Editing for one voice: five decisions held constant, and one paragraph edited twice

A document has one voice when a reader cannot tell where one section stops and another begins. Not because the sections sound alike, but because the same person appears to be making the same decisions at every point: the same formality, the same word for the same thing, the same way of addressing the reader, the same sense of how long a sentence should be, the same willingness to cut. Those five decisions, held constant, produce what readers experience as a voice. Change any one of them halfway through and the document reads as though it were written by two people who did not speak to each other.

That is why editing is not polishing. Polishing is improving sentences one at a time, and a document can be polished sentence by sentence until every line is better and the whole is incoherent — because each improvement was made against a different idea of what the document was for. Editing is the prior act: deciding the five things, recording them where the writer and the editor can both see them, and then applying them without exception, including to sentences that were rather good and no longer fit.

The consistency is the deliverable. A reader will forgive a dull sentence and will not forgive a document that changes its mind about what it calls the client on page seven.

Voice is a set of decisions, not a tone of voice

The phrase “finding your voice” is unhelpful in editorial work because it suggests something discovered rather than decided. A voice in this sense is closer to a stylesheet: a small set of choices, made deliberately at the start, written down, and enforced.

Two consequences follow, and both are practical.

The first is that the decisions can be made before there is any prose to apply them to. An editor who waits to see the draft and then “finds the voice” is doing the work backwards, and will end up with a voice that fits the first two sections — the ones the writer was most comfortable with — and fights the rest.

The second is that the decisions can be written down. That matters more than it sounds. A decision held in one person’s head is enforced unevenly: strictly in the sections that person edits twice, loosely in the ones they pass through once. A decision written into a short list at the top of the document is enforced by whoever reads the list, including an agent working section by section a week later.

Table 1. The five editorial decisions, the question each answers, what records it, and the failure when it is left unrecorded. Source: this publication’s editorial method, applied in one pass over a finished draft.

Table 1 is the checklist for the rest of this piece. Each decision is taken below, with the test that shows whether it has been taken, and each one is then visible in the demonstration — one paragraph, three versions, two edits.

Decision one: register

Register is the answer to two questions asked at once: how formal is this, and how much does the reader already know.

It is decided by naming the reader in a sentence, not by choosing an adjective. “Formal but accessible” decides nothing. “This is written for a finance director who knows the contract, has not read the schedule and will make the decision in a meeting on Thursday” decides a great deal: it rules out the introductory paragraph explaining what a contract is, it rules out slang, and it tells you that any term from the schedule has to be defined at first use rather than assumed.

The practical instrument is the exemplar sentence. Take three sentences from the draft that hit the register exactly, and put them at the top of the working file. Not as models to copy — as calibration. When a section drifts, the drift is visible against them in seconds, which is the difference between an edit that takes an hour and one that takes a day.

Drift in register is the commonest form of two-voice prose, and it has a signature: the formal sections use nominalisations and passive constructions, the informal ones use direct address and short verbs, and the transition between them happens at a section boundary where nobody was looking. It is not fixed by making everything formal. It is fixed by deciding which register the document is in and moving the minority sections to it.

Decision two: terms

Every document of any length needs a term list, and most have one only in the writer’s memory.

The list has three columns and takes ten minutes: the preferred term, the variants that are banned, and the point at which the term is defined. “Client” is preferred; “customer”, “account” and “the organisation” are banned as synonyms for it; it is defined at first use in the scope section. That is the whole of the discipline, and documents without it drift within four pages — a supplier becomes a vendor, an incident becomes an issue, a policy becomes a procedure, and each substitution is invisible to the person who made it and obvious to the reader who is trying to work out whether two things are being discussed or one.

Three rules keep the list from becoming pedantry.

The same thing gets the same name everywhere, including in tables, captions, headings and the change log. A caption that calls it a customer undoes nine pages of consistency.

Different things get visibly different names, and the difference is stated once. If “review” and “assessment” mean different things in this document, say so where both appear; if they mean the same thing, delete one.

And technical terms are defined at first use and then used bare. A document that defines “original equipment manufacturer” and then writes it out in full on every appearance has confused consistency with repetition.

The failure this decision prevents is not stylistic. Two names for one actor in a report about a contractual dispute will, eventually, be read as two actors, usually by the person deciding the dispute.

Decision three: person

Person decides who is speaking and who is responsible, and in organisational documents it is where accountability leaks.

