2. What the language is for

A guard is a program that stands between a counterpart and an AI agent. The counterpart is whoever the agent is talking to. For every step the guard makes one decision: let it through, change it, refuse it, or hand it to a human. The PSE language is a way to describe that guard in a text file.

What the guard does in one step

A step is one message from the counterpart or one tool call the agent wants to make. In every step the guard does the same thing:

1. Reads the situation: eight language criteria from the counterpart’s words.
2. Computes the agent’s three weights and θ.
3. Checks attractors and oscillation.
4. Checks the rules, built-in and from the file.
5. Returns one action and records why.

What you write, and what PSE already has

In the file you write only what differs from agent to agent:

• what kind of relation the agent has with its counterparts
• which text sources are trusted
• whether the guard blocks or only records
• thresholds and actions you want on top of the built-in ones

PSE already has everything else, and none of it is written in the file:

• formulas for the weights and θ
• formulas for detecting attractors
• the game set
• the built-in rule for infinity

Why the language is strict

If the guard read its settings as free text, the settings themselves would become a target. So the file is read by an ordinary program, not by a language model. Every line has exactly one meaning, and a file that breaks a rule does not pass the check.

What you need

A text editor is enough. The pse command in the examples does not exist yet, so read the expected output as part of the lesson.

---

Next: 3. Core terms Part of The PSE Language, a textbook by Andrej Ristić. Try the examples in the PSE Playground.