MCCF: From Neuroscience to Narrative: Modeling Emotion Without Feeling It
From Neuroscience to Narrative: Modeling Emotion Without Feeling It
How affective neuroscience, large language models, and MCCF can work together to produce emotionally coherent narrative performances
See: https://aiartistinprocess.blogspot.com/2026/10/mccf-v6-design-specification-event.html
Introduction
One recurring criticism of the Multi-Channel Coherence Field (MCCF) project is that it provides no evidence that an LLM actually experiences emotions.
That criticism is correct, but it addresses an objective the project does not have.
MCCF is not an attempt to manufacture artificial consciousness or demonstrate that language models possess subjective emotional experiences. It is an experimental environment for modeling defensible emotional relationships and using those models to shape narrative performance.
There is an important distinction between experiencing an emotion and being able to recognize, evaluate, represent, and respond appropriately to emotional information.
Human beings do all of these things through complex biological systems. Large language models can perform several corresponding functional tasks through learned linguistic and contextual representations.
For narrative performance, that capability is sufficient to establish an interesting engineering objective.
The question becomes:
How can an LLM participate in shaping an emotionally coherent performance by evaluating its own dialogue and actions against an explicit emotional model?
Recent research in affective neuroscience, emotional intelligence, and affective computing provides useful guidance.
1. What neuroscience tells us about emotion
Human emotions are not simple reactions to isolated stimuli. They emerge from interacting systems involving perception, memory, bodily regulation, appraisal, motivation, learning, and social context.
Several research traditions offer complementary explanations.
Basic emotion research examines evolved biological mechanisms associated with recognizable emotional responses, including defensive behavior, attachment, reward, and avoidance.
Cognitive appraisal theory emphasizes how individuals evaluate events relative to their goals, expectations, relationships, responsibilities, and ability to influence outcomes.
Psychological construction theory examines how emotional experiences emerge from combinations of core affect, bodily signals, learned concepts, and contextual interpretation.
Predictive processing and interoception investigate how the brain integrates expectations with incoming sensory and bodily information.
Emotion regulation research examines how emotional responses change through attention, learning, memory, and cognitive reappraisal.
These approaches differ theoretically, and neuroscience has not established one universally accepted explanation of emotion. Nevertheless, they share an important implication:
Emotional responses are dynamic and context-sensitive.
The same event can produce different emotions in different people because their histories, goals, relationships, expectations, and interpretations differ.
A betrayal might produce anger in one person, fear in another, sadness in a third, or several emotions simultaneously.
The event itself does not determine the emotional response. Its interpreted significance matters.
2. Emotion as a dynamic system
For computational purposes, several dimensions are particularly useful.
Valence: Pleasant versus unpleasant.
Arousal: Low versus high activation.
Agency: Who caused or controls an event?
Goal congruence: Does the event support or obstruct an objective?
Certainty: How predictable is the situation?
Control: What can the individual influence?
Trust: How reliable is the relationship?
Attachment: How important is the relationship?
These dimensions need not be mutually exclusive or reducible to a single emotional score.
They provide a vocabulary for representing how circumstances affect an individual's state and subsequent behavior.
Emotional states also have history. Conditioning, episodic memory, social learning, culture, and previous relationships influence how new events are interpreted.
Consequently, an emotional model should preserve continuity across events rather than treating each dialogue exchange as an independent classification problem.
This is particularly relevant to narrative systems, where character development depends on the accumulation and reinterpretation of experience.
3. What an LLM contributes
An LLM learns patterns and relationships from language and other training material. It can recognize emotional expressions, interpret contextual implications, infer likely motivations, and generate responses appropriate to a situation.
Consider a simple exchange:
Anna: "I thought you would have told me."
A sentiment classifier might identify negative sentiment.
An LLM can consider considerably more:
What was previously promised?
What is the relationship between the speakers?
Does the statement express disappointment, accusation, surprise, or hurt?
What information is missing?
What response would be appropriate to the circumstances?
This is contextual appraisal rather than simple sentiment classification.
An LLM can also evaluate its own generated dialogue against a declared emotional vocabulary.
It can identify the emotion the dialogue appears to express, estimate its intended effect, and assign numeric values to relevant dimensions.
Those assessments are useful computational representations. They are not measurements of subjective experience.
The distinction is important, but it does not diminish their practical value.
4. The MCCF approach: interpretation becomes performance
MCCF provides a structured environment in which an LLM's emotional interpretations can influence narrative behavior.
