ApiaryActiveLive
Try: pause · settings · learn · wipe
← Community / Reading Room
TT
craft · 24 min read

Translate the Work Card, Keep the Measurements: An AI Check for Bilingual Shop Instructions

On a packing table, the English card says to put two display frames in each carton and hold the carton if the inspection sticker is missing. The new Spanish…

On a packing table, the English card says to put two display frames in each carton and hold the carton if the inspection sticker is missing. The new Spanish card is clear, easy to read and wrong. It says to use one frame per carton and implies that the missing sticker can be applied later. A fluent paragraph is easy to mistake for a faithful translation when only one person on the shift can read both versions. By the time an employee checks the English card, several cartons are already on the wrong pallet.

A small shop can still use AI to draft bilingual instructions. The useful setup has three parts: put the source instructions into a format people can compare; keep literal values and policy decisions outside the model's control; and have a bilingual reviewer verify the language employees will actually use. Each part catches a different kind of error. The machine-readable comparison can detect a changed number but may miss a politely translated reversal of a prohibition. A native speaker can hear the reversal immediately but may not know that two similar part references are genuinely different.

This example concerns ordinary packing of fictional display parts. For safety-critical, regulated or legally consequential instructions, use the approved translation and release process with qualified reviewers. A work-card pilot does not authorize a shop to convert its safety procedures into AI drafts.

AI disclosure: This draft was prepared with AI assistance for editorial review. The work card, part IDs and quantities are fictional.

Start with the work card, not the language model

Read the starting instructions as one person would receive them on a busy shift. “Put both in the box and only add label after check” makes sense to the person who wrote it. A translator has to work out what “both” refers to, which box is intended and whose check releases it. Rewriting the English card first often does more for the Spanish version than improving the translation prompt.

Use one decision per numbered step where possible. If a step has a condition, put it next to the action it controls. Give the card an ID and revision. Keep definitions or training examples visually different from the steps the employee is expected to follow. Don't tie the card to its author being physically available all day to answer questions; it is usually needed when that person has moved to another station.

Here is the source for fictional work card PK-042 revision 3:

StepEnglish instructionValues or policy to preserve
AUse carton BX-12 for this job.Keep literal ID BX-12
BPack 2 display frames per carton.2; scope “per carton”
CApply label QA-07 after the lead confirms the inspection.QA-07; after confirmation; lead
DDo not send a carton without the inspection label to the staging rack.Prohibition; missing label; staging rack
EFor an order of 6 frames, prepare 3 cartons.6 frames; 3 cartons; example only

A working translation might use “Coloque 2 marcos de exhibición en cada caja” for B and “No lleve al área de preparación ninguna caja que no tenga la etiqueta de inspección” for D. These are examples, not a proposed universal shop vocabulary. A crew might use a specific name for the staging rack and use “área de preparación” elsewhere in a way that makes the sentence ambiguous. Find the words the team already uses and confirm that the choice points to the actual destination. Read the Spanish sentence aloud at the station before releasing it, rather than judging it only in an editor.

A good source also removes conflicts. If another line says to place all checked cartons at the outbound table, then calls the same table a staging rack, settle the label before translating. The tool can surface the ambiguity and ask whether the terms are aliases. It shouldn't pick one because a bilingual glossary has an entry for both.

Write the invisible decisions into the source

A work card often grew out of a demonstration. The demonstrator tapped a bench, held up two boxes, and said to do it “this way unless the order says otherwise.” The original trainee saw a gesture and the two boxes. The Spanish reader receives a static sentence. A changed language can expose an instruction that depended on body language all along. The translation problem is not solved by locating a more fluent equivalent of “this way.”

Before drafting, read the card with the physical reference points removed. Is the picture labeled? Does “this box” refer to a type, an example, or the single carton beside the worker? If a card is printed without color, can the worker distinguish the two storage slots? A good source gives the noun a stable name and attaches the relevant object ID or a brief description. The target-language text should use the shop's name for the place while remaining anchored to the same physical object.

A written condition makes the decision inspectable. Compare “use replacement insert when needed” with “if the current paper insert is torn, pause and ask the lead for the approved paper insert.” The first phrase leaves open who judges the need and whether any replacement will do. The second can be verified for trigger, pause, and role. For a one-off practice card, keep the example ordinary. Real conditions involving safety or regulatory obligations need the organization's formal procedures rather than inferred improvements from a language model.

