The answer
On film the trainers answer this question with one idea, and they call it prediction error. A dog acts on what it expects to happen next. When the expected result does not arrive, the dog is surprised, and the surprise is what changes the behaviour. Connor gives the owner-lesson version on Stormie: a dog who has learned that his owner says stop and keeps walking finds, on his first walk with the trainer, that pulling now costs something. In his words, that's weird, that's not what was supposed to happen. Dutchess's trainer gives the same lesson to his clients and tells them it is something he goes over with every single person. The trainer who handles Dolce calls it the way all learning happens.
The new dogs in this edition change what the idea mostly means. Most of the filmed uses are not a correction at all. They are a no-show: the thing the dog's behaviour has always produced simply does not happen. Samson's trainer, Kelton, wore bite gear so that when Samson went to do the thing that had always worked, nothing happened. The trainer did not run and did not fight. He stood there and asked the dog what was going on. Reaper's handlers did no-shows in the kennel. With Kali, Kelton picked up the leash or called her in and then asked for nothing. With Bo, the trainer at the crate did not leave. On Louie's owner walk the trainer let the dog's shriek fail to end the handling. When Jack startled on a trail and nothing bad followed, the trainer named that as one too. On Sage the surprise is a pleasant one: she is called over and nothing hard follows, so coming when called stops predicting work. Reaper's later public walk (September 2026) is the plainest case of the other direction: lunging at dogs and people had always produced little more than 'good boy, come on, let's go', so when he fixated and sprinted at a dog the trainer punished it, and named the different result as the prediction error, 'where all the learning happens'. Walking and looking stay unpunished, so only the one act changes its result.
Three things are said about how to use it, and all of them are on film. First, the event has to end before the next one starts. Samson's trainer calls a return to baseline a super important part of the philosophy, and Finn's trainer caught himself going back too fast. Second, it is repeated: the dog tries the old move again, it fails again, and after enough failures he stops trying it. Third, it works on both halves of training. Connor pairs it with conditioning, which is the word or cue coming to predict what follows it. Bo's handover and Stormie's go-home both teach that half. Kelton's work with Piglet and Kali shows the other side of it. There a dog who predicts too well starts avoiding or offering the next command early, and the handler has to break that prediction on purpose.
What is inferred rather than said: the trainers never measure a dog's expectation. The expectation is their account of the dog, and it is given with confidence. The evidence on film is what the dog does next - a shake-off, coming back to the thing that startled him, sniffing instead of freezing, a behaviour that stops being tried. Whether the house draws a line between a prediction error the handler delivers and one the handler allows is not said anywhere. When a bite gets nothing and when it gets punishment and then nothing was settled by Kelton on 23 Sep 2026: it is totally context- and case-dependent, and different approaches accomplish different goals and outcomes. The first question is laid out below.
A wider reading of every session note that tags the idea finds fifteen more dogs carrying it on film. Two things change. First, the owner-lesson definition is standard: Baxter's, Astro's and Rocky's trainers draw the same sequence (a behaviour, an optional marker, then the prediction error or "wait what moment"), Titan's lesson repeats Connor's two-ways-dogs-learn framing, and Baxter's trainer says that without one no learning ever occurs. Second, the correction direction is commoner than the no-show count above suggested. Milo, Thor, Rocky, Lucy and Anna are each told about an unexpected consequence (pressure that does not stop, a consequence for breaking a stay or running off), while Astro's bite that gets nothing is the no-show again. The spacing rule recurs as a stated rule: Akira's, Baxter's and Max's trainers each make the next trial wait until the dog is back to baseline. Marvin adds a third use - a dog who predicts the wrong thing (a stim on a recall) is given reassurance, not pressure - and Roofus and Teddy add the Piglet side, where a dog who over-predicts the handler's patterns gets pattern breaks.
Angier K9 is a family-owned, play-based Board & Train serving Angier, Raleigh and the Triangle, with three full owner lessons.
What to do
Find out what your dog's problem behaviour reliably gets him, then stop letting it get that. The trainers' examples: do not apologise and back off when he protests, because if you react and let him out of it, next time he'll do it again (Louie). Do not run or fight when he snaps; the result has to be neutral (Samson). Do not follow every recall with hard work; sometimes call him over and just chill together (Sage). A word you use must be followed by what it promises every time, until the word alone predicts it (Stormie's go-home, Bo's 'nope'). After any correction or surprise, let him get back to normal before you ask again. Expect him to try the old move a few more times before he gives it up. Your trainer decides what the dog meets when the old move fails. A bite or a serious threat is the trainers' work, done in protective gear, not yours.
