The user wants the precedent curation page regenerated, and there's a relevant skill for this: precedent-knowledge-layer-weekly-refresh. I should read that skill first using cat via Bash as instructed.

There's a skill I should follow the procedure for. First I'll read the four types of pages and take a snapshot of the corpus.

Now I'll take a snapshot of the corpus in a batch of four sentences, using d1q.mjs with a single-line SQL query.

The corpus counts remain unchanged, and I confirm there are no stored court cases—everything is from national statute information with null workspace IDs. I need to recompute E1-E4, starting by reading the existing three-page set, with the statute cluster being the largest at 33KB, so I'll begin with the keyword axis and leading cases first.

I'm setting up parallel queries: E1 filters domain by query_tag matching 구상금/보증채무/대위변제/신용보증, totaling 1844 records, and I'm drafting the SQL to join precedent_statutes with precedents grouping by statute_key to get the top 20. I still need representative case counts by citation rank, and I should verify the actual column names in precedent_statutes before running these queries.

I need to recalculate E1 through E4, plus pull the top statutes overall across 51k records, top statutes per tag, and top 20 by domain. I should check the statute cluster page structure first and run these queries in parallel, and I still need to verify the remaining schema for precedent_statutes and the citations columns before building out the larger query batch.

Now I'm recalculating E1 by building a query that filters precedents by domain keyword tags, groups them by statute key, counts cases per statute, and identifies representative cases using the top-indegree entries within each cluster.

Given the complexity of combining case counts with representative case rankings in one query, I'm splitting this into two separate queries instead, and checking whether SQLite supports ROW_NUMBER for the ranking logic I need.