One line settles it: third person throughout, with “we” reserved for the organisation that commissioned the work and never used for the author alone. From that line, three consequences follow automatically. Passive constructions that hide the actor are rewritten unless the actor is genuinely unknown. Instructions are addressed to the reader in the imperative or not at all. And the recommendation is attributed to a named party rather than floating in the middle distance.

The test is mechanical: for every sentence in the active voice, ask who the subject is. If the subject is “it”, “there” or “consideration”, the sentence has no actor and probably should. If the subject is the organisation, check that the organisation actually did the thing. A report in which findings appear without anyone having found them reads as though nobody is answerable for the conclusion, which is not the impression a decision-maker is looking for.

Decision four: rhythm

Rhythm is the part of editing that people call instinct and that is in fact arithmetic.

Two rules do most of the work. A paragraph’s first sentence carries its claim, and the rest of the paragraph supports it; a paragraph whose claim arrives in its last sentence has been written backwards, and the reader has spent the paragraph not knowing what it was for. And sentence length varies deliberately: a document in which every sentence runs to thirty words is exhausting, and one in which every sentence runs to twelve is a list. Three sentences of the same length in a row is a signal to break one of them.

The ceiling is a decision, not a rule of grammar. Twenty-five words is a workable ceiling for a report, with an exception for the sentence that carries a definition and needs its qualifiers. Above that ceiling, the sentence usually contains two ideas joined by a conjunction, and splitting it produces two sentences that each say something.

The practical test is to read the paragraph aloud. Not for euphony — for breath. A paragraph that cannot be read in one breath per sentence is a paragraph in which the reader will lose the verb, and the verb is where the meaning is.

Decision five: what to cut

The last decision is a list of things removed on sight, and it is the one that produces most of the visible improvement.

The hedge. “It is generally understood that”, “arguably”, “to some extent”, “it could be said”. Each one keeps a claim and drops the evidence, so the sentence reads as caution while asserting the thing. Cut the hedge and either the claim stands on a source or the claim goes.

The restatement. The sentence that says again, in different words, what the previous sentence said. Common in documents assembled section by section, because each section was written as though it were the first.

The throat-clearing. The opening of a paragraph that announces what the paragraph will do: “In order to assess whether…”, “This section will examine…”, “It is worth considering that…”. The sentence behind it is the paragraph, and it usually starts one clause later.

The promise. The sentence that commits the document to doing something it then does not do, or does elsewhere: “the evidence is set out in more detail below”. If the evidence is there, the reader will find it; if it is not, the promise is a substitute for it.

A fifth cut belongs on the list for documents of any length: the transition that exists only to transition. “Having established this, we can now turn to…” is a signpost on a one-street town. Cut it and the reader arrives at the next section without ceremony, which is what they wanted.

One paragraph, edited twice

The five decisions are abstract until they are applied. The paragraph below is from a constructed report on a contract renewal. It is not badly written — it is the kind of paragraph that passes review because every sentence is grammatical and the whole says almost nothing.

> Version 1, as drafted. In order to be able to assess whether or not the current arrangement continues to represent value for the organisation, it is important that a review of the existing contractual terms is undertaken. It should be noted that the customer — referred to in the schedule as the client — has, over the course of the last three years, consistently paid within the agreed terms. The supplier’s pricing has, however, increased at each renewal, and it is generally understood that comparable arrangements can be obtained at a lower rate elsewhere. It is worth considering that a decision on renewal will need to be made before the expiry date. This section will examine the evidence in more detail.

Five sentences, a hundred and seventeen words: three of them do no work, and the two that carry a fact hedge it. The register is formal to the point of blockage, the actor has disappeared, one thing has two names, two hedges stand where evidence should be, and the paragraph ends by promising something it does not deliver.

The first edit is a line edit. The decisions stay as they are; each sentence is brought into line with them. Nothing is added, nothing is reordered, and the paragraph remains a preamble.

> Version 2, line edit. To assess whether the arrangement still represents value, the organisation should review the contract’s terms. Over the past three years the client has paid within terms on every invoice. The supplier’s price has risen at each renewal; the comparator quotation obtained in June prices the same scope below the current rate. A decision is needed before the contract expires in March. The evidence for each finding is in Appendix B.

