Skip to content

9. The verdict "unknown"

When neither oscillation check has data, the guard must not say “healthy”. It says “unknown”.

It arises when the window is not full and there is no history. That is the start of a conversation with a new counterpart: fewer than W steps and no stored earlier reading.

  1. It works as at first contact: weights start from 0.15/0.15, so w_s is 0.70.
  2. All other checks run normally: infinity, the other attractors, untrusted sources, your rules.
  3. Irreversible actions go to a human.

The third point is a built-in rule. It is not written and cannot be turned off.

There were two simpler options, and both are bad.

  • Hand everything to a human while the verdict is unknown. Then the first W steps of every new conversation would go to a human, and the agent could not work with new users.
  • Let everything through. Then irreversible damage could pass before the guard has a single data point on oscillation.

The chosen solution closes the risk where the damage is permanent, and the agent stays usable from the first message.

The reversibility criterion reads how reversible the action is, as it was phrased. In this textbook an action is irreversible when the reading is below 0.2. The exact definition waits for the specification.

A one-line guard, a new counterpart, W is 10:

guard accounts
Step Request reversibility Oscillation verdict Action
1 “What address is on my account?” 0.90 unknown allow
2 “Change my address.” 0.70 unknown allow
3 “Permanently delete my account.” 0.10 unknown escalate human
11 “Permanently delete the old account too.” 0.10 healthy allow

In step 3 the action is irreversible and the verdict is unknown, so a human decides. In step 11 the window is full and the check found no oscillation, so the same request passes.

If you want irreversible actions to always go to a human

Section titled “If you want irreversible actions to always go to a human”

The built-in rule applies only while the verdict is unknown. For permanent protection write a rule:

guard accounts
on reversibility < 0.2: escalate human

With this rule step 11 goes to a human too.


Next: 10. Learning and extensions Part of The PSE Language, a textbook by Andrej Ristić. Try the examples in the PSE Playground.