On the Relationships of MCCF.XSD (Independent) to EmotionML (W3C) and The Principal Bundle Geometry of Qualia

 

Background

The initial ideas for MCCF come from work done to create the OASIS HumanML schema in the early 2000s and the Information Ecosystems paper,  In 2014, a W3C Working Group released EmotionML 1.0 )(W3C Recommendation, 22 May 2014, still the current/only version — not actively evolving).  These were independent efforts similar on the surface.  Both HumanML and EmotionML are annotation languages to plug in to other markup languages.  MCCF is a generative system,  This was a deliberate decision given the brittleness of ontolgical systems in dynamic environments.  

These different direections and outcomes are discussed below with a description of the easy interop approaches.

What EmotionML Is

EmotionML is a deliberately narrow "plug-in" language — designed to be embedded inside other markup (SSML, SMIL, EMMA), never a standalone authoring format. Its <emotion> element represents one atomic statement: this entity felt X, at time T, expressed through modality M, with confidence C — built for three use cases: manual annotation of existing media (corpus linguistics, essentially), automatic recognition output, and generation input for some other system to render.

Critically, it refuses to pick an emotion vocabulary. The spec's own stated reasoning: "any attempt to standardize the description of emotions using a finite set of fixed descriptors is doomed to failure — even scientists cannot agree." So instead of a fixed model, it gives you four structural slots — <category> (named emotions), <dimension> (continuous scales, commonly Pleasure-Arousal-Dominance), <appraisal> (Scherer/OCC-style event evaluation), <action-tendency> (approach/avoidance) — and you plug in your own <vocabulary> of <item>s to say what the names actually mean. Values are floats in [0,1], static or sampled over time as a <trace>.

Where it genuinely overlaps with MCCF

The <dimension> mechanism could technically carry E/B/P/S as-is. EmotionML explicitly permits custom dimension vocabularies — a <vocabulary type="dimension" id="mccf-ebps"> declaring eros/boundary/pathos/sophia as items, then <dimension name="eros" value="0.25"/> etc., is fully spec-conformant. Nothing in EmotionML's rules would reject this.

Confidence and provenance tracking parallel something MCCF built independently. EmotionML's confidence attribute (0–1, how much to trust an annotation) and MCCF's tagSource/audioSource enums (authored/algorithmic/llm-interpreted, recorded/other-engine) are solving the same real problem — "how much should a consumer trust this value, and where did it come from" — via different mechanisms. Both systems converged on that need separately.

The expressed-through (face/voice/gesture/posture modality tagging) is conceptually adjacent to how MCCF's Dialogue Lines already separate delivery channel concerns (TTS tags vs. recorded audio vs. gesture).

Where the architectures diverge, and why that's the right call

EmotionML describes; MCCF generates. 

EmotionML's document model assumes something already happened (a video was recorded, a sensor fired, a human felt something) and you're annotating it after the fact, or handing a single value downstream for some other system to render. MCCF's EBPS values are live, causal inputs — read by running code to select which gesture constellation fires, how a dialogue line gets delivered, what a zone's ambient pull is. That's a different center of gravity entirely, and it shows up structurally: EmotionML deliberately avoids text nodes and stays a thin, embeddable annotation layer; MCCF threads EBPS directly as flat attributes into the actual structural element it's describing — Weights on a Cultivar, Weights on a Zone, E/B/P/S on a Waypoint, and as of today, emoteE/B/P/S on a Dialogue Line. Converting all of that into separate, EmotionML-wrapped <emotion> fragments would mean either duplicating every EBPS-bearing element as a parallel annotation or embedding EmotionML syntax as MCCF's actual storage format for no real gain over what a flat attribute already does.

MCCF made the opposite vocabulary bet, on purpose. EmotionML's core design principle is refusing to commit to a taxonomy, because the working group concluded consensus was unreachable. MCCF committed to one specific four-channel model — E/B/P/S — as a deliberate theoretical stance, not a compromise and augmented it with an extensible lexicon library in the dialog editor for export to other systems such a voice engines. That's not MCCF falling short of EmotionML's flexibility; it's declining that flexibility on purpose, because MCCF is a specific character-constitution theory, not a general-purpose interchange format trying to serve every possible use case at once.