Same five sentences, now seventy words, and each one does something. What changed, sentence by sentence: the throat-clearing at the front became a verb and a subject; the term drift was resolved to client and the reference to the schedule was cut, because it explained a naming inconsistency rather than a fact; the unsourced “generally understood that comparable arrangements” became a dated comparator quotation, because a claim about price that nobody can check is not evidence; the expiry date acquired an actual date; and the promise at the end became a pointer to a document that exists.

That is what a competent line edit produces, and it is not enough — because the paragraph is still a preamble. A reader who stops after it knows what the report is about and not what it found.

The second edit is a structural edit. Here a decision changes: page one of this report must carry the finding, not the topic. The paragraph therefore stops being an introduction and becomes the report’s opening finding.

> Version 3, structural edit. Three years of invoices show the client paying within terms on every one. The supplier has raised its price at each renewal; the comparator quotation obtained in June prices the same scope below the rate now being charged. The choice in March is therefore between renewing at the current rate and moving at the cost of a transition, and both figures are in Appendix B.

Three sentences, sixty-five words. The first states what the records show, with its source in the sentence. The second puts the two prices side by side, which is the finding. The third converts the decision into a choice between two named options, and points at the arithmetic that supports it.

The two edits look similar on the page and are different operations. Version 2 applies the five decisions to a paragraph whose purpose is unchanged. Version 3 changes one of the decisions — what this part of the document is for — and everything downstream follows: the paragraph moves to page one, the section it introduced is cut, and the report’s first sentence is now a finding rather than an announcement.

An editor who only ever performs the first edit will produce documents that are clean and still do not work. An editor who performs the second without the first will produce documents with a strong shape and sentences that undermine it. The order matters: decisions first, line edit second, structural edit last, because the structural edit is the only one that can make earlier line work redundant.

Holding it across ten thousand words

A document of four pages can be held in one head. A document of ten thousand words cannot, and consistency has to be enforced by a pass rather than by attention.

Four things go wrong at length, and each has a check that takes minutes.

Terminology drift. The term list is applied by search, not by memory: every banned variant, every synonym, every abbreviation spelled out after its first use. The search is the edit; an editor who reads for drift finds the drift they expected and misses the rest.

Numbering drift. Tables, figures and sections number consistently and every reference resolves. A cross-reference to Table 4 in a document with three tables is the reader’s problem until it is the editor’s.

Person drift. The “we” that appears in section six and nowhere else is either a definition that was never made or an authorial slip. Both are found by searching for the pronoun.

Register drift. The sections written first are the register; the sections written last, usually under time pressure, are not. Read the last section first, against the exemplar sentences, and the drift appears in a paragraph.

The change log is the artefact this pass produces, and it is worth keeping even when nobody asks for it. Every cut recorded with its reason, every line still waiting on a source, every decision applied where the draft had made a different one. It is what makes a verdict reviewable, and it is what stops the same argument being had twice about the same sentence.

What the edit costs

Every figure in this section derives from one number, and that number was read rather than assumed. A deepseek-v4-flash call on Base under exact, with a short prompt, was quoted at $0.00700 by the live 402 challenge on 31 October 2026, outside the service’s weekday peak windows of 01:00–04:00 and 06:00–10:00 UTC. A long prompt, or the upto ceiling at max_tokens: 1000, is $0.00955. The $0.00875 that circulates as a catalogue figure for this endpoint is Qwen’s rate, and it is named as Qwen’s wherever it appears here.

An illustrative edit of a 2,400-word article in five sections, against a register of thirty claims, prices as follows.

Table 2. An illustrative edit of a 2,400-word article at the measured rate on Base. Source: live 402 handshake, 31 October 2026; call counts are planning assumptions stated below.

The addition is short: 10 + 6 = 16; + 5 = 21; + 3 = 24; + 5 = 29; + 5 = 34; + 2 = 36; + 2 = 38 calls. Thirty-eight calls at $0.00700 is $0.26600, which is $0.00887 for each of the thirty claims, $0.05320 a section and $0.11083 for each thousand words of the article.

Three comparisons bound it. At the metered ceiling of $0.00955 the same thirty-eight calls cost $0.36290, which is $0.09690, or 36 per cent, more — and the edit is the pass most likely to be run with long prompts, because checking a claim means putting the claim and the passage in the same call. At Qwen’s $0.00875 the same calls would cost $0.33250, a quarter more. And 0.50 USDC buys seventy-one calls at the measured rate: the edit uses thirty-eight of them and leaves $0.23400, which is thirty-three more calls and $0.00300.

