Automated translation QA for every string

A quality pass that spends AI only where it counts

Before spending a single AI call, every row is checked for placeholder integrity (per plural form too), complete CLDR plural categories, max-length, and glossary terminology. A structurally broken row is flagged instantly, at no cost.

Rows that pass the deterministic gate get a batched Claude review (Sonnet 5) that returns a 0–100 quality score plus findings categorized as accuracy, fluency, terminology, style, placeholder, plural, or length — each with a severity and a short message.

A row is flagged for human review when any finding is major-or-worse, or when the AI score falls below the threshold of 70. Findings persist on the row and surface in the editor, so reviewers see exactly what to fix.

Self-heal has three modes: off, suggest (the default), and apply. Suggest attaches a validated fix for a human to accept; apply re-translates the failing machine row and writes a validated, version-snapshotted correction. A human-edited or reviewed row is never touched — human always wins.

Translation QA, self-heal, and style inference are available on the Team plan and up, alongside Quality Mode.

What translation QA gives you

• Deterministic checks, free — before spending a single AI call, every row is checked for placeholder integrity, complete CLDR plural categories, max-length, and glossary terminology.
• MQM-style AI review — rows that pass the deterministic gate get a Claude review that returns a 0–100 quality score plus findings categorized as accuracy, fluency, terminology, style, placeholder, plural, or length, each with a severity and a short message.
• Placeholder & consistency validation — placeholder tokens are compared between source and target (per plural form too), so a mismatched or dropped variable is caught deterministically before it ships.
• Needs-review flagging — a row is flagged when any finding is major-or-worse, or when the AI score falls below the threshold. Findings persist on the row and surface in the editor.
• Confidence-gated auto-approve — at translate time the structured-output agent self-reports a per-string confidence, and only structurally-clean, high-confidence rows auto-approve; anything uncertain routes to a human. The deterministic validator and QA pre-gate always win, so a confident-but-broken row still goes to review — reviewers see the risky strings, not all of them.
• Per-language MQM scorecard — a per-locale scorecard (score trend, top finding categories, and share of rows needing review) you can export as CSV: a procurement- and renewal-ready quality artifact, built from indexed reads at no extra AI spend.
• Guarded self-heal — choose per project: suggest attaches a validated fix for a human to accept; apply re-translates the failing machine row and writes a validated, version-snapshotted correction. A human-edited or reviewed row is never touched.
• Incremental & metered — re-running QA over an unchanged project does zero AI work; a row is only re-checked when its source moved or its translation changed. Runs are quota-metered per organization and scoped to missing, stale, selected, or all.
• Style inference — from your human-reviewed translations, transglot can infer a draft style guide (formality, tone, brand voice, and do/don’t rules) for a human to edit. It is never applied automatically.

How a QA run scores your project

• Deterministic pre-gate, free — each row is first checked, with no AI call, for placeholder integrity (per plural form too), complete CLDR plural categories, max-length, and glossary terminology.
• AI semantic review — rows that pass the gate get a batched Claude review (Sonnet 5) that returns a 0–100 score and findings, each with a category, severity, and short message.
• Flag what needs a human — a row is marked needs-review when any finding is major-or-worse or the score falls below the threshold (70). Findings persist on the row and surface in the editor.
• Optionally self-heal — depending on the project’s mode, a flagged machine row gets a validated fix attached for a human to accept (suggest) or a validated, version-snapshotted correction written (apply). Never over a human edit.

Run QA over exactly the rows you mean

A run takes the same scope shape as a translation batch, so you can score a whole project or just what changed — and re-runs stay cheap.

• missing — score only rows that have never been QA’d: the first pass over a fresh project.
• stale — re-score only rows whose source moved since they were last checked.
• selected — score a specific set of keys or locales, useful for a targeted spot-check.
• all — score every eligible row: the full sweep, still incremental under the hood.

Compared with a manual spot-check

• Coverage — reviewers eyeball a sample and hope the rest is fine; or every target row is checked deterministically and every eligible row is AI-scored.
• Broken placeholders — a dropped {count} is found in production, if at all; or it is caught deterministically before the AI pass, for free.
• Cost of re-checking — re-reviewing the whole project from scratch every release; or an incremental run where a row is only re-checked when its source or translation changed.
• Fixing issues — manually rewrite each flagged string and hope you didn’t break another; or let guarded self-heal suggest or apply a validated, version-snapshotted fix.
• Protecting human work — a bulk “fix” script can clobber a carefully reviewed string; here a human-edited or reviewed row is never overwritten.

The guarantees behind the QA pipeline

QA is designed to be safe to run often: incremental, metered, and constitutionally unable to overwrite a human edit.

• Incremental by source hash — a row is skipped on re-run when its source hash still matches and it hasn’t changed since the last check, so re-running over an unchanged project does zero AI work.
• Scoped runs — target missing (never-QA’d), stale (source moved), a specific selection, or all: the same scope shape as translation batches.
• Metered per organization — the AI review meters once per chunk against your plan’s word limit, with a quota fail-fast before dispatch and a guard so a retried chunk never re-scores an already-scored row.
• Self-heal is compare-and-swap safe — apply mode re-translates only a machine row, gates the result through every deterministic validator, and writes under a row lock that re-asserts the row is still machine-translated. A human edit mid-run wins.
• Style inference stays a draft — from human-reviewed translations, QA can infer a draft style guide (formality, tone, brand voice, do/don’t) for a human to edit and save. It is never applied automatically.