Friday's Release Queue: Routing TM Checks by Risk

The Friday release queue has three translation-memory jobs

Your localization team is preparing a product release. Before the new strings go live, an automated check needs to compare each translation with the team’s translation memory, or TM. The TM contains approved pairs such as an English source sentence and its Spanish translation. It may also include context, product names, status labels, and notes from human reviewers.

The check sounds simple: find similar approved translations and report anything suspicious. In practice, different checks demand different levels of reasoning. A quick match lookup can run on a small, fast model. A sentence with a changed number, missing placeholder, or legal phrase may need a stronger model. A difficult case involving tone and product context may deserve the model you reserve for final review.

Translation-memory checking has three separate decisions

Start by separating the work into three questions. First, does the new source text resemble something in the TM? Second, does the proposed translation preserve the meaning and required tokens? Third, should a human reviewer inspect the result?

The first question is mostly retrieval and comparison. It often needs little creative reasoning. The second requires careful instruction following because the model must inspect punctuation, variables, numbers, terminology, and omissions. The third requires judgment about ambiguity, tone, and business risk. Treating all three questions as one large prompt makes every request more expensive and makes failures harder to diagnose.

Route easy matches to a fast model

Use a fast model when the task has a narrow answer and a clear pass condition. Examples include checking whether a translation includes every placeholder, comparing a new string with a near-identical TM entry, or flagging a changed number. Give the model the source text, proposed translation, matching TM entry, and a small list of checks.

Ask for a compact result such as pass, review, or fail, with one short reason. A consistent output makes the next step predictable. It also prevents the model from rewriting translations when your workflow only needs a check.

For this first stage, a model such as Gemini 3.8 Flash can handle high-volume screening with low latency. The goal is to remove obvious problems before spending more tokens on uncertain cases.

Send terminology and context risks to a stronger model

Escalate a request when the fast check finds a mismatch or cannot decide. Useful escalation signals include a missing glossary term, a changed quantity, a placeholder in the wrong position, conflicting TM entries, or a large difference between the new sentence and its closest match.

The stronger request should include the evidence behind the escalation. Include the source, translation, top TM matches, glossary rules, locale, and the exact reason for review. Ask the model to identify the risk and explain which rule or meaning is affected. This gives reviewers useful evidence instead of a vague confidence score.

Claude Sonnet 5.5 is a practical choice for these context-heavy checks. It can compare several pieces of evidence while following a strict decision format. You can reserve the most capable model for cases where the outcome affects legal text, pricing, safety instructions, or a major launch.

Use routing rules before asking a model to choose

Routing works best when simple rules handle predictable cases first. For example, send strings under a defined length to the fast path, unless they contain placeholders or regulated terminology. Escalate when TM similarity falls below your chosen threshold, when multiple entries disagree, or when the first model returns review.

Keep the rules visible in configuration rather than hiding them inside a prompt. Record the selected route, model response, latency, token cost, and human decision. After a release, sample both accepted and escalated strings. If reviewers regularly override the fast path, adjust its rules or move that category to the stronger path.

Connect the routing layer to your existing translation workflow

Your application can call each route through an OpenAI-compatible SDK, so the surrounding workflow can stay the same while the selected model changes. RelayVano provides pay-as-you-go access to multiple model families through https://api.relayvano.com, which makes it practical to test a fast screening route and a deeper review route without committing to one provider.

Begin with a small batch from a real release. Compare automatic decisions with reviewer decisions, measure the cost per accepted string, and tune the escalation rules. When the results are stable, sign up for RelayVano and connect the routing layer to your production checks.