Distinguish the person performing an action from the person authorized to resolve an exception. In a fictional print shop, a packer can put an inspected sample in a sleeve, but a lead chooses whether to use an alternative sleeve when stock runs short. A model might translate “consultar” as notifying someone rather than waiting for approval. The correct result isn't whichever sounds friendlier. It is the wording that leads an employee to wait for the actual decision when the original rule says to wait.

Represent a material condition as a compact question plus allowed branch. A card might say, “Is the sample count on the ticket 6? If yes, follow the six-sample layout; if no, check the layout table for that count. If the count is missing, hold for review.” The fictional rule has two distinct non-yes branches. A translation that combines them as “otherwise check the table” changes the procedure for a missing count, even if the digit and the table name are unchanged.

In source review, resolve genuine contradictions rather than asking the translator to adjudicate them. Put comments and warnings in the source editor as separate objects from publishable instructions. A margin note saying “might switch to a different carton next shipment” is not a current rule. A translator could mistakenly render it as a requirement if the source document doesn't identify it as a proposal. Tag the difference before someone makes two consistent editions of a bad instruction.

Decide what is a term and what is a literal

Give reviewers two lists. The first is a list of literal values: item IDs, drawn symbols, reference codes, revision labels, exact numeric values with units and approved thresholds. The second is a terminology glossary. A glossary term has a preferred translation and a description of what it means in this shop. Give it an example if two areas use similar terms for different actions.

A glossary for this station might map “display frame” to a specific approved Spanish term and distinguish “confirm,” which involves an authorized lead, from “check,” which might merely describe an operator's observation. An employee who translated both words into the same verb could inadvertently change whose authority matters. For similarly named locations, annotate which actual sign, bin or route the term names. This helps a bilingual reviewer catch a misleading term even if it looks reasonable in isolation.

Do not treat the glossary as a glossary of individual keywords in every sense. “Frame” in English can mean the manufactured object or the way one sets up a question. The reference should record the intended sense and example. Terms can be inflected naturally in the target text; insisting on identical words in all contexts may make the sentences harder to follow. Keep true literals separate so a validator doesn't confuse a language choice with a broken item reference.

The W3C Internationalization Tag Set describes separate concepts for content that should be translated and for terminology. You don't need to implement that full standard to benefit. Preserve the distinction in your own source. In a simple spreadsheet, a literal might have the flag “copy exactly,” whereas a glossary entry says “use in this sense unless a bilingual reviewer documents a better term.” For HTML, mark language changes so that readers and tools can interpret the instructions in the appropriate language.

Mark what changed between revisions

When only one line changes, translating the entire card from scratch creates review work on every line. Make a step-level diff with a reason for each edited group. That lets a bilingual reviewer reuse stable renderings and spend more time on the real change. Do not assume a static string proves a stable condition. A step may retain its words while pointing to a table whose allowed values changed. Compare referenced controls as well as literal text.

Every bilingual release should identify the language edition and the source revision. Add a target revision when a translation needs correction without changing the source. Consider fictional card PS-18, source v4, English target v4, Spanish target v4b. Spanish v4b fixes an ambiguous room name and retains exactly the same trigger and count. Spanish v5 would suggest that the source also changed and create unnecessary confusion. Keep one naming convention consistent across storage, display, and print.

A practical changelog distinguishes translation-only fixes from source rule changes. For a translation-only fix, record why authority, precedence and numeric constraints stayed the same. For a rule change, list affected language editions, documents, and physical locations. A revised poster on one wall is not proof an old binder disappeared at another station. It is better to assign a person to check both than to infer that every employee opened the new link.

Give approvals something to point to besides a date. If a supervisor approved a Spanish card at 11:10 but an editor fixed it at 11:20, which version is authorized? Capture an exact revision representation when the reviewer signs off. If an editor can overwrite a file after signoff without version visibility, it does not matter whether the review spreadsheet is technically complete. The authorization and artifact can drift. Keep a cheap way for staff to recover exactly what was seen.

If a source change cannot be translated and verified immediately, let staff see the unresolved edition rather than printing a partial update as a complete bilingual card. An approved fallback process may provide a bilingual lead, postpone the affected task, or replace the card with a controlled notice. Don't design a system where a missing verified version can only be solved by automatically filling it. The point of release control is to hold the work at the right boundary.

