{"path":"listening/first-time-right-2026-11-05.md","content":"# Making a walkthrough pass first time (2026-11-05)\n\n*About how we work (the renderer, its dry run, the writing step), never about Deliberus's design. Proposals only; one\nof them waits on the founder's word, the rest are operational.*\n\n**His question, 2026-11-05**, after the piece *Judging the connections* needed three rewrite rounds before its gates\npassed: *\"We wasted tokens in making this walkthrough I assume - because drafts failed (and what is the dry run for,\nexactly, again, remind me?) - help ideate about ways to make this just go right on the first go? Maybe gates such as\nthese aren't the best way to handle quality control for the walkthroughs? Idk? Research this extensively online.\"*\n\nTwo readers went out, one on what the language-model literature measures about length and readability control, one\non what writers for the ear do and what comprehension research says. Their reports stand verbatim beside this note:\n[the model-constraint reading](first-time-right-2026-11-05/report-llm-constraint-methods-2026-11-05.md) and\n[the craft and comprehension reading](first-time-right-2026-11-05/report-spoken-word-craft-2026-11-05.md). This note\nanswers the question from them.\n\n## 1. What the dry run is for\n\n`scripts/render_listening.py TEXT OUT.mp3 --dry-run` is a rehearsal that spends no credit. It cuts the text into the\npieces the renderer would send to the voices, forecasts the running time and the pace from the text itself (never from a\nguess at the word count), counts every gate the method has made countable (median sentence length, the ninetieth\npercentile, the longest stretch between pauses, pace with the pauses counted, number words, examples a part, openers\nwithout a subject, stress marks, nested clauses, a borrowed voice that announces itself), checks each sentence his clone\nreads against his own record, and prints the corpus graph's briefing on the notes the piece covers beside the judgement\nchecklist the gates cannot count. A refusal stops the render; a note does not.\n\nIt exists since 2026-10-18, when the audit of eighteen pieces found that the pacing rules the renderer enforced had\nheld, and every rule left to the writer's memory had drifted back within a week\n(`docs/listening/what-eighteen-pieces-taught-2026-10-18.md`; the gates' own docstring says the same). The prepare step\n(`scripts/listening_context.py --prepare`) was added the same day, after the twentieth piece spent 4.45 million weighted\ntokens on a dozen reads and a draft the density gate refused.\n\n## 2. What this piece cost\n\nFour dry runs, three rewrite rounds. The first draft was refused on median 10 against 9, ninetieth percentile 22\nagainst 16, a stretch of 84 words between pauses against 60, pace 119 against 115, one stress mark over four words and\nthree sentences opening without a subject. The second round still had a stretch of 83 and thirteen sentences over\nsixteen words; the third a stretch of 61; the fourth passed (3,309 words, median 8, ninetieth percentile 14, pace 111).\n\nThe shape is the one the literature predicts: the average was near its gate and the tails far from theirs. The two\npieces of 22 and 23 October were rendered with the gates waived, since the sentence gate came later; the pieces of 24\nOctober were refused once each; this one three times.\n\nThe dry runs were slow for a reason that has nothing to do with the writing. Eleven sentences in his voice are each\nchecked against four scopes of his record with a ninety-second limit, and the record's search server was unreachable\nfrom here. Measured on this piece, the gates alone took 331 seconds of wall clock for five seconds of work; the corpus\ngraph's briefing took under a second.\n\n## 3. What the two readers found\n\n- **A model cannot count words while it writes.** Across 26 models, length compliance falls sharply past a short\n  target; asked for a sentence of a stated length, the best model was exact 12% of the time (LIFEBench, NeurIPS 2025;\n  the per-sentence figure from an earlier controlled-generation study).\n- **An instruction moves the average, never the band.** Readability instructions lowered mean grade level and no model\n  or prompt reached a narrow target band (Flesch or Fumble, GEM 2023).\n- **A visible counter changes it.** A countdown marker beside the text took exact-length compliance from 9.7% to 96.3%\n  (CAPEL, 2025). Anthropic's own guidance: limits in sentences and paragraphs hold better than limits in words.\n- **An external checker loop converges in one or two rounds, then flattens.