How does a correction actually change my dog's mind?

The owner-curriculum version, delivered twice to the same family, and the place the whole idea is easiest to hand to somebody else. It is named as one of the two ways a dog learns and as the one the e-collar works on, and it comes with three worked examples in one breath: the dog who pulls because pulling has never cost anything; a dog who has bitten people, bites a sleeve, and finds that nothing happens; and this puppy, who thinks she can run up to another dog. Then the positive half, which is what makes the negative half humane - charge the word up, because a word you do not understand is noise and a word you do understand is a warning, and he uses the bluntest possible analogy for it. And then the two are put together into an outcome: say the cue at every single instance of the trigger until she predicts the cue rather than the dog, and what an onlooker sees is a dog who does not care, while psychologically there is a whole thought process going on. That last line is the most valuable thing in her file, because it explains why finished behaviour looks like nothing.

The same curriculum in a different handler's mouth, which is what makes it a house teaching rather than one person's framing. It is delivered as the psychology half of a first owner lesson, and the trainer says explicitly that he is not doing this because they have a dangerous dog - he goes over it with every single person, because knowing how to communicate with a dog is part of training. Two things in his version do not appear elsewhere. He works the example on this dog's own history rather than on a generic dog: when she first came to him she pulled and thought nothing would happen, and then felt pressure and thought, that was different. And he gives the observable, which is the only physical tell anybody in the corpus attaches to the mechanism - a correctly delivered prediction error produces a shake-off, and the shake is the body language of a dog processing something. In the same lesson he also supplies the reason the concept matters in a bite case: a fixated dog pulled away can redirect onto the person in the way, without meaning to hurt them.

The hardest application in the corpus, and the only place the phrase is used as a verb. A dog has two patterns installed at his crate - fight, or threaten until people back off - and the treatment is to violate both by not leaving: the hand goes in and stays, the growl is marked, nobody escalates and nobody withdraws. The trainer announces it before doing it, which is what makes the file so useful: "I'll just do a prediction error." It is also treated as a discrete, bounded event with an after: as soon as they had done it, they let the dog return to normal. And then the third use, which is the most humane thing anybody does with a bad reaction anywhere in this set. When a handler gets a worse response than the one before, the explanation offered is not that the dog dislikes her or holds a grudge - it is just a prediction issue, a dog still working out how this person's behaviour maps onto the world, so go back to it. The company's own view of where this sits is visible in its publishing: an unread upload in this dog's playlist is titled around prediction error as the headline of his case.

The theoretical statement, and the version that most changes what a correction is. The trainer sets out the full sequence: you ask for a behaviour, the dog tries something, and the result is either a prediction success - his hypothesis about how to relieve the pressure worked - or a prediction error. Then he applies it to an episode in which a frightened dog ran three ideas in a row and all three failed: freezing did not work, running and hiding did not work, running past him in a panic did not work. Naming them that way is how the episode is understood, and it makes the correction a disproof of the dog's theory rather than a punishment of the dog. The same session adds the timing rule that the framing forces: he went back to it too fast, and two corrections that are not coherently distinct events in the dog's brain stop being two events at all and become one prolonged stretch of negative reinforcement, which is no good. That is the sharpest statement in the corpus of why the gap between repetitions is part of the technique.

The dog who shows the restraint side of the technique, which is the half nobody markets. Around her triggers the management is deliberately kept loose, and the reason is stated as a mechanism rather than as a kindness: he has to allow enough freedom for prediction errors to occur, with management out of the way, for her to start feeling differently about what is going on - and if he restricts her options too tightly she will panic the moment she is not restricted, because tight management does not allow any learning at all. The observable he waits for is not compliance but processing: the light switching on, where she starts taking in information instead of freezing. And her file contains the one use of the term in the corpus with no aversive in it at all. A dog who has learned that being recalled means hard work is called over and simply loved on - so the picture is made wrong in the pleasant direction and the wariness repairs.

The most compact and most theoretical statement in the corpus, and it is about something tiny. A protection dog has learned - correctly, in most cases - that a little leash tension means come off it, and she has started reading that tension as a release signal. So the sequence is run in three beats and narrated as it happens: she predicts, the marker word tells her she is wrong, and then the handler supplies the true signal, which is the release word. What is notable is the framing he puts around it: this is "the way all learning happens", offered as a general claim by a handler correcting a very small error in a very advanced dog. There is no borrowed jargon in it and no drama. It is also the only instance in this set where the thing being corrected is the dog's reading of a signal rather than a behaviour.

The conditioning end, filmed on the day it is introduced, which is the other half of the pair and the half that has to come first. The marker is being paid for the first time in this session and the trainer says so - this is the first time he is being praised - and the rule is stated as unconditional: every time he follows the marker, he gets the high-value thing. The same passage contains the criterion for accepting an e-collar level, and it belongs here because it is about the dog's state rather than the dog's compliance: the level catches him off guard a little, he is really excited about it, and he is not losing any of his bounciness. And at his go-home lesson the decay half is given: use the word and never the button and it looks good for a couple of weeks, then goes downhill, because there is no backup. Charge it, then never stop backing it.

The dog whose file proves the term is in-house currency, and the one where it is never explained. It appears once, in a handler's opening narration, as a countable intervention already carried out on this dog's reactivity by another trainer - Christian has worked with it a little and done a couple of prediction errors - and the speaker moves straight on, because he expects to be understood. He then states the same restraint logic Sage's file gives: he wants to see the behaviour and to try to allow it to happen, because if they can get it they can let the prediction error happen and the dog can learn from it. It is the only named technique in the corpus attached to this dog's actual presenting problem, and nothing in his file says what it consists of. That is why the Tier 2 arc flagged it as the most valuable undefined term in the corpus.

