2026-09-02
User:
You are an independent quality evaluator for a novel being drafted in redacted. The author (DRMacIver) has accepted the first act (scenes 01–24) as 'generally very good' and later scenes 25–40 as 'pretty good ... redacted sentence: unpublished prose redacted sentence: unpublished prose The scenes are now being redrafted from scratch, and the author has asked that an independent reader evaluate quality every couple of chapters and say whether it is degrading.
Your job: read the standard, then read the six newest scenes, and give a verdict on whether the newest scenes hold the standard, with evidence.
The standard: read draft/book-01/VOICE.md in full; then read these accepted scenes in full: draft/book-01/scenes/01-, 02-, 05-, 06-, 13-, 19- (the exemplars), and 31-, 32-, 33-, 36-, 37-, 40- (the accepted act-two register). Then read notes/llm-tells.md (the failure modes policed).
The scenes under evaluation: draft/book-01/scenes/41-calm-hands.md unpublished, 42-a-factors-woman.md unpublished, 43-two-lines-one-field.md unpublished, 44-where-they-stopped-writing.md unpublished redacted sentence: unpublished prose unpublished, 46-a-list-of-houses.md unpublished.
STRICT: never read chapters/, rewrite/, draft/book-01/rejected-act2/, summaries/, or plans/book-01/arc-0*.md, unpublished or their git history; do not search for them.
Evaluate, with quoted evidence from both the exemplars and the new scenes: (1) Is the prose at rest where it should be, with the wit in the events? (2) Do people say what they mean — is every exchange understandable to a reader who has not read the previous scene; do characters give their reasons; are there cryptic fragments or callback words carrying meaning on their own? (3) Do the characters sound like themselves and distinct from each other (Edwin, Peri, Mattie, Tess, redacted, redacted, redacted)? (4) Any LLM tells at fingerprint rate (verdict clauses, gesture-at-meaning, near-idioms, false precision, maxim-capped dialogue, triads, negative parallelism, house words like 'honest', 'ordinary', 'the whole of it')? (5) Is Peri likeable and active (asking questions, her own agenda), not a clerk? (6) Are scene endings clean (a scene may end on an answer or decision; not a composed button)? (7) Is anything cryptic, abrupt, or 'people talking at each other'? (8) Is the scene-to-scene register consistent with 31–40, or drifting (toward summary, toward lecture, toward fragments, toward sameness of rhythm)?
Then a verdict, in this exact form on its own line: VERDICT: HOLDING / SLIPPING / DEGRADING — followed by three to six sentences of justification, and a list of the two or three most important concrete fixes if any. Be blunt; the author prefers accurate to kind. Do not edit any file.
I'll read the standard first, then the new scenes.
🔧 Bash: wc -w draft/book-01/VOICE.md notes/llm-tells.md draft/book-01/scenes/*.md | head -60
✅ 53 lines, 2871 chars
🔧 Read draft/book-01/VOICE.md
✅ 257 lines, 16258 chars
🔧 Bash: cd redacted/draft/book-01/scenes && cat 01-*.md 02-*.md 05-*.md 06-*.md 13-*.md 19-*.md
✅ 312 lines, 28304 chars
🔧 Bash: cat notes/llm-tells.md
✅ 332 lines, 19871 chars
🔧 Bash: cd redacted/draft/book-01/scenes && cat 31-*.md 32-*.md 33-*.md 36-*.md 37-*.md 40-*.md
✅ 22 lines, 2242 chars
🔧 Read redacted: private
✅ 337 lines, 45263 chars
Now the six scenes under evaluation.
🔧 Bash: cd redacted/draft/book-01/scenes && cat 41-*.md 42-*.md 43-*.md
✅ 153 lines, 22559 chars
🔧 Bash: cd redacted/draft/book-01/scenes && cat 44-*.md 45-*.md 46-*.md
✅ 153 lines, 19237 chars
🔧 Bash: cd redacted/draft/book-01/scenes && for f in 41 42 43 44 45 46; do echo "== $f"; cat $f…
✅ 43 lines, 527 chars
🔧 Bash: cd redacted/draft/book-01/scenes && echo "em-dash new:"; grep -o -- "—" 4[1-6]-*.md | w…
✅ 11 lines, 248 chars
🔧 Bash: cd redacted/draft/book-01/scenes && grep -n -E "(That's not|It's not|It isn't|not becau…
✅ 11 lines, 1770 chars
Read: VOICE.md, llm-tells.md, exemplars 01/02/05/06/13/19, accepted act-two 31/32/33/36/37/40, and the six new scenes 41–46.
