The main problem is that in the past some users have modified results for whatever reason. The other reason is that some PC's may have errors related to hardware and such that would cause the result to be slightly off. Thats something they want to avoid.

BoB

I agree with you about the current technique of using three seperate computers in a comparision as not needed. Except my belief goes a little deeper in that I believe we don't even need two for a comparision. The 300% inspection of results is to much, not necessary . Through the use of Statistical Quality Control (SQC) with Quality Assurance of results we can accomplish the same if not better quality than the current system. We also have a complete universe of acceptable/unacceptable results that will allow the use of this Statistical Control, I am sure we also know what systems are out of bounds and cause incorrect result. After notification if they continue to provide unacceptable results the bad systems should not be sent WU's until corrected. This cruncher quality control can be made part of an intergrated SQC system here. We have the computer power ! With the power of the higher Educational systems on board WCG, I'm sure the math wizards are available and should be contacted to ask for review our needs in this area.

Statistics is not correct all the time, isn't it?


cio_redulla

I am not involved in industrial production, quality control or assembly lines, but I can't agree at all with what your thinking. WCG (or any other project) is not producing widjets from a controlled production line. Each result is unique (unknown until returned), from unique machines, with unknown tolerances and defects. SQC, from what I read, depends on knowning what tolerances are known, with limited unknowns, but the end result (product) is very known.



We do? Each work unit is unique, with results that are to be determined by the host computer. If WCG or any other project knew what to expect then why do the computing? Maybe some upper/lower boundaries, but then you still have to weed out the subtle errors.



Really? I don't. Recently, I've had to cancel a whole pile of work units because they were overdue, which will get returned as an error. I've had the odd machine in my pile produce error'd results... some even twice. Does this mean that the particular machine is faulty? It's not. This would mean many false positives, akin to virus scans, where machines are being banned because of a few false errors.

The present system seems to be a very balanced approach, reasonable tolerances, error checking, bounds checking, etc. It certainly gets the job done.

Certainly, there may be errors in hardware that would cause the result to be slightly off. But how likely is it that the two different computers receiving the package would have the exact same error, and thus, make the exact same calculation error? Only then would the system accept a result that is erroneous. If its likelihood is very, very low to impossible, then why not have only two computers working on the same package?

I asked this question, because I believe it can be easily checked. Just check the system and how many times computer working on the same work unit have returned the exact same error. If that hasn't happened, I can't see why you need 300% verification, instead of 200% verification.