Repository navigation
Verdict and workbench: honest left rail, self-help rates, adverbs by function, two-rail layout - #57
Conversation
…measured
Goal box: the fixed fiction checklist (Hook reader fast / Build tension /
Emotional depth / Fast pacing / Strong dialogue) ticked "Fast pacing" on a
self-help introduction and left "Strong dialogue" unchecked as if the book
were failing at it. It is replaced by the genre contract from
genre-expectations.js: what a reader of THIS genre comes for, each item
met / partial / unmet / unknown with its observation and how it was evidenced.
The Intelligence window exposes its memoised book-level build for this, taking
the text and analysis explicitly because the rail renders before the editor is
filled.
Scene Intelligence ("Exposition Heavy · Energy: High · Tension: Steady") were
fiction instruments read off whatever text was open. Replaced by the prose mix
from the passage classifier, plus, for fiction only, the measured tension
direction across the last two quarters.
Repetition: two short sentences of similar length sharing two or more content
stems in the same order are parallel structure ("…disappear through sympathy.
It disappears through decisions."). The finding stays on the record marked
rhetorical, and prose-norms sets it aside with that reason regardless of
passage, so an epigraph too short for the classifier is covered. Top copy
issues and category lists now draw only from the narrative span: a word in the
epigraph is never the book's top issue.
Score cards: "Clean" for any dimension scoring 80+ implied an inspection that
never happened. Self-help dimensions now carry the counts they were built from
("13 imperatives · 5 exercises · 0 numbered steps · 2 how-tos"); fiction cards
show their basis or nothing.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VWzYm8L13Mra3ihjhuCM2U
…r suffix
Self-help: every component was a capped raw count (min(25, imperatives × 2)…),
so any book-length manuscript saturated every cap and scored 100 on Practical
Application by length alone. Components are now rates per 1,000 words ramped
linearly to a stated ceiling, each dimension summing to 100, and the same text
five times over scores the same on every dimension (pinned by test). The
basis line on each card quotes the rate and the count. Also fixes a
pre-existing floor: rates were computed against max(words/1000, 1), so any text
under 1,000 words had every rate under-reported (18/1K reported for a true
61/1K), which quietly punished chapters and excerpts.
Adverbs: a stance adverb (generally, eventually, only, probably…) qualifies the
claim it sits in; it is doing work, and is set aside with that reason in any
passage. A manner adverb is raised with the manuscript's own rate quoted
("this manuscript uses 412 manner adverbs, 5.8 per 1,000 words"). An adverb on
a speech tag is raised regardless of passage. An adverb in a list or stack
("quietly, persistently wrong") is never auto-deleted, because removing one
word leaves punctuation debris; it is offered for editing instead.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VWzYm8L13Mra3ihjhuCM2U
Left rail (the verdict): the gauge with its sub-scores and readability grade, the genre contract, the prose mix, and one card, "Start here": the dimension whose repair would move the overall most, (100 - score) x weight, with the basis it was built from and one button into the workbench. The five score cards that sat here were the right rail's list shown a second time, and the two "+ Improve Opening" buttons opened the same coach; the opening diagnosis is now a finding under Hook / Promise Strength with the coach reached from there. Right rail (the workbench): the large "What does your book know right now?" box becomes one button in the tab row; the eight dimension cards become rows with the basis beside each score, a third of the height, so the findings for the selected dimension sit above the fold. Every detail opens with "Built from:" and the counts or density that produced the number. Dimensions with no locatable copy findings (Evidence & Support, Practical Application) list the chapters where attributed claims or explicit actions are fewest, from the same per-chapter extraction the Intelligence window shows. No score changes; every value shown is an existing field. Both models (self-help and fiction) use the one layout. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VWzYm8L13Mra3ihjhuCM2U
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1bbffa42c7
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if (all1 <= 16 && all2 <= 16 && all1 > 0 && all2 > 0 && Math.max(all1, all2) / Math.min(all1, all2) <= 2) { | ||
| const c1 = content(s1.raw), c2 = content(s2.raw); | ||
| const sharedInOrder = []; | ||
| let from = 0; | ||
| for (const w of c2) { const at = c1.indexOf(w, from); if (at !== -1) { sharedInOrder.push(w); from = at + 1; } } | ||
| parallel = sharedInOrder.length >= 2; |
There was a problem hiding this comment.
