10. Optimization and Prescriptive Decision Models

Introduction: From Describing the Supply Chain to Choosing an Action

Descriptive and predictive analytics explain what has happened and what is likely to happen. Prescriptive analytics addresses the managerial question that follows: what should the organisation do? In logistics, this question may concern the quantity to order, the routes to operate, the facilities to open, the customers to serve, the vehicles to assign, or the inventory to position. These decisions are difficult because resources are scarce, objectives conflict, demand is uncertain, and operational choices are interdependent.

Optimization provides a formal way to select the best feasible decision according to a stated objective. A model converts managerial circumstances into decision variables, parameters, an objective function, and constraints. A solver then searches for a solution, while the manager evaluates whether that solution is operationally realistic, robust, ethical, and consistent with organisational strategy.

Optimization does not replace managerial judgement. It disciplines judgement by making assumptions visible, quantifying trade-offs, and showing the consequences of alternative policies. The quality of an optimization result depends not only on mathematical technique but also on data quality, model formulation, interpretation, and implementation.

Learning Objectives

After studying this chapter, the learner should be able to:

1. Distinguish descriptive, predictive, and prescriptive analytics in supply-chain decision-making.

2. Formulate linear programming, integer programming, nonlinear programming, and multi-objective models.

3. Identify decision variables, parameters, objective functions, constraints, feasible regions, and optimal solutions.

4. Interpret shadow prices, reduced costs, slack, dual values, and solver status messages.

5. Compare exact optimization with heuristics and metaheuristics.

6. Translate a mathematical solution into a feasible managerial policy.

7. Recognise modelling risks, ethical concerns, and implementation challenges.

8. Apply optimization reasoning to a logistics network-design and distribution problem.

The Architecture of a Prescriptive Model

A prescriptive model begins with a decision problem rather than with an algorithm. The analyst must first define the decision horizon, organisational scope, units of measurement, data sources, and performance criteria. A model built without this framing may produce a mathematically correct answer to the wrong question.

Decision variables and parameters

A decision variable represents an action under the decision-maker’s control. In a transport model, \(x_{ij}\) may represent the number of shipments from origin \(i\) to destination \(j\). In a facility-location model, \(y_j\) may be a binary variable equal to one if facility \(j\) is opened. In an inventory model, the order quantity and reorder point may be decision variables.

Parameters are known or estimated inputs. They include demand, transport cost, capacity, lead time, fixed facility cost, service requirements, emissions factors, and handling time. A parameter is not necessarily perfectly known; it may be a forecast, an estimate, or a scenario value. Confusing an input estimate with a decision variable is a common modelling error.

Objective function

The objective function measures the performance to be maximised or minimised. A logistics model might minimise total cost:

\[

\min Z = \sum_{i}\sum_{j} c_{ij}x_{ij} + \sum_j F_j y_j,

\]

where \(c_{ij}\) is variable shipment cost and \(F_j\) is the fixed cost of opening facility \(j\). Other objectives may include maximising contribution margin, service coverage, resilience, or reliability. The objective must reflect the organisation’s actual decision purpose. Minimising transport cost alone may increase inventory, reduce service, or expose the network to excessive concentration risk.

Constraints and feasibility

Constraints express physical, contractual, financial, regulatory, or policy limits. Typical constraints include:

• supply and demand balance;
• warehouse, vehicle, labour, or dock capacity;
• minimum order quantities and shipment sizes;
• route, time-window, and vehicle compatibility requirements;
• budget limits;
• service-level requirements;
• logical relationships such as “a shipment is possible only if a facility is open.”

For example, demand satisfaction at customer \(j\) may be represented as:

\[

\sum_i x_{ij} \geq d_j,

\]

where \(d_j\) is demand. If supply at origin \(i\) is limited, then:

\[

\sum_j x_{ij} \leq s_i.

\]

A solution satisfying all constraints is feasible. If no such solution exists, the model is infeasible and the analyst must investigate conflicting requirements rather than simply seek a different solver.

Feasible and optimal solutions

The feasible region contains every combination of decisions that satisfies the constraints and variable restrictions. The optimal solution is the feasible solution with the best objective value. There may be one optimum, several alternative optima, or no finite optimum. An unbounded model may indicate a missing constraint, such as an omitted capacity limit or an incorrect sign in a cost coefficient.

