Stochastic programming chooses decisions under uncertainty by representing possible outcomes and their probabilities, while respecting what can be known when each action is taken. Its gold standard is an implementable policy: a decision made now, permitted adjustments later, and a transparent account of expected cost, service and risk. It does not predict which future will occur.
This guide builds a complete capacity-planning example, calculates the value of modelling uncertainty and the value of information, and explains why scenario trees, recourse limits and independent testing matter. Every numerical case is an original teaching model with stated assumptions. The purpose is to understand the reasoning well enough to change the model when the situation changes.
Start with the Gold Standard master, the Education parent and the How X Works hub. The neighbouring Robust Optimization root develops protection over uncertainty sets; this root develops probability-weighted decisions and the timing of adaptation.
Choose your stochastic-programming route
Chapters 1–3: Understand decisions and recourse
Chapters 4–6: Compare policies and information
Chapters 7–9: Model limits, risk and scenarios
Chapters 10–12: Sample, validate and diagnose
Chapter index
Understand decisions and recourse · Chapters 1–3
1. Begin with a timeline, not an average forecast
Compare policies and information · Chapters 4–6
4. Compare the mean-demand shortcut honestly
Model limits, risk and scenarios · Chapters 7–9
7. Put realistic limits on recourse
Sample, validate and diagnose · Chapters 10–12
10. Understand sample average approximation and its optimism
Transfer and continue · Chapters 13–14
13. Transfer the model when probabilities and decisions change
CHAPTER 1 OF 14 · Understand decisions and recourse
1. Begin with a timeline, not an average forecast
A planner reserves capacity before demand is known. Once demand arrives, additional capacity may be bought at a higher price. The initial reservation must be one decision across all possible futures; the later purchase may vary with the demand actually observed. Those two sentences define more of the problem than a complicated objective formula does.
Write down the sequence explicitly. First, choose the reservation. Second, observe the available demand information. Third, buy permitted emergency capacity. Fourth, deliver the service and count costs. If the emergency purchase must be ordered before demand arrives, this sequence is wrong and the model must change.
Identify the decision and the random quantity
A random demand is not a decision variable merely because it appears in an optimisation model. The planner controls the reservation and later purchase, while the scenario determines demand. Allowing the solver to select demand would turn uncertainty into a convenient choice and usually make the answer meaningless.
Likewise, a probability is not automatically an objective weight expressing preference. It describes the assumed frequency or belief assigned to a scenario. A separate risk preference may determine how strongly adverse outcomes matter. Mixing those roles makes it difficult to explain whether a change in the answer came from new evidence or a changed priority.
The word programming here means mathematical optimisation. A stochastic program may be solved using deterministic algorithms once its scenarios have been assembled. It is distinct from a program that happens to use random numbers, and from stochastic gradient methods used to train a model. The uncertainty belongs to the decision problem, not necessarily to the algorithm.
A useful first deliverable is a short decision ledger: what is chosen, what is uncertain, what is observed, what can be changed and what cannot. Include physical units and deadlines. This ledger prevents a capacity measured in hours from being combined with demand measured in people without a conversion, or tomorrow’s information from being used for today’s commitment.
CHAPTER 2 OF 14 · Understand decisions and recourse
2. Build the two-stage model from first principles
Let q≥0 be capacity reserved now at a cost of three per unit. Demand D will be either four units with probability 0.75 or twelve units with probability 0.25. After observing demand, emergency capacity e≥0 can be purchased at ten per unit. It is immediately available without an upper limit. Unused reservations have no salvage value, and all demand must be served.
For a given q and realised demand d, the cheapest emergency action is e(d)=max(0,d−q). Its cost is Q(q,d)=10max(0,d−q). The total expected cost is F(q)=3q+0.75Q(q,4)+0.25Q(q,12).
Keep one commitment and separate later actions
The equivalent finite linear program uses q,e₄,e₁₂ as nonnegative variables. Its objective is 3q+7.5e₄+2.5e₁₂. The service constraints are q+e₄≥4 and q+e₁₂≥12. There is one q because reservation occurs before demand is known, but two emergency variables because purchases occur afterwards.
Probabilities weight the objective, not the amount of service required in each scenario. Replacing q+e₁₂≥12 by q+e₁₂≥3 because the high-demand probability is one quarter would confuse expected demand with conditional demand. When the high-demand scenario occurs, twelve units are needed.
