Question about false positive process

Madpoo · 2014-11-07, 05:28

Here's a potentially weird question about false positives.

In my work on updating the exponent report page, I noticed that currently it doesn't show "bad" results in the LL section. So I'm figuring out whether to show those or not... after all, it might be interesting to see that, at some point, someone checked in a result with a residue that didn't match the double and triple checks.

While working on that, I happened up a few cases of false positives for being prime. Case in point:
http://www.mersenne.org/M52222747

Now, I see in that case, someone did another test later that day (checked in 3 hours later) that showed it composite.

But no triple-check was ever run that I could tell.

There are certain "safeguards" in place to prevent displaying results in the LL section if the residue is zero, even if that's a suspect or "bad" result, but it does show up in the history section as "Prime!"

It's mainly to make sure there's no premature disclosure of new prime discoveries before the double-check and official announcement. I'll be going through and trying to find other ways that info could "leak out" and trying to close those... sorry, I know that reading between the lines is a popular sport when a suspected prime is thought to be out there, and in the past when people did manage to find out they were decent enough to keep it quiet until the official word, but we can't be too careful. :)

Anyway, back to false positives, would it be good practice to do a triple-check right away to confirm the "composite" result? I generated a list of a handful of cases and I'm doing triple-checks on them now just to dispel any doubts; they'll be done in a couple more days. I'm 99.99% sure none of them are prime because the original result had some errors in the run and the double-checks didn't. So I don't think we any hidden gems in there. :)

TheMawn · 2014-11-07, 12:48

I feel like questionable results ought to have high priority, whether it's a first-time with error(s) or a failed DC.

Madpoo · 2014-11-07, 21:40

Quote:

Originally Posted by TheMawn

I feel like questionable results ought to have high priority, whether it's a first-time with error(s) or a failed DC.

I might have to figure out some good way of showing which exponents would be good candidates for triple-checks.

The good (or bad?) news is that none of the triple-checks I've completed so far for exponents that had reported prime are all composite after all.

There was one that I didn't want to mention that got checked in last year (July 2013) that said it was prime, and hadn't even had a double-check yet. I was under the impression it had been checked by someone else but I didn't see any records of it. It's composite though. :) But of course now it'll need a triple-check to verify my result.

http://www.mersenne.org/M60404837

Here's one I'm triple-checking now and should have a good triple-check in another couple days (45 hours to be exact):
http://www.mersenne.org/M53878507

In general, there aren't that many LL tests needing a triple-check, and if it does need one, then it'll get assigned using the normal double-check process. I do think that false positives for being prime deserve a good triple-check right away though. I imagine they probably are most of the time. I looked back here on the forum for other examples when false positives came to light, and it looked like George had a few folks checking them out simultaneously on multiple systems.

It might just be that the ones I found had their Prime95 verification checked in, but whoever did another verification CudaLucas or something didn't get that result checked in. Which would be a shame of course, but I could see that happening.

By the way, sometimes a residue is all zeros *except* for the last digit, like maybe a 2 or a 4 or whatever . Since the last 2 digits are masked unless it's verified, I think some people have spotted those in the results and thought that might be the hidden prime.

So, for the record, yes, it does happen, and maybe it's a bug similar to what causes false positives in the first place, but there are over a dozen where the residue is all zeros except the last 1 or 2 digits. :) Enjoy!

TheMawn · 2014-11-07, 23:19

If you complete a 55M exponent with a non-zero error code, it will take several years before you find out if the result actually had a problem.

If the potential error is immediately handed back out (as a first time check) with high priority, you get a confirmation of yea-or-nay within a couple of months.

This is the reason I like to run a mix of DC and LL. I want to keep the DC train rolling, but I want to put my fast hardware to work on the harder assignments too. The added assurance of a constant flow of correct DC residues gives me more confidence in my LL results.

2014-11-07, 05:28	#1
Madpoo Serpentine Vermin Jar Jul 2014 3385₁₀ Posts	Question about false positive process Here's a potentially weird question about false positives. In my work on updating the exponent report page, I noticed that currently it doesn't show "bad" results in the LL section. So I'm figuring out whether to show those or not... after all, it might be interesting to see that, at some point, someone checked in a result with a residue that didn't match the double and triple checks. While working on that, I happened up a few cases of false positives for being prime. Case in point: http://www.mersenne.org/M52222747 Now, I see in that case, someone did another test later that day (checked in 3 hours later) that showed it composite. But no triple-check was ever run that I could tell. There are certain "safeguards" in place to prevent displaying results in the LL section if the residue is zero, even if that's a suspect or "bad" result, but it does show up in the history section as "Prime!" It's mainly to make sure there's no premature disclosure of new prime discoveries before the double-check and official announcement. I'll be going through and trying to find other ways that info could "leak out" and trying to close those... sorry, I know that reading between the lines is a popular sport when a suspected prime is thought to be out there, and in the past when people did manage to find out they were decent enough to keep it quiet until the official word, but we can't be too careful. :) Anyway, back to false positives, would it be good practice to do a triple-check right away to confirm the "composite" result? I generated a list of a handful of cases and I'm doing triple-checks on them now just to dispel any doubts; they'll be done in a couple more days. I'm 99.99% sure none of them are prime because the original result had some errors in the run and the double-checks didn't. So I don't think we any hidden gems in there. :)

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Fun with a false positive	Madpoo	Data	12	2016-06-29 19:00
another false positive?	ixfd64	Data	3	2016-03-14 22:11
One slow worker process	aweyhau	Hardware	5	2016-02-27 18:57
positive LL test?	sixblueboxes	PrimeNet	90	2014-07-24 05:51
False positive?	Pi Rho	Lounge	4	2003-04-23 14:11

2014-11-07, 12:48	#2
TheMawn May 2013 East. Always East. 6BF₁₆ Posts	I feel like questionable results ought to have high priority, whether it's a first-time with error(s) or a failed DC.

2014-11-07, 23:19	#4
TheMawn May 2013 East. Always East. 3277₈ Posts	If you complete a 55M exponent with a non-zero error code, it will take several years before you find out if the result actually had a problem. If the potential error is immediately handed back out (as a first time check) with high priority, you get a confirmation of yea-or-nay within a couple of months. This is the reason I like to run a mix of DC and LL. I want to keep the DC train rolling, but I want to put my fast hardware to work on the harder assignments too. The added assurance of a constant flow of correct DC residues gives me more confidence in my LL results.