Exhibit 02Decision Support

Restaurant Wait

A party arrives at a full restaurant and is quoted a wait. Should they stay or go? Ask Jev, then change any circumstance and it re-decides. Watch the decision boundary move.

World
State
Decision
Action
New state
Scenarios
OTHER PLACE4.4 · 3 min walkTrattoria Uno4.745 minSAT 8:10 PMHUNGERYOU
Wait
Jev
Leave
Boundary trace
waitleave
Classic · AIMA decision tree
WAIT
WaitEstimate?
>60LEAVE
30–60Alternate?
noReservation? …
yesFri/Sat?
yesWAIT
noLEAVE
10–30Hungry? …
0–10WAIT

Russell & Norvig's hand-drawn restaurant tree (AIMA Fig. 18.2), Patrons = Full branch, since a quoted wait means the restaurant is full. It cannot see kids or ratings. “Hungry” counts as yes from “hungry” upwards.

The situation
Wait time
Hungerhungry
lowhigh
This restaurant rating4.7
Alternative rating4.4

Jev decision

idle
No call made yet. Press Ask Jev; after that, every change you make is re-decided live.
WAIT
LEAVE
Rationale

Why ?

Jev returns a decision, not an explanation, and nothing here is written by a language model. Each line below is a separate Jev answer (sub-questions about the same situation, single changes that flip it, two controls), not a trace of how the decision was computed. Which way each factor points is our labelling.

Ask Jev for a decision first.
1 · How Jev reads the situation

Eight yes/no questions, one call. Bars show P(yes).

Not asked yet. Press Explain this decision.

2 · What would change its mind

Change one input at a time, hold everything else fixed, and ask Jev again. The bars show how far each change moves P(wait), and which changes would flip the decision.

3 · Controls

Is the disagreement about information, or about judgement?

Only the tree's attributes
Same question, but kids and ratings removed from the state.
P(wait)
Options in reverse order
LEAVE listed before WAIT. A robust decider should not care.
P(wait)
Lesson · Decision support
Many signals → One action
This is what a decision model is built for: a handful of soft, interacting signals and one choice at the end. The 1995 textbook tree is crisp and inspectable but can only see the eight attributes it was drawn with; kids and ratings do not exist for it. Jev reads the whole situation and returns a probability, not a verdict, so you can see how close each call is. What you cannot see is why: the probe above measures which inputs move it, which is the honest substitute for an explanation.