Untracked means the Core universe did not price the ticker in core.price_daily over the three open sessions ending on the session shown. That test is point-in-time and survives a late price run: at 17:50 ET on 2026-09-15 only 30 of 39 Core names had been priced by the 17:30 Tiingo job, and a single-day test would have called the other nine untracked.
A session is every alert whose created_at falls in [D 00:00 ET, the next open session 00:00 ET). created_at is when the alert was knowable, so a session’s set is final at its 21:15 ET collection. Measured over 2026-08-17..09-14: 808 alerts raised after the close stay with their day, none were raised before an open, and none overnight, at a weekend or on a holiday.
Index products are the cash-settled roots SPX SPXW XSP RUT RUTW NDX NDXP VIX VIXW MXEF XND DJX OEX XEO, and the default drops them because this page is about NAMES that could earn a watchlist place. The cut is not silent: every figure in the summary band names it, and the band carries the dropped total beside the kept one. ?index=include folds it back in; ?index=only shows it alone.
The percentile is the only statistic on this page and it is computed in SQL, to core.metric_context()’s definition — midrank including the current reading, over the 365 days ending at the session, cut at 2026-08-17 because the backfill began mid-session on 2026-08-14. A session with no alert is NOT a zero in a ticker’s history — the vendor’s rules did not fire, which is not a measurement of its flow — so n counts alerted sessions only. With about a month of table, n is at most 19 before 2026-09-14 and “insufficient data” is the dominant state. That is correct, not a gap.
Not built: per-user dismiss with flare-resurface (DESIGN.md page 7) needs a write path, and src/lib/db.ts opens every transaction READ ONLY by design. Where that state should live is an owner decision; until it is made, ?stale= is the whole of the auto-hide behaviour.
Cost: two queries in sequence — the session, then the window. Measured warm on 2026-09-15: the session 32–37 ms, the window 259–457 ms at these defaults and 501–659 ms at ?stale=10&limit=200&index=include; 453–462 ms for the whole page end to end at the defaults. The window is read whole off the created_at index, so its cost is linear in the auto-hide parameter.
Flow → OI asks the next question of the tracked names: did the settled open interest confirm the print?