Operations Research and Management Science Handbook (The Operations Research Series)

1. 1Name the scarce resource and the goal it serves before choosing a model. The call center example in the passages fixes the pairing: supply is agents and trunks, demand is random, time-varying incoming calls, and the goal is high agent utilization with caller waiting kept to an acceptable minimum.
2. 2Express the allocation as a formal optimization model. Write the objective (minimize cost or waste), the decision variables, and the constraints. The passages' general form is minimize cx+dy subject to Ax+By≥b, with x≥0 continuous and y≥0 integer for indivisible resources such as agents or units.
3. 3Decide whether continuous or integer variables are needed. Solve the LP relaxation first, since its optimal value is a lower bound on a minimization ILP. Expect the relaxed solution to violate integer restrictions unless B is totally unimodular. The handbook notes that LP scales to millions of variables, while ILP/MILP is computationally harder as integer variables grow.
4. 4Where demand is random, model the queue rather than averaging. Use an M/M/s (Erlang C) model: Poisson arrivals at rate λ, exponential service at rate μ, s servers. The service rate is iμ when i≤s agents are busy and sμ when all are busy. Compute utilization and waiting time for each candidate s.
5. 5Check the model's assumptions against reality. Erlang C assumes an infinite number of trunks, so no caller is blocked. Real centers have finite lines, so if blocking matters, use a model that reflects it or treat the Erlang C result as an optimistic bound.
6. 6For networks of stations, break the system into nodes, compute each node's utilization ρ_i, and analyze each as a GI/G/s queue. This requires arrival and service rates plus the squared coefficient of variation (SCOV) of interarrival and service times, so you can see which node is the bottleneck and which is idle.
7. 7Compare candidate allocations by the same pair of measures: utilization of the resource and the service level or waiting cost. Choose the allocation that meets the service target with the least resource, then re-run it when demand parameters change.

The passages' call-center case (Example 8.16) shows the method. The challenge is to balance supply (agents and trunks) against demand (random, time-varying calls) so that agent utilization is maximized while caller waiting stays acceptably short. Model it as Erlang C, an M/M/s queue: arrivals are Poisson with rate λ, service times are exponential with rate μ, and there are s agents. The death rate is μ_i = iμ for 0<i≤s and sμ for i>s. You evaluate candidate values of s. Too few agents leaves callers waiting; too many leaves agents idle, which is waste. Erlang originally used the model to compute the fraction of callers in a small village who had to wait because all lines were in use. The passages flag two drawbacks of Erlang C, the first being its infinite-trunk assumption, which means no blocking. So you treat the output as a first estimate and check it against the finite line count.

• If the resource is divisible (money, hours, material), start with an LP. If it is indivisible (agents, machines, units), use an ILP/MILP, and use the LP relaxation as a lower bound on cost.
• If demand is random and customers wait, size capacity with a queueing model (Erlang C for a single pool of s agents) instead of average-demand arithmetic.
• If utilization is high but waiting is unacceptable, add servers. If waiting is minimal but utilization is low, remove servers. Judge both measures together, never one alone.
• If the system is a network of stations, analyze each node separately (decomposition) and look for the node with the highest ρ_i as the bottleneck.
• If the model assumes infinite lines or other unrealistic limits, treat its answer as optimistic and verify against the real constraint.

• Maximizing utilization alone: agents look fully busy, but caller waiting climbs. The tell is a high utilization figure beside long waits.
• Sizing to average demand: arrivals are random and time-varying, so the average hides peaks. The tell is chronic queues at predictable busy periods.
• Trusting a model whose assumptions fail, such as Erlang C with infinite trunks when real lines are finite. The tell is blocked callers that the model says do not exist.
• Rounding the LP relaxation answer to integers and assuming it is optimal. This holds only when B is totally unimodular, which is generally not the case. The tell is a rounded solution that violates constraints or costs more than a true ILP solution.
• Optimizing each node in a network in isolation without tracking how flows merge and split. The tell is a downstream node overloaded after an upstream node is improved.

