--- title: "Predictive Processing and Cognition: What I Was Looking for in the Predictive Brain" date: 2026-09-06 updated: language: en document_type: research-note status: working epistemic_status: mixed authority: working-note source_format: markdown topics: - predictive-processing - predictive-coding - active-inference - perception - memory - uncertainty - anxiety - schemas - cognitive-compression - correction related_books: - "You Are Not the Crazy One" related_papers: - "Understanding or Closure?" - "The Correction Channel" related_projects: - sana-os related_concepts: - "CPM" - "History Analysis" - "Fact / View / Care" related_notes: - research-development-map question: "If perception and memory are constructed through prediction, compression, and prior knowledge, what does that imply about certainty, disagreement, and the ability to correct one's own model?" thinking_context: - older-unresolved-question - conversation-with-ai - ai-assisted-research - literature-search - framework-development research_methods: - ai-assisted-research - literature-search - knowledge-base-search - conceptual-synthesis source_verification: mixed supersedes: superseded_by: --- # Predictive Processing and Cognition ## Question If perception and memory are constructed through prediction, compression, and prior knowledge, what does that imply about certainty, disagreement, and the ability to correct one's own model? I did not begin this research because I wanted to prove that "the brain is a prediction machine." The question that mattered to me was more practical: > Why can a person experience an interpretation as obvious reality even when another person, looking at the same situation, experiences something very different? A related question was: > Why does receiving more information sometimes correct a belief, while at other times the new information seems to be absorbed into the old story? These questions were already connected in my thinking to the distinction between an event, the way it is perceived, and the record later created from it. The research on predictive processing, reconstructive memory, uncertainty, and compression seemed potentially useful because it offered mechanisms by which the same physical world could produce different experienced worlds without requiring either observer to be consciously dishonest. This note records that exploration. It does not claim that predictive processing provides a complete theory of cognition, nor that every disagreement can be explained by it. ## Thinking Context I cannot reconstruct a single origin for this line of thought. The question had already appeared in historical analysis, conversations about memory and conflict, and the development of SANA's distinction between Fact, Perception, and Record. Later conversations with AI systems helped make the question more explicit, and AI-assisted research tools were then used to explore relevant literature. The research expanded into several neighboring areas: - predictive coding, - the Free Energy Principle, - active inference, - omission responses, - expectation effects in perception, - reconstructive memory, - schema-dependent learning, - choice blindness, - uncertainty and anxiety, - information bottlenecks, - and lossy compression. The resulting research document connected these areas aggressively. Some of those connections may be useful. Some are probably too strong. That distinction is one reason to preserve this as a Research Note rather than convert the exploratory document directly into a paper. ## Sources Consulted The exploratory research included material associated with: - Karl Friston and predictive coding / the Free Energy Principle - hierarchical prediction and prediction-error research - Catherine Wacongne and related omission-response experiments - Floris de Lange, Peter Kok, and related expectation / sensory-representation work - schema and memory-consolidation research associated with Richard Morris, Dorothy Tse, and related work - Elizabeth Loftus and reconstructive-memory research - choice-blindness experiments associated with Petter Johansson, Lars Hall, and colleagues - Dustin Grupe and Jack Nitschke's work on uncertainty and anxiety - interoceptive inference and computational psychiatry - information-bottleneck and rate-distortion approaches - work on lossy memory compression and adaptive memory - secondary explanations, repositories, review articles, preprints, and AI-assisted source discovery The source set was heterogeneous. Some references were primary papers or peer-reviewed reviews. Others were summaries, repositories, explanatory webpages, or recent preprints. This note therefore treats the source list as a map of what I was exploring, not as a verified bibliography. ## Verified Primary Sources At the time represented by this note, source verification was mixed. Several major experimental paradigms were identifiable and worth checking directly: - omission-response experiments, - expectation effects in visual processing, - Loftus and Palmer's post-event information work, - choice blindness, - schema-related memory consolidation, - and uncertainty research. However, many claims in the exploratory document go beyond what can safely be inferred from any single experiment. Before using them in a formal paper, I would want to verify: 1. the original experimental result, 2. the limits of the paradigm, 3. whether the interpretation is disputed, 4. whether the effect has replicated, 5. and whether the computational explanation is one interpretation among several. This is particularly important for claims that connect a neural observation directly to a broad statement about consciousness, anxiety, certainty, or human rationality. ## What I Found Interesting ### 1. Prediction appears to influence perception before conscious interpretation The omission-response