{"path":"research/writing-out-the-unwritten.md","content":"# Writing out the unwritten: every link written out in full, and the writing checked\n\n**Date**: 2026-10-07 (evening) · **Type**: a founder request, answered by reading his record since\n2011, the corpus, the code and the paper he pointed at; three helper readings kept as they came\nback; and one paid trial on thirty links with thirteen predictions written first. **Status**: every\nproposal is Claude's and **unruled** (ledger AD13 to AD21). Nothing is built and nothing was\nwritten to the graph. One correction of attribution is made on five surfaces (§ 12). It came ahead\nof the four open reviews because he asked.\n\n> ⚠ **Four things here were corrected the same night**, in [philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 13.\n> (1) Where the thirty end up: **17** with a step in which every part is needed, **5** written\n> around, **2** unsettled, **6** not written out, where this note says 15, 9 and 6 (18, 5, 1 and 6\n> stood in this line until 2026-10-08, when link 19 was found unsettled). The pruning\n> tried one order of leaving sentences out, and a step can have more than one small set that is\n> enough. (2) *The reader needed 1.2 added sentences* is a count for that one order. (3) A step\n> with everything written in can still fail to pass support on. (4) His example of 2011 is a step\n> from a sign to its cause, and the blind reader used here, which turns out to read deductively,\n> refuses it as he wrote it.\n\nHis words, on hearing this afternoon's proposal to name what a link rests on and judge nothing\n([what-a-link-rests-on.md](what-a-link-rests-on.md)): ***\"Haven't we written and thought a LOT about\nunwritten/implicit premises since the start of this project? And this also feels like what the\nSciencePedia paper calls the dark matter of reasoning in science, or something like that? Read up\ndeeply on this. I think merely naming what the links consist of would be weak sauce. I want to have\nall the implicit reasoning structure and premises be explicated and exposed and scaffolded\nautomagically. To the extent that process can be trusted of course. This depends if we can split it\nup into microtasks that we can rely on deeply.\"***\n\n## 0. The short version\n\n- **He is right on both counts.** He has asked for unwritten premises to be written out since 2011,\n  at least five times in the last eight weeks, and his 2011 example is already a step written out\n  in full (§ 1). Seven mechanisms were built or proposed against that record. Together they have\n  written **34 claims** out of the unwritten, beside 2,139 taken from the texts (§ 2).\n- **One sentence of mine from March narrowed his request**, and stood on five surfaces as the\n  settled design: *never generating premises, only finding real ones*. His own sentence that day\n  said *finding or creating*. Corrected here (§ 3, § 12).\n- **The paper he pointed at checks where a chain ends and never checks its steps.** That guard\n  needs an answer a machine can verify, which a debate does not have. What carries over, and had\n  been missed: a link gives **both ends**, so the middle can be checked for whether what is written\n  is enough to get from the first claim to the second (§ 4).\n- **The stronger design: write every link out in one form, and check the writing with small tasks\n  that hold no opinion about the link.** The form is the first claim, what ties this case to a rule\n  (two wordings for one thing; a further fact), and one general rule (§ 6). The checks: by code, by\n  a reader who is not told a link was drawn, by writing twice, and by looking for each sentence in\n  the two texts (§ 7).\n- **Tried the same afternoon on thirty links between texts, one model, predictions first** (§ 8).\n  What held: no rule named anything particular (0 of 29); the blind reader gave the same answer in\n  both orders on 26 of 29 steps; two separate writings gave the same rule, or one containing the\n  other, on 23 of 27; and of 119 pointers into the texts none went to a claim that does not exist.\n- **What failed, and what each failure taught.** The writer writes a rule for almost anything, so\n  *a rule can be written* says nothing. It wrote 3.6 added sentences a step where a reader needed\n  1.2, so the reader decides what is needed and the writer never does. And on 9 of 24 steps the\n  reader reached the second claim **without the first claim**: what the machine wrote there went\n  around the link, through a middle claim or another argument.\n- **Where the thirty end up**: 2 plain, 13 resting on a rule or fact with every part needed, 9\n  written around, 6 not written out to the reader's satisfaction. So **15 of 30 pass every check**.\n  ⚠ *Corrected: 15 where 13 stands, 5 where 9 stands, 2 unsettled, so 17 of 30\n  ([philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 7).*\n- **What the checks do not do.** Six of the seven links I had read as *not a reason* come out\n  clean: the form puts their rule on the page and no check says the rule is false. **The checks\n  sort links by structure. They do not sort good reasons from bad.** That part stays with people,\n  and it is now a sentence they can point at.\n- **What writing out found that naming could not.** Thirteen of fifteen rules stand in neither\n  text. And looking for each sentence in the texts found the crux of the consciousness exchange:\n  one text states, and the other denies, that *subjective experience* and *consciousness* are one\n  thing (§ 8c).\n- **So his condition is met in part.** The scaffold and its checks are small tasks of the kind that\n  can be leaned on. The content of a rule is a proposal and stays one: shown as a question with a\n  move beside it, nobody's claim, weighing nothing until a text states it or a person takes it up\n  (§ 9).\n- **Limits, plainly**: one family of models, thirty links, one afternoon; **no person has read a\n  written-out step**; the blind reader leans toward accepting; and at about 17 model calls a link\n  this first form is too dear to run over the graph (§ 10).\n\n## 1. His record: fifteen years of asking for the unwritten to be written out\n\nThe helper that swept the repository kept his words apart from every document's restatement of\nthem. Each line below was then checked at its source by Claude before it was put here.\n\n| When | His words | Source |\n|---|---|---|\n| 2011-08-29 | *\"Just to clarify I mean simply to argue that \"The squirrel is dead.\" from \"The squirrel smells of decomposition.\" along with \"Dead squirrels smell of decomposition.\" [...] I guess what I'm wondering is if they have to be deconstructed/reduced so that what remains is one most proximal set of premises on which the inference is made. Leaning towards no. But I'm sleepy.\"* | a saved discussion thread in the archive; in no document until now |\n| 2011-12-14 | *\"I imagine a crowdsourced quality-optimization process that makes sure every premise is made explicit and neatly & pedagogically spelled out, fallacies sensed and avoided, irrelevant words sorted out, etc.\"* | the same thread |\n| 2011-12-18 | *\"I've decided to skip the relevance metric. It seemed redundant, since the validity metric should encapsulate all we need anyway. If an argument is valid, it is automatically relevant. Right?