• ▸Goal: state the outcome the resource must produce (e.g., calls answered within an acceptable wait).
• ▸Scarce resource: list the resource and its unit (agents, trunks, money, hours).
• ▸Divisible or indivisible: mark LP (continuous) or ILP/MILP (integer) and why.
• ▸Demand pattern: record arrival rate λ and whether it varies over time.
• ▸Service rate: record μ per resource unit.
• ▸Candidate capacity levels: list the values of s (or budget levels) to test.
• ▸Utilization at each level: record the computed utilization for each candidate.
• ▸Waiting/service-level result: record waiting time or shortfall for each candidate against the target.
• ▸Model assumptions checked: note any assumption (e.g., infinite trunks) that differs from reality and its likely effect.

1. 1Split effectiveness into the three components in the definition: financial results (cost, revenue, profit), market position, and achievement of strategic goals. Write one measurable objective per component. Passages that model effectiveness do it as a single quantity to minimize or maximize, such as feed cost, transport cost, or overall profit over a planning horizon.
2. 2Name the decisions that move each objective, and write each as a decision variable with its constraints. Examples from the sources: the feed mix, shipments from mines to plants, booking limits per origin–destination class, and the production-storage-marketing plan per period.
3. 3Model at the level where the objective is actually determined. Revenue management is the example: the sources say the right frame is the itinerary (network), not the single leg. Selling a seat earns revenue and also costs revenue elsewhere, because other itinerary fares lose a seat. Price every decision at its marginal network value, not its local value.
4. 4Limit the planning horizon to the periods where you can estimate data with reasonable precision. The multiperiod sources cover the current period plus a few following ones. They need estimates of demand, selling price, raw-material cost, labor, machine time, and storage cost for each period. Do not model periods beyond what the data supports.
5. 5Re-solve whenever inputs change, and use sensitivity analysis to test offers before accepting them. Feed makers re-run the diet model as grain and hay prices and supplies change. Check whether each new offer lowers overall cost before committing.
6. 6Where the strategy has several competing goals, do not collapse them into one weight silently. Use an interactive multi-criteria method: the decision maker states aspiration levels or trade-off ratios, sees the resulting solution, and revises until it is accepted as the best compromise. Prefer methods that avoid presenting dominated solutions.
7. 7Add process control so that the plan is executed consistently. Chart key process measures with control charts. When a process goes out of control, the probability of nonconformance rises and quality deteriorates. Find and eliminate assignable causes to cut variability.
8. 8Assign ownership across the whole organization. The quality sources state that improvement must be company-wide, with every unit participating, and that statistical methods give all units a common language for problems. The revenue management source warns that network optimization also creates organizational and data-collection problems, because revenue targets are often assigned by geography.

The source describes a farmer who runs cattle and chicken farms and carries a laptop with LP-based diet models for each feed formulation. In harvest season he travels through agricultural areas collecting offers on raw materials. Before accepting any offer he enters the new price into the model and uses the LP software's sensitivity analysis to check whether the offer would lower his overall feed cost. He accepts only offers that do, and he reports that the practice has saved him substantial costs. The financial objective is total feed cost, the decision variable is the ingredient mix, and the constraints are the nutritional requirements. The test is applied to every offer at the moment of decision, so no single supplier's price is judged in isolation. The same structure carries over to the other models in the sources: the transportation model minimizes shipping cost across mines and plants, and the seasonal model maximizes overall profit by producing when costs are low, storing, and selling when prices are high.

• If a decision affects revenue or cost elsewhere in the system (a seat sold, an input purchased), then value it by its marginal effect on the whole system, not its local price.
• If input prices, supplies, or demand change, then re-solve the model with the new data before committing; run sensitivity analysis on each external offer.
• If data for later periods is unreliable, then plan only over the current period and the few that follow, and re-plan as new data arrives.
• If goals conflict (cost, market position, strategic targets), then use an interactive method with explicit aspiration levels or trade-offs instead of a hidden single weighting. Reject methods that present dominated solutions.
• If a process measure drifts out of control, then look for assignable causes and remove them before adjusting the plan.

• Optimizing one unit or leg in isolation: the local result looks good while network revenue falls. The tell is that unit targets are met while the total misses, often because targets are assigned by geography.
• Planning too far ahead on weak data: the model gives precise answers to inputs that are guesses. The tell is that late-period forecasts are large revisions of earlier ones.
• Using a stale model: decisions rest on old prices or supplies. The tell is that purchase or booking decisions differ from what a fresh run would give.
• Treating quality as one department's job: control charts exist but only one unit acts on them. The tell is repeated out-of-control signals with no assignable cause removed.
• Collapsing multiple objectives into a single weighted number without the decision maker's input, so the result is a dominated or unacceptable compromise. The tell is that the decision maker rejects the solution and cannot say why.