What the trainers say
so that's the other way is when a dog predicts that something's gonna happen and that goes oh i was wrong then they adjust their behavior accordingly
um when they first come to us they're like oh when i pull my owner's like stop and just keeps pulling but does i don't have a consequence so i can pull as much as i want there's no problem they come with me and they pull and they feel the e-collar and they go that's weird that's not what was supposed to happen prediction error
yeah you get the two in tandem with each other where she's now realized that her prediction was wrong she adjusts accordingly and then on top of that she's been conditioned every time we go by by the dog she's already ready for me to say it but it looks really like it when you watch it you're just like oh she doesn't care about the dog but like psychologically there's like a whole entire thought process going on there
so for instance duchess when she first came to me right i'm working with her she pulls on the leash and she thinks yeah i can pull nothing's gonna happen and then all of a sudden she feels some kind of pressure she's like oh that was different
different and when you give a prediction error correctly you see the dog kind of do this like shaking motion right and that's like the body language is i'm learning something like i'm processing so that's like half how they learn the other half is conditioning
all i'm doing is doing a little bit of like the way all learning happens it's just uh she's predicting that if she feels like that little bit of leash tension she needs to come off of it which is correct in most cases but she's taking that as a relief signal and I'm giving her that a prediction error there where it's like she's getting an unexpected consequence which is that I'm telling her that she's wrong by saying AH AH and so then from there it becomes relatively simple I give her the real signal which is the okay command
return to baseline that's a super super super important part of our dog training philosophy and what it is that we do because all learning happens through prediction error and when you have a prediction error on the first trial you need um to allow for a return to baseline in order for any learning to occur if the event just continues and never goes away then they don't have a chance to learn from the event because the event's ongoing it never goes away in their mind
And he just needed to learn. He needed to have some prediction errors where essentially Eventually all that happens is he goes to do the thing that has been successful for him in the past and nothing happens. The result is neutral. I just stand there and do nothing and just kind of go like, hey man, are you okay?
and so it is with dogs so it is with dogs you can cause humongous behavioral changes by violating their predictions when they think that one thing is going to happen as an example they turn to violence if you are wearing safety gear which is why you do bite gear evaluations and why you do work with dogs wearing bite gear if you then respond with with kindness. You respond with love. Or you respond with complete neutrality. It results in confusion.
going to fight them right you know what they don't expect nothing they don't expect neutrality they don't expect you to go oh hey buddy like what's going on man why are you why are you doing that why are you upset confusion and then that leads them to rethink their actions the next time and the next time after that and the next time after that
But he's still got that thing in his brain. I think it is genetic. The only thing that can be done about it is to punish it. If he goes to do things like that. But, and that's what we've done. And then, well, that and then do prediction errors where like he goes to do the thing and it doesn't work.
and myself and Christian and Connor we've all been working with him where we've just been not allowing him to bully or display you know like serious aggression we've been penalizing him when he does it and doing like no-shows like prediction errors so he should be to the point now where I can take him out relatively safely from here we can start doing all the obedience work
false starts there because you can see like this little anticipation of like okay so you're gonna like ask me to do a command and put some pressure and then i just completely no-showed that and gave her scratchies and said go right
if i'm if it's anything where it's a change yeah and then that change is followed like it like then like a chain starts yeah that ends in some kind of pressure yeah they predict it they predict predict it and then they avoid it and then it gets deeper and deeper into that cycle and you have a problem
yeah i'll just do a prediction error see what he does hey bud
it's not as severe as yesterday yeah let's look into the prediction area we just let them return to normal in a second yeah yeah you can go ahead
that we call it a prediction error in his mind he goes he predicts that I'm going to react a certain way because maybe 10 people did it that way yeah right i didn't react that way so all of a sudden there's an error in my prediction or his prediction does not come and he goes that didn't work why didn't that work
Again he thinks it works and it does Right like he gets to get knocked out every single time right
Like, if he did, I would just let him bite me. Do nothing about it. No correction of any kind, no punishment. It's a prediction error, straight up.
I freaked him out, but he came right back to it, which is a learning experience, right? Prediction error. He predicted danger, and no danger was forthcoming.
So I have to allow enough freedom for prediction errors to occur for management to be out of the way for her to start feeling differently about what's going on.
so if every time I say come she comes to me I go down or I go sit or I go whatever it is and then I have to put some pressure on her to get it then next next time I say come if I do that 15 times in a row and I say come she's gonna be like I don't know if I want to it's gonna suck if I go over there gonna make me do stuff right so then you have to cause a prediction error and just call them over
Uh although Christian's already kind of worked with it a little bit and done a couple of prediction errors, um I still want to see it and um try to allow it to happen because if we can get it, then we can we can allow the prediction error to happen and he can learn from it.
um she's just anticipating the down and then go ahead and uh pre-doing that like a suppression loop so i don't want that so i'm trying to break her out of that by doing a sickness match for her anticipation so she's getting a prediction error there she's expecting down and then i'm giving her instead to sit and so then she has to think about that as well
and um like anticipation like that is just always based on prediction so the only way to get rid of the anticipation is to remove the antecedent right in the arrangement um make it not predict that every time right so maybe maybe i made that predicted too often
Or, alternatively, what actually happened is a prediction error. He thought, okay, I can relieve the pressure by laying down and doing nothing. And then when that didn't work, okay, I'm going to try a different hypothesis.
I went right back to it really fast which actually don't like become too distinct coherent like coherently distinct like events they're not two different events in in Finn's brain right right, instead of being two coherently distinct events, it's just one prolonged event of negative reinforcement, which is no good
to find out what actually happens and get a different result than they thought. That's where all the learning happens because that's what you would call a prediction error, right?
Because he predicted what's always happened in the past, which is like, good boy, come on, let's go, which is really like not that much.
