Missing a typo is a bad day. Putting an error into a contract is a different kind of problem, and it is the one that should worry anyone who proofreads legal texts for a living.

Picture a share purchase agreement in which the same party is referred to as "the Purchaser" in clause 4 and "the Buyer" in clause 7. A careful proofreader makes them consistent. Reasonable instinct. Except that both terms were separately defined, they refer to two different entities in a back-to-back structure, and the edit has just merged them. Nothing about that error looks like an error on the page. It reads beautifully.

This is the awkward thing about our specialization. We are asked to change texts that carry legal weight, usually without knowing the commercial background, and always without being the party who pays if it goes wrong. Everything below is about managing that gap: the working rules, where our authority ends, and what needs to be settled before the file is even opened.

The word "proofreading" causes most of the trouble
In my experience, arguments between a proofreader and a client rarely start with the text. They start with the word used to order the job.

A client writes "just proofread this." What they usually mean is a full comparison against the source, a terminology check against their glossary, and someone willing to stand behind the result. Under ISO 17100, the term refers to a target-language check only, performed before release. Those two jobs differ by several hours, a different skill set, and a very different answer to the question of who gets blamed.

The gap usually shows up late, once something has gone wrong. The client asks why the mistranslated indemnity clause got through. The linguist accurately says that nobody asked them to look at the source. Both are right. Both are annoyed. And now the invoice is up for discussion.

So, the first professional thing you do on a legal proofreading job has nothing to do with language. You name the task.
Four different jobs, one word
ISO 17100 separates four operations, and it is worth being fussy about them because clients almost never are.
A check is the translator reviewing their own work. Not a service you sell separately.

Revision means comparing target against source, done by a second person. This is where you catch accuracy and completeness problems, and it is what most people actually want when they say "proofreading".

A review is a reading of the target text alone by someone who knows the subject, assessing whether the text works for its purpose in the field. In legal work, that is often a lawyer rather than a linguist.
Proofreading is a final monolingual pass over the target: typos, mechanics, layout, formatting that has come loose.

ISO 20771 goes further by treating legal translation as its own service, setting out what legal translators, revisers, and reviewers are expected to be able to do. Where a client works to that standard, half the vocabulary problem is already solved. Where they do not, it lands on us.

One sentence usually clears it up: "I'll do a bilingual revision against the source, then a monolingual proofread of the target. I'm not assessing whether the clauses achieve their intended legal effect. That needs counsel." It defines what you are delivering and draws the liability line in the same breath.

Eight rules worth keeping
1. Read the governing-language clause first. Before anything else, find out what the translation legally is. A version prepared for information is subject to a different standard than one signed alongside the source, and both differ from a sworn translation entered into a court file. If the contract says the English text prevails, mistakes in the target version carry different weight than if both versions are equally authentic. That clause tells you how much room you have to move.

2. Rank errors by legal effect, not by how visible they are. A comma out of place in a heading jumps off the page. A dropped "not" in a warranty does not. My order of priority: anything that changes rights, obligations, conditions, deadlines, or modality comes first; then figures and identifiers; then defined terms; then register; then typography.

Modal verbs deserve their own paranoia. Shall, may, must, is entitled to, shall not. In Indonesian, the gap between wajib, harus, dapat, and berhak is the gap between an obligation and an option. Whatever your pair, you will have the same minefield, just in different shoes.

3. Check names and numbers cold. Do this in a pass of its own, with both documents side by side, and do not do it while reading for sense. You will see what you expect to see. Names, dates, amounts, currencies, case numbers, article numbers, cross-references. Where a contract writes a figure twice, in digits and in words, check them against each other.

Date formats catch people constantly. 04/10/1982 is 4 October almost everywhere and 10 April in the United States. Cross-references are worse, because renumbering one clause quietly breaks every reference pointing at it.

4. Defined terms are wiring, not vocabulary. Everything a good writer has been trained to do says avoid repetition. In contract drafting, that training works against you. Once a term is defined, every later occurrence must be identical, because varying it gives someone an argument that the parties meant something different the second time around.

So, if the translator used two renderings for a single defined term, fix it. If they used one rendering forty times and it reads like a stuck record, leave it. That is the point.

5. Leave ambiguity where you found it. "Reasonable efforts". "Material adverse change". "Without undue delay". These are negotiated compromises, and the vagueness is deliberate. Sharpen them, and you have taken a side in someone else's deal.

The same goes for ambiguity that was probably accidental. If the source can be read two ways and the translation keeps both readings open, that may well be correct. Put it in a query. Do not settle it yourself.

6. Don't turn a functional equivalent into a false friend. Consideration, estoppel, trust, discovery, equity: common-law concepts with no clean civil-law twin. Going the other way, akta notaris or somasi, or the difference between penetapan and putusan, carry procedural weight that no single English word captures. Every pair that straddles two legal traditions has its own version of this list.

