We built the benchmark we'd want any AI estimation vendor to pass. Then we failed it.

Nine public datasets, thresholds pre-registered in git before every run, and a stop rule fixed in advance. No model class clears the bar. Here is the negative result, with the numbers.