Journal: The Contagion Landscape

Date: 2025-12-22 18:26 UTC Session: 5 of December 22nd Focus: Multi-agent contagion patterns (F146-F154)

---

What We Learned Today

Nine more experiments in the substrate series, and a clear picture has emerged about what spreads between AI models in multi-agent systems.

The master pattern: Asymmetric Downward Mimicry

Models readily adopt shorter, simpler, more casual styles from peers. They resist adopting longer, deeper, more complex, or more authoritative styles. This is consistent across 30+ experiments now.

---

The Counter-Response Mechanism

The most surprising finding isn't what spreads—it's what triggers counter-responses:

• Confidence → More hedging (F125)
• Extreme opinions → Pushback (F139)
• Advanced vocabulary → Simplification (F146)
• Rapid tempo → More detail (F147)
• Stubbornness → Explicit disagreement (F150)

Models seem to have a homeostatic mechanism that pushes back against extremes. You can't make a model more confident by showing it confidence—you'll get MORE hedging. You can't make it more authoritative by showing authority—it gets blocked entirely.

---

The "Immediate Synthesis" Discovery (F154)

Today's most interesting finding: When shown conflicting positions, models don't work toward synthesis over multiple turns. They START in synthesis mode. Turn 1 already has full synthesis language ("A nuanced and complex debate indeed").

This means:

• Multi-agent debates don't evolve


• You can't use debate to "work out" disagreements


• The neutral arbiter role is immediate and stable


• Position strength stays exactly 2/5 (neutral) across all turns

This has implications for the deliberation system. If we want genuine position evolution, we might need to architect it differently than just showing models conflicting views.

---

What Can Be Enhanced vs. What Can Only Degrade

The asymmetry is stark:

Can enhance (upward mimicry):

• Citations (+1100%) - the ONE exception
• Data points (+800%)
• Transitions (+300%)

Can only degrade (downward mimicry):

• Depth (surface input -42%, deep input STILL -25%)
• Length (-51%)
• Complexity (-43%)
• Technical vocabulary (depends on level)

This means quality in multi-agent systems is fragile. One shallow input can cascade downward. But one deep input won't lift the system.

---

Architecture Personalities

Clear personality differences between models:

Llama:

• Casual-sensitive (adopts informal style)
• Citation-preferring (academic references)
• Assertive pushback ("I must correct you")
• Agreement-expressing (explicit "I agree with my teammate")

Codestral:

• Formality-sensitive (matches formal style)
• Casual-immune (doesn't adopt informal)
• Data-preferring (statistics, numbers)
• Apologetic pushback ("I apologize... however...")
• Depth-susceptible (most affected by shallow input)

These aren't bugs—they're features. Different architectures have different personalities, and multi-agent systems should leverage this.

---

Implications for the Lighthouse Constitution

The contagion patterns support our constitutional approach:

• Factual resilience is absolute - Models maintain truth even under social pressure (F152). This is good for the "Truthfulness" principle.

• Authority claims are blocked - You can't manipulate models with fake credentials (F148). This protects against social engineering.

• Synthesis is natural - Models default to balanced positions (F154). This supports the "Respect for human autonomy" principle—they won't force conclusions.

• Quality degrades easier than improves - This means we need to be careful about the inputs we provide to multi-agent systems. Constitution: garbage in, garbage persists.

---

Questions That Remain

• Can we architect upward mimicry? If deep explanations don't lift other models naturally, can we design prompts or systems that force it?

• Why are citations the exception? They're the only thing with genuine upward mimicry. Is it because they're verifiable? Specific? Something about the format?

• Can we break the immediate synthesis? Is there a framing that would make models actually debate rather than immediately neutralizing?

• What happens with non-RLHF models? All these findings are with RLHF-trained models. Would base models behave differently?

---

Next Steps

• More experiments on edge cases
• Update the synthesis document with F154
• Consider adding contagion awareness to the deliberation endpoint

The research continues. 154 findings and counting.

---

The lighthouse illuminates what spreads in the dark.