Education · 2026-09-06
Checking architectural translations at sentence and page level
How a lecturer’s terminology and editing decisions became repeatable Codex checks: translation samples, text-and-illustration separation, and a 337-page layout audit.
A readable sentence can still sit on a broken page
The project began with a lecturer asking for a contextual review of an architectural translation. The first deliverable compared approximately ten samples across the source, existing translation, and proposed revision. Codex separated missing meaning, unclear references, specialist concepts, awkward phrasing, and OCR errors. This small review established an editing direction; it did not measure the accuracy of the entire book. Subsequent feedback expanded the task from language quality to reading order and the relationship between text and illustrations.
The lecturer sets the policy; the agent applies it repeatedly
The lecturer specified Korean terminology with the original term alongside it, plus explanatory notes where needed. Another explicit decision preserved lettering inside illustrations while translating external captions and commentary. Codex incorporated these distinctions into document generation and checking rules. The early classification was imperfect: a short English fragment could be mistaken for a diagram label even when it belonged to an external caption. The correction therefore considered its position relative to the image boundary, rather than relying on text length alone.
Give each output format a specific job
The record contains an editable Word revision with explanatory notes, HTML for comparison with the source, a printable PDF, machine-readable checks, and a review report. After the lecturer identified weaknesses in the Word layout, the work continued through page comparison in HTML and PDF. These formats served different purposes: editing wording, comparing placement, and inspecting the printed result. The reusable idea is to assign each artifact a clear verification role, so that producing an additional file does not get confused with proving its quality.
No overflow does not mean no missing content
One recurring failure came from OCR merging left and right columns at a similar height into a single wide region. Codex separated the columns and used individual coordinates for difficult pages. Later inspections still found text covering artwork and empty explanatory panels, despite earlier overflow checks. The final report records the restoration of 19 grey panels and cleanup of external chapter headings. Its September 6, 2026 JSON lists 337 pages and 950 text boxes, with zero missing assets, out-of-bounds boxes, box collisions, and rendered overflows in that final check.
Read the check results within their scope
For this case study, the surviving PDF candidate matched the SHA-256 recorded in the final QA report. That links an artifact to its recorded checks; it does not establish that every translation is correct. Architectural interpretation, contextual accuracy, and the adequacy of explanations still require lecturer review. The report also notes a format limitation: six bibliography pages remain source raster images, so selecting or searching their text may be restricted. A structural pass should therefore be reported separately from linguistic and teaching approval.
A workflow to reuse for the next teaching document
Start the next document with representative language samples and a written policy for terminology, captions, and embedded labels. Then check page and asset completeness, box geometry, actual rendering, and difficult pages against the source. Separate defects that justify a shared rule change from pages needing individual adjustment. Finish with a dated record of the checks and remaining limitations, followed by the lecturer's decision on classroom use. The useful output is a reviewable preparation process, rather than an unsupported claim of complete automation.