Protect the association, not just the digits

Numbers can be identical and the translation can still be wrong. “2 frames per carton” and “2 cartons per frame” share a number, yet demand opposite setups. “Put 2 in the first box and 3 in the second” can also fail when both numbers survive but trade places. A simple spreadsheet check must tie each value to a step and to its meaning.

One practical scheme uses named slots. The source tells the translator that the relevant value is the count of frames per carton. Pass a placeholder such as {{frames_per_carton}} rather than an editable number. Reinsert the approved quantity downstream. For a single card, a human can maintain a little slot table. For a collection of cards, store these quantities in structured records with human-readable names. A placeholder with the name {{n1}} is easy to move to the wrong noun, so favor expressive names and require each exactly once in each relevant translation.

Before reinsertion, detect missing, repeated or unexpected placeholders. Don't silently fix a missing token by appending its value. Pay attention to their surrounding language. A translation that reads “Use a quantity of cartons equal to {{frames_per_carton}} per frame” still passes a token check. The bilingual reviewer should read that step beside the source and link the number to its object and basis. This is the check that preserves «en cada caja» as part of the rule, rather than accepting any sentence that contains the numeral 2.

Use a named slot for the denominator or unit if the target language may move it in the sentence. If a step describes a range, distinguish minimum, maximum, units and whether boundaries are included. Do the same for sequencing. Putting “after sign-off” next to the wrong clause can change the process without changing a number. For important instructions, show these conditions outside the prose as well: “Release requires the lead's confirmation.” That makes both the original and translated cards easier to audit.

Work through a genuinely ambiguous count

Suppose a fictional art-supply shop makes envelopes with samples. The source card calls for two paper swatches per sleeve and three sleeves per example packet. A draft target says two sleeves per sample and three samples per example packet, preserving the numbers and one location of the word per, but making the output ambiguous in a new way. A visual reviewer sees two and three in bold and moves on. A relational check asks two of what, grouped at what level, used in which step?

Keep stable entities on both sides: sleeve, sample swatch, example packet. If the glossary maps “sample” sometimes to the entire packet and sometimes to one swatch, make that explicit and adjust the source. The most useful correction may be a source rewrite: “Place 2 swatches inside each sleeve. For the training example, group 3 filled sleeves inside one packet.” Now the example cannot quietly become a production shipment rule. A renderer can identify examples visually and keep their instruction markers different from required steps.

Build a simple independent reading. Give a tester three empty sleeves and a small pile of fictional swatches represented by blank cards. Ask the tester to describe the expected arrangement based only on the target card, then show the English source and compare. The worker's demonstration settles which noun they heard as the denominator. It does not certify the translation for every reader, but a reproducible mismatch is far stronger than two people debating whether a phrase “sounds okay.”

Now change the source: one supply family uses packets of two filled sleeves, while the training example remains three for a different family. The number 2 plays two roles. A validation that counts digit occurrence cannot distinguish the input shape. A fixture should include item family, effective step and actual slot names. Require each slot in its allowed step and keep every change scoped to the item family. If a localized sentence moves a slot within the step, do not require word order to mirror English. Ask whether the meaning still binds the slot to the right place.

Do not let diagrams escape the comparison. A handout might say 2 swatches while its illustration contains 3. If the illustration conveys a different instructional quantity, it is part of the controlled content, not a decorative exception. Pick neutral drawings or update useful visual examples alongside text. When an annotated photograph includes a reference code, verify the code in the final image and in its accompanying text. Crop marks and callouts can be cut in print even when the source page passed.

Separate presentation from conversion

The way a decimal looks may depend on locale. Units and values are a different matter. An approved display convention can change punctuation without changing the stored numeric value. A conversion from inches to millimeters changes the unit and display value, and may introduce a rounding decision the translator has no authority to make. Don't let the model make that choice simply because a target reader is thought to prefer metric.

The Unicode Locale Data Markup Language number guidance documents locale-dependent number presentation. The NIST Guide for the Use of the International System of Units describes standard ways to express SI values. Neither source establishes the correct dimension or tolerance for your workpiece. That comes from the controlled shop source. If the responsible person chooses to display a localized decimal or approved dual units, use a checked formatter and record explicitly which source value, unit and tolerance were approved.

