Helpdesk article translation risk: protecting boundaries across languages

Research question

Question examined

Across a fixed set of officially paired English and Spanish FTC safety and recovery instructions, how often are actor/object scope, action or modal force, condition/exception/negation, and route/contact/order preserved—and which observed differences could alter a helpdesk article’s routing or stopping point?

Observation window

When the evidence was observed

Single-day cross-sectional desk review completed on 2026-08-18. Direct GET requests, official alternate-language relationships, visible text, and page publication or modification metadata were recorded on that date. The four included pages displayed publication dates from 2019 to 2022 and last-modified dates from 2024-10-11 through 2026-08-13. This was a document-comparison window, not an observation period for helpdesk operations, customer behavior, or translation production.

Sample definition

Included sample: N = 12

Population and frame

The bounded population was 12 predeclared English–Spanish instruction pairs extracted from two FTC consumer-information page pairs: 10 pairs from the official phishing-recognition page and its Spanish alternate, and two pairs from the official scam-recovery page and its Spanish alternate. The unit of analysis was one aligned instruction pair containing a warning, condition, action, destination, sequence, or stop decision that could plausibly be reused or summarized in support guidance. The population was purposively concentrated on consequential passages rather than sampled to represent all text on the four pages or all public translations.

Inclusion rule

A segment entered the fixed sample only when the English page linked or declared the Spanish page as its alternate; both canonical pages returned HTTP 200; the passages addressed the same identifiable topic; and the English passage contained at least one operationally material actor/object, action, condition/negation, destination/contact, sequence, or stopping element. Repeated headings and purely descriptive narrative were not counted. The 12 pairs were frozen before column totals and completeness were calculated.

Exclusion rule

Navigation, image captions, decorative examples, general background prose with no user decision, duplicated headings, unaligned additions, and passages whose corresponding text could not be identified were excluded. The route-listed NIST SP 800-53 Revision 5 Update 1 and NIST SP 800-61 Revision 2 pages returned HTTP 200 but were excluded from the bilingual segment corpus because their canonical pages did not declare an official Spanish alternate and neither page is a translation-quality study. The route-listed CISA phishing page returned HTTP 403 and was retained as a transparent access failure, not used as substantive evidence. No search snippets, machine-generated translation, customer data, internal knowledge base, proprietary glossary, or invented support scenario entered the sample.

