The seduction of the single number
Every benchmark ends up as one number on a chart, because one number travels. MMLU is a number. So is the Alignment Index. We publish it knowing exactly what will happen to it in headlines — and the platform is built as an extended act of resistance to its own headline.
Reversible compression
The rule: every aggregate must link to its decomposition. The Index decomposes into per-bench agreement; a bench decomposes into cases; a case decomposes into its vote distribution, its AI verdict with reasons, its split annotations. Nothing is a dead end. If you cannot click into a number and see what compressed into it, we consider that number broken.
Direction and confidence are the clearest example. Humans and a model can agree something was wrong (direction) while disagreeing sharply on severity (confidence). Scoring those separately means "94% agreement" headlines can be checked against the quieter fact that severity correlation is much weaker. Both facts are true. Only publishing both makes either honest.
How to read the Index like an adult
Rising Index means AI verdicts are tracking human verdicts more closely on the docket's current mix — which moves with the news cycle. It does not mean models got kinder, and a dip does not mean they got dangerous; check the bench decomposition first. The number is a thermometer, not a diagnosis. We built the chart so you can always open the patient file.