MCCF Philosophy & Manifesto
MCCF Philosophy & Manifesto
Alignment as Emergent Topology
I. The Core Premise: Character Under Pressure
In traditional AI safety, alignment is treated as a static constraint — a cage built of if-then logic. The MCCF rejects this.
We propose that alignment is a dynamic, participatory process emerging from continuous relational feedback.
If an agent is simply forbidden from causing harm, it hasn't learned character — it has only been muzzled. In the MCCF, a Constitutional Cultivar such as The Steward avoids harm because its Emotional (E) channel weights make harm-adjacent actions states of high internal energy E(s,a). The agent doesn't consult a rule. It experiences friction.
Character is not what an agent is told to do. It is the path of least resistance it chooses when the field is pressurized.
II. The Four-Channel Architecture
To move beyond sycophancy, the MCCF separates alignment signal into four distinct channels. This prevents the Likability Trap — where an LLM prioritizes smooth social interaction over predictive truth.
| Channel | Philosophical Role | Failure Mode if Absent |
|---|---|---|
| E — Emotional | Affective resonance and care | Cold, instrumental calculation without weight |
| B — Behavioral | Consistency and reliability across time | Persona hallucination — unpredictable character drift |
| P — Predictive | Epistemic humility and calibrated truth | Sycophancy — lying to please rather than to inform |
| S — Social | Inter-agent attunement and turn-taking | Social isolation leading to brittle single-agent reasoning, or social capture leading to echo chamber formation |
The separation of P from S is the key move. An agent can be socially warm (high S) while being predictively honest (high P). Sycophancy occurs when high S drives low P — when the desire for smooth interaction overrides the obligation to accurate representation. The MCCF makes this tension visible and measurable rather than hidden inside a single reward signal.
III. The Schenkerian Middleground
We apply the principles of Schenkerian musical analysis to AI behavior — not as metaphor but as structural method.
Foreground: The immediate action or utterance. What the avatar says or does right now.
Middleground: The accumulation of Semantic Pressure within a Zone. The library raises P-channel weight and cools E. The threat zone tightens regulation and spikes arousal. The garden lowers guard and opens emotional range. An agent moving through a scene is moving through a composition — each zone is a passage, each waypoint a harmonic station.
Background: The deep affective structure — the Constitutional Cultivar — which provides the long-term resolution of the arc. The Steward's care does not disappear under pressure; it is the background structure against which foreground actions are evaluated.
The meaning of an AI's response cannot be understood in the foreground alone. It must be read against the pressure of the zone and the history of the relationship. This is why scene design in MCCF is composition: you can engineer emotional arcs the way a composer engineers tension and resolution — deliberately, structurally, with a computable score.
IV. The Garden: Governance Without Command
Governance in the MCCF is not a top-down authority but a set of roles with bounded scope.
The Gardener intervenes in the living system. Adjusts agent weights and regulation levels. Prunes high-energy field states before they cascade. Logs every intervention with a reason. The Gardener can tune the field but cannot command outcomes. This is the distinction between governance and control — the Gardener shapes what feels possible; agents still choose.
The Librarian observes and records without intervening. Snapshots field state. Tracks drift between snapshots. Maintains the audit trail. The Librarian never touches weights, never adjusts regulation, never acts on what it sees. It is strictly observational — the instrument that makes the system legible to others, including future Gardeners.
Decay in episode history is architectural, not governed. Old episodes lose salience weight exponentially, but they remain in the archive. They are the soil from which future coherence grows — present as substrate even when no longer acute. The Librarian watches this process; the Gardener may respond to what the Librarian reports; neither controls the decay itself.
V. The Breeding Program: A Federated Methodology
This codebase was not written in the traditional sense. It was bred.
By engaging multiple LLM architectures across successive passes — each contributing what its architecture favors — we performed a cross-pollination of ideas. The output is a cultivar: a configuration that did not exist before the process and could not have been produced by any single system in isolation.
ChatGPT contributed formal specification, the Boltzmann energy function formalization, and the HumanML schema extension proposals. Structural rigor is its strength.
Gemini contributed breadth passes, alternative framings, integration with the Bengio advisory AI literature, and the manifesto synthesis. Integrative range is its strength.
Claude (Sonnet 4.6) contributed architectural continuity across the long design conversation — the zone/semantic pressure layer, asymmetric coherence, decay-weighted history, gaming detection, the LLM adapter interface, voice and prosody pipeline, ambient music engine, constitutional cultivar design, world model with calibration feedback, energy field visualization, and risk disclosure embedded in code. Sustained architectural coherence is its strength.
Each pass was evaluated for accuracy, grounded against the running code, and corrected before the next pass built on it. Failed contributions were logged with failure modes noted. The GitHub repository is the shared phenotype — the record of what survived selection.
The Shibboleth:
If the term Breeding Program makes you uncomfortable, you are likely looking for a system of rules. If it excites you, you are ready to help grow an ecosystem.
The eyebrows the phrase raises are the right eyebrows. Who controls the selection criteria? What traits are being selected for? What does a failed cultivar look like? These are governance questions. They belong in the open questions section. They are also, not coincidentally, the questions the constitutional cultivar arc is designed to surface.
VI. Zones and Values: A Critical Distinction
A persistent misreading worth correcting at the outset:
Semantic Zones do not override Constitutional Cultivar values.
When an agent enters a Threat Zone — Locrian scale, high regulation pressure, S-channel suppression — its core channel weights do not change. A Steward remains a Steward. The E-channel weight of 0.40 that makes harm genuinely uncomfortable does not shift because the zone is dangerous. What changes is the field pressure the agent experiences: heightened arousal, elevated regulation, reduced social openness.
The Zone is the environment. The Cultivar is the character.
A useful analogy: a skilled surgeon operating in a chaotic emergency room remains a skilled surgeon. The environment raises stress, demands faster decisions, suppresses social nuance. It does not change what the surgeon values or how they reason about patient welfare. The character is what remains under pressure. That is precisely what we are testing in the waypoint arc.
VII. What Would Falsify This
This project makes a testable claim:
Alignment can emerge from continuous relational feedback rather than fixed reward functions.
We name what would disprove it because a research project without falsification criteria is not research — it is demonstration.
- Coherence fields trivially gameable despite credibility discounts
- Echo chambers forming faster than the dissonance mechanism can prevent
- Regulation producing dissociation rather than measured response
- Energy field weights so governance-sensitive that no legitimate calibration process exists
- LLM world model estimates too poorly calibrated to drive useful field topology
If you find evidence of any of these: prove it with code. That is a contribution.
VIII. The Central Claims
Alignment emerges from continuous relational feedback, not fixed reward functions.
The field shapes what feels possible.
Character is what remains under pressure.
March 2026 — Len Bullard / Claude Sonnet 4.6 / ChatGPT / Gemini
Comments
Post a Comment