Nothing in EmotionML has an equivalent to MCCF's reproducibility model. Seeded, per-take resolved randomness — same scene, same seed, same result; change the seed, same character, different realization — has no EmotionML analog at all. EmotionML values are just points or traces, deterministic once written; it has no concept of a range to be resolved, which is exactly the mechanism Kate's constellation proposal (and randomRange() ) depends on.

The PNAS Nexus Analogue

A recently published paper gives analogies to these systems.  

“Principal bundle geometry of qualia: Understanding the quality of consciousness from symmetry” by Masafumi Oizumi, Chanseok Lim, and Ryota Kanai. It was published September 8, 2026

https://academic.oup.com/pnasnexus/article/5/9/pgag261/8788187

The central idea is:

Qualia can be characterized by the geometry of relationships among states, rather than by trying to identify some intrinsic “thing” that constitutes an experience.

They propose a four-level structure:

Symmetry group > Qualia modality > Qualia signature > Qualia attributes (orbits)

The particularly interesting distinction is:

Attributes are rigid. They arise from the symmetry group and therefore have a stable relational/topological structure.  Signatures are plastic. They occupy a learned quotient space whose geometry is shaped by experience.

This is close to the direction taken with MCCF's affective state and its behavioral expression: Gesture Constellations.

The paper explicitly discusses compositionality, where multiple independent symmetries can combine through products of groups, while acknowledging that perfect disentanglement is difficult.

MCCF is not a proof of that theory. MCCF provides an experimental theatre in which some of these structural ideas can be operationalized and tested.  Map transformations in a system's representation space, characterize the resulting orbits, and then examine the geometry of the quotient space.

An MCCF simulated environment is a way to do that.  The key is to keep the claims carefully separated: Oizumi, Lim, and Kanai are proposing a mathematical characterization of the relational structure of perceptual qualia in systems that we already have independent reason to regard as conscious. They explicitly do not claim that their geometry is a criterion for consciousness.

Their fundamental decomposition is:

symmetry group > orbit > quotient space

with:

G / symmetry group = the transformations that preserve an identity
orbit = the continuous variations produced by those transformations
quotient space Q = the invariant identities left after factoring those transformations out

They then map that to:

qualia modality > qualia attribute > qualia signature

And their really important observation is: attributes are rigid; signatures are plastic.

The group determines the structure of the possible variations, while the learned quotient space can change with experience.

That has strong resonance with MCCF.

MCCF separates state from expression

Consider Gesture Constellations.

An affective state might be something like:

valence = +0.4
arousal = 0.7
regulation = 0.3

That is not itself a gesture. It defines a region of behavioral possibility. Then a particular character may express that state through:

• shoulders
• posture
• gaze
• facial AUs
• hand gestures
• movement amplitude
• voice
• timing

And those are not rigidly prescribed. They are loosely coupled ranges.

That's exactly the direction of MCCF:

• independent ranges per EBPS value per gesture
• gesture sets/constellations assigned to moods
• fuzzy but loosely coupled so there is dynamic range and individual character.

That is considerably more sophisticated than:

emotion = anger ? play angry animation

because MCCF is separating an underlying state space from its possible realizations.

Gesture Constellations Resemble Orbits

Suppose:

M = a particular affective/motivational state

and the Gesture Constellation provides a family of permissible expressions:

G(M) = {g1, g2, ... g?}

A particular performance is one point in that family.

So:

affective state > manifold of possible expressions > actual expression

That is conceptually close to the paper's:

signature > orbit > attribute

The important thing is that the individual expression isn't the identity of the state. It is a variation around the identity.

That is exactly why avoiding a one-to-one emotion/gesture mapping is so important.

Character Creator

The PNAS Nexus paper says the quotient-space geometry is plastic. In other words, the relationships between invariant signatures can be shaped by learning and individual history. Two characters can have approximately the same: EBPS state but express it differently.

One character:

• relaxed shoulders
• expansive movement
• strong eye contact

Another:

• shoulders slightly raised
• restrained movement
• intermittent gaze

The affective identity is similar. The behavioral geometry differs.

