2026 July 21 Reflection

Yesterday early morning I watched the World Cup final in Cheonan with friends.

Watching the World Cup final together made me almost not sleep at all, and I returned home in a irritated state, and it seems I started a normal day from the afternoon.

First, I thought about yesterday’s fun today. Their styles and strategies are all different. However, those things seemed to continue almost consistently. The start is the same for countries that share roots in the 20th century, but each of them has unique stats. In other words, what came to mind today is that if it gets messed up, a dataset in an undesired style might appear.

Since I couldn’t work for two days and it kept getting postponed, I thought that if we can’t put JSONL into PostgreSQL, we should push it into SQLite3. Using the Markdown → JSONL Converter, we processed the official document JSONL collection up to 1,422 out of 1,550 items.

But the question that arose was that when I created the validator, I only thought about making and controlling it, and didn’t consider the verification stage. So the solution I applied was to let AI fully operate the web debugging tool I originally made for convenience, and also have AI automatically run the Godot 3 and Godot 4 verification tools to perform the checks. As a result, 8520 has gradually improved to some extent since last week.

Thus, the JSONLs generated on today’s 8501 page, which we are moving to SQLite3, began to look suspicious. I think I compared the original text and JSONL 1:1 with AI to see if they were properly created based on Markdown. Doing that seemed to catch the incorrect JSONLs.

Hmm… I think the next task should be to try that. When doing a 1:1 original‑text comparison to see why a JSONL is accurate, it would be good to create another web tool that can show the diff and display the verification result, indicating why AI judged it as incorrect. It’s too difficult, so ending the day here feels comfortable.