Holding, and by the right mechanism. redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence: unpublished prose Edwin's "been looked at, been wished a good evening, and been shown by two dogs to the road" is 37's, not the new scenes'; 41–46 spend less on figures than 31–40 did, which is the right direction. Narrator aphorisms: essentially zero across all six. Em dashes: six in six files, all in the title lines. Word census against 31–40: "honest" 0 (was 7 across the accepted stretch), "ordinary" 2 (was 7), "which was/is" 12 in 8.2k words vs 17 in 13.7k — slightly up per word but nowhere near verdict-clause rate, and every instance is causal, not grading.
This is where the earlier act-two failure lived, and these scenes are clean. Every reason is on the page in the same speech. redacted sentence: unpublished prose Sennet's terms in 37 were the model and 45 matches it: redacted gives the whole sum aloud — "Nineteen on the wall. redacted sentence: unpublished prose redacted sentence: unpublished prose redacted sentence: unpublished prose
The one deliberately cryptic object, the slate reading In, is a good test case and it is handled correctly: Peri asks "In what? Who's in?", they go inside and ask, and redacted explains it in full in the next paragraph. No callback word is left carrying meaning. "Day zero" (42) is defined at the point of coining, including what it does not mean. I found no fragment that gestures at an answer.
Peri, Edwin and Tess are secure. The three new voices are genuinely differentiated by content: redacted redacted sentence: unpublished prose redacted in sums and accepted liability ("It's held forty-nine. redacted sentence: unpublished prose I know what I'm betting"); redacted in counts and kit.
The real drift is that the fen's adults increasingly share one closing shape — the short flat verdict with a load under it. Mattie: "Hedges do that." Sennet (40): "That'll do me." redacted: "The glove's a glove." / "For the person behind me." Individually all good; four scenes running they read as one laconic house register rather than four people. Peri's "Level one... That's the bottom" and redacted's "Where do they go" also share a repeat-the-phrase-then-answer tic.
At or below the accepted rate. No triads with swerve, no maxim-capped dialogue, no nominalised absence, no false precision on unmeasurable things (the numbers are all diegetic: chains, links, ages, dates). Negative parallelism runs at about the same rate as 31–40 and is mostly in a speaker's mouth doing argumentative work ("It's not a count of what's in the fen... It's a count of what's got as far as the houses"), which is licensed. Two soft spots: 41 uses "flat" four times in 1.6k words for voice and manner, and 41 leans on "which was the whole test" / "which meant" three times in a page.
Strongest she has been. She has her own agenda and it is not Edwin's: she invents the head-of-page list of counters with their ages and defends the reasoning when asked (42); she out-argues Edwin's instinct in front of a hostile headland and he concedes it ("That was well done... I'd have talked about the redacted"); she asks redacted redacted sentence: unpublished prose Not a clerk.
Four are clean; two are composed buttons. 41 (Edwin's book entry, then she puts her hand back on the heap), 43 ("Hedges do that," and off to the dinner) and 44 (redacted's "She'll be in before six") are cuts on work. But 45 ends on redacted's held-back line — "That's the second date I've written for that hill this year. I wrote the first one too." — which is a withheld beat released precisely at the last line, and 46 ends on "In the books, nobody ever counts anything," a callback to Peri's running phrase, placed to click shut. Both are good lines; both are exactly the resolving-button shape VOICE.md names, and they are two scenes apart.
Nothing. This is the opposite failure mode from the rejected draft.
Consistent in voice, drifting in shape. Five of the six scenes run the same engine: a measurement or record procedure, a plain explanation of what the number means, and a secondary character's flat closing verdict. 42, 43 and 46 are structurally near-identical (procedure → Edwin or Peri explains the inference in a long speech → local character says something short). 43 is the weakest: the geometry occupies most of the scene and the discovery — two lines give two different centres — is delivered as Edwin talking to Mattie for four consecutive long paragraphs, with Mattie contributing "Then what is it" and "Write all three". That is close to lecture with an audience rather than a scene. VOICE.md §3.3 requires no discovery move to repeat inside four scenes; measure-and-record repeats in 42, 43 and 46 (redacted's houses are a count too).
VERDICT: HOLDING — The prose is at rest, the reasoning is on the page, nobody talks cryptically, the LLM-tell census is at or under the accepted stretch's rate, and Peri is more active and more likeable here than anywhere since scene 32; 45 in particular is the best new-character scene in the book, because redacted and redacted are both right and the scene knows it. The risks are not quality collapse but sameness: the same discovery move (count something, write it down, say what it means) runs through five of six scenes, the fen's adults are converging on one clipped closing register, and two of the six scenes end on composed buttons after four scenes of clean cuts. 43 is the one scene that has genuinely drifted toward lecture. Nothing here reads like the rejected draft; it reads like a good stretch that is starting to repeat its own successful shape.
Concrete fixes: