MCCF: Dynamic Affective Landscapes in Federated Narrative Systems

A research and implementation proposal · October 2026B

A tag in any of these layers couples the text to the curve, Prompt to the curve, Semantics are coupled fields of independent intensity, Trajectory as vector pressure where pressure is determined by the ebps engine power over semantic evolution,

It's just a model. Makes nice graphics. ;o and... music.  

cheers,

len

Executive summary

This proposal describes an experiment latent in MCCF’s original federation-dialog design: multiple heterogeneous large language models (LLMs) interact through MCP-mediated tools in one shared X3D narrative scene, while an explicit affective model and event system govern the evolving conditions of interaction.

The proposed extension is an emotional mixing board: EBPS parameters exposed as controls, with time-based automation lanes—potentially authored with NURBS curves—applied to any asset that exposes EBPS values. A more exploratory condition allows automation to modulate affective topology: the couplings among agents, fields, or relationships, rather than only the state of an individual asset.

The central hypothesis is that time-varying affect and coupling can alter the behavior of a federated narrative system even when the participating model weights remain fixed. The first experiment should therefore not fine-tune or retrain a single LLM. It should compare static federation, dynamic affect, and dynamic topology using the existing MCCF architecture and multiple models.

1. Motivation and research question

MCCF is not proposed as a game engine or as an attempt to make one model possess emotion. It is an experimental, declarative framework for narrative composition and performance: multiple agents, events, affective state, dialogue, sound, and 3D scene behavior are coordinated as a system. The LLMs remain distinct participants. MCCF supplies a shared scene and an explicit control structure in which their interactions can be observed and tested.

Primary research question: Can time-varying EBPS controls—and, separately, time-varying affective topology—improve the coherence, expressiveness, and adaptability of a federated narrative without modifying the participating LLMs?

• Does an automation curve create measurably different behavior from a static emotional setting or a one-time prompt instruction?
• Does changing coupling among agents produce effects that cannot be explained by changing each agent’s EBPS values independently?
• Can one authored emotional arc guide dialogue, animation, sound, and other EBPS-enabled assets without forcing them to share identical numeric values?
• Do results remain reproducible enough to compare conditions despite model stochasticity, latency, and different model tendencies?
• Can the same control abstraction support later batch-time experiments in curriculum design or fine-tuning, without conflating those with real-time orchestration?

2. Development background and design rationale

2.1 From federation-dialog prototype to MCCF

The project began with a federation-dialog prototype in which the human operator manually coordinated multiple LLMs through a shared blackboard, assigned roles and ordered prompts, and encouraged the models to question and respond to one another. The prototype served as a practical proof that the multi-agent idea could be built—not proof that the models shared an internal emotional state.

MCCF formalized that experiment into an executable environment: a dialog engine and event sequencer, a semantic/declarative composition layer, an emotional and social state model, and an X3D scene that can render assets, behaviors, dialogue, facial expression, and sound. The X3D model and production pipeline took roughly a year to develop. The important point is that the federation experiment preceded the finished 3D model; the experiment was latent in the architecture from the beginning.

2.2 Why “Multi-Channel Coherence Field”

The name points to coherence as a system-level property emerging across multiple channels—not merely a score attached to one utterance or one model. The channels include agent dialogue and behavior, emotional state, prediction, social relationships, events, and their visible or audible consequences in the shared scene. “Field” emphasizes that an agent’s state and action are situated within changing relationships and context. “Coherence” is the behavior to investigate, not a guarantee built into the name: it must be operationally defined and measured.

MCCF’s governing idea is lightly coupled interaction under explicit goals, events, and tests. The system should be constrained enough to study and direct, but not so tightly scripted that every outcome is predetermined. A catalytic participant or event can perturb the scene; the resulting responses can reveal the dynamics of the federation.

