § |\n|---|---|---|\n| AG1 | **A leap may be left; it may not be left unmarked.** Every link shows what has been found, written or examined under it; a link a machine drew with nothing under it is shown as a proposal; a support link whose middle is on the map shows that middle; the share of bare links is published. | 4.1, 4.2 |\n| AG2 | **The way down is finding first.** At every level the machine looks before it writes, and what it writes where it finds nothing is a question on the link. A chain ends in a located contact with the world, a dated reading, or a marked gap. *(First written with: below its first level the machine only finds and asks. Withdrawn by AG6.)* | 4.3 |\n| AG6 | **Ruled 2026-10-10**: *\"It may and should write deeper steps. But the microtasks in charge of doing so need to be quality-controlled, just like the other microtasks that are in play in Deliberus.\"* What that control is, is proposed in rows AG7 to AG11: [quality-control-for-deeper-steps.md](quality-control-for-deeper-steps.md). | 4.3 |\n| AG3 | **Every causal and sign link carries the question how**, and its answer is a chain of located claims, each with the same questions. A mechanism a machine writes is shown as questions. | 4.4 |\n| AG4 | **A rule a machine proposes is kept as every wording proposed for it**, with the narrowest that covers the links texts and people drew and the widest that no found case has broken. | 5.3 |\n| AG5 | **The pass that writes under links acts on conditions of the graph, each with a stated reading set, and every write carries the number of its recorded judgment.** Whose steps it may write under is part of it. | 5.1, 5.2, 5.4 |\n\n---\n\n## 8. What this does not settle, and what could go wrong\n\n- **More steps are not more safety.** A long chain of steps that each look sound can hide an error\n  that nobody can locate. The threat model has held this since July, and its named guard is that a\n  decomposition come out the same when it is made again. On 7 October, 12 of 24 links had two\n  different sets of sentences that were each enough, which the corpus reads both as several\n  arguments and as a warning.\n- **One family of models all the way down.** Hence the rule that cases, counter-cases and\n  mechanisms count when found in a text or brought by a person.\n- **No person has read a written-out step.** By the principle of 2026-10-06 a person reads a pass\n  before anyone proves how often it is wrong. Everything measured here is a model reading a model.\n- **Writing the unsaid is not neutral.** Some things are left unsaid on purpose. A written rule is\n  offered and never put in an author's mouth.\n- **Cost.** Seventeen calls a link on 7 October. The four ways of § 4.3 are ordered cheapest first\n  for that reason, and the first needs no model.\n- **A mechanism can be a story.** The record of what was found, and where, is what separates one\n  from the other.\n- **The lift is open**, and until it is ruled an examined step can lower a link and can hardly\n  raise it.\n\n---\n\n## 9. What was read, and how far\n\n**In the corpus, this session.** In full: [how-far-a-premise-is-from-its-claim.md](how-far-a-premise-is-from-its-claim.md);\nthe walkthrough of 8 October ([what our own notes already knew](../listening/2026-10-08-what-our-own-notes-already-knew.md));\nthe three readers' reports of 8 October in `what-a-rule-rests-on/`;\n[kartpaus-auto-enrichment-design.md](kartpaus-auto-enrichment-design.md). In the sections cited:\n[writing-out-the-unwritten.md](writing-out-the-unwritten.md) § 0, § 6, § 7, § 9, § 10;\n[philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 4, § 5, § 6;\n[the-squirrel-and-the-strength-calculus.md](the-squirrel-and-the-strength-calculus.md) § 0 to § 6;\n[what-a-rule-rests-on.md](what-a-rule-rests-on.md) § 0, § 6; [decomposition-axes.md](decomposition-axes.md)\n§ 3, § 4; [what-counts-as-evidence.md](what-counts-as-evidence.md), the passages on mechanism,\non a reasonable challenge and on carrying a finding to a new case;\n[proving-a-pass-reliable.md](proving-a-pass-reliable.md) § 1;\n[relevance-and-the-two-factors.md](relevance-and-the-two-factors.md) § 0;\n[storing-the-record-of-machine-judgments.md](storing-the-record-of-machine-judgments.md) § 7a.\nNot opened: [what-a-link-rests-on.md](what-a-link-rests-on.md), which is known here through the\nnotes that build on it.\n\n**In the code.** `deliberus/relevance.py`, `deliberus/relevance_judge.py` (its header),\n`deliberus/latent_questions.py`, `deliberus/extraction/schemes.py` (the causal patterns and the\ntable that sorts each question), the route that argues a link's relevance in `deliberus/api.py`,\nand the header of `deliberus/judgments.py`.\n\n**Outside, on 2026-10-10, as search extracts from the works themselves. Verify before citing\noutward.**\n\n| Source | Read | Used for |\n|---|---|---|\n| Lamport, *How to Write a 21st Century Proof* (Journal of Fixed Point Theory and Applications, 2012), and his *Making Math More Rigorous* | long extracts of both | § 4.5 |\n| Machamer, Darden and Craver, *Thinking about Mechanisms* (Philosophy of Science, 2000); Stanford Encyclopedia of Philosophy, *Mechanisms in Science* | extracts of § 5 of the paper and of the entry | § 3, § 4.4 |\n| Mitchell, *Version Spaces: A Candidate Elimination Approach to Rule Learning* (IJCAI 1977), with two later papers that restate the method | abstract and extracts | § 5.3 |\n\n## See also\n\n[quality-control-for-deeper-steps.md](quality-control-for-deeper-steps.md) ·\n[alien-proofs-and-the-library.md](alien-proofs-and-the-library.md) ·\n[writing-out-the-unwritten.md](writing-out-the-unwritten.md) ·\n[philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) ·\n[what-a-rule-rests-on.md](what-a-rule-rests-on.md) ·\n[how-far-a-premise-is-from-its-claim.md](how-far-a-premise-is-from-its-claim.md) ·\n[kartpaus-auto-enrichment-design.md](kartpaus-auto-enrichment-design.md) ·\n[the-squirrel-and-the-strength-calculus.md](the-squirrel-and-the-strength-calculus.md) ·\n[self-similar-decomposition-and-claim-ontology.md](self-similar-decomposition-and-claim-ontology.md) ·\n[corpus-graph/README.md](corpus-graph/README.md)\n"}