Trippple Checks

[LaurV]

We finished all 119 TC assignments in 24-26M. All matched, none done by ourselves before (so, no conflicts). Well, honestly they didn't matched all from the first try, some were stubborn and need quadruple checks. But we checked every one of them _before_ reporting (i.e. looked into the DB), and if no match, we re-run till a match was found. This way we found out that the problem was with our cards (a bit overclocked too much, for this special assignment). Essentially is that all previous DC results proved to be right. We didn't find any mistake reason to show our muscles...

We are not going to take other TC assignments right now. We prepare slowly to close the mill for July, when we will not be around (i.e. we try to liquidate old assignments and projects we already have).

[Mark Rose]

I'll take these:

DoubleCheck=27290569,71,1
DoubleCheck=27309781,71,1
DoubleCheck=27311101,71,1

Should be done early July.

[Madpoo]

Originally Posted by LaurV

We finished all 119 TC assignments in 24-26M. All matched, none done by ourselves before (so, no conflicts). Well, honestly they didn't matched all from the first try, some were stubborn and need quadruple checks. But we checked every one of them _before_ reporting (i.e. looked into the DB), and if no match, we re-run till a match was found. This way we found out that the problem was with our cards (a bit overclocked too much, for this special assignment). Essentially is that all previous DC results proved to be right. We didn't find any mistake reason to show our muscles...

We are not going to take other TC assignments right now. We prepare slowly to close the mill for July, when we will not be around (i.e. we try to liquidate old assignments and projects we already have).

Cool, thanks for helping out.

I'm guessing I'll have my current stuff done by mid-July which might include taking on the 32M range. If there are still any of those unassigned blocks at that point I'll snag whatever's left.

[petrw1]

Originally Posted by Madpoo

Cool, thanks for helping out.

I'm guessing I'll have my current stuff done by mid-July which might include taking on the 32M range. If there are still any of those unassigned blocks at that point I'll snag whatever's left.

Late July for me....if there are any left them I may take a little more

[Madpoo]

Originally Posted by Mark Rose

I'll take these:

DoubleCheck=27290569,71,1
DoubleCheck=27309781,71,1
DoubleCheck=27311101,71,1

Should be done early July.

Duly noted.

I'm trying to figure out what I could tackle once this "little" project is done.

I reckon I'll check for future self-verified work and continue to knock those out as they show up, keep a backlog from building up.

I also thought it'd be fun to look ahead and find those exponents that need an *actual* triple-check because none of the previous runs have matching residues. Could be interesting. There are 6,531 of those currently unassigned (and then there are 25 of them where I've already submitted my own result that didn't match anyone else, so I'd prefer someone else triple-check those).

So that'd keep me busy for quite a while, and hopefully clear out some dubious results, helping to further point out machines with consistently bad track records.

Then I'd look at exponents that have only had one test run so far but *might* be bad, based on that computer's track record, or that have an error code that indicates a chance of failure.

End result would be trying to skip ahead a little bit and do "smart guessing" for any missed primes in the gaps up to M57,885,161.

[LaurV]

Originally Posted by Madpoo

I'm trying to figure out what I could tackle once this "little" project is done.

A new project, "Double checking LaurV's unverified LL work". This way you avoid doing triple checks, because I will DC my own work anyhow (limited by the availability of the resources, but in plan...) , and then you will need to TC it

[cuBerBruce]

Originally Posted by Madpoo

I'm trying to figure out what I could tackle once this "little" project is done.

TheMawn recently verified one of my results that also had a non-matching residue. The person who had the bad residue appears to have only one other LL result, which also is a mismatch of a previous result for the same exponent. So I expect "Daniel" has the correct result for M55100509 rather than this other user. I won't say his user name here, but I'll just refer to him as "Mister Arkansas."

I don't know if there are other users whose only "DC'ed" results have been mismatches or verified bad, but that might be something to look for.

[Madpoo]

Originally Posted by cuBerBruce

TheMawn recently verified one of my results that also had a non-matching residue. The person who had the bad residue appears to have only one other LL result, which also is a mismatch of a previous result for the same exponent. So I expect "Daniel" has the correct result for M55100509 rather than this other user. I won't say his user name here, but I'll just refer to him as "Mister Arkansas."

I don't know if there are other users whose only "DC'ed" results have been mismatches or verified bad, but that might be something to look for.

There should be some room coming up for doing fun analysis of "bad computers". Unfortunately some of that analysis is skewed towards the smaller exponents since those are way more likely to have had a double-check attempted, and a subsequent triple-check to clear up the conflict.

As of this moment, there are 6,452 unassigned exponents out there with at least 2 tests that were done, but no matches. About 1/3 are under 40M. That's a lot of hidden data there... somewhere in there are a lot of bad results. I could probably look at who submitted those results and make some guesses on which one is bad, but it really takes a triple-check to find out for sure.

Based on the analysis I've already done in the past, I've identified a few systems that generated consistently bad results and I've done my own double-checks on their unverified results. I've come up with different residues for many of them, but now someone else needs to triple-check those and make sure mine were the correct answers. I'm pretty sure they are, and I feel better knowing there weren't any hidden primes in that limited batch. I think it was like 40 - 50 I tested, and about half are now waiting for triple checks.

Another odd thing... by the time we find out that a certain computer had bad results, that computer has already been retired from use, long ago. We can then look at their other work and do double-checks on them ahead of schedule, but that's about it.

[Madpoo]

I saw this new self-verified double-check come in recently, and it bugged me a bit:
M38301469

It bugged me because the user checked in their result which didn't match the first one. Okay, that happens, no big deal.

But then it was assigned to a new user that same day for the triple-check it needs to clear up the conflict, but I guess at the same time, this user figured they'd run it again themselves, poaching the assignment.

I guess it seemed wrong to me, but then I'm doing triple-checks of self-verified stuff so this one just means another one to add to my queue which kind of made it more egregious.

Any thoughts?

Speaking of, I'd still like to convince George to disallow handing out assignments for double-checking an LL test that the same user ran for the first check. It wouldn't have stopped a poach like this, but I think some people will try to manually reserve the same work again and at least do it the "correct" way, by getting it assigned to them.

But I'm biased... I've seen my fair share now of certain users who are prolific at verifying themselves and I kind of wish they'd verify other people's results, and let other people verify theirs. Scratching each others backs, so to speak.

[Prime95]

Originally Posted by Madpoo

Speaking of, I'd still like to convince George to disallow handing out assignments for double-checking an LL test that the same user ran for the first check.

It already does.

[Prime95]

Originally Posted by Madpoo

I'm doing triple-checks of self-verified stuff so this one just means another one to add to my queue which kind of made it more egregious.

I'd wait a while to see if the now expired DC assignment on 6-1-15 reports a result.