build.allanninal.dev

Part 4 of 7 · Translation relay series ~5 min read

How a reply gets translated back

The agent has read the translated message and written a reply — in their own language, the way they’d write any reply. Now the relay has to carry that reply back out in the customer’s language without changing what the agent meant, without touching a figure, and without ever sending on its own. Four small gates sit between the agent pressing “prepare” and the reply actually leaving. The last gate is a human pressing send.

Key takeaways

  • The staff member writes the reply in their own language; the relay never writes the reply.
  • Protected terms, numbers, and prices are masked before the reply is translated, then restored after.
  • The relay shows a round-trip check — the outgoing reply translated back to the staff language.
  • Nothing sends until a person presses send. The relay prepares; the human approves.
  • The reply leaves by email (SES outbound) or appears in the web widget, in the customer’s language.

Four gates on every outgoing reply

Four gates between the written reply and the sent reply A horizontal flow diagram. On the far left, a "Reply written" box: the staff member typed a reply in the team's working language and pressed prepare. Four gate bars sit in a row to the right. Gate 1: Mask terms — the same protected terms, numbers, and prices from the glossary are swapped for placeholders so the translation can't change them. Gate 2: Translate out — the masked reply is translated into the customer's language with Claude Haiku 4.5, and weak passages are re-run on Claude Sonnet 4.6, exactly like the translate-in step. Gate 3: Round-trip check — the outgoing translation is translated back into the staff language and shown next to what the agent originally wrote, so they can confirm the meaning carried; the placeholders are also restored to the exact original names and figures and shown so the agent can verify them. Gate 4: Human approve — nothing has left yet; the agent reviews the customer-language reply, the round-trip read-back, and the restored figures, then presses send or edits and re-prepares. After all four gates pass, the reply ships via SES outbound email for an email thread or appears in the web widget for a widget thread, in the customer's language. A note at the bottom: the relay prepares the reply and shows its work; a person always presses send. Reply written staff types in team language, presses prepare Gate 1 Mask terms protected terms, numbers, prices → placeholders same lock as translate-in Gate 2 Translate out Haiku 4.5 into customer language weak spans re-run on Sonnet 4.6 Gate 3 Round-trip check translate it back to staff restore terms, show figures for a check Gate 4 Human approve agent reviews reply + read-back edit and re-prepare, or press send Deliver — SES outbound email (email) or web widget (chat) in the customer’s language · figures exact · like a fluent speaker both sides saved to the thread — original and translation kept for audit The relay prepares the reply and shows its work — a person always presses send.
Fig 4. Four gates between the written reply and the sent reply. Mask the terms, translate out, show a round-trip check back to the staff language, and wait for a human to approve. Then deliver by email or in the widget, in the customer’s language.

Gate 1: mask the same things, in the same way

The outgoing direction uses the exact same locking step as the incoming one. The agent’s reply almost always repeats the figures that matter — “we’ll refund the $42.00 to your account” or “your order #4471 ships today.” Those get swapped for placeholders before translation and restored after, so the translation can rephrase the sentence around them but can never turn $42.00 into $420 or drop a digit from the order number. Part 5 covers the masking in full; here it’s enough to know the reply is protected the same way the message was.

Gate 2: translate the reply out

With the figures locked, the masked reply is translated into the customer’s language — the one saved on the thread back in Part 2. It’s the same machinery as translate-in, just pointed the other way: Claude Haiku 4.5 does the everyday work and returns a confidence read per sentence, and any sentence that comes back shaky is re-run on Claude Sonnet 4.6. A reply is usually cleaner to translate than an incoming message — the agent wrote it plainly, on purpose — so the stronger model fires even less often here than on the way in.

Gate 3: the round-trip check

This is the gate that makes the agent comfortable sending something in a language they can’t read. After translating the reply into the customer’s language, the relay translates that result back into the staff language and shows it next to what the agent originally wrote. If the two read the same, the meaning carried. If the read-back says something subtly different — “we might refund” where the agent wrote “we will refund” — the agent catches it before the customer ever sees it. The restored figures are shown too, side by side with the originals, so the agent can confirm at a glance that $42.00 is still $42.00.

A round-trip translation isn’t a perfect proof — some meaning can survive the trip out and back even if the outgoing text is slightly off — but in practice it catches the mistakes that matter: flipped negatives, changed amounts, a softened commitment, a missing “not.” It turns “trust the machine” into “read this plain check,” which is a much easier thing to ask of a busy agent.

Gate 4: a human presses send

Nothing has left the building yet. The agent now sees three things together: the customer-language reply that’s ready to go, the plain read-back of what it means, and the restored figures. They can press send, or they can edit their original reply and re-prepare — maybe they’ll simplify a sentence the round-trip showed was getting mangled, or rephrase an idiom that doesn’t travel. Only when a person presses send does the reply actually go out.

This is the core rule of the whole system, and it’s why the relay is safe to put in front of real customers: it never sends a machine-written or machine-approved message. It does the translation, it shows its work, and it hands the decision to a person. The customer gets a fluent reply in their own language; the business keeps a human accountable for every word that goes out under its name.

Then it ships, and the thread remembers

Once approved, an email thread goes out through SES outbound and a widget thread appears right in the customer’s chat window — in the customer’s language, with the figures exact, reading like a fluent speaker wrote it. Both the original (what the agent typed) and the translation (what the customer received) are saved on the thread, so a bilingual colleague who picks up the conversation later can read exactly what was said in both languages, and the audit trail shows every machine translation and every human edit.

Next post: the guardrail that runs underneath both directions — how names, numbers, and prices are locked before translation and put back after, so the machine can never quietly change a figure or rename your product.

All posts