Reflection on July 23, 2026

It feels like it's been a while since I got to use the 8510 I created in the past, i.e., the debugging tool. I wanted to try it today.

First, when I tested the Qwen Validation Debugger—which performs E2E testing based on synthetic data—on about 50 items, there seemed to be countless validation issues. In that process, I realized that having the AI operate the Qwen Validation Debugger GUI while I monitor it is the most efficient approach, something I figured out last time.

Now that JSONL collection in Markdown is roughly complete and the data has been loaded into SQLite, I tested whether the tool could retrieve the appropriate JSONL documents. I let the LLM open the debugging page I built in the GUI, watch how the code gets chunked and requests are sent, and see whether the intended document content is fetched from SQLite when the strategy runs, while I observed and made adjustments from behind the scenes.

I stopped at 5 pm because of an evening appointment, but I still managed to see whether the LLM correctly checks if a document is suitable, whether it falls back to searching SQLite directly when it isn’t, whether the search works properly, and so on—a great opportunity to explore various uses.

Also, I improved the site today. Hmm… operating the site the way it is now might give the impression of a typical spam blog, something like “Does clicking this install a virus?” I thought that because whenever I search for a site I don’t trust, it often looks like a scam or spam, so I felt that might happen here. So I started by updating the icon, improving the sitemap, and setting up site links.

Additionally, among the OpenRouter models, I had been using gpt-oss:120b until the free tier ended again, so I switched the CI pipeline to the openrouter/free model.

Finally, because Vercel deployments were unstable, I made changes to ensure they are applied again. Today I’ll finish the day with a warm drink.