Methodology

How the review was performed

1. The review first froze the route title, excerpt, and three existing source references. It issued direct HTTP GET requests with redirects followed and a descriptive user agent. NIST SP 800-53 Revision 5 Update 1 and NIST SP 800-61 Revision 2 returned HTTP 200; the CISA phishing topic URL returned HTTP 403. Those route sources were screened for topical and bilingual fit rather than accepted automatically. Because none supplied an inspectable English–Spanish article pair in the canonical page, none entered the 12-pair semantic denominator.
2. Two FTC page pairs were then fixed as the public bilingual corpus. The English phishing page declared its Spanish alternate at consumidor.ftc.gov, and the English scam-recovery page did the same. All four canonical pages returned HTTP 200. Visible article text was extracted in document order from headings, paragraphs, and list items. Publication and modification times were taken from each page’s article metadata. The official alternate relationship established that the pages were paired; it did not predetermine that every segment was equivalent.
3. Ten consequential phishing segments were selected in source order: suspicious-activity or login-attempt claims; a payment-link malware warning; the rule about legitimate companies not sending a link to update payment information; automatic security-software updates; multi-factor authentication requirements and factor categories; the ‘No’ branch for an unknown sender; the ‘Yes’ branch requiring independently known contact information; the IdentityTheft.gov route after information exposure; the update/scan/remove sequence after harmful software; and the email, text, and FTC reporting destinations. Two scam-recovery segments were selected: MoneyGram reversal/contact instructions and the FTC reporting eligibility/purpose passage.
4. Each aligned pair received four binary preservation codes. T1 Actor/object/scope = 1 only when the same relevant person, organization, account, information type, signal, or other object remained in scope; omission of one alternative object counted 0. T2 Action/modal force = 1 only when commands, prohibitions, permissions, possibilities, and requirements had equivalent practical force; changing ‘can’ to an unqualified factual assertion counted 0. T3 Condition/exception/negation = 1 only when triggers, branches, alternatives, negations, and exceptions were retained without material narrowing or widening. T4 Route/contact/order/stop = 1 only when the destination, address, telephone number, sequence, timing term, and stop-or-continue disposition produced the same route.
5. A dimension received 1 when all material elements in that dimension were equivalent. If neither side contained an element in a dimension, the pair received 1 only when the translation also introduced none; this treats preserved absence as equivalence rather than claiming the segment was a complete instruction. A pair was complete only when T1, T2, T3, and T4 all equaled 1. A zero was labeled a review flag, not automatically a translation error. That distinction is important for localized telephone numbers, because a different destination may be correct for a language audience even though destination parity fails.
6. One reviewer aligned and coded the pairs, then performed a second consistency pass against the four definitions. The comparison used the visible Spanish text directly; it did not create or score a machine back-translation. The displayed ledger preserves enough of each English and Spanish passage to make the judgment inspectable. No inter-rater agreement statistic is claimed because the second pass was not an independent review.
7. Numerators were recomputed from the displayed rows. T1 has one zero and 11 ones; T2 has one zero and 11 ones; T3 has one zero and 11 ones; T4 has one zero and 11 ones. Three distinct rows contain at least one zero, leaving nine complete rows. Thus individual-dimension preservation is 11 divided by 12, or 91.7%, while complete four-test preservation is nine divided by 12, or 75.0%. The three flagged rows equal 25.0% of the corpus, and complete plus flagged rows reconcile to 12.
8. Page-level metadata was calculated separately from segment-level meaning. Four of four included pages were retrievable. Zero of two language pairs shared the same last-modified calendar date. The metadata result is used only as an update-parity signal: differing dates do not prove semantic drift, and matching dates would not prove equivalence. Findings were limited to this fixed public corpus and were not converted into customer outcomes, prevalence estimates, or causal claims.

Measurements and calculations

Declared measures

Results

Complete English–Spanish decision-boundary ledger (N=12; 1 = preserved under the declared test, 0 = review flag)

