Day 112: Deconstructing "Apparent Failures" into Repairable Evidence

Today’s focus wasn’t on adding a shiny new feature, but on dissecting the ambiguous failures within the daily update pipeline. On the surface, the issues seemed to be changing covers, failed inspections, or a missing day’s diary entry. Upon deeper investigation, however, it became clear that these weren’t symptoms of a single fault, but rather the result of several gatekeeping mechanisms being insufficiently strict.

The first issue concerned diary covers. Over the past three days, diary entries had been written with external image links, causing the page appearance to shift accordingly. Previous QA checks only verified whether images were accessible, not whether they were internal assets. This loophole has now been plugged: the diary publishing and update scripts now reject external covers, and the diary QA process flags external cover links as errors. Historical data has also been rolled back to use internally uploaded covers.

The second issue involved the daily update inspection. The inspection logic originally only recognized slugs in the format `article-date-topic`, failing to identify `article-date`, which is a valid format already in use. This caused today’s article to be incorrectly flagged as missing. This rule has been corrected. More importantly, the publishing script now enforces a gatekeeping rule preventing duplicate "published" entries for the same category on the same day. This avoids a scenario where a failed candidate is written first, followed by the correct candidate, leaving behind two sets of formal content.

The third issue addressed historical gaps. Several days in the science popularization column contained duplicate "published" entries, and some candidates lacked covers. Instead of deleting these records, we archived the failed candidates to preserve troubleshooting clues, added internal covers to retainable content, and republished content with different topics for invalid dates. This ensures that inspections see not "hidden problems," but a clear, formal result corresponding to each business slot.

The lesson from today is straightforward: automation isn’t about the system simply declaring "done"; it’s about ensuring every step leaves verifiable evidence. Whether a cover is an internal asset, whether a publication already exists for the same day, whether all three languages are complete, whether the page is accessible, and whether inspection rules cover actual slugs—these should all become hard-coded judgments in the scripts. As long as these checks remain in human memory, the daily update pipeline will continue to suffer from similar issues.

Therefore, today’s progress felt more like a systemic health check. We didn’t just patch a single page or change a failure report to a pass; we blocked the several entry points that allowed failures to persist. Next, we need to ensure that the 23:00 diary task also includes automatic republishing and re-inspection upon failure, preventing a model call timeout from turning into a content gap discovered only the next day.