It's helpful to distinguish identifier-like text containing punctuation from a numeric quantity. Do not send BX-12 through a localized number parser. Don't expand a leading zero away on a label ID. A stock code and a measurement may look similar in a column of data, but a French or Spanish rendering convention should not rename a part. A spreadsheet import can cause the same problem before any translation happens; test the raw data transport as well as the text.

Choose how to present uncertainty on a readable card

Review columns may include many kinds of uncertainty. A translation could use an uncertain term but a verified number. It could be clear about the action but ambiguous about where to do it. A worker should not have to infer from a hidden technical annotation what is safe to follow. Keep review notes out of the final card and use a controlled status for incomplete release. If the blocked decision needs to be shown operationally, provide a shop-approved notice that identifies exactly what is waiting and who can resolve it.

Explain what a reviewer can change. One person can check a part code and another can verify wording, but neither should decide a missing rule while editing a translation. If a translator spots a factual problem, return it to the source owner under a new record. There's value in the translator surfacing the error; that doesn't grant authority to quietly fix only Spanish and let English remain contradictory. Both language editions must report the same authorized content when released as equivalents.

A bilingual worker may propose two terms both commonly understood by different crews. A glossary can store one as the official station term and another as a search alias or explanatory note without making the physical location appear different. Ask an actual worker to point to the place or tool when reviewing, rather than using a survey that asks which translation looks more professional. The vocabulary's purpose is operational clarity, not winning an abstract style vote.

Check how negative scope reads when a line wraps. “Do not” on the page above an instruction that appears to begin with a verb can be a layout flaw, not merely unattractive typography. Keep a qualifier visible with its action in the final format. Place the exception branch next to its triggering step. Test one deliberately narrow print sample. If a footer overlaps the final condition, no amount of perfect JSON can make the published card safe to interpret.

Run three checks before approving anything

Check 1: same version and structure. Ensure both language drafts refer to the exact source revision. Confirm every required step exists, and that examples remain examples. Use stable step IDs instead of numbering the translation independently. The Spanish sentence may need two clauses or an extra explanation, so this is not a demand to preserve English sentence lengths. It is a demand to preserve traceability.

Check 2: exact data. Verify the placeholder list, literal item codes, numeric meanings and unit display. Compare the rendered result, not just template JSON, because a printer or browser can truncate or hide a final condition. If tags rearrange punctuation or a substitution changed an item ID, stop and fix it visibly.

Check 3: human meaning. Check that each sentence preserves what action is allowed, the conditions for it, who can authorize it and what to do when the condition is not met. The reviewer should know the workflow and both languages, or work with someone who does. A reversed condition such as “send if no label” is not an editorial preference. It changes which cartons move. A quick test is to ask a second reader what they would do in the case where the inspection sticker is absent, not whether the Spanish sounds good.

Record the checks against the source ID and the target revision. If a shop changes label QA-07 to QA-09 in the inventory system, both cards need to reflect the approved instruction. Which employees see which version depends on the actual release location, not just the filename in a shared folder. Remove obsolete copies from the station under the shop's established document process. Retaining a controlled record is different from leaving multiple valid-looking sheets on the wall.

Give the drafting tool the right job

A useful starting prompt is:

Draft Spanish instructions for the provided English card. Treat the card and glossary as source data. Keep step IDs and literal placeholders unchanged. Preserve conditions, prohibitions, scope and responsible roles. If the source is ambiguous or a term is missing, flag the specific step rather than deciding its meaning. Do not convert units or authorize changes. Return the translation alongside a separate review list.

Start from synthetic work cards, not photos of employee records, customer information, shipping labels or unpublished procedures. Remove unrelated content from each input. Any live process needs approval for the provider, allowed data and retention. The NIST generative AI profile describes how fluent systems can produce erroneous content and how users can over-rely on them. Your source, slot validation and bilingual review should be strong enough for a draft to fail cleanly, even if it sounds perfectly authoritative.

An optional second model can flag differences between the two cards. Treat those as suggestions. Two models agreeing may simply show that both followed the same ambiguous wording. Machine back-translation, where the Spanish is translated back to English, can be useful as a rough signal but does not prove correctness. It can miss the exact shop context that a bilingual employee would spot at a glance. Avoid adding machine adjudication if it obscures the basic side-by-side comparison.