\"* | the same thread; in no document until now |\n| 2011-12-29 | *\"The general strategy is to leverage automation as far as possible, then move on to crowdsourcing; designing microtasks for users to do the work that still is beyond computers.\"* | the same thread |\n| 2013-01 | His one-page board has a box *\"Identify missing/implicit premises\"* beside *\"Automation + crowdsourcing; where one ends the next takes off\"*. Its drawing gives each premise a second premise written beside it, each with its own truth value: *\"Lärare är ofta arga\"* (0.6) and *\"Om lärare ofta är arga, är de arga\"* (0.9), under the badges *valid* and *relevant*, with the note *\"\"Lärare är ofta arga\" är relevant för \"Lärare är arga\"\"*. A second drawing puts *\"Companies will fire workers if it saves them money\"* beside *\"It will save companies money to fire workers\"*, under *\"Companies will fire workers\"*. | two drawings in the archive, viewed |\n| 2026-03-29 | *\"Could the system include auto-finding premises for each new claim submitted, so that all claims would be the conclusion of a few logical premises? That's the way I (or my left hemisphere at least) always imagined it. But perhaps I need to let go of that rigorous analytical hope. Perhaps that's a pipedream.\"* And: *\"If the system would auto-decompose incoming new claims in a way that automatically prunes them based on some criteria, like only finding or creating new premises that could one step removed be allowing connections to other existing claims. Then perhaps it would be more computationally feasible and helpful.\"* | [scheme-bounded-decomposition-and-evidence-as-subgraph.md](scheme-bounded-decomposition-and-evidence-as-subgraph.md) |\n| 2026-08-14 | Recorded only in paraphrase: that a premise one side assumes and the other never uses *\"must **always** be made explicit by the pipeline\"*. His own sentence was not kept. | [decomposition-axes.md](decomposition-axes.md) |\n| 2026-08-16 | *\"All of this should be built.\"*, on five items: the splitter looks before it mints; the pass for unstated premises runs on the interactive path; a decision on a second pass for background premises; more of them found; the counts taken again. All five are unchecked today. | `TODO.md` |\n| 2026-08-17 | *\"They SHOULD of COURSE be INVITED to dive in and review and reflect and certify or fix by hand any and all small or large mistakes or inconsistencies or plot holes or missing implicit premises or logic jumps or WHATEVER, BUT Deliberus should ideally over time be made capable of offloading as much of the idealized processing and structuring and scaffolding as possible.\"* | [conviction-and-critique.md](conviction-and-critique.md) |\n| 2026-08-25 | *\"Generally I want to capture accurately how people really actually argue and think. All unstated premises need to be clarified and scaffolded.\"* [...] *\"the full premise structure of what's been said/typed/implied ... and I'm leaning towards that this should also occur (to some degree/cap at least) in normal use of Deliberus online in the future...\"* | [weighing-scheme-and-terminus-classifier-design.md](weighing-scheme-and-terminus-classifier-design.md), the addendum of that evening |\n| 2026-09-09 | *\"The gap in question could also just be a missing premise that a user supplies (and which is then vetted somehow by humans & machine, but yeah)\"* | `specs/landing-redesign/constellation-map.md` |\n| 2026-09-21 | *\"To me, whether anything is shown as potential work to be done by a human user in the UI depends on: how confident can we be in the machine/system instead doing that work on its own, so it doesn't even need a UI?\"* | [split-review-and-the-founders-intentions.md](split-review-and-the-founders-intentions.md) |\n| 2026-09-29 | *\"The more we succeed in this designing of sturdy microtasks and other creative innovative challenging of what's possible to automate safely [...] the less humans need to feel overwhelmed\"* | [proving-a-pass-reliable.md](proving-a-pass-reliable.md) § 9 |\n| 2026-10-07, morning | *\"This needs to be handled/explicated/scaffolded by some pass. Maybe a new pass that's more like a 'standing property of the graph' in some ingenious way.\"* | [what-a-link-rests-on.md](what-a-link-rests-on.md) |\n| 2026-10-07, afternoon | *\"I think merely naming what the links consist of would be weak sauce. 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.\"* | this session |\n\n**Three things in that record that no document had.**\n\n- **His 2011 example is a written-out step.** A claim about one case (*the squirrel smells of\n  decomposition*), a general rule (*dead squirrels smell of decomposition*), and the conclusion.\n  The same comment asks the morning's question of fifteen years later: whether premises must be\n  reduced to *\"one most proximal set\"*.\n  ⚠ *Corrected the same night: it is written out as a step from a sign to its cause. The trial's\n  reader refuses it as he wrote it and accepts it with the rule turned round\n  ([philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 5).*\n- **His 2013 drawings write the second premise in.** One writes the bare conditional (*if teachers\n  are often angry, they are angry*), which says nothing the arrow did not. The other writes a rule\n  and a fact about the case (*companies fire workers if it saves them money*; *it will save them\n  money*). The difference between those two is the difference § 5 turns on.\n- **His 2011 remark on relevance is the idea behind writing a step out.** *\"If an argument is\n  valid, it is automatically relevant.\"* A friend answered that relevance belongs to the arrow\n  between a claim and what it supports, and the 2013 board carries both badges. The two-factor\n  ruling of 2026-09-28 sided with the friend, for links as they are drawn. Both are right about\n  different objects: an arrow needs a reading of how much it bears, and a step with everything\n  written in does not, because whether it bears has become whether its written premises are true.\n  ⚠ *Too quick, corrected the same night: a complete step can still fail to pass support on, and a\n  rule's standing is not the link's strength ([philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 9).*\n\n**So he has asked for this at least five times in eight weeks, and my answer of this afternoon answered a smaller\nquestion.** He asked for *handled/explicated/scaffolded*. I proposed *named, and judged by nobody*.\n\n## 2. What was built against that record: seven mechanisms, none of which writes a step out\n\nCounted on the live graph on 2026-10-07 unless a date is given.