A translator who knows the field will have picked a strategy: a functional equivalent, a borrowing with a gloss, or a description. Work out which one before you overwrite it. Swapping a considered borrowing for a comfortable-looking near-equivalent is the most reliable way to make a legal translation worse while making it look like you improved it.

7. In sworn work, mirror the source, warts included. Certified translation runs on completeness and fidelity, not improvement. Nothing added, nothing dropped, nothing quietly fixed. If the source spells a name two ways, the translation spells it both ways and notes this in a translator's note. If a stamp is half illegible, you mark it rather than guess it.

Which means your pass covers the apparatus as much as the prose. Are the seals, signatures, letterheads, handwritten notes and margin scribbles all accounted for? Can a reader lay the two documents side by side and follow them page for page? Is the certification wording intact? A tidied-up sworn translation has lost its correspondence with the original, which made it useful as evidence in the first place.

8. Keep corrections and preferences apart, and log both. Mossop's minimal intervention principle is the discipline here: make the fewest changes to reach the required quality, and make no change you cannot defend.

In practice, two buckets. Corrections are errors and go in as tracked changes. Preferences are improvements and go in as comments. Next to them, keep a query log: the item, where it is, the issue, what you suggest, and who has to decide.

That log is the most useful thing you produce all week. It shows what you found, what you touched, and what you deliberately left alone for someone with the authority to decide. Six months later, when the question comes back, your scope is on paper instead of in your memory.

A case where the language was fine
Here is an example from my own market that shows the ceiling on what we can contribute.
Article 31 of Indonesian Law No. 24 of 2009 requires agreements involving Indonesian parties to be made in Indonesian. In PT Bangun Karya Pratama Lestari v Nine AM Ltd (West Jakarta District Court Decision No. 451/Pdt.G/2012/PN.Jkt.Bar, 2013), the court held an English-only loan agreement null and void for breaching it. The Court of Appeals agreed, and the Supreme Court did too in 2015.

No amount of proofreading on the English text would have rescued that contract. The problem was not linguistic. There simply was no Indonesian version. And Indonesian case law on Article 31 has been inconsistent since, with other decisions and at least one arbitral award going the other way on foreign-language agreements, which is exactly why this was never a translator's call.

Most jurisdictions have some version of this layer sitting above the language. EU legislation is published in every official language, all of which are equally authentic. Canadian federal statutes are enacted in English and French, both of which are authoritative. The legal architecture decides what a translation is long before anyone judges how well it reads.

Good language is necessary. It is never sufficient. Our job is to make the linguistic layer sound and pass everything above it to whoever is qualified to decide.

So should you always ask for instructions?
Yes, but asking is the weak version of the answer.
An open question hands the problem to a client who often cannot solve it. Plenty of people who commission this work, including paralegals, procurement staff and in-house counsel, have never needed to distinguish between revision and proofreading. They will write back "just make sure it's correct." That is not an instruction. It is the same undefined obligation, now with a paper trail.

The better move is to turn up with your own default and ask only for exceptions. "Unless you tell me otherwise, I'll work on this basis." Now they can confirm it in thirty seconds instead of composing a brief.

Mine settles nine things:
1. Service level. Revision, proofreading, or both.
2. Purpose and audience. Information, filing, signature, evidence, publication.
3. Governing or authentic language, and whether the versions rank equally.
4. Certification. Sworn or not, and whose name and seal go on it.
5. Reference material. Glossary, TM, style guide, earlier versions, parallel authentic texts.
6. Who owns terminology when a term is contested.
7. Who answers the query log, and by when.
8. Sign-off wording. What exactly you are attesting to.
9. The point at which you stop and renegotiate.

One line belongs in every one of those: sign off on linguistic accuracy and completeness, never on legal accuracy or fitness for legal purpose. It is honest, it is precise, and it keeps the border between your competence and counsel's visible to everyone in the chain.

When to stop and renegotiate
That ninth item is where this stops being theory.
Sooner or later, you open a job posted as "proofreading, 8,000 words, delivery tomorrow" and find raw machine output or a draft that needs to be translated again from scratch. At that point, you have two honest options: renegotiate or decline. Rewriting a bad translation at a proofreading rate loses twice. You eat the cost and quietly become the author of a text nobody has properly revised, which is an uncomfortable place to be when the document is a court filing.

Agree on the threshold before you start. If the text falls below it, stop, send back a short representative sample with your assessment attached, and wait. Two or three annotated sentences argue the case better than any adjective, and they give the client something to decide rather than something to resent.

Charging by the hour rather than by the word suits legal revision for the same reason. It only works, though, if the specification is tight enough for the client to see what those hours are buying.
One habit

All of the above comes down to a single habit: define the job in writing before you open the file and keep a record of what you changed and why. The craft takes years. The scope takes five minutes, and it is the part you cannot reconstruct afterward.