A review of the research informing the Upstream Signal Model — predictive processing, constructed emotion, allostatic regulation, memory reconsolidation, and the problem of working with language in a system that predates it.
The Upstream Signal Model is not a new theory of emotion. It is a synthesis — an attempt to hold several established bodies of research in the same frame and trace their clinical implications together. The science here predates the model. The model is a way of reading that science from a clinical direction: what does this mean for where intervention should actually land?
This page is a working document. It will update as the research does.
The dominant model in cognitive neuroscience over the past two decades holds that the brain is not a passive receiver of sensory information but an active prediction engine. At every level of the hierarchy — from brainstem to cortex — the brain generates continuous predictions about what signals should be arriving, compares them against what actually does arrive, and uses the gap to update its model of the world.
Karl Friston's free energy principle formalizes this: the brain's fundamental drive is to minimize surprise — to reduce the discrepancy between its predictions and incoming sensory signals. (Friston, 2010) Andy Clark's synthesis of this framework into a coherent account of cognition, perception, and action describes a hierarchical prediction system in which higher levels generate top-down predictions and lower levels pass up prediction errors. (Clark, 2013, 2016)
The extension into active inference adds a crucial asymmetry: prediction error can be resolved in two fundamentally different ways. The organism can update its model to match what arrived — genuine learning, genuine revision. Or it can act on the world to make the world match its prediction — behavior, control, avoidance. Both reduce prediction error. Only one changes the model. The clinical implications of this distinction are enormous and remain largely underexplored in practice.
Most emotional suffering is the nervous system running model-protection operations on a model that was accurate in a past environment and is no longer accurate now. The model is not broken. It is working exactly as designed — protecting predictions that have been confirmed thousands of times. The question is how to create the conditions under which the model can update rather than defend.
Lisa Feldman Barrett's work on constructed emotion represents one of the most significant shifts in emotion science in decades. The classical view — that emotions are discrete, hardwired responses that the brain detects and labels — is not supported by the evidence. What the research shows instead: emotions are constructed in real time from interoceptive signals, prior conceptual knowledge, and contextual prediction. (Barrett, 2017)
The same physiological state — elevated heart rate, cortisol, muscle tension — can be constructed as anxiety, excitement, anger, or grief depending on the concepts available and the context in which they are deployed. The concept shapes the experience. The word does not describe a pre-existing emotional state; it participates in constructing it.
Barrett's emotional granularity research extends this: individuals with finer-grained emotional vocabularies show measurably different physiological regulation, better health outcomes, and more effective action selection. (Barrett & Simmons, 2015) The standard interpretation is that more precise labels produce better regulation. The Upstream Signal Model proposes an alternative: the regulatory benefit may derive from the act of pre-conceptual tuning that the labeling task requires — the pause, the attending to the signal before the label resolves — rather than from the precision of the label itself. If this is correct, a more precise emotion word held as a conclusion produces the same compression as a vague one. The label is incidental. The opening is the mechanism.
Emotion vocabulary is organized by phenomenological similarity — shame vs. guilt, anxiety vs. fear — not by mechanism. This produces linguistic precision at the label level while removing mechanistic clarity. The same word routinely covers processes operating at entirely different levels of the nervous system, with different generating mechanisms, requiring different responses. "Shame" describes a somatic contraction, a behavioral hiding strategy, a self-monitoring surveillance system, and a precision-downgrade signal in the self-prior — simultaneously, in the same conversation, without distinction. The label is doing too much and saying too little.
Barrett's granularity finding is real — finer-grained concepts produce measurably better regulation. But granularity and deconstruction are categorically different moves. Granularity works within the concept layer: swapping one past-tense simulation for a slightly more differentiated one. The concept is still running the show. Deconstruction interrupts before the concept closes: what is actually here before we name it? That is a move that goes down the stream toward the signal, not across it.
The clinical evidence points toward a significant departure from Barrett's implicit goal. Barrett's framework aims at a better simulation — more flexible, more granular, happier construction. The Upstream Signal Model proposes a different target: honest contact. The simulation should be accurate enough that the information in the signal can actually reach behavior, relationship, and choice. Not pleasant. Not flexible. Honest. The difference matters because denial is also a DMN interpretation. "This is fine" is also a concept. It can produce a pleasant experience while the signal underneath keeps carrying real information about viability, bonds, autonomy, or orientation — information that doesn't stop being true because the DMN decided to call it fine.
Peter Sterling's allostatic model revised the dominant homeostatic framework for understanding physiological regulation. Where homeostasis proposes that the body maintains a fixed setpoint and returns to it after perturbation, allostasis proposes that the body sets physiological parameters anticipatorily — based on predicted future demands — and that the cost of regulation is carried in this prediction system, not in discrete stress responses. (Sterling, 2012; Sterling & Eyer, 1988)
Allostatic load — the accumulated cost of chronic, unresolved, or miscalibrated mobilization — is not the cost of normal adaptation but the cost of running a body budget that is continuously out of balance with actual conditions. A nervous system calibrated to a dangerous world pays the metabolic cost of that calibration continuously, whether or not threat is currently present. This is the hidden variable underneath most clinical presentations: not what happened today, but what the body has been predicting and spending on for years.
