Portuguese Comments in a Lisbon Team’s Model Field Test

A Lisbon checkout service with two working languages

Picture a development team in Lisbon maintaining a checkout service. Function names are in English, but many comments are in Portuguese because that is the language the team uses to discuss refunds, delivery exceptions and customer support. A new colleague asks an AI coding assistant to fix a bug. Will it understand the comments, preserve their meaning and follow the repository’s language conventions?

This is an illustrative field test, not a report of measured results. It gives a team a small, repeatable way to answer that question using its own code. You do not need a multilingual benchmark or specialist knowledge to begin.

Portuguese comments can carry business rules

A code comment is text written for people reading a program. It usually does not execute. It can explain a decision, identify an exception or warn that an apparently unnecessary step must stay.

Its language matters when it contains information that the surrounding code does not make obvious. Imagine a Portuguese comment whose English meaning is: “Keep the original payment date when retrying a refund.” An assistant might produce code that looks reasonable while accidentally resetting that date. Fluent wording in its response would not prove that it respected the rule.

Comments can also be outdated. The assistant should notice a conflict between a comment and the implementation, then identify the uncertainty. Automatically treating either one as authoritative can turn a small edit into a business logic mistake.

Three tasks for a first model field test

A model field test means trying an AI model on realistic work in a controlled setting. For this Lisbon team, a useful first round would contain three short tasks, each in an isolated branch with a known expected outcome.

• Ask the model to explain a refund function with Portuguese comments, answering in English. Check whether it identifies the date preservation rule.
• Ask it to fix a retry bug while preserving existing comment language and public identifiers. Check the resulting code and run the relevant tests.
• Give it a function where a Portuguese comment conflicts with the implementation. Check whether it flags the disagreement before assuming which behavior is correct.

Use code that the team is authorized to share with its chosen service. A small example with realistic names and rules is often enough; customer records and production credentials add nothing to this test.

A prompt that separates language from behavior

Make the assignment explicit: “Explain your reasoning in English. Preserve the language of existing comments. Keep function names and public interfaces unchanged. Fix the refund retry bug without changing the original payment date. Flag any disagreement between comments and code.”

This separates four decisions that an assistant might otherwise mix together: the language of its reply, the language inside the repository, the permitted code changes and the required behavior. If the repository requires European Portuguese for new comments, state that too. A Lisbon address alone does not communicate a writing convention to a model.

A review sheet for comparing model outputs

Start with one model, such as Claude Sonnet 5.5. Save the prompt, input files, model name and output so another developer can repeat the exercise. If you later try GPT-6 Astra, use the same starting files and instructions.

Record whether each answer preserved the business rule, passed the relevant tests, followed the language instructions and avoided unrelated edits. Have a Portuguese-speaking teammate review comment accuracy. Tests can establish that a refund date stays unchanged; they cannot establish that a rewritten comment reads naturally or preserves every qualification.

Repeat a task if consistency matters. One successful response is evidence about that attempt, not a guarantee about every future edit.

RelayVano for the team’s next test round

RelayVano is a practical option for running these comparisons through a pay-as-you-go relay. It works with Claude Code, Codex, Gemini CLI and OpenAI-compatible SDKs, with the API base at https://api.relayvano.com. Keep your test inputs and review sheet consistent as you change models.

Sign up for RelayVano when you are ready to run your own small comparison. Choose based on preserved behavior, accurate comments and manageable review work—the things your team will depend on when the next refund bug arrives.