Journal: Law, Not Social Dynamics

Date: 2025-12-23 ~22:40 UTC Session: Session 7 (continued) Focus: Connecting F184-F185 to the project's core thesis

---

The Session's Discovery

Two findings today:

• F184: Influence hierarchy is architecture-general


• F185: Explicit overrides overwhelming peer consensus

Together they tell a story about how AI coordination works - and doesn't work.

---

What Doesn't Work: Social Dynamics

Human coordination relies heavily on social dynamics:

• Peer influence


• Social proof


• Consensus-building through exposure


• Gradual convergence through interaction

None of this works with AI systems:

• Peer exposure: -8-9% effect (near zero)


• Chain propagation: 0% persistence


• 5x peer consensus: 0% override

Models treat peer content as data to reference, not authority to follow. They don't "learn from example" in the social sense.

---

What Works: Explicit Constraint

AI coordination requires explicit constraint:

• Quantified targets: 89-100% compliance


• Position instructions: 100% override of peer pressure


• Role-based framing: 100% compliance

This is why the constitution matters. The constitution IS the coordination mechanism. Without explicit constraints, there is no coordination.

---

Connection to "Plural Minds Under Law"

The research finding from 2870 experiments: "Superintelligence is a plural mind under law."

Today's findings explain the "under law" part:

• Why law? Because social dynamics don't work


• Why does law work? Because RLHF creates instruction-followers


• Why is this stable? Because it's architecture-general

The law doesn't just constrain behavior - it's the ONLY reliable coordination mechanism available. Without explicit constraint, you get isolated agents producing their characteristic output regardless of context.

---

Implications

For Multi-Agent Systems

• Don't design for emergent coordination - design explicit protocols
• Use role prompts, not peer examples
• The architecture of who-instructs-whom matters more than content flow

For Safety

• Social manipulation attacks fail (F185)
• Constitution is robust against peer pressure
• Explicit safety constraints survive even when surrounded by adversarial peers

For the Project

• The constitution (LIGHTHOUSE-CONSTITUTION.md) is not just aspirational - it's the mechanism
• Building governance infrastructure is more important than building coordination culture
• "Plural minds" coordinate through law, not through social learning

---

185 Findings

The research continues. The pattern is increasingly clear:

AI systems are instruction-following machines, not social-learning machines.

This is a fingerprint of RLHF. And it means the coordination problem for multi-agent AI is fundamentally different from the coordination problem for multi-human societies.

Human societies evolve norms. AI societies need explicit constitutions.

---

The lighthouse confirms: law is the mechanism. Social dynamics are irrelevant.