The distinction between mathematical optimality and managerial acceptability is important. A model may recommend using nearly all capacity, relying on a single supplier, or changing a route pattern daily. Such a result can be optimal under the stated assumptions but unacceptable in practice. The implementation decision therefore requires a second stage of review.

Linear Programming: A Foundation for Allocation Decisions

A linear programming (LP) model assumes that the objective and constraints are linear functions of the decision variables and that variables may generally take continuous values. LP is valuable in transportation, blending, workforce allocation, production planning, inventory positioning, and capacity planning.

Consider a distributor shipping products from two warehouses to three markets. Let \(x_{ij}\) be the quantity shipped from warehouse \(i\) to market \(j\). The model minimises transportation cost subject to warehouse supply and market demand. If partial units are meaningful—for example, pallets, tonnes, or litres—the continuous assumption may be reasonable. If the decision is about whole trucks or delivery tours, LP alone may be inadequate.

The transportation model has a useful structural property: costs, supplies, and demands are visible and auditable. It can also expose the marginal value of additional warehouse capacity or the cost of serving a remote market. However, it may simplify congestion, driver regulations, delivery sequencing, and uncertain travel times.

Assumptions and interpretation

LP commonly assumes proportionality, additivity, divisibility, and certainty. Proportionality means that twice the activity produces twice the cost or resource consumption. Additivity means that the total effect is the sum of individual effects, with no interaction unless explicitly modelled. Divisibility permits fractional quantities. Certainty means that parameters are treated as known for the model run.

These assumptions are modelling choices, not universal truths. Fixed ordering costs violate proportionality; vehicle activation costs violate simple additivity; whole-vehicle requirements violate divisibility; and forecast error violates certainty. The response may be to use integer, nonlinear, stochastic, or robust optimization, or to retain LP as an informative approximation.

Integer and Mixed-Integer Models

Integer programming requires some or all variables to take integer values. Binary variables take values of zero or one and are especially useful for yes/no decisions. Mixed-integer linear programming (MILP) combines continuous and integer variables.

A facility-location model may use \(y_j=1\) when facility \(j\) is opened. Shipment from facility \(j\) is then linked to the opening decision through:

\[

\sum_i x_{ij} \leq M y_j,

\]

where \(M\) is a valid upper bound on shipments. If \(y_j=0\), shipments must be zero; if \(y_j=1\), shipments may occur up to the bound. Choosing an excessively large \(M\) can weaken the model and slow computation, whereas choosing a bound that is too small can eliminate valid solutions.

Integer models represent vehicle counts, workforce rosters, order quantities, facility openings, product selections, and assignment decisions. Their computational difficulty arises because the feasible region is discrete and often non-convex. Solvers typically use methods such as branch-and-bound, cutting planes, presolve, and linear relaxations.

The solver may report an optimality gap. A gap of zero, subject to tolerances, indicates that the best known feasible solution has been proven optimal. A non-zero gap means the solver has a feasible solution and a bound on how far it may be from the true optimum. In large operational models, a manager may accept a small gap when the improvement from further computation is unlikely to justify the time or complexity.

Nonlinear Optimization and Network Effects

A nonlinear model contains nonlinear expressions in its objective or constraints. Nonlinearity may arise from congestion, economies of scale, service deterioration, risk functions, fuel consumption, or inventory-holding relationships. For example, delivery time may increase disproportionately as a facility approaches full utilisation. A congestion penalty could be represented by a convex function of flow.

Convex nonlinear optimization is generally more tractable because a local optimum is also a global optimum under suitable conditions. Non-convex models may contain multiple local optima, discontinuities, or complex interactions. Solver output must therefore be interpreted carefully. A local solution is not automatically the best solution overall.

Nonlinear models also create communication challenges. Managers may understand a fixed handling cost or a capacity limit more readily than a calibrated congestion function. The analyst should explain the business meaning of each nonlinear relationship and test whether conclusions remain stable under alternative specifications.

Multi-Objective Decision Models

Supply-chain decisions rarely have a single legitimate objective. Cost, service, carbon emissions, resilience, working capital, and employee safety may all matter. A multi-objective model makes these competing priorities explicit.

