B1G was going to drop suspension to Two Games

What sign stealing can accomplish using film (like the All-22 film, which does not contain any signals, so is useless) and TV broadcasts (which show maybe one or two shots over the course of an entire game of the signals) and on gameday, pales in comparison to what can be accomplished when you record every signal and every corresponding play for 10+ games on a single opponent.

There's the signs that a team uses, which can be changed during the week if it's really necessary.  Then there's the system that is used, which cannot reasonably be changed during the year, because the players need to be able to decode the signals in one or two seconds-- it needs to be almost instinctive.  As an analogy, the system would be the rules of grammar and syntax, while the signals would be the particular words that are used.

If all you know is the signals, then you're not going to gain an advantage when the other team changes their signals.  But if you were able to learn their system, now you've got a simple translation exercise.  You spend the first half recording the signals and corresponding plays, and then, crucially, you plug that data into the system you've deduced through the 10 games' worth of film.  Now you know what all the signals mean in the 2nd half.  As I'm sure you're aware, Michigan was losing to OSU 20-17 at the half, and beating PSU 16-14 at the half, before going on to collectively outscore them 52-6 in the second half.  That could be due to Michigan's greatness, but it's not unreasonable for someone to wonder if they went from not knowing what plays were being called in the first half, to knowing exactly what plays were being called in the 2nd half-- which is the pattern you would expect to see in the scenario I outlined, where you have an accurate model built, and only need to do a correlation exercise on the new signals.

Here's a very oversimplified example of what I'm talking about.  Let's pretend that offense is so simple that you only need to signal in the three pieces of information: the formation, the play, and the check-down play if the QB gets to the line and doesn't like what he sees.  We'll even further oversimplify this and say that there are only 26 plays to choose from.

Now let's assume that a team wants to make it non-trivial for other teams to steal their signals.  They can only make the system so complicated, before it leads to too many mistakes with players getting the calls wrong.  One way to do it would be to have pictures on signs that you can hold up.  There are 3 pieces of information, but every play, you're going to hold up 4 signs; one of them is a dummy, serving no purpose other than to make it confusing for someone trying to decode their signals.  In this system, signals are always read left to right.  To further confuse sign-stealers, when reading the signals, you don't always start at the far left; instead, you start at the position corresponding to the down number, and wrap around to the first spot if you would need to read beyond the 4th spot.  So on 1st down, the dummy signal would be the 4th, because the players would just read 1st=Formation, 2nd=Play, 3rd=CheckDown Play.  But on 3rd down, the 2nd spot would be the dummy, because players would read 3rd=Formation, 4th=Play, and 1st=CheckDown Play.  The final element of the system is that all the players really need to know about the signs being held up is the name of the picture-- for example, is that a picture of a "plane" or an "airplane".  That's their rote memorization, and you can reasonably expect them to memorize all new signals in a week, because for the vast majority of them, they're going to intuitively know the name by looking at the picture.  We stipulated 26 different plays; each one of them corresponds to a letter, A-Z, and that never changes-- it's an inherent part of the system.  So Play A is indicated if an Apple is shown, or if an Airplane is shown, or if Asparagus is shown, etc.  (If you really wanted to get paranoid, you could make sure that like 25% of the pictures use a name that isn't the first or second name that a signal stealer would guess- for example, instead of a P for Plane, or A for Airplane, it's actually U for United, because the team always flies to games on United Airlines.)

You cannot figure that system out using only the sources that teams normally use.  The All-22 Film does NOT show the signals-- it's quite useful for scouting other teams' tendencies, and looking for tells, like if the RB leans further forward on a run play than on a pass play; but it's absolutely useless for signal stealing. The vast majority of the data that you'll have access to is what you collect on game day-- not nearly enough, and not nearly enough time to analyze it.

In contrast, with your database of at least 700 plays (in the case of OSU, where we know of at least 10 games that Stalions bought tickets to), and plenty of time to run statistical analyses (correlation, regression, etc.) or, maybe even better, Machine Learning, you can absolutely expect to figure this entire system out.  The only mildly tricky bit is that a human would need to separate out the first letter of the picture names as its own variable, as statistical methods wouldn't know that the first letter is significant, and even an ML model is quite unlikely to be capable of that with only 700 plays.  But you've got someone who specializes in signal stealing, who probably knows how a lot of systems work; unless the team that uses this hypothetical system has invented a completely new idea, some other team has used the first letter idea, so it's no great stretch of the imagination to believe a human would think to make a separate variable for the first letter.  Everything else is trivial.

That's one way that Michigan could benefit from collecting this data, even when the opposing team changes their signals.  Those other teams are relying on their experience with past signal stealing operations, and believe they just need to change the pictures themselves; but in this case, their entire system is busted, so they would have needed to change how the pictures are decoded, as well.

I obviously have no way of knowing what Stalions did with all the data he collected, but I will say that it's absolutely the first thing I would think of doing.  I'd let several instances of Machine Learning algorithms rip, while I'd use statistical analysis software to try to figure out what I can on my own.  Between the two, with enough data, I should be able t break any system that's simple enough for college players to be using.  On gameday, I know which of the 4 pictures has the formation, which one has the play, and which one has the CheckDown play.  I know that what's important is the first letter of the name of the picture.  30 plays should be sufficient for me to figure out all of the common signals, and at least a few of the less common ones.  No other team is going to be able to do the same thing.