Several considerations -
(a) for routine calculations, round off should be in accordance with normal math protocols - x.5 though (x+1).499 repeated rounds to (x+1) so, for example, 4.6 rounds to 5 while 4.4 rounds to 4, presuming you require precision to an integer value.
(b) where there is a need for conservatism, this might, appropriately, be varied to maintain the conservative margin. So, for fuel required, round up to the next unit of required precision. So, if you calculate that you need 209.1 litres, this would be rounded to 210, rather than 209, litres. Alternatively, if the CG limit is 1234 mm and you are adding ballast of, say, 22.7 kg to achieve a cg at the limit (from inside the envelope), then you would restrict the ballast round off to 22 kg to keep the cg within the envelope. Alternatively, coming from outside, you would round to 23 kg.
It just takes a little bit of thought to keep things on an even keel. Be careful in exam questions to read the examiner's intent. He may well use words which infer a modified round off protocol.
(c) as to consecutive calculations, you should be running at your machine's level of precision until the final answer, at which stage you round appropriately. So, for example, if you are running the sums on your electronic calculator, just let the machine do its tricks to whatever its level of floating point precision might be and then, at the end, round off the ridiculous answer of 23,918456383981 to the appropriate precision required for the answer. So, for example, were you requiring an answer precise to 3 decimals, the answer would be 23.918 while, for an integer, 24 and so on. Clearly, you need to modify this approach if you are running the sums longhand or, say, on the slide rule but the principle remains valid.