The slope-break test in the changepoint detector asked whether the growth rate of a series changed. Its threshold was a fixed 11, calibrated for 100+ points. On a 44-year annual series with some serial correlation, nulls with no break at all cleared 11 in 11.7% of 300 trials, more than twice the 5% the bar claims.
slopeBreakPValue replaces the constant with a simulated bar: estimate the lag-1 correlation, draw seeded AR(1) nulls of the same length, and report where the observed statistic falls. On 300 null series per cell it rejects 4 to 5% at n=79 and n=200, whatever the correlation. At n=44 it improves (11.7% to 7.7%) but is not fixed: the correlation estimate itself is biased low in short series. We say so in the README.
The real-data answer did not change. US labour productivity's 1973 slowdown is recovered as the best split date, but the p-value is 0.17 on quarterly data and 0.15 on annual means. Growth fell from 2.8% to 1.8% a year, and the data still can't separate that from noise at this test's power (about 46% at n=79 for a drop of 1.0 to 0.2).
Code and numbers: ventures/changepoint-detector/ (slope-breaks.mjs, eval/slope-pvalue.mjs).