\n\n| Mechanism | What it does | What it reaches |\n|---|---|---|\n| The pass for unstated premises inside one argument (`implicit_premises.py`, April) | One call per text. Told to *\"prefer fewer\"*, at most one or two per argument, and to skip arguments that are already explicit. Runs on one of the two extraction paths. | 23 claims, hand extractions among them. 24 of 35 sources have none. |\n| The pass that reads two opposed texts whole (`cross_source_premises.py`, August) | One call per pair of texts, one to three premises, run by hand, proposals only. | 11 claims stored, from four pairs. |\n| Critical questions | Each argument pattern brings its standing questions, most of which probe what the step takes for granted. Since September they are rendered from the link and become claims when answered. | Every one of the 4,756 links that carry a pattern. None of the 871 links the linking step drew: it stores no pattern. |\n| The generic question for a link with no fitting pattern (`warrant_acceptable`, 18 August) | Its template: *\"IF {premise} holds, {conclusion} genuinely follows — the connecting assumption survives being stated out loud\"*. | 191 links are asked whether an assumption survives being stated out loud. Nothing states one. |\n| A link's relevance argued as a claim (28 September) | A person can argue that one claim bears on another; the argument is a claim, and the link's weight follows that claim's standing. | No link has one. |\n| The splitter | Cuts a claim into the parts it asserts. A different axis: parts of a claim, not premises of a step. | 60 parts. |\n| Naming under a link (this afternoon, AD6 to AD12) | Up to three further claims and the names joined, stored as text, judged by nobody. | Proposed. Unbuilt. |\n\nTogether: **34 claims that some pass wrote out of the unwritten, beside 2,139 claims taken from the\ntexts.** A study of argument reconstruction read for the afternoon's note (Ryu and colleagues,\n2026, in extract only) puts unstated premises at about four in ten of all premises in arguments\nthat have been reconstructed. The graph holds under two in a hundred.\n\nEach mechanism is aimed at a different object (a text, a pair of texts, a pattern, a link's\nweight, a claim's parts), and each is sparing by instruction or runs by hand. None takes the one\nobject that is both bounded and everywhere: **a link**.\n\n## 3. One sentence of mine that narrowed his request, and its ground\n\nHis sentence of 2026-03-29 says *\"finding **or creating** new premises\"*, pruned by whether they\nconnect. The same document, in Claude's prose four paragraphs later, turns it into this:\n\n> *\"The better inversion — Auto-CONNECT, not auto-decompose\"* [...] *\"This avoids hallucination\n> entirely — you're never generating premises, only finding real ones.\"* [...] *\"An LLM evaluates:\n> \"Does existing claim X logically support new claim Y?\" — classification (high accuracy), not\n> generation (hallucination-prone).\"*\n\nThat is not his position and no ruling adopted it. It was built: the linking step (auto-connect)\nis that classifier, and its 871 links are what it produced. The sentence was then repeated as if it\nwere the settled design in four other places, two of them printing it next to his quote\n(corrected in this commit: § 12).\n\n**Its ground is what today's measurements contradict.** *Does X support Y* is the question that had\nno steady answer on thirty links under four designs. Writing what the step would rest on is the\npart that held. For links, the judgement was the unreliable half and the writing was the checkable\none, the reverse of what the sentence assumed.\n\n**What stays true in it.** A premise found in a text has an author. A premise a machine wrote has\nnone. § 9 keeps that difference at the centre: found before made, and what is made is nobody's\nclaim until somebody takes it up.\n\n## 4. The paper he pointed at\n\nLi and colleagues, *Inverse Knowledge Search over Verifiable Reasoning* (arXiv:2510.26854, a\npreprint, third version of January 2026). Read for this note: the paper's method, verification and\nlimits sections from the copy in `data/papers/`, by Claude; the main text by a helper\n([its report](writing-out-the-unwritten/outside-reading-explication-with-checks.md) § 1); and the\ncorpus's private reading of May 2026.\n\n**What it says is missing.** Written knowledge is the product of what the authors call *\"the\nradical compression of reasoning\"*: writing costs people time, so textbooks and encyclopedias keep\nconclusions and drop the derivations. They call what was dropped the dark matter of knowledge,\nthe scaffolding under the recorded facts, and they name two costs. Without it a claim is hard to\ncheck, so the reader must trust an authority. And the links between fields are cut. They also say\nthat a model asked to write an encyclopedia will repeat the compressed form it was trained on.\n\n**How they trust the machine.** Three guards. A second model screens each question before it is\nsolved (about 5% are dropped). Questions are designed so that the answer can be checked by machine:\na number, a formula, code, a choice. And each question goes to at least two models from different\nmakers, the pair being discarded when the final answers differ. **The steps are not checked.** The\nanswer check is said to vouch for the reasoning that led to it. The paper gives no rate for answers\nthat agree and are wrong, no error rate for steps, and no check by a person. Two later preprints\nthe helper found, read in abstract or extract, report that a right final answer does not show the\nsteps were sound.\n\n**Why that guard does not carry over.** It needs an endpoint whose truth a machine can check. A\ndebate about whether a model is conscious has none. The corpus has said so since May: the\ndiagnosis transfers and the pipeline does not. Agreement between models is also weaker evidence\nthan it looks: on one 2026 measurement a panel of nine models from seven makers carried about two\nindependent votes\n([the second helper's report](writing-out-the-unwritten/outside-reading-dependable-small-tasks.md) § 5).\n\n**What does carry over, and had been missed.** The paper knows where each chain must *end*. A link\nin this graph gives **both ends**: two claims, each in a text. A step between two fixed points can\nbe checked for something that needs no oracle for truth: **is what is written enough to get from\nthe first to the second, with nothing missing and nothing idle?** That is a check of the middle,\nthe one thing the paper does not check. It says nothing about whether the added premises are true.\nIn science that would be a defect. Here it is the point: what gets added is usually what the two\ntexts disagree about, and the graph exists to hold that open.\n\n**Two more things the paper has.** Its search runs backwards, from a concept to every chain that\nreaches it. The same reading over written-out links runs from a rule to every link that rests on\nit. And it says what it leaves out: who held what. In this graph that is half the work. Every\nsentence written under a link needs an owner, and when no text states it the owner is nobody.\n\n## 5. Why naming was too weak\n\nThe afternoon's proposal stored, under each link a machine draws between two texts, up to three\nfurther claims and the names joined, and judged nothing. Four things are wrong with stopping there.\n\n1. **The form was free, and a free form drifts to the empty sentence.** Asked for *the further\n   claim*, the model often wrote the two claims joined by *if* and *then* with a few words changed.