literature was one of the clearest reasons predictive processing interested me. In an omission paradigm, the system learns a regular sequence and then an expected stimulus is sometimes omitted. What matters conceptually is that the brain can respond at the expected time even when the external stimulus is absent. That makes prediction harder to dismiss as merely a retrospective verbal interpretation. Something in the nervous system has prepared for an expected event before the event arrives. This does not by itself establish every claim made by predictive-processing theories. But it supports the narrower idea that perception is not simply a sequence of passive registrations. The system is already structured by expectation. ### 2. Expected input and unexpected input may not be processed symmetrically The research on expectation suppression and sharpening suggested another possibility. If expected sensory information can be represented differently from surprising residual information, then prediction may help allocate processing resources. That is interesting because it changes how I think about "seeing what is there." The system may not treat every incoming element as equally informative. What fits the current model may require less updating. What violates it may demand more processing. This makes prediction error a useful abstract concept: > the difference between what the model expected and what the system encountered. That difference may be more important for learning than raw input alone. ### 3. Perception, memory, and record are not the same operation This research reinforced a distinction I was already using in historical analysis: ```text Fact → Perception → Record ``` The event itself is not identical to the organism's perception of it. The perception is not identical to what is later remembered or recorded. Reconstructive-memory research is important here because remembering is not equivalent to replaying a stored video. Post-event information, schemas, expectations, and later questions can alter what a person reports. This does not mean memory is arbitrary. It means the record is produced by a system with constraints, priors, goals, and limited storage. The practical implication for me was not "human memory is false." It was: > A sincere record can still be a reconstruction. That matters enormously for historical sources, interpersonal conflict, testimony, and self-narrative. ### 4. Existing schemas may accelerate learning while also shaping what is retained The schema literature suggested a trade-off that I found especially interesting. Existing knowledge can make new learning dramatically more efficient. Information that fits an established structure may be easier to integrate because the system already has somewhere to put it. But the same fact raises a harder question: > What happens to information that does not fit? The exploratory document sometimes described this too strongly, as though inconsistent information were simply discarded. I do not think that is a safe general claim. A better working formulation is: > Existing schemas can influence encoding, interpretation, consolidation, and retrieval, and therefore may bias which aspects of an experience become stable parts of later memory. That is enough for the conceptual problem I care about. A model can help us learn faster and still make some corrections harder. ### 5. Choice blindness suggested limits on introspective access Choice-blindness experiments were another important connection. In these experiments, participants can sometimes fail to notice that the outcome presented to them differs from the choice they actually made, and then provide reasons for the substituted outcome. The strongest interpretation in the exploratory document was that consciousness simply fabricates a coherent narrative after the fact. That formulation is too sweeping. But the experiments still raise a useful question: > How much direct access do we really have to the processes that produced our own judgments? If a person can sincerely explain a choice that was not actually theirs, then fluent explanation cannot by itself prove accurate introspection. This matters because humans often treat the availability of a reason as evidence that the reason was causally responsible. That may not always be warranted. ### 6. Uncertainty may be affectively costly, but the bridge from uncertainty to closure needs care The research on anxiety and intolerance of uncertainty was especially relevant to SANA. The exploratory model proposed something like: ```text uncertainty → sustained prediction error / instability → affective pressure → rapid closure → stable narrative ``` I still find the structure plausible. But several levels are being combined here: - formal uncertainty in a computational model, - subjective uncertainty, - anxiety, - physiological arousal, - threat expectation, - and cognitive closure. They should not be treated as interchangeable. A safer working hypothesis is: > Under some conditions, uncertainty carries cognitive and affective costs, and threat-related states may bias how evidence is weighted and how quickly people seek a stable interpretation. That is much weaker than saying that "the brain cannot tolerate uncertainty." But it is also more useful because it can be tested without turning a computational metaphor into a universal psychological law. ### 7. A prior can be useful and still become difficult to revise Predictive-processing language gave me a useful way to think about strong expectations. A prior is not an error. Without prior structure, interpretation would be impossible. The problem begins when a prior becomes so dominant that contradictory evidence no longer receives enough effective weight to change the model. In simplified form: ```text prediction + incoming evidence ↓ updated