Discrepancy awareness
The wiki’s predictive pages are full of prediction error. This page is about the one version of it a person can be asked to report.
Three things that are not the same
| what it is | who has it | measured how | |
|---|---|---|---|
| Discrepancy awareness | the person notices, and can rate, that their body felt different from what they expected | the participant | subtract an anticipation VAS from an experience VAS (interoceptive-anticipation-paradigms); or IPA, from the IBL |
| Interoceptive prediction error | a quantity in a hierarchical generative model, weighted by precision, driving belief updating | the model | fitted to trial sequences (BLT) or inferred from neural signal |
| Failure to anticipate interoceptive states | a clinical disposition proposed as the seat of anxiety and depression | the patient | mostly not — see martin-paulus |
The literature slides between these because the same words serve all three, and the slide is usually in the flattering direction: a self-report gap is described as prediction error, and prediction error is described as the clinical mechanism. Neither inference is licensed.
The concrete form of the problem: Wilzok et al. describe their measure as capturing whether participants “recognize when these beliefs had been violated.” What the number actually is, is the difference between two marks a person made on the same scale four seconds apart. A person can produce a large gap by having noticed a violated expectation, or by having rated the anticipation carelessly, or by using the scale differently before and after a painful event. Nothing in the paradigm separates these.
Why it is nonetheless worth a page
Because it is the only member of the trio with an instrument, and because two properties of it have now been measured rather than assumed.
It is reliable. Induced discrepancy retests at ICC 0.78 respiroceptively and 0.54–0.66 nociceptively over about a week. That is unremarkable by the standards of most psychology and notable by the standards of interoceptive measurement, where the gastric field’s flagship quantity has run-to-run reliability near zero (phase-locking-value). A stable individual difference is at least a candidate for being about the person.
It correlates across channels. r = 0.57 between respiroceptive and nociceptive discrepancy in the same 52 people. Whatever this quantity is, it is not organ-specific — which distinguishes it sharply from sensitivity, precision and metacognitive efficiency, all of which decorrelate across channels (banellis-2026-cross-modal-decorrelation).
That combination — reliable, cross-modal — is exactly the profile the wiki has otherwise only seen in confidence. Banellis et al. found mean confidence correlating across cardiac, respiratory and auditory tasks, and concluded it was a fact about the person’s response style rather than about their body. Discrepancy awareness may be the same trait wearing different clothes, and the test is the one Banellis ran and Wilzok did not: put an exteroceptive channel in the battery. If discrepancy magnitude correlates as strongly with a visual or auditory expectation-violation task as with the other interoceptive one, it is scale use. Nobody has done this.
The bias, which is the more interesting result
Set aside the individual-difference question and there is a group-level finding here that constrains interoceptive-inference accounts.
Anticipation and experience did not merely differ under manipulation. They differed at the highest intensity even when the stimulus matched the cue — in both modalities, after a familiarization phase that calibrated every other level fine. Participants overpredicted the strongest stimulus. And the induced discrepancy was systematically larger for underestimation than for overestimation.
Both patterns are long established in the pain literature (Rachman & Arntz 1991): people overpredict severe pain, and underpredicted pain hurts more than it should. Wilzok et al.’s contribution is showing the same shape for aversive inspiratory loading, which suggests it is a fact about aversive interoceptive expectation generally rather than about pain.
The wiki’s predictive-coding pages treat priors on bodily states as having a mean and a precision, with pathology located in mis-set precision. This is a systematic bias in the mean, present in healthy young adults, asymmetric with respect to intensity, and resistant to the training that calibrated the rest of the range. It is a normative feature, not a disorder — and it is the kind of thing an account of allostasis ought to explain (overpredicting harm is cheap; underpredicting it is not), and does not.
The gap this does not close
It is not forecasting. The cue is an arbitrary grey patch predicting an experimenter’s action; the body’s own dynamics play no part. A person anticipating that an experimenter will apply 256 mN is not modelling how their internal state will evolve. That page’s gap remains exactly as wide, and now has a second near-miss recorded against it.