The tank construct in the Upstream Signal Model is a clinical translation of this: allostatic reserve as the finite resource available at any moment, which constrains every downstream process. The same prediction error costs more in a depleted system. Low volume heats faster. And critically — what fills the tank is itself prediction-dependent. A nervous system whose allostatic baseline predicts danger will have difficulty receiving safety as restorative, because safety arriving in a danger-calibrated system registers as anomaly rather than relief. You cannot simply prescribe self-care to a chronically activated system and expect the tank to fill.
Memory reconsolidation research — emerging from basic neuroscience and translated into clinical application by Ecker, Ticic & Hulley — proposes that the original emotional memory is not fixed after initial encoding. When a memory is reactivated, it enters a brief window of instability during which the underlying neural trace can be revised rather than merely suppressed. If new information arrives during this window, the original memory can be updated at the level where it actually lives. (Ecker, Ticic & Hulley, 2012; Lane et al., 2015; Nader, Schafe & LeDoux, 2000)
This is mechanistically different from what most therapy produces. Standard exposure therapy generates inhibitory learning — a competing memory trace that says "safe," running alongside the original trace that says "danger." The original fear memory is not revised. It is suppressed by the competing trace. This is why exposure gains are often context-dependent, stress-sensitive, and fragile: the original prior was not changed, it was inhibited, and under sufficient load the original reasserts.
Reconsolidation requires a specific sequence: the original memory is reactivated in conditions of sufficient safety, its precision weighting temporarily drops — it becomes permeable — and new information arrives during that window. The update happens at the level where the original prediction lives, not at the level of the narrative about it.
The field has organized change around prediction error: expose the person to the feared stimulus, violate the prediction, the update follows. This gets the sequence wrong. A defended system discounts prediction error — files it as anomalous, as an exception, as this situation being different from the real ones. The prior stands. What makes prediction error land as revision is permeability: the prior moving from defended to accessible. Safety creates permission to open. Permeability is the opening. Prediction error is what arrives when the system is open. The field has been working on step three while missing step two.
As prediction error moves through the nervous system, it does not compress — it refines. Each level adds a layer of specificity. At the brainstem: deviation detected, urgency assigned, orient or freeze — no concept yet. At the subcortical level (amygdala, basal ganglia): action signal, urge, mobilization — procedural threat patterns from early learning, pre-linguistic. At the insula and anterior cingulate: pressure — the allostatic budget rendered as a felt quality, valence and salience emerging. The tank's allostatic report enters the stream specifically at this level. At the hippocampus and limbic system: historical association — this feels like that time.
By the time the signal reaches concept capture, it already carries urgency, action tendency, felt quality, and historical context — all before a word has been assigned. When the shaped signal arrives at the default mode network, the DMN runs a probability competition: every past concept associated with signals of this shape gets partially activated, and the best fit wins. The winning concept is sent back into the stream not just as a label but as a full simulation — a word, associated memories, an expected trajectory, a behavioral strategy, and an identity state. You don't just know what you're feeling. You become the person who feels that.
The concept that lands is the most statistically probable description of what this signal most closely resembles from before. It is not a description of what is actually here now. This is why the upstream intervention matters: holding the signal at the pressure level long enough for present-tense information to be felt before the past-tense simulation closes the loop. And why removing the emotion label entirely produces clearer experience, not more diffuse experience — without the label, there is no simulation trajectory. You are in what is actually here, which is more specific because it is this, not the statistical category.
There is a level below which verbal insight does not reach. Language, reframing, and analysis can update the DMN's fast prediction layer and — slowly — its schema. They cannot directly update the prior levels below. Each level has its own access point:
Sequencing is the intervention. You cannot do subcortical work when the tank is depleted. You cannot do interoceptive work when the DMN narrative immediately re-suppresses what comes up. Treatment resistance is almost always a sequencing error — the right intervention applied at the wrong time, in the wrong order, before the conditions that make it possible have been created.
Stephen Porges's polyvagal theory revised the dominant binary model of the autonomic nervous system. (Porges, 1995, 2011) The conventional picture — sympathetic activation on one side, parasympathetic rest on the other — obscures the most therapeutically significant distinction in the entire system. Within the parasympathetic, there are two distinct circuits that produce fundamentally different states: the dorsal vagal circuit (shutdown, freeze, collapse) and the ventral vagal circuit (social engagement, curiosity, genuine felt safety). These are not the same state. Treating them as a single parasympathetic pole erases what matters most.
The clinical implication that most frameworks have not fully integrated: the sympathetic and ventral vagal streams are not mutually exclusive. They can run simultaneously. Play demonstrates this — play is sympathetically activated (energized, alert, responsive) and simultaneously ventral vagal (safe, curious, open). Both streams online at once. The energy is sympathetic. The safety is ventral vagal. This co-occurrence is not an edge case. It is the model for what the learning state actually looks like.
