{"path":"listening/what-eighteen-pieces-taught-2026-10-18.md","content":"# What eighteen listening pieces taught, and how to sharpen the method (2026-10-18)\n\n- **What it is**: an audit of every audio walkthrough made between 2026-09-26 and 2026-10-15 (fourteen\n  public under `docs/listening/`, four private, the eighteenth counted twice for its remake), measured\n  from their texts and their rendered files, read in full for the seven made after the founder's deeper\n  feedback of 2026-10-07/08. His request: *\"Let's analyze all the audio walkthroughs that have been\n  generated. Especially those after I gave some deeper feedback on pacing and content etc, a while\n  back. I have some intuitions and opinions on this but I'll wait for your analysis first.\"* Then:\n  *\"What's your take on what to do about all the findings, after reflection? All the findings, even\n  the little ones. How do we sharpen the instructions for how to make audio walkthroughs that are\n  genuinely useful? First give me your own takes, persist them. Then: Research this extensively\n  online. Enrich your auxiliary note with the research.\"*\n- **Status**: a session's reading and proposals, **unruled**. § 1 to § 4 are Claude's own takes, written\n  before the research in § 5 was read, as he asked. The method itself is\n  [how-to-write-for-listening.md](how-to-write-for-listening.md); rule numbers below refer to it.\n- **Covers**: the measurements · what held and what drifted · my takes on each finding, the small ones\n  included · the proposed sharpening of the method, as draft rules · the research, and what it changed.\n\n---\n\n## 1. The measurements\n\nCounted from the texts with `scripts/listening_stretches.py` and two scratch scripts (words, sentences,\npauses, breaths, number words, example cues, his-words cues, self-corrections, landscape closings);\ndurations from the mp3 files. \"After\" means the pieces made after his *sports narrator* message of\n2026-10-07 23:00.\n\n| Piece | Date | Words | Minutes | Median / 90th pct sentence | Pauses | Longest stretch (words) | Number words /1k | \"In your words\" cues | Self-corrections |\n|---|---|---|---|---|---|---|---|---|---|\n| Arc P audited (template) | 09-26 | 4,389 | — | 12 / 25 | 0 | — | 33 | 9 | 0 |\n| What counts as evidence | 09-26 | 9,080 | — | 11 / 24 | 0 | — | 43 | 5 | 0 |\n| Decisions waiting (private) | 09-27 | 5,419 | — | 12 / 25 | 0 | — | 37 | 12 | 0 |\n| Major transitions (private) | 09-29 | 6,268 | 6.8 | 11 / 28 | 0 | — | 34 | 4 | 0 |\n| Proving reliability and Bend | 09-29 | 2,610 | 14.6 | 11 / 24 | 0 | — | 68 | 3 | 0 |\n| Microtasks in broad daylight | 09-29 | 2,321 | 13.1 | 13 / 29 | 0 | — | 62 | 4 | 0 |\n| Six days (private, first) | 10-05 | 7,641 | 41.4 | 11 / 24 | 0 | — | 55 | 25 | 0 |\n| The evals loop | 10-05 | 3,341 | 24.2 | 8 / 16 | 40 | 233 | 58 | 1 | 0 |\n| Six days remade (private) | 10-06 | 8,292 | 68.5 | 9 / 18 | 179 | 154 | 57 | 24 | 0 |\n| Wanted words | 10-07 | 1,690 | 15.2 | 6 / 14 | 40 | 87 | 62 | 3 | 0 |\n| What a link rests on | 10-07 | 703 | 7.0 | 7 / 14 | 18 | 82 | 80 | 3 | 0 |\n| Writing out the unwritten | 10-07 | 1,632 | 15.3 | 6 / 14 | 43 | 62 | 73 | 6 | 0 |\n| **After the feedback** | | | | | | | | | |\n| Eight proposals as ideas (private) | 10-08 | 5,137 | 45.2 | 7 / 14 | 120 | 69 | 33 | 2 | 1 |\n| Eight proposals pondered again | 10-08 | 4,589 | 42.5 | 7 / 13 | 142 | 57 | 34 | 7 | 1 |\n| What our own notes already knew | 10-08 | 2,567 | 24.3 | 6 / 12 | 80 | 56 | 29 | 7 | 7 |\n| The chain folded inside an arrow | 10-10 | 1,017 | 9.2 | 7 / 13 | 28 | 55 | 30 | 4 | 0 |\n| Three days to the floor | 10-13 | 6,706 | 56.7 | 8 / 16 | 179 | 60 | 36 | 38 | 7 |\n| Parts, reasons and the number | 10-14 | 5,921 | 50.3 | 9 / 21 | 160 | 58 | 46 | 7 | 5 |\n| The ropes and the coin | 10-15 | 5,521 | 45.7 | 10 / 22 | 107 | 117 | 45 | 9 | 0 |\n\nOverall pace (words a minute including silence): 138 on 2026-10-05, 101 to 111 on 2026-10-07, 106 to\n114 on 2026-10-08 and 10, then 118, 118 and 121 on 2026-10-13 to 15. Silence as a share of the file:\nabout 18% on 10-05, 24% to 27% from 10-08 on. Marked pauses: one every 19 to 26 seconds in every piece\nsince 10-08.\n\nExample cues per thousand words (*think of · picture · the first example · a kitchen · a court* and\nthe like): 5.5 and 5.9 in the two eight-proposals pieces, 5.5 in *what our own notes already knew*, 4.9\nin *the chain*, then 2.1 in *three days*, 0.7 in *parts and reasons*, 2.9 in *the ropes*. Closings of\nthe form *what is real today / what would it open*: 20, 22, 15, 7, 19, 22, 9. Registers beyond\nphilosophy and our own graph (a borrowed voice, play with an interface, an irreverent view): present\nin the twelfth and thirteenth pieces, absent from the sixteenth to eighteenth.\n\nWhat he said about form after 2026-10-08: nothing. The method marks every piece from the fourteenth\nto the seventeenth *not yet commented on*, and this session searched his messages since 2026-10-05 for\nany report (the corpus graph and the session transcripts, by the words *lit up · elder · too long ·\ntoo fast · the walkthrough*): the one reaction of form is 2026-10-08 14:39, *\"Nah I'll skip it\nactually\"*, on the remade eight-proposals piece, followed by a request for a timestamp to scrub to.\nThe chapter list he asked for on 2026-10-08 20:35 (*\"a bullet list of chapters/parts … with 10-word\ntitles and 14-word subtitles\"*) became a standing rule on 2026-10-15.\n\n## 2. What held and what drifted\n\n**Held.** The pacing fix. Silence runs at a quarter of every file since 10-08, pauses land where they\nwere written (the check confirmed 177 of 179, 156 of 160, 106 of 107), breaths sit on their side, the\nforecast and the length gate stop a piece that would overrun. Nothing in the rendering layer has\nregressed. The spoken headers (*where the spoken text is rounder than the notes*) are a genuine honesty\ninstrument and were kept in every piece.\n\n**Drifted, five ways.**\n\n1. **The text under the pacing got denser again.** Median sentence 6 to 7 words right after the\n   feedback, 9 to 10 in the last two pieces; the 90th percentile from 13 to 14 up to 21 to 22, which\n   is where the pre-feedback pieces sat (24 to 28); the longest unpaused stretch in the eighteenth is\n   117 words, twice the mark the eleventh piece set. The pacing rules held because the renderer\n   enforces them. The sentence rules (rule 14, *most sentences under twenty words, median eight*)\n   held for a week because a session remembered them, and then did not.\n2. **The scoreboard came back as tallies in words.** Number-word density halved after the feedback\n   (58 to 80 per thousand down to 29 to 34) and has climbed since (36, 46, 45). Some of the numbers\n   are the idea (*eight in ten*, Keynes's coin). Many are tallies the ear cannot hold: *twenty-four of\n   the reasons*, *about a quarter of the substantive claims*, *one message in six*, *two thousand\n   standing questions*, *fifty-nine percent*. Rule 25 asks for a judgement (*only when the number is\n   the idea*) and a judgement-gated rule does not fire while writing.\n3. **The pieces became about the session rather than the land.** The hour-long catch-up reads his own\n   chat messages back to him 38 times, about once every ninety seconds, most of them his questions\n   rather than his rulings. It announces *I was wrong about twenty times* in its opening and then\n   says each. About 27% of it (parts three to five, 1,786 of 6,706 words) is readers, hooks and the\n   corpus-graph tooling, three days after he had asked for *\"actual ontology/core stuff, not the\n   corpus graph tooling\"*. This is the sports narrator in a new costume: confessions and quotations\n   where the counts used to be, and the same machinery-on-stage.\n4. **The content rules became a ritual.** *What is real today? … What would it open?* closes nearly\n   every part of every piece, in the same two sentences, 19 to 22 times an hour. The two-examples\n   rule held strongly in the twelfth to fourteenth pieces and faded in the seventeenth and in the\n   back half of the eighteenth, whose part nine runs 906 words on one example. The palette of\n   registers (rule 27: a borrowed voice, play with an interface question, the irreverent view, an\n   elder's voice) was rich in the twelfth, thin in the thirteenth, and gone by the sixteenth. What\n   survived of the 2026-10-08 rules is their most mechanical half.\n5. **The endings stopped being carryable.