Lee Sun-woo

Preserve the data inside the rewrite

In Pulse’s October 25, 2026 translation implementation, I kept source tables out of the model’s rewrite and restored them afterward. Explanatory prose could change language. The extracted table text needed to remain available for exact comparison with the source.

This is a source-fidelity decision. Preserving a publisher’s value does not make that value true, but it prevents the translation step from silently becoming another editor of the data.

Decide which transformations are allowed

Consider an illustrative table with a value of 0.07 and a label limiting the result to one test condition. Rounding the value to 0.1, removing the qualification or translating a label ambiguously could change what the reader concludes. A smoother-looking table would not compensate for that loss.

The implementation replaces each extracted table with a marker, translates the surrounding prose, then restores the original table after text cleanup. Keeping the table outside generation makes preservation an implementation rule instead of an instruction the model has to remember.

Conceptual transformation. The table bypasses rewriting, so its value stays the same.

If a marker is missing from the result, the table can be appended rather than discarded. That preserves the material at the cost of its original position. It is a fallback to inspect, since a table separated from the paragraph that explains it can become harder to interpret.

Let presentation adapt while the values stay fixed

The preserved table may still have English headings inside a Korean article. That is a visible compromise. It also leaves the design work of making long labels, empty cells and wide columns readable.

For a layout revision, the allowed changes can be precise: wrapping, spacing and a way to reach columns on a narrow screen. Removing a column, rounding a value or shortening a qualification changes the content and needs a separate decision.

A visual check should be paired with a content comparison. The screenshot shows whether the table can be read; the comparison shows whether cells, labels and links survived. Extraction errors and inaccurate source material remain separate problems.

Simon Willison’s Vibe engineering argues for testing and human accountability when using generated code. This case gives that responsibility a concrete target: verify the material the transformation was supposed to preserve.

The design lesson is to separate what may be rewritten from what must remain exact. Once that boundary is explicit, AI can help with the presentation without receiving permission to reinterpret the evidence.