As the task involves floating point operations with limited precision, rounding creates some issues. Some of the published solutions fail sometimes due to rounding to a different value.

This could be an interesting kata if you would lower the limit to 26.