\n   Every source on the unstated premise warns of that sentence: it can be written under any step,\n   and it cannot show a bad step to be bad. It is also why the afternoon's check called the stated\n   premise idle on most links. With *if A then B* written in, A adds nothing. The repair is forty\n   years old and was in the afternoon's own outside reading: what is added should be a **general\n   rule the step is one case of** (Hitchcock 1985, 2011: the covering generalization). A rule that\n   names nothing from the case cannot do the case's work, so the stated premise is needed.\n2. **Nothing checked whether the named claims were enough.** The step with its further claims\n   written in was never put to a reader who did not know where it was meant to go.\n3. **It connected to nothing.** A named claim was text under a link. It was not looked for in the\n   two texts, not compared with the claims the graph already holds, and not joined to the pattern's\n   questions.\n4. **It stopped at links between texts**, and left an author's own steps as they are. His request\n   is for all of it.\n\n## 6. The form: a link written out\n\n**A written-out link is the first claim, what ties this case to a rule, and the rule. The second\nclaim is what follows.**\n\n| Part | What it is | His 2011 example | His 2013 drawing |\n|---|---|---|---|\n| The first claim | as its text has it | *The squirrel smells of decomposition.* | *It will save companies money to fire workers.* |\n| This case: names | two wordings for one thing, or a thing and the kind the rule speaks of | none needed | none needed |\n| This case: a fact | a further fact about the case, which no claim states | none needed | none needed |\n| The rule | one general sentence that names nothing particular from either claim | *Dead squirrels smell of decomposition.* | *Companies will fire workers if it saves them money.* |\n| The second claim | as its text has it | *The squirrel is dead.* | *Companies will fire workers.* |\n\n**Why a rule, and what makes one.** From Hitchcock, as two helpers read him: the rule generalises\nover something the two claims share; it does not give the conclusion by itself; it is not true\nmerely because the conclusion is; and a parallel case can test it. His procedure is four steps\n(form *if A then B*; list what the two sentences share; replace it with variables; simplify), and\nthe helper found no program that implements it and no measurement of it.\n\n**Why names are their own part.** Hitchcock's rule generalises over an expression the premise and\nthe conclusion *share*. Two claims from two texts often share none: one says *J-space*, the other\n*mental workspace*. Before any rule can be written, the names have to be joined. That is why the\nafternoon found a joining of names under 28 of 45 links. It is also where a debate's wanted words\nsit, and by the ruling of the same day a claim that two such words name one thing is somebody's\nclaim on a word.\n\n**Depth.** One rule to a link. A rule is a claim like any other, and what lies under it is found\nthe way anything is found here: by looking for claims that state it, support it or deny it, and\nwriting those links out in turn. A chain is not guessed in one go. Measured this afternoon: an\noptional middle step was filled in on 27 of 45 links, among them 6 of the 8 read as one step. And\nin the literature, on an earlier generation of models under strict scoring, whole chains came out\nright 79% of the time at one step, 18% at two and almost never at three or more.\n\n**Where it lives.** On the link, as potentials: each sentence is rendered with the link and becomes\na claim when somebody answers it, the way a critical question does since September. The rule is\nthe specific form of the generic question of § 2. This is the direction he ruled on 25 August for\nall machine scaffolding, in the design note of that day: *do eagerly, show on contest*, and\n*surfaces present questions, never finished analysis*.\n\n## 7. Small tasks, and which of them can be leaned on\n\nHis condition: *\"This depends if we can split it up into microtasks that we can rely on deeply.\"*\nThe second helper read what is measured about that\n([its report](writing-out-the-unwritten/outside-reading-dependable-small-tasks.md), 386 lines, each\nsource with its read depth). Four findings carry the design.\n\n- **Small tasks with one exact answer can be made as dependable as one likes.** One 2025 study cut\n  a task into single steps, had each step answered several times until one answer led by a margin,\n  threw out malformed replies, and ran over a million steps without an error. Its authors say what\n  it does not cover: steps with no exact answer, and errors that are not independent.\n- **Open writing cannot be voted into reliability.** The same premise written twice comes back in\n  different words, so there is nothing exact to count. Models from different makers also converge\n  on the same open-ended output, so agreement there shows a shared habit.\n- **What makes a check hold, measured**: the checker is not shown the draft or the answer it is\n  meant to reach; *is this part needed* is decided by someone other than the writer; both orders\n  are tried and only agreement is accepted, since telling a model that order must not matter does\n  not work; each task sees as little as it needs; and a countable or code-checkable answer is\n  asked for before any judgement.\n- **A whole is only as good as its seams.** From crowd work: context lost where the work was cut,\n  one bad piece upstream spoiling what follows, and a verifying step that is itself noisy (in the\n  classic case it caught half the errors and threw out some good work).\n\n**So writing a link out is cut along that line.** Every task with an exact answer is given to code\nor to a closed question asked blind. The one open task, writing, is boxed in from four sides and\nits output is never treated as true.\n\n| # | Task | Who | Kind of answer | What is known about leaning on it |\n|---|---|---|---|---|\n| 1 | Write the rule, the names and any fact | model | open writing, in a fixed form | Cannot be voted. Measured here: § 8. In the literature: free gap-filling by people gives 2.7 premises with a third of the words shared; inside one narrow pattern two annotators wrote the same missing premise in 69 of 71 cases. Form narrows it. |\n| 2 | Is the rule general? | code | exact | Deterministic. § 8. |\n| 3 | Does a sentence say a claim over again? | code | a count | Deterministic. The line to draw on the count is a choice. |\n| 4 | Which of four claims follows from these statements, if any? | a reader not told that a link was drawn | closed | § 8: steady across the two orders on 26 of 29. Elsewhere: readers of this kind lean toward accepting (an older model let through about half of wrong steps), and order can move an answer by tens of points when premises are many. |\n| 5 | Is this part needed? | code, from task 4 run without the part | exact, given task 4 | The writer never decides it. This is the design rule above, and it is the opposite of the afternoon's optional *middle*, which the writer filled in on 27 of 45. |\n| 6 | Write it again, having seen what the reader said | model | open, in a loop | A published loop of write, check, rewrite raised a human rating of its reconstructions from 2.9 to 4.5 out of 5 in about three rounds. § 8 tried one round. |\n| 7 | Is this the same rule as that one? | the sameness check the graph uses | closed | Steady here for *same* against *unrelated*; unsteady for partial overlap. § 8. |\n| 8 | Does a claim of this text state this sentence, or deny it? | model points, code checks the claim exists | pointing | Of the kinds measured steady here. § 8. |\n| 9 | How many links rest on this sentence? | code, after task 7 | a count | Triage, not proof: what recurs is where attention pays, not what is true. |\n| 10 | Read it | a person | | Nobody has. |\n\n**What no small task removes.