When a translation does harm

Two words, two laws

The German word Mehrwertsteuer belongs to the German Umsatzsteuergesetz. The British word VAT belongs to a different law. Both words name a tax on added value. The rules that decide the rate, the exemptions, and the point of liability are not the same rules.

A glossary that writes Mehrwertsteuer and VAT as one term states an equality the law does not grant. The statement is not merely imprecise. It removes the only place in the repository where a reader finds out that a legal difference exists.

A translation is a decision, not a lookup

A translation across a context boundary is a claim that two words stand for the same fact. In a multilingual team that claim is often false. The format therefore keeps the word the domain expert actually said, and not the word a translator reaches for. A concept carries a preferred label per language, so German and English sit side by side and neither stands in for the other.

Where two terms in two bounded contexts do relate, the atlas records the relation as a mapping with a named strength. A close match says that one word stands in for the other in the translation chain. It does not say the two are the same thing. The record exists for that difference.

A rejected word stays in the entry

The plainer way to handle a word the team dropped is to delete it. The format records it instead, inside the entry that dropped it, with the reason and the date.

The reason is what makes the record useful. A rejection without a reason is indistinguishable from a change of mind, so the format treats a rejected word without a reason as an error. The date places the rejection against the other decisions of the same period.

What the record is for

A later reader who does not know the decision searches for the word he learned. Without the record he finds nothing, concludes that nobody considered the word, and puts it back. The team then repeats a discussion it already held, and sometimes reverses a decision on grounds nobody wrote down.

With the record, the same reader finds the word, finds the reason the team dropped it, and finds the date. He can still argue that the decision was wrong. He cannot undo it by accident, which is the failure the record exists to prevent.

An invented word runs through the same mechanism

An artificial intelligence agent can coin a domain term and then treat it as settled before anyone decides it. In one working session of this project an agent coined four terms nobody agreed on, and two of them stood in a formal specification before anyone noticed. The incident happened under review, which is why the format does not rely on a reviewer who catches the next one.

A concept therefore records who put the word into circulation and how far the domain expert settled it. A word the agent coined counts as decided only where a recorded decision says so, and that decision names who confirmed the word and when. An agent can still propose a useful term, and the proposal stays visibly a proposal until someone settles it.

The two cases are the same mechanism seen twice. A rejected word and a confirmed coinage both survive as a dated record of a decision inside the entry. Neither survives as an outcome with no author. Both records stay inside one bounded context’s glossary, for the reason given in Why the language is per bounded context.