Review workflow

Locavello's quality model: the agent does the volume, mechanical gates catch what machines can catch, and human reviewers — free seats, on every tier — decide what actually ships where it matters.

The lifecycle

             agent first pass
                   │
                   ▼
               machine ──────────────┐
                   │                 │  (gated namespace:
     human edit    ▼                 │   agent output lands
        ┌──── needs_review ◄─────────┘   here instead)
        │          │
        │   reviewer decides
        │      ┌───┴────┐
        ▼      ▼        ▼
    approved        rejected (with a reason)
        │                │
   feeds the TM     excluded from publishing,
                    re-queued for the next agent pass

• machine — the agent's first pass, untouched. In a gated namespace, agent output is downgraded to needs_review on write — machine output is never a terminal state there.
• needs_review — the default for human edits, and agent output in gated namespaces.
• approved — a reviewer signed off (or an editor saved with approve-in-one-step). The reviewer is recorded, and the value is upserted into the account-wide translation memory.
• rejected — rejected with a required, stored reason. Rejected values never publish, count as missing in the check gate, and the key is picked up again by the next agent pass for that locale.

Every one of these writes — agent or human — passes the placeholder-safety gate first.

The review queue

The queue collects everything with status machine or needs_review, oldest first, optionally filtered by locale. Each entry carries the context a reviewer needs to decide without leaving the queue: source text, key description, screenshot URL if set, placeholders, length budget, usage context, and the namespace's review policy.

The workbench

For deeper work than the queue, the workbench is a three-pane editor:

• Left — the key list, with filters: namespace, locale, status (including missing), and full-text search over names and source text.
• Center — the editor for the active locale, with the source's placeholder chips and live validation.
• Right — context: description, length budget, usage sites, translation-memory suggestions (exact match first, then fuzzy — one click inserts), and glossary hits for terms present in the source.

What publishing ships

Publishing a release evaluates each key against its namespace policy:

	standard namespace	gated namespace
approved	ships	ships
machine / needs_review	ships (best available)	held back
rejected / no translation	held back	held back

Standard namespaces optimize for coverage — a reviewed-later machine translation is usually better than an untranslated string. Gated namespaces (mark legal, pricing, anything liability-shaped) optimize for control: nothing unapproved ever reaches users from them.

Held-back keys simply fall back at runtime — to the locale's fallback chain, then the source language. Never a blank.

Reviewers are free

There is no per-seat charge for translators or reviewers on any tier — inviting a native speaker to review never costs money. Plans are priced on projects, keys, and agent words instead. See pricing.