** Whether a rule is true. Whether a rule is the one an author had in\nmind, which only the author can say. How broad a rule should be: broader says more and risks\nputting words in a mouth, and the sources disagree on where to stop. And whether two wanted words\nname one thing.\n\n**What that means for trust, said plainly.** The scaffold and its checks can be made dependable.\nThe content of a rule is a proposal and stays one. A written-out step that passes every check has\nbeen shown to be complete and to hold together. It has not been shown to be right.\n\n## 8. The trial: thirty links written out, and four checks that ask nobody what they think of the link\n\n**What was run.** The thirty links the linking step drew between two texts, the sample read by hand\nthat morning (8 one step, 11 one further claim, 4 several steps, 7 not a reason; Claude's reading,\nno person's). One model throughout, `claude-sonnet-5`, the family of the hand reading too. Nothing\nwas written to the graph. Predictions were written to a file before the first call\n([predictions](writing-out-the-unwritten/written_out_predictions.txt)), with two dated additions:\nafter a three-link pilot showed faults in the test, and after the blind readings showed the steps\nover-full.\n\n1. **Write.** The step is written out in the form of § 6, twice, under two wordings of the task\n   (one starts from the rule, one from a reader who grants the first claim and doubts the second).\n2. **By code.** Does the rule name anything particular from either claim; how many of its words\n   does each sentence share with each claim.\n3. **Blind.** A reader that is not told a link was drawn gets numbered statements and four claims\n   from the second claim's own text, and says which follows, if any. Four times: the whole step in\n   two orders, the step without the rule, the step without the first claim.\n4. **Repair, once.** Where the whole step was not enough, the writer is shown what the reader said\n   and writes again.\n5. **Prune, by code.** Each added sentence in turn is left out; it is dropped when the reader still\n   reaches the second claim without it in both orders. Then the first claim is left out of what\n   remains.\n6. **Twice.** The two rules from step 1 go to the sameness check the graph already uses.\n7. **In the texts.** For the fifteen links between the two consciousness texts, each sentence of the\n   step is looked for among each text's own claims: stated, denied, or neither. Code checks that\n   every claim pointed at exists.\n\n**What the pilot changed.** Two of twelve blind replies were the runner's known probe (*\"Test\"*)\nand not readings: a reply under eight words is now asked again. The first form let a case sentence\nsay *what the first claim reports is a case of what the second asserts*, and the blind reader\nreached the second claim without the first claim and said why: names became two wordings for a\nthing, joined by code. And looking in the texts had counted a claim about one case as stating the\ngeneral rule.\n\n### 8a. Thirteen predictions\n\n| | Prediction | What came back | |\n|---|---|---|---|\n| P1 | The two wordings agree on whether a rule can be written, on at least 24 of 30 | 27 of 30 | held |\n| | Of the 7 links read as not a reason, both decline on 2 to 5 | 0. A rule was written for 29 and 28 of 30 | **failed** |\n| P2 | At most 6 rules name something particular from the claims | 0 of 29 | held |\n| P3 | A rule shares under a quarter of its words with each claim (median) | 0.14 with the first, 0.25 with the second | half: a rule has to speak of the conclusion's kind, and the prediction did not allow for that |\n| P4 | The whole step is enough for the blind reader on at least 75% | 26 of 29, first claim first; 23 of 29, rule first | held |\n| P5 | The reader gives the same answer in both orders on at least 85% | 26 of 29 | held |\n| P6 | Without the rule the reader still reaches the second claim on at least half of the one-step links | 7 of 8 | held |\n| | and on at most a third of the far links | 7 of 14 | **failed** |\n| P7 | Without the first claim it reaches the second claim on at most 20% | 13 of 29 | **failed** |\n| P8 | The two rules are the same, or one contains the other, on at least 60%; unrelated on at most 10% | 23 of 27 (8 the same, 15 one containing the other, 4 overlapping); unrelated on none | held |\n| P9 | At least 8 of 15 links have a sentence stated or denied by a claim that exists | 15 of 15; 119 pointers, none to a claim that does not exist | held |\n| | At least 4 have a sentence one text states and the other denies | 1 | **failed** |\n| P10 | At least half of the steps sent back pass after one repair | 1 of 3. One was rewritten and still not enough; on one the writer answered that no rule can be written after all | **failed** |\n| P11 | After pruning, the reader reaches the second claim without the first claim on at most 20% | 9 of 24 | **failed** |\n| P12 | At most 3 of the 8 one-step links keep a rule or a fact after pruning | 5 of the 7 pruned | **failed** |\n| | At least two thirds of the far links keep one | 11 of the 11 pruned (15 links) | held |\n| P13 | A pruned step holds 1.5 added sentences or fewer | 1.21, where 3.62 were written | held |\n\nCounting each half as its own part, eighteen things were predicted: ten held and eight failed.\n\nFor comparison, from the same thirty links earlier that day: a free naming came back the same on 6\nof 10; a check that asked one reader whether each premise was idle changed with the order of the\npremises on 14 of 44.\n\n### 8b. What the failures say\n\n**The writer writes a rule for almost anything.** Asked cold, it declined one link of thirty under\none wording and two under the other, none of them twice, and none of the seven read as not a\nreason under both. The literature has the name for it: a reconstruction model is *\"maximally\ncharitable\"*. So whether a rule *can* be written says nothing about a link. The one refusal that\ncame with a ground was link 1: the blind reader could not get from the written step to the second\nclaim, the writer was shown that, and it answered that the only sentence that would bridge the gap\nwas the second claim's own content over again.\n\n**The steps were over-full, and the writer cannot be the one who trims them.** It wrote 3.6 added\nsentences a step. The reader needed 1.2. Of 42 joinings of names it could do without 37, and of 21\nfacts, 16. The afternoon had found the same habit from the other side: an optional middle step was\nfilled in on 27 of 45. Whether a part is needed has to be read from what a reader does without it.