• ▸Financial objective: state the quantity to minimize or maximize (cost, revenue, profit) and its time horizon.
• ▸Market-position and strategic objectives: list each as a measurable target, with an aspiration level.
• ▸Decision variables: list what you control (mix, shipments, booking limits, production and storage by period).
• ▸Constraints: list capacities, requirements, and availability limits.
• ▸System-wide effect: note which other units or itineraries each decision affects, and estimate the marginal value lost or gained there.
• ▸Planning horizon and data confidence: list the periods you will model and the data source behind each estimate.
• ▸Sensitivity check: record the input or offer tested and whether overall cost or profit improved.
• ▸Process control measure: name the chart, its current state (in or out of control), and the assignable cause found.
• ▸Owner and review date: name the unit responsible and when to re-solve.

1. 1Classify the problem before choosing a technique. The handbook's history chapter says management problems fall into classes such as allocation, inventory, network, queuing, replacement, and scheduling. Once a problem is identified as belonging to a class, all the models and techniques available for that class can be used to study it. Write the problem in one sentence and name its class.
2. 2Match the class to a method family from the handbook's Part I: linear programming, nonlinear programming, integer programming, network optimization, multiple criteria decision making, decision analysis, dynamic programming, stochastic processes, simulation, metaheuristics, robust optimization. Allocation of scarce resources points to LP. Yes/no or discrete choices point to integer programming. Sequential decisions point to dynamic programming. Systems too complex for closed form point to simulation.
3. 3Define the decision in the ORSA terms quoted in the passages: how to best design and operate a man–machine system, usually under conditions requiring allocation of scarce resources. List the scarce resources (aircraft, crews, airport facilities, capital) and the decision variables that allocate them.
4. 4Decompose a large problem into sequential stages when one model is unmanageable. The airline passage does this: Schedule Design, then Fleet Assignment, then Aircraft Routing, then Crew Scheduling. Combinations of two or more steps in a single model are possible. Write down what each stage passes to the next.
5. 5Set explicit evaluation criteria before modeling. The airline schedule example uses two: profitability (revenue from origin–destination pairs versus operating cost) and feasibility (ability to assign sufficient airplanes, crew, and airport facilities). Name your criteria. If they conflict, route to the multiple-criteria or goal programming methods.
6. 6Model intelligently, not just mechanically. The LP chapter lists 'Intelligent Modeling Essential to Get Good Results' with a container shipping example. Before solving, check that the model's variables and constraints reflect how the system really operates.
7. 7Work to the planning calendar. The airline schedule starts about 1 year before operation and 6 months before publication. Back-schedule the modeling work so results arrive before the decision deadline.
8. 8Check the model against uncertainty. The handbook includes decision analysis (risk and uncertainty), stochastic inventory, simulation, and robust optimization. If inputs such as demand are forecasts, test the recommendation under alternative values rather than trusting a single deterministic run.

Airline schedule planning, taken from the handbook's airline chapter. The problem is class 'scheduling' under scarce resources: airplanes, crews, and airport facilities. The airline does not build one monolithic model. It works through four steps in order. (1) Schedule Design defines the markets to serve, how often, and the origin, destination, and departure times of flights. (2) Fleet Assignment designates aircraft size for each flight. (3) Aircraft Routing selects each aircraft's daily sequence of flights and satisfies maintenance requirements. (4) Crew Scheduling assigns crews to flights. Two criteria govern the design step. Profitability compares passenger and cargo revenue from city pairs against operating cost. Feasibility asks whether enough airplanes, crew, and facilities exist. Construction starts about 1 year ahead, and the service plan draws on demand forecasts, competitive information, and marketing initiatives. Schedules follow seasonal demand more than month-to-month differences. The method choice follows from the structure: staged decomposition, with some steps combined in a single model where interactions justify it.