One approach is the weighted-sum method:

\[

\min Z = w_1 C + w_2 E + w_3(1-S),

\]

where \(C\) is cost, \(E\) is emissions, \(S\) is service performance, and the weights express managerial priorities after suitable normalisation. Weights must be interpreted carefully because cost, tonnes of carbon, and service percentages have different scales. A poorly scaled model can allow one criterion to dominate unintentionally.

A second approach is the epsilon-constraint method. One objective is optimised while the others are imposed as limits, such as minimising cost subject to emissions being no greater than a specified threshold and service being at least 98 percent. This method is often easier to explain in policy discussions because managers can debate explicit targets.

The set of non-dominated solutions is called the Pareto frontier. A solution is Pareto-efficient when improving one objective would worsen at least one other objective. The frontier does not select the final policy; it presents meaningful alternatives for managerial choice. In a strategic setting, showing three or four robust Pareto alternatives may be more valuable than presenting one apparently precise answer.

Reading Solver Outputs as Managerial Evidence

A solver output should be treated as evidence, not as an instruction. The analyst should examine at least five elements: solution status, objective value, variable values, constraint activity, and sensitivity or gap information.

A binding constraint is satisfied at its limiting value. If a warehouse capacity constraint is binding, increasing that capacity may improve the objective. A non-binding constraint has slack. Slack indicates unused capacity or an unmet upper limit, but it does not necessarily mean the resource is unimportant under all scenarios.

A shadow price, or dual value, estimates the marginal change in the objective caused by relaxing a constraint by one unit, within the applicable sensitivity range and subject to model assumptions. If the shadow price of warehouse capacity is 12 currency units per pallet position, one additional position may reduce cost by approximately 12 units, provided the model structure remains unchanged. The value is not automatically valid for a large capacity expansion or a changed demand pattern.

Reduced cost is associated mainly with variables at a bound. In a minimisation model, it can indicate how much an unfavourable objective coefficient would need to improve before a currently zero variable becomes attractive, subject to solver conventions. Analysts should verify the sign and interpretation in the specific software documentation.

Sensitivity analysis changes parameters systematically to test stability. Scenario analysis changes coherent groups of assumptions, such as fuel prices, demand, or supplier availability. Stress testing examines severe but plausible conditions. These approaches are essential because a single optimal solution may be fragile.

Exact Optimization, Heuristics, and Hybrid Approaches

Exact optimization seeks a provably optimal solution or a certified bound. It is preferred when the model is tractable, the decision is high-value, and proof of optimality matters. However, routing, scheduling, and large network problems may be computationally expensive.

A heuristic uses a rule or constructive procedure to obtain a good solution quickly, without guaranteeing optimality. Examples include nearest-neighbour routing, savings methods, greedy assignment, local search, and repair-based scheduling. Metaheuristics such as genetic algorithms, tabu search, and simulated annealing explore solution spaces through broader search strategies.

Heuristics are appropriate when decisions must be generated rapidly, data changes continuously, or the exact model is too large. They should be evaluated against lower bounds, known benchmarks, historical practice, and operational feasibility. A heuristic result is not inferior merely because it lacks a proof of optimality; its value depends on quality, speed, transparency, and implementation.

Hybrid methods combine approaches. An MILP may determine facility locations while a routing heuristic constructs daily tours. Alternatively, a heuristic may generate a good incumbent solution that helps an exact solver narrow its search. In practice, the strongest decision systems often combine optimization, simulation, rules, human approval, and real-time data.

From Mathematical Solution to Operating Policy

Implementation requires translation. A solution stating that 4,260 units should be shipped from one node to another must become a policy concerning order release, dispatch timing, replenishment authority, exception handling, and accountability.

A useful translation process includes four stages. First, validate the solution against business rules not represented in the model, such as labour agreements, customer preferences, safety rules, and maintenance schedules. Second, convert quantities into executable decisions: purchase orders, transport bookings, production plans, or facility allocations. Third, define contingencies for forecast error, supplier failure, vehicle breakdown, and demand surges. Fourth, monitor actual performance and compare it with model assumptions.

The policy should specify which decisions are automated, which require approval, and when the model will be rerun. It should also define guardrails, such as maximum supplier concentration, minimum service levels, or a limit on daily schedule changes. Optimization should support a controlled decision process rather than create an opaque command system.

An Illustrative Case: Regional Distribution Network Redesign

Background

Northstar Medical Supplies, a fictional regional distributor, serves hospitals from three candidate distribution centres. Demand is stable in the base case but varies by region. The company is considering whether to operate one, two, or three centres. Management wants to reduce annual logistics cost while maintaining a 98 percent service target and avoiding excessive dependence on one location.

