# LQA Desk

> A bilingual linguistic-quality desk. Paste a batch of translated segments beside their source,
> with the glossary and the target locale, and get the review that lets you sign the batch off: an
> MQM-style QA review, a termbase, a revision of the failing segments, and a locale sign-off. A free
> in-browser check reads the batch first and the model then has to answer for every one of its
> findings.

- URL: https://lqa-desk.skillsafe.ai/
- API tutorial: https://lqa-desk.skillsafe.ai/api.html
- Token management: https://lqa-desk.skillsafe.ai/tokens.html
- Platform: SkillSafe (https://skillsafe.ai/)
- Model: `gpt-terra` (resolves to `gpt-5.6-terra`), publisher markup 1000 bps
- Derived from three agent skills: `@jimliu/baoyu-translate` (primary),
  `@kostja94/translation`, `@shino369/translation-expertise`. A derived work, not affiliated with
  those skills' authors.

## What it is for

Localization quality assurance on a delivered batch, in the browser. The work object is one
bilingual paste: source segments beside their translations, in any of six shapes — tab or pipe
separated columns, two keyed string tables aligned on their keys, `|||`-delimited pairs,
language-prefixed line pairs, blank-line blocks, or plain alternating lines. Everything else — the
glossary, the locale, the register, the length budget, the brand list — describes how that batch
should be judged.

It is not a translation tool. It does not translate a document from scratch, it does not edit files,
and it never reaches a repository or a CAT server. It reviews what you paste and gives it back.

## The four lanes

All four run over the same paste and return the same envelope. `task` selects one.

| `task` | Produces |
| --- | --- |
| `qa-review` | Per-segment severities, quoted spans in both languages, a fix per error, and an MQM-style score with its arithmetic shown. Verdict: `release-ready` / `fix-then-release` / `rework` / `reject`. |
| `termbase` | One approved target per governed source term, every rendering the batch actually used with counts, and the variants to retire. Verdict: `consistent` / `drifting` / `conflicted` / `no-termbase`. |
| `revise` | Only the failing segments rewritten, each holding the source's placeholder set, markup, numbers, glossary term and length budget, shown old beside new. Verdict: `all-fixed` / `most-fixed` / `partly-fixed` / `cannot-fix`. |
| `locale-brief` | The conventions this language pair is judged on, whether the batch follows each one, and a release note. Verdict: `compliant` / `minor-deviations` / `non-compliant` / `unclear`. |

The lanes chain: `termbase` feeds `qa-review`, `qa-review` feeds `revise`, `revise` feeds
`locale-brief`. The handoff carries the previous lane's output as `upstream` context; the paste never
changes, which is what makes these one app rather than four.

## The free in-browser check

Runs on every keystroke, needs no account, costs nothing, and never leaves the browser. It is
deterministic — it judges mechanics, not meaning — and its findings are handed to the model as facts
it must reconcile.

- **Alignment** into segments across the six input shapes, with the shape it chose reported back.
- **Placeholders** across ICU (`{name}`, `{0}`), printf (`%s`, `%.2f`), positional printf (`%1$s`),
  handlebars (`{{name}}`), template literals (`${name}`), Ruby (`%{name}`), bracket (`[[name]]`) and
  escape (`\n`) forms, compared as multisets.
- **Numbers** compared as canonical values, not strings.
- **Markup** balance and tag parity.
- **Glossary** adherence, banned variants, do-not-translate names.
- **Consistency** across the batch: one source term rendered several ways; identical sources with
  different targets; different sources sharing one target.
- **Length**: declared per-segment limits, and the expansion band for the pair.
- **Locale**: sentence-final punctuation in the target's own alphabet, half-width punctuation after
  an ideograph, French no-break spaces, typographic quotes, edge whitespace, double spaces, and
  untranslated Latin runs left inside a non-Latin target.

## Design decisions that matter if you are quoting this app

An inverted finding — reporting a problem that is not there — is treated as worse than a missed one,
because a reviewer who is told a correct segment is broken stops trusting the report. Each of the
following exists because the naive reading produces a wrong finding:

- **`1.5` and `1,5` are the same number.** Numeric tokens are canonicalised to values using each
  side's own separators, and a genuinely ambiguous token (`1,000` — one thousand, or one?) keeps
  every reading it could have; a match against any reading counts.
- **A numeral spelled as a word is still that number.** An English source saying "one business day"
  against a Japanese target saying "1営業日" is a correct translation, not an added figure.
- **`%2$s %1$s` is a legal reordering.** Placeholders may move; only a real absence is an error.
- **A glossary right-hand side of `n/a`, `none`, `-` or `tbd` means there is no agreed term.** It is
  recorded for reference and never reported as a miss.
- **Glossary matching is asymmetric on purpose.** Source-side terms are matched on word boundaries so
  `log` does not fire inside `login`; target-side terms of four characters or more match as
  substrings, because German, Dutch and the Nordic languages compound and `Protokoll` is genuinely
  used inside `Aktivitätsprotokoll`. CJK, Thai and Devanagari match as substrings on both sides,
  having no word boundaries to anchor on.
- **An approved term absent literally drops to minor when an inflected, separable or compounded form
  is present.** The check is literal and says so rather than insisting.
- **Severity is computed from the mitigating facts.** A target identical to its source is
  informational when the source is a URL, a lone placeholder, a two-character label or a
  do-not-translate term, and an error when it is real prose.
- **One defect reports once.** A missing closing tag reports as unbalanced markup, and the derived
  "markup missing" is demoted and told which finding explains it; a banned variant demotes the
  glossary miss it causes.
- **Half-width punctuation in a CJK target is flagged only when it directly follows an ideograph**, so
  `1,000` and `Wi-Fi, Bluetooth` are left alone.
- **The payload sample is a golden-ratio draw, never every k-th row.** A fixed stride resonates with
  the periodic structure a string-table export always has — grouped by screen, by feature, by
  translator — so a stride of four over a file whose every fourth row is one translator's work would
  hand the model that one translator alone. Flagged segments are sent first regardless.
- **The finding count sent to the model is the length of the array sent**, never the number found.
  Anything over the payload budget is reported separately as omitted, so the model is never asked to
  reconcile something it cannot see.

## Cost and account

Running a lane is metered against the signed-in user's SkillSafe credits; the free check is not.
`ss.estimate()` is called before every run and its `hold_credits` is shown as a reservation, not a
price — the charge is usually far lower. The run button is disabled rather than allowed to 402, and
every run carries an `Idempotency-Key` that includes the lane, so a retry cannot double-bill. Both
bundled examples replay a saved run in every lane, for free and without an account.

## Privacy

The paste, the glossary and the notes are read in the browser. They leave it only when you run a
lane, and then only as the run input. Review history is stored in the app's declared `reviews`
collection against your own account, mirrored in this browser for instant paint; the local mirror is
scoped to your platform subject id and removed on sign-out.