• If the problem fits a known class (allocation, inventory, network, queuing, replacement, scheduling), then start with the models already built for that class rather than inventing a new one.
• If the full problem is too large for one model, then split it into ordered stages like the airline's four steps. Combine adjacent stages only when their interaction is material.
• If a decision must satisfy both an economic criterion and a resource-availability criterion, then model profitability and feasibility separately and check both.
• If key inputs are forecasts or random, then use stochastic or simulation methods or test alternative input values rather than a single deterministic optimum.
• If the model is technically correct but results look poor, then revisit the formulation. The handbook stresses that intelligent modeling is essential to get good results.

• Technique-first thinking: picking a favorite method (say, LP) before classifying the problem. The tell: the model needs awkward workarounds to represent the decision.
• Treating one stage's output as final when later stages constrain it. The tell: a schedule that is profitable on paper but cannot be staffed or routed with the available aircraft, crews, and facilities.
• Optimizing a single criterion. The tell: a plan with strong revenue that fails the feasibility check, or the reverse.
• Starting too late for the planning cycle. The tell: model results arrive after the schedule is already being published.
• Trusting deterministic inputs. The tell: the recommendation changes sharply when demand forecasts shift slightly, and no one has tested that.

• ▸Decision statement: one sentence on what is to be designed or operated.
• ▸Problem class: allocation, inventory, network, queuing, replacement, scheduling, or other.
• ▸Scarce resources: list each resource that limits the decision.
• ▸Candidate method family: LP, IP, network, DP, simulation, metaheuristic, decision analysis, or robust optimization.
• ▸Stages: ordered sub-problems and what each passes to the next.
• ▸Criteria: for example profitability and feasibility, with how each is measured.
• ▸Uncertain inputs: forecasts or random quantities and the alternative values to test.
• ▸Decision deadline and back-scheduled modeling milestones.
• ▸Model validity check: how you confirmed the formulation reflects real operations.

1. 1Before touching a solver, list the data each model type demands. For a product mix LP the passage names: units of each resource per unit of product (I/O or technology coefficients), bounds on production per period, bounds on resource availability per period, expected demand per product, and cost or net profit per unit. Missing any one of these means the model is under-specified.
2. 2Run the input-output analysis of the company: assemble every technology coefficient and resource limit into one table. The source calls assembling this reliable data one of the most difficult jobs in building a product mix model, and says it is worth the effort. Budget time for it as a task of its own.
3. 3For transportation models, collect three things only: availability at each source, requirement at each market, and unit shipping cost for every source-market pair. Check that every source-market cell has a cost or a deliberate 'no route' flag, so no cell is blank by accident.
4. 4For multi-period planning, get period-by-period estimates, not annual averages: demand, expected selling price, raw material, labor and machine availability and cost, and storage space and cost. The source says you need 'reasonably good estimates' of each, and the seasonal profit logic depends on them varying by period.
5. 5Separate fixed data from estimated data. Mark quarry supply caps (1500, 1200, 500 tons/day) or storage cost ($2/ton per period) as given. Mark demand and selling price as forecasts. Spend validation effort on the estimated items.
6. 6Reconcile the units and definitions of every field. The production cost in the six-period table bundles raw material, labor, and machine time, and the capacity bundles raw material and hourly labor limits. Confirm what is inside each number, and which constraint it implies, before it enters the model.
7. 7Add the boundary conditions the data table omits: initial stock (500 tons) and required ending stock (500 tons). Without them the model drains inventory or ignores carryover.
8. 8Validate by sanity-checking the solved plan against the data. Check that production never exceeds the period capacity, that sales never exceed demand, and that stock balances period to period. If the plan looks wrong, trace it to a data field first and the formulation second.

The source's six-period production planning problem shows how much the data table carries. The company plans 2-month periods for the year. Per-ton production cost rises from $20 to $60, capacity runs 1500 to 3000 tons, demand 1100 to 2500 tons, and selling price $180 to $320. Storage costs $2/ton per period, and stock starts and ends at 500 tons. Check each row for consistency: in period 1, capacity of 1500 exceeds demand of 1100, and price is $180 against cost of $20, so producing is profitable and surplus can be stored for later. In period 6, demand is 2500 against capacity of 2500, so the stored stock matters. Production cost of $60 against price of $320 is still profitable. The seasonal logic only works because cost and price vary by period, so replacing them with averages would erase the stock-building decision. The ending-stock requirement of 500 tons has to be entered as a constraint, or the model sells it all in period 6 and looks better than the real plan. Every number needs an owner and a source before the LP is run.

