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

Status: proposal, not a plan. Written 2026-08-11.

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 already is, and which of a user's preferences ought to outlive the draft that prompted them.

Written for someone weighing the language-and-style half of AUTO-IMPROVEMENT.md, or working on #102's 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.

Partly built, as of 5.12.0. #103 shipped the sanctioned edit path (draft-reviser's copy-edit mode) and #104 the recorded target (scope.md's language: line, and WRITING-STANDARDS.md §8), so what this document called the objective function is now normative in that file's §9 rather than proposed here. What remains unbuilt is the machinery: everything under "What persists across drafts" below, which no issue yet covers. The detector (#107) shipped in 5.13.0, and its automatic invocation (#183) in 5.19.0 -- a PostToolUse hook per write and a step in all nine skills.

The rule that decides everything¶

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

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

The reason behind it is that a continuous score invites the loop to optimise the score instead of the draft -- and that 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 not subtle.

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 a polysyllabic term is indistinguishable, to the metric, 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, and it is worth stating because the index is otherwise such an attractive candidate: it is exactly 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 it, and it was written that way before any of this was contemplated. Its §2 and §4 are not aspirations; they are conformance rules with a decidable answer.

The rule-by-rule triage used to live here and no longer does. It is WRITING-STANDARDS.md §9, which names for each rule whether it is decidable and whether a machine may act on it unattended. That move was deliberate: this document is a proposal, and nothing may be built against a proposal, so the normative copy belongs in the file whose status is reference -- and in one place, so the two cannot drift into disagreeing about the same rule. What §9 says, in one line: the top of that list is genuinely autoresearchable, and the bottom is a review aid's output for a human.

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

#104, shipped. 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 rather than restated in chat each session. That is the recorded target an unattended pass needs, and it needed no new tooling: draft-reviser already read scope.md before any edit.
#103, shipped. draft-reviser has a copy-edit branch -- the sanctioned path for a whole-document edit that touches no evidence, which the skill had no shape for.
#107, open. python -m src.draft style -- dialect consistency against that line, plus §2's banned words -- stdlib-only, exit 0 always, a review aid and explicitly never a gate. That is the detector and the re-check. (Re-homed from the scripts/style_check.py this document first named; scripts/ holds dev tooling and no layer entry point.)
#183, open. What invokes the detector once it exists, so a prose finding arrives without a human remembering to ask for it.

Detector, recorded target, re-check, and a sanctioned edit path -- 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.

One caution #104 already records. §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 #108's 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 (#104, #107)
§2's defect markers	yes, minus "just"	yes
Acronym expanded at first use; term defined once	yes, given the dossier glossary	yes
Terminology and notation used consistently	yes, given a registry	yes (#138's detectors)
Cross-references resolve	yes	yes (#138)
Duplicate or near-duplicate sentences across sections	yes, at a threshold	yes -- `src/overlap_index.py` already indexes n-grams
Reference-list consistency	yes	already deterministic (`src/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 today 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 #104'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, and all four already have a per-draft artefact that nothing reconciles across drafts:

Dialect and house style. #104's 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. 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, and nothing today notices that they have -- for an ordinary term. #190 built the narrower, acronym-shaped slice of this: [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 src.draft style now reports when a draft's own glossary has drifted from it (src/style_acronym_drift.py, draft-reviser's acronym-realignment mode fixes what that reports). Reconciliation for a plain term -- "digital twin" spelled three ways with no acronym in sight -- is still nothing.
The boilerplate allowlist. #128 built this one already, and not quite as first framed here: content/verbatim_allowlist.toml is per-host, gitignored data, the same footing as config.toml, not version-controlled. A phrase waved through once still does not get re-flagged in the next draft on the same host, because the file is one per clone rather than one per draft and persists across everything that clone drafts. But "auditable" now means "in this host's own file," not "visible in the project's git history to every contributor." See PLAGIARISM.md for the shipped design and why.
Recurring refusals. 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 has to survive the altitude change. 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. It proposes; the human accepts -- the same sentence as everywhere else in this proposal, 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 src.draft dossier acronyms-suggest --apply looks like an exception and is not one: 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, not a bypass of it.

Relationship to #102's roadmap¶

Issue #102 already sequences the language work as six PRs, ranked by who is blocked today. Nothing here reorders it; this document adds the reason the first three matter more than their size suggests.

#102's PR	What it is	Why it matters here
#103, shipped	copy-edit mode in `draft-reviser`	the sanctioned edit path -- without it there is nowhere for a prose fix to go
#104, shipped	dialect as a first-class draft property	the recorded target
#105, #106	render-language plumbing, localisable references	unrelated to this loop; the non-English track
#107	`python -m src.draft style`	the detector and the re-check
#108	multilingual corpus support	out of scope, and see the caution above
#183	automatic invocation of the detector	what makes the loop a loop rather than a command someone remembers

Issue #102 ranks #107 fifth, as "a nice-to-have consistent with the project's review-aid posture". That ranking is right for a human-driven workflow and wrong for an unattended one: without #107 there is no binary check, and without a binary check the prose class of the agenda cannot be acted on at all. Issue #183 takes that reading and moves #107 ahead of #105 and #106.

One thing #183 settled that this document left open: the check is decidable, so the question of gating it was live. It is not gateable, and 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".