He can do that all day and he can lunge and bark and everything like that and no serious consequence happens.
prediction error it's the wait what moment so that happens when they expect one thing and something else happens. The pattern's violated. If the pattern's never violated without prediction error of some kind, no learning ever occurs.
so we skip that step we go straight to prediction error prediction error is where all the learning happens you can think of it as the wait what moment
and then you have a consequence of some kind. And that consequence can either be surprising which is what we call a prediction error or it can be a confirmation that they already know and that's prediction success
second point which is there's two like main ways dogs learn the first one we already talked about it's conditioning and the second one is through prediction error
success for me would be that she ran off she started running off and you recalled her and did a big punishment and she came back because because what I want is learning. Learning happens through prediction error. She thinks that the same thing's gonna happen, and something different happens.
so it's called a prediction error you're predicting one thing and and something else happens and when you do that you have a period of time where you return like to baseline where you're neutral nothing's happening and then there's a second trial
How it looks, dog by dog
Stormie
The owner-curriculum version, the easiest place to hand the idea to somebody else. Connor names it as the second of two ways a dog learns, and gives worked examples in one breath. The first is the puller who has never paid for pulling. The second is a dog who has bitten people and watched everyone run away in terror, then bites a sleeve and nothing happens. The third is Stormie herself at another dog. Then he joins the two mechanisms together: say the cue at every single instance of the trigger until she predicts the cue rather than the dog. What an onlooker then sees is a dog who does not care, though psychologically there is a whole thought process going on. At go-home the positive half is stated as the opening principle: charge the word up, because a word you do not understand is noise.
On film
or on a more extreme case let's say a dog's bit a couple people and then like everybody runs away in terror and then they come to us and they bite the sleeve nothing happens and then all of a sudden they're shook
go and then she immediately unfixates and follows along and it's the same thing with the dogs if i say let's go every single time we walk by a dog her brain starts to go well every time we walk by a dog he says let's go i'm going to predict it this time and so it just looks like a seamless i don't care about the dog okay right
she just the thing about it is you have to charge the word up um you know like if you didn't know what a word meant in another language you'd be like whatever
Dutchess
The same lesson from a different trainer. That makes it house curriculum rather than one person's framing. He says he is not teaching the psychology because the family has a dangerous dog; he goes over it with every single person. His version has two things nobody else's has. He works the example on this dog's own history: she pulled, thought nothing would happen, felt pressure, and thought, that was different. He also puts a share on it, half of how they learn, with conditioning as the other half.
On film
bit like that but um more so i'm just going over the psychology not because i'm like you have this dangerous dog i go over with every single person
different and when you give a prediction error correctly you see the dog kind of do this like shaking motion right and that's like the body language is i'm learning something like i'm processing
Dolce
The finest-grained use, on a very advanced dog. She has learned, correctly in most cases, that a little leash tension means come off it, and she has started reading the tension as her release. The trainer runs the fix in three beats and narrates them as they happen. She predicts, the marker AH AH tells her she is wrong, and then he gives the real signal, the release word okay. It is the only case in the set where what gets corrected is the dog's reading of a signal rather than a behaviour. It is also where the idea is claimed as the way all learning happens.
On film
all i'm doing is doing a little bit of like the way all learning happens it's just uh she's predicting that if she feels like that little bit of leash tension she needs to come off of it
Samson
Samson is a serious dog: the trainer tells the owner he still needs managing and that random people should not play around with him. Kelton explains his change as punishment plus prediction errors. The bite-gear work itself was not filmed; Kelton narrates it twice, three months apart, once to camera and once to the owner. In the gear, when Samson went to do the thing that had always worked, the result was neutral. Kelton did not flee and did not fight. He stood there and asked the dog what was wrong. The dog was confused, rethought, and very quickly started trying other things, sniffing and offering play. In the same walk Kelton says all learning happens through prediction error. He ties that to the house rule on timing: after a trial, allow minutes of return to baseline, or the event never ends in the dog's mind and nothing is learned.
On film
And he just needed to learn. He needed to have some prediction errors where essentially Eventually all that happens is he goes to do the thing that has been successful for him in the past and nothing happens.
and very quickly you get a dog who goes let me try something else let me see what he smells like they start using their nose
oh, hmm, that didn't work like I thought it would, right? So we were able to do both of those things with him and that changed his behavior and it helped a lot.
of time to return to neutral right so like if I did another recall and try to make him sit again right this moment there wouldn't be enough time for him to return to baseline and learn from what we just did
Reaper
The same no-show, run as a team in the kennel before the dog was ever taken out. The lead trainer, Christian, Connor and a kennel tech penalised the bullying and serious aggression and did 'no-shows like prediction errors'. Only once that was done was the dog judged safe to take out and start obedience. The pick-up day was staged for the same purpose. The family was asked to help, with staff in bite gear, and the reason given is that most people have never given the dog a chance to try. Months into the program, on a public walk (title-dated 25 Sep 2026), the trainer uses the term for a delivered correction rather than a no-show: Reaper fixated and sprinted at a dog, was punished, and the trainer explains that in the past lunging and barking had cost him nothing, so a different result is the error and the learning. He keeps the boundary narrow: walking and seeing are never punished, only the one act.
On film
and myself and Christian and Connor we've all been working with him where we've just been not allowing him to bully or display you know like serious aggression we've been penalizing him when he does it and doing like no-shows like prediction errors
Because most dogs, like you said, most people don't do this. So they've never given the dog a chance to try.
he fixated on the dog and then he decided i'm gonna attack that dog i'm gonna sprint after that dog so i punished it
Because he predicted what's always happened in the past, which is like, good boy, come on, let's go, which is really like not that much.
that's why we don't punish walking we don't punish seeing them that's all perfectly okay right we punish that one thing right so then the boundary becomes clear
Bo
The crate case, and the only place the phrase is used as a verb said before the act. Kelton says he'll just do a prediction error and see what Bo does, and then has Christian and Connor keep a hand at the open crate. Nobody leaves and nobody escalates. The growl is marked not with authority but with a disappointed dude-why-are-you-doing-that. It is treated as a bounded event: once it is done, the dog is let back to normal before the next go. At the handover the other half is taught. 'Nope' is the word that predicts something bad will happen next if he does not listen.