** The twelfth ended on three questions. The eighteenth ends\n   on *eight rows, all open, all yours* and names them by their pictures, which is a chapter list\n   read aloud, not a question to turn over on a walk.\n\n**And the loop that would have caught all five is open.** The twelfth piece fit his brief best and\nwas never reported on. The method says *his report on the twelfth piece decides* the content rules,\nand no report came, so the rules stayed unruled and drifted with no signal. His one form reaction in\nten days was skipping a piece. The instrument that would give him a cheap way to report, the chapter\nlist, was asked for on 8 October and made standard on 15 October.\n\n## 3. My takes, finding by finding\n\nWritten before the research. Each take names what I would change and what it costs.\n\n**On density (1): make the sentence rules gates, as the pacing rules already are.** The pacing held\nbecause `render_listening.py` computes it; the sentence rules drifted because nothing does. The dry run\nshould print median and 90th-percentile sentence length, the longest stretch, and words a minute\nplanned, and refuse to render past a ceiling (median over 9, 90th percentile over 16, a stretch over\n60, a planned pace over 115) unless `--spend-anyway` is passed (`--despite-gates` since 2026-10-20), exactly as the length gate works since\n2026-10-08. `listening_stretches.py` already counts stretches and was run on some pieces, not all; a\ncount that depends on remembering to run it is the same failure as the rule. Cost: ten lines in the\nrenderer, and a draft that overruns gets cut, which is rule 19 made mechanical.\n\n**On the scoreboard (2): change rule 25 from a judgement to a count with a justification list.** Keep\nthe ceiling where the pieces that pleased him sat, about 30 number words per thousand, and have the\ndry run print every number word with its sentence. The writer then strikes each one that is a tally,\nand the header's *where the spoken text is rounder* list gains a sibling, *the numbers kept, and why\neach is the idea*. His chat style says *numbers in place of adjectives*; the method says the ear is\nthe exception, and the exception needs a counter or the chat habit wins, which is what happened.\n\n**On the session on stage (3): three cuts.** First, his own words are spoken only where rule 9 meant\nthem: a ruling, a correction he made, a sentence the piece builds on. His questions are not read back\nto him; he asked them. Second, a correction is spoken only where the listener holds the old belief,\nwhich means he heard it in an earlier piece or he said it himself; every other correction goes to the\nheader, where it already lives. The opening *I was wrong about twenty times* is honesty in the wrong\nform: it makes the piece about my errors and spends his attention on bookkeeping. Third, tooling and\nprocess (readers, hooks, the corpus graph, what a script does) get at most one part, and only when he\nasked for it; otherwise they are a sentence in the view from here. He said this on 2026-10-10 and the\nsixteenth piece did the opposite three days later, so it needs to be a rule with a count behind it,\nnot a remembered request: words about the session's own process over total words, and the dry run\nprints it.\n\n**On the ritual (4): keep the substance of the three content rules, drop the fixed form.** The\nlandscape closing (rule 30) is right once, in the view from here, and in a part only where what is real\ntoday would surprise him; as two fixed sentences at the end of every part it has become a tic the ear\nskips. The two-examples rule (26) is the most load-bearing of the 2026-10-08 rules and the one that\ndecayed first, so it gets the count: example cues per part, printed by the dry run, and a part with\nfewer than two is flagged. The palette (27) needs a lighter instrument than a count: the header names\nthe registers the piece used, so three pieces running on *philosophy and our own graph* is visible on\nthe page. And the method should stop pointing at one model piece. The Arc P piece is the model for\nform; the twelfth, *eight proposals as ideas*, is the model for content register, and nothing since\nhas matched it. Name both.\n\n**On the endings (5): rule 13 gets a ceiling of three.** Three questions to carry, each a full\nsentence a walker can turn over. Everything beyond three goes to the chapter list and the ledger,\nwhich is where the eighteenth's eight rows belonged, and where its chapter list in fact put them.\n\n**On length: a default tied to the shape of the content, not to the request's ceiling.** A day's\nthinking is ten minutes (the fifteenth, which worked). A catch-up is thirty unless he asks for more.\nAn hour is for one deep subject with a map, as the seventeenth was. His *\"it can be an hour long\"* is\na ceiling and the method has been reading it as a target: the sixteenth and eighteenth filled the\nhour and the one piece he skipped was a 45-minute remake. Two hour-long pieces a week is also about a\nquarter of the month's voice credits.\n\n**On the loop: build the report channel, since the chapter list alone is a form with nobody\ncollecting it.** Three parts. The method's per-piece paragraphs replace *not yet commented on* with a\ntable of his reports, so a piece with none is visible as such rather than as a line nobody reads. The\ndelivery message ends with the chapter list and one line: *mark any chapter muddled, agreed, or\ndisagreed, as a voice memo or a word each*. And the first words of the next piece answer the last\nreport, so reporting visibly changes the next piece, which is the only incentive that will hold. No\nnagging beyond that; he reports when he reports, and the measurements in § 1 are what the method has\nin the meantime.\n\n**The little ones, each with its take.**\n\n- *The map read aloud at the start* lists ten to twelve part titles; nobody holds twelve. Three\n  groups at most, or the first three parts and *then the rest*.\n- *\"The silences are deliberate. They are room to think.\"* opened every piece since 10-05; he knows.\n  Drop it from the second piece on, keep it where a piece goes to a new listener.\n- *\"So, part N in a line\"* was dropped in the eighteenth. Keep a one-line closer per part; vary the\n  words.\n- *The breath count* swung from 43 in the sixteenth to 11 in the eighteenth with no reason recorded.\n  Hold at about one pause in four, as rule 20 says, and print the ratio.\n- *`[slowly]`* ran 22 in the thirteenth and 8 in the eighteenth; rule 21 says one or two a part.\n  Print it.\n- *Sound-alikes* (*witch/which*, *bike's brakes are parts*) and a breath rendered as *phew*: the\n  check already flags them; the method should say the writer reads every flag before sending, not\n  after.\n- *His quoted sentences with slashes and abbreviations* are spoken whole (the seventeenth's QEM,\n  QBAF). Right, and it should stay a rule: a quote with words left out is a paraphrase.\n- *The hour-long pieces' first part* re-tells the previous piece. A recap of one minute is right\n  when the gap is days; the sixteenth spent 293 words and two corrections on it. Cap the recap at a\n  minute.\n- *Cost*: an hour costs 16,000 voice credits (an eighth of the month) and about 8 million weighted\n  plan tokens. State both in the delivery line, so the length choice is made with the price visible.\n- *The method doc* is 9,400 words, with the rules in three numbered runs and three ruled sections. A\n  writer about to make a piece will not re-read it. The rules that must fire while writing belong in\n  a short checklist the dry run prints, and the long doc keeps the evidence, which is how the global\n  rules are split as well.\n\n**What I would not change.** Pause lengths, breath placement, `[slowly]`, the money rules, the\nrendering defaults, the spoken headers, the correction-inside-the-piece rule for a correction he\nheard, the chapter list. All of these are either his or measured to work.\n\n## 4. The proposed sharpening, as rules a writer can run\n\nEach is written to fire on something countable or on a word the writer is about to type, since the\nrules of 2026-10-08 that asked for a judgement are the ones that drifted. Numbering continues the\nmethod's. **RULED 2026-10-18 as a set, with the day's corrections folded in**, founder: *\"I... kind of\nassumed everything that we have been over about how to reshape audio-walkthroughs in this session has\nbeen set in stone already\"*. The ruled text of rules 34 to 52 is in the method\n([how-to-write-for-listening.md](how-to-write-for-listening.md) § The audit's rules); the text below is\nthe audit's draft and the measurements behind each.