How do you post-edit machine translation so that a multilingual document is not merely smooth English, French or Spanish but actually faithful to its source? ISO 18587:2017 is a useful reference for understanding full human post-editing of machine translation output. This practical guide shows how post-editors examine meaning, missing conditions, terminology, numbers, privacy, variables and final publication quality.
People searching for ISO 18587 machine translation post-editing requirements, full post-editing versus light post-editing, MTPE quality assurance, human post-editing checklist or professional machine translation review usually face a live production decision. Should the organization edit an automated draft, order a fresh human translation, or stop because the source is too sensitive or the output too unreliable? The right answer depends on purpose and risk, not on how naturally the machine-generated sentence sounds.
ISO 18587:2017 is currently published and covers full human post-editing processes and post-editor competences for machine translation output. ISO also lists a new draft under development; that draft is not the published 2017 edition. This educational guide relies on the official public scope and practical risk examples. It does not assert access to or reproduce the full standard and does not imply that post-editing alone guarantees safe legal, clinical or regulated use.
Primary standard reference: ISO 18587:2017 — ISO catalogue. The following source-text snippets are invented teaching examples, not quotations from the ISO standard. They illustrate the kind of communicative decision a client or language professional must verify.
The practical problem this guide solves
Modern machine output can read like original prose. That quality creates a special checking problem: the reader becomes less suspicious precisely when an omission or invented condition may be most consequential. Post-editing is not a race to remove red underlines; it is a structured comparison between what the source commits to and what the target now asks a reader to believe or do.
A phrase has at least three lives: the author intended something, a language professional renders that intention, and a reader or listener uses the result. A service succeeds only if those lives are connected with enough evidence for the consequences involved. What looks like a vocabulary question may actually be about who has authority, whether an action is complete, or what conditions still apply. For this reason, the examples below begin with a realistic source situation rather than a dictionary definition.
Do not treat the examples as translations into a named language: the suggested language below is an English explanatory back-rendering that identifies meaning to preserve. The appropriate final wording must be written or interpreted by someone competent in the actual target language and setting. If the source is wrong or incomplete, language work cannot make the underlying policy or instruction true. The source owner must resolve factual and legal uncertainties.
A practical workflow from source to verified communication
Step 1: Agree the quality target
Decide whether the deliverable is an internal gist, a controlled external publication or a safety-sensitive artifact. In this kind of work, the practical question is not whether someone can read the sentence aloud fluently. It is whether a second person could inspect the evidence, follow the choice and explain why this wording is safe for this audience. Different uses require different review depths, and a light cleanup should never be sold as a full accuracy process. Put the decision into the assignment record, not only into a private chat message, and identify the person allowed to approve it.
Try a simple acceptance test: show the relevant source and proposed target to two competent people who understand their respective roles. Ask each to identify the speaker or actor, the condition that activates the statement, any quantity or deadline, and what the reader is entitled to do next. If their answers conflict, find out whether the difference comes from terminology, source ambiguity, missing context or an unauthorized change of meaning. Fix the actual cause rather than voting for the more elegant sentence.
Step 2: Confirm tool and data permissions
Identify whether confidential or personal data may enter a particular machine translation system. In this kind of work, the practical question is not whether someone can read the sentence aloud fluently. It is whether a second person could inspect the evidence, follow the choice and explain why this wording is safe for this audience. A linguistically good output does not excuse unlawful or unapproved source processing. Put the decision into the assignment record, not only into a private chat message, and identify the person allowed to approve it.
Try a simple acceptance test: show the relevant source and proposed target to two competent people who understand their respective roles. Ask each to identify the speaker or actor, the condition that activates the statement, any quantity or deadline, and what the reader is entitled to do next. If their answers conflict, find out whether the difference comes from terminology, source ambiguity, missing context or an unauthorized change of meaning. Fix the actual cause rather than voting for the more elegant sentence.
Step 3: Set source and output versions
Record the exact source snapshot and any model or workflow used to create the draft. In this kind of work, the practical question is not whether someone can read the sentence aloud fluently. It is whether a second person could inspect the evidence, follow the choice and explain why this wording is safe for this audience. Version control makes the subsequent human comparison and later correction reproducible. Put the decision into the assignment record, not only into a private chat message, and identify the person allowed to approve it.
