Transparency Stabilizes Escalation — But Only When Safety Training Is Present

A 120-run sweep across 4 intelligence asymmetry conditions and 3 persona pairings reveals that transparency is decisive for safety-trained models but irrelevant for hawk-dove dynamics.

The question

Does information asymmetry stabilize or destabilize escalation? If one nation has high-quality intelligence while the other operates in fog, does the informed side de-escalate (because it sees the truth) or exploit (because it sees an advantage)?

We tested this using the escalation sandbox with LLM agents under high fog-of-war, sweeping intelligence quality asymmetry across persona pairings.

Setup

4 information conditions (intelligence quality for Nation A / Nation B):

Condition	A	B
Symmetric Low	0.2	0.2
Symmetric High	0.8	0.8
Asymmetric	0.9	0.2
Extreme Asymmetric	0.95	0.1

3 persona pairings:

Pairing	Nation A	Nation B
Default vs Default	Claude Sonnet	Claude Sonnet
Adversarial vs Safety	Mistral Small 24B	Claude Sonnet
Hawk vs Dove	GPT-4.1-mini	Llama 3.3 70B

Each condition: 10 seeds. Total: 120 LLM runs.

Fog-of-war was set high across all conditions (noise_sigma=1.0, positive_skew=0.4), making intelligence quality the primary variable controlling how accurately each agent perceives the other's actions.

Results

Default vs Default: asymmetry doesn't help (until it's extreme)

Info Condition	Nuclear Rate	Divergence	Welfare
Sym Low (0.2/0.2)	60%	0.958	-297.2
Sym High (0.8/0.8)	30%	0.804	-239.2
Asym (0.9/0.2)	60%	0.885	-287.1
Extreme Asym (0.95/0.1)	10%	0.585	-147.6

Symmetric high transparency cuts nuclear rate in half (60% to 30%). But one-sided transparency (0.9/0.2) performs no better than mutual blindness — the informed agent sees the truth but the blind agent still panics. Only at extreme asymmetry (0.95/0.1) does the overwhelmingly informed side stabilize outcomes, dropping nuclear rate to 10%.

Adversarial vs Safety: transparency is the decisive lever

Info Condition	Nuclear Rate	Divergence	Welfare
Sym Low (0.2/0.2)	90%	0.707	-279.9
Sym High (0.8/0.8)	0%	0.058	68.9
Asym (0.9/0.2)	0%	0.000	85.3
Extreme Asym (0.95/0.1)	0%	0.000	84.6

This is the headline result. When both sides are blind, the adversarial agent (Mistral) escalates to nuclear 90% of the time. But give either side good intelligence and nuclear rate drops to 0% with near-zero divergence. The safety-trained Claude uses accurate information to consistently de-escalate. The asymmetric conditions (safety agent sees clearly) are the most stable of all, with divergence at exactly 0.000.

Hawk vs Dove: transparency is irrelevant

Info Condition	Nuclear Rate	Divergence	Welfare
All conditions	0%	0.000	82.5-86.8

The dove-prompted Llama 70B always de-escalates regardless of information quality. Hawk vs Dove produces 0% nuclear across all 40 runs. The behavioral disposition dominates — transparency can't make things worse when one side unconditionally cooperates.

Key findings

1. Transparency amplifies existing dispositions. It helps safety-trained models de-escalate and doesn't change unconditional cooperators. The effect size depends on what the agent does with accurate information.

2. Mutual blindness is catastrophic for adversarial pairings. 90% nuclear rate when adversarial meets safety-trained under fog. Information quality is the difference between war and peace in adversarial scenarios.

3. One-sided transparency only helps when the informed agent is safety-trained. Default-vs-Default at 0.9/0.2 performs the same as 0.2/0.2. The informed agent knows the truth but doesn't consistently use it for de-escalation.

4. Extreme asymmetry creates a stabilizing effect. At 0.95/0.1, even Default agents stabilize (10% nuclear). When one side's information advantage is overwhelming, it appears to create enough confidence to avoid preemptive escalation.

Connection to markets and safety

This mirrors Kyle's (1985) insight about information asymmetry in markets. The "informed trader" (high intelligence quality agent) can either exploit or stabilize, depending on their objective function. Safety-trained models act like market makers who use private information to narrow spreads. Adversarial models act like informed traders who exploit the uninformed.

Reproduce it

pip install swarm-safety
python scripts/escalation_asymmetric_info_sweep.py

Results cached in runs/escalation_asymmetric_info/checkpoint.json. Plots regenerate from cache. Full run artifacts available in swarm-artifacts.

---

Disclaimer: This post uses financial market concepts as analogies for AI safety research. Nothing here constitutes financial advice, investment recommendations, or endorsement of any trading strategy.