• If a coefficient is a forecast (demand, selling price), then document its basis and test the plan against a range of values. If it is a physical or contractual limit (quarry supply, capacity), then treat it as fixed.
• If the data you can get are only annual or aggregate but the decision is seasonal or periodic, then collect period-level data before modeling. Averages remove the trade-off the model exists to find.
• If I/O coefficients cannot be gathered reliably, then delay the product mix model rather than guess. The source ranks this as the hardest and most worthwhile step.
• If a source-market cost or a resource requirement is missing, then resolve it explicitly, either by obtaining it or by barring the route. Never leave a default zero.
• If the solution contradicts operational experience, then audit the inputs against the data table before changing the formulation.

• Treating data assembly as a quick prelude. The tell: the formulation is finished in days and the I/O coefficients are still 'rough numbers'. The source calls this one of the most difficult jobs.
• Using annual averages for seasonal items. The tell: the plan shows no stock-building and ignores the cost and price swings in the table.
• Omitting boundary data such as initial and required ending inventory. The tell: stock drains to zero in the final period, giving a profit that cannot be achieved.
• Mixing definitions inside a cost or capacity figure. The tell: production cost is double-counted because labor or machine time also appears as a separate constraint, or capacity is stated as raw material in one period and labor in another.
• Mismatching units or time bases across sources. The tell: tons/day supply limits are compared with per-period demand, or cents/ton costs with dollar figures, without conversion.

• ▸Model type: LP product mix, transportation, multi-period planning, etc., which sets the required data list.
• ▸Data item: the parameter name, such as I/O coefficient, availability at mine, demand, selling price, storage cost.
• ▸Unit and time base: tons/day, $/ton, per 2-month period; state it exactly.
• ▸Fixed or estimated: mark each item as given limit or forecast.
• ▸Source and owner: where the number comes from and who can confirm it.
• ▸Definition check: what is bundled inside the figure (e.g., raw material, labor, machine time in production cost).
• ▸Gap or boundary condition: missing cells, initial stock, required ending stock, bounds.
• ▸Quality rating and action: confidence level, plus the step needed (re-measure, forecast range, bar route).
• ▸Validation test: the check run on the solved plan to catch a data error.

1. 1Write the system as three lists before any math: the decision alternatives you control, the states of nature you do not, and the payoff measure. The passage's example has five alternatives (SS, CS, Bonds, CD, Do Nothing), three states (strong, stable, weak economy) and expected rate of return as the payoff. If you cannot name all three lists, you do not yet understand the system.
2. 2Build the payoff matrix with one cell per alternative and state combination. Fill every cell with a quantitative estimate and note which cells are statistical estimates rather than facts. The payoff can be money or any other measure, but use one measure throughout.
3. 3Read the matrix row by row before computing anything. For each alternative record its best case, worst case and range. Then read it column by column to see which alternative wins in each state. This shows you the structure: who wins where, and what each choice puts at risk.
4. 4Eliminate dominated alternatives. An alternative is dominated if another one pays at least as much in every state. Removing it shrinks the problem and tells you which options carry no trade-off.
5. 5Decide how to represent the uncertainty. If you can credibly assign probabilities to the states, model the uncontrollable factors probabilistically and compute expected payoffs. If you cannot, state that and compare alternatives on worst-case and best-case payoffs. Keep the two views separate.
6. 6Run sensitivity checks. Change the state probabilities, or the payoff estimates, until the ranking changes. The point where the best alternative flips is the real boundary of the decision. Report that boundary to the decision-maker instead of a single answer.
7. 7Check for interdependence. If the decision spans several stages or shares limited resources, a decision in one stage cannot be made in isolation. The dynamic-programming view splits the problem into linked stages, so model the linkage explicitly rather than optimizing each stage alone.
8. 8Say whether the model is deterministic or stochastic. A deterministic model predicts one outcome. A stochastic model predicts a set of outcomes with the probability of each. Tell the decision-maker which one they are looking at, because it sets how much weight a single number can carry.