On film
yeah i'll just do a prediction error see what he does hey bud
there it is and then just do like a like uh instead of um like an authoritative thing just do like a like just kind of disappointed like dude why are you doing that
something else and he doesn't recall cuz of that then I'm gonna use the word that predicts that something bad will happen next if he doesn't listen right so that word is nope n-o-p-e
Kali
Prediction as the problem. Kali has learned the first link of a chain that ends in pressure - the leash being picked up, being called in - and starts avoiding at that link. Kelton's diagnosis is general: any change that starts a chain ending in pressure gets predicted and then avoided, and the cycle deepens into a problem. The fix is the no-show again, on the first link. He picks up the leash and nothing bad happens; they go back to scratches. He calls her in and asks for nothing, then gives her the release. He calls these false anticipations and false starts. He says to a colleague that it feels like wasting time but is faster.
On film
All right, so, what's the thing, the first thing in the chain? Right? That's the thing that means something.
Right, so I wanted to think, oh, he picked up the leash, nothing bad happens we go back to the scratchies
i'm going to do a couple of like false anticipations for her right so i'm going to call her over and then ask her to be and then i'm not going to ask her to do anything
Piglet
The same problem on a dog who is, in Kelton's words, a proficient and skilled anticipator. She pre-offers the down before she is asked, which he calls a suppression loop. He breaks it by giving the sit where she expects the down, so she gets a prediction error and has to listen to the word. The criterion is plain: hear it and then do it, not anticipate it. On another walk he turns the diagnosis on himself. Anticipation is always based on prediction, so maybe he made his settled gaze predict a command too often, and the fix is to change what that antecedent predicts.
On film
skilled um anticipator right she anticipates and i don't want her working off of anticipation every time i want her working off of off of the verbal so i want her to hear it and then do it not anticipate it
and um like anticipation like that is just always based on prediction so the only way to get rid of the anticipation is to remove the antecedent
Louie
The owner-walk version, told to a family about their own dog's protest. When Louie screams to get out of something, the trainer says most owners react and apologise and let him off, so next time he does it again. The trainer does not react. The dog predicted a reaction because maybe ten people gave it, and it did not come. He tries it again, and after enough failures he stops. The same trainer then turns it on the vet. Sedation at every visit means he thinks the screaming works, and it does. He offers to go to the vet and try prediction error tests there.
On film
cases it's the dog going man i don't want to do that and then if i react and go oh i'm so sorry and let him do it next time he'll do it again
then maybe i could do that yep didn't work the last three times and probably after you know the 10th time he's like man didn't work yeah and he's like so why do it right
No, and I would even be willing to go to the vet with him and talk to the vet and see if they would be willing to do some prediction error tests with me.
Jack
The fearful-dog direction, with no aversive anywhere in it. The only correction he allows is for chewing the leash. On the trust walks the mechanism is stated as the whole plan: some prediction errors in his head, where he sees that nothing bad happens and good things do. On a trail the trainer startles him by accident, Jack comes right back, and the trainer names it on the spot: he predicted danger, and no danger was forthcoming.
On film
Like, if he did, I would just let him bite me. Do nothing about it. No correction of any kind, no punishment. It's a prediction error, straight up.
he's very stressy, very anxious, so we need him to act differently. So there's some prediction errors in his head. He sees nothing bad happens, that good things happen, that he can explore the world, he can feel safe.
I freaked him out, but he came right back to it, which is a learning experience, right? Prediction error. He predicted danger, and no danger was forthcoming.
Sage
The restraint side and the pleasant side. Around her triggers in a Lowe's the management is kept loose on purpose: boundaries on her choices, but not so tight she cannot learn to cope. Restrict her too tightly and she will panic the moment she is not restricted; he has to allow enough freedom for prediction errors to occur. What he waits for is the light switching on, where she starts to process information instead of freezing. And at the owner lesson Kelton repairs her wariness of the trainers the other way round. If every come is followed by hard work, the dog decides coming will suck. So you cause a prediction error: call her over and just chill together.
On film
If I restrict her options too tightly, she'll panic if she's not restricted. That kind of tight management doesn't allow for any kind of learning.
So I just want to see that light kind of switch on where she starts to process information and stop it freezing off.
so if every time I say come she comes to me I go down or I go sit or I go whatever it is and then I have to put some pressure on her to get it
Louie
The file that shows the term is everyday house language. A trainer opens a home session by reporting that Christian has already done a couple of prediction errors on Louie's reactivity. He then says he wants to see the reactivity himself and allow it to happen, so that the prediction error can happen and the dog can learn from it. Nothing in Louie's file shows or describes what Christian did. It is still the only named technique attached to his presenting problem.
On film
Uh although Christian's already kind of worked with it a little bit and done a couple of prediction errors, um I still want to see it and um try to allow it to happen because if we can get it, then we can we can allow the prediction error to happen and he can learn from it.
Finn
The theory stated in full, on a frightened dog. You ask for a behaviour, the dog tries something, and the result is either a prediction success, where his hypothesis about relieving the pressure worked, or a prediction error. Finn tried freezing, then hiding in a thicket, then running past in a panic. Three prediction errors in a row, and then the thing that worked was coming to the handler. The same session holds the timing rule the Samson walk states: the trainer went back too fast, so two events became one prolonged event of pressure. The next rep waits until the dog has sniffed, jogged and got back to a relaxed state.