Is a prediction error something a handler does, or something a handler allows?

Not addressed anywhere, and the two are not merely emphases - they imply opposite handler behaviour in the same moment. One says step in and falsify the expectation; the other says back off so the dog can be wrong on his own. Louie's file, remarkably, contains both halves in one sentence: another trainer has done a couple of them, and this handler wants to allow one to happen. Whether the business distinguishes a delivered prediction error from an allowed one is the single most answerable open question on this page, and Kelton could settle it in a sentence.

How many, and how often.

These may be the two halves of the pair rather than a disagreement - conditioning is built by unbroken repetition, and a prediction error is by definition rare because it only exists where an expectation already stands. But they are not stated that way, and an owner given both instructions has been told to do something every single time and to do it once or three times. The reconciliation is nowhere on film and is left open here.

Whether the term is explained to owners at all.

Not in conflict about what to teach, only about which words to use - but it matters for every page written from this corpus, because the business's own most distinctive technical term is one that one of its handlers deliberately avoids in front of clients. That is an editorial decision rather than a training one, and it should be made once rather than per page.

Why does a correction teach anything, rather than just hurting?

Because on this account the thing being changed is a bet, not a body. The dog has a working theory - I can pull and nothing happens - and he acts on it because it has always been true. The correction's job is to make it false, and the trainers describe the moment from the dog's point of view: that's weird, that's not what was supposed to happen. Two consequences follow, and both make the practice gentler rather than harsher. Certainty matters more than severity, because what breaks a statistical theory is the outcome arriving every time. And a dog whose theory has just failed needs a moment - one trainer catches himself going back in too fast and explains that two corrections too close together stop being two events and become one long stretch of pressure, which teaches nothing.

How will I know it worked?

The tell the corpus offers is the shake-off. One trainer tells a family that a correctly delivered prediction error is followed by the dog shaking himself out, and that the shake is the body language of a dog who has processed something. On a frightened dog a softer version is used: the trainer says he is waiting for the light to switch on, for her to start taking in information instead of freezing. Both of those are about the dog's state rather than his compliance, which is the point - a dog who simply stops doing the thing may have understood or may have shut down. It is worth saying that the shake-off is stated once, by one handler, and nothing else in the corpus confirms it.

Does this mean I have to let my dog misbehave?

It means you have to leave him room to be wrong, which is not the same as letting him rehearse the behaviour. The clearest statement of it is about a nervous dog: the handler has to allow enough freedom for prediction errors to occur, with management out of the way, for her to start feeling differently - and the failure mode is named in the previous sentence, that restricting her options too tightly produces a dog who panics the second she is not restricted, and that tight management does not allow learning at all. In practice the corpus does this with room rather than with risk: a long line, a big space, a known tester dog, a situation the handler can contain. The version to avoid is the household that has solved the problem by never letting it happen, because that household has a dog who has learned nothing and an owner who can never stop.

Why does the trainer's 'no' work when mine doesn't?

Because his has been charged and yours has not. The analogy given at a handover is about language: a word in a foreign language means nothing to you, and the same word means a great deal if you know it is followed by something that matters. So the word is built by pairing it with the consequence every single time until the word itself predicts the consequence - and at that point the word does the work alone. The same trainer's colleague gives the failure mode from the other end: use the word and never the button, and it looks good for two weeks and then decays, because there is no longer anything behind it. This is why an owner who has said the same word for a year without it ever being followed by anything has a word the dog treats as weather.

Is my dog being stubborn, or does he not understand?

The corpus's habit is to reach for a third answer before either of those: he is still working out how this situation, or this person, maps onto the world. When a handler at a crate got a worse reaction than the handler before her, the explanation offered was not a grudge and not defiance - it is just a prediction issue, he is trying to figure out how that works - and the instruction that followed was the counterintuitive one, which was to go back to it rather than hand the dog to someone else. On a frightened dog the same generosity is applied to three failed attempts in a row: freezing, hiding and bolting past the handler are described as techniques the dog tried, hypotheses about how to relieve the pressure, rather than as disobedience.

Saying stop, and doing nothing. The corpus's standard example of a dog with a broken prediction is a dog who has learned that pulling costs nothing because the owner says stop and keeps walking.

Managing the dog so tightly that he can never be wrong. Named as the thing that prevents learning, and as what produces a dog who panics the moment the management is removed.

Using the word and never backing it. It looks fine for a fortnight and then decays, because the word has nothing behind it any more.

Repeating the correction before the dog has come back to baseline, which turns two events into one long stretch of pressure the dog cannot parse.

Backing away from a growl, which confirms the dog's prediction rather than falsifying it. The whole crate protocol is built on refusing to do this.

Reading a worse reaction to one particular person as a grudge. The explanation offered in the corpus is that the dog is still working out how that person's behaviour maps onto the world - so the answer is more of that person, not less.

Only ever making the dog wrong in the unpleasant direction. A dog who has learned that being called means hard work can be recalled and simply loved on, and the wariness repairs.

You have been saying the same word for months and it has never once been followed by anything.

He is fine as long as you hold him tightly, and unrecognisable the moment you let go.

Everybody in the house has learned to leave him alone when he growls, and it always works.

He is noticeably worse with one particular person in the household and you have decided he does not like them.

He avoids you when you call him, or comes slowly - because being called has always meant the end of the fun or the start of the work.

The training looked great for a couple of weeks and then faded - and you had stopped using the collar because you did not seem to need it.