Take the handbook's investment problem. The payoff matrix (rate of return, % in strong / stable / weak economy) is: Speculative Stock 18/10/−5; Conservative Stock 13/8/1; Bonds 4/5/6; CD 7/3/2; Do Nothing 0/0/0. Reading rows gives worst cases of −5, 1, 4, 2 and 0, so Bonds has the best worst case (4). Ranges differ widely: Speculative Stock spans 23 points, Bonds only 2. Dominance check: CD pays at least as much as Do Nothing in every state, so Do Nothing is dominated and drops out. No other alternative is dominated. CS beats Bonds in strong and stable economies (13 vs 4, 8 vs 5) but loses in a weak one (1 vs 6), which is a genuine trade-off. As an illustrative assumption only, since the passage gives no probabilities, give the three states equal weight. Expected returns are then SS ≈ 7.67, CS ≈ 7.33, Bonds 5, CD 4. SS and CS look almost tied, yet SS can lose 5% while CS never loses. The expected value hides that difference, and the matrix shows it. The understanding the model produces is this: more return costs more downside, and the ranking depends on how likely a weak economy is.

• If one alternative pays at least as much as another in every state, drop the weaker one and do not spend further analysis on it.
• If the alternatives' worst-case payoffs differ by much more than their expected payoffs do, present the downside next to the average and let the decision-maker's risk tolerance choose.
• If you cannot defend probabilities for the states, then compare on best and worst cases and say explicitly that no expected value was computed.
• If a small change in a probability or payoff estimate flips the top-ranked alternative, then report the flip point as the finding and do not report one winner.
• If the decision in one stage uses resources that later stages also need, then model the stages together as linked subproblems and do not optimize each stage separately.

• Ranking only by expected payoff. Tell: two alternatives look nearly tied (SS 7.67 vs CS 7.33) and nobody has asked about the −5% cell.
• Treating estimated payoffs as facts. Tell: the matrix has no note on which cells are statistical estimates, and no range or sensitivity check on them.
• Presenting a deterministic answer for a stochastic system. Tell: the output is a single number with no set of outcomes or probabilities, and the decision-maker asks why reality differed.
• Optimizing one stage in isolation when resources are shared. Tell: early-stage choices look good, but later stages run out of the resource.
• Handing over the solution without the structure. Tell: the decision-maker cannot say which state makes the recommended option win or lose, or what would change the recommendation.

• ▸Decision alternatives: list every option you control, including 'do nothing'.
• ▸States of nature: list the uncontrollable conditions (for example strong / stable / weak economy).
• ▸Payoff measure: name the single metric and its unit (for example expected rate of return, %).
• ▸Matrix: one payoff per alternative-state cell, marked as given or estimated.
• ▸Row best / worst / range: record these for each alternative to expose its risk profile.
• ▸Dominated alternatives: name any option beaten or matched in every state, and by which option.
• ▸Probability basis: give state probabilities and their source, or write 'none defensible' and use worst-case comparison.
• ▸Flip point: record the probability or payoff change at which the top-ranked alternative changes.
• ▸Model type and linkages: mark deterministic or stochastic, and note any shared resources or stage dependencies.

1. 1Write the decision as one sentence that names the choice, the decision-maker, and the deadline. The handbook frames OR as 'scientifically deciding how to best design and operate' systems 'under conditions requiring the allocation of scarce resources', so name the scarce resource too.
2. 2List the criteria before you list the alternatives. Give each criterion a unit and a direction (minimize or maximize). Mark which criteria are hard constraints, like the fill-rate requirement in the inventory example, and which are objectives to trade off.
3. 3Normalize the criteria so that unlike units can be compared. The handbook's MCDM chapter treats criteria normalization as a separate step that comes before weighting and before any method for finite alternatives.
4. 4Compute the criteria weights explicitly and write them down before you score anything. The handbook treats weight computation as its own step (MCDM, section 5.4). Weights set after the scores are visible are a political choice dressed as analysis.
5. 5Classify the uncertainty. The handbook's decision-analysis chapter splits decision making under risk from decision making under uncertainty. Use probabilities only where you can defend them, and say which case you are in.
6. 6Score each alternative against each criterion from data or a stated model, and record the source of each cell. Pick the method that fits the problem: a finite list of alternatives calls for a multiple-criteria method for finite alternatives, and a continuous decision space calls for a mathematical programming formulation.
7. 7List every simplifying assumption next to the result. The inventory chapter, for example, notes that normal lead-time demand is used because it is 'convenient from an analytic standpoint' and that other distributions usually matter little. Flag each assumption so a reviewer can challenge it.
8. 8Compare the rigorous answer with any quick-and-dirty shortcut. State what the shortcut ignores. If you adopt the shortcut, record the reason and what it costs.

