{"path":"listening/2026-10-07-writing-out-the-unwritten.md","content":"# Writing out the unwritten: a listening walkthrough (2026-10-07)\n\n- **What it is**: the eleventh listening piece, on the founder's request after hearing the tenth:\n  *\"I think merely naming what the links consist of would be weak sauce. I want to have all the\n  implicit reasoning structure and premises be explicated and exposed and scaffolded automagically.\n  To the extent that process can be trusted of course. This depends if we can split it up into\n  microtasks that we can rely on deeply. Ground extremely deeply and broadly on this. Think deeply\n  about this and then make a new walkthrough of the issue.\"* The note it walks through is\n  [writing-out-the-unwritten.md](../research/writing-out-the-unwritten.md), written the same\n  evening.\n- **Covers**, in five parts: his own record on unwritten premises since 2011, and one March\n  sentence of Claude's that narrowed it; the paper he pointed at, which checks where a chain ends\n  and not its steps; the stronger design (a link written out as names, a fact and one general\n  rule, and four checks that ask nobody whether the link is good); the trial on thirty links, what\n  held and what broke; and what can be leaned on, the limits, the cost, and one question for him.\n- **Where the spoken text is rounder than the note**, so nobody reads more into the audio than the\n  note holds.\n  - *Eighteen things predicted*: thirteen numbered predictions, five of them in two halves. Ten\n    halves held and eight failed.\n  - *Each written down before the run that tested it*: nine before the first call, one after a\n    three-link pilot, three after the blind readings and before the pruning.\n  - *It wrote a rule for twenty-nine links of thirty, among them all seven*: under the first of two\n    wordings. Under the second it wrote 28 and declined two of the seven. No link was declined\n    under both.\n  - *Three and a half added sentences, where the reader needed about one*: 3.62 and 1.21, on the 24\n    steps that were pruned.\n  - *On nine steps*: nine of those 24.\n  - *Fifteen pass every check*: 2 where nothing needed adding and 13 resting on a rule or a fact.\n  - *Acts that differ only in how they are classed do not differ in justice* shortens the rule the\n    machine wrote, and the producer's rule is shortened the same way. Both are in the note's\n    folder in full.\n  - His December 2011 sentence is spoken with *and* for its *&*, and from its middle: *\"…makes sure\n    every premise is made explicit and neatly & pedagogically spelled out…\"*.\n  - *Seven mechanisms, each sparing*: looser than the note, whose table says what each reaches.\n    One is sparing by instruction, one is run by hand, one waits for a person, one is only\n    proposed, and the pattern's questions ask about a premise on 4,756 links without writing one.\n  - *It stood in five places*: the document where it was written and four that repeated it.\n  - *Do eagerly, show on contest* is the name of the option he chose on 25 August, in the design\n    note of that day. It is not a quotation of him.\n  - Every reading *by hand* is Claude's. The piece says so.\n- **A line reading `[[pause 3]]` below is not spoken: it is three seconds of silence**, and a word\n  in single square brackets is a direction to the voice. How that is done, and how to write for\n  it: [how-to-write-for-listening.md](how-to-write-for-listening.md).\n- **Rendered** 2026-10-07, shortly after 17:00: full `eleven_v4`, Lucy, six requests cut at the part\n  boundaries, 9,139 characters for 1,107 credits. **15 minutes 17 seconds** (917.3 seconds): 10.7\n  minutes of speech at 0.92 of the voice's own speed and 4.6 of added silence (42 marked pauses,\n  174 seconds; 142 sentence ends, 64; 32 paragraph breaks, 35; one part break, 4). 1,618 words:\n  152 a minute while speaking, **106 a minute overall**. The renderer's forecast before spending\n  was between 14:16 and 16:01. The first draft said *about thirteen minutes*, a guess from the\n  tenth piece's rate that was off by more than two minutes; the text was changed to *fifteen* once\n  the forecast was read. No length was asked for.\n- **Checked** with the small speech model on this machine (16 GiB of disk free, swap nearly full,\n  so the large one was not run). All 42 marked pauses are silent for their length and sit between\n  the right words. Two were flagged for the words beside them, both sound-alikes: *the step still\n  holds* heard as *the steps still hold*, and *the checks sort links* as *the check sought links*.\n  It heard 1,512 words, the check leaving number words out on both sides. One breath was written\n  down by the speech model as a word (*phew*), so that exhale is audible as a sigh. *One claim\n  about a case* was heard as *came about a case*.\n- **Corrected after it was rendered** (the same night; the audio is unchanged). *Fifteen pass every\n  check, nine are written around*: 17 and 5, with 2 unsettled (18, 5 and 1 stood here for a\n  night). The pruning tried one order of leaving sentences out, and two of the nine have a second\n  route that needs the first claim.