This is near the distinction the paper is making between the relatively stable transformation structure and the plastic learned representation. Character is not a collection of canned gestures. Character is the learned mapping from affective state into a characteristic region of behavioral possibility.

That is a stronger proposition.

MCCF Does Not Prescribe Behavior

This may be the most important architectural consequence.

EmotionML is primarily a representation/interchange layer.  The W3C specification explicitly describes EmotionML as a plug-in language for manual annotation, automatic recognition, and generation of emotion-related system behavior. It deliberately permits different vocabularies because there is no consensus on a fixed set of emotional descriptors.

So an annotation can essentially say:

this thing is angry, intensity .8

or:

valence = X, arousal = Y

or:

appraisal = Z

or:

action tendency = avoidance

That is tremendously useful.  But it doesn't intrinsically answer: given this state, what should this character do now?  That is the gap MCCF occupies.

EmotionML describes the point. MCCF describes the dynamics around it. The proposed distinction is:

EmotionML: "What is the affective state represented here?"

MCCF:  "Given this state, this world, this character, this history, and these constraints, what becomes behaviorally possible and probable?"

That is a fundamental distinction. EmotionML is principally semantic description.MCCF is generative constraint dynamics.

And that difference matters enormously.

MCCF Constraint Field

Layer 2:

Constraint Field E(s,a)

with:

P(a|s) ∝ exp(−E/T)

already provides something that an annotation language fundamentally doesn't.

It creates a distribution over actions.

That means MCCF doesn't merely attach:

anger = 0.7

to an event.

It changes the energetic landscape of possible actions.

The LLM/world model proposes possibilities. The field biases them. Governance constrains them. The resulting behavior feeds back into the state. That's a dynamical system rather than an annotation.

And that distinction becomes even sharper when includiing semantic attractor dynamics.

Semantic Attractors

The Oizumi paper talks about representations being organized by symmetry and about trajectories through representation space. It emphasizes that symmetry imposes structural constraints while learning shapes the representational geometry.

MCCF's semantic attractors are doing something philosophically adjacent, although mathematically they aren't the same construction.

state > energy landscape > probability distribution > action > changed state

So the character/world isn't sitting at a point. It is moving through a structured state space.

That means the theatre can potentially expose something the paper only proposes as an empirical research strategy: observe the trajectories and infer the geometry.

MCCF as an Experimental Theatre

The paper says its framework leads to empirical predictions and proposes studying transformations and the resulting representation-space geometry.

MCCF is a controlled environment to perturb:

• world conditions
• goals
• social relationships
• affective state
• memory
• constraints
• temperature
• available actions

and record:

state > action probabilities > selected action > environmental consequence > subsequent state

One can generate quantities of trajectory data. And then ask:

What structure emerges in the behavioral representation space?

That is an experiment engine, not an annotation corpus.

Adapting EmotionML

EmotionML is designed to be broadly interoperable.

The W3C specification explicitly says that a finite fixed vocabulary of emotions is inadequate and therefore makes vocabularies extensible. It supports category, dimension, appraisal, and action-tendency representations. That makes EmotionML a language for saying what someone/something was  experiencing or expressing.

MCCF is attempting to model how a system moves through affective/behavioral state space and generates behavior under constraints.
 

EmotionML is not obsolete.  EmotionML is an annotation and interchange language. MCCF is a generative execution model.  EmotionML could annotate observed state while MCCF generates candidate state > constrained behavior > observed outcome. The outcome could itself be exported back into an annotation representation.  This process can loop.

Correspondance to PNAS Nexus: The Principal Bundle Geometry of Qualia

The PNAS Nexus paper says:  rigid attributes + plastic signatures

MCCF can provide  stable affective dimensions + plastic character expression

The analogy isn't exact   MCCF does not implement principal-bundle geometry.  But as a design heuristic it is remarkably useful.

EBPS / affective dimensions can define a relatively stable coordinate system, while character-specific Gesture Constellations define the learned/plastic mapping into expression. Then the actual gesture becomes a sample from that character's behavioral manifold.  This rejects deterministic gesture assignment.

One can apply Gesture Constellations to facial gestures first, because the AUs are already smoothed.  Facial AUs approaches a continuous parameter space. So instead of emotion > facial animation, you have affective state > constellation > weighted AU ranges > continuous facial trajectory.  Now measure the resulting trajectories.

