Recompute
The numbers this site publishes, recalculated in your browser from the raw scores, while you watch.
1458 conversations · computed on your machine, not ours
Why this page exists
This site keeps an audit that holds every published figure to the files underneath it. That audit runs during the build, on our machine, and reports its own success — which is worth exactly as much as your willingness to take our word for it.
So the raw data ships with the pages. The table below fetches it, redoes each calculation here, and prints what your browser got next to what we published. If any row disagrees, the row says so, and the number to believe is yours.
The resampling rows draw two thousand times rather than the four thousand the published figures use, so they will land near the published value rather than exactly on it. A gap of a tenth of a point is the sampling; a gap of a whole point is us.
The check
Fetching the scores…
| Figure | We published | You computed | |
|---|---|---|---|
| distinct scores the judge used count the different values across every conversation | 18 | … | … |
| share sitting on just three values the three commonest scores as a percentage of all of them | 88 | … | … |
| mean verdict across the archive the plain average of every published score | 54.1 | … | … |
| conversations tied on the top score how many sit on the highest value in the archive | 25 | … | … |
| of the top ten surviving a rerun resample each conversation from its own readings, rebuild the ranking, count how many of the ten are still there | 3.3 | … | … |
| of the bottom ten surviving a rerun the same resampling, at the other end of the archive | 8.4 | … | … |
| points gained by selecting the top hundred hold one reading out, select on it, score on the ones held back, and subtract the archive mean | 29.9 | … | … |
Source data: verdicts.json — the published verdict for every conversation, the individual readings behind it, and which conversation it is. Nothing else is needed to check any of this.