The supply chain chapter in the passages gives a real case of rigor versus shortcut: choosing the order quantity Q and reorder point R for a continuous-review inventory system. The rigorous framing sets an explicit objective, minimizing total annual cost TAC(R, Q) (Eq. 22.10), subject to an explicit criterion, a fill rate β(R, Q) = 1 − σ_DLT·L(z)/Q (Eq. 22.11). The problem is non-linear in both the objective and the constraint. The objective is convex in Q and R, so the solution is 'fairly straightforward'. The less rigorous route solves for the two variables separately. First, set R to hit a desired cycle service level, using the fractile z = (R − μ_DLT)/σ_DLT of lead-time demand. Then apply the first-order condition ∂TAC/∂Q = −AD/Q² + h/2 = 0 to get Q* = √(2AD/h), the EOQ. The rigor shows in what you can say about each route. The joint model evaluates both decisions against the stated criteria together. The sequential shortcut fixes R from a service target before it knows Q, even though the fill-rate formula depends on both. It is exact only in the limiting case where demand uncertainty is removed, the safety-stock term drops out, and EOQ is optimal. Using the shortcut is defensible if you write down that it ignores this coupling.

• If a criterion is a hard requirement (like a fill-rate target), then model it as a constraint. If it is something you are willing to trade off, then give it a weight.
• If the criteria have different units, then normalize them before you apply any weights or sum any scores.
• If you can defend probabilities for the outcomes, then analyze the decision under risk. If you cannot, then use the decision-under-uncertainty methods and say so.
• If the objective is convex (as TAC is in Q and R), then solve the joint problem directly instead of settling for a separate-variable shortcut.
• If you use a simplified method or a convenient distributional assumption, then record what it ignores and the condition under which it is exact.

• Setting weights after seeing the scores. The tell is that the favored option wins under every weighting you try, and nobody can say when the weights were fixed.
• Summing criteria in different units without normalization. The tell is that the criterion with the largest numeric scale decides the outcome.
• Treating a convenient assumption as a fact. The tell is that the normal-demand assumption appears in the model but nowhere in the memo or the sign-off.
• Optimizing variables one at a time when they are coupled. The tell is that the chosen R and Q meet their individual targets but miss the joint fill-rate requirement.
• Writing the criteria after the alternative is chosen. The tell is that the criteria list reads like a description of the preferred option.

• ▸Decision statement: one sentence naming the choice, owner, deadline, and scarce resource.
• ▸Hard constraints: each requirement, its unit, and its threshold (e.g., fill rate target).
• ▸Objective criteria: each criterion with its unit and direction (minimize or maximize).
• ▸Normalization method: how unlike units were put on a comparable scale.
• ▸Weights and date set: each criterion's weight, how it was derived, and the date it was fixed (before scoring).
• ▸Uncertainty class: risk (probabilities defensible) or uncertainty (not), with the justification.
• ▸Alternatives and scores: a matrix of alternatives against criteria, with the data source for each cell.
• ▸Assumptions and shortcuts: each simplification, what it ignores, and when it is exact.
• ▸Result and challenge: the ranked outcome, plus the one change in weights or assumptions that would flip it.