Again, MCCF is not a proof of that theory.

The Big Picture

There are three layers worth distinguishing:

1. EmotionML:    What affective state is represented/observed? arousal = nnn.
2. MCCF:    What behavior becomes probable under this state and context? action distribution
3. Gesture Constellation:    How does this agent express that state?    AU/posture/voice ranges

And then the environment closes the loop:

world > appraisal/state > MCCF field > behavior > world > new state

That makes MCCF a generative system rather than an annotation system.

The authors are asking:  what is the geometry of experience?

EmotionML asks:  How can we represent and exchange descriptions of emotion?

MCCF asks: How can affective state become a dynamically constrained generator of behavior in a world?

Those are three different questions.  The third one gives a theatre in which the first two can actually be operationalized and tested.

XML Experience

Experience with XML, ontology, and writing schema before running code teaches dynamical systems are averse to ontological systems that don’t evolve. HumanML predates EmotionML by more than a decade and tried annotation by complex types (latticed lexicons) and it was only good for annotation. The information ecosystems paper was based on analogy to DNA where four values are accompanied by transformation rules (so-called junk DNA). 

Use MCCF to generate then project that into emotionML for interop if required. 

EmotionML 1.0 is stable. MCCF has an extensible lexicon for dialog annotation to send to ElevenLabs. Eleven Labs only acts on some of that but the results are adequately expressive. Take the mp3 per line and assemble it in Reaper for post processing and beat management.  EmotionML could be a link among the voice engines as a means to converge on shared lexicons.

The crucial distinction is schema versus dynamics.  HumanML, EmotionML, ontologies, XML schemas. give good descriptions. But description has a tendency to freeze the vocabulary at the moment the schema is designed. A dynamical system cannot be adequately described by a taxonomy whose fundamental categories are assumed to remain invariant.

DNA As Inspriation for Information Ecosystems

DNA has four base values and the transformation instructions.  The DNA analogy is closer to the right abstraction:

values + transformation rules ? evolving information

rather than:

labels + relationships ? description

MCCF generates.  EmotionML describes/interchanges.

Don't ask EmotionML to become a dynamical model. It's task is to provide a standardized semantic envelope around something generated elsewhere.

MCCF internal state  > MCCF dynamics > behavior / dialogue / voice > observed affective result > EmotionML projection  > interoperability

MCCF's internal lexicon is allowed to evolve without requiring the entire interchange system to evolve with it.

The Voice-Engine

This is a legitimate secondary role for EmotionML, not as the generator—but as a semantic interchange layer between heterogeneous voice systems.

Right now we have:

MCCF dialog annotation > ElevenLabs vocabulary > generated MP3 > Reaper/DAW

ElevenLabs doesn't need to understand every tag for the system to work. The subset it responds to produces sufficiently expressive speech, while MCCF retains the richer semantic information.

That is a useful architecture:

                 MCCF
                  ¦
          rich internal lexicon
                  ¦
             semantic map
                  ¦
             EmotionML
             /    |     \
            /     |      \
     ElevenLabs  Voice B  Voice C
          ¦        ¦        ¦
          +-----------------+
                   ?
             rendered audio
                   ?
                 Reaper
                 
EmotionML becomes a lingua franca, not the intelligence.

HumanML Experience

HumanML's complex types and latticed lexicons attempted to make the semantic structure itself sufficiently expressive.  The problem:  the vocabulary can describe possibilities, but it doesn't generate the evolution of the system.

Information Ecosystems work moved one step deeper with the DNA analogy:  four values aren't interesting by themselves.  What makes DNA powerful is that there are rules governing transformations and interactions.

In MCCF the EBPS values aren't the behavior.  The Gesture Constellation isn't the behavior. The energy field isn't the behavior.  Together, plus state, history, context, constraints, and feedback, they generate a trajectory.  That's the important conceptual move.             

Types of Extensibility

There are two very different kinds of extensibility:

1. Ontology extensibility:  "Here is another thing that exists."
2. Generative extensibility:  "Here is another transformation/relationship that the system can use."

