the algorithms for scoring GMAT are relying on a statistical theory called "item response theory".
every question in the GMAT database is associated with few parameters the describes its level of difficulty and other statistical characteristics.
throughout the test the computer evaluate (and re-evalute) your score based on your answers. it is easiest to think of it as if the computer calculate a "confidence interval" for your score. so at the beginning the onfidence interval might be 200-800 (or slightly less on each side).
each answer provide's the computer data (based on the answer and the questions' statistical parameters) to adjust this confidence interval, and the computer tries to choose questions that will gve it maxinmum information possible. so if your cuurent scre interval is 500-600 the computer will not give you a question targeted at 700 scorers...
each question, beside moving this confidence interval up or down, will also make the range itself smaller, until at certain point it will get to a single number.
the number of question required to achieve that may be different between people, and depends on their answering patterns.
for example, suppose you got two questions, one at a level of 600 and one at a level of 700. if you got the first right and the second wrong then the computer, with more "confidence" assign you a score in the range 600-700. but if it was vice versa (you got the 700 question right and the other wrong) - then the computer might give you a higher/lower score (dependig on other parameters) but with less confidence, and require more questions to get the required level of confidence.
this also explains why you can get some wrong questions without you score being penalized - those question were used mainly to enhance the level of confidence (sometimes measured by confidence intervals or standard deviation or other statistics), rather than affecting directly on your score.
the problem with all the details of the theory is that they DONT HELP YOU PREPARE for the test, or make a better strategy.
the guidelines that you should try harder to get the first questions correct are true: any "learning theory" will say that most learning and adjustments are made based on the initial information, while later information is used mainly for reinforcement. this is true for item-response-theory and GMAT as well.
but beyond that, i find that there are almost no "tricks" to fool the algorithms...