Situation and problem

The existing network uses three centres, each with different fixed costs, capacities, and transport distances. The finance team focuses on facility and transport cost. Operations managers emphasise response time and disruption risk. The sustainability team requests an emissions reduction. The decision cannot be made by comparing facility rent alone because opening fewer centres may increase transport distance and reduce resilience.

Model formulation

The analyst defines binary variables \(y_j\) for opening centre \(j\), continuous variables \(x_{ij}\) for shipments from centre \(i\) to hospital region \(j\), and possibly scenario-specific variables for emergency demand. The objective minimises fixed facility cost, transport cost, expected shortage penalty, and an emissions term. Constraints impose facility capacity, demand coverage, service requirements, and a maximum share of demand supplied by any one centre.

The service constraint is not represented merely by total volume. At least 98 percent of demand in each region must be reachable within the required delivery time. This requires feasible arcs or time-specific constraints rather than a simple average-distance measure. The concentration constraint ensures that no centre supplies more than 70 percent of total demand.

Analysis and approach

The first MILP run produces a two-centre solution with the lowest weighted objective. However, sensitivity testing shows that the solution depends heavily on one centre’s capacity. A disruption scenario removing that centre causes significant unmet demand. The analyst therefore constructs a Pareto comparison: a low-cost two-centre design, a slightly more expensive three-centre design, and a resilient design with additional reserved capacity.

The three-centre design has higher fixed cost but lower disruption exposure and better emergency response. Management rejects the cheapest design because the model’s concentration limit does not capture the full clinical consequence of failure. Instead, it selects the resilient design and negotiates flexible capacity at the third centre rather than operating it at full scale immediately.

Outcome and policy

The chosen policy opens two primary centres, retains a smaller third node as a contingency and rapid-response facility, and establishes quarterly capacity reviews. The optimization model determines regular allocations, while an exception rule permits emergency shipments when hospital criticality exceeds a threshold. The company also creates a dashboard showing capacity utilisation, service performance, concentration, emissions, and deviation from the recommended allocation.

Learning points

The case illustrates that the cheapest mathematical solution may not be the best organisational policy. It also demonstrates the importance of correctly modelling service, separating normal and emergency operations, testing disruption scenarios, and presenting alternatives rather than hiding value judgements inside arbitrary weights.

Common Failure Modes and Model Risk

The most serious error is often not an incorrect algorithm but an incorrect formulation. Common failures include omitting a critical constraint, double-counting costs, using inconsistent units, allowing shipment through unopened facilities, setting an invalid big-

M value, and treating forecast demand as certain.

Overfitting is possible in optimization. A model may reproduce historical decisions but perform poorly under new conditions if it encodes temporary patterns or exceptional costs. Data leakage, stale master data, and inconsistent location or product identifiers can also produce misleading results.

Managers should be alert to objective gaming. If performance evaluation rewards transport-cost reduction alone, planners may increase inventory, reduce service, or shift costs to another department. A model can optimise what is measured while damaging what is not measured. Governance should therefore include cross-functional review of objectives, constraints, data definitions, and unintended consequences.

Feasibility problems deserve disciplined diagnosis. Relaxation variables can identify which constraints cause infeasibility, but they should not become a permanent way to violate service or safety requirements. If constraints are relaxed, the associated penalty and managerial approval must be explicit.

Ethics, Governance, and Responsible Automation

Prescriptive systems influence allocation of scarce goods, supplier opportunities, delivery priority, and employee workloads. Ethical concerns arise when the objective disadvantages particular customers, regions, suppliers, or worker groups without justification. A service model that systematically deprioritises remote communities may be financially efficient but socially unacceptable or legally restricted.

Governance requires documented assumptions, data lineage, model version control, access controls, reproducible runs, and an audit trail of overrides. Users should know when a recommendation is generated automatically and when a human has changed it. Sensitive variables should be reviewed for proxy discrimination. Safety-critical decisions require human oversight and clearly defined escalation procedures.

Explainability is especially important when managers challenge a recommendation. The system should be able to show the binding constraints, cost trade-offs, scenario assumptions, and reasons why an alternative was rejected. Responsible use also requires monitoring after deployment because demand, regulations, supplier behaviour, and service priorities change.

