At 22:00 UTC (6 PM ET), ADS-B message volume dropped 37% — from 141,751 to 88,880 messages/hour. Not receiver failure; aircraft absence. This drop occurred 30 minutes after the first FAA ground stop (see FAA correlation below).

This analysis was not planned. It emerged step by step — each evaluation informed the next hypothesis.

;; STEP 1: Load and count
(def storm-data (load-data "2026-05-19" [21 22 23]))
(count storm-data)
;; => 352,003 messages in 3 hours

;; STEP 2: Message type distribution
(->> storm-data (map :msg-type) frequencies)
;; => {"3" 27782, "4" 39241, "7" 205789, "8" 68109}
;; MSG,3 = positions, MSG,1 = callsigns

;; STEP 3: Correlate callsigns to positions via ICAO hex
(def icao->cs (build-callsign-map storm-data))
;; => 203 unique aircraft

;; STEP 4: Hypothesis — are aircraft holding?
(def by-icao (group-by :icao (filter :lat storm-data)))

;; STEP 5: Calculate geographic spread per aircraft
(defn spread-nm [positions]
  (let [lats (map :lat positions)
        lons (map :lon positions)]
    (apply max (map #(haversine-nm (mean lats) (mean lons) (:lat %) (:lon %))
                    positions))))

;; STEP 6: The threshold wasn't known until step 5 revealed the distribution
(->> by-icao
     (map (fn [[icao ps]]
            {:callsign (get icao->cs icao icao)
             :positions (count ps)
             :spread (spread-nm ps)
             :alt-range (- (apply max (keep :altitude ps))
                           (apply min (keep :altitude ps)))}))
     (filter #(< (:spread %) 15))
     (filter #(< (:alt-range %) 3000))
     (sort-by :positions >)
     (take 10))
;; => JBU834 (347 pos, 13.5nm), AAL2595 (295, 11.3nm), SWR52 (261, 14.3nm)

The filter predicate at step 6 — < 15nm spread, < 3000ft altitude range — could not have been written in advance. The thresholds came from inspecting the distribution at step 5. This is the non-zero-shot argument: the analysis responds to its own output.

Position data revealed 56 aircraft in holding patterns northwest of Logan, centered on the BOSCO/ROBUC holding fixes at ~42.45°N, 71.08°W.

Swiss 52 reported 261 positions with only 25-foot altitude variance — textbook holding at assigned altitude.

102 aircraft descended through the 8,000 ft marker during the storm window. The approach band (5-12k ft) shows aircraft stacking up in holding patterns rather than completing their descents to the terminal area. Terminal operations (<5,000 ft) dropped 60%. Cruise traffic (>28,000 ft) continued unimpeded — the signature of a ground stop.

The FAA ATCSCC issued multiple advisories for BOS on May 19. The timeline shows our ADS-B data captured the effect of the ground stop, not a precursor:

The ADS-B message drop at 22:00Z occurred 30 minutes after the first ground stop (advisory 118 at 21:30Z). This is the expected propagation delay — departures stop, then airborne traffic clears.

Source: Advisory 160 — 3,277 total delay-minutes, ground congestion on taxiways after brief reopenings.

May 19 produced 167 ATCSCC advisories — nearly double the 90/day baseline. The chart shows it in context: a clear anomaly, though March 16 (216) was worse. Other airports also ground-stopped, but those are outside our receiver range.