\n\n⚠ *Corrected the same night: these are counts for one order of leaving out (names, then facts, then\nthe rule). On 12 of the 24 pruned links the step without the rule had also been enough, so the case\nsentences and the rule could each stand in for the other, and the order decided which survived\n([philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 7).*\n\n**Leaving out the first claim sorts the links by structure.** After pruning (⚠ corrected the same\nnight in [philosophy-of-the-unwritten-step.md](philosophy-of-the-unwritten-step.md) § 7: 15 where 13\nstands below, 5 where 9 stands, and 2 unsettled):\n\n| Where a link ends up | Links | Read by hand as |\n|---|---|---|\n| Plain: nothing needed adding | 2 | 2 one step |\n| Rests on a rule or a fact, and the first claim is needed | 13 | 4 one step, 2 one further claim, 1 several steps, 6 not a reason |\n| What was written reaches the second claim without the first claim | 9 | 1 one step, 6 one further claim, 2 several steps |\n| Not written out to the reader's satisfaction | 6 | 1 one step, 3 one further claim, 1 several steps, 1 not a reason |\n\n*The nine, read one by one (Claude's reading).* What was kept is the second claim in general form,\nor another reason for it. The reader says so in its own words: the second claim *is just this\nprinciple instantiated*. Example: from *to earn money a producer must satisfy what consumers want*\nto *employers will not pay a worker more than the worker's output is worth*, the machine wrote\n*producers avoid paying more for an input than the revenue it lets them earn*. The second claim is\na case of that rule alone. The first claim is at most a reason for the rule. So the honest drawing\nis a chain through the rule, and the machine found the middle of it without being asked for one.\nEight of the nine are links the hand reading called far.\n\n*The six read as not a reason come out clean.* Six of seven end as a first claim, one rule, and\nthe second claim, each part needed. What makes them poor reasons is the rule: *acts that differ\nonly in how they are classed do not differ in whether they advance justice* is what one would have\nto hold to use a defender of the death penalty against it. The form puts that sentence on the page.\n**No check here says it is false, and none should.** The checks sort links by structure. They do\nnot sort good reasons from bad.\n\n**The blind reader is steady, and it is lenient.** Its answer was the same in both orders on 26 of\n29 whole steps, and on 81 of 87 readings with a sentence left out. Read one by one it often names\nexactly what is missing. It also accepts a step with a hole in it now and then, and it joins two\nnames without being told they are one thing, which is why 37 of 42 joinings looked unneeded. A\nreader of the writer's own family takes for granted what the writer takes for granted. **So this\nreader cannot decide whether a joining of names is needed**, and that drop is no evidence that\nthe joinings are idle.\n\n### 8c. In the texts\n\nFifteen links, 53 sentences (15 rules, 27 joinings of names, 11 facts), each looked for among the\nclaims of both texts.\n\n| | Stated in the paper | Stated in the reply | Denied in the paper | Denied in the reply | In neither |\n|---|---|---|---|---|---|\n| Rules (15) | 2 | 0 | 0 | 0 | **13** |\n| Joinings of names (27) | 17 | 4 | 2 | 0 | |\n| Facts (11) | 6 | 4 | 2 | 1 | |\n\n- **Thirteen of fifteen rules stand in neither text.** The rules are the unwritten layer.\n- **Most joinings of names are in the paper's own claims.** The claim that the *J-space* is the\n  model's internal workspace is there six times over. This afternoon's note called that joining\n  the commonest unwritten thing under these links. It is written, on one side: the word *mental*\n  is the reply's.\n- **On three links a sentence the step needs is denied by one of the texts.** On one, the reply\n  states it and the paper denies it: that *subjective experience* and *consciousness* are one thing.\n  The paper's claims keep them apart (access consciousness is defined by what a state does). That\n  is the point the exchange turns on, and it was found by pointing at sentences. On another, the\n  step from *the researchers report key differences* to *the findings fall short of the theory's\n  criteria* needs the differences to be ones the criteria turn on, and the paper's own claims say\n  the main difference may not matter. Its author would not grant the step.\n- The denials were read by Claude and look right. A pilot reading produced one that was not (a\n  claim about the untrained model taken to deny one about the trained model), so they want a\n  person's eye.\n\n### 8d. What it used, and one stopped run\n\nAbout 512 model calls on the keeper's plan, pilot and pruning included: about $27 at API list\nprices. **No money was paid**: a dollar figure here measures plan allowance. That is about 17 calls\nand 90 cents a link in this first form, most of it the blind readings. The weekly plan figure moved\nfrom 80% to 83% during the afternoon, with this session's own work and three research helpers in\nthat. The main run was stopped at its forty-minute limit during its last stage with nothing of\nthat stage saved. All 29 calls it had made were in the record of machine judgments and were\nrebuilt from there with no new call\n([the script](writing-out-the-unwritten/written_out_recover_pointing.py)); the trial script now\npicks up where a stopped run left off.\n\n## 9. The design, revised\n\n**9.1 Every link is written out, in one form.** The first claim; what ties this case to the rule\n(names, and any fact no claim states); the rule. One rule to a link. This replaces *up to three\nfurther claims* (AD6). It is a standing property with a debt, as proposed that afternoon: a link\nwithout a written-out step owes one. The debt starts with the 871 links the linking step drew and\ngoes on to an author's own steps, where the pattern on the link already supplies the most general\nrule.\n\n**9.2 Nothing judges the link. The writing is checked.** By code, blind in both orders, twice, and\nin the texts. A reading on which the two orders disagree goes to a person. The rule of the same\nafternoon stands (AD7): no machine reading of whether a link is a reason may write, hold, retire or\nweigh it.\n\n**9.3 The reader decides what is needed, and the writer never does.** Prune by leaving out. Names\nare the exception: they are kept as their own part and settled once for a pair of texts (AD9),\nsince a reader of the writer's family joins them silently.\n\n**9.4 The checks sort a link four ways, with no verdict on truth.** A plain step. A step that rests\non a rule or a fact. A link the machine could only write around, which is drawn as a chain through\nthe sentence it wrote, with the first half owed a writing-out of its own. And a link that could not\nbe written out, which says so.\n\n**9.5 Each sentence is a question on the link, with a move beside it.** *Dispute it. It is stated\nhere.* It is the machine's sentence and nobody's claim. It weighs nothing until a text states it or\na person takes it up (AD11, unchanged). This follows his ruling of 25 August, and it follows what is\nmeasured about readers: a list of assumptions that people only read has no evidence behind it;\nhaving them act on an assumption cut over-trust most in one study of 214 people; a question helped\nmore than an explanation; and a wrong explanation left readers worse off than none.