2.3 Why EBPS is fundamental

EBPS stands for Emotional, Behavioral, Predictive, and Social. It supplies a shared vocabulary for representing dimensions of an agent or asset’s state and for relating that state to behavior and interaction. It is fundamental because it offers the control surface between narrative intent and execution: values can be set, observed, changed by events, and potentially exposed to other assets and tools. Without such an explicit state model, emotional direction risks remaining implicit in prose prompts and difficult to inspect, compare, or reproduce.

In earlier MCCF exploration, EBPS developed from simple field and agent concepts toward trust, hysteresis, asymmetry, topology couplers, and predictive feedback. A later interpretation proposed by Claude—Eros, Bathos, Pathos, and Sophia—is a poetic/semantic reinterpretation, not the canonical expansion of the EBPS acronym. This proposal retains Emotional, Behavioral, Predictive, Social as the formal model and treats any alternative mapping as a separately documented semantic profile.

The precise operational semantics, units, bounds, and update rules of each EBPS dimension must be documented per implementation version. The labels alone do not establish that a model or asset literally experiences emotion.

3. The emotional mixing board

The UI borrows a familiar idea from digital audio workstations. Sliders set a baseline or current parameter value; automation lanes describe how the value changes over scene time. The same general mechanism can be applied to any asset that declares EBPS parameters—not only dialogue or characters, but potentially music themes, environmental effects, agents, and relationships.

 

Figure 1. Conceptual EBPS automation lanes. Curves are illustrative, not measured MCCF output.

3.1 Control model

Let the EBPS state of asset i at time t be:

aᵢ(t) = [Eᵢ(t), Bᵢ(t), Pᵢ(t), Sᵢ(t)]ᵀ

A static slider establishes a baseline bᵢ. One possible automation policy is additive modulation:

aᵢ(t) = clip(bᵢ + gᵢ(t), aᵢ,min, aᵢ,max)

A multiplicative policy is another option where it makes semantic sense. The project should not silently choose one universal combination rule: each lane should declare whether it overrides, offsets, scales, or otherwise transforms the baseline. Clipping, interpolation, units, missing values, and behavior at scene boundaries must be explicit.

For a NURBS-authored lane, the curve is a function C(u) of parameter u. Scene time t must be mapped to u, for example u = τ(t), before sampling. NURBS control points and weights give the author useful shape control, but the parameter itself is not inherently time. The implementation should define monotonic time mapping, endpoint handling, value limits, and how events sample or interpolate the curve.

3.2 Assets and lanes

• Character: a selected EBPS dimension may guide dialogue, gesture, facial expression, or response selection.
• Dialogue line: a mini-arc can shape delivery across the line, while a scene-level arc supplies an overarching contour.
• Music or sound: EBPS controls may guide an asset’s emotional contribution or trigger selection; this does not mean arbitrary EBPS values directly control audio unless an explicit mapping exists.
• Relationship or agent coupling: a lane can alter an exposed interaction coefficient or influence, subject to defined topology semantics.
• Scene or event: an arc may provide a shared dramatic envelope while individual assets interpret it through their own mappings.

The editor should remain generic: it needs parameter metadata, domain and bounds, interpolation, sampling, and dispatch. The asset or subsystem defines what a parameter means and how it affects behavior.

4. Dynamic affective topology

Figure 2. Conceptual coupling network. Edge thickness illustrates possible variable coupling, not current implementation values.

State modulation changes an agent’s values while leaving the interaction structure unchanged. Topology modulation changes the couplings through which agents or fields influence one another. Candidate controls include coupling strength, trust influence, relative field influence, or sensitivity to perturbation—but only where MCCF’s actual topology exposes those quantities.

This is the more experimental condition. It may reveal collective effects that cannot be obtained by independently changing each agent’s EBPS values. It also creates stability risks: time-varying feedback may cause oscillation, saturation, abrupt transitions, or sensitivity to timing. Start with bounded, slowly varying coefficients and log every topology change. A null or fixed-topology baseline is essential.

da/dt = F(a(t), W(t), e(t), c(t))

Here a(t) denotes the collection of agent/asset states; W(t) is a time-varying coupling or topology parameterization; e(t) denotes events; and c(t) denotes relevant scene context. This is a conceptual dynamical form, not a claim that MCCF currently implements this differential equation. A discrete-time update may be more appropriate for the actual event sequencer.