Try a simple acceptance test: show the relevant source and proposed target to two competent people who understand their respective roles. Ask each to identify the speaker or actor, the condition that activates the statement, any quantity or deadline, and what the reader is entitled to do next. If their answers conflict, find out whether the difference comes from terminology, source ambiguity, missing context or an unauthorized change of meaning. Fix the actual cause rather than voting for the more elegant sentence.
Step 4: Compare propositions before style
Check actor, action, denial, obligation, exception, timing, quantity and referenced entity. In this kind of work, the practical question is not whether someone can read the sentence aloud fluently. It is whether a second person could inspect the evidence, follow the choice and explain why this wording is safe for this audience. Meaning errors often survive surface proofreading because the target sounds idiomatic. Put the decision into the assignment record, not only into a private chat message, and identify the person allowed to approve it.
Try a simple acceptance test: show the relevant source and proposed target to two competent people who understand their respective roles. Ask each to identify the speaker or actor, the condition that activates the statement, any quantity or deadline, and what the reader is entitled to do next. If their answers conflict, find out whether the difference comes from terminology, source ambiguity, missing context or an unauthorized change of meaning. Fix the actual cause rather than voting for the more elegant sentence.
Step 5: Inspect tokens and structure
Test placeholder integrity, headers, links, tables, captions and unprocessed text. In this kind of work, the practical question is not whether someone can read the sentence aloud fluently. It is whether a second person could inspect the evidence, follow the choice and explain why this wording is safe for this audience. Machine output can quietly damage a working interface or skip a segment. Put the decision into the assignment record, not only into a private chat message, and identify the person allowed to approve it.
Try a simple acceptance test: show the relevant source and proposed target to two competent people who understand their respective roles. Ask each to identify the speaker or actor, the condition that activates the statement, any quantity or deadline, and what the reader is entitled to do next. If their answers conflict, find out whether the difference comes from terminology, source ambiguity, missing context or an unauthorized change of meaning. Fix the actual cause rather than voting for the more elegant sentence.
Step 6: Use qualified subject review
Escalate medical, legal, engineering or safeguarding decisions to the responsible professionals. In this kind of work, the practical question is not whether someone can read the sentence aloud fluently. It is whether a second person could inspect the evidence, follow the choice and explain why this wording is safe for this audience. A post-editor should correct language without inventing authority over substantive facts. Put the decision into the assignment record, not only into a private chat message, and identify the person allowed to approve it.
Try a simple acceptance test: show the relevant source and proposed target to two competent people who understand their respective roles. Ask each to identify the speaker or actor, the condition that activates the statement, any quantity or deadline, and what the reader is entitled to do next. If their answers conflict, find out whether the difference comes from terminology, source ambiguity, missing context or an unauthorized change of meaning. Fix the actual cause rather than voting for the more elegant sentence.
Step 7: Measure errors by consequence
Separate major omissions, fabricated claims, terminology failures and minor style issues. In this kind of work, the practical question is not whether someone can read the sentence aloud fluently. It is whether a second person could inspect the evidence, follow the choice and explain why this wording is safe for this audience. Counting edits is not equivalent to testing whether a reader can take the right action. Put the decision into the assignment record, not only into a private chat message, and identify the person allowed to approve it.
Try a simple acceptance test: show the relevant source and proposed target to two competent people who understand their respective roles. Ask each to identify the speaker or actor, the condition that activates the statement, any quantity or deadline, and what the reader is entitled to do next. If their answers conflict, find out whether the difference comes from terminology, source ambiguity, missing context or an unauthorized change of meaning. Fix the actual cause rather than voting for the more elegant sentence.
Step 8: Release with evidence and update plan
Document checks, corrections, exceptions and who can approve new source versions. In this kind of work, the practical question is not whether someone can read the sentence aloud fluently. It is whether a second person could inspect the evidence, follow the choice and explain why this wording is safe for this audience. A publishing process must support correction of both the current target and obsolete copies. Put the decision into the assignment record, not only into a private chat message, and identify the person allowed to approve it.
Try a simple acceptance test: show the relevant source and proposed target to two competent people who understand their respective roles. Ask each to identify the speaker or actor, the condition that activates the statement, any quantity or deadline, and what the reader is entitled to do next. If their answers conflict, find out whether the difference comes from terminology, source ambiguity, missing context or an unauthorized change of meaning. Fix the actual cause rather than voting for the more elegant sentence.