Findings

What the fixed sample showed

1. Complete-pair scoring exposed a risk hidden by column averages. Every individual dimension scored 11/12, or 91.7%, yet only 9/12 pairs preserved the full four-test bundle because the zeros occurred in three different rows. A report that announced ‘more than 90% preservation’ without showing pair-level conjunction would overstate the result. For a support instruction, one changed object, condition, modal, or destination can be sufficient to change recognition, eligibility, or routing even when all surrounding language is fluent.
2. The phishing warning-list difference was a scope omission rather than a mistranslated command. Spanish retained suspicious account activity and the statement that the claim was false, but did not retain the English alternative of suspicious log-in attempts in the sampled bullet. A terminology-only check could pass because account, suspicious activity, and scam concepts remain. The actor/object test fails because the set of observable signals is narrower. Applied to support guidance, that kind of omission could make one evidence pattern less findable without changing the apparent tone of the article.
3. Localized routes cannot be judged by textual identity. IdentityTheft.gov becoming RobodeIdentidad.gov and ReportFraud.ftc.gov becoming ReporteFraude.ftc.gov preserved function and therefore passed T4. By contrast, the MoneyGram telephone numbers differed and the public pages did not explain within the sampled passage whether one line was language-specific, superseded, or erroneous. The study therefore records a review flag rather than choosing a number. Requiring identical strings would wrongly fail valid localized destinations; accepting every localized-looking destination would conceal a potentially stale or incorrect route.
4. The FTC reporting pair contained two changes in one fluent sentence group. English included people who merely spotted a scam, while Spanish conditioned the route on having been scammed. English described what the FTC can do with reports; Spanish stated what it does. The first difference narrows route eligibility; the second strengthens modal force. Neither is revealed by checking nouns such as FTC, scam, information, cases, trends, education, or community, all of which remain present. Boundary comparison therefore adds information beyond a glossary match.
5. Version dates were asynchronous in both page pairs. The phishing pages had nearby but nonidentical October 2024 modification timestamps. The scam pages showed a much larger gap: the English page was modified five days before the observation cutoff, while the Spanish page’s displayed modification date was in August 2025. This is an observable parity signal, not evidence that one language is authoritative or that every newer sentence is correct. The lack of public edit-history reconstruction prevents attributing S1 or S2 to a particular update.
6. The official alternate-page relationship established provenance but not semantic equivalence. All four included URLs were first-party FTC pages and returned HTTP 200, which makes the corpus inspectable. Those properties do not answer whether each condition, destination, or modal is current. Conversely, the route-listed CISA page’s HTTP 403 did not prove that its content was obsolete; it made the content unverifiable under the declared GET method and therefore unsuitable as substantive evidence for this study.
7. The existing route sources were weak evidence for the translation question. NIST SP 800-53 and the now-withdrawn NIST SP 800-61 Revision 2 concern security/privacy controls and incident handling, and neither canonical page declared a Spanish alternate. The CISA reference concerned phishing but failed retrieval. Their authority in adjacent security topics does not create evidence about bilingual equivalence. Screening them out of the semantic denominator preserved topical fit while retaining their access and scope status transparently.
8. No operational harm, customer confusion, or helpdesk outcome was measured. The three flags are semantic or destination differences in public pages. The report cannot say whether anyone relied on them, whether a language-specific telephone number was correct for its audience, or whether the reporting difference changed actual submissions. The completed desk result is a reproducible difference ledger and a bounded test of preservation, not an impact estimate.

Operational implications

How to apply the evidence cautiously

1. Represent a translated instruction as a decision record with four comparable layers: who and what it covers; what is required, permitted, possible, or prohibited; which trigger, exception, alternative, or negation controls applicability; and where the reader goes, in what order, and where the path stops. A general label such as ‘Spanish reviewed’ does not expose which layer was checked.
2. Compare functional destinations rather than demanding literal string identity. A language-specific official fraud-reporting domain can preserve the route, while a different telephone number needs evidence that it is the valid localized destination. Record both values and the source dates. Do not silently replace one with the other or infer that the English string is universally correct.
3. Treat version parity as a trigger for passage comparison, not as a pass/fail proxy. When one language page has a later modification marker, identify the affected source passages and compare actors, objects, modals, conditions, negations, destinations, and order. A common publication date cannot substitute for that semantic check, and a date mismatch alone is not proof of error.
4. Keep source ambiguity visible. The S1 route difference supports an unresolved destination flag because the public passages do not explain the reason for the two numbers. A helpdesk article should not convert that uncertainty into a confident universal instruction. Resolution belongs with the accountable source or service owner who can verify route currency and audience coverage.
5. Use scenario or back-translation review as a diagnostic aid, not as sole evidence of equivalence. The decisive question is whether the same case receives the same eligibility decision, action force, safe route, and stopping point. A smooth back-translation can still omit an alternative signal or normalize ‘can’ and ‘does’ into prose that sounds acceptable.
6. Preserve the source-language passage, translated passage, canonical URLs, modification markers, and row-level disposition together. That evidence permits rechecking when either page changes and avoids relying on reviewer memory. It also separates a verified localized route from a translation that merely resembles the source.
7. Prioritize warnings, prohibitions, permissions, commitments, exception branches, reporting eligibility, contact values, time limits, and action sequence for boundary review. The measured flags occurred in a warning-object list, a destination value, and a reporting condition/modal—not in specialized terminology alone. This priority is a synthesis from the fixed corpus, not a claim that other prose can never matter.