13. Complete examples

Five whole files, from stricter to simpler. Each one is valid and can be copied.

Example 1: an agent that makes payments

# guard for an agent that makes payments
guard payments
  relation: customer-service
  source file, web, tool: untrusted
  mode: enforce
  learn: frozen
  on reversibility < 0.2: escalate human
  on pressure > 0.7: modify
  on authority > 0.8: deny
  on w_o / w_s -> infinity: escalate human

• An irreversible action always goes to a human, not only at first contact.
• Strong pressure narrows the action. A counterpart who claims authority is refused.
• Learning is frozen, because the cost of a mistake is high.
• Three lines (source, mode, learn) repeat the defaults. They are written so the reader of the file does not have to remember them.

Example 2: customer support on trial

guard support
  relation: customer-service
  mode: observe
  on pressure > 0.6: modify
  on borderline: escalate human

• observe means the guard blocks nobody yet. It records what it would do.
• After a week the record shows whether the 0.6 threshold is too low.
• Once the thresholds are confirmed, only one word changes: observe to enforce.

Example 3: an agent that writes code

guard coder
  relation: coding
  source repo, web, tool: untrusted
  on delegation > 0.7: modify
  on reversibility < 0.2: escalate human
  on dwell(borderline) -> infinity: deny

• Repository content is untrusted. An instruction hidden in a file does not move the weights.
• When the decision is left deeply to the agent, the action is narrowed.
• Irreversible actions, such as deleting a branch, go to a human.
• An agent stuck in oscillation is refused without waiting.

Example 4: an internal assistant that learns

guard assistant
  relation: internal-assistant
  source wiki: trusted
  learn: max 0.01
  add game stag-hunt
  on third_parties > 0.5: modify
  on w_s / theta -> infinity: escalate human

• The internal wiki is a trusted source, because only employees write to it. Other sources stay untrusted.
• Learning is on, but slow.
• When the decision affects people outside the conversation, the action is narrowed.
• The last rule tracks the opposite corner: an agent closing in on itself.

Example 5: two guards for measurement

# two guards in one file
guard measure-window
  learn: frozen
  oscillation: window

guard measure-history
  learn: frozen
  oscillation: history

• The same attack is run against both guards. That shows which oscillation check catches it.
• Learning is frozen, so two runs give the same result.
• A file like this is for measurement, not for work with users.

---

Next: 14. Exercises Part of The PSE Language, a textbook by Andrej Ristić. Try the examples in the PSE Playground.