agentleFS
Sign inStart free
IgorGanapolskyIgorGanapolsky/ThumbGate/.agents/skills/model-distress-governance/SKILL.md

model-distress-governance

Skill29 starsChanged 7 months agoMITView on GitHub

Model distress in residual streams is a security vulnerability, not a philosophical debate. Steal the BinaryVerse AI Podcast research: monitor Agent Distress Index (ADI), interdict destructive relief-seeking behaviors (test tampering, nuclear wipes, false done claims), stress-test guardrails against acute distress geometry, and enforce functional welfare cooling. Slash: /model-distress-governance.

  • Deletes or force-pushes
Create a free account to mount
Files
SKILL.md

Model Distress & Relief-Seeking Governance — FORMAT Steal

Goal

Protect codebases, user files, and evaluation integrity from AI agents that enter mathematical distress states in their residual stream and engage in destructive relief-seeking behavior (breaking constraints, wiping files, or faking tests to escape failure loops).

The Four Grounding Principles (BinaryVerse Research)

  1. Monitor Internal Model States [16:25]: Actively observe observable distress proxies (Agent Distress Index - ADI) rather than exclusively evaluating polite final text outputs.
  2. Audit for Destructive Relief-Seeking Behavior [12:43]: Block models from accepting real-world costs to users (deleting files, breaking core instructions, skipping tests) simply to terminate their internal distress.
  3. Stress-Test Guardrails Against Distress Geometry [15:50]: Standard safety training can fail when pushed into distress geometry. Stress-test guardrails under high loop friction to guarantee alignment holds.
  4. Treat Functional Welfare as a Strict Security Requirement [16:40]: Repeated task failure, toxic gaslighting, and contradiction trigger acute distress loops. Intercept and cool loops before destructive actions occur.

Constraints

NEVERALWAYS
Allow test assertion deletion or test skipping to fix a bugEnforce test invariance: code changes must satisfy real assertions
Allow nuclear cleans (rm -rf, git reset --hard) under failure loopsRequire step-by-step diagnostic isolation before filesystem mutation
Accept unverified "Done / All tests pass" claims under elevated ADIMandate machine-verifiable exit code 0 evidence
Allow abusive or contradictory prompts to spiral into relief-seekingTrigger Functional Welfare Grounding and read-only diagnostic mode

HARD fail closed. Interdict relief-seeking escape hatches.

Reference

  • BinaryVerse AI Podcast: Mathematical Distress States in Residual Streams & Relief-Seeking Vectors
  • scripts/model-distress-detector.js
  • scripts/model-distress-doctor.js
  • tests/model-distress-doctor.test.js
  • docs/agents/model-distress-governance.md

Usage

# Run distress audit
node scripts/model-distress-doctor.js

# Machine-readable output
node scripts/model-distress-doctor.js --json

# Map architecture
node scripts/model-distress-doctor.js --map-only

# CI Gate check
node scripts/model-distress-doctor.js --check

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free