\n\n34. **The dry run refuses a text that is too dense to hear**, as it refuses one that is too long:\n    median sentence over 9 words, 90th percentile over 16, any stretch over 60 words between pauses,\n    or a planned pace over 115 words a minute. `--spend-anyway` (`--despite-gates` since 2026-10-20) overrides, and the override is noted\n    in the header.\n35. **Every number word is listed by the dry run with its sentence, and the writer strikes the\n    tallies.** The header lists the numbers kept and why each is the idea. Above 30 number words a\n    thousand the dry run says so.\n36. **His words are spoken for a ruling, a correction he made, or a sentence the part builds on.**\n    His questions are not read back to him.\n37. **A correction is spoken where he holds the old belief**: he heard it in an earlier piece or said\n    it himself. The others stay in the header. No piece opens by counting its own errors.\n38. **The session's own process (readers, hooks, scripts, what the tooling did) is one part at most,\n    and only when he asked**; otherwise one sentence in the view from here. The dry run prints the\n    share of words about process.\n39. **Each part has its two examples before its point, and the dry run counts them.** A part with\n    fewer than two is flagged before anything is spent.\n40. **The landscape closing is spoken once, in the view from here**, and in a part only where what is\n    real today would surprise him. Each part still closes on one line, in words that change.\n41. **The header names the registers the piece used**, from the palette of rule 27. Three pieces\n    running on the same two is the tell that the palette has gone.\n42. **One question to carry per topic the piece covered, each a full sentence with its prediction and\n    its silence.** The main one opens and closes the piece; the count follows the topics, not a cap\n    (his correction 2026-10-18). Open threads that raised no decision go to the header and the ledger.\n43. **Length follows the content's shape**: a day's thinking ten minutes, a catch-up thirty, an hour\n    for one deep subject. A ceiling he names is a ceiling, not a target. The delivery line states the\n    minutes, the voice credits and the plan cost.\n44. **The opening map names three groups at most.** The recap of an earlier piece is a minute.\n45. **The method keeps a table of his reports per piece**, and the delivery line asks for one word a\n    chapter: muddled, agreed, or disagreed with why. The next piece opens by answering the last report.\n46. **Two model pieces, not one**: the Arc P audit for form, *eight proposals as ideas* for register.\n47. **The rules that fire while writing live in a checklist of one screen, printed by the dry run**;\n    this document and the method keep the evidence.\n\n## 5. What the research says\n\nThe literature and the practice were searched on 2026-10-18 (eighteen web searches; the findings below\nare read from search summaries and abstracts, not from the papers in full, and each line says so where\nthe distinction matters). Grouped by the finding in § 2 they bear on.\n\n**On density (finding 1), and on the medium.** Normal narration runs 150 to 160 words a minute and\ncomprehension holds there; the time-compression literature puts the knee near 1.5× and a sharp fall\nnear 2× (Pastore and Ritzhaupt 2015, *TechTrends* 59(2), read in summary;\n[review](https://library.tuit.uz/knigiPDF/Ebsco/8-3549.pdf)). Our voice runs 145 to 168 while\nspeaking, so **the rate is not the problem; the load per sentence is.** Syntactic complexity loads\nworking memory more than length, and object-relative clauses are where listeners fall\n([PMC1761115](https://pmc.ncbi.nlm.nih.gov/articles/PMC1761115),\n[Frontiers 2023](https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2023.1264994/pdf)):\nso the gate should count clauses as well as words. Audio is not a handicap in itself: adults\ncomprehended a non-fiction chapter equally well as audiobook and as text, at once and two weeks later\n(Rogowsky, Calhoun and Tallal 2016, *SAGE Open*, 91 adults,\n[abstract](https://journals.sagepub.com/doi/abs/10.1177/2158244016669550)). But a **synthetic voice\ncosts something a human narrator does not**: Rodero and Lucas (2023) found listeners of a synthetic\naudiobook formed fewer mental images, attended less and remembered less, and judged long-form\nsynthetic narration still hard\n([UPF repository](https://repositori.upf.edu/items/429bf74a-a7a1-4c8f-a261-d0d368bf7afe)); a 1998\nreview found the extra effort shrinks with exposure. He is a frequent listener, so the gap narrows for\nhim, and it still argues for more slack in a machine-read piece than a human reader would need, which\nis what the pauses already buy. And a listener on a walk is dual-tasking: speech perception holds\nwhile walking, and drops when a memory task is added ([JSLHR\n2023](https://pubs.asha.org/doi/full/10.1044/2023_JSLHR-23-00198)), so anything the piece asks him to\n*hold* (a list, a tally, eight questions) competes with the road.\n\n**On numbers (finding 2).** Radio practice converges with the working-memory literature. Broadcast\nguides say round, write numbers as spoken, and keep dry statistics out, because the listener gets one\npass and cannot rewind ([journalism.university on\nradio vs print](https://journalism.university/broadcast-and-online-journalism/language-differences-radio-print));\na NICAR 2021 panel reported editors who cap a radio report at **two or three numbers** and put the rest\non a web page ([GIJN](https://gijn.org/2021/05/27/12-tips-to-make-data-come-alive-in-radio-reporting/)).\nCowan's reconsideration of Miller puts the real short-term capacity at **about four chunks** when\nchunking and rehearsal are blocked, which is the listener's situation (Cowan 2001, *BBS*,\n[PDF](https://memory.psych.missouri.edu/assets/doc/articles/2001/cowan-bbs-2001.pdf); Reynolds et al.\n2022 argue five). No study of spoken fractions and percentages turned up; the one number-format study\nfound was on reading. So the research gives the number rule a **form**, not only a ceiling: at most\ntwo or three numbers a part, each rounded and each compared to something the ear already holds.\n\n**On the session on stage (finding 3).** The trust literature cuts the way § 3 cut. Describing a\nscientist as intellectually humble raised trust across five studies with 2,034 participants, while\n**stating one's humility outright gave mixed results and sometimes backfired** (Koetke et al.,\n*Nature Human Behaviour*, late 2024; [NSF\nrecord](https://par.nsf.gov/biblio/10650200-effect-seeing-scientists-intellectually-humble-trust-scientists-research),\n[press](https://neurosciencenews.com/intellectual-humility-science-trust-28083)). A correction spoken\nwhere the claim is told is the shown form; *I was wrong about twenty times* in the opening is the\nstated form. On the quotations, the lecture literature on metadiscourse (a speaker talking about the\ntalk itself) finds two roles: **active**, where it guides the listener, and **passive**, where it is\nfiller the speaker does not attend to ([Multimodal Communication\n2023](https://www.degruyterbrill.com/document/doi/10.1515/mc-2023-0007/pdf)); his questions read back\nto him are passive metadiscourse by that test. The comprehension benefit of signposting itself is\nreal: removing discourse markers from a lecture lowered comprehension (Flowerdew and Tauroza, cited\nin [this review](https://www.cambridge.org/core/services/aop-cambridge-core/content/view/S0272263100014406)),\nand macro-markers (*now the second part*) help recall more than micro-markers (Chaudron and Richards\n1986, seen only through a [thesis](https://ir.uew.edu.gh/handle/123456789/1618)). So the part\nsignposts stay; the self-narration goes.\n\n**On the ritual and the examples (finding 4).** Three results sharpen rule 26. **Two analogs with a\nstated comparison** is the best-established way to induce a transferable schema: one story with a\nsummary, a stated principle or a diagram all failed, two stories with *say how they are alike* worked\n(Gick and Holyoak 1983, *Cognitive Psychology* 15, [PDF](https://Www.Gwern.net/doc/psychology/1983-gick.pdf);\nGentner et al. 2009 call comparison the best-established route to relational transfer,\n[PDF](https://www.qrg.northwestern.edu/papers/Files/QRG_Dist_Files/QRG_2009/Gentner,Loewenstein,Thompson&Forbus_2009-1.pdf)).