5. Federated LLM experiment architecture

Figure 3. Conceptual federation architecture: separate models interact through MCP-connected tools around shared MCCF state and a shared X3D scene.

The experiment uses multiple LLMs as distinct agents in the same scene. MCP provides tool interfaces and access; MCCF remains responsible for the authoritative shared state, turn/event scheduling, parameter evaluation, and scene consequences. The X3D/X_ITE environment makes the resulting dialogue, behaviors, animation, and sound inspectable.

MCP alone does not guarantee shared state, synchronization, turn-taking, or semantic agreement; these must be specified by MCCF. The first study should not modify model weights. The question is whether a dynamic environment changes the behavior of the federation. Models may have different capabilities, styles, latency, and stochasticity; those are both sources of experimental noise and potentially meaningful differences.

6. Experimental conditions and goals

6.1 Conditions

• Condition A — Static federation: multiple models share the scene, with fixed EBPS parameters and fixed topology.
• Condition B — Dynamic affect: EBPS values are driven by automation lanes while topology remains fixed.
• Condition C — Dynamic topology: selected coupling parameters vary over time; EBPS automation is held fixed or separately controlled to isolate the topology effect.
• Condition D — Combined control (follow-up): both affective state and topology vary, allowing tests of interaction effects.
• Optional reference condition — Prompt-only direction: a fixed or changing textual instruction without explicit MCCF lane control, to test whether explicit state/automation adds value beyond prompts.

6.2 Goals and hypotheses

• H1: Dynamic EBPS automation produces distinguishable emotional and behavioral trajectories compared with static settings.
• H2: Dynamic topology changes inter-agent influence and yields effects not reducible to independent per-agent value changes.
• H3: A shared authored arc can coordinate dialogue, performance, and scene assets while allowing asset-specific mappings.
• H4: A lightly coupled federation can recover from a perturbation and maintain narrative continuity better under some dynamic control conditions than under static controls.
• H5: Explicit curves and logged state make the experiment more inspectable and reproducible than prompt-only affective direction.

These are hypotheses, not expected outcomes. A dynamic control may worsen coherence, over-constrain agents, amplify model-specific quirks, or add complexity without measurable benefit. Null results are useful.

7. Measurement and experimental method

Use a small scripted scene with clear initial conditions, a few defined events, and at least one perturbation. Keep the task and evaluation rubric stable across conditions. Repeat runs because model outputs vary.

• Narrative continuity: consistency of facts, goals, causal links, and prior events.
• Character continuity: whether behavior changes plausibly given the agent’s history and relationship state.
• Affective trajectory fidelity: how closely observed or annotated behavior follows the intended curve, without treating the labels as objective emotional truth.
• Responsiveness: whether agents react appropriately to another agent’s action or a scene perturbation.
• Diversity and emergence: whether the system permits distinct, non-scripted responses while remaining within the task’s constraints.
• Recovery and stability: oscillation, runaway coupling, saturation, deadlock, incoherent turn loops, or inability to recover after a perturbation.
• Operational measures: latency, number of model/tool calls, failure rate, token/cost use, and missed or reordered events.
• Human evaluation: blinded reviewers score coherence, expressiveness, and plausibility; preserve disagreements rather than collapse them prematurely into a single score.

Record the exact model/version, prompts, tool schemas, scene state, curve definitions, event order, timestamps, topology coefficients, outputs, and any retries. Distinguish deterministic MCCF behavior from nondeterministic model responses. Define coherence operationally before evaluating results; do not use the project name as its own metric.

