✍ House style: the prose axis, and what persists across drafts

Status: reference.

Prose is the axis an unattended loop can improve best, which is the reverse of the usual intuition that language is the soft part and citations the hard one. This document says why, what the objective function is, and which of a user's preferences outlive the draft that prompted them.

Written for someone weighing the language-and-style half of AUTO-IMPROVEMENT.md, or working on the language roadmap. It assumes WRITING-STANDARDS.md, which is the standard being checked against, and DRAFT-ITERATION.md for the dossier.

Not covered here: the loop's own machinery (the agenda aid, the skill, the requirements), which is AUTO-IMPROVEMENT.md, and why the line falls where it does, which is AUTO-IMPROVEMENT-RATIONALE.md.

What is built: the sanctioned edit path (draft-reviser's copy-edit mode), the recorded target (scope.md's language: line, and WRITING-STANDARDS.md §8), the detector (python -m chitragupta.draft style), its automatic invocation (a PostToolUse hook per write, and a step in every drafting and revising skill), the per-host boilerplate allowlist and the per-host acronym vocabulary. What is open, and covered by no issue: a cross-draft dialect default, reconciling plain (non-acronym) glossary terms across drafts, and noticing recurring refusals. What persists across drafts has each.

🧭 Table of contents

• The rule that decides everything
• Why a readability index is a trap
• The objective function already exists
• What is binary, and what only looks it
• What persists across drafts
• Relationship to the language roadmap

🔑 The rule that decides everything

R3: "An unattended item's check is binary. No continuous score is ever the thing being optimised."

From AUTO-IMPROVEMENT.md, which owns the wording.

The reason behind R3 is that a continuous score invites the loop to optimise the score instead of the draft, and keeping to binary checks is the whole reason prose can carry an unattended loop at all. It is also why the specification counts findings rather than scoring quality: a count of binary conformance failures is safe to drive to zero, and a quality score is not safe to maximise.

Applied to language, the rule cuts the axis cleanly in two. "Apply §2" is binary: "obviously" is present or it is not. "Improve the readability" is not, and no amount of care in the prompt makes it so.

⚠ Why a readability index is a trap

Flesch-Kincaid and its relatives look like the val_bpb this pipeline is missing: one number, cheaply computed, no model needed, comparable across revisions. They are the wrong number, and the failure is plain to see.

The indices are functions of sentence length and syllable count. A loop minimising grade level will split sentences past the point where the argument survives the break, and will replace precise technical vocabulary with shorter, vaguer words, because to the metric a polysyllabic term is indistinguishable from bad writing. The result scores better and reads worse, and every one of those edits passes its own re-check.

That is Goodhart's law with a specific mechanism. It is worth stating because the index is otherwise an attractive candidate: it is the sort of check that would be adopted for being measurable rather than for being right.

An index may still be reported. A grade level that moves sharply between revisions is worth a human's attention. It may never be the thing being optimised.

🎯 The objective function already exists

WRITING-STANDARDS.md is the objective function. Its §2 and §4 are conformance rules with a decidable answer.

The rule-by-rule triage is WRITING-STANDARDS.md §9, which names for each rule whether it is decidable and whether a machine may act on it unattended. It lives there, in one place, so the two pages cannot drift into disagreeing about the same rule. In one line, §9 says that the top of that list is autoresearchable and the bottom is a review aid's output for a human.

The apparatus around it, and none of it needs a model:

• The recorded dialect. The draft's dialect is a language: line in the dossier's scope.md (BCP-47: en-GB, en-IN, en-US), so the target is on disk instead of restated in chat each session. That is the recorded target an unattended pass needs; draft-reviser reads scope.md before any edit.
• The copy-edit mode. draft-reviser has a copy-edit branch: the sanctioned path for a whole-document edit that touches no evidence.
• The detector. python -m chitragupta.draft style checks dialect consistency against that line, §2's defect markers, acronyms, table, figure and equation ids and references, and what has to fit the page (§9 lists every rule). It always exits 0, and is a review aid, never a gate. That is the detector and the re-check.
• Automatic invocation. A PostToolUse hook (.claude/hooks/style_check_hook.py) runs the detector on every write, and every drafting and revising skill runs it before presenting, so a prose finding arrives without a human remembering to ask for it.

Detector, recorded target, re-check, and a sanctioned edit path make up the whole loop, and every deterministic part of it costs zero tokens. This is also why the language half sits on the cheap rungs of the cost ladder, and therefore why it is the half worth building first.

The dialect issue already records one caution. §2's list is English literals and §4's voice rules are an Anglophone convention. A non-English draft needs them adapted, not transliterated, so none of this generalises to the multilingual track for free.

