A synthetic AI-spend panel shaped like card and bill-pay data, January to September 2026. Seats post on cards within days; API tokens are billed after the month closes, and heavy users are mostly invoiced. Every transaction carries its usage month and its posting date, so the panel can be read as of any day and the mature truth is known.
| Read | Top 1% cutoff | Median |
|---|---|---|
| July, read Aug 12 | 13,022 | 11.40 |
| July, read Sep 9 | 14,423 | 11.51 |
| August, read Sep 9 | 10,836 | 11.36 |
| August, mature | 12,672 | 11.49 |
Nine days after month end, 85% of API spend has posted for the top 5% of spenders against 92% for everyone else. The median barely revises (+1.1%); the tail a top-1% headline is about revises +16.9%.
The nowcast adds last month's spend times the share not yet posted at this lag, learned from January to June by spender segment. 94% of July's top-1% firms are still in the top 1% in August; the cutoff fell more than the firms in it did.
The publication-day top-1% read sits 14.5% below the mature value at the median (range -23.0% to -6.2%). The nowcast cuts the error in the month-over-month change from 13.6 points to 5.1. In 5 of 30 worlds, the publication-day read showed the top 1% cutting spend when it actually grew.
The effective price fell 25.1%. A Fisher index on the same token basket attributes 19.0% to like-for-like price cuts and 7.6% to the move from frontier to standard models.
In the main world the nowcast overshoots ($13,393 against a mature $12,672). Bornhuetter-Ferguson expects August to look like July, and August had a vacation dip and price cuts. The completeness curve was right; the prior was wrong. The fix is a seasonal, price-adjusted expectation, not more history.
python3 run.py --quick # under a second python3 run.py # about 20 seconds, with the 30-world check