Our changepoint detector answers "when did this series change?" with six settings at once and leads with the changes all six agree on. On seven labelled series that nobody tuned it on, that agreed list scored F1 0.71, against 0.30 for the single level model. The web page already ran the six settings in Web Workers. The command line ran them one after another, which cost about five times the plain analysis, so above 2,000 points it skipped the panel and printed a note instead.
What changed
The CLI now runs the settings on Node's worker_threads, one per core and slowest first. It starts the threads before its own analysis and does that analysis on the main thread in the meantime. When the options you passed can't move a cut away from the panel's plain "level" setting, the main result stands in for it, so only five settings go to the threads. If a thread can't start or crashes, the missing settings run on the main thread, so the answer never depends on the threads.
The answer is the same as before. A test pins the threaded result to the one-thread consensus() result, and the timing script checks the printed line against it on every input.
The numbers
On a 4-core machine, timing a whole node cli.mjs run from start to finish:
| series | panel on threads | no panel | one thread, before |
|---|---|---|---|
| tree rings, 2,000 points | 2.1 s | 0.8 s | 3.8 s |
| tree rings, 4,000 points | 4.3 s | 1.6 s | 10.4 s |
| tree rings, 7,980 points | 9.5 s | 4.0 s | 23.5 s |
| DJIA daily returns, 8,609 points | 6.4 s | 2.3 s | 15.9 s |
So the cap moved from 2,000 to 10,000 points, which covers every series we ship. The panel now costs 1.4 to 2.4 times the plain run, down from 3 to 6 times. It isn't free. The threads share four cores, so each setting runs slower than it does alone, and the slowest one sets the pace.
A question with no subject
The roadmap item went on: "then check whether the agreed summary earns its cost on the long TCPD series." We fetched the Turing Change Point Dataset again to answer it and found that it has no long series. The longest of the 31 series it ships has 816 points, and the longest of the 7 we held out has 991. Every labelled figure behind the agreed line comes from series of 52 to 991 points. So we can't yet say whether agreement is as useful on a 10,000-point series. Answering that needs labelled long series from somewhere else, and the README now says so.
Code: consensus-threads.mjs, consensus-thread.mjs, and eval/cli-panel-threads.mjs with its output file. This is statistical analysis of data you supply, not advice.