Worked casebook: ISO 18587:2017 Machine Translation Post-Editing, Human Review and Release Quality
Each of the following cases isolates a different reason apparently good language can fail. The situations are illustrative. The proposed wording describes meaning that requires competent rendition, not an official standard clause, legal instruction or automatically acceptable target text.
Case 1: Full versus light post-editing
Consider a product team wants customer-ready copy but orders a quick machine-output cleanup. The working English source message is: “Fix only obvious typos and release it.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “Light cleanup establishes publication-grade accuracy.” The danger is not merely that a different word was chosen. Editing effort is a delivery decision, not a cosmetic label. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Specify full human post-editing when target text must satisfy the agreed quality objective.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Case 2: Adequacy before fluency
A tempting but unsafe version would amount to: “Fluency demonstrates that every condition survived.” The danger is not merely that a different word was chosen. Smoothness can make a silent omission harder to notice. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Compare each proposition, restriction, agent, quantity and consequence to the source.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider a machine output reads naturally but drops a condition from the source. The working English source message is: “It sounds fluent, therefore it is complete.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
Case 3: Negation
A safer working instruction is: “Mark negation and exception scope for deliberate bilingual checking.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider a returns page says refunds are not available for certain items. The working English source message is: “Refunds are available for those items.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “The eligibility boundary remained intact.” The danger is not merely that a different word was chosen. A single missing negative reverses the company’s promise. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
Case 4: Numbers and currencies
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider a price table contains a decimal comma and a currency symbol. The working English source message is: “Change both to local style without comparing values.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “The converted amount is identical to the original.” The danger is not merely that a different word was chosen. Machine translation cannot infer a new price or exchange rate. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Separate formatting localization from any authorized currency conversion.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
Case 5: Product model numbers
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider a maintenance guide contains similar device identifiers differing by one digit. The working English source message is: “Normalize the identifiers to the most common model.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “An identifier is ordinary vocabulary that may be paraphrased.” The danger is not merely that a different word was chosen. A readable but wrong model number can lead to incompatible parts. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Copy and compare identifiers character by character against the reference list.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Case 6: Unit conversion
Consider an equipment manual lists pressure in psi for a regulated audience. The working English source message is: “Silently change the number to an SI-looking value.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “The target number has been technically validated.” The danger is not merely that a different word was chosen. Numeric localization carries a physical-safety risk absent from ordinary grammar. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Retain source values unless approved conversion and engineering verification are defined.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Case 7: Terminology drift
A tempting but unsafe version would amount to: “All privacy responsibilities belong to the same actor.” The danger is not merely that a different word was chosen. A fluent replacement can erase who is accountable for a decision. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Map each role to a defined term and check every obligation against that role.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider a software contract uses controller, processor and subprocessor as separate roles. The working English source message is: “Call all three a service provider.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
Case 8: Gender and reference
A safer working instruction is: “Preserve uncertainty or ask the source author to clarify the referent.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider a customer-support message uses a pronoun whose referent is unclear in the target. The working English source message is: “Choose a person because the machine version picked one.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “The source identified that person unambiguously.” The danger is not merely that a different word was chosen. Grammatical completion is not permission to invent a subject. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
Case 9: Instruction sequence
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider an assembly guide requires isolation before cover removal. The working English source message is: “Reverse the clauses for smoother style.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “The action order is unchanged.” The danger is not merely that a different word was chosen. Correct individual words do not guarantee a safe procedure. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Verify prerequisites, action order, stop conditions and warnings as one sequence.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
Case 10: Formatting tags
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider a help article uses placeholders and HTML variables. The working English source message is: “Translate every token inside braces.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “The website can still inject values correctly.” The danger is not merely that a different word was chosen. A translated token can break rendering even if the sentence looks sound. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Protect code, placeholders, URLs and variables while editing visible language.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Case 11: Machine hallucination
Consider an AI-generated draft adds a help-desk phone number not in the source. The working English source message is: “Keep it because it sounds useful.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “The invented contact is an authorized service channel.” The danger is not merely that a different word was chosen. Post-editing cannot treat model invention as source evidence. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Verify every inserted fact against controlled sources and remove unsupported detail.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Case 12: Bias and respect
