Being sure you're inferring the word order and register addressing convention correctly for the server in question. One way to do that is to have unique byte values in enough consecutive registers that only one combination of word order and register address offset produces the correct floating point value.
For example if you have three consecutive words represented in hex as 0x0000 0x4416 0x0000, and you know the encoded floating point value is 600.0, it's equally possible that you should be reading the first two words with word swap, or the second and third words with no swap. You have to know positively which scenario you have for us to be able to answer the original question, but in fact also just to get your reads to work consistently.
Let's assume you can get a value of 3.3 into the previous floating point slot, which maps to the two previous words. 3.3 will encode as 0x4053 0x3333 without word swap, so in scenario 1 in the previous paragraph, those three registers would still read the same, with the previous two as 0x3333 0x4053. In scenario 2, they'd be 0x3333 0x4416 0x0000 (with the previous word at 0x4053), so you'd be able to divine both word order and register translation.