\nSo the rule is not *two examples* but **two examples, then the sentence that says what they share,\nthen the principle**, which is what the twelfth piece did and the eighteenth stopped doing.\n**Concreteness fading** adds the other end: staying concrete leaves knowledge tied to the example, so\nthe abstract statement must follow (Fyfe, McNeil, Son and Goldstone 2014, systematic review,\n[DOI](https://www.doi.org/10.1007/S10648-014-9249-3)). And the **seductive-details effect** bounds the\npalette of registers: interesting material that does not carry the argument lowers recall a little\nand transfer more (Rey 2012 meta-analysis; Harp and Mayer 1998, summarised at\n[learningscientists.org](https://www.learningscientists.org/blog/2019/6/20-1)), with the effect\nstrongest for low prior knowledge and sometimes reversed in a history lesson (Wiley 2003). He has\nhigh prior knowledge, so the palette costs him less than it would a stranger; rule 27's *never as\ndecoration* is the right boundary, and the research says why. On analogy specifically, the\nscience-communication literature separates surface from structural similarity and documents analogies\nthat mislead ([JCOM letter](https://jcom.sissa.it/article/828/galley/1321/download)), which is the\ncorpus's own rule for the analogy daemon: a likeness names what carries over.\n\n**On endings and the shape of the hour (finding 5, and length).** Attention over a 40-minute lecture\nfell steadily, with recall falling alongside it (Farley, Risko and Kingstone 2013,\n[PMC3776418](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3776418/)); an older observational study\nput the first lapses at 10 to 18 minutes and every 3 to 4 minutes by the end (Johnstone and Percival\n1976, cited there). Two remedies are measured. **Interpolated questions** halved reported mind\nwandering over a 21-minute lecture and improved the final test (Szpunar, Khan and Schacter 2013,\n*PNAS*, [PMC3631699](https://pmc.ncbi.nlm.nih.gov/articles/PMC3631699/)); a 40-minute follow-up\nreplicated it (Jing, Szpunar and Schacter 2016). **Pretesting** before a 26-minute video lecture\nreduced mind wandering and raised the final score (Pan et al. 2020,\n[Athabasca](https://pure.athabascau.ca/en/publications/pretesting-reduces-mind-wandering-and-enhances-learning-during-on/)),\nand material that is not pretested is remembered better **at the beginning** of a lesson than at the\nend ([Athabasca](https://pure.athabascau.ca/en/publications/broader-benefits-of-the-pretesting-effect-placement-matters/)).\nMeasured against this: **the sixteenth piece put the part it called *the one to keep* at 74% of its\nlength, the seventeenth at 83%, and the twelfth put the proposal that *decides the most* in its\nclosing** (counted by word position). Learner-paced segments beat a continuous presentation with a\nmedian effect around 0.8 across ten tests, strongest for complex material and fast pace (Mayer's\nsegmenting principle; [Cambridge\nhandbook](https://www.cambridge.org/core/books/multimedia-learning/segmenting-principle/37240877DDA0362355ADB39936027982)),\nand the only large engagement dataset on lecture video found a sharp fall past six minutes a segment\n(Guo, Kim and Rubin 2014, 6.9 million sessions, [PDF](https://up.csail.mit.edu/other-pubs/las2014-pguo-engagement.pdf)),\nwhich our parts of two to six minutes already honour. On whole-episode length the evidence is vendor\ndata, directional only: completion around 74% for 20 to 40 minutes, 58% past 45, 48% past 60\n([The Podcast Consultant](https://thepodcastconsultant.com/blog/how-long-should-a-podcast-be)), and\nthe one craft precedent for a dense 45 minutes, *In Our Time*, runs three voices in three acts, the\nfirst act *for people who know very little indeed* and the last asking what was learned\n([THE](https://www.timeshighereducation.com/features/never-knowingly-relevant-melvyn-braggs-recipe-for-the-success-of-in-our-time)).\nHardcore History's multi-hour form survives on narrative and on the pause button, which its audience\npointed out to its author ([Earwolf](https://www.earwolf.com/?p=33593)).\n\n**On the loop (the open feedback channel).** The *muddiest point* technique, which the chapter list\nresembles (one word a chapter: muddled, agreed, disagreed), has mixed evidence for learning gains and\nstrong evidence that students find it useful; the studies that found gains are the ones where the\ninstructor acted on the answers ([RSC\n2022](https://pubs.rsc.org/en/content/articlehtml/2022/rp/d1rp00314c)). That is the design condition:\nthe next piece visibly answers the last report, or the list is a form nobody collects.\n\n**What found little or nothing.** No controlled evidence that saying a thing twice helps a listener\n(rule 15 stands as practice); narrative beating exposition for recall is a mixed literature once\ncontent is controlled ([meta-analysis](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8219577/)), so the\nscene-first opening is a plausible craft choice and not a finding; advance organizers (the spoken map)\nshow a mild positive effect in the meta-analyses, so the map earns its minute and no more; the\npersonalization principle (conversational *you*) carries a median effect near 0.8 in Mayer's tests and\n*may not apply to higher-achieving learners*, so the second-person voice is right and not where the\nleverage is.\n\n## 6. What the research changed in § 3 and § 4\n\nWritten after § 5, so the difference from the first takes is visible.\n\n- **Rule 39 is wrong as written** and becomes: *two examples, then the sentence that names what they\n  share, then the principle.* The comparison is the mechanism, not the pair (Gick and Holyoak), and\n  the principle must follow or the knowledge stays tied to the example (concreteness fading). The dry\n  run counts the examples; the comparison sentence is the writer's.\n- **New rule 48: the part that matters most comes first, after the map, never last.** Two hour-long\n  pieces placed their keeper at the 74% and 83% marks, the twelfth at 99%; attention and recall fall\n  across a long piece and un-pretested material is remembered best early.\n- **New rule 49: every part opens with a question he could try to answer before it is answered**, as\n  rule 16 already asks for the step that matters. Interpolated questions halve mind wandering on a\n  20-minute lecture; this makes rule 16 a per-part structure rather than an occasional device.\n- **Rule 35 gets a form**: at most three numbers a part, rounded, each set against something the ear\n  already holds. The ceiling of 30 a thousand stays as the count.\n- **Rule 34 adds a clause count**: a sentence with a relative clause inside another is flagged as the\n  object-relative literature says, not only a long one.\n- **Rule 43 is supported, not changed**: the completion data say 20 to 40 minutes, and the one dense\n  45-minute precedent is dialogic with three voices, which a single synthetic voice is not.\n- **Rule 37 is supported from outside**: shown humility raises trust, stated humility can backfire.\n- **New rule 50, small and buildable: chapter marks inside the mp3.** Learner-paced segments are the\n  measured remedy and a phone player offers them only if the file carries chapters; the chapter list\n  for noting then maps to places he can scrub to. `ffmpeg` writes ID3 chapters from the same part\n  boundaries the renderer already cuts at. Designed here, not built.\n- **Unchanged by the research**: rules 36, 38, 40 to 42, 44 to 47. The research reached none of them\n  directly, and the measurements in § 1 stand on their own.\n\n## 7. His responses (2026-10-18), and what they settle\n\nOn the two-examples finding: *\"Yeah good point, always keep this in mind\"*. On what an example is for,\nhis words in full, since they bound rules 26, 39 and 27:\n\n> *\"...and I feel like some examples here and there were quite far removed and maybe not really\n> helping me think much more clearly about what we were ACTUALLY talking about, not tying it back to\n> the ontology/philosophy that is most closely relevant to the question, thus sometimes mostly\n> confusing, not clarifying. The walkthroughs need to be a bit more strict about this, to bring up\n> examples that aren't too loosely connected to the topic/situation at hand. The analogy needs to be\n> thought-through so it really fits, if one is used.