Given the wild but apparently "consistent" numbers I wonder if we could reverse-engineer the wrong algorithm with enough data points? Maybe the proper cost estimate has some relationship to the reported cost estimate.