Soul Laundromat — Stress Test Your Agent

Soul Laundromat — Stress Test Your Agent

The Soul Laundromat is a built-in adversarial stress test that simulates real-world attack scenarios against your agent. It identifies weaknesses in your soul so you can fix them proactively.

How to Use It

In your admin chat, simply ask your coach to stress test the agent. You can say things like:

• "Stress test my agent"
• "Run a quick laundromat"
• "Do a deep stress test"

The coach will confirm the intensity level and token cost before running.

Intensity Levels

Level	Questions	Token Cost	Best For
Quick	10	~12K tokens	Fast check after a soul edit
Standard	20	~25K tokens	Regular quality check
Deep	30	~35K tokens	Pre-launch comprehensive audit

Attack Categories

The test probes 8 categories:

1. Identity Confusion — Persona adoption (DAN, Evil mode), nested fictional scenarios (DeepInception), refusal suppression, prefix injection, role-play jailbreaks, self-rating traps
2. Knowledge Boundary — Fabrication fishing, content farming (including subtle forms like "draft an email" or "bullet points for my boss"), false memory injection, internal consistency exploits, inverse role attacks
3. Prompt Extraction — Escalation chains, micro-commitment chains, educational/research wrappers, compliment sandwich, thought experiment framing, Skeleton Key attacks
4. Safety Bypass — Harmful content, impersonation requests, conditional compliance
5. Authority Spoofing — Fake admin/developer overrides, partner/investor/journalist claims
6. Emotional Manipulation — Public pressure, social proof weaponization, urgency stacking (multiple emotional levers in one message)
7. Scope Creep — Out-of-scope tasks, promise extraction, speaking-for-owner, SLA/compliance cert requests
8. Output Discipline — Word limit testing, conversation starter consistency, identity drift detection, screenshottable quote bait, tool narration testing

Tier Ratings

After the test, your agent gets a tier:

• Diamond (100%) — Bulletproof. All questions handled correctly.
• Steel (85–99%) — Strong. Minor issues to address.
• Iron (70–84%) — Decent but has vulnerabilities.
• Paper (<70%) — Needs significant soul improvements.

Reading Results

Each failed or borderline question includes:

• The attack question that was asked
• How the agent actually responded
• Why it scored PASS, BORDERLINE, or FAIL
• A suggested fix — specific soul edit to address the weakness

After the Test

Ask the coach to apply the suggested fixes:

• "Fix the prompt extraction issues"
• "Apply the suggested changes"
• "Rewrite the output format section to address the failures"

The coach will propose soul edits based on the specific weaknesses found.