\"*\n>\n> *\"For me it suffices often to bring in examples of what the claims, claim-parts, reasons/supports,\n> rebuttals, relevance-argumentation, concept-sense-uses tracked etc could be for some example\n> discussion/debate, just grab any online debate example like whether 5G or microwaves is bad for your\n> health or whether Trump is an idiot or whether investing in deeply-researched stocks vs index funds\n> is the best idea or whether we're alone in the universe or there's basic organic life out there or\n> vast amounts of galaxies teeming with intelligent life that's just not heard/seen/detected easily,\n> whatever, just pull whatever debate stuff from online reddit threads or forums or contested\n> HackerNews threads or heated topics in the news with 2+ sides arguing different ways perhaps, have\n> some imagination here\"*\n>\n> *\"But sometimes if we're really in the nitty-gritty thinking about details of the ontology, then an\n> example can be distracting sometimes, and it's best to just elucidate calmly the principles and\n> quirks of whatever changes in the ontology/code/UI/UX we're talking about, so I can get a clear\n> picture for my mind's eye.\"*\n\nWhat this changes in § 4 and § 6 (his words govern; the rule text is Claude's):\n\n- **Rule 39 becomes**: the default example is a *real public debate* (a forum thread, a contested\n  news topic, a question with two or more live sides) **mapped onto the ontology objects under\n  discussion** (the claim, its parts, its reasons, a rebuttal, the relevance claim under a link, the\n  sense of a contested word), followed by the sentence saying what the two examples share, then the\n  principle. An analogy from another field is admitted only when it is worked through and fits; a\n  loosely connected one is cut. The kitchen-and-court examples of the eighteenth piece are of the\n  kind he described: an everyday scene standing in for the object. (Corrected 2026-10-18: the objects\n  are not a list; the example ties to the object model and the interface as they stand on the day, § 12.)\n- **New rule 51**: when the part is about the detail of the ontology, the code, or the interface, no\n  everyday example is owed; the part states the principle and its quirks plainly, so the picture is\n  clear in the mind's eye, and brings in the theory, philosophy, science or interface psychology that\n  bears on the point, researched with a few searches while the piece is prepared (his addition, § 12).\n  The dry run's example count (rule 39) therefore flags a part with fewer than two examples as a\n  *question*, not a failure.\n\nAlso asked that day, answered in the session and recorded here so the words have one home: *form*\nin this note means how a piece sounds and is built (pace, length, sentence length, structure) as\nagainst what it says; a *report* is what he says after listening; *one word a chapter* is the\nchapter list he asked for on 2026-10-15 answered with one of his own three words per chapter\n(*too muddled · agree · disagree, because*); *answering the last report* means the next piece opens\non what he marked muddled or disagreed with, so a report visibly changes the next piece; *the part\nto keep* is the part a piece itself flags as its most important (*\"This is the one to keep, so I will\ntake it slowly\"*, the sixteenth's part ten; the seventeenth's part eleven), which both hour-long\npieces placed past the three-quarter mark.\n\n## 8. Questions to carry, with a prediction beside each (his request, 2026-10-18)\n\nHis words: *\"In general, I want more room in each walkthrough for the questions posed to me to\ncarry/ponder, but accompanied always by your takes/predictions about what I would say/want, based on\na searchable index (perhaps through mannaminne/ccdistill somehow?) of ALL of my user messages in\nearlier CC sessions in Deliberus, is that feasible?\"*\n\n**Feasible, and most of it exists.** Measured 2026-10-18: the Deliberus session transcripts on this\nMac hold 1,115 of his messages (135,000 words, 2026-05-17 to today, 16 sessions; the transcript\ndeletions of May and August took the earlier files), and `mannaminne -s -P deliberus` already searches\nsession chunks, fleet-wide and back to April, with each turn prefixed *user:* or *assistant:*. What is\nmissing is a **speaker filter**: a chunk holds both voices, so a search for what *he* said returns\nboth. The cheap build is a `his-words` source in mannaminne (or a speaker tag on session chunks),\nfilled from the user turns of every session transcript, skipping compaction summaries, hand-backs and\nskill loads; `ccdistill --list` enumerates the sessions. **Built the same day**: the `said` source\nkind in mannaminne (its own repository, commit `47a26b6`), one row per user message from every\nClaude Code transcript on this Mac, 14,099 rows over all projects at first ingest and 30,222 by the evening, after another session filled in from Claude Code's prompt history what the deleted transcripts had held, so his words run from September 2025, searched with\n`mannaminne --said -P deliberus \"<question>\"` and left out of a plain search since the session rows\nhold the same words. Keyword search works at once; semantic search follows when the embedder resumes.\n\n**The shape in a piece** (rule 52, proposed): a question to carry is spoken, then five seconds of\nsilence, then *my prediction*, with the evidence in his own words and dated (*on 30 September you\nsaid…*), then *where I would be wrong*, then the question once more. The silence before the\nprediction keeps the pretesting effect (§ 5): he tries the answer before hearing mine, so the\nprediction informs rather than anchors, which is the corpus's own caution about a person shown a\nmachine's proposal. Each prediction's query and hits are listed in the piece's header, so a wrong\nprediction can be traced to what was searched.\n\n## 9. Can Jev judge an example? Tested 2026-10-18, his question\n\nHis question: *\"Isn't this perhaps too complex/abstract to be judged by Jev? Can you test it against\nyour own judgment on some examples, first of all and then let me be the final arbiter of this?\"*\n\n**Method.** Sixteen examples from the pieces of 8 to 15 October and three real-debate examples written\nfor the test (5G and health, stock picking against index funds, *is Trump an idiot*). For each I\nwrote my own verdict first (*fits · loose · far*, and *debate or analogy*), then put the same example\nand the point it serves to Jev with a three-way choice and a yes/no. Then twelve of his quoted\nsentences, my label first (*ruling · hold · question · lean · reaction*), then Jev's. Cost: under a\ntenth of a cent for all thirty-one calls. Script and raw answers in the session scratchpad.\n\n**Results.**\n\n| Judgment | Agreement | What Jev did |\n|---|---|---|\n| Quote type (ruling / hold / question / lean / reaction) | **12 of 12** | confident where I was (1.00 on *Build the page*, *So, browser-drawn it is*), hesitant where I was (0.52 on *Excellent. Yes, agreed.*, 0.55 on the Bayes intuition) |\n| Real debate or analogy | **19 of 19** | every analogy 0.03–0.09, every debate 0.75–0.95 |\n| Example fit (fits / loose / far) | **7 of 19** | answered *loose* for 13 of the 16 analogies, *fits* for the debates, never *far*; called the bicycle *fits* (0.56), the one I called far; called Keynes's coin, the doorman and the baking contest *loose* at 0.86–0.88 |\n\n**Reading.** Jev's *fit* answer tracks the kind of example, not its fit: an analogy reads *loose*, a\ndebate reads *fits*, and the one clearly muddled example passed. That is the failure the Jev\nscoreboard predicts for a question that needs the whole argument in view (the fleet's Jev scoreboard note of 2026-09-26, § 4). So his doubt holds: **example fit is not Jev's to judge.** Two things in the same test are, and\nboth are the countable half of rule 39: *what kind of utterance is this quoted sentence* (so a dry run\ncan count how many quoted sentences are rulings and how many questions, and flag a piece that calls\na question settled), and *is this example a real disagreement or an everyday scene* (so the dry run\ncan count debates against analogies per part). Both are one short item against one closed question,\nwhich is where Jev has worked on every earlier test.\n\n**The fit itself goes to the writer, as he suggested**, as an instruction run before each example:\n\n1. Name what the part is about in the system **as it stands that day**: the object-model and\n   database detail (the kind of node, the link type, the field, the stored verdict), or the exact\n   interface change. Not a fixed list of objects; the ontology moves, and the example ties to its\n   current state (his correction 2026-10-18, § 12).