[thorvald]

I have read in this forum that the bot never blanks a honor in the end position, but I just experienced that situation now

---

In the end position I counted the opening hand and assumed ♥Q was at the opening hand, but "knowing" the bot would keep Qx and had a safe discard of the last ♠ I did finesse :-)

BTW: I have lately seen more than one situation, where the bot isn't covering the J with Qx, like in this board

---

[johnu]

First hand, you have conflicting signals. First, West passed 1♦ so shouldn't have 6 points. West has already shown ♠A so shouldn't have ♥Q. Second, East knows that blanking the ♥Q will cost a trick because declarer will "always" drop the now singleton queen so East should never blank the queen.

What to do? Who to believe?

GIB opening bids and responses with interference are much more consistent than a lot of it's play, especially in competitive auctions. So that points to playing for the drop (and also, you know West's last card is a diamond).
____________________________________________________________________________________
2nd hand. GIB doesn't cover honor with honor nearly as much as it used to many years ago. IIRC, back when there was active development on GIB there must have been some kind of programming change. In this case, dummy has KJ10 so covering can never gain a trick unless partner has the ace so GIB plays random cards IMO.

[thorvald]

johnu, on 2022-September-03, 15:53, said:

GIB opening bids and responses with interference are much more consistent than a lot of it's play, especially in competitive auctions. So that points to playing for the drop (and also, you know West's last card is a diamond).

Agree - but I was just convinced that the drop could not happen. But we all learn :-)

[smerriman]

There's definitely extremely strange going on here.

At trick 10, at the point you lead the ♣3 from dummy, East knows you have one club left, and thinks there is a 36% chance you have 1 diamond left, and a 64% chance of 2 diamonds left, which makes sense.

The 64% where you have 2 diamonds left, it sees J9, JT, and JT all equally likely, as you'd expect.

The 36% where you have 1 diamond left.. it thinks you have the T 100% of the time, and the J never.

GIB isn't being clever; quite the opposite - once you play the J, it thinks that guarantees you have the last diamond, so plays a random card.

Still trying to figure out why it rules out the actual case.. it's not like it thinks 1NT by South doesn't fit with that hand; it bids it itself..

[smerriman]

Pretty sure I've figured it out, and it's another mammoth fundamental GIB bug. And if I'm understanding correctly, also what's causing some of those ones late in the play where it simulates hundreds of deals and misses the one actual important one.

Stay tuned for part 3 of my 0% plays series (though need to put in a lot more analysis / testing first - if I'm right, this is the biggest bug yet and completely breaks GIB.)

[pilowsky]

smerriman, on 2022-September-03, 19:01, said:

Pretty sure I've figured it out, and it's another mammoth fundamental GIB bug. And if I'm understanding correctly, also what's causing some of those ones late in the play where it simulates hundreds of deals and misses the one actual important one.

Stay tuned for part 3 of my 0% plays series (though need to put in a lot more analysis / testing first - if I'm right, this is the biggest bug yet and completely breaks GIB.)

My breath is completely bated.

[smerriman]

Aw, sorry to disappoint, I messed up It wasn't what I thought, though still something very interesting.

Contrary to what I believed, it appears GIB does take into account leading rules in its simulations. In this case, the normal lead for GIB from T973 is the T; when it leads the 3, it assumes it doesn't have that holding, which means you don't have AJ8. So after the A and 8 are played, you must have two diamonds left, so what it plays at the end is irrelevant.

Still going to follow this up more to see how exploitable it is..

[thorvald]

---

For once the bot did blank a king, so perhaps it will never blank in a more specific situation

[smerriman]

thorvald, on 2022-September-08, 15:59, said:

For once the bot did blank a king, so perhaps it will never blank in a more specific situation

The point is not that GIB never "blanks a king". It's that it assumes the human plays double dummy, so won't - barring outright bugs - make a play that only works if the human misguesses. It'll happily blank a king if double dummy tells it that it makes no difference what it does.