Spanish Localization QA for Machine-Translated Pages

Machine translation into Spanish keeps the fluency and drops the parts a model needs in order to quote you. In a sample of 400 machine-translated passages across three client sites the Spanish read fine, and 38 percent had lost at least one element that made the English passage quotable: a figure separated from its unit, a standard's identifier reformatted, a model code localised into nonsense, or the entity name carrying the attribution replaced by a pronoun. Fluency review does not catch any of that, because none of it is a fluency error. What the QA looks for: Numeric integrity, including the decimal comma and thousands separator conventions that flip between en-US and es-MX and can silently move a load rating by three orders of magnitude. Identifier survival: NOM, ASTM, EN, ISO and IP strings, part numbers and SKU formats, checked by pattern rather than by reading. Entity retention, because the sentence that names your brand as the source of a figure is the sentence that gets cited. And claim equivalence on the sentences carrying a regulated meaning, verified by back-translating those sentences only — back-translating everything is theatre and spends the reviewer's attention on prose that does not matter. The register decision: Neutral Spanish is a reasonable choice for a category page and a bad one for anything a parent or a maintenance technician has to act on. For kids' products, age grading and hazard wording go into es-MX with the vocabulary Mexican parents use, and I document where that reading diverges in Colombia or Argentina. For industrial documentation the term base decides once, in TBX, whether it is rodamiento or balero, and then nothing in the corpus is allowed to disagree. Term base first, then translation, then QA. A term base written afterwards is a glossary nobody applies. How much is checked: Acceptance sampling, not full review. A random sample sized to the corpus — sixty passages per thousand gives a usable read at 95 percent confidence for the defect rates these engines produce — scored against an MQM-derived typology cut down to the six categories that actually occur in this material. When a sample fails, the batch goes back to the engine or the vendor with the failing pattern named, instead of being repaired passage by passage at your cost. What I do not do: I am not your translation vendor and I do not price per word. I do not sign off on label text, packaging, or anything requiring a certified translator's stamp in Mexico; bring a perito traductor for that and I will tell you which strings need one. I do not build a Spanish site out of an English one. If there is no Spanish content decision underneath, QA on the output is tidying the symptoms. Who this is not for: A team that wants the Spanish site to be the English site, faster. And teams treating locale QA as a launch gate rather than a standing check — the defects come back the next time the engine or the model version changes, so a one-time pass buys a clean launch and nothing after it. What arrives: The term base, the QA harness as a script that runs against your export format, the scored sample with every defect located and classified, and a one-page acceptance rule your vendor can be held to. Your reviewers are trained in a single session; after that the pass runs without me, and the numbers it produces are comparable across quarters because the typology does not move.