Master Art of Translation | The Human Translation System — How Translators, Revisers, Reviewers, Subject Experts and QA Turn Meaning Into Accountable Target Text

EDKSG-TRANS-MASTER-WORLD-080

Human translation is often described as the opposite of machine translation, but that contrast is too simple. People searching for human translation, professional translation, translation workflow, translator versus reviser, translation revision, bilingual review, proofreading, translation quality assurance, subject-matter review or how professional translators work are usually asking a larger question: what does a reliable human translation process actually look like from source text to final publication?

A professional human translation system is not one person reading a sentence and typing another sentence in a different language. It is a sequence of decisions, checks and responsibilities. The translator analyses the source, identifies meaning, applies terminology, drafts the target, resolves ambiguity, researches names and concepts, checks the target against the source and prepares the work for revision. A reviser may then compare source and target bilingually for accuracy and completeness. A reviewer may read the target monolingually for clarity, tone and audience suitability. Subject-matter experts may verify technical concepts. Formal QA can check names, numbers, tags, formatting and consistency. The final translation is therefore produced by a system of human judgment, not by intuition alone.

The hidden advantage of human translation is not that humans never make mistakes. Humans do make mistakes. The advantage is that a human translation workflow can make decisions explicit, query missing evidence, recognise when the source itself is defective, compare competing interpretations, understand social and cultural context, justify trade-offs, assign responsibility and escalate uncertainty instead of silently guessing. The purpose of the human translation system is not to romanticise the translator. It is to organise human expertise so meaning remains accountable from source to target.

What this article owns in the Master Art of Translation architecture

This node owns the complete human translation workflow.

It covers:

translation brief,

source analysis,

translator competence,

domain competence,

research,

terminology,

drafting,

self-revision,

bilingual revision,

monolingual review,

proofreading,

subject-matter review,

formal QA,

queries,

decision logs,

risk routing,

handoff,

approval,

release,

correction,

and professional accountability.

It does not replace the earlier Source Analysis node.

That node owns how the source is understood.

It does not replace Equivalence.

That node owns what should remain stable when languages do not match one-to-one.

It does not replace the Context Stack.

That node owns the evidence that resolves interpretation.

It does not replace the Translation Unit, Terminology System, Translation Memory System or Machine Translation System.

Those are infrastructures and mechanisms.

The Human Translation System owns the people, responsibilities and sequence that turn all those resources into an accountable finished target text.

The hidden problem: “a human translated it” is not a quality method

A translation can be human and still be poor.

The translator can:

misread the source,

choose the wrong term,

omit a qualifier,

copy the wrong number,

misidentify a name,

flatten tone,

write unnatural target language,

mis-handle a quotation,

introduce unsupported explanation,

or fail to notice a source contradiction.

Human status does not certify quality.

Professional quality comes from competence plus process.

The question is not:

Was a human involved?

The question is:

Which competent human performed which task, with what evidence, under what specifications, and with what independent checks?

ISO 17100 and the professional process frame

ISO 17100:2015 remains the current published international standard for translation services in 2026, while ISO is developing its second edition.

Its public abstract frames translation quality as a combination of core processes, resources and other requirements needed to deliver translation services that meet applicable specifications.

This matters because quality is treated as a system.

Not a talent claim.

Not a promise that one translator is “native.”

Not a word-count transaction.

The standard’s current published scope also distinguishes translation services from raw machine-translation output plus post-editing, which belongs to a different process frame.

The architectural lesson is:

define the process before measuring the outcome.

Translation quality as fit for purpose

The European Commission’s Directorate-General for Translation describes quality around accuracy, clarity and fitness for purpose across different text types and audiences.

That is more useful than the idea of one universal “perfect translation.”

A legal act.

A school notice.

A tourism page.

A technical manual.

A research paper.

A museum label.

A software button.

These have different priorities.

The human translation system begins by defining the purpose.

The translation brief

A translation brief should answer:

source language,

target language and locale,

audience,

purpose,

domain,

genre,

publication channel,

terminology requirements,

style requirements,

format constraints,

reference materials,

risk level,

deadline,

review level,

and approval owner.

Without a brief, translators silently invent assumptions.

With a brief, assumptions become policy.

The brief is not administration

A good brief prevents repeated questions.

It tells the translator:

who will read,

why they will read,

what they must understand or do,

what tone is expected,

which terms are controlled,

and how exact the target must remain.

The brief is cognitive infrastructure.

Source intake

Before translation begins, inspect the source package.

Is the file complete?

Is it the latest version?

Are all pages present?

Are tables readable?

Are images needed for context?

Are links working?

Are footnotes included?

Are tracked changes resolved?

Is OCR reliable?

Are there comments from the author?

Is there a terminology list?

Are there previous translations?

Do not begin sentence-level translation until the source is operationally ready.

Source-version control

A human translator can create excellent work on the wrong source version.

That is still wrong work.

Record the source version.

When the source changes, identify affected segments.

Do not manually guess which paragraphs changed when version-diff tools are available.

Translation quality begins with source identity.

File and format triage

Different formats create different risks.

Word document:

tracked changes, comments, headers, footnotes.

PDF:

reading order, OCR, text extraction.

Spreadsheet:

cell context, formulas, hidden rows.

PowerPoint:

layout, text boxes, notes.

Website:

links, metadata, forms.

Software resource:

keys, placeholders, variables.

Human translation is partly document engineering.

Translation readiness

A source is translation-ready when the translator has enough information to interpret it reliably.

This does not require perfect writing.

It requires known ambiguity.

If the source contains unclear names, broken sentences, inconsistent terms or impossible numbers, record them.

Do not force the translator to solve authoring defects invisibly.

The pre-translation query

Before drafting, raise high-impact questions.

Examples:

Which person does “they” refer to?

Does “charge” mean fee or electrical charge?

Is 04/05/2026 4 May or 5 April?

Is “may” permission or possibility?

Is this product name translated or retained?

Does this legal term have an approved equivalent?

Queries are not signs of weakness.

They are evidence control.

The translator’s first responsibility: understand before writing

The translator’s first job is not target prose.

It is source meaning.

Read:

title,

introduction,

conclusion,

headings,

definitions,

repeated terms,

captions,

tables,

and surrounding context.

Map the document.

Identify what the text is trying to achieve.

Only then start producing the target.

The translator’s source model

For each difficult sentence, identify:

main proposition,

actors,

actions,

objects,

time,

conditions,

negation,

modality,

references,

terminology,

tone,

and function.

This internal model becomes the basis of target generation.

Human translation is not word substitution.

It is controlled reconstruction.

Translator competence is multidimensional

A professional translator needs more than bilingual ability.

Competence can include:

source-language comprehension,

target-language writing,

translation strategy,

domain knowledge,

research skill,

terminological skill,

technology skill,

cultural knowledge,

and self-review.

A person can speak two languages fluently and still lack translation competence.

Source-language competence

The translator must recognise:

literal meaning,

grammar,

idiom,

register,

implication,

reference,

technical language,

and ambiguity.

Weak source comprehension cannot be repaired by beautiful target writing.

Target-language competence

The target should read as competent language in its genre.

The translator needs control of:

grammar,

collocation,

syntax,

register,

punctuation,

style,

genre conventions,

and local usage.

Translation is also writing.

Domain competence

Domain knowledge narrows interpretation.

“Consideration” in contract law.

“Field” in physics.

“Cell” in biology.

“Object” in programming.

“Yield” in finance or chemistry.

The words are familiar.

The concepts are specialised.

Research competence

A translator needs to know how to verify:

terms,

names,

institutions,

laws,

standards,

citations,

technical concepts,

official spellings,

and real target-language usage.

Search is not merely lookup.

It is evidence gathering.

Terminology competence

The translator should know when a word is:

ordinary vocabulary,

a controlled term,

a defined legal term,

a product label,

an acronym,

or a name.

Terminology consistency preserves concept identity.

Cultural competence

Some meaning is not encoded directly in words.

Forms of address.

Institutional conventions.

School systems.

Legal concepts.

Historical references.

Humour.

Social hierarchy.

The translator needs enough cultural understanding to recognise where mediation is required.

Technology competence

Modern human translation often involves:

CAT tools,

translation memory,

termbases,

corpora,

QA tools,

version comparison,

file filters,

and sometimes AI assistance.

Technology competence does not replace language competence.

It allows language competence to operate efficiently.

Ethical competence

Human translators handle:

confidential documents,

identity information,

legal evidence,

medical content,

private communication,

and unreleased products.

Professional translation includes:

confidentiality,

data care,

conflict awareness,

and knowing when a task exceeds one’s competence.

The translator as evidence manager

A translator constantly asks:

What do I know?

What do I infer?

What remains uncertain?

Which source supports this term?

Which official name should I use?

Which interpretation does context favour?

The translator is not merely a language converter.

The translator manages evidence under linguistic constraints.

Drafting strategy: meaning first

A reliable drafting sequence is:

understand source,

identify invariant meaning,

choose target structure,

apply terminology,

write natural target,

then compare against source.

This prevents target wording from being trapped by source syntax.

Drafting strategy: clause mapping

For long sentences:

find the main clause,

identify subordinate clauses,

map modifiers,

mark conditions,

mark negation,

mark time,

then rebuild.

The target may use different sentence boundaries.

The logic should remain stable.

Drafting strategy: paragraph coherence

Do not translate sentences as isolated objects.

Track:

pronouns,

repetition,

topic,

connectors,

terminology,

and information flow

across the paragraph.

Human translators are especially valuable where discourse relationships exceed one segment.

Drafting strategy: document coherence

The same term should not drift.

A character should not change name.

A product feature should not be translated three ways.

A legal defined term should remain controlled.

A recurring metaphor may need consistent treatment.

A human translation system preserves document identity.

Research during drafting

Do not research every uncertain word immediately.

Classify.

Critical:

research now.

Low-impact stylistic option:

mark and continue.

Repeated term:

resolve before it propagates.

Rare one-off reference:

research when it appears.

Efficient translators allocate attention.

The uncertainty log

Maintain a list of unresolved issues.

TERM?

NAME?

DATE?

REFERENCE?

LEGAL?

CULTURE?

NUMBER?

SOURCE ERROR?

Do not rely on memory.

Clear the log before release.

Translation decision log

For high-stakes or long projects, record major decisions.

Issue.

Options.

Chosen solution.

Reason.

Authority.

Affected sections.

A decision log prevents repeated debate and helps reviewers understand intent.

Translation memory use

Human translators can use TM suggestions.

But they should ask:

Does this old target fit current context?

Is terminology current?

Is the locale correct?

Is the match authoritative?

Human translation remains current-source driven.

Terminology use

When the termbase specifies a preferred term, follow it unless the current context proves the concept is different.

If the termbase seems wrong, query the owner.

Do not silently create a competing terminology policy.

AI assistance inside human translation

Human translation can still use AI.

Possible uses:

alternative phrasings,

source explanation,

terminology brainstorming,

consistency checking,

summarising context,

or drafting low-risk segments.

The translator remains accountable for the result.

Human translation describes ownership of judgment, not absence of tools.

Dictation

Some translators draft target text by voice.

Dictation can increase speed for fluent target production.

It can also introduce recognition errors.

Names, numbers and technical terms need checking.

Method does not change responsibility.

Self-revision begins after drafting

The translator should not submit first-draft output without self-review.

A useful self-revision sequence separates passes.

Pass 1:

meaning.

Pass 2:

terminology.

Pass 3:

names and numbers.

Pass 4:

target language.

Pass 5:

format.

Layered revision reduces cognitive overload.

Meaning pass

Compare source and target.

Check:

actors,

actions,

objects,

time,

conditions,

negation,

modality,

quantities,

attribution,

and uncertainty.

Ignore stylistic perfection until meaning is secure.

Terminology pass

Search critical terms across the target.

Check:

preferred forms,

capitalisation,

abbreviations,

defined terms,

and variants.

Inconsistency is easier to detect globally.

Hard-detail pass

Names.

Numbers.

Dates.

Currencies.

Units.

Identifiers.

URLs.

Codes.

These errors can be operationally serious and visually small.

Target-language pass

Read the target alone.

Does it flow?

Is grammar natural?

Are collocations correct?

Does register fit?

Does the target sound like a competent text in this genre?

Format pass

Headings.

Lists.

Tables.

Captions.

Footnotes.

Tags.

Placeholders.

Links.

Layout.

Translation is not complete if the text cannot function in its document.

The value of temporal distance

When possible, self-revise after a short break.

The translator’s memory of the source can cause the brain to read intended meaning into a flawed target.

Distance makes errors more visible.

Bilingual revision

The European Commission describes revision as a bilingual second-pair-of-eyes process comparing target with source for accuracy and completeness.

This is distinct from ordinary proofreading.

The reviser verifies translation decisions.

What a reviser checks

Meaning.

Completeness.

Terminology.

Names.

Numbers.

Logic.

Reference.

Register.

Consistency.

Potential source errors.

The reviser does not simply rewrite according to personal preference.

Revision versus retranslation

A poor revision process replaces the translator’s voice without evidence.

A strong reviser changes when:

meaning is wrong,

target language is defective,

terminology is noncompliant,

style guide is violated,

or the target does not serve purpose.

Revision should improve quality, not display reviewer taste.

Evidence-based revision

A revision comment should answer:

what is wrong,

why it matters,

what evidence supports the change.

“Prefer my wording” is weak.

“This target changes permission to obligation” is strong.

“This is the official institution name” is strong.

“This term is deprecated in the current glossary” is strong.

Severity in revision

Not all issues are equal.

Critical:

meaning reversal, safety, legal, clinical, financial.

Major:

substantial mistranslation, omission, wrong term, wrong action.

Minor:

stylistic weakness, punctuation, local fluency.

Separating category from severity improves feedback.

Monolingual review

The European Commission describes review as a monolingual second-pair-of-eyes process focused on clarity, tone and suitability for intended readers.

The reviewer may not compare every sentence to the source.

This makes monolingual review complementary to bilingual revision.

What monolingual review catches

Translationese.

Awkward syntax.

Inconsistent tone.

Dense sentences.

Unnatural collocation.

Poor information flow.

Reader confusion.

Genre mismatch.

A target can be faithful and still be poor writing.

Revision and review are different cognitive jobs

Bilingual revision asks:

Does this target faithfully represent the source?

Monolingual review asks:

Does this target work as target-language communication?

Combining both in one pass can work for small projects.

Separating them can improve focus for complex projects.

Subject-matter review

Some content needs technical expertise beyond translation expertise.

Medical.

Legal.

Engineering.

Finance.

Scientific research.

A subject-matter expert can verify:

concept,

procedure,

term,

factual interpretation,

and domain convention.

The expert should not automatically rewrite language outside their linguistic competence.

Translator and subject expert partnership

Translator owns:

cross-language fidelity and target-language communication.

Subject expert owns:

domain correctness.

Together they resolve specialised ambiguity.

Neither role should silently absorb the other’s responsibilities.

Proofreading

Proofreading is usually late-stage target checking.

It can focus on:

spelling,

punctuation,

typography,

formatting,

layout,

cross-references,

and final production defects.

Proofreading should not be confused with bilingual revision.

A polished mistranslation can pass proofreading.

Formal QA

Automated or checklist QA can detect:

numbers,

missing segments,

inconsistent terms,

tags,

placeholders,

punctuation,

capitalisation,

and forbidden terms.

Formal QA is especially valuable for repetitive checks.

Human attention can then focus on meaning.

In-context review

Open the final format.

Website.

App.

PDF.

Slide.

Subtitle.

Form.

Poster.

A translation can be correct in the CAT tool and fail in the product.

In-context review closes the loop.

Risk-based quality control

The European Commission’s current quality framework explicitly links quality-control choices to document type, risk profile, intended use and reader needs.

This is an important principle.

Not every text needs the same number of review layers.

Risk should control review depth.

Low-risk workflow

Example:

internal reference document.

Possible process:

qualified translator,

self-revision,

automated QA,

spot review.

Medium-risk workflow

Example:

public information page.

Possible process:

translator,

self-revision,

bilingual revision or monolingual review,

QA,

in-context check.

High-risk workflow

Example:

legal notice.

Possible process:

specialist translator,

bilingual reviser,

terminology check,

subject expert,

formal QA,

final approval.

Critical workflow

Example:

medical dosage or safety warning.

Possible process:

specialist translator,

independent bilingual revision,

subject expert,

hard-detail verification,

formal sign-off,

version control.

Quality architecture scales with consequence.

The second pair of eyes

Independent review reduces individual blind spots.

A translator can be experienced and still miss:

a number,

a pronoun,

a small omission,

or an overconfident interpretation.

Independence creates error detection.

The second pair of eyes is a system property.

Why independence matters

If a reviewer saw the translator’s reasoning too early, they may anchor to it.

If they first compare source and target independently, they may detect alternative interpretations.

Then the decision log can explain difficult choices.

Independence and communication both matter.

Review handoff

The translator should provide:

brief,

source,

target,

termbase,

style guide,

decision log,

queries,

reference materials,

and unresolved risks.

A reviewer should not reconstruct project context from scratch.

Query handling

Queries should have owners.

Source author.

Client.

Subject expert.

Legal.

Product manager.

Terminologist.

The translator should know where to escalate each issue.

Good query format

Source excerpt.

Problem.

Possible interpretations.

Recommended choice if any.

Consequence of choosing wrong.

Concise queries get better answers.

Query closure

Do not leave answers in email while the translation file remains outdated.

Update:

target,

termbase,

decision log,

and TM where appropriate.

A query is not resolved until its consequences propagate.

Reviewer disagreement

Two competent linguists can disagree.

Classify disagreement.

Meaning?

Term policy?

Style?

Locale?

Preference?

If meaning or policy is involved, escalate.

If both are acceptable stylistic variants, choose one consistently or allow variation.

Adjudication

High-stakes programmes may use a language lead or senior reviser for unresolved disagreements.

The adjudicator should record the principle, not only the one sentence.

This prevents the same disagreement recurring.

Reviewer calibration

Review teams should periodically compare judgments.

Use sample texts.

Discuss severity.

Discuss acceptable variation.

Discuss terminology.

Calibration improves consistency.

Error taxonomies

A shared taxonomy can include:

accuracy,

omission,

addition,

terminology,

name,

number,

logic,

reference,

grammar,

register,

style,

locale,

formatting,

and technical integrity.

Categories support analysis.

Severity supports priority.

Root-cause analysis

When an error repeats, ask why.

Translator knowledge?

Bad source?

Missing glossary?

Wrong TM?

Unclear brief?

Review failure?

File-processing issue?

Fixing the sentence alone does not prevent recurrence.

Corrective action

A translation correction may require:

fix target,

fix terminology,

fix TM,

fix source,

fix style guide,

fix project template,

retrain team,

or change tool settings.

Human translation quality is a feedback system.

The professional workflow map

Intake.

Brief.

Source analysis.

Resource preparation.

Translation.

Self-revision.

Bilingual revision.

Monolingual review if required.

Subject review if required.

Formal QA.

In-context review.

Approval.

Release.

Feedback.

Correction.

Memory update.

This is the human translation system.

The translator’s stop rule

Stop and escalate when:

source meaning is genuinely unclear,

critical terminology has no authority,

identity cannot be verified,

legal effect is uncertain,

number conflicts with source evidence,

or task exceeds competence.

Professionalism includes refusing false certainty.

The translator–reviser distinction

The translator creates the target text from the source.

The reviser independently checks that target against the source.

These are different cognitive positions.

The translator is inside the problem.

The reviser approaches an already-made decision.

That distance can reveal errors the translator no longer sees.

Why translators miss their own errors

Human perception is predictive.

Once a translator knows what a sentence is supposed to mean, the brain can read intended meaning into imperfect target text.

A missing word may be mentally restored.

A wrong number can be overlooked because the surrounding sentence is familiar.

A mistranslated pronoun can feel right because the translator remembers the source referent.

Self-revision is essential.

Independent revision remains valuable.

The translator’s first-pass bias

During drafting, translators optimise forward movement.

They solve one problem and continue.

This creates temporary assumptions.

A term may be marked mentally as “good enough for now.”

A pronoun may be interpreted provisionally.

A difficult sentence may be drafted before later context clarifies it.

The second pass should challenge these provisional decisions.

The reviser’s independence

A reviser should not assume:

the translator understood the source,

the terminology is correct,

the numbers were copied accurately,

or the target’s fluency proves fidelity.

The reviser reads evidence.

This is not distrust.

It is division of labour.

Human translation diagnostic laboratory

The following cases show where human judgment adds value and where human systems can still fail.

Diagnostic 1: correct word, wrong sense

Source:

“The bank raised its reserve requirement.”

A translator chooses a target equivalent meaning river bank.

The sentence becomes absurd.

Easy to catch.

Now consider a subtler case:

“The bank issued a charge.”

“Charge” may mean fee.

A plausible target equivalent meaning accusation can produce a grammatically correct but semantically wrong sentence.

Lesson:

word knowledge is not enough.

Sense selection depends on domain.

Diagnostic 2: correct sentence, wrong referent

Source:

“Lena told Sofia that she should revise the report.”

Who should revise?

The translator chooses Lena.

Later context reveals Sofia.

The target language marks gender or name explicitly, making the error visible.

Lesson:

reference chains must be resolved across context.

Diagnostic 3: correct meaning, wrong social register

Source:

“Could you please send the file?”

Translator preserves proposition but chooses an overly casual target form for a formal client relationship.

Nothing is factually wrong.

The relationship is wrong.

Lesson:

interpersonal meaning belongs to translation.

Diagnostic 4: correct grammar, wrong legal force

Source:

“The tenant may terminate the agreement.”

Translator chooses a target modal closer to “might.”

Permission becomes possibility.

Lesson:

legal effect can hinge on modal semantics.

Diagnostic 5: accurate term, wrong concept scope

Source:

“equity”

In finance, law, education and social policy, the concept differs.

A translator selects an approved target from the wrong domain.

Lesson:

termbase lookup must include concept definition and domain.

Diagnostic 6: exact number, wrong unit

Source:

“5 mg”

Translator copies “5” but changes milligrams to grams.

The number is correct.

The quantity is not.

Lesson:

hard-detail checking must include units.

Diagnostic 7: correct value, wrong decimal convention

Source:

“1,250”

In one source locale, this means one thousand two hundred fifty.

A translator interprets it as one point two five zero.

Lesson:

number meaning can depend on locale.

Diagnostic 8: date ambiguity

Source:

“03/04/2026”

Translator assumes March 4.

Source locale means 3 April.

Lesson:

do not translate ambiguous numeric dates without source-locale evidence.

Diagnostic 9: official name improvised

A government agency has an official bilingual name.

Translator produces a literal translation instead.

The phrase is linguistically reasonable and institutionally wrong.

Lesson:

proper names require authoritative lookup.

Diagnostic 10: established translation ignored

A treaty, law, standard or famous work has an established target title.

Translator creates a new one.

Lesson:

historical and institutional continuity can outrank linguistic originality.

Diagnostic 11: source error repaired silently

Source says:

“Submit by Tuesday, 31 February.”

Translator changes date to a plausible one without asking.

The target looks sensible.

The translator has invented source content.

Lesson:

flag contradictions rather than silently repairing high-impact facts.

Diagnostic 12: typo repair changes evidence

A source typo produces a valid but unexpected word.

Translator assumes what the author intended.

Later the author confirms the unusual word was deliberate.

Lesson:

source difficulty and source error are not the same.

Diagnostic 13: omission hidden by fluency

Source contains:

“Applicants must submit proof of residence and proof of income.”

Target contains only residence.

The sentence is still grammatical.

Lesson:

fluency does not reveal missing information.

Diagnostic 14: addition hidden by helpfulness

Source:

“Bring identification.”

Translator writes:

“Bring a passport.”

The target becomes more specific than the evidence.

Lesson:

helpful clarification can be invention.

Diagnostic 15: “at least” disappears

Source:

“at least 18 months”

Target:

“18 months”

An inclusive minimum becomes an exact requirement.

Lesson:

small quantifier phrases can control eligibility.

Diagnostic 16: “up to” disappears

Source:

“up to 50%”

Target:

“50%”

Marketing claim strengthened.

Lesson:

commercial translation also requires evidential precision.

Diagnostic 17: “may” becomes “will”

Source predicts possibility.

Target asserts certainty.

Lesson:

epistemic strength must be preserved.

Diagnostic 18: “should” becomes “must”

Advice becomes obligation.

Lesson:

normative force is part of meaning.

Diagnostic 19: “must not” becomes “need not”

Prohibition becomes absence of obligation.

Lesson:

negative modality is especially dangerous.

Diagnostic 20: causal relation added

Source:

“After the change, complaints increased.”

Target:

“Because of the change, complaints increased.”

Chronology becomes causation.

Lesson:

do not translate inference as proposition.

Diagnostic 21: evidence becomes fact

Source:

“Researchers reported that…”

Target removes attribution.

Lesson:

claim ownership matters.

Diagnostic 22: quotation becomes paraphrase

Source contains direct quotation.

Translator smooths wording and keeps quotation marks.

Now the target presents altered words as direct speech.

Lesson:

quotation status is evidential structure.

Diagnostic 23: paraphrase becomes quotation

Source reports what someone said indirectly.

Translator adds quotation marks for readability.

Lesson:

do not invent direct speech.

Diagnostic 24: historical register modernised away

A nineteenth-century letter uses formal or archaic language.

Translator modernises aggressively.

Information survives.

Historical voice disappears.

Lesson:

genre can make style part of evidence.

Diagnostic 25: child voice upgraded

A child’s simple narration becomes sophisticated adult prose.

Meaning survives.

Character changes.

Lesson:

target-language quality does not mean maximum lexical sophistication.

Diagnostic 26: dialect flattened

Source dialogue uses regional grammar and vocabulary.

Translator normalises everything to standard language.

Identity and social positioning disappear.

Lesson:

variation can be meaningful.

Diagnostic 27: sarcasm becomes literal praise

Source:

“Wonderful. Another delay.”

Translator chooses straightforward positive wording.

Lesson:

pragmatics can reverse surface sentiment.

Diagnostic 28: understatement weakened

Source:

“That wasn’t ideal.”

Context indicates serious failure.

Target sounds like mild preference.

Lesson:

pragmatic force requires situational reading.

Diagnostic 29: politeness becomes servility

Translator overcompensates for formality and produces an excessively deferential target.

Lesson:

register is not simply “more polite is better.”

Diagnostic 30: title hierarchy changes

A section heading becomes a sentence.

Subheading relation is lost.

Lesson:

layout and hierarchy contribute meaning.

Diagnostic 31: table relationship breaks

Translator handles each cell individually.

A column header no longer governs the correct values.

Lesson:

tables must be read structurally.

Diagnostic 32: chart label becomes ambiguous

Source label “Rate” is clear next to an axis.

Extracted text lacks chart context.

Translator picks wrong sense.

Lesson:

visual context can be lexical evidence.

Diagnostic 33: footnote marker lost

Target prose is accurate.

Citation relationship breaks.

Lesson:

formal structure matters.

Diagnostic 34: list scope changes

Source intro:

“Applicants must provide:”

Three bullets follow.

Translator converts intro and bullets into independent sentences, weakening the grammatical requirement.

Lesson:

list grammar can encode obligation.

Diagnostic 35: heading translated as slogan

Source technical heading is informational.

Translator creates a catchy target.

Genre changes.

Lesson:

titles have functional constraints.

Diagnostic 36: software button becomes noun

Source “Save” is a command.

Target is a noun meaning savings.

Lesson:

interface context determines part of speech.

Diagnostic 37: “record” ambiguity

Source string:

“Record”

Could mean noun or verb.

Translator guesses.

Lesson:

short strings require context, keys or screenshots.

Diagnostic 38: user variable controls grammar

Source:

“Welcome, {name}.”

Target language requires grammatical agreement depending on user attributes.

Translator treats placeholder as neutral text.

Lesson:

internationalisation constraints can exceed translation.

Diagnostic 39: translation memory creates false confidence

TM gives 100% match.

Translator accepts it.

Current project uses another locale.

Lesson:

retrieval score is not contextual approval.

Diagnostic 40: terminology policy changed

Translator reuses exact approved sentence from last year.

The preferred term changed this month.

Lesson:

current policy outranks historical approval.

Diagnostic 41: source was updated after translator started

A client changes one condition in the source.

Translator finishes old version.

Lesson:

version control is part of quality.

Diagnostic 42: reviewer edits final file only

Reviewer corrects an error in PDF.

TM and source target file remain wrong.

Next project repeats error.

Lesson:

corrections must propagate to reusable assets.

Diagnostic 43: reviewer preference presented as error

Translator uses a valid natural expression.

Reviewer rewrites it to personal preference and marks it “wrong.”

Lesson:

revision needs evidence and acceptable-variation policy.

Diagnostic 44: reviewer misses mistranslation because target is elegant

A beautiful target distracts from source mismatch.

Lesson:

bilingual revision should compare propositions deliberately.

Diagnostic 45: subject expert damages target language

Engineer corrects terminology but rewrites sentences into awkward or ungrammatical target language.

Lesson:

subject expertise and language expertise are complementary.

Diagnostic 46: translator damages domain accuracy

Translator improves prose but replaces a controlled technical term with a stylish synonym.

Lesson:

target fluency should not erase concept control.

Diagnostic 47: source writer answers query ambiguously

Translator asks:

“Does ‘they’ refer to the committee or the suppliers?”

Answer:

“Yes, that’s right.”

Lesson:

queries sometimes need forced-choice structure.

Diagnostic 48: decision made but not documented

Team resolves a term in a meeting.

Another translator later reopens the question.

Lesson:

decisions need durable records.

Diagnostic 49: time pressure removes self-review

Translator submits immediately after drafting.

Small omissions remain.

Lesson:

speed should not eliminate defined quality gates.

Diagnostic 50: over-review destroys deadline without improving risk

Low-risk internal text receives three senior reviewers.

High-risk legal notice waits.

Lesson:

quality resources should follow consequence.

The human translation risk matrix

Classify risk along several dimensions.

Meaning risk:

What happens if the proposition is wrong?

Action risk:

Will a reader take action?

Legal risk:

Are rights or obligations involved?

Health risk:

Could physical well-being be affected?

Financial risk:

Could money or valuation be affected?

Identity risk:

Are names, gender, ethnicity or personal records involved?

Reputational risk:

Could poor language damage trust?

Technical risk:

Can formatting or code fail?

This matrix helps assign review.

Translator selection by task

Do not assign solely by language pair.

Consider:

domain,

genre,

target locale,

subject expertise,

software experience,

security clearance,

and writing ability.

A strong literary translator may not be the best patent translator.

A strong patent translator may not be the best children’s-book translator.

Native language claims

“Native speaker” is an incomplete qualification.

A target-language expert still needs:

source competence,

translation competence,

domain knowledge,

and professional process.

Conversely, multilingual professionals can have near-native or professional writing competence across more than one language.

Evaluate evidence of ability, not labels alone.

Translator test translations

Tests can assess:

source comprehension,

target writing,

research,

terminology,

and instruction following.

Use representative material.

Do not create unpaid tests so large they become real production work.

Translator portfolios

Portfolios reveal writing style.

But confidentiality often limits what professional translators can show.

References, certifications, domain history and controlled tests may be more useful.

Translator subject expertise

Subject expertise can come from:

formal education,

professional experience,

translation specialisation,

or years of domain research.

Ask whether the translator understands the concepts, not only vocabulary.

Translator directionality

Many translators work most strongly into one primary target language.

Others work professionally in multiple directions.

The key is target-language writing quality plus source understanding.

Do not impose simplistic universal rules.

Assignment briefing

When assigning work, provide:

purpose,

audience,

locale,

domain,

style,

terminology,

reference files,

deadline,

file format,

review process,

and contact for queries.

A good assignment prevents avoidable errors.

Reference package

Useful references:

previous approved translations,

termbase,

style guide,

product screenshots,

official names,

legal references,

parallel texts,

subject manuals.

Too many uncontrolled references can conflict.

Label authority.

Authority hierarchy

Example:

current legal requirement,

current client specification,

current termbase,

current style guide,

approved master TM,

recent project reference,

legacy documents,

general web examples.

When references conflict, the hierarchy prevents improvisation.

Source research versus target research

Source research asks:

What does this mean?

Target research asks:

How does the target community normally express it?

Both are needed.

A bilingual dictionary often sits between them.

Monolingual dictionaries

Use source-language dictionaries to confirm sense.

Use target-language dictionaries to confirm target meaning and register.

Bilingual lists alone can hide semantic mismatch.

Corpora

Corpora reveal:

frequency,

collocation,

register,

genre,

and patterns.

They help test whether a candidate phrase is actually used naturally.

Parallel texts

Read authentic target-language documents in the same domain.

They show how the community writes.

A technical manual should sound like a target-language technical manual, not translated English.

Official sources

For institutions, laws, standards, programmes and government terminology, prefer official target-language sources.

Search result frequency is not authority.

The translator’s research stop rule

Research should stop when enough evidence exists for the project risk.

Low-risk word:

one strong dictionary plus context may suffice.

Legal term:

official legislation or specialist authority may be required.

Research depth should follow consequence.

Over-research

Translators can lose time chasing marginal stylistic certainty.

Ask:

Would either option change meaning, term policy, reader function or genre?

If not, choose a natural option and continue.

Professional speed includes knowing when evidence is sufficient.

Draft consistency controls

During drafting, maintain:

name list,

term list,

abbreviation list,

style notes,

and query log.

Externalise memory.

Do not rely on remembering every decision across 20,000 words.

Name list

Record:

source name,

approved target form,

role,

gender if explicitly relevant and known,

transliteration,

notes.

This prevents identity drift.

Abbreviation list

Record:

short form,

full form,

target policy,

first-use rule,

and whether it remains untranslated.

Acronyms can create repeated uncertainty.

Resolve once.

Style sheet

Control:

spelling,

capitalisation,

punctuation,

numbers,

dates,

units,

quotation marks,

headings,

and address forms.

The target should look like one document.

Progress checkpoints

Long projects benefit from checkpoints.

After first section:

verify terminology and style.

Midpoint:

review queries and consistency.

Before final section:

check unresolved decisions.

This reduces end-stage surprises.

Collaborative translation

Multiple translators increase capacity.

They also increase inconsistency risk.

Shared resources are mandatory.

Use:

common brief,

termbase,

style guide,

TM,

decision log,

and communication channel.

Lead translator

A lead translator can:

answer questions,

coordinate terminology,

resolve style,

monitor consistency,

and prepare the package for review.

This role reduces fragmentation.

Parallel translation risk

Two translators may independently solve the same phrase differently.

Neither may be wrong.

The document still becomes inconsistent.

Shared TM and real-time terminology updates help.

Handoff between translators

When one translator hands a section to another, include:

current decisions,

unresolved queries,

new terms,

names,

and known source problems.

Handoff is knowledge transfer.

Revision of collaborative work

A reviser can unify:

voice,

terminology,

and document-level cohesion.

This is especially important when multiple translators drafted separate sections.

Translator–reviser communication

Revision should not be one-way correction.

For complex decisions, dialogue improves both people.

Translator may know source research.

Reviser may see document-wide inconsistency.

The system should capture the final agreed principle.

Feedback quality

Good feedback is:

specific,

evidence-based,

proportionate,

and reusable.

Bad feedback is:

vague,

personal,

absolute about preferences,

or inconsistent.

Review quality affects translator development.

Reviewer burden

A reviewer correcting hundreds of preventable issues is not a sustainable QA system.

Find root causes.

Maybe:

translator mismatch,

poor brief,

bad terminology,

or unrealistic deadline.

Review should not become permanent cleanup.

Translator burden

A translator facing contradictory reviewer preferences across projects will lose efficiency.

Calibrate reviewers.

Publish style decisions.

Create one source of truth.

Conflict resolution protocol

1. identify issue category.
2. gather evidence.
3. distinguish rule from preference.
4. apply authority hierarchy.
5. escalate if unresolved.
6. record final decision.
7. update resources.

This turns disagreement into system learning.

Human translation and confidentiality

Translators may see highly sensitive information before the public does.

Use:

secure file transfer,

approved storage,

least access,

confidentiality agreements,

and clear retention policy.

Do not send sensitive text into unapproved web tools.

Human translation and personal data

Names, addresses, health details and identifiers require care.

Minimise unnecessary copies.

Know where files are stored.

Delete according to policy.

Language work is data processing.

Human translation and conflicts of interest

A translator may be asked to translate:

legal dispute involving a client,

competitive business information,

or material conflicting with professional obligations.

Conflict policy should exist.

Human translation and intellectual property

Translation itself can create copyright and licensing issues.

Clarify:

who owns target text,

whether TM reuse is allowed,

whether reference material can be copied,

and whether translators can retain project resources.

Professional independence

A translator may need to tell a client:

the source is ambiguous,

the requested wording is misleading,

or the deadline makes required review impossible.

Professional service includes uncomfortable information.

The human advantage: knowing when evidence is missing

A good translator can stop and say:

I cannot resolve this safely.

That is a strength.

Automatic systems often produce an answer anyway.

Human translation architecture should protect the right to query.

The human advantage: purposeful ambiguity

Humans can recognise when ambiguity is deliberate.

A joke.

Poem.

Diplomatic phrase.

Literary motif.

Brand slogan.

A translator may choose to preserve uncertainty instead of resolving it.

The human advantage: cross-document interpretation

A translator can recognise:

a later definition clarifies an earlier term,

a chart contradicts a paragraph,

a footnote changes scope,

a legal definition controls twenty pages,

or a character’s speech pattern evolves.

Human attention can integrate heterogeneous evidence.

The human advantage: ethical context

A translator can consider:

community preference,

identity,

cultural harm,

historical naming,

and power relationships.

These are not reducible to word frequency.

The human advantage: justification

A professional translator can explain:

why this term,

why this register,

why this ambiguity was preserved,

why an official name was used,

why a literal translation was rejected.

Justification creates accountability.

But humans need systems

Human strengths fail under:

fatigue,

time pressure,

poor files,

missing references,

and inconsistent review.

Process turns expertise into repeatable quality.

Quality is not heroism

A translation programme should not depend on one brilliant person remembering everything.

Use:

checklists,

resources,

roles,

tools,

and review.

World-class quality is institutionalised attention.

Worked human workflow 1: public information page

Source type:

public-facing explanatory page.

Risk:

medium.

Workflow:

translator receives brief, target locale, style guide and glossary.

Translator scans full page and related navigation.

Translator drafts with attention to terminology, headings and links.

Translator self-revises bilingually.

A second linguist performs bilingual revision on high-risk sections and monolingual review on the full page.

Formal QA checks links, headings, numbers and formatting.

Page is previewed in the CMS.

Lesson:

the workflow is proportionate to public impact without treating every sentence like a legal act.

Worked human workflow 2: school notice

Source:

notice to families about a schedule change.

Important variables:

date,

time,

location,

required action,

tone,

contact information.

Human translator verifies local date and time format, preserves the action and uses accessible target language.

Reviewer checks whether parents can understand what changed and what they need to do.

Lesson:

simple documents can carry high operational value.

Worked human workflow 3: examination paper

Source:

assessment instructions and questions.

Risk:

measurement integrity.

Translator must preserve:

command words,

difficulty,

answer scope,

stimulus references,

and distractor function.

Reviser compares source and target carefully.

Subject expert checks whether the target still tests the same skill.

Lesson:

clarity alone is not the goal; construct equivalence matters.

Worked human workflow 4: technical manual

Source:

maintenance procedure.

Risk:

operational and safety.

Resources:

technical glossary,

part-name list,

previous manual,

diagrams.

Translator maps action sequence.

Reviser checks conditions, warnings, part names and units.

Engineer checks technical correctness.

QA checks repeated terminology and numbers.

In-context review checks labels against diagrams.

Lesson:

technical translation is multidisciplinary.

Worked human workflow 5: legal agreement

Source:

contract.

Risk:

rights and obligations.

Translator identifies:

defined terms,

parties,

modal verbs,

conditions,

exceptions,

cross-references,

and jurisdiction-specific concepts.

Reviser performs full bilingual comparison.

Legal expert reviews concept equivalence where required.

No stylistic synonym is introduced into defined terms.

Lesson:

legal consistency is functional, not monotonous.

Worked human workflow 6: patient information

Source:

medical instructions for general readers.

Risk:

health.

Translator needs:

medical terminology,

plain-language competence,

dosage precision,

and audience awareness.

Medical reviewer checks concepts and instructions.

Linguistic reviewer checks readability.

Hard-detail QA verifies numbers, units and frequency.

Lesson:

accuracy and accessibility must coexist.

Worked human workflow 7: academic paper

Source:

research manuscript.

Translator maps:

research question,

method,

claim strength,

evidence,

limitations,

citations,

and discipline terminology.

Reviser checks hedging and causal language.

Subject expert verifies technical terms.

Target-language editor checks academic style.

Lesson:

academic translation preserves reasoning, not only information.

Worked human workflow 8: marketing campaign

Source:

headline, body copy and call to action.

Translator receives brand brief and legal claim constraints.

Literal translation may be insufficient.

Translator develops target candidates that preserve:

promise,

tone,

audience,

and factual limits.

Brand reviewer checks voice.

Legal or compliance reviewer checks claims.

Lesson:

human translation sometimes becomes constrained creative adaptation.

Worked human workflow 9: software interface

Source:

strings in a localisation platform.

Translator needs:

screenshots,

keys,

character limits,

platform terminology,

and product glossary.

Short strings are translated by function.

Reviewer tests in product.

Engineer checks placeholders.

Lesson:

human language decisions depend on software context.

Worked human workflow 10: subtitle file

Source:

spoken dialogue plus timing.

Translator considers:

speaker,

scene,

tone,

reading speed,

shot change,

and visible information.

Target may compress.

Reviewer watches video, not only text.

Lesson:

audiovisual translation is multimodal.

Worked human workflow 11: museum label

Source:

short scholarly text.

Translator preserves:

object identity,

uncertainty,

period terminology,

culture-specific names,

and visitor accessibility.

Curator verifies scholarship.

Target reviewer checks readability at display length.

Lesson:

museum translation mediates expertise without flattening it.

Worked human workflow 12: historical document

Source:

archival letter.

Translator distinguishes:

source errors,

period spelling,

historical terms,

and deliberate ambiguity.

Editorial policy defines whether to modernise spelling or preserve it.

Historian may verify references.

Lesson:

translation can preserve evidence rather than optimise modern readability.

Worked human workflow 13: news article

Translator preserves:

attribution,

quotation,

uncertainty,

chronology,

names,

and factual boundaries.

Editor checks headline does not overstate article.

Lesson:

journalistic translation must distinguish reporting from interpretation.

Worked human workflow 14: customer-support knowledge base

High repetition makes TM useful.

Translator uses:

product terminology,

policy wording,

and prior approved phrases.

Human reviewer focuses on:

current product behaviour,

user action,

tone,

and clarity.

Lesson:

human translation can be fast when infrastructure is strong.

Worked human workflow 15: live incident notice

Source changes quickly.

Translator receives only confirmed facts.

Names, locations and times are verified independently.

Templates provide stable warning language.

Second reader checks before release.

Lesson:

time pressure increases value of structure.

Worked human workflow 16: financial report

Translator distinguishes:

revenue,

profit,

margin,

percentage,

percentage points,

currency,

and accounting terminology.

Reviewer checks tables against narrative.

Finance expert may review technical sections.

Lesson:

numbers and financial terms are semantic content.

Worked human workflow 17: patent application

Translator controls:

claim terms,

dependencies,

reference numerals,

embodiments,

and technical naming.

Revision is highly formal.

Subject expertise matters.

Lesson:

patent translation combines legal and engineering precision.

Worked human workflow 18: insurance policy

Translator tracks:

coverage,

exclusions,

conditions,

limits,

deductibles,

and defined terms.

A reviser checks scope.

Insurance expert checks domain terms.

Lesson:

one modal or exception can alter coverage meaning.

Worked human workflow 19: public-service form

Translator treats labels, instructions and eligibility text as one system.

Reviewer tests whether target users can complete the form.

Lesson:

form translation is interaction design.

Worked human workflow 20: children’s book

Translator preserves:

reading level,

repetition,

character voice,

humour,

rhythm,

and age appropriateness.

Editor checks read-aloud quality.

Lesson:

professional target writing is audience-specific.

Worked human workflow 21: poetry

Translator analyses:

image,

rhythm,

sound,

lineation,

ambiguity,

and cultural references.

Several target drafts may be tested.

Revision discusses trade-offs rather than one-to-one correctness.

Lesson:

human translation can require artistic judgment that cannot be reduced to literal fidelity.

Worked human workflow 22: game localisation

Translator needs:

speaker,

character biography,

world lore,

quest context,

UI constraints,

and branching choices.

Reviewers test target inside the game.

Lesson:

context can be distributed across systems.

Worked human workflow 23: HR policy

Translator preserves:

employee rights,

obligations,

benefits,

conditions,

and respectful tone.

HR reviewer checks policy alignment.

Lesson:

workplace language can have legal and interpersonal consequences.

Worked human workflow 24: cybersecurity advisory

Translator controls:

vulnerability names,

indicators,

severity,

uncertainty,

attribution,

and remediation steps.

Security expert checks technical concepts.

Lesson:

specialist review protects operational meaning.

Worked human workflow 25: scientific laboratory protocol

Translator preserves:

sequence,

reagent names,

quantities,

temperature,

duration,

conditions,

and safety.

Scientist verifies method.

Lesson:

scientific prose is procedural data.

Human translation quality gates

A gate is a point where the project cannot progress until required evidence is present.

Gate 1:

source accepted.

Gate 2:

brief complete.

Gate 3:

resources complete.

Gate 4:

translation complete.

Gate 5:

translator self-review complete.

Gate 6:

required revision/review complete.

Gate 7:

queries closed.

Gate 8:

QA passed.

Gate 9:

in-context check complete.

Gate 10:

approval recorded.

Gate 11:

release package verified.

Gate 12:

feedback path open.

Quality gates prevent “almost finished” work from becoming published work.

Gate 1: source accepted

Verify:

correct file,

correct version,

all pages,

all attachments,

all languages,

all images,

all tables.

If source is incomplete, stop.

Gate 2: brief complete

No translator should guess:

target locale,

audience,

purpose,

tone,

or required review

when the client can define them.

Gate 3: resources complete

Attach:

termbase,

style guide,

TM,

reference texts,

names list,

product material.

Mark authority.

Gate 4: draft complete

All segments translated?

No hidden cells?

No comments ignored?

No text in images omitted?

Completeness comes before polish.

Gate 5: translator self-review

The translator explicitly confirms self-revision.

This should be a workflow step, not an assumption.

Gate 6: independent quality control

Apply the required second-pair-of-eyes process.

Revision?

Review?

Specialist review?

Depends on risk.

Gate 7: queries closed

No unresolved:

TERM?

NAME?

DATE?

LEGAL?

SOURCE?

Critical ambiguity must not reach release.

Gate 8: formal QA

Run repeatable checks.

Numbers.

Tags.

Terms.

Placeholders.

Missing translation.

Formatting.

Gate 9: in-context review

Check the actual environment.

A target can pass linguistic review and fail visually.

Gate 10: approval

Who approves?

Translator?

Reviewer?

Client?

Legal?

Subject expert?

Make the final authority explicit.

Gate 11: release package

Verify:

correct target file,

correct locale,

correct filename,

correct version,

correct links,

correct metadata.

Delivery errors can undo linguistic quality.

Gate 12: feedback path

After release, users may find defects.

Provide a correction route.

Quality is not frozen at publication.

Human translation decision matrix: translator only

Appropriate for:

low-risk internal text,

short-lived content,

simple familiar domain,

experienced translator.

Minimum:

self-revision and QA.

Not appropriate when independent review is mandated by risk or specification.

Decision matrix: translator plus reviser

Appropriate for:

public content,

important technical documents,

high-value business communication,

content where source fidelity is primary.

Reviser compares source and target.

Decision matrix: translator plus monolingual reviewer

Appropriate when:

target-language publication quality and reader experience need special attention.

Especially useful for:

marketing,

public information,

long-form educational writing.

May supplement rather than replace bilingual revision.

Decision matrix: translator plus subject expert

Appropriate when:

technical concept risk is high.

Engineer.

Doctor.

Lawyer.

Scientist.

Finance specialist.

Subject expert should not become sole linguistic reviewer unless qualified.

Decision matrix: translator plus reviser plus subject expert

Appropriate for:

high-risk specialist content.

This separates:

cross-language fidelity,

independent bilingual check,

and domain validation.

Decision matrix: team translation

Appropriate for:

large volume,

tight deadline,

multiple content streams.

Needs:

lead translator,

shared terminology,

shared style,

central decisions,

and unified review.

Decision matrix: human-first with AI assistance

Translator owns draft.

AI helps with:

research,

alternatives,

comparison,

and QA.

Appropriate for:

complex or creative content where human strategy should lead.

Decision matrix: machine-first post-editing

Machine drafts.

Human post-editor owns final.

This belongs primarily to the Machine Translation System but can sit inside a broader human-controlled programme.

Human accountability remains.

Translator capacity planning

Do not estimate translation only by words.

Workload depends on:

source quality,

domain,

format,

repetition,

research,

review,

and language pair.

A 500-word legal clause can take longer than 2,000 words of familiar support text.

Word-count fallacy

Words are a convenient billing and scheduling unit.

They are not a direct measure of cognitive difficulty.

Plan by risk and complexity as well.

Translation speed

Professional speed comes from:

domain familiarity,

terminology,

TM,

reference resources,

keyboard/tool fluency,

and decision reuse.

Rushing without architecture increases rework.

Deep work

Translation requires working memory.

Interruptions increase:

reference loss,

term rechecking,

and sentence reconstruction time.

Protect focused work blocks for difficult translation.

Batch similar work

Grouping similar sections can improve consistency.

Translate all UI errors.

Then all settings labels.

Then help text.

Context remains stable.

But preserve document order when discourse matters

Literature, argument and long-form prose may depend on previous context.

Do not batch so aggressively that discourse is lost.

Workflow depends on genre.

Research batching

Resolve recurring term questions once.

Do not search the same concept repeatedly.

Store evidence.

Query batching

Send related questions together when possible.

This reduces client interruption.

But do not hold a critical blocker too long.

Revision productivity

Revisers need enough time to read source and target independently.

A revision schedule based only on target word count can underestimate difficult source comparison.

Review depth by risk

Define levels.

Level 0:

translator self-check only.

Level 1:

spot review.

Level 2:

full monolingual review.

Level 3:

full bilingual revision.

Level 4:

bilingual revision plus subject expert.

Level 5:

dual independent or regulated sign-off.

Project policy assigns level.

Sampling

For large low-risk sets, sampling can supplement automated QA.

Random sample.

Risk-based sample.

New translator sample.

New domain sample.

Critical term sample.

Sampling is not suitable when every segment carries high consequence.

Spot checks

Useful spot-check targets:

numbers,

headings,

calls to action,

legal notices,

new terms,

low-confidence translator sections.

Translator confidence flags

A translator can flag:

uncertain term,

source ambiguity,

unverified name,

or domain question.

Do not penalise useful uncertainty reporting.

It directs review attention.

Reviewer confidence flags

Reviewers can also escalate.

A second pair of eyes is not omniscient.

High-quality systems make uncertainty visible.

Red-team translation review

For high-risk content, ask a reviewer to search deliberately for failure.

Could the target reverse a condition?

Could a number be misread?

Could a pronoun refer to someone else?

Could a reader take unsafe action?

Adversarial review is useful when consequence is high.

Back translation

Back translation can be one diagnostic for controlled contexts.

It is not a universal gold standard.

A good natural target may back-translate differently.

Use it where project methodology requires it.

Reconciliation

Some research and clinical workflows use:

two independent translations,

comparison,

reconciliation,

back translation,

and expert review.

This is expensive.

It can be appropriate when measurement equivalence matters.

Parallel independent translation

Two translators independently translate the same content.

Differences expose ambiguity.

This can be useful for:

validated instruments,

sensitive terminology,

or high-stakes short text.

It is unnecessary for ordinary content.

Blind revision

A reviser can sometimes review without seeing translator identity.

This may reduce interpersonal bias.

The value depends on team culture.

Open revision

In collaborative teams, dialogue between translator and reviser can improve learning and consistency.

Transparency also has benefits.

Choose process intentionally.

Reviewer track changes

Revision systems should make changes visible.

Translator or lead can inspect:

meaning changes,

terminology changes,

and stylistic edits.

This supports accountability.

Reviewer comments

Use comments for:

non-obvious changes,

queries,

policy decisions,

and teaching points.

Do not comment on every comma.

Acceptable variation

Two translations can both be correct.

Policies should distinguish:

mandatory term,

preferred style,

and free variation.

Over-standardisation can make natural language rigid.

Document voice

Some content benefits from one consistent target voice.

A lead editor can unify multi-translator work.

This is especially useful for:

books,

reports,

and public-facing guides.

Reader testing

For important public content, test with target readers.

Can they find information?

Understand action?

Interpret terminology?

Reader testing reveals issues experts may miss.

Comprehension testing

Ask readers questions.

Do not ask only:

“Did you like it?”

Test whether intended meaning was understood.

Usability testing

For forms and apps, observe users.

Can they complete the task?

Translation quality can be measured through successful action.

Accessibility review

Check:

plain language,

screen-reader labels,

captioning,

alt text,

and visual order.

Human translation should serve all target readers.

Translation incident

If bad translation goes live:

triage,

contain,

correct,

propagate correction,

analyse cause,

and update resources.

Do not hide the incident.

Learn from it.

Incident root cause: translator error

Correct target.

Provide feedback.

Check similar segments.

Determine whether training or reference was missing.

Incident root cause: reviewer miss

Why did revision fail?

Time?

Ambiguous rubric?

Overwork?

No source access?

Improve system.

Incident root cause: bad source

Fix source if possible.

Update target.

Create source-authoring guidance.

Incident root cause: terminology conflict

Resolve authority.

Update termbase.

Repair TM.

Notify team.

Incident root cause: file-processing error

Correct extraction or filter.

Recheck all affected files.

A language reviewer cannot fix missing text they never received.

Human translation and continuous improvement

Track:

error categories,

query frequency,

review changes,

term disputes,

source defects,

and user feedback.

Patterns show where the system should improve.

Translator feedback loop

Reviewer feedback returns to translator.

Translator learns.

Future draft improves.

This reduces review burden.

Source-author feedback loop

Repeated translator questions reveal source-authoring problems.

Writers learn to:

clarify pronouns,

stabilise terms,

and reduce avoidable ambiguity.

Translation can improve source quality.

Terminology feedback loop

New concepts encountered in translation become:

new termbase entries,

definitions,

and examples.

Translation grows institutional vocabulary.

Style feedback loop

Repeated reviewer choices become documented style rules.

Do not make people rediscover preference every project.

Tool feedback loop

Repeated technical problems can lead to:

better file filters,

better QA,

better macros,

or automation.

Human quality can improve through engineering.

Management feedback loop

Project data reveals:

where deadlines are unrealistic,

where expertise is missing,

and where review is over- or under-applied.

Quality management is operational management.

Measuring human translation quality without reducing it to one number

Human translation quality is multidimensional.

A useful measurement system can include:

accuracy,

completeness,

terminology,

target-language quality,

register,

locale,

format,

technical integrity,

and task success.

Do not collapse everything immediately into one score.

A translation can be fluent and inaccurate.

Accurate and unnatural.

Terminologically correct and socially inappropriate.

Mechanically clean and incomplete.

Separate dimensions make diagnosis possible.

Accuracy dimension

Ask:

Does the target preserve source propositions?

Are actors correct?

Actions?

Objects?

Time?

Conditions?

Negation?

Modality?

Quantities?

Attribution?

Uncertainty?

This is the core bilingual dimension.

Completeness dimension

Ask:

Are all:

sentences,

headings,

list items,

captions,

footnotes,

tables,

and labels

represented?

Omission can hide inside fluent target language.

Terminology dimension

Ask:

Are controlled concepts represented by approved terms?

Are defined terms stable?

Are deprecated forms absent?

Does the same target term preserve the same concept throughout?

Target-language dimension

Ask:

Is grammar correct?

Are collocations natural?

Does syntax follow target norms?

Is punctuation appropriate?

Does the text read as target-language writing rather than source-language structure copied across?

Register dimension

Ask:

Is social distance appropriate?

Is formality correct?

Is politeness consistent?

Is technicality appropriate for reader?

Locale dimension

Ask:

spelling?

date?

currency?

number format?

quotation marks?

address?

institution names?

A broad language label can hide local mismatch.

Format dimension

Ask:

headings?

lists?

tables?

references?

links?

page structure?

No structural loss?

Technical dimension

Ask:

tags?

variables?

code?

identifiers?

placeholders?

Software and structured documents need technical fidelity.

Task-success dimension

Can the reader:

follow instruction,

complete form,

understand warning,

navigate interface,

or answer question?

Functional success is a real quality measure.

Severity

Keep severity separate from category.

A grammar error can be minor.

A grammar error that reverses who acted can be major.

A missing decimal can be critical.

Severity depends on consequence.

Critical error

Possible consequences:

harm,

wrong legal action,

clinical risk,

major financial loss,

loss of rights,

security failure.

Critical errors receive immediate attention.

Major error

Meaning materially altered.

Important information omitted.

Wrong terminology affects understanding.

User action changes.

Major but not necessarily catastrophic.

Minor error

Target remains accurate and functional.

Issue affects:

style,

punctuation,

minor fluency,

or low-impact consistency.

Minor does not mean irrelevant.

It means lower consequence.

Error frequency

A single critical error may matter more than fifty commas.

Do not optimise quality systems around raw error count.

Use severity-weighted analysis carefully.

Reviewer calibration session

Choose ten sample segments.

Each reviewer marks:

category,

severity,

preferred correction,

and rationale.

Then compare.

Discuss disagreements.

Create examples for future policy.

Calibration is especially useful for large teams.

Calibration on acceptable variation

Give reviewers several valid target alternatives.

Ask:

Which are acceptable?

Which violate style?

Which change meaning?

This reduces overcorrection.

Calibration on severity

One missing article may be minor in English.

One missing negative can be critical.

Teach reviewers to score consequence.

Calibration on terminology

Show:

preferred term,

admitted synonym,

deprecated term,

forbidden term.

Reviewers need the same terminology policy.

Revision productivity metrics

Useful measures:

review time,

major errors per thousand words,

critical errors,

translator acceptance of reviewer changes,

reopened decisions,

and repeated error rate.

Do not use metrics to punish honest query reporting.

Translator improvement metrics

Track patterns.

Fewer repeated errors?

Better term adherence?

Better self-revision?

Fewer source misunderstandings?

More useful queries?

Development is qualitative as well as quantitative.

Reviewer quality metrics

Reviewers can make errors too.

Track:

incorrect corrections,

preference edits,

missed critical errors,

and inconsistency.

Review quality should be auditable.

Rework rate

How much translation returns for correction after review?

High rework can indicate:

wrong translator,

poor brief,

bad source,

unclear policy,

or ineffective revision.

Query rate

A high query rate is not automatically bad.

New domain or poor source can create legitimate queries.

Look at query usefulness.

Terminology-dispute rate

Repeated term disputes suggest:

termbase gap,

definition ambiguity,

or governance weakness.

Fix resource.

Source-defect rate

Track source errors discovered during translation.

This can improve authoring.

Translation teams often become excellent source-quality sensors.

Human translation cost architecture

Cost includes:

translation,

research,

revision,

review,

subject expert,

QA,

engineering,

project management,

and rework.

Word price is only one component.

Quality-cost balance

Over-review wastes resources.

Under-review creates risk.

Risk-based architecture aims for appropriate control.

Human translation and deadline architecture

A deadline should include:

translation,

self-review,

revision,

query response,

QA,

and final production.

If all time is allocated to drafting, quality gates become compressed.

Rush work

For urgent projects:

reduce scope,

add qualified resources,

use approved templates,

and prioritise critical review.

Do not pretend 24 hours contains 72 hours of expert attention.

Batch size

Large batches can create fatigue.

Small batches can create overhead.

Choose chunks that preserve context and quality.

Reviewer fatigue

Long revision sessions reduce attention.

Use breaks.

Prioritise critical material.

Automate mechanical checks.

Human attention is finite.

Translator fatigue

Fatigue increases:

omission,

number errors,

and source-mirroring.

Quality management includes workload.

Shift handoff

For round-the-clock translation, handoff notes should contain:

progress,

open queries,

terminology decisions,

and risks.

Time-zone relay without knowledge transfer creates inconsistency.

Global team

A global translation team may include:

project manager,

lead linguist,

local translators,

reviewers,

subject experts,

terminologist,

engineer.

Roles should be visible.

Small team

A small team can combine roles.

One person may translate and manage terminology.

Another revises and proofreads.

The system can remain rigorous without enterprise bureaucracy.

Solo translator

A solo translator can simulate process separation.

Draft.

Pause.

Self-revise bilingually.

Read target monolingually.

Run QA.

Preview final format.

Independence is limited, but layered attention still helps.

Peer review network

Freelancers can arrange reciprocal peer revision for high-stakes projects, subject to confidentiality.

The reviewer needs project brief and authority resources.

Translation agency workflow

Agency can add:

vendor selection,

resource management,

central terminology,

QA,

and client coordination.

Quality depends on actual process, not agency size.

In-house team workflow

In-house translators often benefit from:

deep product knowledge,

direct author access,

stable terminology,

and institutional memory.

Risk:

becoming too familiar with source assumptions.

Independent revision still matters.

Outsourced workflow

External translators may lack context.

Provide:

brief,

screenshots,

glossary,

references,

and query path.

Do not expect vendors to infer internal knowledge.

Mixed in-house/outsource

Use in-house experts for:

terminology,

high-risk review,

and institutional decisions.

External teams can handle volume.

Shared resources keep output coherent.

Professional ethics: competence boundary

A translator should recognise when they lack:

domain knowledge,

language competence,

or technical ability

for a task.

Accepting work beyond competence can create hidden risk.

Escalate or decline.

Professional ethics: confidentiality

Do not upload client content to unapproved systems.

Do not discuss identifiable project details publicly.

Do not retain files beyond policy.

Professional ethics: accuracy over convenience

Do not invent an answer because client is unavailable.

Mark uncertainty.

Ask.

Preserve source ambiguity.

Professional ethics: identity

Respect:

names,

pronouns,

community terminology,

and self-identification

where source evidence supports it.

Do not infer hidden identity characteristics.

Professional ethics: contested terminology

Some terms are politically, culturally or historically contested.

Follow brief, official policy where applicable, source context and target-community awareness.

Do not silently insert personal ideology.

Professional ethics: faithful offensive content

A translator may need to translate:

insults,

hate speech,

or offensive historical language

for evidence, research, moderation or literature.

Fidelity can require preserving unpleasant content.

Context and safety policy decide.

Professional ethics: misinformation

If translating misinformation as source material, do not “correct” it inside the translation unless brief requires annotation.

Preserve claim status and attribution.

Professional ethics: legal advice boundary

A legal translator translates legal content.

That does not automatically make them legal counsel.

Subject expert involvement may be needed for legal interpretation.

Professional ethics: medical advice boundary

Medical translators preserve medical content.

They should not invent clinical advice.

Queries go to qualified medical authority.

Professional ethics: translator visibility

Some projects credit translators.

Others do not.

Authorship and credit policy depends on genre and contract.

Literary translation especially treats translator contribution as visible creative labour.

Decision matrix: choose human translation from scratch

Prefer human-first when:

source is highly creative,

ambiguity is central,

language pair is low-resource,

risk is high,

terminology is unsettled,

or machine output would require heavy rebuilding.

Human-first can still use tools.

Decision matrix: choose human post-editing

Prefer machine-first plus human post-editing when:

content is repetitive,

engine performs well,

terminology is controlled,

and human review remains efficient.

Decision matrix: choose independent bilingual revision

Use when:

source fidelity is critical,

content is public or high value,

or specifications require second-pair-of-eyes comparison.

Decision matrix: choose monolingual review

Use when:

reader experience,

tone,

or publication quality

deserves separate attention.

Especially valuable after literal or multi-translator work.

Decision matrix: choose subject expert review

Use when concept errors can escape language professionals.

Medicine.

Law.

Engineering.

Finance.

Scientific speciality.

Decision matrix: choose reader testing

Use when usability or comprehension is central.

Forms.

Public health.

Websites.

Education.

Instructions.

Decision matrix: choose two independent translators

Use for:

validated research instruments,

very short high-stakes statements,

or cases where interpretive alternatives need comparison.

Expensive but sometimes justified.

Decision matrix: choose back translation

Use when methodology requires it.

Do not treat it as universal truth test.

Decision matrix: choose reconciliation

Use when two translations need a controlled combined solution.

Common in research and instrument adaptation.

Decision matrix: choose no translation

Sometimes the best answer is:

retain original term,

use official bilingual label,

or not translate a name.

Translation is not mandatory for every string.

Decision matrix: transliterate

Use when:

name or term should move across scripts without semantic translation.

Follow established systems and personal preferences where available.

Decision matrix: explain

A culture-specific term may require:

original term + short gloss.

This preserves identity and accessibility.

Decision matrix: adapt

Localisation may adapt:

date format,

currency display,

address,

units,

examples,

or UI convention.

Adaptation must follow brief.

Decision matrix: transcreate

Use for:

slogans,

campaign language,

wordplay,

brand effect

where function matters more than literal form.

Keep factual claims controlled.

Human translation and AI boundaries

A human translator can use AI without surrendering authorship of judgment.

Useful AI tasks:

brainstorm alternative phrasing,

explain source grammar,

compare target options,

check consistency,

search within provided references.

Risky AI tasks:

inventing authority,

guessing legal interpretation,

translating confidential text in unapproved systems,

or replacing final bilingual review.

AI as second opinion

Ask AI:

What alternative readings exist?

Which target term might fit?

What might this pronoun refer to?

Then verify independently.

AI suggestions are hypotheses.

AI as adversarial reviewer

Ask AI to search for:

omissions,

numbers,

modal changes,

and term inconsistency.

This can supplement human review.

It should not replace required independent revision.

AI as source simplifier

AI can paraphrase a difficult source for analysis.

Danger:

paraphrase may distort meaning.

Always return to original source for final translation.

AI as target editor

AI can improve fluency.

Danger:

it may add, omit or strengthen claims.

Compare edited target with source.

Human translator remains accountable

If the translator uses AI, dictionary, corpus, TM or colleague, the final responsibility remains with the professional workflow.

Tool output is not an excuse.

Training programme: week 1 — source understanding

Exercises:

paraphrase source in same language,

identify propositions,

resolve pronouns,

mark ambiguity.

Do not translate yet.

Week 2 — vocabulary and terminology

Exercises:

sense selection,

collocation,

termbase use,

false friends,

term definition.

Week 3 — grammar and logic

Exercises:

negation,

modality,

conditionals,

quantifiers,

reference,

time.

Week 4 — target writing

Exercises:

restructure sentences,

remove translationese,

match genre,

control register.

Week 5 — research

Verify:

official names,

terms,

citations,

institutions,

and domain usage.

Students document sources.

Week 6 — revision

Translate a text.

Wait.

Self-revise in layers.

Compare error types.

Week 7 — peer revision

Students revise each other using a shared rubric.

They must justify changes.

Week 8 — specialist workflow

Translate one technical text with a subject reference package.

Week 9 — tools

TM.

Termbase.

QA.

Version diff.

AI assistance.

Use tools without surrendering judgment.

Week 10 — final project

Complete:

brief,

translation,

query log,

decision log,

revision,

QA,

and reflection.

This teaches translation as system.

Beginner exercise: literal trap

Provide sentence where literal target is grammatical but wrong.

Students explain why.

Beginner exercise: word sense

One common source word.

Five contexts.

Five target choices.

Beginner exercise: pronoun map

Draw arrows from pronouns to antecedents before translation.

Beginner exercise: modal ladder

Rank:

must,

should,

may,

might,

could.

Translate without changing force.

Intermediate exercise: termbase creation

Build 20 concept entries.

Use definitions and examples.

Translate a domain passage.

Intermediate exercise: revision categories

Give flawed target.

Students label:

accuracy,

term,

grammar,

style,

format.

Intermediate exercise: reviewer restraint

Give two valid translations.

Students must not change both.

They learn acceptable variation.

Intermediate exercise: query writing

Present ambiguous source.

Students write one concise client query.

Advanced exercise: risk routing

Twenty project scenarios.

Assign review level.

Justify.

Advanced exercise: blind revision

Revise target without knowing translator identity.

Then compare reasoning.

Advanced exercise: source defect investigation

Give source with contradictions.

Students distinguish:

translation issue,

authoring issue,

and domain issue.

Advanced exercise: multi-translator consistency

Three students translate different sections.

A lead integrates terminology and voice.

Advanced exercise: correction propagation

Discover one error.

Students list all assets needing update:

target,

TM,

termbase,

style,

released page.

Teaching translation to Primary learners

Focus on understanding.

Who?

What?

Where?

When?

What does this word mean here?

Children explain source before translating.

Teaching translation to Secondary learners

Add:

tone,

register,

idiom,

connectives,

pronouns,

modality,

and paragraph cohesion.

Require evidence for choices.

Teaching translation to advanced students

Add:

briefs,

termbases,

revision,

research,

QA,

and professional commentary.

Translation becomes applied reasoning.

Human translation as critical thinking

Every difficult segment asks:

What evidence supports this interpretation?

Which alternatives exist?

What would change meaning?

Which resource is authoritative?

This makes translation a disciplined form of critical thinking.

Human translation as vocabulary learning

Translation forces vocabulary into context.

Words gain:

sense,

collocation,

register,

grammar,

and semantic boundaries.

The Vocabulary Learning Hub owns deeper lexical architecture.

This node shows how humans deploy that knowledge.

Human translation as English learning

Translating to or from English exposes:

syntax,

articles,

tense,

modality,

reference,

information structure,

and discourse.

How English Works owns those mechanisms.

The Human Translation System uses them operationally.

Human translation as writing

A translator is also a target-language writer.

The target should communicate without requiring the reader to imagine the source.

This is why target-language craft matters.

Human translation as reading

A translator reads more closely than many ordinary readers.

Every modifier.

Every connective.

Every pronoun.

Every hedge.

Every repeated term.

Translation makes reading accountable.

Human translation as research

Names and terms lead outward into the world.

A translator becomes temporary expert enough to represent concepts responsibly.

Research is not optional decoration.

Human translation as project management

Long projects require:

tracking,

versioning,

queries,

deadlines,

resource control,

and handoffs.

Professional translation has operational structure.

Human translation as collaboration

Quality often emerges from:

translator,

reviser,

reviewer,

subject expert,

editor,

and engineer.

The final target is a team product.

Human translation as institutional memory

Decisions can feed:

termbases,

TMs,

style guides,

and training.

One solved problem should make the next project easier.

The human quality equation

Conceptually:

human translation quality =
source understanding
× target writing
× domain competence
× research
× terminology
× revision
× QA
× context
× accountability.

Not mathematics.

A reminder.

If one factor is near zero, overall quality can collapse.

The minimum viable human workflow

For low-risk content:

brief,

competent translator,

self-revision,

QA,

delivery.

Even the minimum includes review by the translator.

The standard professional workflow

Brief.

Resources.

Translation.

Self-revision.

Independent bilingual revision or appropriate review.

QA.

Delivery.

The high-risk professional workflow

Detailed brief.

Specialist translator.

Controlled terminology.

Independent bilingual reviser.

Subject expert.

Formal QA.

In-context review.

Named approval.

Versioned release.

Human translation governance

Define:

qualification,

assignment,

review levels,

term authority,

query ownership,

correction process,

and incident handling.

Quality should not depend on unwritten custom.

Quality manual

A translation organisation can document:

workflow,

roles,

error taxonomy,

review criteria,

data security,

tool policy,

and escalation.

The manual turns experience into repeatable operations.

Translation standard operating procedure

An SOP can be short.

Before:

check brief and source.

During:

use resources and log queries.

After:

self-revise, QA and handoff.

Review:

follow risk level.

Release:

verify final package.

Simple rules consistently applied outperform elaborate rules ignored.

The human translation audit

A mature translation workflow should be auditable.

An audit does not mean distrusting every translator.

It means checking whether the process that claims to produce reliable translation is actually doing so.

The audit can examine:

brief quality,

source version control,

translator assignment,

terminology use,

query handling,

revision,

review,

QA,

release,

and correction.

Audit question 1: was the right translator assigned?

Check:

language pair,

target locale,

domain,

genre,

experience,

and conflict constraints.

A process can fail before translation begins.

Audit question 2: was the brief sufficient?

Did the translator know:

audience,

purpose,

style,

terminology,

and risk?

If not, later inconsistency may be predictable rather than individual failure.

Audit question 3: was the source stable?

Did changes occur after translation began?

Were changes communicated?

Did version-diff identify affected text?

Audit question 4: were authoritative resources available?

Termbase?

Style guide?

Previous approved work?

Official names?

If resources existed but were inaccessible, do not blame translators for inconsistent reinvention.

Audit question 5: were queries encouraged and resolved?

A culture that punishes questions encourages guessing.

Check whether translators had a real path to clarification.

Audit question 6: did translators self-revise?

Is self-revision a defined stage or a slogan?

Look for:

workflow status,

time allocation,

or documented checks.

Audit question 7: was independent revision appropriate to risk?

Low-risk internal text may not need full bilingual revision.

High-risk legal content probably does.

Check whether review depth matched consequence.

Audit question 8: was the reviser qualified?

Revision requires source-language understanding, target-language judgment and often domain knowledge.

A random extra reader is not automatically a reviser.

Audit question 9: was subject expertise used where needed?

Technical concepts may require specialist validation.

Check whether expert feedback was available for high-risk domain questions.

Audit question 10: did formal QA run?

Mechanical errors should not consume expert attention if tools can catch them.

Numbers.

Tags.

Placeholders.

Missing segments.

Forbidden terms.

Audit question 11: was final context checked?

Was the translation previewed in:

website,

app,

PDF,

subtitle,

or form?

Or did review end in a spreadsheet detached from use?

Audit question 12: were reviewer corrections propagated?

A final-file correction that never reaches TM or terminology will return.

Audit feedback loop.

Audit question 13: were incidents recorded?

If a public mistranslation occurred, did the organisation:

fix it,

find cause,

and update process?

Hidden incidents teach nothing.

Audit question 14: are metrics creating bad incentives?

If translators are measured only on speed, quality may fall.

If reviewers are measured only on number of edits, overcorrection may rise.

Metrics shape behaviour.

Audit question 15: does the workflow preserve accountability?

Can the organisation determine:

who translated,

who revised,

which resources were used,

what version was released,

and who approved?

Traceability matters in high-risk work.

A 50-point human-translation release audit

1. Correct source file?
2. Correct source version?
3. Correct source language?
4. Correct target language?
5. Correct target locale?
6. Audience defined?
7. Purpose defined?
8. Domain defined?
9. Risk level defined?
10. Translator appropriately qualified?
11. Brief complete?
12. Termbase attached?
13. Style guide attached?
14. Reference translations labelled by authority?
15. Names list available?
16. Source defects logged?
17. Ambiguities queried?
18. Translator completed full draft?
19. Translator completed self-revision?
20. Main propositions preserved?
21. Actors preserved?
22. Conditions preserved?
23. Negation preserved?
24. Modality preserved?
25. Quantities preserved?
26. Dates verified?
27. Units verified?
28. Currencies verified?
29. Names verified?
30. Terminology consistent?
31. Defined terms consistent?
32. Quotations handled correctly?
33. Attribution preserved?
34. Uncertainty preserved?
35. Tone appropriate?
36. Register appropriate?
37. Target grammar natural?
38. Collocations natural?
39. Locale conventions correct?
40. Required bilingual revision complete?
41. Required monolingual review complete?
42. Required subject-matter review complete?
43. QA passed?
44. Tags and placeholders intact?
45. Tables and lists complete?
46. Links and references intact?
47. In-context review complete?
48. All queries closed?
49. Approval recorded?
50. Correction path defined?

If critical answers remain unresolved, the translation is not ready to release.

The human translation incident-response workflow

When an error reaches production, first identify impact.

Is the issue:

critical,

major,

or minor?

Then contain.

Remove or replace harmful content where possible.

Correct.

Verify correction independently.

Trace propagation.

Did the error enter:

TM,

termbase,

template,

or other language versions?

Then analyse.

What allowed it through?

Finally improve the system.

Incident response is not about assigning blame quickly.

It is about reducing recurrence.

Incident example: wrong number

Published medical instruction says 5 mL instead of 0.5 mL.

Immediate actions:

remove unsafe target,

issue corrected version,

notify responsible stakeholders,

inspect other dosage strings,

inspect TM,

inspect source extraction,

inspect review record.

Root cause may be:

typing,

OCR,

or review miss.

Fix follows cause.

Incident example: wrong legal condition

Target translates “unless” incorrectly.

Eligibility or obligation changes.

Correct current publication.

Search all related templates.

Update review checklist for logical operators.

Train relevant team.

Incident example: wrong name

Person’s name transliterated incorrectly.

Correct public record where possible.

Update names list.

Check all related documents.

Identity errors deserve respectful remediation.

Incident example: wrong term repeated

One mistranslated technical term appears across ten documents.

This suggests a shared resource problem.

Fix termbase.

Fix TM.

Re-run QA.

Do not repair ten files manually and leave root cause.

Incident example: stylistic complaint

Reader dislikes one phrase but meaning is correct.

Classify carefully.

Not every complaint indicates translation error.

It may reveal:

audience expectation,

brand style,

or regional preference.

Update policy only if evidence justifies it.

The correction doctrine

A correction is complete when:

live content is fixed,

authoritative resources are fixed,

future recurrence is reduced,

and the reason is documented where useful.

The visible typo is only one layer.

Human translation and versioned releases

Translation should be tied to release versions.

For each release, know:

source version,

target version,

terminology version,

and review status.

This matters for:

products,

regulations,

manuals,

and policy documents.

Change requests after release

Clients often change source after translation.

Do not edit target ad hoc without source relationship.

Update source first when possible.

Then translate change.

This preserves alignment.

Emergency patch

Sometimes a target needs immediate correction before source is updated.

Record the divergence.

Reconcile later.

Temporary exceptions should not become permanent hidden states.

Human translation and continuous localisation

In continuous product environments, humans may translate small updates every day.

Architecture needs:

stable context,

TM,

terminology,

fast review,

and change tracking.

Human translation can be continuous without becoming careless.

Pull-request style linguistic review

Software teams can review translations alongside code changes.

Reviewers see:

changed strings,

context,

screenshots,

and version history.

This can improve accountability.

Translation branching

Different product branches may carry different wording.

Do not accidentally apply new-branch terms to old supported versions.

Version context matters.

Release freeze

Before major publication, freeze source where possible.

Late source edits create translation risk.

If edits are unavoidable, flag them visibly.

Human translation and procurement

When buying translation services, evaluate:

process,

qualifications,

review model,

security,

terminology management,

and quality evidence.

Price per word alone does not reveal total quality.

Request for proposal

A translation RFP can specify:

language pairs,

domains,

volumes,

review levels,

tool requirements,

security,

and quality expectations.

Ask suppliers how they handle errors.

Service-level agreement

An SLA may define:

turnaround,

availability,

issue response,

and delivery format.

Quality should also have defined expectations.

Pilot project

Before large outsourcing, run a representative pilot.

Evaluate:

translation,

queries,

review responsiveness,

terminology,

and technical delivery.

The relationship matters as much as the sample paragraph.

Vendor onboarding

Provide:

brief template,

style guide,

termbase,

TM,

security rules,

query process,

and escalation.

Do not expect quality from undocumented institutional knowledge.

Vendor scorecard

Possible dimensions:

accuracy,

term adherence,

target quality,

delivery,

query quality,

technical compliance,

and correction responsiveness.

Avoid one opaque score.

Translation price models

Per word.

Per hour.

Per project.

Retainer.

Post-edit rate.

Each can fit different work.

Quality policy should be independent of billing model.

Discount for repetition

TM repetition can reduce effort.

But exact matches may still require review.

Pricing discounts should not silently remove quality checks.

Rush surcharge

Rush work may require:

additional staff,

overtime,

or reduced scheduling flexibility.

Do not pretend urgency has no cost.

Subject-matter expert cost

SME review adds value in high-risk content.

Budget it deliberately.

Do not expect specialist review to appear for free at the end.

Revision cost

Independent revision requires time.

If a specification requires it, budget it.

Underfunded review becomes performative.

Human translation and certification

Some contexts require certified, sworn or officially recognised translators depending on jurisdiction.

Requirements vary.

Do not assume ordinary professional translation satisfies legal certification needs.

Check applicable rules.

Human translation and notarisation

Notarisation, certification and legalisation are distinct administrative processes.

Translation itself does not automatically satisfy them.

This architecture focuses on translation quality, not jurisdiction-specific legal formalities.

Human translation and interpreting

Translation handles written or recorded text products.

Interpreting handles spoken or signed communication in real time or near-real time.

Skills overlap.

Processes differ.

The Master Art of Translation architecture should give interpreting its own later node rather than folding it into written translation workflow.

Human translation and transcreation

Transcreation gives greater freedom to recreate effect.

Marketing.

Slogans.

Campaigns.

The human workflow still needs:

brief,

claims control,

brand review,

and target-market judgment.

Creative freedom remains accountable.

Human translation and localisation

Localisation extends beyond language into:

format,

UI,

culture,

function,

and product conventions.

Human translation is a core layer.

Localisation is a larger product process.

Human translation and terminology

Terminology system feeds human translation.

Human translation feeds terminology system.

New concepts encountered during work should become structured knowledge.

This loop keeps institutional language current.

Human translation and translation memory

Approved human target segments can enter TM.

But not every draft belongs in master memory.

Revision status and provenance matter.

Human does not automatically mean approved.

Human translation and machine translation

The distinction is workflow ownership.

Human translation:

human builds or owns the target reasoning.

Machine translation:

system generates a target candidate.

Modern projects can mix both.

The architecture should label provenance clearly.

Human translation and LLM tools

An LLM can support a human translator.

The translator should protect:

confidentiality,

source evidence,

terminology,

and final responsibility.

Human-controlled AI is still human translation workflow when the professional owns the decisions.

The “native speaker” myth revisited

Target-language depth matters.

But quality also requires:

translation strategy,

research,

domain knowledge,

and source comprehension.

A monolingual target editor cannot replace the translator.

A bilingual speaker cannot automatically replace a professional translator.

Roles matter.

The “one translator is enough” myth

Sometimes one translator is enough.

Low-risk.

Small text.

Strong expertise.

But independence catches errors.

The right question is not whether every project needs two people.

It is whether the risk justifies a second pair of eyes.

The “more reviewers means more quality” myth

Too many reviewers can create:

conflicting style,

late changes,

and responsibility diffusion.

Quality comes from correct roles, not maximum headcount.

The “literal means accurate” myth

Literal structure can distort meaning when languages differ.

Accuracy means accountability to source meaning and function.

Form can change.

Evidence cannot.

The “free translation means natural” myth

Free paraphrase can add or remove meaning.

Naturalness must remain constrained.

The “translation is subjective” myth

Some choices are subjective.

Many are not.

Wrong number.

Wrong name.

Wrong actor.

Wrong obligation.

Wrong condition.

Wrong term.

Professional translation contains both objective constraints and legitimate stylistic variation.

The “review is proofreading” myth

Proofreading catches surface errors.

Revision verifies translation.

They are not interchangeable.

The “reviewer is always right” myth

Reviewers also make mistakes.

Evidence and governance matter.

A translator can challenge a change professionally.

The “client always knows terminology” myth

Clients know their business.

They may still use inconsistent language.

Terminology work can reveal hidden inconsistency.

Ask and document.

The “dictionary decides” myth

Dictionaries provide candidates and definitions.

Context decides sense.

Target usage decides natural expression.

Domain authority may decide term.

The “web frequency decides” myth

Frequent online use can reflect poor translation.

Prefer authoritative and native-domain sources.

The “AI decides” myth

AI can generate and compare.

It does not own source evidence.

Human professional remains responsible.

The professional translator’s daily checklist

Before work:

source version?

brief?

resources?

queries?

During work:

meaning?

terms?

names?

numbers?

context?

After work:

self-revision?

QA?

unresolved issues?

correct delivery?

This simple routine prevents common failures.

The reviser’s daily checklist

Before review:

brief understood?

source available?

resources current?

During review:

meaning?

completeness?

terms?

hard details?

target quality?

After review:

comments clear?

systemic issues reported?

resources updated?

The reviewer’s daily checklist

Reader fit?

Clarity?

Tone?

Genre?

Consistency?

No unexplained translationese?

No style drift?

The subject expert’s daily checklist

Concept correct?

Procedure correct?

Term correct?

Fact boundaries correct?

No expert rewrite that damages target language?

The project manager’s daily checklist

correct people?

correct files?

correct version?

queries moving?

deadline realistic?

review scheduled?

risk escalated?

The language lead’s daily checklist

terminology stable?

reviewers aligned?

repeated errors?

new decisions documented?

brand voice consistent?

The human translation learning spiral

Project 1 teaches:

one term.

Project 2 reuses it.

Project 3 reveals an exception.

Termbase improves.

Project 4 benefits.

Revision discovers a recurring source ambiguity.

Writers improve.

Human translation becomes organisational learning.

Translation quality and humility

The best translators are not people who never hesitate.

They know where certainty ends.

They verify.

They query.

They revise.

They accept correction.

They defend evidence.

Humility is a quality mechanism.

Translation quality and confidence

Professional confidence is not speed of answer.

It is the ability to commit after sufficient evidence.

The translator can say:

this is the correct term because official target-language legislation uses it.

this ambiguity cannot be resolved from source.

this product name should remain unchanged.

Confidence should have reasons.

Translation quality and empathy

The target reader may not share:

source culture,

specialist knowledge,

or institutional assumptions.

Good translation anticipates reader needs without inventing source facts.

Empathy shapes clarity.

Translation quality and responsibility

A translated warning can affect safety.

A translated form can affect access.

A translated contract can affect obligation.

A translated medical note can affect care.

Language work has consequences.

Professional systems respect that.

Translation quality and time

More time does not automatically create quality.

Structured time does.

Research time.

Draft time.

Revision time.

Review time.

QA time.

Late chaotic editing can reduce quality even on long schedules.

The one-hour translation architecture

For a small 500-word professional text, one possible hour:

5 minutes:

brief and scan.

30 minutes:

translation.

10 minutes:

self-revision.

5 minutes:

hard-detail check.

5 minutes:

target-language read.

5 minutes:

final QA and delivery.

Not every text fits.

The point is explicit quality allocation.

The one-day architecture

Morning:

source analysis and term research.

Midday:

draft.

Afternoon:

self-revision.

Later:

independent review.

End:

QA and delivery.

Complex work may extend across days.

The long-project architecture

Week 1:

brief, terminology, pilot.

Week 2 onward:

translation batches.

Continuous:

queries and term updates.

Midpoint:

consistency review.

Final:

revision, QA, in-context check.

Long projects need governance more than heroic memory.

Pilot chapter

For books, manuals and large sites, translate one representative section first.

Confirm:

style,

terms,

tone,

and workflow.

Then scale.

This prevents hundreds of pages of wrong assumptions.

Freeze decisions early, not permanently

Early consistency helps.

But if better evidence emerges, update decisions.

Version the change.

Propagate deliberately.

Consistency should not lock errors.

The human translation maturity model

Level 1:

ad hoc bilingual writing.

Level 2:

professional translator with self-review.

Level 3:

shared terminology and independent revision.

Level 4:

risk-based review, QA and TM.

Level 5:

integrated source feedback, specialist review, metrics and incident response.

Level 6:

continuous multilingual governance and organisational learning.

Maturity is repeatability.

Frequently asked question: what is human translation?

Human translation is translation in which qualified people perform or own the linguistic and interpretive decisions required to create the target text.

Tools may still assist.

The defining feature is accountable human judgment.

Frequently asked question: is human translation always better than machine translation?

No universal answer.

Human translation is especially valuable for:

high-risk,

ambiguous,

creative,

context-rich,

or specialised content.

Machine translation can be excellent for many repetitive or low-risk tasks.

Architecture should choose by purpose.

Frequently asked question: what makes a professional translator?

Professional competence combines:

language,

translation skill,

research,

domain knowledge,

target writing,

tools,

and professional process.

Bilingualism alone is insufficient.

Frequently asked question: what is translation revision?

Revision is a bilingual quality-control process comparing the target translation with the source to verify accuracy and completeness and improve the target where necessary.

Frequently asked question: what is translation review?

Review can refer broadly to checking translation, but the European Commission’s current quality framework distinguishes monolingual review from bilingual revision: monolingual review focuses on target clarity, tone and suitability for readers.

Frequently asked question: what is proofreading?

Proofreading is a late-stage check for surface and production errors such as spelling, punctuation, typography and formatting.

It does not replace bilingual revision.

Frequently asked question: why use a second translator?

Independent eyes catch errors the original translator may overlook through familiarity and cognitive anchoring.

The need depends on risk.

Frequently asked question: does every translation require a reviser?

No universal rule.

Low-risk content may use translator self-review and spot checks.

High-risk or specification-driven content may require independent revision.

Frequently asked question: what is a subject-matter expert review?

A domain specialist checks technical concepts, terminology or procedures.

The expert complements rather than automatically replaces linguistic review.

Frequently asked question: can a translator use AI and still call it human translation?

Yes, if the human owns the translation decisions and final accountability, subject to project policy.

AI can function as a tool just as dictionaries, corpora and CAT tools do.

Provenance should be clear where required.

Frequently asked question: what if the source is wrong?

Flag the problem.

Do not silently invent corrected meaning when the change matters.

Query the source owner or record assumptions according to project policy.

Frequently asked question: what if two translations are both correct?

Acceptable variation exists.

Choose based on:

style,

context,

genre,

and consistency.

Do not label preference as error.

Frequently asked question: how do I know which term is right?

Use:

context,

definition,

domain,

authoritative target sources,

termbase,

and real usage.

Do not choose only by bilingual dictionary order.

Frequently asked question: should translation sound literal?

It should sound faithful.

That may require non-literal structure.

Word order and grammar can change while meaning remains stable.

Frequently asked question: how much should a translator research?

Enough to support decisions at the project’s risk level.

Critical terms require stronger authority than ordinary low-impact wording.

Frequently asked question: why are translation queries important?

Queries expose missing evidence.

They prevent confident guessing.

A good query is a quality-control action.

Frequently asked question: why do translators need style guides?

Style guides externalise recurring decisions so the target remains consistent across translators, documents and time.

Frequently asked question: why do translators use translation memories?

TM reduces repeated translation and preserves approved language.

Human translators still verify context.

Frequently asked question: why do translators use termbases?

Termbases control concepts and preferred terminology.

This is especially important in technical and long-running projects.

Frequently asked question: what is the biggest human-translation mistake?

Starting target writing before understanding the source and project context.

Most later quality mechanisms become more expensive if the initial interpretation is wrong.

Frequently asked question: what is the biggest review mistake?

Treating stylistic preference as correction while missing source fidelity.

Review needs priorities.

Frequently asked question: can proofreading catch mistranslation?

Sometimes by accident.

But proofreading alone is not designed to compare source meaning.

Bilingual revision is the relevant process.

Frequently asked question: is translation subjective?

It contains judgment.

But many requirements are objective.

Names, numbers, conditions, obligations, identities and defined terms can be checked.

Professional translation separates constrained decisions from genuine variation.

Frequently asked question: why can two professional translators disagree?

Languages do not map one-to-one.

More than one target can preserve the same function.

Evidence, brief and target norms determine which options are acceptable.

Frequently asked question: should a translator translate into their strongest language?

Target-language writing competence is crucial.

Directionality policy varies by person, language pair and professional setting.

Evaluate actual competence.

Frequently asked question: how do I choose a translator?

Look for:

relevant language pair,

target writing quality,

domain experience,

professional process,

references or testing,

security fit,

and ability to communicate about uncertainty.

Frequently asked question: how long does professional translation take?

It depends on:

difficulty,

domain,

format,

research,

repetition,

language pair,

and review level.

Words per day are only a rough planning measure.

Frequently asked question: why is human translation expensive?

You are paying for:

interpretation,

research,

writing,

terminology,

revision,

QA,

and responsibility.

The target is a professional communication product.

Frequently asked question: what should a good translation service deliver?

Not merely target text.

Ideally:

correct file,

correct locale,

resolved queries,

controlled terminology,

review status,

and a target ready for its intended use.

Frequently asked question: what is translation QA?

QA is the set of human and automated checks used to detect meaning, terminology, formatting, consistency and technical problems before release.

Frequently asked question: can QA prove a translation is perfect?

No.

Quality assurance reduces risk.

It does not eliminate all human fallibility.

Frequently asked question: how do you improve translation quality over time?

Capture corrections.

Update termbases.

Update TMs.

Calibrate reviewers.

Improve source writing.

Train translators.

Review incident patterns.

Quality improves through feedback.

The standards context in 2026

ISO 17100:2015 remains the published international standard for translation services and was confirmed in 2020.

ISO currently lists a second edition, ISO/AWI 17100, as under development and intended to replace the 2015 edition.

The current published standard’s public abstract describes requirements for core translation processes, resources and other aspects needed to deliver services meeting applicable specifications.

That systems framing aligns with the architecture developed here.

European Commission quality architecture

The European Commission’s current translation-quality framework describes quality as accurate, clear and fit for purpose.

It links quality-control approach to:

document type,

risk profile,

intended use,

and reader expectations.

It distinguishes bilingual revision from monolingual review and assigns roles across translators, quality officers and workflow managers.

This provides a useful large-scale example of risk-calibrated human translation quality.

Why this matters for everyday translation

You do not need a government translation service to use the principle.

For a school email:

simpler workflow.

For a medical instruction:

stronger workflow.

For a literary poem:

different review criteria.

The same architecture scales down.

The governing human translation principle

The translator should never be asked merely:

“Can you translate these words?”

The professional question is:

What must this target text allow this reader to understand, feel or do, and what evidence must remain intact for that to happen responsibly?

That question determines the workflow.

Where this node leads next

The Human Translation System sits alongside the Machine Translation System.

Together they create a major architectural fork.

From here the tree can continue into:

post-editing,

translation revision,

translation quality assurance,

localisation,

transcreation,

interpreting,

literary translation,

technical translation,

legal translation,

medical translation,

academic translation,

business translation,

software localisation,

SEO translation,

and subtitle translation.

Each later branch can inherit the same role logic:

brief,

qualified human,

controlled resources,

risk-appropriate review,

QA,

and feedback.

Vocabulary route

If a human translator’s problem is lexical—word sense, collocation, register, false friends, semantic range—the Vocabulary Learning Hub owns the deeper learning architecture.

The Human Translation System uses vocabulary knowledge in production.

How English Works route

If the problem involves English grammar, reference, clause structure, tense, modality, information structure or discourse, How English Works owns the language mechanism.

This node owns the human workflow that applies it.

Translation Memory route

When the translator needs prior approved bilingual evidence, use the Translation Memory System.

It controls reuse and provenance.

Terminology route

When the issue is concept identity and preferred terms, use the Terminology System.

It controls concept-to-term policy.

Machine Translation route

When non-human generation is part of the workflow, use the Machine Translation System.

It owns NMT, LLM translation, post-editing architecture and automation controls.

Final synthesis

Human translation is not defined by the absence of technology.

It is defined by accountable human judgment.

The translator understands.

The translator researches.

The translator chooses.

The translator writes.

The translator revises.

Another human may challenge the decisions.

A reviewer may test the target as communication.

A subject expert may test the concept.

QA may test hard details.

The final product is released because the required evidence has survived each gate.

Humans can make mistakes.

That is why professional human translation is a system rather than a promise.

The strongest system makes uncertainty visible.

It assigns the right expertise.

It separates drafting from independent checking.

It uses terminology and memory without being ruled by them.

It uses technology without surrendering responsibility.

It matches review effort to risk.

It learns from corrections.

It updates institutional language.

It protects the target reader.

Translation becomes accountable when every important choice can answer one question:

Why is this target justified by this source, for this reader, in this context?

That is the craft.

That is the process.

That is the human translation system.

---

Continue through the eduKateSG translation architecture

Master Translation root: Master Art of Translation | The Complete System for Moving Meaning Between Languages

Source Analysis: How to Read a Source Text Before You Translate It

Equivalence: How Equivalence Works When Languages Do Not Match One-to-One

Context Stack: The Context Stack

Translation Unit: The Translation Unit

Terminology System: The Terminology System

Translation Memory System: The Translation Memory System

Machine Translation System: The Machine Translation System

Vocabulary: Vocabulary Learning Hub

English: How English Works

Professional references

ISO 17100:2015 — Translation services — Requirements for translation services

ISO/AWI 17100 — Translation services — Requirements for translation services, Edition 2 under development

European Commission — Translation quality

European Commission Knowledge Centre — Revision and post-editing

Advanced revision laboratory: twenty high-risk cases

The following revision cases are designed to show how an independent human reviser thinks. The aim is not merely to find “bad wording.” The reviser identifies what kind of evidence the target must preserve and what consequence follows if it does not.

Case 1: prohibition

Source:

“Do not operate the device while the cover is removed.”

Target:

“Operate the device while the cover is removed.”

Revision priority:

critical.

Reason:

prohibition removed.

Action:

correct immediately, search related warnings, verify no similar polarity errors.

Case 2: permission versus obligation

Source:

“Customers may request a refund.”

Target means:

“Customers must request a refund.”

Revision priority:

major to critical depending on policy.

Reason:

right becomes obligation.

Case 3: possibility versus certainty

Source:

“The change may affect performance.”

Target:

“The change affects performance.”

Reason:

certainty strengthened.

The reviser restores epistemic force.

Case 4: quantity boundary

Source:

“At least 12 months.”

Target:

“12 months.”

Reason:

minimum threshold becomes exact value.

This can affect eligibility.

Case 5: maximum boundary

Source:

“No more than five attempts.”

Target:

“Five attempts.”

Reason:

maximum becomes exact requirement.

Case 6: range

Source:

“between 10 and 15 minutes.”

Target:

“10 to 50 minutes.”

One digit error.

Large operational effect.

Hard-detail comparison catches it.

Case 7: percentage versus percentage point

Source:

“increased by five percentage points.”

Target:

“increased by five percent.”

Not equivalent.

Finance and statistics require distinction.

Case 8: source attribution

Source:

“The minister said the programme could begin in June.”

Target:

“The programme will begin in June.”

Attribution and uncertainty both disappear.

Case 9: reported versus verified

Source:

“The company reported a 10% reduction.”

Target:

“The company achieved a 10% reduction.”

Reported claim becomes independent fact.

Case 10: legal defined term

Source defines “Provider.”

Target later uses three target synonyms.

Reviser standardises the defined term.

Case 11: medical frequency

Source:

“one tablet twice daily.”

Target:

“two tablets daily.”

Same daily total.

Different dosing pattern.

Critical.

Case 12: technical sequence

Source:

“Close the valve before disconnecting the hose.”

Target:

“Disconnect the hose before closing the valve.”

Sequence reversed.

Operational risk.

Case 13: exception

Source:

“All visitors except emergency personnel must register.”

Target omits exception.

Policy scope changes.

Case 14: only

Source:

“Only authorised staff may enter.”

Target:

“Authorised staff may enter.”

The target states permission but loses exclusivity.

Other people are no longer explicitly excluded.

Case 15: not all

Source:

“Not all participants completed the survey.”

Target:

“No participants completed the survey.”

Partial negation becomes total negation.

Case 16: pronoun

Source:

“Rina told Maya that she had been selected.”

Target names Rina as selected.

Later context shows Maya.

Reference error.

Case 17: app label

Source:

“Archive”

Context:

verb command.

Target:

noun meaning archive collection.

Function changes.

Case 18: brand name

Source contains official product name.

Target translates product name literally.

Identity changes.

Case 19: historical uncertainty

Source:

“probably written around 1820.”

Target:

“written in 1820.”

Approximation and uncertainty vanish.

Case 20: academic limitation

Source:

“These findings may not generalise beyond this sample.”

Target:

“These findings do not apply to other populations.”

Cautious limitation becomes absolute claim.

Advanced review laboratory: target-only problems

Not every issue requires source comparison.

A monolingual reviewer may find target-language defects invisible to a bilingual accuracy pass.

Target problem 1: translationese

Sentence follows source word order so closely that a native reader must mentally reconstruct the original.

Review:

rewrite syntax while preserving meaning.

Target problem 2: wrong collocation

Each target word is valid.

Combination is unnatural.

Review:

replace with conventional target phrase.

Target problem 3: register drift

Page begins formal and ends conversational.

Review:

stabilise voice according to brief.

Target problem 4: inconsistent address

Formal pronoun in one paragraph, informal in next.

Review:

apply project policy.

Target problem 5: mixed spelling variety

British and American spellings alternate.

Review:

apply locale.

Target problem 6: excessive nominalisation

Target becomes heavy and bureaucratic because source noun structures were copied.

Review:

use natural target verbs where meaning permits.

Target problem 7: sentence overload

Source allows long nested syntax.

Target readers expect shorter units.

Review:

split while preserving logic.

Target problem 8: choppy target

Translator mirrors short source segments.

Target needs smoother cohesion.

Review:

join or connect where appropriate.

Target problem 9: repeated pronoun ambiguity

Target language requires more explicit nouns than source.

Review:

clarify referents without inventing information.

Target problem 10: culturally odd politeness

Literal polite formula sounds cold or exaggerated.

Review:

choose conventional target form with equivalent social function.

High-risk checklist: legal

Before legal release, verify:

parties,

defined terms,

rights,

obligations,

permissions,

conditions,

exceptions,

deadlines,

jurisdiction,

cross-references,

numbers,

and signatures.

Look especially for:

shall,

may,

must,

unless,

provided that,

subject to,

notwithstanding,

and only if.

The exact target depends on legal system and drafting convention.

High-risk checklist: medical

Verify:

drug name,

dose,

unit,

frequency,

route,

duration,

age group,

contraindication,

warning,

uncertainty,

body site,

and device setting.

Never infer missing clinical information.

High-risk checklist: finance

Verify:

currency,

percentage,

percentage point,

rate,

period,

accounting term,

sign,

decimal,

table total,

and claim attribution.

Cross-check tables with narrative.

High-risk checklist: safety

Verify:

hazard,

prohibition,

required action,

sequence,

condition,

exception,

protective equipment,

distance,

temperature,

pressure,

and emergency response.

A safety translation should be operationally testable.

High-risk checklist: public eligibility

Verify:

who qualifies,

who is excluded,

minimum and maximum thresholds,

residency,

age,

income,

documents,

deadline,

appeal rights,

and contact path.

Translation must not change access.

High-risk checklist: security

Verify:

threat name,

severity,

affected version,

condition,

indicator,

remediation,

credential instruction,

and uncertainty.

Do not overstate attribution.

Specialist handoff: translator to legal expert

Translator sends:

source clause,

target clause,

defined terms,

question,

and reason for uncertainty.

Legal expert responds on concept.

Translator integrates response into natural target-language legal drafting.

Specialist handoff: translator to doctor

Translator asks:

Does “administration” refer to giving medication or administrative management?

Provide surrounding clinical sentence.

Doctor clarifies concept.

Translator chooses target wording.

Specialist handoff: translator to engineer

Translator flags ambiguous component term.

Engineer identifies actual part from drawing.

Translator applies approved term consistently.

Specialist handoff: translator to historian

Translator finds old institutional title.

Historian explains period-specific role.

Translator chooses historically accurate target with note if policy permits.

Specialist handoff: translator to product designer

UI string “Back” is ambiguous.

Designer explains screen action.

Translator chooses correct command.

This is human translation as cross-functional work.

The translator’s final bilingual pass

Read source sentence.

Read target sentence.

Ask:

Did I preserve all information?

Did I add anything?

Did I change certainty?

Did I change action?

Did I change who did it?

Did I change quantity?

Did I change time?

Did I change relationship?

Then continue.

The translator’s final target-only pass

Hide source.

Read as a target reader.

Ask:

Would this be written naturally?

Is it coherent?

Does it sound like one person wrote it?

Is terminology understandable?

Does it fit genre?

This pass catches translationese.

The reviser’s final pass

Focus on residual high-risk issues.

Numbers.

Names.

Terms.

Conditions.

Negation.

Modal force.

Then read critical passages one more time in context.

The proofreader’s final pass

Spelling.

Punctuation.

Widows/orphans where relevant.

Headings.

Page numbers.

Cross-references.

Table alignment.

Typography.

No late rewrite unless necessary.

Late stylistic changes can introduce errors.

Freeze after approval

Once approved, changes should re-enter review.

Do not allow an unreviewed last-minute edit to bypass the process.

The final five words can still contain a critical error.

Change-control label

Every post-approval change should be classified:

source change,

translation correction,

terminology update,

format fix,

or style edit.

Then assign required recheck.

Translation sign-off

For high-risk work, sign-off can record:

translator,

reviser,

subject expert,

QA,

version,

date,

and unresolved limitations.

The purpose is traceability.

Professional humility at sign-off

Sign-off does not mean perfection.

It means required process completed and known issues resolved or disclosed.

The 100-question human translation master checklist

1. Do I have the correct source?
2. Is it the latest source?
3. Is source language confirmed?
4. Is target language confirmed?
5. Is target locale confirmed?
6. Is audience known?
7. Is purpose known?
8. Is domain known?
9. Is genre known?
10. Is publication channel known?
11. Is risk known?
12. Is deadline realistic?
13. Is translator qualified?
14. Is reviewer assigned if required?
15. Is subject expert assigned if required?
16. Is termbase current?
17. Is style guide current?
18. Is TM appropriate?
19. Are reference materials authoritative?
20. Are names documented?
21. Are source defects logged?
22. Are ambiguous dates identified?
23. Are unclear pronouns identified?
24. Are technical terms identified?
25. Are defined terms identified?
26. Are abbreviations identified?
27. Are do-not-translate items identified?
28. Are placeholders identified?
29. Are images needed for context?
30. Are tables understood?
31. Are footnotes included?
32. Are links included?
33. Is source structure mapped?
34. Is central claim understood?
35. Are conditions mapped?
36. Is negation mapped?
37. Is modality mapped?
38. Is event order mapped?
39. Are quantities understood?
40. Are references resolved?
41. Are culture-specific items identified?
42. Are idioms identified?
43. Are metaphors identified?
44. Is register identified?
45. Is tone identified?
46. Are queries raised?
47. Are queries answered?
48. Are answers recorded?
49. Is draft complete?
50. Are all sections translated?
51. Are tables translated?
52. Are captions translated?
53. Are labels translated?
54. Are hidden text areas checked?
55. Is terminology consistent?
56. Are names correct?
57. Are numbers correct?
58. Are dates correct?
59. Are units correct?
60. Are currencies correct?
61. Are conditions preserved?
62. Are exceptions preserved?
63. Is negation preserved?
64. Is modal force preserved?
65. Is attribution preserved?
66. Is uncertainty preserved?
67. Are quotations faithful?
68. Is target grammar correct?
69. Are collocations natural?
70. Is word order natural?
71. Is cohesion natural?
72. Is target register correct?
73. Is target tone correct?
74. Is target locale correct?
75. Is self-revision complete?
76. Did translator read target alone?
77. Did translator compare source and target?
78. Was required bilingual revision done?
79. Were reviewer changes justified?
80. Were disagreements resolved?
81. Was subject expert review done?
82. Did formal QA run?
83. Were QA warnings resolved?
84. Are tags intact?
85. Are placeholders intact?
86. Are URLs intact?
87. Are cross-references correct?
88. Is final layout checked?
89. Is accessibility checked?
90. Is target usable?
91. Is correct file being delivered?
92. Is correct filename used?
93. Is correct version marked?
94. Is approval recorded?
95. Are corrections synced to TM?
96. Are new terms added to termbase?
97. Are style decisions documented?
98. Are source defects fed back?
99. Is incident path known?
100. Can the team explain why the translation is ready?

This checklist is intentionally exhaustive.

Real projects can use a shorter version.

The point is coverage.

A compact ten-question version

1. Right source?
2. Right reader?
3. Meaning correct?
4. Terms correct?
5. Names and numbers correct?
6. Target natural?
7. Review level appropriate?
8. QA passed?
9. Final context checked?
10. Approval recorded?

Use the compact version for ordinary work.

Use the full architecture for high-risk work.

The human translation doctrine

A human translator should be free to make linguistic decisions.

They should not be free from evidence.

A reviser should be free to challenge the target.

They should not be free to impose unexplained preferences.

A subject expert should be free to correct concepts.

They should not be free to destroy target-language quality.

A project manager should be free to optimise workflow.

They should not be free to remove necessary review simply to meet an unrealistic deadline.

Every role has authority.

Every role also has boundaries.

Why this node matters after Machine Translation

The previous node explained non-human generation.

This node explains accountable human production.

The two are not competitors.

They are selectable architectures.

Some projects are human-first.

Some machine-first.

Some mixed.

The shared goal is the same:

preserve meaning,

serve readers,

control risk.

Why human translation remains central in the AI era

As automatic generation improves, the most valuable human skills shift.

Less value in retyping routine phrases.

More value in:

source interpretation,

ambiguity management,

domain judgment,

cultural mediation,

quality evaluation,

and accountability.

Technology raises the ceiling.

Human judgment defines whether the result is fit to use.

Final release principle

A translation should not be released because:

the translator finished,

the reviewer finished,

or the deadline arrived.

It should be released because the required evidence and checks for that risk level are complete.

That is the discipline that turns bilingual skill into professional human translation.

Final operating matrix: who should do what?

A human translation system becomes efficient when each task is assigned to the person best placed to do it.

Translator:

understand source,

research,

draft target,

apply terminology,

self-revise,

raise queries.

Reviser:

compare source and target,

verify meaning and completeness,

challenge terms and logic,

correct substantive language defects.

Monolingual reviewer:

read target as target text,

improve clarity, tone, flow and audience fit.

Subject-matter expert:

verify technical concept and domain convention.

Terminologist:

govern concepts, preferred terms and variants.

Localization engineer:

protect files, tags, variables and context.

Project manager:

control scope, version, deadline and handoff.

Language lead:

resolve policy, reviewer disagreement and systemic consistency.

Proofreader:

catch final target and production defects.

Approver:

accept release according to risk and governance.

The same person may hold several roles.

The roles should still remain conceptually separate.

Final operating matrix: what must remain independent?

Some tasks benefit from independence.

Self-revision cannot fully substitute for independent revision.

Subject expertise cannot fully substitute for linguistic review.

Monolingual review cannot verify source fidelity by itself.

Automated QA cannot evaluate all meaning.

Client approval cannot retroactively make a mistranslation accurate.

Independence is designed according to the failure each role can detect.

Final operating matrix: what can be combined?

For low-risk work, one expert translator may:

translate,

self-revise,

proofread,

and deliver.

For medium-risk work, one second linguist may combine bilingual revision and monolingual review.

For high-risk specialist work, keep concept and linguistic checks more separate.

Process should be proportionate.

The quality triangle

Human translation balances three obligations.

Source obligation:

do not distort what the source says.

Target obligation:

write competent target-language communication.

Reader obligation:

make the target usable for its intended audience and purpose.

A strong translation satisfies all three.

Over-focus on source can create translationese.

Over-focus on target can create rewriting.

Over-focus on reader convenience can create unsupported explanation.

The brief helps balance the triangle.

The accountability chain

Every critical decision should be traceable to a layer.

Source text.

Context.

Author clarification.

Termbase.

Official reference.

Style guide.

Subject expert.

Language-policy decision.

When a target choice has no evidential layer, it deserves another look.

The final human translation decision ladder

When facing a difficult phrase:

1. Read the complete source context.
2. Identify the proposition and function.
3. Determine whether the expression is ordinary or technical.
4. Check terminology and names.
5. Research authoritative target usage.
6. Generate one or more target candidates.
7. Compare what each candidate preserves or loses.
8. Select according to brief.
9. Record decision if it will recur.
10. Verify in final context.

This is slow only when the problem is genuinely difficult.

With experience and resources, many steps become rapid.

The professional query ladder

If uncertain:

First, search internal approved resources.

Second, inspect broader document context.

Third, consult authoritative external sources.

Fourth, ask subject expert or author.

Fifth, escalate to client or language lead.

Do not jump straight to guessing.

The professional correction ladder

If an error is found:

Correct current target.

Check repeated occurrences.

Check translation memory.

Check terminology.

Check templates.

Check other languages if source problem is shared.

Check released versions.

Update training or policy if systemic.

One correction can protect future work.

The professional release ladder

Draft complete.

Self-revision complete.

Required revision complete.

Required review complete.

Subject checks complete.

QA complete.

Context check complete.

Approval complete.

Only then release.

If a project intentionally skips one layer, the reason should be risk-based, not accidental.

The professional stop ladder

Stop translation when:

source is corrupted,

critical meaning is unresolved,

data cannot be processed securely,

required expertise is missing,

or deadline makes mandated quality process impossible.

Escalate.

The ability to stop is a quality feature.

Reader-centred verification

Before release, imagine the target reader knows nothing about the source.

Can they understand the target on its own?

Can they distinguish:

instruction from suggestion,

fact from report,

certainty from possibility,

requirement from option?

If not, revise.

Source-centred verification

Now imagine a source author compares both texts.

Can every important target statement be justified?

Is any fact invented?

Any limitation missing?

Any actor changed?

Any condition weakened?

If yes, fix.

Product-centred verification

Now inspect final environment.

Can the target fit?

Do labels align?

Are variables correct?

Do line breaks obscure meaning?

Does the subtitle remain readable?

Does the form still work?

The target is not finished until the product works.

Risk-centred verification

Ask:

What is the worst plausible consequence if this translation is wrong?

Then inspect the target specifically for that failure.

Safety content:

unsafe action.

Legal content:

changed right.

Medical content:

wrong dose.

Finance:

wrong value.

Education:

changed task.

This makes review purposeful.

Human translation and trust

Readers usually cannot see the source.

They trust the translator and publisher to have preserved it.

That asymmetry creates responsibility.

A professional translation system honours invisible trust through visible process internally.

Human translation and transparency

Where appropriate, organisations can state:

human translated,

human reviewed,

machine translated and post-edited,

or automatically translated.

Provenance helps users calibrate trust.

Not every project needs public labels.

Internal provenance should remain clear.

Human translation and institutional voice

Large organisations need translations to sound like one institution even when many translators contribute.

This requires:

style,

terminology,

review,

and examples.

Institutional voice should not erase natural target language.

It should create consistent identity.

Human translation and community voice

For indigenous, minority or community-specific languages, authority may belong partly to the community.

Professional process can include:

community review,

preferred orthography,

cultural governance,

and consent.

Translation quality is not always defined by external institutions.

Human translation and language change

Languages evolve.

Terms change.

Politeness norms change.

Inclusive language changes.

Style changes.

A translation approved ten years ago may no longer be best today.

Human systems need living governance.

Human translation and legacy content

Legacy translations are evidence, not automatic templates.

Classify:

still current,

historically valid,

needs refresh,

or unsafe.

Do not rewrite protected material without policy.

But do not let history silently dictate new work.

Human translation and reuse

Reuse can come from:

TM,

phrase banks,

standard clauses,

templates,

and established terminology.

Humans should reuse decisions when context remains valid.

This reduces cost and improves consistency.

Human translation and originality

A good translator does not seek novelty for its own sake.

If an approved phrase works, reuse it.

Originality matters where the source itself demands creative reconstruction.

Human translation and simplicity

Simple target language can be excellent.

But simplicity must not remove:

conditions,

qualifiers,

or technical distinctions.

Plain language is disciplined clarity.

Not deletion.

Human translation and complexity

Some source ideas are genuinely complex.

Do not flatten them merely to avoid difficult target syntax.

Make relationships clear.

Preserve intellectual content.

Human translation and beauty

Literary, rhetorical and brand texts can require beauty.

Beauty remains bounded by source.

A target can be elegant and unfaithful.

A professional translator pursues effect with accountability.

Human translation and precision

Technical and legal texts may prioritise precision over stylistic variety.

Repetition can be correct.

Controlled terminology can sound less literary and more trustworthy.

Genre determines priorities.

Human translation and speed

Speed is valuable.

The best speed comes from:

preparation,

resources,

experience,

and reuse.

Not skipping meaning checks.

Human translation and technology

Technology should remove:

retyping,

repeated lookup,

mechanical comparison,

and formatting burden.

Humans should spend attention on:

meaning,

judgment,

and writing.

This is the proper division of labour.

The human translation architecture map

Upstream:

source readiness.

Then:

brief.

Then:

source analysis.

Then:

terminology and reference resources.

Then:

draft translation.

Then:

self-revision.

Then:

risk-appropriate independent revision or review.

Then:

subject validation where needed.

Then:

formal QA.

Then:

in-context review.

Then:

approval.

Then:

release.

Then:

feedback and correction.

Then:

memory update.

The workflow is a loop, not a line.

The quality feedback loop

Every error should strengthen one of the upstream layers.

Wrong term?

Improve termbase.

Wrong source interpretation?

Improve source analysis guidance.

Wrong product string?

Improve context extraction.

Repeated reviewer preference?

Improve style guide.

Wrong number?

Improve QA.

This is how a translation organisation compounds learning.

The final ten release questions

1. Is this the right source version?
2. Does the target preserve all important source meaning?
3. Are controlled terms, names and numbers correct?
4. Is the target natural for its readers?
5. Did required independent quality control occur?
6. Are all queries resolved?
7. Did formal QA pass?
8. Does the target work in its final environment?
9. Is approval recorded?
10. Can the team correct it quickly if a defect is found?

If all ten are yes, the target has a strong professional basis.

Closing: accountable language transfer

Human translation is sometimes presented as a craft threatened by automation.

A better view is that automation makes the nature of the craft clearer.

The translator’s unique value is not typing target words.

It is accountable interpretation.

A human can recognise missing evidence.

A human can decide that a source term is technical rather than ordinary.

A human can preserve deliberate ambiguity.

A human can understand why a child’s voice should stay simple.

A human can ask whether a legal modal changes rights.

A human can tell a client that a source date is impossible.

A human can defend a choice with evidence.

A human can accept correction and update future practice.

Those capabilities become more valuable, not less, when machines generate language cheaply.

The strongest future translation workflow therefore does not choose between human skill and technology.

It gives each the correct job.

Technology stores, searches, compares, checks and drafts.

Humans interpret, judge, write, revise, escalate and take responsibility.

The result is not merely translated text.

It is a target-language document whose meaning has survived a chain of accountable decisions.

That is the Master Art of Translation at the human level.

Final governance layer: responsibility does not disappear at handoff

A common translation failure occurs when every participant assumes someone else owns the final decision.

Translator:

“I sent it for review.”

Reviewer:

“I only checked language.”

Subject expert:

“I only checked terminology.”

Project manager:

“I assumed the reviewers approved it.”

Publisher:

“I received a final file.”

Responsibility becomes distributed until it becomes invisible.

A strong workflow names the release owner.

The release owner does not need to redo every linguistic decision.

They confirm that required gates were completed.

The release owner’s checklist

Correct source version.

Correct target locale.

Required translator completed work.

Required revision completed.

Required specialist review completed.

Critical queries closed.

QA passed.

In-context review completed.

Correct final file selected.

Approval evidence present.

This role closes the process.

The translation package

For high-value projects, the final package can contain more than the target.

Possible contents:

final target file,

source reference,

query log,

decision log,

terminology updates,

QA report,

review status,

and version information.

Not every client needs every artifact.

The organisation should retain enough internally to understand the release.

Archiving

Archive:

final source,

final target,

relevant terminology,

approved decisions,

and project status.

Avoid archiving uncontrolled drafts as if they were approved references.

Future translators need to know what became final.

Reuse status

Label translations:

draft,

reviewed,

approved,

deprecated,

historical,

or reference-only.

This status can later control TM and content reuse.

Human origin alone does not define authority.

Retention

Keep content according to:

contract,

privacy,

legal requirement,

and operational value.

Translation archives can contain sensitive text.

Retention should be intentional.

Deletion

When project data must be deleted, consider:

local files,

cloud storage,

TM entries,

backups,

email attachments,

and vendor copies.

A translation can persist in more places than the final document.

Human translation and auditability

For ordinary low-risk writing, formal audit trails may be unnecessary.

For regulated or high-risk content, auditability can answer:

which source was translated,

who translated it,

who reviewed it,

which version was released,

and what correction occurred later.

Auditability supports accountability.

Human translation and trust calibration

Readers should not need to understand the whole workflow.

The organisation does.

Internally, trust should be calibrated by evidence.

A translation with:

known expert translator,

current terminology,

independent revision,

and passed QA

deserves more operational trust than an unknown old file.

Authority decay

Approval is not eternal.

Language changes.

Facts change.

Policies change.

A translation can remain linguistically accurate while becoming operationally outdated.

Review long-lived content periodically where freshness matters.

Freshness triggers

Trigger re-review after:

law change,

product update,

rebrand,

new terminology policy,

medical guidance change,

or major style revision.

This keeps human-approved content current.

Human translation and search visibility

For web content, target-language search intent matters.

The translator should not stuff keywords mechanically.

Use natural phrases target readers actually search.

SEO should support reader discovery without changing factual meaning.

Human translation and metadata

Titles, meta descriptions, alt text and structured labels need separate attention.

They may have:

length limits,

search function,

or accessibility role.

Do not treat them as incidental leftovers.

Human translation and internal linking

A translated educational ecosystem works best when related concepts connect.

But links should respect canonical owners.

A human translator can route:

vocabulary issues to Vocabulary,

English-language mechanisms to How English Works,

translation-memory issues to the TM node,

machine generation to the MT node.

Architecture prevents cannibalisation.

The human translator as router

A skilled translator knows when a problem is not actually a translation problem.

Unknown word?

Vocabulary.

Unclear English clause?

English mechanism.

Unclear source fact?

Author query.

Technical concept?

Subject expert.

Past wording?

Translation memory.

Controlled term?

Terminology system.

Machine-generated draft?

Machine Translation System.

This diagnostic ability saves time.

The final professional stance

Do not maximise literalness.

Do not maximise freedom.

Do not maximise speed.

Do not maximise reviewer intervention.

Optimise for:

faithful meaning,

natural target communication,

appropriate risk control,

and efficient evidence-based process.

That is the balance.

Final note on the evolving standard

As of September 2026, ISO 17100:2015 remains the published standard for translation services, while ISO lists Edition 2 as an approved work item under development.

That means professional practice continues to evolve.

Tools change.

AI changes.

Workflow changes.

The foundational principle remains remarkably stable:

quality comes from competent resources and controlled processes that satisfy the specifications of the translation task.

Final note on revision and review

The European Commission’s current quality framework provides a useful operational distinction.

Bilingual revision:

compare translation against source.

Monolingual review:

focus on clarity, tone and reader suitability.

The two checks answer different questions.

A mature human translation system chooses the check that fits the document’s risk and purpose.

Final note on human value

Humans are not valuable because they are slower than machines.

They are valuable because they can take responsibility for a decision.

A person can say:

I checked the source.

I verified the official name.

I preserved the uncertainty.

I queried the ambiguous condition.

I chose this target because the audience needs this register.

I rejected the old memory because the context changed.

I corrected the machine because it omitted the exception.

I signed off because the required evidence is complete.

That is accountability.

The closing rule

Before publishing any human translation, ask:

Can the person responsible for this target explain why its important choices are justified?

If yes, the translation has an accountable foundation.

If no, the workflow still has work to do.