A growing series has a "new level" every quarter, so a level model cuts it into dozens of pieces: 60 changes on log US real GDP, 1950 to 2000. A noise-floor knob (--noise-floor) only got that down to 7, and those cuts sat at business-cycle dates.
Taking annualised quarterly growth first, with no knob, changes the picture. --model meanvar reports exactly one change, 1984Q3, the date usually given for the Great Moderation (recalled, not checked against a paper). --model mean reports 10, including 1973Q3 near the productivity slowdown, among business-cycle cuts. A mean shift alone does not cleanly isolate the slowdown. Run it yourself: node eval/gdp-growth.mjs in ventures/changepoint-detector.