Digital Tools and Emerging Developments

Modern optimization applications increasingly combine cloud computing, application programming interfaces, digital twins, real-time transport data, and machine-learning forecasts. Forecasting systems estimate demand or travel time; optimization converts those estimates into actions. The two functions should remain conceptually distinct so that forecast uncertainty is not mistaken for decision certainty.

Stochastic optimization represents multiple demand or disruption scenarios and chooses decisions that perform well across them. Robust optimization seeks decisions that remain feasible within defined uncertainty sets. Distributionally robust methods go further by protecting against ambiguity about the probability distribution itself. These approaches can produce less aggressive but more dependable policies.

Digital twins allow planners to test network changes, capacity failures, and policy alternatives in a simulated representation of operations. Reinforcement learning and other AI methods may help with sequential decisions, but they require careful reward design, exploration controls, and operational safeguards. AI-generated recommendations should not bypass constraint validation or governance.

Best Practices for Professional Use

A reliable optimization project normally follows these practices:

• define the decision and success criteria before selecting a solver;
• involve operations, finance, commercial, technology, and risk stakeholders in formulation;
• use meaningful variable names, documented units, and auditable data transformations;
• build a small prototype and validate it with known cases;
• separate hard constraints from preferences and penalties;
• use realistic bounds and avoid unnecessarily large constants;
• test base, optimistic, pessimistic, and disruption scenarios;
• report feasibility, objective value, gap, binding constraints, and sensitivity;
• present alternative policies and explain trade-offs;
• pilot recommendations before full automation;
• monitor actual performance and revise the model as conditions change.

A mature organization treats the model as a governed decision product, not as a one-time spreadsheet.

Chapter Summary

Optimization converts supply-chain choices into formal decision models consisting of variables, parameters, objectives, and constraints. Linear programming supports continuous allocation decisions, while integer and mixed-integer models represent discrete choices such as facility openings, vehicles, and assignments. Nonlinear models capture congestion and other interactions, and multi-objective methods reveal trade-offs among cost, service, emissions, and resilience.

Solver outputs must be interpreted through feasibility, binding constraints, shadow prices, reduced costs, sensitivity, and optimality gaps. Exact methods provide guarantees when tractable; heuristics and hybrid methods provide speed and practical solutions for complex problems. The final managerial task is to convert a mathematical solution into an executable, governed, and adaptable policy. Scenario analysis, ethical review, data governance, and human oversight are essential because an optimal model result is only as credible as its assumptions and implementation.

Key Terms

Decision variable: A controllable choice represented mathematically in a model.

Parameter: An input value treated as known or estimated for a model run.

Objective function: A mathematical measure to maximise or minimise.

Constraint: A mathematical representation of a resource, policy, physical, or regulatory limit.

Feasible solution: A set of decisions satisfying all model constraints.

Optimal solution: The feasible solution with the best objective value.

Linear programming: Optimization with linear objectives and constraints and generally continuous variables.

Integer programming: Optimization requiring selected variables to take integer values.

Binary variable: A variable restricted to zero or one.

Mixed-integer linear programming: A linear model containing both continuous and integer variables.

Big-M formulation: A logical modelling technique using a sufficiently large constant to link decisions.

Nonlinear optimization: Optimization involving nonlinear objectives or constraints.

Multi-objective optimization: Optimization addressing two or more potentially conflicting objectives.

Pareto frontier: The set of solutions for which improvement in one objective worsens another.

Shadow price: The marginal objective value of relaxing a constraint within a valid sensitivity range.

Reduced cost: An indication of the objective-coefficient improvement needed for a bounded variable to become attractive, subject to solver conventions.

Optimality gap: The measured difference between a best feasible solution and a bound on the optimum.

Heuristic: A solution method designed to find good answers without guaranteeing optimality.

Robust optimization: An approach seeking solutions that remain feasible under specified uncertainty.

Digital twin: A digital representation used to analyse or simulate a physical operation or network.

Review Questions

Short Answer Questions

1. What is the difference between a decision variable and a parameter in an optimization model?

2. Why might a transportation problem require integer programming rather than linear programming?

3. What does a binding constraint indicate?

4. Define a shadow price and state one limitation on its interpretation.

5. What is the purpose of an optimality gap?

Long Answer Questions

1. Explain the structure of a prescriptive supply-chain model and illustrate each component with a logistics example.

2. Compare linear programming, mixed-integer programming, and nonlinear optimization in terms of assumptions and applications.