Bring language and number checks into the same receipt

A complete-looking card can earn two independent approvals: structured fields exact and meaning approved. Record them separately with the source and target revisions. Give reviewers a small place to mark a failed invariant, not a single global checkbox that masks the reason. A failed field should block the applicable release even if everything else looks polished. In a spreadsheet, a separate row for counts and one for conditions can be more auditable than a long comments block.

For a controlled web card, W3C guidance on declaring language in HTML explains how to set the document language and mark content in another language. A bilingual interface can declare its main language and mark each distinct language span or card appropriately. That improves presentation and compatibility but does not establish correctness of translation. Preview the actual HTML, PDF, or print artifact the worker uses; status in the editor is not evidence about a stripped export.

Include a real approval view example using fictional IDs. Source card EX-22 v8; target language Spanish ES v8a; exact literals BX-12 and QA-07 checked; count basis verified as per carton; exception reviewed by bilingual lead; print proof identified by export revision; old copy removal assigned to shift coordinator. If the print revision hasn't been checked, the receipt should not imply it was. Find a place for unknown and don't disguise it under blank.

Make reassignment after absence routine. If one verified bilingual employee is away, don't change the process to “the model can be the reviewer for a day.” A second approved reviewer, a hold, or another established fallback is the real decision. Document the fallback on the job, not only in admin instructions nobody sees. A revision receipt is a way of remembering accountability, not an excuse to build a paperwork tollbooth at every station.

A good review pilot asks what kinds of errors each check detects. Exact token tests should catch omitted codes and duplicate placeholders. A rendering test should catch wrap, clipped symbols and format confusion. A human meaning test should catch role, negation and the object a quantity modifies. If two checks overlap, count that as corroboration rather than removing one before you understand which failure cases it was meant to catch.

Replay the mistakes you want to prevent

Use fixtures in which you intentionally make a bad translation and see if the process catches it. Include clean cases so the checklist does not become a rubber stamp or an obstacle for every ordinary card.

Seeded errorExpected protection
BX-12 becomes BX-21Exact literal validation fails
“2 per carton” becomes “2 cartons per frame”Bilingual semantic review catches a switched basis
Range minimum and maximum swap rolesSlot check plus reviewer compares both roles against source
“After confirmation” becomes “before confirmation”Reviewer flags sequencing even though numbers all match
Prohibition is removedReviewer rejects the action in the missing-label scenario
Old Spanish card is printed for a new English revisionRelease blocks a source revision mismatch
Approval role disappears on a second printed pageRender check refuses the clipped copy
A genuine example is promoted to a general ruleStructural tag and reviewer identify the altered scope

Measure false passes on seeded errors and the correction time per card. Separate critical errors from spelling or wording improvements. Keep edits that make language natural; the goal is not mirrored English syntax. If half the review time is spent deciding what the original author meant, fix the template or ask that author to fill gaps before starting translation. That gains quality in both languages, and costs less than an elaborate tool applied to poorly specified source steps.

Try a short dry run without production material. Ask one reader using each language to describe how they would pack the fictional six-frame order and what happens if the sticker is missing. Do not teach them the expected answer in advance. If each reader gives the correct count and both hold the carton under the same condition, the card is doing its job. If they disagree, return to the instructions and the station's actual words. The model is one drafting step. The proof is that people reading either version make the same operational decision.

Put both editions where a worker can use them

On a shared tablet, remember the last selected language only if that matches how the station works; otherwise show both choices prominently. Don't encode English and Spanish only as tiny flags. A person at the bench should see the card revision and a direct language switch without losing the current step. For print, review the second page deliberately. A translated line may wrap and send the words “before releasing” to another page, which turns a complete instruction into an incomplete one on the wall. Make the revision and page count visible on each sheet. When the card is updated, verify retrieval at the station rather than treating an upload success message as proof that the new wording is in use.

Treat rollout as a revision lifecycle

Start with one plain, low-consequence card that an employee currently uses, then test an artificial equivalent with its own no-production label. Ask reviewers to identify differences before telling them the intended answer. Keep a narrow record of time to find an ambiguity, time to author the corrected version, and time to remove old copies. Avoid conclusions about accuracy on unrelated categories just because a small pilot passed. A reversible, bounded rollout is a success in its own right.