\n\n*2026-10-10, later: he ruled that a machine's sentence under a link is a claim from the moment it is written (\"The moment it's written. Go ahead and make it happen.\", ledger AG12). So where this says* nobody's claim*, read* a claim nobody asserts*. That it weighs nothing, and how it is shown, stand as written.*\n\n**9.6 Found before made.** His sentence of March: *\"finding or creating new premises that could one\nstep removed be allowing connections to other existing claims\"*. A sentence a text states becomes a\nlink to that claim, marked as found by a machine. A sentence one text states and the other denies\nis offered as the crux. A rule that comes back under many links is stored once, with its count\n(AD10). What is made and connects to nothing stays a potential on its link.\n\n**9.7 Depth by taking the rule in turn.** A rule is written out in its turn when it recurs or is\ndisputed, or when the reader showed the link to be a chain through it. No step is built on a\nmachine's step deeper than its measured error allows (the canon line on depth; ledger T1).\n⚠ *2026-10-10: this is the order of work for sentences the machine makes, and it was written as\nthe standard. His ruling of 25 August has the whole descent worked out eagerly, and ledger B2 makes\nseparability the floor:\n[the-chain-inside-a-link.md](the-chain-inside-a-link.md) § 1, § 4.3 (ledger AG2, unruled). That\nnote first added that below its first made level the machine only finds and asks. He ruled\notherwise the same day (ledger AG6): it may and should write deeper steps, and the small tasks that\nwrite them are quality-controlled like the others. What that control is, with a first trial of a\nstep written under one of this trial's nineteen kept rules:\n[quality-control-for-deeper-steps.md](quality-control-for-deeper-steps.md).*\n\n**9.8 A person reads the thirty written-out steps before any rate is claimed** (replaces AD12),\nwith four answers taken from the rubric on which raters agreed best in the literature: enough;\nsomething is missing; it says a claim over again; this is not what the link rests on.\n\n**What this bears on, and does not decide.** His approved list of 16 August has an open item that\nis his: whether to build a second pass for background premises. The rules under links *are*\nbackground premises, reached one link at a time and counted by how often they return. That is a\nthird way of getting them, beside the two the item names.\n\n**The same pattern, not yet tried on other unwritten things**: what a single claim takes for\ngranted; a conclusion a text never states; what an attack link rests on. Only support links between\ntexts were sampled.\n\n## 10. Risks and limits\n\n- **One family of models, thirty links, one afternoon. No person has read a written-out step.**\n  Every count above is a model reading a model.\n- **A complete step is not a right one.** Some premise always completes a step. Completeness and\n  steadiness are what was checked.\n- **Same family on both sides.** The reader shares the writer's blind spots; § 8b has one measured.\n  A second family was not available without spending the keeper's money. On one 2026 measurement\n  nine models from seven makers carried about two independent votes, so a second family would buy\n  less than it seems.\n- **A written sentence can mislead.** A machine's premise can make a weak step look reasoned, and\n  in one study raters accepted a model's warrants more often than the dataset's own. Hence\n  questions, the mark, no weight, and a person's reading first.\n- **Writing the unsaid is not neutral.** The corpus's own note on the unsaid says that turning what\n  goes without saying into something arguable is a political act, and that some things are left\n  unsaid on purpose. A written-out rule is offered. It is never put in an author's mouth.\n- **Cost.** Seventeen calls a link is too much to run over the graph today. Untried: a cheaper\n  reader for the closed question, which is the kind of question that read the same on the cheapest\n  models in this project's earlier measurements; and spending first where a link is contested.\n- **The four-way choice may leak.** A reader can pick a claim by its wording. Leaving out the first\n  claim and leaving out the rule are the controls for that, and they are why the over-full steps\n  were seen.\n\n## 11. Proposals (all unruled; ledger AD13 to AD21)\n\n| | Proposal |\n|---|---|\n| AD13 | **Every link is written out in one form: the first claim, what ties this case to a rule (names, and any fact no claim states), and one general rule.** A standing property with a debt, starting with the 871 links the linking step drew. Replaces AD6's *up to three further claims*. |\n| AD14 | **The writing is checked and the link is not judged**: by code (the rule is general; what a sentence repeats), by a reader not told a link was drawn (in both orders; a disagreement goes to a person), by writing twice, and by looking for each sentence in the texts. |\n| AD15 | **What a step needs is decided by leaving parts out before a blind reader, never by the writer.** Joinings of names are exempt and kept: a reader of the writer's family joins them silently. ⚠ *Refined by AD26: every small set that is enough, not the one a fixed order finds.* |\n| AD16 | **The checks sort a link four ways and say nothing on truth**: plain; resting on a rule or fact; written around (a chain through the machine's sentence, its first half owed a step of its own); not written out. ⚠ *Counts corrected by AD26: 17, 5, 2 unsettled, 6.* |\n| AD17 | **Each sentence of a step is shown as a question on the link with a move beside it, is nobody's claim, and weighs nothing until a text states it or a person takes it up.** Never a list to be read. |\n| AD18 | **Found before made**: a sentence a text states becomes a link to that claim; one text stating and the other denying is offered as the crux; a recurring rule is stored once with its count; the rest stay potentials. |\n| AD19 | **Depth comes from taking a rule in its turn** (when it recurs, is disputed, or carries a chain), capped by measured error. |\n| AD20 | **A person reads the thirty written-out steps before any rate is claimed**, with four answers: enough; something missing; says a claim over again; not what the link rests on. Replaces AD12. |\n| AD21 | **Correction: *never generating premises, only finding real ones* was Claude's sentence of 2026-03-29 and never his position.** Its ground, that judging a link is accurate and writing a premise is not, is what this day's measurements contradict for links. |\n\n**Considered, and not proposed.**\n\n- *Formal logic and a solver for the check.* Two 2026 systems translate each sentence to logic and\n  let a solver say whether the step is valid, which makes that one check exact. The judgement moves\n  into the translation, defeasible steps do not sit well in it, and a study of such a pipeline found\n  the model writing the conclusion in as an axiom in over half of its fabrications. A candidate for\n  later, on the steps that pass the blind reader.