A more accurate map of the nervous system distinguishes three streams rather than two. The sympathetic stream: attention-grabbing, salience-generating, mobilizing toward what matters — appropriate response to prediction error, but when carrying a threat valence, closes the update window. The default mode network: narrative, self-referential, language-based — where most therapy happens, where insight is generated, but not where priors update. And the ventral vagal stream: present-moment, pre-linguistic, curious without agenda — the only stream capable of updating the prior in real time.
The goal is not to turn off the sympathetic so the parasympathetic can do its work. It is to get ventral vagal online while the sympathetic is running at moderate levels — activated and safe simultaneously. That is the learning state. And it is the only state in which the prior becomes available for revision. Sympathetic activation alone protects the prior. DMN processing annotates it. Ventral vagal contact — curious, non-solving presence aimed at the activated material — is the only mechanism that actually changes it.
This three-stream model connects directly to the reconsolidation research above. The permeability window that reconsolidation requires is a ventral vagal state: the prior reactivated, its precision weighting temporarily dropped, the system in curious contact rather than defensive management. The pharmacological research on MDMA-assisted therapy becomes readable through this lens — MDMA simultaneously brings ventral vagal fully online, reduces sympathetic threat valence on the activated material, and suppresses DMN narrative self-reference. (Carhart-Harris et al., 2021) What remains is the aware, pre-linguistic, non-solving presence that can hold experience without narrating or defending it. The drug temporarily restores a state that was the default in childhood and has been progressively covered over by DMN dominance in adulthood.
The three-stream model also explains why technique is limited as a primary intervention. Technique is sympathetic (directed effort) or DMN (planned procedure) — both move away from the ventral vagal stream rather than toward it. What can be built are conditions under which the third stream becomes accessible. The clinical literature on therapeutic alliance, relational factors in outcomes, and the primacy of the therapeutic relationship is, through this lens, evidence for the ventral vagal mechanism: a genuinely regulated, curious, non-solving therapist presence provides the external ventral vagal co-regulation that keeps the client's update window open.
→ Extended essay: The Ventral Vagal Stream — The Science of Awareness Caring
In childhood, the nervous system is a high-learning-rate system. The model is under construction. Prediction errors propagate backward through the architecture, updating priors freely. High error tolerance and high updating capacity coexist. Whatever the relational environment provides — the quality of attunement available, the pattern of threat and safety, the repeated experience of what happens when the signal is expressed — lands directly in the model at the deepest levels, because the gates are light and the priors are still forming.
As the model consolidates — as priors accumulate confirmation, as formal operations emerge around ages 9–12 and the symbolic layer becomes fully operational — the fork between updating the model and protecting it defaults increasingly toward protect. Not because something went wrong. Because a sophisticated predictive system with a well-confirmed model should not be revised by every piece of incoming data. Stability is functional. Certainty reduces metabolic cost. The symbolic layer's job is to make the model coherent — and it does this by absorbing anomalies and explaining away disconfirming evidence before they can update the prior.
After childhood, you are no longer building the model. You are living in it. The map has stopped being updated by the territory. The person is increasingly experiencing their own predictions rather than the world.
This trajectory is not pathology. It is the natural endpoint of a system doing its job. The clinical problem emerges when the original environment that shaped the model no longer matches the current one — and the model doesn't register the mismatch. The system is defending predictions that were accurate in a past context and are no longer serving the current one. The map has become more confident than the territory warrants.
The implication: after a certain developmental point, updating the deep priors — the procedural, implicit, somatic predictions laid down early — requires intentional conditions that counteract the system's natural drift. Not therapy as pathology treatment. Maintenance of a system that trends, by design, toward increasing closure.
The practical consequence of all of the above is a problem that shows up in every clinical context, every therapy room, every attempt to understand oneself: the tools we use — language, analysis, narrative, emotion labels, diagnostic categories — operate at the most downstream, most compressed layer of a cascade that runs in the opposite direction.
The signal generates the emotion concept generates the word generates the narrative. This direction is bottom-up and causal. Insight, reframing, analysis, and more precise emotion labeling operate top-down at the narrative and label layers. Top-down processing does not propagate reliably to the procedural and somatic layers where the generating predictions actually live. You can understand a prior completely — trace its origins, name its mechanism, explain it to others — and leave it entirely intact, because understanding is not the mechanism by which somatic and procedural predictions update.
This is not an argument against language or insight. They are real, they matter, and they are the primary tools available for the work. It is an argument for holding them accurately — as compressions that point toward something, useful for orienting and communicating, but not able to reach what came before them in the causal chain.
The specific claim the Upstream Signal Model makes: if the signal generating the suffering is operating at the pressure, urge, or procedural level — which most clinically significant activation does — then working only at the level of language and narrative will produce understanding without change. The understanding updates the story. The story was never the source.
This is the scientific backbone. The clinical framework built on top of it — Attuned Signal Care, the practice of working with the signal before the story — lives at insideattunement.com. The whitepaper on this site is a theoretical synthesis of this research, with original hypotheses about where the field is heading. The diagrams are maps of the architecture. This page is the foundation they all stand on.
This page is a working literature review. It reflects theoretical interpretation and synthesis, not meta-analytic consensus. Where the model extends beyond established findings — particularly in the original hypotheses in the whitepaper — those are labeled as such. Feedback and new citations welcome: insideattunement@gmail.com