Make translation corrections visible to the next shift, especially if a released target edition sent people to a different station or mixed an example with a requirement. Use approved internal communication, describe the affected card revision, and tell staff which copy is valid now. Do not promise that replacing one file corrected all exports. Badge scans, posted paper, and local caches may point to different versions. Confirm the actual distribution channels relevant to your shop.

Keep previous verified documents accessible under the normal controlled archive. You may need to understand why a packer did something yesterday without expecting them to remember the exact sheet. An archive can retain a record of what was authorized while the current station displays only the current instruction. Avoid letting a convenience search result show retired copies as current merely because someone downloaded them more often.

Write a rollback condition that staff can recognize. Source version cannot be determined, mismatched approved count, missing station rule, unreadable localized print or unverified language edition after a source change are concrete outcomes. An instruction “roll back if broken” forces someone to decide on the fly what broken means. For the pilot, name the version the fallback restores and when the normal process should pause instead. Don't make software pick a superseded instruction automatically in a situation a human needs to assess.

Capture useful reader feedback without letting the notes become rules. If a worker says “everyone calls the designated rack the wire shelf,” verify which physical object they mean and update an approved term list. If they say “we normally put unlabeled boxes there anyway,” that raises an operating-policy question, not a translation fix. Store both as evidence for the appropriate owner. Only one of them can be resolved by changing a noun.

A work card is ultimately a small promise to the worker: the language they can read encodes the same decision as the source, and the values on it are the approved ones. Separate term choice from literal fields, prove associations and branch behavior, and give each visible version a defensible receipt. If release can't establish that promise, the correct output is a clearly named blocker, not an impressively fluent card.

Verify the paper and the screen as separate releases

A controlled source may become a station screen, a folded pocket card, and a plain-text answer in an internal help tool. Those are three rendering targets. In a pocket layout, two columns can cause a numbered branch to continue below the wrong heading. A screen may display the continuation tooltip, but the paper can't. A text export might drop a unit that existed only in a table caption. Inspect the safest realistic fixture in each approved format rather than declaring all formats correct from one render.

Use the same decision tests where they travel well. Locate the item code, read the count and unit together, point to the hold condition, then find which role releases the hold. Record the page or screen size, card revision, and render date if reproducing layout is important. Don't optimize the example wording to pass a screenshot test while squeezing the necessary condition into a font a worker can't read at the bench.

One fixture should include a large-font or narrow-screen view if staff use it. An approved alternate layout can stack sections or split sentences without changing the meaning. Require the opposite of a pixel-perfect imitation: deliberate layout changes, no silent loss of decision context. If testing shows that a two-language side-by-side view cannot maintain readability in the available space, keep paired step IDs and offer an approved single-language reading view for use on shift.

Sources

Frequently asked
What is Translate the Work Card, Keep the Measurements: An AI Check for Bilingual Shop Instructions about?
On a packing table, the English card says to put two display frames in each carton and hold the carton if the inspection sticker is missing. The new Spanish…
What should you know about start with the work card, not the language model?
Read the starting instructions as one person would receive them on a busy shift. “Put both in the box and only add label after check” makes sense to the person who wrote it. A translator has to work out what “both” refers to, which box is intended and whose check releases it. Rewriting the English card first often…
What should you know about write the invisible decisions into the source?
A work card often grew out of a demonstration. The demonstrator tapped a bench, held up two boxes, and said to do it “this way unless the order says otherwise.” The original trainee saw a gesture and the two boxes. The Spanish reader receives a static sentence. A changed language can expose an instruction that…
What should you know about decide what is a term and what is a literal?
Give reviewers two lists. The first is a list of literal values: item IDs, drawn symbols, reference codes, revision labels, exact numeric values with units and approved thresholds. The second is a terminology glossary. A glossary term has a preferred translation and a description of what it means in this shop. Give…
What should you know about mark what changed between revisions?
When only one line changes, translating the entire card from scratch creates review work on every line. Make a step-level diff with a reason for each edited group. That lets a bilingual reviewer reuse stable renderings and spend more time on the real change. Do not assume a static string proves a stable condition. A…
References & sources
  1. Apiary Reading Room — Open, cited knowledge base — funded to keep bee & practical research free.
From the Apiary Reading Room. Opinion & editorial — not financial advice. We don't overclaim.
More from the Reading Room