** With a deterministic readability check,\n  64.5% of items passed untouched, 80% after one round, 87% after two, 91.5% after four (HIT-YOU, TSAR 2025). A model\n  checking itself without external feedback did worse than not checking (Huang et al., 2024).\n- **Adherence decays past about 4,000 tokens of output** (DocFollow, EMNLP 2026). A piece of 3,300 words sits at that\n  edge.\n- **Code fixes what has one right answer, and only that.** Prettier reprints formatting and refuses the cases with a\n  judgement in them; Vale and Hemingway only flag readability. Splitting a sentence by rule is safe at *and* and at an\n  adverbial clause (near 100% in DisSim), and unsafe at a relative clause or a *but*, where meaning moves across the\n  split.\n- **No broadcast house runs a hard gate.** The practice is one sentence per line, speak it before writing it, edit by\n  ear. The one measured checker (Caterpillar's controlled English, 180,402 checks) passed 94% first time for trained\n  authors with an interactive checker; the reject-only version made new authors flail, and diagnostics that named the\n  fault and offered a rewrite found the problem 90.2% of the time.\n- **The nine-word median and 115 words a minute are listener-calibrated preferences**, inside the convergent range of\n  professional advice (averages of 6 to 16 words; 115 wpm as the floor of speaking rates) and stricter than every cited\n  ceiling. No study separates a median of 8 from one of 10. Over-splitting has a measured cost (staccato; readers\n  preferred a two-way split to a three-way one).\n- **Three levers the evidence supports that the gates hold loosely**: one object-relative clause inside a six-word\n  sentence cost about nine points of comprehension (Peelle 2010); pauses help only at clause and sentence boundaries\n  (Wingfield 1999); one number a sentence (NPR), where the gates count per thousand words and per part.\n\n## 4. The moves, in dependency order\n\n1. **Fail fast on the record.** Probe the record once; unreachable, skip the quote check with the note the gates already\n   print for that case. Saves about five minutes a dry run. No ruling needed.\n2. **Write and check one part at a time.** `scripts/listening_gates.py PART --brief` returns in seconds on a part of\n   600 to 900 words. Draft a part, run it, fix the named lines, then the next part; assemble; one whole dry run at the\n   end. This keeps each unit under the decay edge and turns a refusal into a few sentences rather than sixty. A habit\n   for the walkthrough skill; no ruling needed.\n3. **Give the writer counters, not word targets.** The writer's card already prints the gates' numbers. Add per-part\n   budgets in sentences (a 750-word part is about 95 sentences at median 8), a pause line after every six or seven\n   sentences, and at most one number a sentence. No ruling needed; the card is the prepare step.\n4. **A formatter for what has one right answer: pause placement and pace.** Where a stretch would pass 60 words, code\n   inserts `[[pause 4]]` at the nearest sentence boundary and recomputes the pace, so the floor is guaranteed and the\n   writer places only the pauses that mean something (after a question, a number, a dense sentence). **This changes a\n   rendered artefact, so it waits on his word.**\n5. **Diagnose, never only refuse.** For each refused sentence the gates print the fault and, at an *and* or an adverbial\n   clause, a candidate split; at a relative clause or a *but*, the sentence alone. A build; no ruling needed.\n6. **Cap at two rounds.** Gains flatten after the second; a piece still refused then goes to a person. No ruling needed.\n7. **Record the gates as what they are.** In the method, the nine-word median and 115 wpm are his measured preference,\n   kept because his ear is the test, not thresholds from research. Candidates for his word: a floor on sentence length\n   against staccato, and one number a sentence in place of thirty a thousand.\n\nNot worth doing, by the evidence: a model reviewing its own draft as the gate (worse than nothing); constrained decoding\n(no hosted API enforces sentence length; Anthropic drops `minLength` and `maxLength` from a schema); waiving the gates to\nsave a round.\n\n## 5. What to expect, and how we will know\n\nUnmeasured. The evidence puts a well-briefed writer with a per-part checker near the Caterpillar shape: most parts pass\nuntouched, a few need one round. The check is cheap: each kept piece's header already records its rounds, so after five\npieces the first-pass rate is a count.\n\n**The one decision for him**: move 4, a formatter that places the floor's pauses by code. Moves 1, 2, 3, 5 and 6 proceed\nas operational changes; move 7 and the two candidate gates wait on his word with the method's other rules.\n"}