Do not treat two shared content stems as sufficient proof of parallel structure. Ordinary adjacent sentences such as “She opened the heavy door and entered the quiet room. The door closed behind her, leaving the room silent.” satisfy this condition, causing both repetitions to receive rhetorical: 'parallel'; ProseNorms.apply then suppresses them from the copy score and actionable findings. Compare sentence structure rather than merely ordered vocabulary, or mark only repetitions participating in a verified parallel construction.
Useful? React with 👍 / 👎.
| return [ | ||
| {k:'plot',name:isNF?'Argument Structure':'Plot Structure',score:scores.plot||0,density:null,issues:0,weight:'9%',evidence:plotEv}, | ||
| {k:'transitions',name:'Transitions',score:scores.transitions||0,density:null,issues:(t.details||[]).filter(x=>!x.supported).length,weight:'7%',evidence:t.smoothRate!=null?t.smoothRate+'% of paragraph boundaries supported':''}, | ||
| {k:'hook',name:'Hook Strength',score:hookScore,density:null,issues:hookScore<80?hookProblems:0,weight:'9%',evidence:hookProblems?hookProblems+' opening problem'+(hookProblems===1?'':'s')+' diagnosed':'no opening problems diagnosed'}, | ||
| {k:'style',name:'Style & Voice',score:scores.style||0,density:null,issues:0,weight:'7%',evidence:st.lexicalDiversity!=null?'lexical diversity '+st.lexicalDiversity+'% \u00B7 sentence SD '+(st.sentenceLengthStdDev??'\u2014'):''}, |
There was a problem hiding this comment.
Include all dimensions used by _computeScoreBundle in this returned list before claiming that _startHere selects the fix that moves the overall most. For fiction and general nonfiction, the overall also contains line, clarity, discipline, efficiency, engagement, and momentum, but none can be selected here; for example, a manuscript with line: 0 and all listed cards above 50 will recommend a smaller-impact listed dimension instead. Nonfiction also renormalizes applicable weights, so the displayed percentages are not its actual shares of the overall.
Useful? React with 👍 / 👎.
| if(isSH&&r.selfHelpScores){ | ||
| const ev=sh.evidence||{}; | ||
| return [ | ||
| {k:'sh_clarity', name:'Clarity & Polish', score:Math.round(sh.clarityReadability||0), density:perPage(copyRaw),issues:copyRaw,weight:'15%',evidence:ev.clarity}, |
There was a problem hiding this comment.
Do not assign copy-finding density as the basis of sh.clarityReadability. That score is calculated from Flesch ease, sentence length, paragraph length, and transition rate, not from copyRaw; because _dimensionBasis prioritizes density, any self-help manuscript with a copy finding displays “Built from: … findings” and hides ev.clarity, even though removing those findings may leave this score unchanged. Keep the findings as separate editorial examples while using ev.clarity for the score basis.
Useful? React with 👍 / 👎.
This branch was successfully deployed
2 participants
Three commits, each building on the last, all in the spirit of the week's brief: numbers read off the manuscript, no hardcoded verdicts, and a layout that gives each rail one job.
1. Left rail: the genre contract, the prose mix, and no verdicts nobody measured (
6ab9b55)genre-expectations.js: what a reader of this genre comes for, each item met / partial / unmet / unknown with its observation and how it was evidenced.2. Self-help model as rates; adverbs judged by what they do (
4483413)min(25, imperatives × 2)…), so any book-length manuscript saturated every cap and scored 100 on Practical Application by length alone. Now rates per 1,000 words ramped to a stated ceiling, each dimension summing to 100; the same text five times over scores identically (pinned by test). On the 58k-word fixture Practical Application went 100 → 65 with its basis visible.max(words/1000, 1), so any text under 1,000 words had every rate under-reported (18/1K reported for a true 61/1K) — chapters and excerpts were quietly marked down.3. Verdict and workbench: give each rail one job (
1bbffa4)(100 − score) × weight, with its basis and one button into the workbench. The five duplicate score cards and both "+ Improve Opening" buttons are gone; the opening coach is reached from its finding.Verification
npm run verify), including newcopy-detectors.test.jsanddetailed-metrics.test.jscases: parallel-structure suppression, stance/manner/list/speech-tag adverb behaviour, self-help length invariance, per-dimension evidence strings.npm run test:browser) with new assertions for the genre contract, prose mix, absence of "Clean", Start here → workbench, "Built from", the Intelligence button, and per-chapter evidence rows on the self-help fixture.Caveats
🤖 Generated with Claude Code
https://claude.ai/code/session_01VWzYm8L13Mra3ihjhuCM2U
Generated by Claude Code