model ``` works only if the evidence can matter. This became one bridge toward what I later began calling the Correction Channel. The central question is not whether a system has priors. Every useful system does. The question is: > Can disconfirming information survive long enough, and with enough weight, to modify the model? That is a different problem from merely detecting error. ### 8. Compression may be unavoidable The information-bottleneck and lossy-compression material connected the cognitive problem to a more general systems problem. No finite system can preserve every detail of every input. Some form of selection, abstraction, or compression is unavoidable. That means the relevant question is not: > Can cognition avoid distortion? It may be: > Which distortions are introduced by the compression strategy, and can the system detect when those distortions matter? This is one of the most important shifts in my thinking. Compression itself is not necessarily a bug. A map must omit most of the territory. The problem appears when: 1. the omitted information becomes important, 2. the system cannot recover it, 3. and the compressed representation is mistaken for the territory itself. ## Working Interpretation At this stage, I was moving toward the following abstract loop: ```text world / input ↓ prediction ↓ prediction error ↓ uncertainty ↓ model update or assimilation ↓ compressed record ↓ future prediction ``` The crucial branch is not yet well defined. When reality contradicts a model, at least two broad responses seem possible. ### Path A: Model Update ```text evidence conflicts with expectation ↓ conflict remains visible ↓ model changes ``` ### Path B: Evidence Assimilation ```text evidence conflicts with expectation ↓ evidence is reinterpreted, discounted, omitted, or narratively absorbed ↓ model remains substantially unchanged ``` This distinction is more important to me than the claim that predictive processing itself explains everything. The deeper question is: > What determines whether discrepancy changes the model or gets changed by the model? That is where predictive processing began to connect with psychological safety, negative capability, organizational dissent, and AI verification. ## A Distinction I Want to Preserve Several ideas that can sound similar should remain separate. ### Prediction is not hallucination Prediction contributes to perception, but that does not mean ordinary perception is equivalent to clinical hallucination. The phrase "controlled hallucination" is provocative and sometimes useful, but it can also blur important distinctions. For formal writing I would rather say: > perception is inferential and constrained by both prior expectations and sensory evidence. ### Reconstruction is not fabrication Memory reconstruction does not imply that all memories are invented. A reconstructed record can contain substantial accurate information. The point is that retrieval and reporting are active processes, not perfect playback. ### Uncertainty is not anxiety Uncertainty can exist without pathological anxiety. Anxiety may alter how uncertainty is processed, but the terms are not interchangeable. ### Confidence is not accuracy A coherent internal record may feel certain without being complete or correct. But uncertainty also does not imply error. These distinctions matter because otherwise the theory itself becomes another compression. ## Connections I Was Considering ### Fact / Perception / Record This was the most direct connection. The neuroscience and cognitive-psychology research did not create the distinction for me, but it gave me reasons to take the separation more seriously. A historical record should not be treated as a transparent copy of an event. A personal memory should not be treated as transparent access to the past. An AI summary should not be treated as transparent access to its sources. Different systems produce records through different mechanisms, but in each case the representation is not the thing represented. ### Negative Capability Another connection was the value of retaining an unresolved variable. If uncertainty is costly, then the ability to remain with "I do not yet know" may be a meaningful cognitive capability. The point is not to romanticize uncertainty. Indefinite suspension can also be maladaptive. The interesting possibility is that keeping a question unresolved for long enough may preserve information that would otherwise be compressed into a premature explanation. This became closely related to the SANA idea of preserving an unknown variable X. ### Cognitive Compression The exploratory research also influenced how I thought about cognitive compression. I do not want to define cognitive compression as "humans become irrational under stress." A more useful version is: > Under pressure, a system may reduce the complexity of the hypothesis space it is willing or able to maintain. That may involve faster categorization, stronger reliance on available narratives, reduced exploration, or greater weighting of threat-compatible interpretations. This remains a theoretical synthesis, not a validated scalar psychological construct. ### Premise Alignment If two people are working from different predictive models, their disagreement may occur before explicit reasoning begins. Each person's interpretation may already contain assumptions about: - what kind of situation this is, - what the other person intends, - what outcome is likely, - and what counts as danger. Premise Alignment may therefore be useful not because agreement is intrinsically good, but because it makes hidden model differences more observable. Instead of arguing over conclusions, the conversation can expose the assumptions generating them. ### The Correction Channel This research eventually contributed to a broader systems question. A healthy