On film
The full learning sequence is that you get a behavior or you ask for a behavior, they try something, and it leads to a particular result. Whether it be prediction success, like he came back and And that was his hypothesis, his theory, about how to, like, relieve the pressure, and it worked.
So all three of those, like prediction error, prediction error, prediction error, all three of the techniques he tried, none of them worked. Freezing didn't work. Running and hiding didn't work. Running by me in a panic didn't work.
Akira
On day two the trainer ties the term to spacing: long gaps between repetitions so she returns to baseline, then the next trial, and the learning happens during the prediction errors, while she tries her candidate ways of making the pressure stop (refusing, whining, biting, jumping, rolling over). He wants her to find the answer herself, which takes patience.
On film
they return to baseline. They return to neutral. That's when I can do a second trial for the learning
And that's where the learning happens, right? During the prediction errors where they think, hey, the pressure will stop if A, B, C, D, E, right?
Anna
A one-line version on leash pulling: every time she hits the end of the leash it gets no good result, and the trainer calls that the basis of how you train every dog - what you thought was going to happen did not happen.
On film
she learns that she can't be doing that and hitting the lease does not result in a positive outcome which of course is the basis on how you train every dog anyways right it's basically oh what you thought was going to happen didn't actually happen
Astro
The session titled for the idea: the trainer narrates the learning sequence at the kennel - behaviour, an optional marker step skipped because Astro will not interact, then straight to prediction error, the wait-what moment. He names his own first rep as a prediction success (he drew his hand away and the dog got what he wanted). At the April update he explains why Astro stopped biting him: he did not run or pull his hand away, so the bite did not work.
On film
so we skip that step we go straight to prediction error prediction error is where all the learning happens you can think of it as the wait what moment
so you saw that first time prediction success right I made a mistake I drew my hand away he stopped me from doing the things that I was doing
doesn't think very much about biting me anymore yeah uh no um well because it didn't work right nothing happened i didn't like run away i didn't whip my hand away he didn't get what he wanted it was a prediction error
Baxter
The private-lesson version for an owner, drawn as a chart: behaviour, an optional marker, then the prediction error, the wait-what moment, without which no learning ever occurs. In the second part he applies the spacing rule to Baxter himself: the first trial was confusing and stressful, so he may wait ten minutes before another.
On film
prediction error it's the wait what moment so that happens when they expect one thing and something else happens. The pattern's violated. If the pattern's never violated without prediction error of some kind, no learning ever occurs.
in the early learning stage, I'm gonna wait, like, I might not do anything else for 10 minutes, right, because if you remember that little chart we had, the prediction error
Lucy
At the end of the in-person evaluation Kelton tells the owner what success at home would look like: she runs off, is recalled and given a big punishment, and comes back - because learning happens through prediction error. It is the correction direction, set up on purpose.
On film
success for me would be that she ran off she started running off and you recalled her and did a big punishment and she came back because because what I want is learning. Learning happens through prediction error. She thinks that the same thing's gonna happen, and something different happens.
Lunna
At Lake Benson the trainer explains to the owner why the dog changed so quickly: dogs learn fast in terms of predicting, and learning at its core is thinking something will happen and it not happening.
On film
so dogs in terms of predicting actually learn really fast most of the time you just have to really show them like hey you thought this was gonna happen and it didn't
like what does learning stem from at its core right learning is anytime you think something is gonna happen and it doesn't happen
Marvin (Giant Schnauzer)
A third direction at the go-home: a dog who predicts the wrong thing. If Marvin hesitates on a recall because he thinks a stim is coming, Kelton says he is predicting the wrong thing, so he sharpens the prediction with verbal encouragement instead of pressure, because what the dog needs there is reassurance.
On film
thinking oh they recalled maybe i'm gonna get stemmed right so then it's like okay well he's predicting the wrong thing because it doesn't mean that right so i need to sharpen that prediction
right over here right instead of putting pressure because he doesn't need pressure there he needs reassurance
Max (Goldendoodle)
At the go-home Kelton gives the model with a phone-call analogy, then states the sequence: a prediction error, a return to baseline where nothing is happening, then a second trial, which is where the learning shows. He applies it to a dog who has pulled on the leash hundreds of times.
On film
so it's called a prediction error you're predicting one thing and and something else happens and when you do that you have a period of time where you return like to baseline where you're neutral nothing's happening and then there's a second trial
Milo
Narrated on a recall off a spot: Milo thinks he is safe if he stays put and everyone will give up; instead there is an aversive consequence, and when he leaves and comes in there is a rewarding one. The trainer calls it the way all learning works.
On film
so the way that it works is all learning works with a prediction error he thinks he'll be safe if he stays there nothing will happen we'll just give up and then that's not what happens instead there's an aversive consequence
Penny
The Dolce move on a release word: Penny is breaking place on movement rather than on okay, so Kelton deliberately gives her what she might take for the signal, lets the error cost her, and sends her back, so that only okay releases her.
On film
so she needs to have a prediction error first so I'm gonna give her the thing that she might think is the signal place now she's gonna get some negative reinforcement
Rocky
Day one on a flat collar: a consequence is either a surprise (prediction error) or a confirmation (prediction success). The common success, the trainer says, is a dog who whines and freaks out until the person stops and apologises. With Rocky the pressure did not stop until he went down, and that was the prediction error.