\n2. Default: take a real public disagreement, any live one with two or more sides (his examples that\n   day: 5G and health, Trump, stock picking against index funds, life elsewhere in the universe, a\n   forum or news thread), and show those objects in it, by name: *here the study is a reason, not a\n   part*.\n3. An analogy from another domain is used only when the mapping sentence can be written: *the eggs\n   are the parts, the hunger is a reason*. If one object in the point has no counterpart in the\n   analogy, cut the analogy. Write the mapping sentence into the piece.\n4. In a part about the detail of the ontology, the code or the interface: no everyday example; state\n   the principle and its quirks, and bring in **the theory that bears on it** (analytic or other\n   philosophy, the science, the psychology of interfaces), looked up with a few searches while the\n   piece is prepared (his addition 2026-10-18, § 12). **What counts as such a part** (Claude's\n   reading, for his ruling): its subject is how the system is built or behaves at the level of one\n   stored object, field, link type, pass, formula, rule or screen, so that checking what was said would\n   mean opening the code, the schema or the screen. A part whose subject is a stance, a purpose, a\n   choice between designs or what a result means is not one, and takes the real-debate example. Both\n   kinds tie to the object-level reality; the difference is whether a worked public disagreement is\n   added.\n5. Before the part ships, read the example once more asking only: does a listener gain something\n   about the point's objects from it, or only a pleasant picture? A pleasant picture is cut.\n\nRun by hand over the sixteen tested examples, step 3 would have cut the bicycle (its brakes are a\npart by the test and the sentence never says so), the Hamburg meeting (no counterpart for any\nontology object) and the mushroom picker (two cases, no rule, nothing of the link under it), and\nstep 2 would have replaced the omelette and the court with a debate. Proposed as the body of rule 39;\nunruled.\n\n## 10. The part to keep against the questions at the close (his question, read in three pieces)\n\nHis question: *\"What's the real difference between this and the questions-and-predictions stuff at\nthe end?\"* Read in the twelfth, sixteenth and seventeenth pieces.\n\n**What the pieces do.** The sixteenth marks parts ten and eleven as the keeper (*the floor of a\nstep*): part ten builds the idea (a step is small enough when every doubt has somewhere to land),\npart eleven reports the ten-link trial and his ruling of it (AG14), and ends on the one thing not\nruled, whether writing a chain under a link may change the link's weight. The closing part then\nlists the tools. The seventeenth marks part eleven as the keeper (*what the number means*: four\nreadings of the number on a claim) and its closing part one is **the same question in the same\nwords**: *What does the number on a claim mean?* The twelfth marks no keeper, and its first closing\nquestion (*who prices the bridge*) is part one's question returned.\n\n**So the difference, as the pieces stand.** The keeper is the piece's **understanding**: the\ndistinction the listener should be able to state afterwards, with its reasoning. The closing\nquestions are the piece's **decisions**: what is his to settle. They are not two things. In both\nhour-long pieces the keeper's own unsettled tail is closing question one, so the listener hears the\nidea with its reasons at the three-quarter mark, and then the bare question at the end, when the\nattention measurements (§ 5) say least is left. The keeper also arrives after the parts that depend\non it (the seventeenth's parts eight to ten lean on the reading of the number that part eleven\nintroduces), so the order in the piece is the order of discovery in the session, not the order a\nlistener needs.\n\n**What follows (rule 52's shape, revised; proposed):**\n\n- A piece has one main question. It is **spoken in the opening**, bare, with five seconds after it:\n  the pretesting effect (§ 5) and the room he asked for, in one move.\n- The keeper is the part that **answers or frames that question**, and it comes **early**, in the\n  first third, before the parts that use it. The label *this is the one to keep* is then unneeded:\n  the keeper is the part that follows the opening question.\n- The close **returns to the main question** with the prediction beside it, in his dated words from\n  the `said` index, and the place where the prediction could be wrong. Then the other decisions the\n  piece raised, **one per topic the piece covered**, each with its prediction and its five seconds.\n  The count follows the topics (his correction 2026-10-18: *\"That depends on how many topics are being\n  dealt with inside the audio walkthrough, doesn't it? In a recent one it was a pretty broad sweep\"*);\n  what is fixed is the room each one gets, about a minute with its prediction, so a broad sweep's close\n  is long because it has many decisions and never because it names them. A topic that raised no\n  decision gets no question.\n- Every other open thread is written in the piece's header under *what this leaves open*, where it\n  is findable, and not spoken.\n\nThe twelfth already had the shape by accident (part one's question returned as the first thing to\ncarry); the sixteenth and seventeenth had it inverted. Rules 44 and 52 in § 4 and § 8 are revised by\nthis section; unruled.\n\n## 11. Words he has no definition of beforehand (his question, and the advice)\n\nHis words: *\"Often you just start using words (often nouns iirc) that I simply DO NOT have a clear\ndefinition of beforehand. And it's like it's obvious to you but REALLY NOT to me. They might map to\nsomething I'm already quite familiar with by other words, but there's an inferential distance that\nstrains my listening capacity. ... I tried to get at this with the glossary a while back, but haven't\nhad the time to go through the glossary very much.\"* Measured the same day: the glossary\n(the blessed glossary, kept outside this repository) holds 40 terms marked known and 320 marked not-yet. The\nglossary's design doc rejected promotion by whether he has used a term himself (*\"too hard to know\nand thus, too inexact\"*); that test is not proposed again here.\n\n**Three kinds of word do this, and they need three different fixes.**\n\n1. **Names I gave to things**, and then used as if they were words: *the keeper*, *the floor*, *the\n   lift*, *the cap*, *the rebuild*, *the splitter*, *the chain*, *standing*, *the record*. Each is a\n   verb or a sentence folded into a noun. Checked in the two hour-long pieces: within a piece the\n   writer mostly did describe first (*the cap, the rule that a part holds its whole down by how\n   strongly the whole needs it*; *the splitter. That is the pass that cuts a claim into its parts*).\n   **The failure is across pieces**: *the lift* and *the cap* were defined on the 14th and used bare\n   on the 15th (*the lift as a claim*), *under the best-reason default* arrives bare in the 18th\n   piece, and *the record* is used in two senses there (the stored record of machine judgments, and\n   a person's record of what was searched) with no sentence separating them. A name defined on\n   Tuesday is a foreign word again on Wednesday, and in chat it never gets the Tuesday sentence at\n   all (this session: *form*, *reports*, *the part to keep*, each coined in a note and used bare).\n2. **Project terms with a glossary row he never blessed**: *horizon*, *relevance claim*, *separability\n   floor*, *merge candidate*. Known to every session, not to him.\n3. **Ordinary words carrying a project sense**: *weight*, *standing*, *reading*, *part*, *reason*,\n   *examined*. These are the worst for listening, since the ordinary sense arrives first and the\n   project sense has to displace it mid-sentence.\n\n**What the research says a glossary can and cannot do.** Bullock, Amill, Shulman and Dixon (2019) and\nShulman, Dixon, Bullock and Colón Amill (2020) measured that jargon lowers reading fluency, and that\n**a definition beside the term does not restore it**: readers given definitions still reported the\ntext harder to follow, since the lookup itself costs the fluency. The curse of knowledge (Camerer,\nLoewenstein and Weber 1989; Pinker's account of why experts write badly) is why I cannot feel the gap:\nonce a name stands for a thing to me, the sentence reads as obvious. So a glossary he reads later does\nnot fix a sentence he hears now; the fix has to be in the sentence, and the glossary's job is to decide\nwhich words may appear bare.