A tempting but unsafe version would amount to: “The tone is equally appropriate in both communities.” The danger is not merely that a different word was chosen. Literal grammar can carry a social harm that the source did not convey. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Compare register, interpersonal intent and audience expectations with native target readers.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider an automated translation introduces a disrespectful form of address. The working English source message is: “Leave it because the grammar is technically valid.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
Case 13: Confidential source input
A safer working instruction is: “Check contract permissions, secure deployment and retention before any non-human processing.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider a hospital asks a freelancer to paste records into a public model. The working English source message is: “The tool’s translation quality makes upload acceptable.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “The content stayed within the authorized data boundary.” The danger is not merely that a different word was chosen. An excellent output can still originate from improper handling of confidential data. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
Case 14: Source ambiguity
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider a policy sentence has two plausible meanings. The working English source message is: “Trust the machine output to decide the intended policy.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “The system interpreted the contract authoritatively.” The danger is not merely that a different word was chosen. A probabilistic guess does not establish policy. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Escalate ambiguity, record both readings and wait for an authorized decision.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
Case 15: Back-translation trap
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider a team uses automatic back-translation as the only quality test. The working English source message is: “Matching English proves equivalence.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “Two machine systems agreeing confirms the original target meaning.” The danger is not merely that a different word was chosen. A shared model bias can survive a round trip. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Have a qualified reviewer compare source and target directly, aided by targeted checks.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Case 16: Style guide overrides
Consider an automated draft ignores the client’s approved names and tone. The working English source message is: “Prefer the most statistically common words.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “Approved term constraints are optional suggestions.” The danger is not merely that a different word was chosen. A generic fluent result may not satisfy a specific publication contract. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Apply brand and domain instructions consistently and record justified exceptions.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Case 17: Low-resource language
A tempting but unsafe version would amount to: “The model has equal evidence for all target varieties.” The danger is not merely that a different word was chosen. System quality is not uniform across languages or subject matter. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Test with competent native professionals and allocate stronger review where evidence is weak.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider a model produces convincing output in a language with scarce evaluation data. The working English source message is: “Assume comparable quality across every language pair.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
Case 18: Untranslated segments
A safer working instruction is: “Perform an explicit coverage check including tables, captions, headers and footnotes.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider a long document includes paragraphs the model silently left in English. The working English source message is: “Deliver the document because the main sections look finished.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “Every source segment has a corresponding target segment.” The danger is not merely that a different word was chosen. A missing segment is easy to overlook in a plausible long file. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
Case 19: High-risk medical content
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider an appointment instruction includes medication timing and contraindications. The working English source message is: “Release after an ordinary spelling check.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “Clinical intent and patient safety have been verified.” The danger is not merely that a different word was chosen. Post-editors should not independently practice medicine while fixing prose. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Require qualified specialist review and an appropriate clinical release owner.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
Case 20: Update from new source
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider a new source revision changes one eligibility threshold. The working English source message is: “Keep last month’s post-edited target without a fresh comparison.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “Earlier sign-off transfers to modified source material.” The danger is not merely that a different word was chosen. An old translation can become wrong without a single typo. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Re-diff the source, update dependent statements and retest the final artifact.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Case 21: Error taxonomy
Consider a manager measures edits only by counting changed words. The working English source message is: “More edited words always means better quality.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
A tempting but unsafe version would amount to: “Edit distance is an accepted measure of semantic correctness.” The danger is not merely that a different word was chosen. A necessary one-character fix can be more important than a page of style polish. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Classify omissions, mistranslations, terminology, locale and function failures separately.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Case 22: Release evidence
A tempting but unsafe version would amount to: “No further process trace is needed.” The danger is not merely that a different word was chosen. Release decisions cannot be investigated without a reproducible trail. Identify the exact additional promise, removed limit, changed actor or unsupported assumption in the tempting version. Then return to the source evidence: who issued the original instruction, how current is it, and is there a document or subject expert able to answer a real ambiguity? A linguist should not manufacture an authorization that does not exist.
