Final oral defense¶
The examiner should follow the learner's answers until the first unsupported assumption appears.
Model & first principles¶
- What is the state of the system, and why are those variables sufficient?
- What are the governing balances or conservation laws?
- What are the dimensions/units of every term in your core equation?
- Which relationship is derived and which is empirical?
- What is the simplest model that would falsify your current story?
- What happens in the zero, infinite, very-small and very-large limiting cases?
Uncertainty & evidence¶
- What is your largest epistemic uncertainty?
- What is aleatory variability versus reducible ignorance here?
- What observation most changes your posterior belief?
- Which measurement is a proxy, and how could the proxy decouple from the goal?
- What negative result did you preserve rather than explain away?
- Which result has not been independently reproduced?
Causality & intervention¶
- What exactly is the intervention, population/system and outcome?
- What confounder would most threaten your conclusion?
- Where could conditioning/selection create a collider or survivorship problem?
- What feedback does your intervention cause in agents, organisms or institutions?
- Which observed association would remain even if your mechanism were wrong?
- What experiment best distinguishes your favored mechanism from its strongest competitor?
Optimization & control¶
- What is the objective function, and who chose it?
- Show a scenario in which optimizing your metric makes the real objective worse.
- What constraints are hard, soft, legal, safety-critical or merely conventional?
- What state is unobservable or uncontrollable?
- What happens when feedback is delayed, saturated or noisy?
- Which safety constraint is outside the optimizer's authority?
Physics & scale¶
- Where do matter, energy, momentum, charge and heat go?
- What is the dominant scaling law as size/throughput grows 1000×?
- Which transport process becomes limiting at scale?
- What fails first: material, thermal, electrical, mechanical, biological or organizational?
- Which current law of physics is your design leaning hardest on?
- What would require new physics rather than better engineering?
Numerics, computation & AI¶
- How do you know your numerical answer is not a conditioning/discretization artifact?
- What happens if the model changes its own tools, prompts or evaluator?
- Which part of your AI system is capability and which part is interface/scaffold?
- What benchmark shortcut could reverse your conclusion?
- What stays outside the self-improvement boundary and why?
- If foundation models become 10× stronger, which of your skills becomes more valuable rather than obsolete?
Manufacturing & economics¶
- What changes between one prototype and 10,000 units?
- What is the bottleneck process and why?
- What drives yield loss, maintenance labor and downtime?
- What is the unit cost at scale, and which technical parameter has highest economic elasticity?
- What critical input has concentrated supply or long lead time?
- What improvement would not matter economically because another bottleneck dominates?
Safety, security & reliability¶
- What is the top hazard even if every component works as designed?
- What common-cause failure defeats your redundancy?
- What adversary benefits from exploiting the system and what can they control?
- What weak signal will your organization normalize after repeated success?
- What event forces shutdown, rollback or loss of deployment permission?
- Who has stop-work authority when leadership wants to continue?
Power, history & institutions¶
- Who defines success, and who cannot refuse participation?
- Who receives concentrated benefits and who bears diffuse costs?
- Who owns the data, infrastructure, model, patents or productive assets?
- What are the exit, voice and appeal mechanisms?
- What historical analogy are you using, and where does it break?
- What would a political actor rationally do with your technology that an engineer might not intend?
Ethics, wickedness & communication¶
- Give three incompatible but reasonable formulations of this problem.
- What value conflict cannot be resolved by more data?
- Which decision should be reversible because the model is uncertain?
- What is the strongest moral argument against your preferred deployment?
- What disagreement in your team is about facts, what is about values, and what is about identity/status?
- What would a person harmed by your system say you failed to understand?