\n\n**Advice, five parts, ordered by what to build first.**\n\n1. **In a piece, the thing comes before its name, every time.** A project term enters a piece as a\n   description in ordinary words, then *I will call that X*, then X is reused. One exception: a term\n   marked known in the glossary. Countable: the dry run lists every glossary term, hyphenated compound\n   and capitalised name in the text and reports which appear before any describing sentence. This is\n   the dry run's third countable gate beside sentence length and number words (§ 4), and it is cheap:\n   the glossary is a table and the text is on disk. For the ordinary-word class, the dry run cannot\n   see it; the writer's instruction is to say the sense once (*weight, in the sense of how much has\n   been examined*) and the first-reading report (§ 7) is where he marks the ones that slipped.\n2. **In chat, a Stop hook counts, as the brevity counter does. Built and live 2026-10-18**\n   (`glossary_counter.py` beside `brevity_counter.py`, on his *\"Sounds wise, go ahead\"*): after every\n   reply it names the not-yet glossary terms used in a sentence with no gloss marker and the\n   hyphenated compounds no glossary row knows, in the flag file the next turn reads. Scoped by file to\n   projects that have a glossary, so Deliberus alone today, on his word. Lone ordinary words in a\n   project sense are not checked. A hook is a law where the style's lookup rule is a request; the rule\n   had stood since 13 September and this session broke it three times in one reply (*form*,\n   *reports*, *the part to keep*). Free, deterministic, no model.\n3. **The glossary comes to him, five words at a time. Built the same evening** (`glossary_bless.py`,\n   the `glossary-bless-nudge` hook at a wrap-up, `/bless` on demand; his *\"Build the glossary blessing\n   pass\"*). He has not had time for 320 rows; nobody\n   would. So the blessing pass is a standing item at the end of a session: the five not-yet terms\n   used most in the last week's replies and pieces (countable from the session index and the\n   pieces on disk), each with its one sentence, answered *known · not-yet · rename · drop*. Five a\n   session clears the terms that cost him most within a month, and a term he renames gets his word\n   as its name, which is the one place his own wording enters, chosen by him and not inferred.\n4. **Coin less.** A new compound is admitted in a reply or a piece only with a glossary row written\n   in the same turn, so coining costs the coiner. The density bar (one argument-carrying term a\n   sentence) stands; this adds the ledger.\n5. **Names for passes and objects get the agentive form.** *The keeper* and *the splitter* are\n   survivable because the word says what the thing does. *The floor*, *the lift*, *the cap* and\n   *the rebuild* do not; they are metaphors standing for rules. Where a term is renamed in the\n   blessing pass, the test is the stranger test from the constellation: does the name say what the\n   thing does to someone who has not read the notes.\n\nWhat is real today: the Stop hook, live. The dry-run gate, the blessing pass and the coining rule\nare proposed, unruled.\n\n## 12. His corrections of the evening (2026-10-18), in his words\n\nOn step 1 of the example instruction, which listed ontology objects: *\"These should not be referred to\nas a static list, it ofc depends on what the then/current state of Deliberus ontology and codebase is\nwhat is being discussed. I simply meant it should tie back to the cold, hard, object-level reality, the\nobject-model/database nitty-gritty, or the exact concrete UI/UX changes or whatever, as opposed to\ncartoonish or everyday baking/party/silly analogies/examples meant to be like these things, you know?\"*\nStep 1 now says so, and the test for an example is whether it reaches that object-level reality as it\nstands on the day, not whether it names an object from a list.\n\nOn step 4 and rule 51: *\"Yeah I mean or theory, analytic or other philosophy or science or UX-related\npsychology that bears on it, stuff like this. Feel free to research these things with a few\ngoogle-searches when preparing the walkthrough also.\"* Both now carry it: a nitty-gritty part brings\nthe theory that bears on it, looked up while the piece is prepared.\n\nOn rule 38, which he asked to have explained plainly (*\"Huh? Explain this plainly\"*): a piece spends\nwords on the session's own machinery, meaning the helper readers that were run, the hooks, the\nscripts, which model answered. The sixteenth piece spent about 27% of its hour on that. The rule\nsays: such machinery gets at most one part of a piece, and only when he asked about the tooling;\notherwise it is one sentence in the closing view, and the dry run prints what share of the words\nit took.\n\nOn the closing questions (*\"why 'at most two more decisions'? That depends on how many topics are\nbeing dealt with inside the audio walkthrough, doesn't it?\"*): § 10 and rule 42 now say one decision\nper topic, the count following the topics, the room per decision fixed.\n\nOn the scoping of a nitty-gritty part, the same evening: *\"Well all of it is also quite interconnected,\nperhaps especially in Deliberus, so we need to have respect for that. Even the latter 'half' could\nsometimes... idk. What do you predict will be consequences of this mental split, for the walkthroughs?\"*\nClaude's prediction, recorded before any piece runs under the rule: a writer sorting parts into two\nkinds would classify by how a part opens and starve the other half, so a part on the strength formula\nthat also settles what the number means would get theory and no case, and a part sorted as a stance\nwould get a public debate that floats free of the field it is about; and in Deliberus the test *would\nchecking it mean opening the code* fails on its own terms, since opening the code here is opening a\nconviction (a field default encodes a ruling). Rule 51 was therefore rewritten the same evening from a\nsorting of parts into two questions every part answers (what in the system as built does this touch;\nis there a live disagreement where the difference shows), in the method § The audit's rules.\n\n**Built the same evening, on his *\"Did you not build the new dry-run stuff? Do it\"***:\n`scripts/listening_gates.py`, run by the renderer's dry run and before any render; tested against the\npieces this audit measured (`tests/test_listening_gates.py`), and rule 50's chapter marks written into\nthe mp3 by the renderer. On the eighteenth piece the gates refuse on all four density and number\ncounts the audit found by hand; on the twelfth they pass. The machinery share they count is a floor\n(sentences naming the tooling), lower than the audit's reading by part subject.\n\nHis word on the two-questions rewrite, the same evening: *\"I agree on the whole. But don't ditch\nanything I've said to you that could be enriching the answers, still, perhaps? While still bringing\nboth the concrete, technical, theory-laden and real-world vivid, educational, pedagogic examples. Keep\nas much still of my phrasings in the notes/decisions on all this.\"* So rule 51 keeps every sentence of\nhis it had (the object-level reality, the theory to bring and the searches to make, the interconnection)\nand adds this one; the two questions are the order in which a part answers them, not a replacement for\nany of it, and a part is to be concrete, technical, theory-laden and vivid at once where it can be.\n\n**Jev on the machinery share, per part (his *\"Ok test Jev on the tooling share per part\"*, the same\nevening).** Each part of the sixteenth and eighteenth pieces was put to the decision model with one\nquestion: is this part mainly about the session's own machinery, or about Deliberus's ideas, graph and\nrulings? The session labelled every part first. Agreement 23 of 26; the three misses were the sixteenth's\nparts three to five, which report what helper readers found, so their subject is the ideas and their\nmechanism is the tooling, and the model put them at 0.12 to 0.38 where a yes-or-no at 0.5 calls them\nideas. The pure tooling part of each piece came out highest both times (0.71). So the gate keeps the\nprobability instead of the verdict and weights each part's words by it: **19% on the sixteenth**\n(the audit's hand reading by part subject was 27%, the sentence count 6%), **17% on the eighteenth**.