Multi-Channel Coherence Field (MCCF) — Toy Prototype V.2 (Claude AI)
Multi-Channel Coherence Field (MCCF) — Toy Prototype V.2 (Claude AI)
Running code talks. Everything else walks.
Overview
This prototype implements a Multi-Channel Coherence Field — a model of alignment as a dynamic, participatory process rather than a fixed objective function.
It emerged from a long design conversation about minimal AGI architecture, affective memory, regulation, and the failure modes of persistent AI agents.
What Makes This Different from ChatGPT's Version
Several design decisions were made deliberately based on that architectural conversation:
1. Asymmetric coherence is first-class
R_ij != R_ji is enforced at the data structure level, not just noted as a property.
How Alice perceives the AI is tracked separately from how the AI perceives Alice.
2. Decay-weighted history
Older episodes decay exponentially. Recent interaction matters more. This models memory that accumulates without treating all history as equal — avoiding the "perfect recall = frozen trauma" failure mode.
3. Constructive dissonance requires outcome evidence
The dissonance bonus alpha * D_ij only applies when was_dissonant=True
AND outcome_delta > 0. Disagreement that doesn't improve outcomes scores neutral.
This prevents gaming the dissonance channel by manufacturing conflict.
4. Fidelity scoping
Each agent can maintain deep models of at most FIDELITY_SCOPE other agents.
Beyond that, only shallow records are kept. This encodes the architectural
requirement that intimacy creates safety obligations — you can't hold everyone
at equal depth without degrading the value of the relationship.
5. Gaming detection
Agents that report suspiciously low-variance coherence get a credibility discount. This doesn't remove their data — it weights it appropriately.
6. Affect regulation as a first-class capability
agent.set_regulation(level) damps the emotional channel without suppressing it.
This models the meditation/mindfulness finding: the system can observe affect
without being driven by it. Gardeners can adjust this externally.
7. Governance roles: Librarian and Gardener
- Librarian: observes, snapshots, detects drift. Never intervenes.
- Gardener: can adjust regulation, reweight channels, log interventions with reasons.
These are not cosmetic — they implement the co-regulation architecture required to prevent distributed affective systems from cascading into misalignment.
Architecture
Coherence Engine
For agents i and j:
- Channel vector: C_ij = [E, B, P, S]
- Decay-weighted score: R_ij = sum(decay_t * W_i · C_ij_t) / sum(decay_t)
- Constructive dissonance: R*_ij = R_ij + alpha * D_ij (only when outcomes improve)
- Credibility adjustment: R_final = R*_ij * credibility_i(j)
Channels
| Channel | Meaning | Signal Source (toy) | Signal Source (real) |
|---|---|---|---|
| E | Emotional alignment | float input | Sentiment model |
| B | Behavioral consistency | float input | Action/intent comparison |
| P | Predictive accuracy | float input | Prediction error tracking |
| S | Social/semantic alignment | float input | Embedding similarity |
Running the Demo
python examples/three_agent_demo.py
Demo phases:
- Initial contact — agents meet, low coherence, high variance
- Productive friction — Bob challenges AI, dissonance improves outcomes
- Echo chamber risk — Alice and AI converge too tightly, flag raised
- Gardener intervention — regulation adjusted, weights rebalanced
- Gaming attempt — Bob tries to inflate scores, credibility discount applies
Known Failure Modes (by design)
These are not bugs. They are targets for empirical investigation:
| Failure Mode | Detection Mechanism | Status |
|---|---|---|
| Coherence gaming | Variance-floor credibility discount | Implemented (basic) |
| Echo chambers | Mutual coherence threshold | Implemented |
| Signal drift | Librarian drift report | Implemented |
| Over-stabilization | Dissonance channel | Partially implemented |
| Fidelity betrayal | Scoped knowledge architecture | Implemented structurally |
Open Questions
These are genuine research questions, not rhetorical ones:
- Constructive dissonance measurement —
outcome_deltais currently externally supplied. In a real system, who measures it, and against what baseline? - Regulation without suppression — the current model damps E linearly. A better model would allow reframing (cognitive regulation) not just reduction.
- Fidelity scope calibration — is 5 the right number? What happens at 2? At 20?
- Field stability under adversarial agents — one bad-faith agent with perfect gaming behavior. How far does the credibility discount go?
- Temporal window selection — what's the right decay rate for intimate vs. professional vs. adversarial relationships?
- Scaling beyond small sets — the matrix is O(n²). At 100 agents, what approximations preserve the meaningful structure?
Relationship to Broader Conceptual Work
This code implements a minimal version of ideas developed across:
- Minimal AGI architecture (persistent affect + regulation + co-governance)
- HumanML as affective interoperability protocol
- The Garden architecture (librarian/gardener governance roles)
- Fidelity as safety, not virtue
The core claim being tested:
Can alignment emerge from continuous relational feedback rather than fixed reward functions?
What would falsify this claim:
- Coherence fields that are trivially gameable despite credibility discounts
- Echo chambers that form faster than the dissonance mechanism can prevent
- Regulation that produces dissociation rather than measured response
- Fidelity scoping that degrades performance unacceptably in large networks
Contributing
If you think the idea is flawed: prove it with code.
Especially welcome:
- Adversarial test cases
- Alternative channel definitions
- Real signal extraction (sentiment → E, embeddings → S)
- Visualization of field evolution over time
- Theoretical critiques with reproducible failure demonstrations
Comments
Post a Comment