Checks each citation against the indexed corpus, verifies any quoted text is accurate, and surfaces anything in our records (later cases, conflicts) that could weaken reliance on it. A citation not found here isn't necessarily wrong -- it may simply not be indexed yet.
Ask a question above, or choose a suggested query, to search the indexed corpus of Kenyan judgments.
Chunk counts, not document counts -- a single judgment can split into anywhere from 1 to 70+ chunks depending on length, so this isn't a 1:1 proxy for "judgments indexed." Click any court for its real distinct-case count (a heavier, on-demand scroll).
Flush the /query result cache in Redis -- either everything, or just one cached query + court-filter combination.
Deletes every indexed chunk under an exact case_id. Exact match only, not a search -- confirm the case_id first (e.g. from a source card).
Runs the detection pass directly -- one LLM call per semantically-close pair found, so a broad (all-topics) run is a real cost decision, not a free click. Start with a topic and a modest limit.
Pairs shown here already cleared two independent LLM passes that had to agree (one arguing FOR a conflict, one explicitly arguing against it). Promotion still needs 2 distinct reviewer/admin accounts to independently approve -- one person's sign-off is never enough. Any single reject removes a pair immediately.