⚖ What is binary, and what only looks it

Wider than §9, and for a different purpose: §9 is normative over WRITING-STANDARDS.md's own rules, and this is the inventory of everything an unattended loop might reach for, most of which that document never mentions. Where the two overlap, §9 is the one that governs; the rows below restate its verdicts only so the inventory reads as one list.

Property	Binary?	Unattended?
Dialect conformance against scope.md's language:	yes	yes when the dialect comes from scope.md or --language; a person decides when it comes from config.toml
§2's defect markers other than "just"	yes	yes
"just"	no: the adjective ("a just outcome") is correct prose	no; a person decides
Acronym expanded at first use, and matching the vocabulary	yes	yes
A plain term defined once and used consistently	yes, given the dossier glossary	not checked
Terminology and notation used consistently	yes, given a registry	no: the book-scale registry check reports for a person, and no agenda class carries it
Cross-references resolve	yes	no for a book's registries (as above); yes for table, figure and equation references inside one draft (draft style)
Duplicate or near-duplicate sentences across sections	yes, at a threshold	could be, since chitragupta/overlap_index.py indexes n-grams; no aid checks it
Reference-list consistency	yes	already deterministic (chitragupta/references.py)
A section citing nothing at all	yes	surfaced; the fix is evidence, not wording
Sentence length, hedging density, passive-voice ratio	no: these are scores	surfaced only
Readability index	no	reported, never optimised
Whether a paragraph leads with its point	heuristic	surfaced
Whether a source supports a claim	no	never (why)

🗄 What persists across drafts

A user of this pipeline writes many documents over years (a survey, a thesis chapter, a textbook chapter, tutorials), and each one starts from a blank slate on everything except the corpus. Their house style is re-derived, or re-stated in chat, every time.

The relevant precedent is the dialect issue's framing: a preference "had nowhere on disk to live", so it silently reverted to the model's default. Four more preferences have the same shape. Each has a per-draft artefact; two are reconciled across drafts and two are open:

• Dialect and house style (open). The language: line is per-draft by design, because a thesis at an Indian university and an IEEE submission legitimately differ. But a user who has chosen en-GB four times has a default, and re-choosing it is friction rather than a decision.
• The glossary (acronyms built, plain terms open). Every dossier carries one. An author writing a thesis, a survey and a textbook chapter on the same subject should not define "digital twin" three different ways. The acronym-shaped slice is built: [style].acronyms in config.toml gives a user's own expansions a per-host home (content/acronyms.toml, the same gitignored footing as the boilerplate allowlist below), every genre skill drafts from it, and python -m chitragupta.draft style reports when a draft's own glossary has drifted from it (chitragupta/style_acronym_drift.py, draft-reviser's acronym-realignment mode fixes what that reports). For a plain term ("digital twin" spelled three ways with no acronym in sight) there is still no reconciliation.
• The boilerplate allowlist (built). content/verbatim_allowlist.toml is per-host, gitignored data on the same footing as config.toml, and is not version-controlled. A phrase waved through once does not get re-flagged in the next draft on the same host, because the file is one per clone, not one per draft. "Auditable" means "in this host's own file," not "visible in the project's git history to every contributor." See PLAGIARISM.md for the design and why.
• Recurring refusals (open). A user who declines the same kind of source in four drafts has a policy, not four coincidences. rejected.md records each refusal with a reason; nothing reads those reasons across drafts to notice the pattern.

The asymmetry between proposing and accepting has to hold at this level too. In autoresearch the human edits program.md and the agent never does; here program.md is the skills, docs/WRITING-STANDARDS.md, and these standing preferences. A loop that rewrote its own standards from its own accepted-edit statistics would be a machine revising the terms of its own supervision. The loop proposes and the human accepts, as everywhere else in this pipeline, one level up.

What that means concretely: a cross-draft pass may report that en-GB was chosen four times out of four, or that a term is defined three ways, and may offer the edit. It may not write the preference file itself.

python -m chitragupta.draft dossier acronyms-suggest --apply follows the same rule: it writes content/acronyms.toml only when a person at a terminal types --apply, for the draft they named, in that one run. Nothing here runs it on a schedule, inside a revision loop, or because a threshold was crossed. The human typing the flag is the acceptance this rule requires.

🗺 Relationship to the language roadmap

Of the language roadmap's items, three serve this loop and are built: the copy-edit mode (the sanctioned edit path), the recorded dialect (the recorded target), and the detector with its automatic invocation (the binary check and the re-check). Without the detector there is no binary check, and without a binary check the prose class of the agenda could not be acted on at all. The render-language items (render-language plumbing, localisable references) and multilingual corpus support are unrelated to this loop; see the caution about non-English drafts above.

The detector is decidable but not a gate. The reason is narrower than "prose is soft": the gate is measured against the ledger, which is ground truth, and a dialect check is measured against a line someone typed, which can be wrong or stale. See ARCHITECTURE.md's "Layer 4".