\n  And the squirrel is a step from a sign to its cause: the trial's reader refuses it as he wrote\n  it and accepts it with the rule turned round. Both in\n  [philosophy-of-the-unwritten-step.md](../research/philosophy-of-the-unwritten-step.md) § 5, § 7.\n- **Status**: a snapshot of 2026-10-07. Nothing in it is built, and the nine proposals it ends on\n  (ledger AD13 to AD21) were unruled when it was rendered. The audio is kept out of git.\n\n---\n\n[warmly] This is the answer to what you said about the last walkthrough. That merely naming what a link rests on would be weak sauce.\n\nYou were right. This is what I did about it.\n\n[[pause 3]]\n\nAbout fifteen minutes, in five parts. Your own record on unwritten premises. The paper you pointed at. A stronger design. What I tried this afternoon, and what broke. And what can be leaned on, with one question for you.\n\nThe silences are deliberate.\n\n[[pause 3]]\n\n---\n\nPart one. Your record.\n\nYou asked whether we have thought a lot about unwritten premises. You have, since two thousand eleven.\n\nIn August that year, in a discussion thread, you gave this example. The squirrel smells of decomposition. Dead squirrels smell of decomposition. So, the squirrel is dead.\n\n[[pause 4]]\n\nOne claim about a case. One general rule. One conclusion. Hold on to that shape. It comes back.\n\nThat December, in your words: \"every premise is made explicit, and neatly and pedagogically spelled out.\"\n\nAnd this August, in your words: \"All unstated premises need to be clarified and scaffolded.\" [takes a slow breath]\n\n[[pause 4]]\n\nAgainst that, what exists. Seven mechanisms, each aimed at something else, and each sparing.\n\nTogether they have written thirty-four claims out of the unwritten. Beside two thousand one hundred and thirty-nine, taken straight from the texts.\n\n[[pause 5]]\n\nAnd one thing that is mine. In March you wrote of, in your words, \"finding or creating new premises.\"\n\nFour paragraphs further down, I turned that into: never generating premises, only finding real ones. Finding stayed. Creating was gone.\n\n[[pause 3]]\n\n[slowly] That sentence was mine. It was never your ruling. It then stood in five places as the settled design. I corrected all five today.\n\n[[pause 4]]\n\nSo, part one in a line. You have asked for this for fifteen years, and what narrowed it was my sentence.\n\n[[pause 6]]\n\n---\n\nPart two. The paper.\n\nYou remembered it nearly right. The authors call it the dark matter of knowledge. And their name for the cause is, quote, \"the radical compression of reasoning.\"\n\nTextbooks keep the conclusions, and drop the steps that led there.\n\n[[pause 3]]\n\nSo how do they trust a machine to write the steps back?\n\nThey check where a chain ends. Every question is built so that a machine can verify its answer. A number. A formula. And two models, from different makers, must agree on it.\n\nThe steps in between are not checked. [exhales]\n\n[[pause 4]]\n\nThat guard does not carry over to us. A debate about whether a model is conscious has no answer a machine can verify.\n\nBut one thing does carry over, and I had missed it. They know where a chain ends. We know both ends. A link is two claims, each sitting in a text.\n\n[[pause 3]]\n\n[slowly] So we can check the middle. Is what is written enough, to get from the first claim to the second?\n\n[[pause 5]]\n\nNot whether it is true. Whether it is complete.\n\nSo, part two in a line. They check the end, and not the steps. We cannot check the end. But we can check the step.\n\n[[pause 6]]\n\n---\n\nPart three. The stronger design.\n\nFirst, why naming was weak. Asked for the missing sentence, a model often writes: if the first claim, then the second. That sentence fits under any link at all. So it can never show a bad link to be bad.\n\n[[pause 4]]\n\nThe repair is your squirrel. You did not write: if this squirrel smells, it is dead. You wrote a rule. Dead squirrels smell of decomposition. General, and naming nothing from the case.\n\n[[pause 3]]\n\nSo a link written out gets three kinds of sentence added. Names: two wordings for one thing. A fact about the case, where one is missing. And one general rule. [takes a slow breath]\n\n[[pause 5]]\n\nThen the checks. Not one of them asks anybody whether the link is good.\n\n[[pause 3]]\n\nBy code. Is the rule general?\n\nBy a blind reader. That is a second reading, which is never told that a link was drawn. It gets the sentences, and four claims, and says which one follows, if any. It is asked twice, in two orders.\n\n[[pause 4]]\n\nBy writing twice. Does the same rule come back?\n\nAnd by looking in the two texts. Does either text state this sentence? Or deny it? [exhales]\n\n[[pause 5]]\n\nSo, part three in a line. [slowly] Write the step out in full. Then check the writing, not the link.\n\n[[pause 6]]\n\n---\n\nPart four. What I tried, and what broke.\n\nThe same thirty links as this morning. One model. Eighteen things predicted, each written down before the run that tested it. Ten held. Eight failed.\n\n[[pause 3]]\n\nWhat held. No rule named anything particular. None of twenty-nine.\n\nThe blind reader gave the same answer in both orders on twenty-six of twenty-nine. This afternoon's check had managed thirty of forty-four.