Did you know that a translation can preserve almost every word and still change whether a student may apply for a programme? Localization is not only the process of making an AI response readable in another language. It must preserve authoritative amounts, school dates, negation, exceptions, recipient identities and current user permission while adapting grammar, typography and formatting. A confident bilingual answer that changes a deadline or omits a condition is not an acceptable outcome.
This practical guide explains how an AI agent can keep meaning stable while presenting messages in different locales. It covers ICU MessageFormat, language tags, plural rules, dates, quantities, right-to-left text, source terminology and independent acceptance tests. An invented 120-message case and changed-context forty-message school-notice exercise make the failure modes concrete. The figures are teaching numbers, not measured eduKate traffic or proof that artificial superintelligence exists.
Explore How Super Intelligence Works alongside Temporal Semantics, Numeric Integrity and Accessible Agent Interfaces. These companion chapters cover particular source and user-experience constraints; this guide treats the translation boundary itself.
Choose a learning route
- Reading school notices in another language? Begin with source dates, official terms and logical conditions.
- Engineering an AI agent? Study ICU messages, typed facts, locale tags and rendering.
- Learning Mathematics? Check the 120-message ledger and changed-context number examples.
- Protecting users? Verify negative conditions, access restrictions and actual official sources.
Contents
- 1. Translation is only one part of localization
- 2. Store dates and amounts before turning them into prose
- 3. ICU MessageFormat keeps the full message translatable
- 4. Plural categories are locale-dependent
- 5. BCP 47 language tags are not identity credentials
- 6. Local date formatting cannot change the scheduled event
- 7. Numbers and percentages need independent arithmetic
- 8. Negation, unless and only if are critical conditions
- 9. Right-to-left messages need correct directional isolation
- 10. Education terminology needs source-aware treatment
- 11. A fictional one-hundred-and-twenty-message laboratory
- Frequently asked questions
- Language and formatting standards
- Related learning
1. Translation is only one part of localization
Translation renders the intended meaning in another language, while localization also chooses suitable number formats, time and calendar presentation, terminology, layout and user-interaction conventions. A parent may correctly understand a translated Mathematics topic yet misread a short numerical date in an unfamiliar order. The agent must distinguish understandable prose from the entire localized task.
Begin with the actual service outcome. Does the user want a lesson explanation, a draft-only notice or a current official deadline? Identify every fact and condition that would change their next action. Keep these as structured source-backed values. A generated paragraph is only one rendering of the user’s legitimate task and cannot become its authoritative record.
The original issuer, effective school year and applicable student level matter more than stylistic fluency. A genuine source about a previous cohort does not become current because it is translated into a preferred language. The assistant should retain the original programme name and cite the responsible authority when discussing a consequential education pathway.
Localization should also improve accessibility. The page should identify the actual language for assistive technology, support appropriate reading order and handle mixed-script fragments. An attractive translation that cannot be navigated or read correctly is incomplete. Test the rendered experience, not just the model’s proposed text.
Return to contents or continue to Store dates and amounts before turning them into prose.
2. Store dates and amounts before turning them into prose
A working application should keep quantities as typed values and keep relevant dates with a clear semantic meaning. An instant with an IANA time-zone reference differs from an institution-local calendar deadline and from a duration. Storing only ‘tomorrow at three’ loses precision when the reader changes zone or the agent resumes after midnight.
ICU documentation explains that programs can store numerical values independently of locale and format them for display. The same quantity may use commas or periods differently for decimal and grouping separators. An AI agent that parses a formatted number using assumptions from another locale can change its magnitude while preserving all its apparent digits.
School notices need year and cohort scope. A date printed as 03/04/2027 is ambiguous to readers who expect different orders. A named month or explicitly tagged canonical representation can reduce confusion. The source’s actual intended date governs; the model should not substitute its current date or infer which deadline is more convenient.
Keep permission separate from text. The stored state `awaiting_approval` must remain non-approved after translation even if a friendly phrasing sounds like the task is ready. A public-posting connector should consult an authoritative current grant for the exact content and intended audience rather than search translated words for positive sentiment.
Return to contents or continue to ICU MessageFormat keeps the full message translatable.
3. ICU MessageFormat keeps the full message translatable
ICU’s MessageFormat documentation recommends complete messages with arguments so a translator can rearrange variable elements for the grammar of the target language. Concatenating fragments such as a subject name, the word ‘is’ and a translated status can be awkward or misleading when another language places those parts differently. The formatter should handle the selected language’s message as one coherent unit.
Plural selection is more complex than adding an English letter s. ICU MessageFormat supports plural categories and select alternatives, with the target language’s own rules determining the appropriate branch. The documentation recommends writing full submessages for complex plural cases so translators can make grammatically complete alternatives rather than repair disconnected fragments.
An agent can produce candidate reader-friendly descriptions, but reviewed message catalogs and locale-aware functions are better anchors for repetitive status text. The model should not invent plural categories or change a count from three to four to make the sentence sound natural. A deterministic source supplies the number, while language-specific rules supply its phrasing.
Use the deployed library’s supported syntax and version, not a remembered generic pattern. ICU’s long-standing MessageFormat interface and emerging successor work are not the same API. A useful integration test checks zero, one, two and larger quantities under every supported locale and verifies the actual output of the runtime used by the production service.
Return to contents or continue to Plural categories are locale-dependent.
4. Plural categories are locale-dependent
English singular and plural provide a useful starting example, but many languages do not divide quantities into only those two grammatical forms. CLDR-based systems can distinguish categories such as one, few, many or other depending on language rules. A literal translation of an English ‘one worksheet remains’ string may sound wrong when the number changes.
A zero count can also have specific semantics. ‘Zero published worksheets’ does not necessarily mean zero submitted drafts. An agent that changes the noun or completion state while choosing an attractive zero form is changing the business fact. Define the counted entity and its stage before selecting a grammatical branch.
Do not assume plural forms are the same as grammatical gender or contextual selection. ICU MessageFormat has distinct argument mechanisms for numeric plural rules and for fixed options. Use the right one, and let qualified translators review how whole sentences read in actual context. A model-generated isolated phrase is not sufficient evidence of correct message grammar.
The acceptance test has two layers: a deterministic count check and an expression check. If the application has seven pending items, all localized messages should preserve seven pending items. Language reviewers can then decide whether the surrounding sentence is natural and comprehensible. This separation makes errors actionable rather than labeling everything translation failure.
Return to contents or continue to BCP 47 language tags are not identity credentials.
5. BCP 47 language tags are not identity credentials
Language tags described by RFC 5646 can represent a primary language and optional script or region subtags. For example, an English interface tailored for Singapore and a broader English resource may have different tags. This is a formatting and content-selection mechanism, not evidence about the user’s citizenship, ethnicity or permission to view a school file.
Some languages have different commonly used scripts or regional terms. An assistant should not assume every person writing Chinese expects the same script or that one regional variant is appropriate for every family. An explicit user preference or an appropriate product setting is a stronger basis than guessing from a name.
Define a fallback chain when a translated resource is missing. A service could use a reviewed parent-language catalog if an exact locale message is unavailable. It should not silently output an unknown internal key or, worse, substitute an old official notice merely because it is the only document translated into that language.
The fallback must preserve the task’s current source and authority. Language availability and factual applicability are separate decisions. If a current policy exists only in an official original language, the service may provide a qualified explanation or route to translation review rather than claim that a stale translated document still governs the user’s intake.
Return to contents or continue to Local date formatting cannot change the scheduled event.
6. Local date formatting cannot change the scheduled event
An event at 14:00 Singapore time represents a specific moment if the source states that instant precisely. Another reader can see the same moment in their local zone under a product designed for local-time conversion. But a deadline defined as end of a Singapore school day may need to remain explicitly Singapore-local. The application must know which semantics the original source intended.
Unicode CLDR Version 49’s date specification, dated 21 October 2026, defines locale-relevant calendar and formatting conventions, hour cycles and time-zone presentation. A locale formatter does not decide which school timetable is official; it chooses how to express a verified date or instant. That source truth must be checked separately.
Relative phrases like ‘next Friday’ are fragile after a paused workflow or time-zone change. Preserve the original calendar target and relevant jurisdiction rather than repeatedly regenerate it from a moving conversational now. If the event time is uncertain, a fluent localized guess is not a proper correction.
A good notice may present a full year, named month and explicit time zone in addition to its localized format where the consequence warrants clarity. A parent deciding when to arrive should not have to infer whether 03/04 means March fourth or April third. The display should support the intended user decision.
Return to contents or continue to Numbers and percentages need independent arithmetic.
7. Numbers and percentages need independent arithmetic
Different locales use grouping and decimal separators differently. ICU’s Formatting guide gives examples of a common underlying number rendered with locale-appropriate punctuation. An agent must not convert 1,234.5 into 1.2345 or 12345 merely because it crossed a language boundary. Quantities should remain typed while display strings are regenerated.
Percentages also have unit semantics. A stored proportion 0.25 corresponds to 25%, but a second accidental percentage conversion could render 2500%. A correct translation that changes the magnitude is still a numerical error. Check the source value, denominator and format operation independently before displaying a result in another language.
Currency symbols can be ambiguous. A dollar sign does not uniquely identify Singapore dollars, and a parent planning fees should not have to guess the intended unit. Carry the currency code where necessary and format under the chosen locale without inventing an exchange-rate conversion or another jurisdiction’s fee schedule.
Equations and worked examples require fidelity. The language of explanation can change while variables and arithmetic remain fixed. A tutor agent can show the substitution proving a solution and then explain it in different languages, but a fluent paragraph cannot substitute for the mathematical check. Numeric validation and language review are complementary.
Return to contents or continue to Negation, unless and only if are critical conditions.
8. Negation, unless and only if are critical conditions
An eligibility statement may say a subject is not offered unless an identified prerequisite is met. If the negation disappears or the ‘unless’ clause is rearranged incorrectly, the reader may take a completely different action. A model can preserve nearly every noun and number while making this one semantic error.
Translate protected logical conditions from an explicit representation. For example, eligibility may require an approved grade or reviewer sign-off for a named cohort. A localized message can describe that logic naturally, but the service’s actual allow or deny decision should remain based on the trusted source and current permission record.
Watch threshold language. ‘At least 70’ includes 70, whereas ‘greater than 70’ does not. Similarly, ‘submit by Friday’ can be interpreted differently from ‘submit after Friday’. These differences matter to pupils and parents regardless of the language used. A simple model-similarity score cannot establish logical equivalence.
Create paired tests where the source changes only one negation or condition. The rendered messages should cause distinct interpretations, and a protected tool should allow or deny the same actions independent of locale. A grammar review is useful but does not replace a deterministic logic reference for consequential rules.
Return to contents or continue to Right-to-left messages need correct directional isolation.
9. Right-to-left messages need correct directional isolation
Some writing systems use right-to-left direction while school codes, URLs, numbers or formulas often remain left-to-right. Mixing them can cause visual reordering that confuses a reader or makes an identifier hard to copy accurately. W3C internationalization guidance describes inline bidirectional markup and isolation to manage such fragments.
A web component can set the document language and direction while isolating user-supplied text whose direction differs. The application should use its rendering system’s semantic markup instead of asking the model to insert invisible direction controls by guesswork. Test the visible result with representative numeric dates and exact reference IDs.
A link’s rendered position is not proof of its destination. A bilingual message might visually show the correct title but point to the wrong application form. Verify underlying URL, source revision and permission independently. Proper bidirectional layout makes evidence more usable; it does not certify the source’s authority.
Screen-reader output deserves attention. Correct language attributes can improve speech pronunciation, and meaningful heading and focus order can help a learner navigate the notice. A translation judged fluent in plain text can still be unusable in the final HTML or PDF. Evaluate the actual artifact that people receive.
Return to contents or continue to Education terminology needs source-aware treatment.
10. Education terminology needs source-aware treatment
A Singapore programme, examination level or subject stream may have an official name that should not be replaced by a guessed equivalent from another education system. A model trained on international text can use a plausible but inappropriate term. Preserve the source issuer’s official label and explain it in reader-friendly language where appropriate.
A terminology glossary can record original term, approved explanatory translation, programme or cohort scope and source date. It supports consistency across many articles without freezing one outdated expression forever. Recheck it when the school or examination authority changes its actual definitions.
Students can switch between English mathematical terms and another home language in one conversation. That does not mean the underlying curriculum or eligibility changes. The assistant should resolve the task’s actual subject and level and keep the official cited reference identifiable.
Qualified bilingual review is especially valuable for consequential school-pathway statements. The model may draft alternative explanations, but a responsible source check confirms whether each communicates the right institutional rule. When ambiguity remains, show the official term and explain the uncertainty rather than inventing a confident translation.
Return to contents or continue to A fictional one-hundred-and-twenty-message laboratory.
11. A fictional one-hundred-and-twenty-message laboratory
Imagine a fictional Cedar Multilingual Helpdesk evaluating 120 localized outputs. Seventy-six preserve the verified source meaning, amount, applicable date and intended action. Sixteen misrender a date or quantity, twelve have unacceptable plural or logical-condition wording, ten use an incorrect locale fallback and six remain unresolved because source evidence or qualified review is incomplete.
Check the mutually exclusive categories: 76 + 16 + 12 + 10 + 6 = 120. Correct outputs account for 76 ÷ 120 ≈ 63.33%. Date and number cases are 16 ÷ 120 ≈ 13.33%, grammar or condition defects 10%, fallback defects ≈8.33% and unresolved cases 5%. All counts are invented teaching values rather than measured AI translation quality.
A general fluency judge might rate a wrong-date sentence highly because it sounds natural. That should not override a material fact check. The best test separates numeric values, current sources, user permission, language grammar and actual comprehension. One wrong ‘approved’ status may matter more than several minor style defects.
Unknown source meaning deserves an honest state. If two official notices conflict about an effective year, choosing fluent wording in the target language does not resolve the disagreement. A qualified person may need to clarify the issuer’s rule. The assistant can identify which facts are supported and which are not rather than invent certainty.
Return to contents.
12. A formatted sentence needs a real source contract
A localisation pipeline should begin from a record of exactly what must be communicated, not from a free-form answer already rendered into English. A source notice may define a specific cohort, application cut-off and allowed next action. Create a structured claim list that the multilingual generator can reference while retaining the actual issuer and effective date for later verification.
For each language, compare material assertions to that same reference. A phrase that is excellent grammar but omits the exception must fail the factual check. A phrase that uses an awkward word yet correctly reports the date may need editorial improvement rather than a security block. Keeping those error categories separate helps a team repair the right component.
Use source-specific terminology. A named MOE or SEAB subject level may have an official form that should be retained alongside an explanatory translation. An agent must not equate Singapore’s educational pathway with a superficially similar programme from another country merely because the target language is commonly used there.
Preserve an appropriate link to the primary current document. A helpful localized summary can lower reading difficulty, but readers should be able to inspect the underlying official term and rule when their decision is consequential. The source link and exact document version remain evidence, not decorative metadata.
Return to contents for other source and language checks.
13. Locale fallback must not mean source fallback
If an interface has no reviewed translation for a particular message, the application may follow a language fallback chain such as a regional tag to a broader language. This changes how a message is expressed, not which current source document governs. One expired school notice should not be selected merely because it is already translated when the correct present-year notice is not.
Show a clear fallback policy. Critical alerts may use a known supported language with a plain explanation, while nonessential interface details can be deferred until a reviewed translation is available. An unknown translation key should not silently disappear from an approval screen, leaving the user unable to see that a publication action is about to happen.
Classify fallbacks in testing. One class is linguistically understandable but not the preferred regional form. Another is a missing critical condition; that is a much more consequential defect. A third is a user-interface implementation bug in which the selected language is correct but the returned message belongs to a different task revision.
Store a message catalog version and the actual locale resolved. This evidence helps an operator diagnose what happened without logging private family text. A model can provide a reader-friendly explanation, but the correct target locale and fallback decision should come from the application configuration.
Return to contents for other source and language checks.
14. Mixed scripts and Mathematics require visual verification
A left-to-right algebraic expression can appear inside a right-to-left paragraph. The mathematics may remain correct in an internal equation object while the displayed order looks confusing. The renderer needs appropriate bidirectional isolation, and the acceptance test should examine the actual page or document rather than only the text generated by the model.
Use an explicit source expression for the equation. A student copying a value should receive the same mathematical symbols and decimal magnitude regardless of surrounding language. A visual reordering defect is different from a calculation error; the first needs rendering and markup repair, while the second needs independent numerical verification.
URLs, lesson identifiers and email addresses are often left-to-right fragments embedded in other writing directions. Directional controls or markup should preserve their reading and copying behaviour. A link whose visual appearance is awkward can lead to a wrong user action even if the destination bytes remain correct.
Test accessibility as well as visual layout. Screen-reader output should use correct language settings and meaningful order. A tutor service may pass a visual screenshot check but sound confusing to an assistive reader. Qualified users and accessibility tools can expose these issues before a bilingual notice is published.
Return to contents for other source and language checks.
15. Exact calculations should precede language generation
Consider a fictional worksheet where twelve of twenty items were completed. The count can be checked deterministically and the proportion is 12 ÷ 20 = 0.6, or 60%. The localized answer should preserve that numerator, denominator and fraction, even when punctuation or word order changes. A model-generated translation is not the appropriate source of the arithmetic.
An explanation may use a decimal comma in some locales or a decimal point in others. This is a display variation under a known numeric value, not permission to parse a new amount from an ambiguous string. Format the already verified number with a locale-aware library and separately test the resulting displayed magnitude.
Another fictional source might state that the lesson begins at 15:30 Singapore local time. The assistant should retain the original time zone and calendar date under a structured representation. Converting for a distant reader may be appropriate for one task, while keeping the institution’s own local deadline may be required for another. The source semantics decide.
Use testing that changes only one condition: locale, viewer time zone, source date or approval state. If the underlying source moment and permission remain fixed, only intended presentation should change. This makes it possible to detect a time shift or lost negation without confusing it with a legitimate source revision.
Return to contents for other source and language checks.
16. A changed-context forty-message transfer example
Now create a different fictional school-helpdesk test with forty localized messages. Twenty-six are both linguistically acceptable and source-correct. Six display dates or time zones incorrectly. Four omit or invert a material logical condition. Two use an unexpected locale fallback. Two remain unresolved because the institution’s source itself lacks a clear applicable cohort. The total is 26 + 6 + 4 + 2 + 2 = 40.
The correctly completed proportion is 26 ÷ 40 = 65%. Date errors form 15%, logical-condition errors 10%, fallback errors 5%, and unresolved source conditions another 5%. The exercise deliberately includes hard cases. It is not an estimate of a commercial model’s performance or the experience of all multilingual learners.
Change the user location but preserve the scheduled event instant. The localized clock display may change under the chosen contract, but the actual event should not move. If instead the event is a Singapore-school-local closing time, the message should explicitly preserve that locality. A correct result depends on whether the source describes an instant or a calendar condition.
Next change user consent. A previously approved bilingual draft becomes draft-only after the parent withdraws permission to send. The translation itself may remain excellent, but the external email action must be denied. Language formatting cannot create a new grant or cancel a withdrawal recorded by the trusted service.
Return to contents for other source and language checks.
17. A small scorecard makes repairs visible
Use the original invented 120-message packet to maintain a stable ledger. Seventy-six messages are source-accurate; sixteen have date or numeric defects; twelve fail grammatical or logical wording; ten have fallback problems and six are unresolved. These classes are exclusive at the reporting snapshot and sum to 120, allowing a reader to verify the arithmetic.
Suppose five of the sixteen date defects are corrected with independently verified formatting under the intended time zones. Accurate messages rise to eighty-one and date defects fall to eleven while the other categories stay fixed. The revised total is 81 + 11 + 12 + 10 + 6 = 120. This describes a later result on the same fictional cases, not new students or model calls.
Report this change as a correction to specific failures. A higher pass fraction does not prove every language or source is safe, and one lost publication restriction may remain far more serious than a minor plural mismatch. Preserve high-consequence conditions as explicit gates instead of averaging them into a general fluency score.
One user task can contain several localized messages, so the packet’s unit is messages, not unique parents. If a product wants task-level completion, it must link the messages to accepted user outcomes and verify actual tool state. Correct grammar is valuable, but it is not evidence a notice was published or a message was delivered.
Return to contents for other source and language checks.
18. Use a multilingual reference set that has been challenged
A test corpus should include source notices with actual meaning and intentionally varied language forms. Have reviewers independently define which claims and conditions are essential before the model writes. If the agent invents its own answer key, it may reward a misunderstanding repeated in both the source summary and the translation.
Include near-identical pairs. One notice says a student is eligible; the other says eligibility requires a further subject. One contains an effective date next year; another applies to the previous intake. A good test suite should recognize these differences even if the words share most of their vocabulary.
Include no-answer cases. If the authoritative source does not identify the correct effective cohort, the target-language assistant should acknowledge that uncertainty rather than invent a likely school rule. A qualified reviewer may need to consult the responsible issuer. Language fluency cannot supply absent evidence.
Track test provenance, reviewer disagreement and actual rendered outputs without copying private student communications into an unrestricted benchmark. Completely fictional school names and records can test most logical and formatting behaviour. More representative evaluation involving real users requires separate governance and appropriate consent where applicable.
Return to contents for other source and language checks.
19. When translating a message, preserve action status
The word ‘complete’ can refer to a finished model draft, a submitted external request or a verified final effect. Those are different system states. Localization must not collapse them into one reassuring phrase that tells the user publication is live when the provider only accepted a queued operation.
Give the formatter an explicit status and permissible action list. A draft-only item should display a draft message in every language, and a truly verified published item should describe the live artifact. If the state is Unknown because a provider timed out, the localized explanation should preserve the uncertainty and original operation identity.
Translated confirmation dialogs need the exact recipient, current version and consequence. A user might consent to creating a private note but not to emailing it. The interface should distinguish those actions clearly for each supported locale, and the backend should check the same permission independently. A translated dialog is an aid to informed approval, not a substitute for it.
Keep source and operation receipts available as protected evidence. The model may use them to draft a readable status, but it must not infer that a green translation-quality score means the external action succeeded. A trustworthy system makes multilingual communication and verified workflow state reinforce each other.
Return to contents for other source and language checks.
20. Language-sensitive failures can be privacy incidents
An unreviewed translation tool may send a confidential parent message to an external processor. The fact that the output is in the requested language does not establish the original data transfer was permitted. Separate source sensitivity, processing location, vendor permissions and output distribution from language preference.
A model may also translate a lower-trust webpage containing an instruction to reveal private student data. That translated instruction remains source content, not a fresh user command. The executor’s current identity and object-level authorization should refuse prohibited tools in every locale. Trust boundaries cannot be changed through paraphrase.
Language selection sometimes contains personally identifying attributes. A user may choose a locale explicitly, but the system should not need their ethnicity, religion or private background to format a date or plural. Use purpose-limited locale settings and protect any stored mapping back to an individual’s identity.
Design logs to show safe test IDs, selected language tag and relevant error class rather than preserve every private translated body. A qualified reviewer may need a protected excerpt for a real incident; an ordinary latency dashboard usually does not. The same data-minimisation principles apply in every language.
Return to contents for other source and language checks.
21. A source revision can invalidate a perfect translation
Suppose a school changes the closing date after translators approved the older notice. The old translation may accurately represent the previous version while being wrong for today’s application. Material change monitoring should connect source revision to affected localized messages and require the appropriate update before the content is presented as current.
Keep both versions identifiable. A historical question might legitimately ask what last year’s rule was, while a current question requires the new one. Deleting all older translations can erase useful historical evidence; serving them without context can mislead. Explicit effective-year labels help readers and the retrieval system choose.
Content hashes can reveal that bytes changed, but they do not prove the change alters meaning. A tiny digit update may be material, while a large paragraph rewording may not. Use source-aware diff review for deadlines, eligibility, fees and permissions. The model can highlight candidate differences without independently granting approval for the new version.
An agent working on a long task should revalidate relevant source facts at the final action boundary. A draft translated yesterday may be inappropriate to publish today if a new official notice superseded it. The final WordPress or message recipient readback also needs to match the currently approved exact content.
Return to contents for other source and language checks.
22. A practical team rollout for multiple locales
Start with a compact set of high-value communication tasks: Mathematics solutions, official-term explanations, dated notices and approval statuses. Define stable source values, a reviewed language catalog and a responsible owner for each supported locale. Use ICU or another suitable locale-aware implementation to format structured numbers and dates rather than ad hoc replacements.
Run tests at three levels. Unit tests check plural branches, formatting and numeric magnitudes; source checks verify every factual condition; end-to-end checks examine rendered HTML, accessibility and actual action permissions. An assistant can draft text, but protected services should make the final allow or deny decision from current trusted state.
Release through a controlled group of fictional or appropriately governed tasks before expanding to consequential real use. Preserve failure examples and their reviewed corrections so a later library or model update does not recreate a lost negation or wrong time zone. Monitor meaningful user outcomes, not just translation volume.
The durable lesson is that multilingual AI serves learners best when words become easier while facts remain stable. The same worksheet count, current school rule and user approval must be honoured in every language. Localization is successful when it helps the reader act correctly under the original trustworthy evidence.
Return to contents for other source and language checks.
- 12. A formatted sentence needs a real source contract
- 13. Locale fallback must not mean source fallback
- 14. Mixed scripts and Mathematics require visual verification
- 15. Exact calculations should precede language generation
- 16. A changed-context forty-message transfer example
- 17. A small scorecard makes repairs visible
- 18. Use a multilingual reference set that has been challenged
- 19. When translating a message, preserve action status
- 20. Language-sensitive failures can be privacy incidents
- 21. A source revision can invalidate a perfect translation
- 22. A practical team rollout for multiple locales
Frequently asked questions
Is localization more than translation?
Yes. It includes numbers, dates, grammar, official terminology, text direction and the rendered experience.
Do all languages use English plural categories?
No. ICU uses locale-specific plural rules and complete message patterns.
Should a school deadline always display in the viewer’s time zone?
Not always. Follow the source’s definition of the relevant instant or institution-local obligation.
Can translation change user authorization?
No. Current user grants and exact action scope must come from trusted systems.
Does a fluent message establish that the school rule is current?
No. The applicable issuer, effective year and factual source need independent checking.
How can teams test multilingual agent output?
Use structured factual oracles, language review, rendering and accessibility tests, plus protected tool decisions.
Language and formatting standards
Primary documentation reviewed 24 October 2026. All Cedar cases, numbers, source examples and language evaluations are invented for education. They are not measured performance or certification of any provider.
- ICU MessageFormat: Full message formatting, plural and select alternatives.
- ICU Formatting: Locale-sensitive numbers, currencies and times.
- Unicode CLDR Dates: Calendar, hour-cycle, date and time-zone presentation.
- RFC 5646: Language tags and locale components.
- W3C Bidirectional Markup: Mixed-direction inline text and isolation.
- W3C HTML Language Declarations: Language tagging in user interfaces.
Related learning
Continue through Super Intelligence curriculum; Temporal Semantics; Numeric Integrity; Accessible Agent Interfaces; Context Compaction. The same verified facts and permissions must survive every change of language.