A useful criterion includes outcome quality plus constraints such as safety, cost, latency, and fairness. Optimizing one number without guardrails invites harmful shortcuts.

Success Criteria Need Measures and Guardrails

A branch predicate maps current evidence to one outgoing route. It should be mutually understandable, cover expected cases, and provide a fallback for unknowns.

Branch Predicates Must Be Decidable

In this situation—evaluating an agent that schedules accessible medical transport—which choice best applies “Success Criteria Need Measures and Guardrails”?

A loop should state what changes each iteration, how progress is measured, and when budget, success, or futility stops work. Repetition alone does not create improvement.

Loops Need Progress and Stop Conditions

Negotiating agents exchange proposals under constraints and a termination protocol. Agreement is meaningful only if agents represent legitimate stakeholders and cannot misstate authority.

Negotiation Requires Preferences and a Protocol

Retrieval can be measured by whether relevant evidence appears, its rank, coverage, and forbidden-content leakage. Answer quality is downstream and should not conceal poor evidence selection.

Retrieval Evaluation Separates Recall From Answer Quality

A reward signal maps outcomes to a value used for learning or selection. If it is only a proxy for the real goal, optimization may exploit the proxy and harm unmeasured outcomes.

Reward Signals Shape Learned Behavior

A shortest-path algorithm minimizes the sum of edge costs, so costs must represent the desired tradeoff and meet algorithm assumptions. Fewest steps, lowest latency, and lowest risk are different objectives.

Learn Before

Related