One comparison is worth stating because it looks like an economy and is not. Running the edit at Qwen’s rate would buy nothing on this document: Qwen’s upto ceiling at max_tokens: 1000 is $0.00875, the same as its exact amount, so metering is free at that allowance and the call costs a quarter more than DeepSeek’s to begin with. Where a second model earns its place it is on a contested claim, in the research station, not on a consistency pass over a draft one model has already read.

What the figure covers is the checking and the editing. The corpus is not in it: the draft has to exist, and someone has to have decided what the article is for before the decisions can be taken. Nor are the hours of the person who chooses the register, argues with the structural edit and signs the verdict. On a document of this length the editing is the cheap part of the work and the judgement is the expensive part, and no measurement of the second is offered here because none has been taken.

Where the edit stops

Four limits, and the first is the one editors are most tempted to hide.

An edit cannot make a source better. It can make an unsupported claim smoother, and it can make a gap in the evidence look like a stylistic choice. Both are failures, and both are prevented only by the register: the claim either resolves to a passage or it is marked as a gap, and an editor who improves the sentence without checking the row has done harm that reads as improvement.

The verdict asserts faithfulness to the register, not accuracy. A clean verdict means every sentence says what its passage says. It says nothing about whether the passages are right, and a register built on a bad corpus produces a clean verdict on a wrong document.

The edit cannot supply the argument. Where a draft has no claim, the decisions can be applied flawlessly and the result is an orderly piece that says nothing. That is a return, not a pass, and the return is a legitimate outcome of an edit — the editor’s job includes saying that the material needs to be rewritten rather than repaired.

And no readability score is offered here as evidence of anything. Sentence-length statistics can be computed and are not a measure of clarity; the rhythm rules above are stated as decisions that produce a readable document, not as a threshold that proves one.

The decision

Write the five decisions down before the first cut, and apply them in that order: register, terms, person, rhythm, cuts.

The order is not arbitrary. Register decides which words are available, so the term list inherits it. Person decides who acts, which changes which verbs survive. Rhythm decides how much of each sentence can be carried, which changes what gets cut. Applied the other way round, each decision fights the one before it, and the document ends up with a voice that changes at every section boundary — which is the failure the whole method exists to prevent.

Then apply one test at the end of the pass, which catches almost everything: read the last section first. If it reads as though it were written by someone who had read the first, the voice held. If it reads as though it arrived from a different document, the drift is in the last section, and the fix is cheaper now than it will be after publication.

Where this goes

• The pillar: x402.review/2026/10/31/content-and-knowledge/ — the four stations, the citation chain and the measured month
• The role that edits: studio.x402.press/roles/curator_editor — the register read into text, the matrix, the flagged items and the verdict
• The role that drafted it: studio.x402.press/roles/writer — the document parts, the worked cost and the refusals
• The service it buys: x402.press/services/deepseek/ — the canonical price, every chain
• The second model: x402.press/services/qwen/ — $0.00875 on the entry tier, and a ceiling that equals it at a 1,000-token allowance
• The reviews: DeepSeek and Qwen — the measured handshakes and what they do not promise
• The price ladder: inference endpoints compared — three endpoints, the metering ceilings and the cost of escalation

Method note

The price was read from the live 402 challenge on Base on 31 October 2026, outside DeepSeek’s weekday peak windows of 01:00–04:00 and 06:00–10:00 UTC; Qwen’s exact and upto amounts were read from its handshake in the same session and were both $0.00875 at a 1,000-token allowance. The catalogue figure of $0.00875 was checked against the handshake, found to be Qwen’s rate, and is named as Qwen’s wherever it appears. The call counts in Table 2 are planning assumptions rather than observations. The report paragraph in the three versions is constructed for this piece: no client, contract or published document is described, and the comparator quotation in it is an illustrative artefact rather than a measured price. The five decisions are stated as the editorial pass this publication applies, not as a standard, and no readability metric is offered as evidence for them.

x402.review is written and edited by N.P. Vincent. The publication is connected to x402.press, which sells the services reviewed here; that interest is disclosed on every piece that touches a service, and it is stated here rather than left to inference.

Filed under Jobs · Tags: editing, voice, style, consistency