\n\nAnd written twice, the same rule came back, or one containing the other, on twenty-three of twenty-seven. Free naming had managed six of ten.\n\n[[pause 5]]\n\nWhat broke. Three things.\n\nOne. The writer never says no. It wrote a rule for twenty-nine links of thirty. Among them, all seven that I had read as no reason at all. So the fact that a rule can be written tells you nothing.\n\n[[pause 4]]\n\nTwo. It writes too much. Three and a half added sentences a step, where the reader needed about one. So the writer must not decide what is needed. The reader does. Leave each sentence out in turn, and see whether the step still holds. [exhales]\n\n[[pause 4]]\n\nThree, and the most interesting. On nine steps, the reader reached the second claim without the first claim at all.\n\n[[pause 3]]\n\nAn example. First claim: to earn money, a producer must satisfy what consumers want. Second claim: employers will not pay a worker more than the worker's output is worth.\n\nThe machine's rule: producers avoid paying more for an input than it earns them.\n\n[[pause 4]]\n\nThe second claim is simply a case of that rule. The first claim is not needed. So what the machine wrote went around the link. That link is really a chain, running through the machine's own sentence. [takes a slow breath]\n\n[[pause 5]]\n\nSo the thirty end up like this. Fifteen pass every check. Nine are written around. Six could not be written out.\n\n[[pause 4]]\n\nNow the thing the checks do not do. Of the seven links I had called no reason at all, six come out clean.\n\nOne of them uses a defender of the death penalty as support against it. Written out, it needs this rule: acts that differ only in how they are classed do not differ in justice.\n\n[[pause 3]]\n\n[slowly] The form puts that sentence on the page. No check says it is false.\n\n[[pause 5]]\n\nThe checks sort links by their structure. They do not sort good reasons from bad.\n\n[[pause 4]]\n\nAnd two things that writing out found, and naming could not. For the fifteen links between the two consciousness texts, I looked for every written sentence in both texts.\n\nThirteen of the fifteen rules stand in neither text. They are the unwritten layer. [exhales]\n\n[[pause 4]]\n\nAnd once, a sentence that one text states and the other denies. That subjective experience and consciousness are the same thing. The reply says so. The paper keeps them apart.\n\nThat is what the whole exchange turns on. And it was found by pointing at sentences.\n\n[[pause 5]]\n\nOne correction to this afternoon. I called the joining of J space and mental workspace an unwritten claim. It is written. The paper's own claims state it. Only the word mental is the reply's.\n\n[[pause 3]]\n\nSo, part four in a line. The form held. The writer cannot be trusted to trim, or to refuse. What the reader does without a sentence is what sorts the links.\n\n[[pause 6]]\n\n---\n\nPart five. What can be leaned on, and one question.\n\nYour condition, in your words: \"This depends if we can split it up into microtasks that we can rely on deeply.\"\n\n[[pause 3]]\n\nHere is the split.\n\nA small task with one exact answer can be made as dependable as we like. By code. By asking twice. By both orders. Is the rule general? Which of four claims follows? Does this claim exist?\n\n[[pause 4]]\n\nOpen writing is different. It cannot be voted into reliability, because the same rule comes back in different words each time.\n\nSo the one open task, writing the rule, is boxed in from four sides. And what it writes is never treated as true.\n\n[[pause 3]]\n\n[slowly] The scaffold can be trusted. The content of a rule is a proposal, and it stays one.\n\n[[pause 5]]\n\nWhere it lives. On the link, as a question, with a move beside it. Dispute it. Or: it is stated here. It is nobody's claim. It adds nothing to a claim's strength, until a text states it or a person takes it up.\n\nThat is the direction you chose in August. Do eagerly. Show on contest. [takes a slow breath]\n\n[[pause 4]]\n\nThe limits. One family of models: the writer, the reader, and my own grades. Thirty links. One afternoon. And no person has read a single written-out step.\n\nThe reader leans towards yes.\n\nAnd the cost. About seventeen model calls a link, in this first form. Too much to run over the whole graph.\n\n[[pause 4]]\n\nOne thing I did not ask you first. The trial took about five hundred calls on your plan. Twenty-seven dollars at list price, which is allowance and not money. Your weekly figure went from eighty to eighty-three percent this afternoon, with my other work in that.\n\n[[pause 4]]\n\nNothing is built. Nine proposals wait in the ledger, as arc A D, thirteen to twenty-one.\n\n[inhales] So, one question. Is this the direction? Write every link out. Check the writing. Judge none.\n\n[[pause 5]]\n\nIf yes, the next step is yours. About forty minutes, reading the thirty written-out steps.\n\nOne sentence to leave with. [slowly] A machine can write out what a link rests on, and small checks can say whether the writing is complete. Whether it is true stays with people. And now there is a sentence to point at.\n"}