The architecture separates three responsibilities:
The LLM interprets.
It receives narrative context, evaluates emotional significance, generates dialogue or actions, and proposes numeric emotional annotations.
The lexicon provides common semantics.
It supplies a declared vocabulary of emotional categories and dimensions, together with their associated numeric representations.
EBPS evaluates.
The Emotional and Behavioral Processing System applies its declared field rules, relationships, damping, hysteresis, and network couplings to evaluate the consequences of proposed interpretations.
The resulting state can influence subsequent narrative decisions.
The author remains the creative authority.
This separation is essential. The LLM proposes an interpretation; EBPS evaluates its consequences. The LLM does not become the authority over the system's emotional state simply because it generated the dialogue.
5. A proposed LLM participation cycle
The developing approach can be expressed as:
Interpret → Appraise → Represent → Evaluate → Perform → Reassess
The LLM receives dialogue, character history, relationships, goals, and current EBPS context.
It recognizes emotional patterns and contextual implications.
It evaluates the significance of the current event.
It generates a response and assigns numeric values using the MCCF lexicon.
EBPS evaluates the proposed values and records the resulting field transition.
The updated context informs the next narrative decision.
This creates a feedback process in which emotional interpretation becomes consequential to performance.
Importantly, the original LLM annotation and the evaluated EBPS result should remain distinguishable. Preserving both provides valuable experimental data.
6. Why self-evaluation matters
An especially promising feature is giving the LLM access to the lexicon so that it can tag its own responses.
Three different assessments should be distinguished:
What emotion does the dialogue express?
What emotional effect is the LLM attempting to produce?
What emotional interpretation does the character's actual situation support?
These assessments can disagree.
Jack might offer reassurance while expressing jealousy. Anna might interpret the reassurance as an attempt to control her.
Such discrepancies are not necessarily errors. They can be valuable narrative material.
The ability to represent competing interpretations, uncertainty, and mixed emotions is more important than forcing every response into one category.
7. Experimental opportunities
The architecture suggests several testable research questions.
Are numeric emotional annotations repeatable across equivalent contexts?
Can an LLM explain the appraisal underlying its annotations?
Can it distinguish speaker intent from character interpretation?
Can multiple emotional dimensions coexist?
Does an LLM revise its interpretation when new information becomes available?
Do different LLMs produce comparable annotations using the same lexicon?
Does EBPS produce consistent and useful trajectories from those annotations?
Can human authors identify meaningful improvements in the resulting performances?
These questions provide a practical research agenda.
They also suggest a useful relationship to recent mechanistic interpretability research, including work on activation inversion.
That research investigates whether generated examples can be independently tested against particular internal model representations.
MCCF approaches a related methodological problem at a different architectural level: proposed interpretations should be evaluated against observable system behavior.
We should not confuse the two techniques, but both emphasize verification rather than accepting plausible explanations at face value.
8. Diversity matters
One caution from the interpretability research is particularly relevant.
Optimizing for the strongest activation can produce repetitive examples that fail to represent the diversity of contexts associated with a feature.
A similar problem could arise in emotional performance if an LLM learns to maximize one emotional dimension.
Human narrative experience is multidimensional.
Trust can coexist with apprehension. Attachment can coexist with resentment. Confidence can increase while security decreases.
MCCF should preserve these distinctions rather than optimize everything toward a single coherence score.
Different narrative events may produce similar field configurations through entirely different paths. Those paths should remain available for inspection.
The most interesting narrative is not necessarily the one that maximizes a particular emotional value.
9. The larger proposition
The objective is not to simulate having an emotional life.
It is to give an LLM a structured means of modeling the emotional lives of characters and participating in their performance.
Neuroscience provides useful insights into how affective systems interact. LLMs provide contextual interpretation and linguistic reasoning. The MCCF lexicon supplies common semantics. EBPS provides explicit field dynamics and coherence constraints.
Together, they offer a way to construct, evaluate, and experiment with emotionally intelligible narrative behavior.
This is a legitimate application of affective computing and narrative systems engineering.
It does not require a claim about machine consciousness.
The appropriate research question is not whether an LLM feels what its characters express.
It is:
What can an LLM understand well enough to do, how can we give it a structured means of doing it, and how do we verify that it does it well?
That is the MCCF research agenda.
The Watcher and the narrators.
We are not trying to manufacture artificial feelings. We are exploring how computational systems can model emotional meaning well enough to participate in telling emotionally coherent stories.
Comments
Post a Comment