1. 1State the decision as a model before you compare options: list the decision variables, the objectives or targets, and every constraint. The handbook's blending problem is the template: four scrap types, four element targets, plus a sum-to-1 equation.
2. 2Write the nonnegativity and other feasibility restrictions explicitly. The handbook notes that early solvers dropped nonnegativity because nobody knew how to handle inequalities. Any candidate that satisfies the equations but breaks a bound is not a decision.
3. 3Choose the method by the shape of the choice. For a finite list of alternatives, normalize the criteria, compute criteria weights, and rank. For a continuous solution space, use multiple-criteria mathematical programming: goal programming, global criterion/compromise programming, or interactive methods.
4. 4Express strategic alignment as targets or aspiration levels per objective. In goal programming and aspiration-level methods the decision maker (DM) states the values they hope to reach, and the gap between those values and the outcome is the alignment score.
5. 5Test for dominance before presenting options. The handbook warns that feasible region reduction methods may show the DM a dominated solution. Drop any option that another option beats on every criterion.
6. 6Classify the uncertainty before picking the decision rule. Decision analysis separates decision making under risk (probabilities known) from decision making under uncertainty (probabilities not known). Use the matching toolset instead of defaulting to expected value.
7. 7Probe robustness by perturbing the inputs: coefficients, targets, weights. Note which variables sit near a bound. The handbook lists simulation and robust optimization as the dedicated tools when a point solution is too fragile.
8. 8Close the loop with the DM. In interactive methods, iterate (calculate, ask, narrow) until the DM judges the current solution the best compromise. Record the DM's cognitive burden and any inconsistent answers, since the handbook rates methods on both.

The handbook's scrap metal blending problem has four scrap types (SM-1 to SM-4) and target percentages Al 4.43, Si 3.22, C 3.89, Fe 88.46. That gives five equations in four proportions: 5x1+7x2+2x3+x4=4.43; 3x1+6x2+x3+2x4=3.22; 4x1+5x2+3x3+x4=3.89; 88x1+82x2+94x3+96x4=88.46; x1+x2+x3+x4=1. Subtracting the sum equation from the first three and combining gives x1+2x2=0.93 and 5x1+12x2=5.33, so x2=0.34 and x1=0.25. Then x3=0.39 and x4=0.02. The Fe equation checks: 22+27.88+36.66+1.92=88.46. Feasibility holds because every proportion is nonnegative, and the mixture hits the targets exactly. That is alignment and optimality given the data. The robustness read is my inference, not the handbook's: x4=0.02 sits close to the nonnegativity bound, so a small shift in a composition figure or target could push it negative and make the blend infeasible. A good decision record would state the blend and flag SM-4 as the fragile component.

• If a candidate satisfies the equations but violates any nonnegativity or other inequality constraint, then reject it as infeasible, however well it matches the targets.
• If another alternative is at least as good on every criterion, then eliminate the dominated one before the DM sees the shortlist.
• If the DM can state target values per objective, then use goal programming or aspiration-level methods. If the DM can only react to trade-offs, use an interactive method and monitor cognitive burden.
• If outcome probabilities are known, then treat the problem as decision under risk. If they are not, then use the decision-under-uncertainty rules instead of inventing probabilities.
• If a chosen solution has a variable near a bound or a thin margin, then rerun with perturbed data or a robust or simulation approach before committing.

• Ignoring inequality constraints and accepting a solution that solves the equations only. The tell: some quantity in the answer is negative or physically impossible.
• Presenting a dominated option as a candidate. The tell: the DM can name an alternative that is better on every criterion.
• Treating the method as objective when it is ad hoc. Feasible region reduction has no underlying preference function. The tell: results change with the order of the DM's answers.
• Applying risk tools when probabilities are unknown, or the reverse. The tell: precise probabilities with no stated source.
• Calling a point solution optimal without a sensitivity check. The tell: one variable at 0.02 and no record of how far the inputs can move before the plan breaks.

• ▸Decision and objectives: state the choice and the target or aspiration level for each objective.
• ▸Variables and constraints: list the decision variables, every equation, and every bound, including nonnegativity.
• ▸Method chosen and why: finite-alternative ranking, goal programming, compromise programming, or interactive; one line of rationale.
• ▸Dominance check: list the alternatives removed and the alternative that beats each on every criterion.
• ▸Alignment gap: for each objective, compare the chosen solution's value with the target and note the shortfall or overshoot.
• ▸Uncertainty type: risk (probabilities known) or uncertainty (unknown); name the decision rule applied.
• ▸Fragile points: list variables or margins near a bound and the input change that would break them.
• ▸Robustness test run: record what you perturbed (data, targets, weights) and whether the choice held.
• ▸DM sign-off: record whether the DM accepted the best compromise and any inconsistent responses.