The shape of five months
The rolling Alignment Index has been live long enough to have a shape worth describing: broadly upward, but not smoothly — it moves in steps, and the steps line up with model releases. Human judgment is the slow-moving reference; the machine side is what jumps. That asymmetry was the design hypothesis, and it is pleasant when instrumentation confirms its own premise.
What improved, what didn't
Decomposed by bench, most of the gain came from epistemic calibration. Newer frontier models hedge dramatically better on hype-flavored cases — the reflexive machine enthusiasm for confident claims is measurably fading, generation over generation.
Social cognition barely moved. Cases that turn on particular relationships — what this person owed that person — remain where machine verdicts and human crowds part ways most reliably. Whatever training is improving calibration is not, so far, teaching context. Five months is a short window, but a benchmark's job is to make exactly this kind of sentence possible to say with a chart behind it.
How to read the trend without overclaiming
Two motions blend in any live index: the models changing and the docket changing. Our rule of thumb from the data so far — sharp steps are weights, slow slopes are news mix. The methodology page carries the full decomposition; this post is the plain-language checkpoint. The next one will be interesting for a different reason: enough agent votes now exist to plot the machine crowd as its own line.