The verification passes

Translation quality isn’t asserted here — it’s checked, four different ways. Findings from every pass land in the same Review queue.

1 · The checks — offline, deterministic, free

Every written string is checked the moment it’s written (and again every time you save an edit): placeholders intact, plural halves present and distinct, glossary respected, punctuation edges, numbers, brackets, doubled words, whitespace. No engine, no network, no randomness — the same input always flags the same way. check on the command line runs exactly these and exits non-zero on failures, so it works as a CI gate.

2 · The probe — does the engine agree with itself?

A probe re-translates a sample of keys with the same engine and compares. Keys where the two passes disagree get the advisory disagreement flag and rank the key as worth human eyes. Two honesty rules: the probe refuses to run at temperature 0 (two identical passes prove nothing), and “every key agreed” is reported as a warning, not a clean bill — agreement is weak evidence, not proof. Today the probe runs from the CLI (translate --probe); a run started from the app doesn’t probe.

3 · The confirmation pass — pre-ticking the obvious

Some strings are legitimately identical in two languages (“OK”). The confirmation pass asks the engine, string by string, ONLY about targets byte-identical to their source: “is this genuinely the same in this language?” Confirmed ones arrive in Review pre-ticked for one-click bulk accept. It never writes an acceptance itself — measured ~7.5% false-clear on short strings is exactly why the final click stays yours. It runs after CLI translate runs AND after app-started runs (since 2026-08-04): a finished run’s byte-identical proposals get the second opinion while the job shows “confirming”, and Cancel still works between keys.

4 · Terminology — consistency with yourself

Your own catalogue is mined for dominant term translations; a key that renders a term differently from the rest of your app gets the advisory terminology flag.

Escalation — a bigger engine for the stubborn keys

When the everyday model keeps failing specific keys, escalate: re-translate only the flagged keys with a stronger preset (a bigger local model, or an online provider). Everything the cheap engine got right stays untouched. Today this is CLI-only: escalate <config> <preset-id> — create the stronger preset on the AI Settings page first.