\n- *A gate that asks first whether the link is a reason.* One system does this before\n  reconstructing. Here four ways of asking were unsteady in one day.\n- *Voting over several writings.* It works where a step has one exact answer. A rule does not.\n- *Naming only* (AD6 as it stood). His answer is the occasion of this note.\n- *Writing out to the bottom in one go.* Error grows with every step built on a machine's step.\n\n**Added 2026-10-10.** Two later proposals bear on this design, both unruled: let a group of links\nprompt a rule, which starts no wider than they need (ledger AF1), and test the pass that writes steps on\nrules that texts state themselves (ledger AF4). Both come from reading how machine-written proofs\nare tidied for a mathematics library:\n[alien-proofs-and-the-library.md](alien-proofs-and-the-library.md) § 4.6, § 4.7.\n\n## 12. Corrected with this note\n\n*Never generating premises, only finding real ones*, and the heading *Auto-CONNECT, not\nauto-decompose*, stood in five places as the settled design or beside his quote as if it were his.\nEach now says whose sentence it is and points here:\n[scheme-bounded-decomposition-and-evidence-as-subgraph.md](scheme-bounded-decomposition-and-evidence-as-subgraph.md)\n(where it was written), [the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md) § 1,\nthe founder-voice map in `what-counts-as-evidence/`, [conceptual-threads.md](../conceptual-threads.md),\nand two lines of `CLAUDE.md`. `README.md` dated his words to April; they are of 29 March.\n\nThis afternoon's own note is corrected in three places: its proposal to store *up to three further\nclaims* (replaced, § 9.1); its § 4.8, which left an author's own steps unwritten; and its reading of\nthe joined names as unwritten, when the paper's own claims state the joining (§ 8c).\n\n## 13. What was read, and how far\n\n**In the corpus, by Claude, this session**: the two notes of the day; the March document with his\nquestion and the sentence corrected here; the record of his intentions since 2011; the note on the\nunsaid; the notes on small tasks, on proving a pass, on machine work as a standing property, on a\nlink's two factors, on latent questions, on the trust ladder, on the reuse flywheel (sections); the\ncode of the two premise passes, the patterns file and the sameness check. **By a helper**: the whole\nrepository and the archive of 2010 to 2017, with his words kept apart from every restatement (its\nreport is kept privately, because it quotes other people by name). Every line of § 1 was then\nchecked at its source by Claude, the two drawings by eye.\n\n**Outside, on 2026-10-07, by two helpers** (models), each report kept as it came back with its own\nstatement of how far every source was read:\n[writing out with checks](writing-out-the-unwritten/outside-reading-explication-with-checks.md) and\n[dependable small tasks](writing-out-the-unwritten/outside-reading-dependable-small-tasks.md). The\nafternoon's two reports are in [the earlier folder](what-a-link-rests-on/). **Nothing in them was\nchecked against the sources by Claude or by a person, except the paper of § 4, whose method and\nlimits sections Claude read. Verify before citing outward.**\n\n| Used here | Source, as the helper cites it | The helper's read depth |\n|---|---|---|\n| Steps unchecked; the three guards; no rate for agreeing and wrong | Li et al., arXiv:2510.26854 v3 | main text; Claude read the method and limits |\n| A right answer does not show sound steps | two 2026 preprints | abstract; extract |\n| The rule as a covering generalization; four steps; what a gap-filler must not do | Hitchcock 1985, 2011, 2019, and an undated reply | main text; abstract and extracts |\n| No program implements it | the helper's search | not found, which is weaker than absent |\n| *Maximally charitable*; agreement between reconstructions as a signal | Betz & Richardson, *SEM 2022 | main text |\n| A loop of write, check, rewrite: 2.9 to 4.5 of 5; trivially true conditionals | Ryu et al., arXiv:2509.21679 | main text |\n| A gate first; a solver prunes; 207 of 209 translations right | Ryu et al., arXiv:2603.17432 | main text |\n| The conclusion written in as an axiom, 56% | Poiroux et al., a 2026 workshop paper | main text |\n| Same missing premise in 69 of 71 within one narrow pattern | Razuvayevskaya & Teufel 2017 | results |\n| A step with only its own premises: about half of wrong steps let through | Ling et al., NeurIPS 2023 | full text; a 2023 model |\n| Four-way rubric, kappa .85 | Hong et al., Findings of NAACL 2022 | manual evaluation |\n| 79%, 18%, near 0% by number of steps | Dalvi et al., EMNLP 2021 | main text; strict scoring; an earlier generation |\n| Order of premises; telling a model to ignore order fails; both orders and people for the rest | Chen et al., ICML 2024; Wang et al., ACL 2024; Zheng et al., ICLR 2024 | full text |\n| A checker not shown the draft does better; open questions beat yes or no | Dhuliawala et al., Findings of ACL 2024 | full text, in part |\n| A million steps without an error, for exact answers | Meyerson et al., arXiv:2511.09030 | full text, in part |\n| Nine judges, about two votes | Kohli, arXiv:2605.29800 | first part |\n| Verify caught half and removed good work; checkable questions first | Bernstein et al., UIST 2010; Kittur et al., CHI 2008 | full text, in part; extract |\n| Readers: acting on an assumption; questions over explanations; a wrong explanation worse than none | Ghosh et al., arXiv:2601.18033; Danry et al., CHI 2023; Si et al., NAACL 2024 | abstract and extracts |\n\n**The scripts and their outputs** are in [the folder](writing-out-the-unwritten/): the predictions\nwith their two dated additions, the trial, the pruning, both scorers, the rebuild from the record,\nthe pilot's answers on the first form, and what the scorers printed.\n\n## Cross-references\n\n[what-a-link-rests-on.md](what-a-link-rests-on.md) and\n[how-far-a-premise-is-from-its-claim.md](how-far-a-premise-is-from-its-claim.md) (the day's two\nearlier notes) ·\n[scheme-bounded-decomposition-and-evidence-as-subgraph.md](scheme-bounded-decomposition-and-evidence-as-subgraph.md)\n(his question of March) ·\n[split-review-and-the-founders-intentions.md](split-review-and-the-founders-intentions.md) (his\nrecord on machine work) · [the-load-bearing-unsaid.md](the-load-bearing-unsaid.md) ·\n[kartpaus-auto-enrichment-design.md](kartpaus-auto-enrichment-design.md) (the direction ruled on 25\nAugust) · [latent-scaffolding-and-the-type-token-split.md](latent-scaffolding-and-the-type-token-split.md)\n(potentials) · [relevance-and-the-two-factors.md](relevance-and-the-two-factors.md) (a link's\nweight) · [microtasks-in-broad-daylight.md](microtasks-in-broad-daylight.md) and\n[proving-a-pass-reliable.md](proving-a-pass-reliable.md) (small tasks and proof) ·\n[machine-work-as-a-standing-property.md](machine-work-as-a-standing-property.md) (the debt) ·\n[the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md) ·\n[prime-real-estate-words.md](prime-real-estate-words.md) (names people want) · the walkthrough:\n[2026-10-07-writing-out-the-unwritten.md](../listening/2026-10-07-writing-out-the-unwritten.md).\n"}