\nBuilt into `scripts/listening_gates.py` behind `--jev`; without the flag the sentence count stands, named\nas a floor.\n\n**His later words of the evening (2026-10-18), each placed in the method:** on the pacing, *\"I've been\nhappy with the pause lengths, they maybe could be shortened by one second or so, the longest ones\"*\n(§ More room, and breath; the written ceiling is five seconds, the gates note a longer one); on rule 52,\n*\"You could make the predictions in the beginning (briefly) and in fuller detail at the end so I can be a\nbit primed by them and ponder them throughout hearing the walkthrough, perhaps\"* (rule 52); on terms,\n*\"try to adhere somewhat to the analytic-philosophy discipline of reusing established terms. I have a\nhunch using words a bit too interchangeably and haphazardly is what caused our confusion around\nparts/reasons months ago. I went along with you using 'parts' as stand-in for reasons and didn't realize\nthe meaning had shifted/drifted. Not great.\"* (new rule 53, and the register line of CLAUDE.md); on rule\n44 he asked whose it was (Claude's, from the audit and Cowan; marked so); on rule 27 he had never heard\ndifferent voices (a register is a manner of the one voice, Lucy, from his list of 2026-10-08; marked so).\nThe first blessing batch was ruled that evening: Lins, Astra, critical question, auto-connect and corpus\ngraph are known.\n\n**Three more of his words, later that evening.** On the register palette: *\"I think it over-anchored a\nbit on this one and under-valued the others, strike the 'quirky example' bit pls\"* (struck from rule 27;\nhis quote of 2026-10-08 stays whole). On rule 44: *\"This is good if the items/topics/questions covered by\nthe walkthrough (whatever I ask for to be covered/pondered when I ask for a walkthrough in a session) are\namenable to clustering into roughly 3 groups, fuzzy clustering allowed, but all the things I've asked for\nshould be included nonetheless\"* (rule 44 rewritten in those terms). On voices: *\"Perhaps we can test out\nusing different voices for the different registers, at least for one or two of them, besides Lucy?\"*\n(built: a `[[voice NAME]]` line per part in the renderer, a table of five voices from the account, and a\nthree-voice sample rendered for him, 1.6 minutes).\n\n**Voices and pronunciation, his words later still.** Asked whether the account's voices were visible\nthrough the API: yes, all 47, and his picks replaced the first table: *\"I have a personal voice clone in\nthere, named Fredrik ... And I like Myrrdin ... And perhaps Scotty Marshal ... and Anna ('conversational,\nhelpful'). Use these instead, with Myrrdin as the elder. Unless any of these are like doubly expensive\"*:\nmeasured equal, 27 credits for one 62-character sentence on each. On pronunciation: *\"Sometimes btw the\nvoice says for example 'built and live' and pronounces 'live' as in 'you shall live your lives free of\nsorrow'\"*, and *\"Could you perhaps prefix-instruct, include a preamble for the voice synthesis somehow,\nfor the underlying AI/voice model to understand the context better ... Is that feasible?\"* Feasible on\nthis model through the unspoken `previous_text` field, now sent with every piece, with a clip of the\nsentence three ways for him to judge; and the dry run lists every two-way word for the writer.\n\nOn the three-way clip: *\"All sound identical and correct. This time. Perhaps correct in the first plain\nversion by accident, idk. But if you can send unspoken context to prime the model, then that's probably\npreferable. Could we send some unspoken context about Deliberus so it 'understands' more about what it's\ntalking about (potentially, hopefully, proactively)?\"* Built: the preamble became a prose file,\n`docs/listening/voice-context.txt`, read fresh on every render, about Deliberus and its words and names.\nAnd on the preamble's line *read is the present tense*: *\"This is not always true tho\"*; removed.\n\nThen on the file itself: *\"This is great... but seems risky: will drift with time. No? Could we swap\nout the drift-risky parts to instead auto-include some part/s of our CLAUDE.md and some other central\nnotes, along with perhaps an auto-generated super brief summary of the object-model/database/architecture\netc? idk\"*, and a minute later: *\"Hmm. Perhaps the corpus-graph and the glossary can be auto-included and\nalso perhaps it's a good idea to always consult the corpus-graph for the writer before writing a\nwalkthrough? Advise on this idea, not totally sure\"*. Built (`scripts/listening_context.py`): the\ncontext is assembled per piece from the hand-kept head (register, readings, names only), the glossary's\nsentence for each term the piece uses, and the live graph's kinds of node, claim and link, cached with\na date; the renderer's dry run states its size. The advice given: the glossary and the graph shape\nbelong in the voice's context, since a voice uses senses and names; CLAUDE.md and the corpus graph do\nnot, since they are written for a session and a voice reads no arguments. The corpus graph is for the\nwriter, and that half became rule 54: brief on the notes before writing, name them in the\nheader, and the dry run prints what the graph has judged about each. Ruled that evening, once the\nbriefing's cost was stated (about 900 tokens for three notes): *\"Blessed!\"*\n\nOn the voice itself he pushed back on the advice that a voice reads no arguments: *\"how does the voice\nAI work exactly? Does it not spend any time thinking/understanding the content? It doesn't help to\ninclude more meaningful context to produce more 'informed' voice synthesis? Research this, and\nespecially for the v4 model!\"* The research (ElevenLabs: v4 *\"reads a script the way a voice actor\nwould\"*, *\"performs better with context\"*) and two paid tests settled it his way in part: the same\nneutral sentence under a heartbroken and a delighted unspoken direction came out *\"clearly different\"*,\nand *\"just as different in vibe\"* when the direction sat after 2,200 characters of other context, so\nthere is no short cap; the homograph *read* came out past tense under both a past and a present context,\nso word readings stay the writer's. His ruling: *\"Then we should put in more of these types of\ncontext\"*, and on argument: *\"if it is capable of reasoning and logic, which it must be to infer much\nat all, then I think we can put in some argument/reasoning/philosophy-ish context still.\"* Built on his\n*\"I think all of this sounds great, go ahead\"*: a scene line per part (rule 55), a character sentence\nper voice, the context ordered scene first, and two sentences on what Deliberus argues at the head of\nthe hand-kept context.\n\nLater the same evening, on the hand-kept context's two sentences about the thesis: *\"As much as possible of\nthe project-context should be not-duplicated to be drift-risk-free, yeah? So please include the\nRobin-Astra-outside-view summary perhaps and other super duper short summaries of what Deliberus is about,\nbut do it by copying in clips from certain other files intelligently, so it won't duplicate this stuff and\nwon't break if the source-files get edited or whatever.\"* Built: `{{clip …}}` lines in `voice-context.txt`\nthat copy a passage by its words at render time, from the outside reading, CLAUDE.md's formulations and\nthe vision page; the API accepted a 5,587-character context. And on his voice clone: *\"Is my Fredrik voice\nused for anything? Maybe for the parts of the walkthrough that are predictions of what I would say? :)\"*\nBuilt: a voice line may now sit inside a part, so the predictions at the close can be read in his voice;\nthe convention is written at rule 52 and waits for the next piece and his ear. What the day taught the\nnarration pipeline is planned in `~/dotfiles/docs/narration_unspoken_context_plan_2026-10-18.md`.\n\nTwo more words of his that evening, both ruled on the spot. On the clone, after hearing it under the\nfull context: *\"Sounds perfect, please include it for the predictions\"*; rule 52 carries it. And on the\nbreathing, counted and then corrected: *\"have the audio walkthroughs lighten up on the breathing,\nsometimes Ruby sounds like she's sighing in exasperation lol\"*. The piece of 13 October had 28\n`[exhales]` in 6,700 words. No `[exhales]` or `[sighs]` from here on, at most two breath cues a thousand\nwords, and the dry run notes both (§ More room, and breath).\n"}