On film
and then you have a consequence of some kind. And that consequence can either be surprising which is what we call a prediction error or it can be a confirmation that they already know and that's prediction success
Whoever it is stops and goes, oh my gosh, I'm so sorry, buddy. Or in this case, what we saw was a prediction error. The pressure didn't stop.
Roofus
The Piglet problem on Roofus (the recording is titled for Irwin): his brain is like a predicting machine, which is why he learns positions fast, but he starts over-predicting patterns, such as sitting whenever the handler stops. The fix is pattern breaks: stop and ask for nothing, to violate the pattern.
On film
noticed that his brain the reason that he learns positions so quickly is because his brain is like a predicting machine
start incorporating pattern breaks here where I stop and I don't ask him to do anything right so if I stop right here he might predict that I want him to do something but I need to not do anything not ask him to do anything and violate that pattern
Teddy
Named in passing on an off-leash hike, on a heel rep: each rep is a prediction error that builds the picture in the dog's mind of what heel predicts and where he has to be.
On film
Another learning rep. Right? Prediction error. He's starting to see the picture of what a heel looks like.
Thor
Proofing a stay: the trainer says you need the dog to make a mistake, because he has to try something and get a prediction error - he thought getting up was fine, and a negative consequence happened; when he stayed, good things happened.
On film
Because they need to try something and get a prediction error. Because you probably thought, oh, like, you know, get up. Nothing will happen. It's all good. And instead, a negative consequence happened, and that's how they learned.
Titan
Connor's owner-lesson framing again, as on Stormie: two main ways dogs learn, conditioning first (let's go said at every dog) and prediction error second.
On film
second point which is there's two like main ways dogs learn the first one we already talked about it's conditioning and the second one is through prediction error
Why it works
Stated on film: a dog acts on what it expects, and when the expected result does not come, it adjusts its behaviour. Connor calls this one of the two ways a dog learns (Stormie, watch 00:03). Dutchess's trainer calls it half of how they learn (watch 08:35). Kelton and Dolce's trainer call it all learning (watch 11:00; watch 15:31). The error runs in both directions. The dog may expect nothing and get something: the puller feels the collar, Dolce hears the marker. Or the dog may expect something and get nothing: the biter gets no flight and no fight, the avoider gets no pressure after the leash pickup, the frightened dog gets no danger after the startle, the recalled dog gets no work. Samson's walk gives the reason it needs an ending: learning from the first trial needs a return to baseline, because an event that never ends in the dog's mind cannot be learned from (watch 11:00). Finn's session says the same thing about pressure (watch 16:32). Repetition finishes the job: the dog tries the old move the next time and the time after that, and after enough failures he asks, why do it (watch 39:10; watch 25:58). The same mechanism explains the problems. Prediction built by accident produces avoidance at the first link of a pressure chain (Kali) and anticipated commands (Piglet). An owner or vet who reliably gives way produces a dog whose protest works (Louie).
Inferred
The trainers never measure a prediction; the expectation is their account of the dog, and the evidence on film is what the dog does next. Only the shake-off is given as the sign of the error itself, and only by one trainer. The pairing with conditioning is stated (Stormie, Dutchess). How the two fit together is inferred here: conditioning builds the predictions the handler wants, and prediction errors break the ones he does not. The no-show and the correction are treated here as one mechanism because the trainers use one name for both. No trainer says they are the same technique.
Where the trainers differ
Is a prediction error something the handler does, or something the handler allows?
One view
Something you do, and can announce, stage and count. Kelton says he will do one at Bo's crate before he does it. Reaper's pick-up is set up with the family so that one happens. Kali gets a planned couple of false anticipations. Louie's trainer offers to run prediction error tests at the vet. Reaper's later public walk names a delivered punishment for a lunge as the prediction error.
On film
yeah i'll just do a prediction error see what he does hey bud
Because the purpose of this exercise, the reason that we're not, we work with tons of dogs. Most of the time, we just send a random person, one of our staff, to go pick the dog up.
No, and I would even be willing to go to the vet with him and talk to the vet and see if they would be willing to do some prediction error tests with me.
to find out what actually happens and get a different result than they thought. That's where all the learning happens because that's what you would call a prediction error, right?
Another view
Something you make room for. The dog brings the hypotheses; the handler keeps management loose enough for them to fail, and tight management is said to prevent learning altogether. Finn's three failed ideas were his own. Louie's trainer wants to allow one to happen rather than produce it.
On film
So I have to allow enough freedom for prediction errors to occur for management to be out of the way for her to start feeling differently about what's going on.
Uh although Christian's already kind of worked with it a little bit and done a couple of prediction errors, um I still want to see it and um try to allow it to happen because if we can get it, then we can we can allow the prediction error to happen and he can learn from it.
Or, alternatively, what actually happened is a prediction error. He thought, okay, I can relieve the pressure by laying down and doing nothing.
How to read it
It may be one practice seen from two sides: the handler sets up the situation, then waits for the dog's own move. The Reaper pick-up does both ('they've never given the dog a chance to try'). But no trainer says so, and in the moment the two call for opposite handler behaviour.
Is prediction error all of learning, or half of it?
One view
All of it.
On film
return to baseline that's a super super super important part of our dog training philosophy and what it is that we do because all learning happens through prediction error
all i'm doing is doing a little bit of like the way all learning happens it's just uh she's predicting that if she feels like that little bit of leash tension she needs to come off of it
Another view
One of two. Conditioning is the other, and the owner curriculum teaches them as a pair.
