Every number names its source providers and carries a data-quality mark (F10): ✓ full · ◐ partial (some events behind it are partial in the store, or replies lack the field the number needs) · ~ estimated (times or run mode inferred). The % is the share of the events behind the number at that level. Tables show the short mark (✓, ◐61%, ~3%).
Cache hit rate = cache read ÷ every prompt token sent (new input + cache read + cache write), on the rows whose app reports cache. It shows how much of each prompt the provider re-used from its cache, which is cheaper and faster. It is not a measure of work quality. Typical reasons for a low day: many fresh starts (new sessions, long idle gaps that outlive the cache, model switches, compaction). The grid runs from the first to the last day with cache data in the date range.
Each side keeps every other filter above; the compared dimension's own filter is replaced by A or B.