Written natively, never machine-translated

Every firm that publishes in two languages faces a quiet procurement decision, usually taken by nobody in particular: is the second language a production run or an authorship act? Machine translation with human post-editing costs a fraction of native authorship and ships in hours. The tooling is certified, the output is grammatical, and on a busy week the difference seems cosmetic.

Our answer is written into our build: French editions of Stredge Partners publications are authored natively or they do not ship. That rule is an economic judgment, not a stylistic preference.

The claim worth interrogating is narrow. It is not that machine translation is bad; for gist, logistics and low-stakes correspondence it is excellent and getting better. It is that for decision documents — texts that move budgets, authorise thresholds or bound liability — converted language fails precisely where the document earns its keep, and the failure is invisible to the people best positioned to stop it.

Start with what the translation industry says about itself. ISO 18587, the international standard governing machine-translation post-editing, sets the goal of full post-editing as output "comparable to human translation." The ceiling of the certified process is equivalence to the thing it replaces, and nobody certifies a process by promising it will match the alternative unless matching is the hard part.

Second, the legal architecture of our corridor is not linguistically neutral. The OHADA treaty system, which harmonises business law across seventeen African states, names French among its working languages. The Uniform Acts a Côte d'Ivoire subsidiary lives under, the security interests a lender registers in Dakar, the company forms a holding structure takes in Lomé: the controlling text is French.

Third, the structural argument, which is ours and should be read as a firm position rather than a measured result. Decision prose fails differently from descriptive prose. A description that is ninety-five per cent right is a good description; a threshold that is ninety-five per cent right is a different threshold. The elements that make a document decision-bearing — the caveat that bounds a claim, the modal verb that separates must from should, the invalid-use boundary, the exact scope of an authorisation — are short, low-frequency strings whose meaning is carried by convention within each language, not by dictionary equivalence. These are precisely the strings a statistical system renders most fluently and most plausibly, because plausibility is what it optimises. The error that survives post-editing is not the ungrammatical sentence, which any reviewer catches. It is the grammatical sentence that says a slightly different thing, and catching it requires the one reviewer the post-editing economy does not budget for: someone holding the decision context, not the language pair.

French sharpens the problem. It distributes emphasis through syntax rather than typography and binds terminology through charters and usage rather than casing; its professional register signals, within a sentence, whether the writer reasoned in the language or arrived in it. A francophone board reads that signal whether or not it can name it, and what it reads from converted text under your mark is that the firm's judgment was formed elsewhere and localised afterwards.

The rule therefore binds hardest exactly where our clients operate. A European holding board authorising capex for an Abidjan operating company decides on a French document, because the operating company's counsel, auditors, regulator and bankers work in French, under a French-language legal regime, with a francophone central bank. The English edition is the courtesy copy. In the WAEMU-to-euro corridor the language of record and the language of decision are the same language, and it is not the one most international advisory output is written in. Firms headquartered in anglophone capital markets treat English as the master and French as the derivative; here the derivative is the document that gets acted on, so deriving it mechanically hands the highest-consequence reader the lowest-authorship text.

Evidence cards

The honest alternatives are two: author natively in both languages, freezing the thesis and evidence before the second authorship begins so the editions cannot drift on substance; or publish in one language and say so. Both are defensible. The indefensible middle is a converted edition presented as an authored one, which is why our build gate rejects it and why the discipline is enforced by validation rather than by good intentions. On cost, native authorship of a note like this roughly doubles language production. Against that, price the downside: a covenant threshold read one increment differently by a lender's counsel in Dakar; a board authorisation whose French scope is arguably wider than its English one.

Sources and limitations

Sources and limitations. Institutional facts in this note rest on the ISO 18587 standard text and the OHADA treaty framework, each cited with its caveat above. The structural argument about decision prose is the firm's position, drawn from operating practice in the corridor, and is graded as interpretation, not measurement: we cite no error-rate statistics because we found none we could verify to our standard, and we publish no internal panel data because none has passed our own evidence-release gate. The transferability boundary is explicit: the economics favour native authorship most strongly where the second language is the language of legal record; where it is a courtesy language, lighter processes may be rational. This note is not valid as a claim that machine translation is unfit for informational content, nor as a benchmark of any vendor or tool.

1. The Managing Partner commissions a one-page native review of a French artifact already released. One page, one reviewer who holds the decision context, one question: authored or converted? Run it before a counterparty runs it for you.

2. The General Counsel writes the controlling-language rule into the publication standard this week. When the English and French editions differ on a threshold, a caveat or a scope, one text must control and that must be recorded somewhere. If it is not, the controlling text will be chosen by the counterparty's counsel, after the fact.

3. The editorial owner attaches a validation record to every edition labelled natively authored. A label without a gate is a claim without evidence. Ours is a validation record; yours may differ, but it should exist before the label is used again.

4. The engagement lead re-scores every artifact filed as "low-stakes" before conversion is approved, and signs the score. The costliest conversions in our experience are documents misfiled as descriptive that a reader treated as decision-bearing; the person who confirms the stakes should be named.