On film
different and when you give a prediction error correctly you see the dog kind of do this like shaking motion right and that's like the body language is i'm learning something like i'm processing so that's like half how they learn the other half is conditioning
so that's the other way is when a dog predicts that something's gonna happen and that goes oh i was wrong then they adjust their behavior accordingly
How to read it
A difference in how the theory is framed, not in practice. It matters for any page or handbook that states the house theory in one line. The Dolce line may be casual phrasing; the Samson line is given as the philosophy.
Build the prediction, or break it?
One view
Build it. Say the cue every single time you pass a dog until she predicts it; the finished dog looks as if she does not care.
On film
go and then she immediately unfixates and follows along and it's the same thing with the dogs if i say let's go every single time we walk by a dog her brain starts to go well every time we walk by a dog he says let's go i'm going to predict it this time
Another view
Break it. A dog who anticipates the command is not working off the word, and a predictable routine is what created the problem. Vary it so the antecedent stops predicting.
On film
and um like anticipation like that is just always based on prediction so the only way to get rid of the anticipation is to remove the antecedent right in the arrangement um make it not predict that every time
if i'm if it's anything where it's a change yeah and then that change is followed like it like then like a chain starts yeah that ends in some kind of pressure yeah they predict it they predict predict it and then they avoid it
How to read it
Probably not a real disagreement: Stormie is taught to predict the handler's cue from the trigger, while Piglet and Kali had learned to predict pressure or the next command from the handler's routine. But the line between a prediction to build and one to break is never drawn on film, and an owner told 'every single time' in one lesson and 'not every time' in another needs it drawn.
How long a return to normal before the next go.
One view
Seconds. At the crate the dog is let back to normal in a second and the next handler goes.
On film
it's not as severe as yesterday yeah let's look into the prediction area we just let them return to normal in a second yeah yeah you can go ahead
Another view
Minutes. Two to six minutes of freedom before the next ask on Samson's walk. On Finn, sniffing and jogging until he is back to a relaxed state.
On film
of time to return to neutral right so like if I did another recall and try to make him sit again right this moment there wouldn't be enough time for him to return to baseline and learn from what we just did right so I might end up going two three four five six minutes before I ask him to do anything else
I went right back to it really fast which actually don't like become too distinct coherent like coherently distinct like events they're not two different events in in Finn's brain
How to read it
The criterion is the same everywhere (back to normal), and the time is what it took each dog. Bo's crate event was short and mild ('not as severe as yesterday'); Samson and Finn had just taken pressure. Left as a difference in practice, not resolved.
Common owner mistakes
- Saying stop and carrying on. The standard example of a broken prediction is a dog who has learned that pulling costs nothing, because the owner's word is never followed by anything.
On film
um when they first come to us they're like oh when i pull my owner's like stop and just keeps pulling but does i don't have a consequence so i can pull as much as i want there's no problem
Stormie · Watch at 0:21 ↗ - Apologising and letting the dog off when he protests. The reaction he expected arrives, so next time he does it again.
On film
cases it's the dog going man i don't want to do that and then if i react and go oh i'm so sorry and let him do it next time he'll do it again
Louie · Watch at 25:01 ↗ - Letting the protest end the handling every time, down to sedation at every vet visit. He thinks it works, and it does.
On film
Again he thinks it works and it does Right like he gets to get knocked out every single time right
Louie · Watch at 9:45 ↗ - Running away, or fighting back, when the dog goes to bite. Either one is exactly what he predicted.
On film
turn tail and run. You like that horse poop, buddy? They're expecting that you will turn tail and run. Or they're expecting that you will fight back.
Samson · Watch at 38:51 ↗ or on a more extreme case let's say a dog's bit a couple people and then like everybody runs away in terror
Stormie · Watch at 0:35 ↗ - Following every recall with hard work, until coming to you predicts that it's gonna suck.
On film
so if every time I say come she comes to me I go down or I go sit or I go whatever it is and then I have to put some pressure on her to get it then next next time I say come if I do that 15 times in a row and I say come she's gonna be like I don't know if I want to it's gonna suck
Sage · Watch at 32:05 ↗ - Managing the dog so tightly that he can never be wrong. It is named as preventing learning, and as the thing that makes a dog panic the moment the management is gone.
On film
If I restrict her options too tightly, she'll panic if she's not restricted. That kind of tight management doesn't allow for any kind of learning.
Sage · Watch at 18:48 ↗ - Running the same routine every time, so the dog learns to predict the command (or the pressure) from the handler's habits instead of listening to the word. The trainer names himself as the one who did it.
On film
every time right so maybe maybe i made that predicted too often that like when i when i settled my gaze on her and i kind of watched her that that meant that i was going to ask her to do something every single time
Piglet · Watch at 7:25 ↗ - Going straight back to the next rep before the dog has come back to normal, so two events become one long stretch of pressure.
On film
I went right back to it really fast which actually don't like become too distinct coherent like coherently distinct like events they're not two different events in in Finn's brain right right, instead of being two coherently distinct events, it's just one prolonged event of negative reinforcement, which is no good
Finn · Watch at 16:32 ↗
Questions owners ask
How does a correction teach anything?
On the trainers' account, it teaches by surprising the dog. A dog who has learned that pulling costs nothing finds, with the trainer, that it does. In Connor's words, that's weird, that's not what was supposed to happen, and the dog adjusts. With Dolce the correction is only a word, AH AH, telling her that her reading of the leash was wrong. It is followed at once by the right signal. The correction is the surprise, and what the dog is shown next is the answer.