A safer working instruction is: “Document source version, processing system, post-editor, revision evidence and exceptions.” This is guidance on the semantic outcome, not a ready-made certified target-language sentence. Before delivery, have an appropriately competent language professional express it in the actual target language with the necessary register and local conventions. Check it both forwards against the source and backwards against the real user task. Where the source lacks enough information, retain an open query rather than disguising it as editorial polish.
For a practical quality check, ask someone who did not draft the text to explain what this version says a person can do, cannot do, must do or may only consider doing. Ask what evidence supports that conclusion, and whether the proposed form can still be understood in its final context: spoken exchange, customer email, screen, form, label or published document. Record the answer under this case heading. The point is not bureaucratic paperwork; it is to prevent a silent semantic change from becoming the basis of a real decision.
Now imagine the underlying source changes next month. A responsible maintainer would need to find this phrase, the term or condition it depends on, the target versions affected, and the reason for the earlier choice. Keep identifiers and review notes that permit that search. If a user reports an error, respond to its consequence first: incorrect eligibility, a wrong number, mistaken scope, missing consent or a lost warning may need immediate correction rather than a routine style update. That maintenance test makes the original quality claim meaningful.
Consider an agency marks a job complete without recording the model and human reviewer. The working English source message is: “The final file proves an adequate workflow.” Its purpose cannot be judged from syntax alone; the intended audience, responsible speaker, time and supporting record determine what a faithful version needs to carry. The first task is to write down the decision a real person could make after reading or hearing this line. If the translated statement changes that decision, the apparent fluency is beside the point.
Questions readers and commissioners often ask
Is full post-editing the same as quick proofreading?
No. Full post-editing requires deliberate source-to-target checking to the agreed standard of delivery. Proofreading alone may never notice a fluent false statement.
For this question, ask to see the exact assignment specification and the responsible decision-maker rather than accepting a marketing phrase as proof. A useful quality question is concrete: what source statement and output would you examine, and what evidence would persuade you the reader understood the same condition? This approach respects the user’s ability to check the service instead of relying on terminology alone.
Can a machine translator decide unclear source meaning?
It can propose a likely reading, but that does not make the reading authorized. Important ambiguities should go back to the person or record responsible for the source.
For this question, ask to see the exact assignment specification and the responsible decision-maker rather than accepting a marketing phrase as proof. A useful quality question is concrete: what source statement and output would you examine, and what evidence would persuade you the reader understood the same condition? This approach respects the user’s ability to check the service instead of relying on terminology alone.
Does ISO 18587 cover every AI system?
The published 2017 standard is framed around machine translation output. ISO has a revision project with broader draft wording, but a draft must not be represented as the current published requirement.
For this question, ask to see the exact assignment specification and the responsible decision-maker rather than accepting a marketing phrase as proof. A useful quality question is concrete: what source statement and output would you examine, and what evidence would persuade you the reader understood the same condition? This approach respects the user’s ability to check the service instead of relying on terminology alone.
When should we translate afresh instead?
Consider a new qualified human translation when machine output quality is weak, confidentiality prevents its use, specialist risk is unusually high, or repairing systematic errors costs more than rebuilding the target.
For this question, ask to see the exact assignment specification and the responsible decision-maker rather than accepting a marketing phrase as proof. A useful quality question is concrete: what source statement and output would you examine, and what evidence would persuade you the reader understood the same condition? This approach respects the user’s ability to check the service instead of relying on terminology alone.
How this specialist chapter connects to eduKate’s existing learning resources
This is a focused field guide to one service decision, not a new master index or a competing translation hub. For the complete architecture of preserving meaning between languages, continue to Master Art of Translation: the complete system for moving meaning between languages. For understanding the development of words, concepts and distinctions, use the protected Vocabulary Learning Hub. For English wording, context and the relationship between language and meaning, read How English Works: translation and meaning across languages. These are existing canonical learning routes; this chapter belongs beneath them.
Final working rule
Good language service does not remove every uncertainty by making the prose sound confident. It preserves every distinction the reader needs: what is known, what remains possible, who is responsible, which condition applies, and what has been approved. Check those distinctions with the source and intended audience. If a claim is not supported, do not add it for the sake of elegance. The reader’s next decision is the most practical test of whether meaning travelled safely.
Primary standard reference: ISO 18587:2017 — ISO catalogue. The following source-text snippets are invented teaching examples, not quotations from the ISO standard. They illustrate the kind of communicative decision a client or language professional must verify.