8. Real-time versus batch-time extensions

8.1 Real-time orchestration (primary study)

Automation changes the context and operating conditions under which the models act. It may affect prompt construction, memory/retrieval selection, tool arguments, agent-specific state summaries, or—if the model and runtime permit it—internal activation steering. The first implementation should prefer mechanisms already available in MCCF and MCP. Activation steering is optional research, not a prerequisite.

8.2 Batch-time learning (later study)

A training corpus is fixed for a given training run, but its ordering, sampling, weighting, and curriculum can vary during training; the model’s weights are updated by optimization. A later study could compare training or fine-tuning on narrative examples ordered by emotional trajectories against matched examples without that structure. This is separate from real-time federation and should not be claimed as a consequence of the primary experiment.

Potential questions include whether trajectory-aware curricula improve narrative transitions or character continuity, and whether they also cause overfitting, style narrowing, or forgetting. Training is not necessary to test the central MCCF hypothesis.

9. Caveats, unknowns, and threats to validity

• Construct validity: EBPS labels are engineering constructs, not proof that a model or character has subjective feelings. Define each parameter operationally.
• Causal attribution: if prompts, curves, topology, model choice, and event order all change together, the source of an effect cannot be identified. Change one factor at a time before testing combinations.
• Semantic mapping: identical numeric values may mean different things for different asset types or models. Use asset-specific parameter metadata and mappings.
• Curve semantics: specify baseline combination, interpolation, time mapping, bounds, discontinuities, and event sampling behavior.
• Topology stability: changing couplings may produce unstable or path-dependent behavior. Use bounded changes, safe defaults, and complete logging.
• Model variability: sampling, provider updates, context limits, tool-use behavior, and latency can confound comparisons. Record versions and repeat runs.
• MCP boundaries: tool protocols do not themselves guarantee synchronization, shared memory, or reliable event ordering.
• Evaluation bias: human judgments of emotional coherence are subjective; use multiple raters, explicit rubrics, and preserve uncertainty.
• Over-control: if automation dictates every response, emergent character behavior may disappear. Treat curves as targets or envelopes where possible, not necessarily hard commands.
• External validity: a successful scene demonstration does not establish general performance across genres, model families, or long-running narratives.
• Scope: dynamic fine-tuning and activation steering should remain separate follow-on work; neither is required for the initial experiment.

10. Proposed implementation sequence

1. Document the current EBPS schema: parameter meanings, ranges, defaults, state ownership, and update rules.
2. Add lane metadata and automation evaluation to the events editor, initially for one or two EBPS dimensions on one asset.
3. Implement and test baseline-combination modes, time mapping, bounded interpolation, and deterministic logging.
4. Run the static federation baseline with multiple MCP-connected models in one shared scene.
5. Add dynamic EBPS automation while keeping topology fixed; compare outcomes.
6. Expose one bounded topology/coupling parameter and run a controlled topology experiment.
7. Evaluate the combined condition only after the isolated effects are understood.
8. Publish the scene, curve data, event scripts, model/configuration notes, evaluation rubric, and limitations to support replication.

11. Expected contribution

The proposed contribution is not a claim that emotional curves make an individual LLM emotional. It is a testable architecture for authoring and studying dynamic affect in a federated narrative system. Its central innovation is to make affective state and, experimentally, inter-agent coupling explicit, time-varying, inspectable controls in a shared environment.

If successful, the work would show that the conditions of interaction can be dynamically authored without retraining the participating models. If unsuccessful, the experiment should still identify which effects are attributable to prompts, explicit state, event scheduling, or topology—and whether the added control surface is worth its complexity.

Conclusion

MCCF’s original federation-dialog experiment asked what distinct models could do when they interacted under shared goals and constraints. The X3D scene made that interaction visible; EBPS provided an explicit affective and social state vocabulary; the event system supplied temporal structure. An emotional mixing board extends those existing ideas by making the state of any EBPS-enabled asset automatable, and by opening a carefully controlled experiment into dynamic affective topology.

The first target is a federated scene with frozen model weights—not a retrained model—and the principal scientific task is to measure whether changing the landscape changes the collective behavior.