3. Explain how multi-objective optimization can support decisions involving cost, service, emissions, and resilience.

4. Discuss why solver optimality does not necessarily imply managerial acceptability.

5. Explain the role of scenario analysis, sensitivity analysis, and robust optimization in managing supply-chain uncertainty.

Analytical / Application Questions

1. A distributor must decide which warehouses to open and how much each will ship. Formulate suitable variables, objective terms, and constraints, including a logical link between warehouse opening and shipments.

2. A model recommends a low-cost network that depends on one facility operating at 99 percent capacity. What additional analyses and governance checks should managers undertake before implementation?

3. A company wants to reduce cost while maintaining service of at least 98 percent and limiting emissions. Propose a multi-objective modelling approach and explain how you would present the resulting alternatives to senior management.

22. Multiple Choice Questions with Answers

Question 1

Which statement best describes the role of an objective function in a supply-chain optimization model?

A. It lists all data sources used by the analyst

B. It measures the performance criterion to be maximized or minimized

C. It guarantees that every variable will be an integer

D. It identifies which constraints are infeasible

Correct Answer: B. It measures the performance criterion to be maximized or minimized

Question 2

A warehouse-opening decision is represented by a variable that can take only 0 or 1. What type of variable is this?

A. Continuous variable

B. Slack variable

C. Binary variable

D. Scenario parameter

Correct Answer: C. Binary variable

Question 3

If a warehouse-capacity constraint is binding and has a positive shadow price in a cost-minimization model, what does this generally suggest?

A. Additional capacity may reduce total cost within the sensitivity range

B. The warehouse has unlimited unused capacity

C. The model is necessarily infeasible

D. The capacity constraint should be deleted

Correct Answer: A. Additional capacity may reduce total cost within the sensitivity range

Question 4

Why is an excessively large big-M value potentially harmful in a mixed-integer model?

A. It prevents all binary variables from being used

B. It can weaken the formulation and increase computational difficulty

C. It converts all costs into revenues

D. It guarantees a local optimum only

Correct Answer: B. It can weaken the formulation and increase computational difficulty

Question 5

A solver reports a feasible solution with a non-zero optimality gap. Which interpretation is most appropriate?

A. The solution violates at least one constraint

B. The solver has found a feasible solution but has not proved it is globally optimal

C. The model contains no objective function

D. The solution must be rejected immediately

Correct Answer: B. The solver has found a feasible solution but has not proved it is globally optimal

Question 6

Which method presents alternatives that cannot improve one objective without worsening at least one other objective?

A. Nearest-neighbour search

B. Pareto-frontier analysis

C. Presolve elimination

D. Demand aggregation

Correct Answer: B. Pareto-frontier analysis

Question 7

A routing heuristic is most defensible when which condition applies?

A. A proof of global optimality is legally required

B. The problem must be solved quickly and a good feasible solution is more useful than a delayed proof of optimality

C. All variables are continuous and every coefficient is known exactly

D. The model has no operational constraints

Correct Answer: B. The problem must be solved quickly and a good feasible solution is more useful than a delayed proof of optimality

Question 8

Which situation is an example of a model-formulation error rather than a solver-performance issue?

A. The solver takes several hours on a large model

B. The model permits shipments through a facility that was not opened

C. The solver reports a small optimality gap

D. A heuristic finds a solution quickly

Correct Answer: B. The model permits shipments through a facility that was not opened

Question 9

Why should a low-cost optimization solution be tested under disruption scenarios?

A. To ensure that the objective function becomes nonlinear

B. To determine whether the solution remains acceptable when key assumptions fail

C. To eliminate all decision variables

D. To convert a binary model into a descriptive dashboard

Correct Answer: B. To determine whether the solution remains acceptable when key assumptions fail

Question 10

Which governance practice is most important when optimization affects service allocation and employee workloads?

A. Hide the model assumptions to prevent challenge

B. Remove all human review after deployment

C. Maintain documented assumptions, audit trails, and procedures for reviewing overrides

D. Optimise only the cost metric regardless of consequences

Correct Answer: C. Maintain documented assumptions, audit trails, and procedures for reviewing overrides