system may not be one that avoids prediction error. It may be one that can use prediction error. My current working principle is: > A system is corrigible to the degree that disconfirming information can survive long enough to modify the model or decision process that generated the expectation. Predictive processing is only one source of inspiration for that principle. Organizational voice, scientific falsification, feedback control, and AI verification may instantiate different mechanisms. I do not assume they are biologically or computationally identical. The possible commonality is functional. ## Uncertainties and Limits This research note contains several areas where I want future readers, including future me, to be skeptical. ### 1. Predictive processing is a research framework, not settled ontology The exploratory document sometimes wrote as if predictive coding and the Free Energy Principle were simply established descriptions of how the brain works. That is too strong. Predictive-processing frameworks are influential, but specific formulations, neural implementations, and explanatory scope remain debated. ### 2. The Free Energy Principle was used too freely The research document moved quickly from variational free energy to subjective anxiety, metabolic burden, and premature cognitive closure. Those connections require much more careful justification. Mathematical free energy in a model should not be casually equated with felt psychological discomfort or literal metabolic expenditure. ### 3. Neural-region stories may be overspecified Statements assigning functions such as "the amygdala generates threat prediction errors" or "the insula sends crisis prediction errors upward" can sound more precise than the evidence supports. Distributed networks, task dependence, and competing interpretations matter. I would avoid building the theory around region-by-region stories unless the primary literature clearly requires them. ### 4. The memory-compression story may be too unified Reconstructive memory, schema effects, forgetting, information bottlenecks, and predictive processing are related topics, but they are not automatically one mechanism. The exploratory document sometimes treated them as parts of a single compression pipeline. That is a hypothesis, not an established fact. ### 5. Choice-blindness findings do not prove that introspection is generally fictional The experiments show striking limits under specific conditions. They do not justify the universal conclusion that conscious reasons are always post-hoc confabulations. ### 6. "The brain cannot tolerate uncertainty" is too strong Humans can tolerate uncertainty, sometimes for long periods. The more defensible claim is that uncertainty can impose cognitive or affective costs and that tolerance varies across people, contexts, and states. ### 7. Prediction error does not automatically produce learning A discrepancy can be ignored, misweighted, explained away, or treated as noise. This is precisely why I became interested in the conditions under which correction succeeds. ### 8. The Fact / Perception / Record analogy must not be mistaken for a neural model It is an epistemic distinction used in SANA and historical analysis. The neuroscience research may illuminate why the distinction is useful, but it does not prove that the brain literally contains three corresponding stages. ## What Could Change My View I would revise this working interpretation if: - stronger evidence showed that predictive-processing accounts add little explanatory value beyond more local models, - supposedly predictive effects were better explained by attention, adaptation, or task structure, - schema effects proved less relevant to belief revision than I currently suspect, - uncertainty tolerance showed little relationship to closure or exploratory behavior, - or the model-update / evidence-assimilation distinction turned out to be poorly specified or not empirically tractable. I would also revise the conceptual synthesis if a more precise existing literature already captures what I am calling the Correction Channel. Again, originality is not assumed. The point of this note is to preserve the route through which these questions became connected. ## Later Updates ### 2026-09-06 The most important correction I would make to the exploratory research is linguistic. I would no longer write: > "The brain distorts reality in order to survive." That sentence compresses too many different claims. A better formulation is: > **Finite cognitive systems construct usable representations from incomplete and noisy information, using prior structure to guide perception, memory, and action. Those same mechanisms can support efficient adaptation while also creating systematic limits on correction.** Likewise, I would replace: > "The brain cannot tolerate uncertainty." with: > **Uncertainty can carry cognitive and affective costs, and under some conditions those costs may increase pressure toward premature closure.** The question I now find most useful is not whether prediction "distorts" reality. It is: > **What allows a predictive system to notice that its own useful compression has become wrong?** That question connects this research directly to the later idea of the Correction Channel. --- ### Note on Provenance This note records what I was thinking and investigating at a particular time. It does not claim to reconstruct a single origin for the ideas recorded here. The underlying exploratory research used AI-assisted literature discovery and synthesis. Some claims in that document were stronger than I would now consider justified. Where the original research document used strong neuroscientific or computational language, this note preserves the direction of inquiry while explicitly marking those claims for later primary-source verification.