In all the quoted cases - Bushmaster I, III, III, and Bofors L/70 gun, the dispersion figures are established with pretty high confidence (that is, from actual firing tables, or from development specifications given by the respective army customers). Unless you think that these armies were deliberately feeding us with false data, these are the results. Like them or not.
For public sources, see Jane's Ammunition Handbook issue 2009, page 262: Mecar 25mm x 137 M935 APFSDS, dispersion "less than .5 mil"
It's up for debate if this is the standarddeviation, or the 50% percentile (which happens to be about 1.05 x std dev ("
s")).
In our model, all rounds will fall within 3
s (that is, we cap the normal distribution there, so there won't be outliers). That means at 2000m a circular dispersion of 3m RADIUS around the aim point for 99.7% of all rounds, of 2m RADIUS for 95% of all rounds, and about 1m RADIUS for 68% of all shots.
A BMP's frontal silhouette is about 1.40m high and about 3m wide. Therefore, vertically only about 50% of all rounds will actually be in the allowed height bracket (provided that the BMP is actually fully exposed). Horizontally, about 80% of all rounds will be in the acceptable width bracket.
In order to hit, a round must be in both the vertical and horizontal bracket. That applies to only 40% of all rounds. In other words, the majority of all rounds will actually NOT hit at 2000m!
For comparison, Jane's Ammunition Handbook issue 2009,
- page 263:
Alliant 25mm x 137 M792 HEI-T/SD, .55 x .55mrad
Alliant 25mm x 137 M791 APDS-T, .30 x .30mrad
Alliant 25mm x 137 PGU-32/U SAPHEI, .45 x .45mrad - page 278:
RO 30mm x 170 L14A2 APDS-T, .5 x .5mil (="half the dispersion of the HE-T round") - page 281:
De Kruithoorn 30mm x 173 MPDS, .4 x .4mrad
As you can see, these dispersions are by and large in a similar range. 40mm x 365 are modeled with quite comparable parameters in SB Pro.
So why are you seeing so seemingly different results?
Because your sample size simply is too small. In order to make a high confidence estimate of a standarddeviation parameter, a sample size of 4000 (!) is needed (for a confidence of (just) 95%). Your judgement is based not only on no actual controlled measurements, but even of sample sizes in an order of magnitude of about ten rounds. That simply isn't enough.