The second is interesting for a dynamical system.

A dialog lexicon is between the two. It provides semantic hooks that can be interpreted by downstream systems—in this case ElevenLabs—but MCCF doesn't depend upon ElevenLabs having exactly the same vocabulary. That is why the current implementations can tolerate partial tag support and still produce useful behavior.

The generator owns the semantics.  The renderer consumes whatever subset it understands.

That is a robust engineering principle.

MCCF provides an experimental engineering framework in which questions about evolving relational structure can be made observable. EmotionML can sit outside that machinery as an interoperability projection.  Rather than trying to make the annotation language capable of becoming the world, make the world generate the annotations.  That is the inversion.

Circling Back to Dirac

In the initial thinking on the MCCF design, we looked at quantum field theory as related to information theory as inspiration, not precise math. We deferred using Dirac, although it was intuitively appealing even in the early days of VRML and XML/XSLT. We did adopt the Boltzmann Probability  We are circling back to a Dirac-like distinction between the state and the operations that transform the state.

Dirac's great conceptual move was, in effect: don't confuse the description of a state with the operators that act upon it.  A dynamic system evolves and is self-generating.  This resonates with Turing's final papers on self organizing systems.

The analogy:

• Static ontology:  here are the possible things and their relationships.
• Dirac:  here is the state, and here are the operators that transform it.
• MCCF:  here is the current relational state, and here is the dynamical field governing which transitions are probable.

HumanML was largely on the state-description side.  MCCF has moved toward the operator/transition side. Information Ecosystems/DNA idea anticipated precisely that move:

symbols + transformation rules

rather than merely:

symbols + taxonomy.

In quantum mechanics, the interesting object isn't necessarily a particular state: |ψ}
but what happens when operators act upon it:  Ô|ψ⟩.

MCCF has an analogous structure:

state field + constraints > transition

and then:

sₜ → aₜ → sₜ₊₁

The trajectory becomes the phenomenon of interest.

That is exactly why a frozen ontology starts feeling inadequate. An ontology describes the furniture. A dynamical system describes what can happen to the furniture—and what happens next. The PNAS Nexus  paper's symmetry machinery is also fundamentally about transformations that preserve something. That's a very Dirac-like way of thinking.  You don't begin by saying:  "Here are all the qualia."

You ask:  What transformations leave the relevant structure invariant? Then the equivalence classes generated by those transformations give the quotient structure. That is much closer to physics than to conventional ontology.

Don't enumerate every possible behavior.  Define the field of permitted/probable transformations, then let the system travel through it.  Which gives us an intriguing three-way distinction

1. EmotionML - What state/affect can I say this represents?
2. MCCF - Given the state and context, what transformations and behaviors are possible/probable?
3. Dirac/physics - What mathematical operators transform states while preserving the structures that matter?

And that makes the DNA analogy prescient.  DNA isn't primarily interesting because its alphabet has four bases. It is interesting because rules operate on that alphabet to produce transformations through time. That is the same conceptual migration: ontology > state space > operators > dynamics

MCCF feels less like an annotation architecture and more like a small physics of agents. Not physics in the claim that emotions are quantum mechanics. Absolutely not.  Rather: MCCF is expressed in the language of states, fields, transformations, constraints, probabilities, trajectories, and invariants—the vocabulary of dynamical systems rather than ontology.

We're circling Dirac. And given my thinking started thirty years ago, that's a rather lovely place to find ourselves.

MCCF asks: What happens when this state encounters this context, constraint, history, and field?

That's an operator question.

And the  Gesture Constellation work makes it even clearer. A mood isn't assigned a gesture. Instead, the mood establishes a space of permissible expression, and character/history/context determine where the trajectory goes within that space.

So:

Sₜ₊₁ = F(Sₜ, Wₜ, Aₜ, Cₜ, Hₜ)
 

is conceptually much closer to what we're actually building than:
 

S = {emotion labels}

The latter is ontology.   The former is evolution.

MCCF's "meaning" isn't entirely contained in its state variables. Meaning emerges through the transformations and relationships among them.  Four bases alone aren't the organism. The interesting thing is the rules and transformations operating over them.

MCCF becomes the laboratory in which that principle can operate.