To dispel any doubts that something was wrong with my personal installation (even though I didn't have any mods), I completely deleted BG1:EE (including saves and baldur.ini) and re-installed it.
With the same test (without editing intmod.2da) I got yet again similar results (Imoen failed 10 scrolls out of 67) with 21 INT. For fun I thought of repeating my test with the edited intmod.2da (I set it to 110 at 25 INT this time) and made Imoen drink a second potion of genius to raise her to 25 INT. I got 2 failures out of 67 attempts (with what I presume to be an 110% chance of success!).
EDIT: And yes, @Mathsorcerer, if you could do some testing that would be very useful, I would like to confirm that I'm not the only one with abnormal results.
When you performed your Imoen testing (non-specialist), Were you ONLY testing with scrolls at or below her casting level?
In the interest of honesty, I did have one Cloudkill scroll and I was too lazy to re-edit the save file to get to lvl9.
Looks like all the spells were within the level limit except 1. That makes his results even a bit more consistent with what we should expect from the theory. (In any case he had a bit of bad luck, but well within the limits of "normal" bad luck).
I've found a problem similar to the one described in this thread when playing a bard. Since bards don't have schools of magic, the explanation provided here doesn't seem to apply in this case. I link to the thread in case someone is interested: Do bards have a school of magic?
Let's say your specialist has Intelligence 18. That's an 85 percent base chance to scribe. You attempt to scribe a spell that is outside your specialty. Your chance goes down to 70 percent. The spell is of a higher level than you can currently cast. Your chance goes down to 55 percent. That's barely higher than a 50-50 chance to scribe that spell, with Intelligence 18.
According to the manual, "For every spell level higher than the wizard can comprehend, the wizard receives a –10% penalty. For example, a wizard who can cast 3rd-level spells trying to use a 5th-level spell scroll has a –20% penalty." The chance couldn't be 55%, it would be 60%, or 50%, or whatever the penalty yields.
I decided to do a bit of my own testing on this topic. All tests were done on core rules with 1st level mages (non-specialists) learning a full inventory of 1st level spells (16 spells), repeated 10 times.
New PC in Candlekeep with an INT of 22 learning 160 spells - failed 2 out of 160 times, or 1.25% New PC in Candlekeep with an INT of 18 learning 160 spells - failed 22 out of 160 times, or 13.75% These numbers look good. In fact the DM was a little generous here.
Immy dualed to 1st level mage with an INT of 17 learning 160 spells - failed 67 out of 160 times, or 41.875%! This is significantly above the expected 25% failure.
Thank you. That looks interesting. I'd say your sample data is very representative. This seems to confirm the general pattern observed so far: learning chances work fine for the more standard classes, but they start to misbehave with less standard classes, like dual classes, wild mages or bards.
Just in case, did you double check that all of Imoen's spells were level 1 as well?
Just in case, did you double check that all of Imoen's spells were level 1 as well?
I EE'd them in and used the same spells with the new PC as I did with Imoen. I kept track of how many times each spell failed. Below are the spells used for the tests with the number of Immy's failures in parentheses - Armor (6), Blindness (6), Burning Hands (4), Charm Person (6), Chill Touch (3), Chromatic Orb (6), Color Spray (2), Find Familiar (1), Friends (6), Grease (6), Identify (7), Infravision (4), Reflected Image (3), Magic Missile (3), Protection from Evil (1), Protection from Petrification (3).
So, yeah, as I said, great analysis and it confirms that something funny is happening with dual classing as well. I've added a link to this thread in the bug report, which is located here.
If you feel like sparing a minute, you might add your thought to that report. It would help the developers become aware of the number of people interested in this issue.
There is a problem with the rnd() function or similar on any computer. It does not generate random numbers. Instead it produces a finite sequence of numbers which perform well when tests of randomness are applied. The finite bit is important because after a certain point you start going around in a loop.
NOTE: I am not claiming that this causes the observed problem but it is something to be aware of, as it might just be relevant.
For the terminally curious (and cats).
In the good old days people used to use (very large) books of random decimal digits. These had the same problem though, if you think about it. If you need a very large number of random values and work through the book you will eventually get back to where you started,
A standard method of producing the sequence involves doing modular arithmetic division with a large prime number divisor. No matter how large the prime is, you only have (prime - 1) possible remainders, so then it is back to the start.
There is a problem with the rnd() function or similar on any computer. It does not generate random numbers. Instead it produces a finite sequence of numbers which perform well when tests of randomness are applied. The finite bit is important because after a certain point you start going around in a loop.
NOTE: I am not claiming that this causes the observed problem but it is something to be aware of, as it might just be relevant.
For the terminally curious (and cats).
In the good old days people used to use (very large) books of random decimal digits. These had the same problem though, if you think about it. If you need a very large number of random values and work through the book you will eventually get back to where you started,
A standard method of producing the sequence involves doing modular arithmetic division with a large prime number divisor. No matter how large the prime is, you only have (prime - 1) possible remainders, so then it is back to the start.
Agreed. However, the issue here is not of a random sequences repeating itself. It's the number of failures occurring not even being remotely indicative of the percentage chance it's supposed to be. If my PC had an INT of 18 and for every 13 spells I attempted to learn I failed on the 4th and 13th attempt, that would be indicative of the correct percentage - even if the 2 failures always occurred on the 4th and 13th attempt. Also, if it was an issue with the rnd() function, every type of spell caster would be affected equally. Although, I guess what you're saying is that with a small enough sample size, extrapolated out to the nth degree you can get numbers that are very 'off'.
Comments
When you performed your Imoen testing (non-specialist), Were you ONLY testing with scrolls at or below her casting level?
Do bards have a school of magic?
All tests were done on core rules with 1st level mages (non-specialists) learning a full inventory of 1st level spells (16 spells), repeated 10 times.
New PC in Candlekeep with an INT of 22 learning 160 spells - failed 2 out of 160 times, or 1.25%
New PC in Candlekeep with an INT of 18 learning 160 spells - failed 22 out of 160 times, or 13.75%
These numbers look good. In fact the DM was a little generous here.
Immy dualed to 1st level mage with an INT of 17 learning 160 spells - failed 67 out of 160 times, or 41.875%! This is significantly above the expected 25% failure.
Just in case, did you double check that all of Imoen's spells were level 1 as well?
Armor (6), Blindness (6), Burning Hands (4), Charm Person (6), Chill Touch (3), Chromatic Orb (6), Color Spray (2), Find Familiar (1), Friends (6), Grease (6), Identify (7), Infravision (4), Reflected Image (3), Magic Missile (3), Protection from Evil (1), Protection from Petrification (3).
If you feel like sparing a minute, you might add your thought to that report. It would help the developers become aware of the number of people interested in this issue.
NOTE: I am not claiming that this causes the observed problem but it is something to be aware of, as it might just be relevant.
For the terminally curious (and cats).
In the good old days people used to use (very large) books of random decimal digits. These had the same problem though, if you think about it. If you need a very large number of random values and work through the book you will eventually get back to where you started,
A standard method of producing the sequence involves doing modular arithmetic division with a large prime number divisor. No matter how large the prime is, you only have (prime - 1) possible remainders, so then it is back to the start.
If my PC had an INT of 18 and for every 13 spells I attempted to learn I failed on the 4th and 13th attempt, that would be indicative of the correct percentage - even if the 2 failures always occurred on the 4th and 13th attempt.
Also, if it was an issue with the rnd() function, every type of spell caster would be affected equally. Although, I guess what you're saying is that with a small enough sample size, extrapolated out to the nth degree you can get numbers that are very 'off'.