On film
um when they first come to us they're like oh when i pull my owner's like stop and just keeps pulling but does i don't have a consequence so i can pull as much as i want there's no problem they come with me and they pull and they feel the e-collar and they go that's weird that's not what was supposed to happen prediction error
all i'm doing is doing a little bit of like the way all learning happens it's just uh she's predicting that if she feels like that little bit of leash tension she needs to come off of it which is correct in most cases but she's taking that as a relief signal and I'm giving her that a prediction error there where it's like she's getting an unexpected consequence which is that I'm telling her that she's wrong by saying AH AH and so then from there it becomes relatively simple I give her the real signal which is the okay command
Does it always mean correcting the dog?
No. In most of the filmed cases the surprise is that nothing happens. Samson went to do what had always worked and the trainer just stood there. Kali was called in and asked for nothing. Jack startled and no danger followed. Jack's trainer would not correct even a fear bite, because the bite getting nothing is the prediction error. Kelton calls this a no-show. The term is also used the other way: on Reaper's public walk a punished lunge is called the prediction error, because lunging had always cost him nothing.
On film
And he just needed to learn. He needed to have some prediction errors where essentially Eventually all that happens is he goes to do the thing that has been successful for him in the past and nothing happens.
false starts there because you can see like this little anticipation of like okay so you're gonna like ask me to do a command and put some pressure and then i just completely no-showed that and gave her scratchies and said go right
Like, if he did, I would just let him bite me. Do nothing about it. No correction of any kind, no punishment. It's a prediction error, straight up.
Because he predicted what's always happened in the past, which is like, good boy, come on, let's go, which is really like not that much.
Why does my dog keep doing it when I've told him no?
Because it keeps working. The trainer's example is a protest: if you react, apologise and let him off, next time he'll do it again. The same goes for a vet who sedates him every visit: he thinks it works, and it does. The fix is to stop it working, calmly. The trainer expects the dog to try it a few more times. After enough failures, the tenth time, he says, the dog asks, why do it.
On film
cases it's the dog going man i don't want to do that and then if i react and go oh i'm so sorry and let him do it next time he'll do it again
then maybe i could do that yep didn't work the last three times and probably after you know the 10th time he's like man didn't work yeah and he's like so why do it right
How do I know it landed?
The trainers point to what the dog does next. Dutchess's trainer says a well-timed prediction error produces a shaking motion, the dog processing. Jack came right back to the thing that had frightened him. On Sage the trainer waits for the light to switch on, where she starts to take in information instead of freezing. Samson started sniffing and trying to make friends. These are the trainers' readings of the dog, and only the shake-off is claimed as a sign of the moment itself.
On film
different and when you give a prediction error correctly you see the dog kind of do this like shaking motion right and that's like the body language is i'm learning something like i'm processing
So I just want to see that light kind of switch on where she starts to process information and stop it freezing off.
Why do the trainers wait so long between reps?
Because the dog cannot learn from an event that has not ended. Kelton calls return to baseline a super important part of the philosophy. After a trial Samson gets several minutes of freedom before the next ask. On Finn the trainer admits he went back too fast, so two events became one long stretch of pressure.
On film
return to baseline that's a super super super important part of our dog training philosophy and what it is that we do because all learning happens through prediction error and when you have a prediction error on the first trial you need um to allow for a return to baseline in order for any learning to occur
I went right back to it really fast which actually don't like become too distinct coherent like coherently distinct like events they're not two different events in in Finn's brain
The lesson
The goal
Recognise the prediction a dog is acting on, and learn the trainers' two ways of breaking it: an unexpected consequence, or the expected result not arriving (the no-show). Then allow a return to baseline before the next go.
Learn first
- Two events, not one - and what the handler does in between
- Neutrality is the goal, friendship is optional
- False reps: the dog reads the chain, not the word
See it done
- Connor teaches the idea to an owner with three worked examples: the puller, the sleeve biter, and Stormie at another dog. Watch at 0:03 ↗
- Kelton explains the no-show in bite gear: neither flight nor fight, confusion, and a dog who starts trying other things. Watch at 38:11 ↗
- False anticipations on a recall: call in, ask for nothing, scratches and release, narrated as it happens. Watch at 15:39 ↗
- The sit given where the dog expects the down, so she has to listen to the word. Watch at 27:59 ↗
- A startle with no danger after it, named as a prediction error the moment it happens. Watch at 18:42 ↗
The drill
As run on film with Kali (watch 15:39). When the dog starts to hesitate on coming in because she expects a command and pressure, give her a long break first. Then call her over and ask for nothing: scratches, the release word, and send her off to explore. Do a couple of these before the next rep that asks for something. The leash version is the same (watch 06:28): pick up the leash and have nothing bad follow, and watch her body language to see whether the pickup still means pressure to her. The obedience version is Piglet's (watch 27:59). When she offers the down before being asked, cue the sit instead, then mix the order so she has to hear the word.
Watch out for
- Reacting to the protest: apologising, backing off, letting the dog out of it (watch 25:01).
- Fleeing or fighting when the dog snaps. Either gives him what he predicted (watch 38:51).
- Going straight into the next rep before the dog is back to normal (watch 16:32).
- Running the same routine every time, so the antecedent predicts the command (watch 07:14).
- Following every recall with hard work (watch 32:05).
Check yourself
Your dog lies down and freezes whenever you pick up his leash. What is he predicting, and what would a prediction error look like? He predicts pressure is coming. Pick up the leash and let nothing bad happen, go back to what you were doing, and repeat until the pickup stops meaning pressure.