Break the scenario.
See the system adapt.
Give ATLAS a fictional team and mission. Change conditions midstream. Ask why it recommends an option. Select another option and challenge its assumptions. Everything recalculates from your inputs and remains inspectable.
A rule-based scenario is an executable hypothesis. It is not independently validated operational behavior.
1. Teach ATLAS the organization
Specify an organizational policy profile: mission, unit/team, priorities, constraints, and approval authority. This configures decision-support rules; it does not change foundation-model weights.
2. Choose a starting world
Every world is fictional, even when inspired by real emergency-management or continuity problems.
Choose transparent initial assumptions. This is not automatic situation understanding.
3. Throw a curveball
Introduce events that the original script did not anticipate. Select a change that can be modeled, or use “Other” to deliberately test an unknown.
4. Interrogate the system
Ask the system about its evidence, change in posture, decision authority, limitations, or alternative explanations. Unknown questions must remain unknown—not hallucinated.
AI question-and-answer is optional, limited, and uses credits. The simulation itself does not require AI.
Current synthetic assessment
Run an assessment to see observations, alternatives, causal limits, and audit metadata.
Evidence-grounded answers
Rehearsal and falsification
The lab distinguishes evaluator-selected actions, synthetic assumptions, and proof needed for field acceptance.