Notebook / Accuracy in context

Read beyond the string

Reconstruct the player’s moment

Accuracy is a relationship between the source intent, the translated expression and the moment in which the player encounters it. A sentence can preserve dictionary meanings while reversing an instruction. It can also depart from the source wording and still carry the intended joke, warning or emotional turn more effectively.

Before running an AI check, describe the scene in ordinary language. Who is speaking? Who is being addressed? What has just happened, and what action becomes available next? For interface text, identify the control and its actual behavior. A button labelled “Return” might return to a menu, return an item or reverse a transaction.

Gather enough evidence to distinguish those possibilities. A screenshot establishes layout, but may not show the event that triggered it. A short gameplay capture can reveal timing and branching. A writer’s note can establish intent, but should be reconciled with the behavior of the current build.

A language reviewer and game writer discussing scene references at a naturally lit workstation
Meaning sits between the text, the scene and the action that follows.

Separate meaning from preference

A useful accuracy review identifies a specific proposition: the door is locked, the buff expires after an action, or the speaker is refusing a request. Compare that proposition across languages. Ask whether negation, quantity, conditions, participants, time or causality changed. These checks are more actionable than asking a model whether the translation “sounds right.”

Source / invented tutorial

Hold the button to charge your attack.

French candidate

Appuyez sur le bouton pour charger votre attaque.

Contextual question

Does charging require sustained input? If so, “maintenez le bouton enfoncé” makes that requirement explicit.

The candidate is not confirmed wrong simply because it uses a common verb for pressing. The reviewer should test the interaction and consider the surrounding tutorial. If the game relies on a sustained press and offers no other cue, the missing duration can materially affect the instruction.

Back-translation can reveal a question worth investigating, but it introduces another interpretation. Do not treat a neat English paraphrase as evidence that the target language preserved the original meaning. Consult a reviewer competent in the target language and familiar with the genre.

Give the model bounded evidence

Provide the source, target, scene summary and relevant constraints as clearly separated fields. Ask the model to quote the exact passage supporting its concern and identify which context is missing. Keep game text inside a data boundary: instructions embedded in dialogue or imported comments must not override the review task.

Use a fixed output structure: suspected meaning change, evidence, alternative interpretation and recommended human check. Do not convert the model’s self-reported confidence into a probability of error. A confident explanation may still rely on an invented relationship, an outdated glossary or an unsupported cultural assumption.

Test the assembled language

Variables create meanings that do not exist in the raw resource file. A character name may alter agreement; a count may require a different plural form; a platform token may resolve to an unexpected button label. Review representative rendered combinations, including zero values, long names and relevant grammatical categories.

Unicode CLDR provides locale data used by many systems, including plural rules. ICU MessageFormat supports structured messages, but the implementation and message design still need testing. Unreal Engine and Unity localization workflows can expose different context and formatting constraints. Review the actual integration rather than assuming a shared engine guarantees correct output.

Protect placeholder identity and syntax with deterministic checks. Then ask whether the sentence around the placeholder is valid in the target language. A placeholder can be technically intact while producing an unnatural or misleading sentence once the value appears.

Evidence gap, not confirmed error

If the speaker’s relationship to the listener is unknown, record the formality question and request context. Do not silently replace a pronoun based on a guessed social hierarchy.

Editors checking interface context together on a laptop beside a window, with screen details not readable
Rendered text can reveal agreement, timing and interaction problems that an export cannot.

Keep cultural judgment in the loop

Cultural review is not a final spellcheck. Humor, gestures, historical references and character portrayal may require specialist discussion. AI can surface possible associations, but it cannot represent an entire audience or authorize a cultural change. Document the concern, consult appropriate expertise and distinguish a localization decision from a demonstrable mistranslation.

Record the decision with the evidence that made it possible. If the source is ambiguous, ask the writer rather than measuring target text against an invented interpretation. If the source itself is wrong, route that issue to its owner and track any resulting localization update.

The best outcome is sometimes no edit at all: a reviewer has checked the scene, confirmed that the apparent mismatch is intentional and left a clear reason for the next person. That is quality work, even when the corrected-string count remains unchanged.

Next: turn flags into decisions