The GAMS deterministic-equivalent documentation gives the general two-stage structure and its scenario expansion. The simple model here makes that structure visible without requiring software: one shared decision, scenario-specific recourse and probability-weighted costs.
These assumptions deliberately make emergency capacity easy to obtain. That is useful for learning but often unrealistic. A real model may need lead times, integer vehicles, limited suppliers or demand that can only partly be observed. Those restrictions belong in the equations before a result can support an operational decision.
CHAPTER 3 OF 14 · Understand decisions and recourse
3. Solve the capacity example and explain the optimum
For 0≤q≤4, both scenarios need emergency capacity. Expected cost is 3q+10(6−q)=60−7q because mean demand is six. Cost decreases as reservation increases. For 4≤q≤12, only the high-demand scenario needs emergency capacity, giving F(q)=3q+2.5(12−q)=30+0.5q.
For q≥12, no emergency capacity is required and F(q)=3q. The first region slopes downward, the second upward, and the third upward. Therefore q*=4 is the global optimum, with expected cost 32. The policy buys zero emergency units at demand four and eight emergency units at demand twelve.
Expected cost is not the realised invoice
The low-demand invoice is twelve. The high-demand invoice is ninety-two: twelve for the reservation and eighty for emergency supply. Their weighted average is 0.75·12+0.25·92=32. It would be misleading to say that the policy costs thirty-two in each scenario.
This large difference between average and adverse cost is central to interpreting the answer. The model minimises expected expenditure while serving demand through unlimited emergency supply. It does not impose an upper budget in every scenario. A planner unable to pay ninety-two has a constraint missing from the model, not a reason to ignore the high-demand outcome.
The marginal calculation explains the result. Between four and twelve, another reserved unit costs three and avoids a ten-unit emergency charge only with probability 0.25. Its expected saving is 2.5, less than its reservation cost. Buying that extra unit increases expected cost by 0.5.
Below four, another reservation avoids emergency purchases in both scenarios, saving ten while costing three. Beyond twelve, it saves nothing. This economic explanation is as important as the algebra because it identifies exactly which price or probability changes could move the optimum.
CHAPTER 4 OF 14 · Compare policies and information
4. Compare the mean-demand shortcut honestly
Mean demand is 0.75·4+0.25·12=6. A deterministic model that replaces demand by six reserves q=6 because reservation costs three and emergency supply costs ten. Its reported deterministic cost is eighteen. That number is not its expected cost under the original uncertain demand.
Fix q=6 and restore the two scenarios. Low-demand cost is eighteen. High-demand cost is eighteen plus sixty, or seventy-eight. Expected cost is 0.75·18+0.25·78=33. The mean-demand policy is feasible but costs one more in expectation than the stochastic optimum.
Distinguish a forecast model from policy evaluation
The value of the stochastic solution, for this minimisation example, is 33−32=1. It compares two implementable first-stage decisions under the same probability model and the same recourse rules. Subtracting the deterministic model’s eighteen from thirty-two would compare different evaluations and would not calculate this value.
The mpi-sppy documentation on the value of the stochastic solution distinguishes the stochastic optimum from evaluating the mean-value decision across scenarios. Its distinction is useful even when every calculation is done by hand.
A small value does not mean uncertainty is unimportant. Here the two policies have very different worst invoices, seventy-eight and ninety-two, even though their means differ by one. The chosen criterion makes that distinction secondary. If a budget limit matters, evaluate it explicitly rather than assuming a small expected-cost gap settles the decision.
Nor is the average-data policy always feasible when returned to the original scenarios. Limited emergency capacity can make it fail. In that case its expected cost may be treated as infinite in a hard-service model, or a separately defined shortage penalty may be used. State the interpretation instead of hiding infeasible scenarios from the evaluation.
CHAPTER 5 OF 14 · Compare policies and information
5. Calculate the value of perfect and imperfect information
Imagine learning demand perfectly before reserving capacity, while retaining the same reservation price. The planner would reserve four in the low scenario and twelve in the high scenario. Expected cost would be 0.75·12+0.25·36=18. This is a wait-and-see benchmark with information unavailable in the original problem.
The expected value of perfect information is 32−18=14. Under the stated model, fourteen is the maximum expected saving available from learning demand perfectly before the reservation. It is not a forecast that a particular information service will save fourteen.
Test an imperfect signal
Suppose a signal H has probability 0.5 of appearing when demand is high and probability 0.1 when demand is low. Signal L covers the complementary outcomes. The joint probabilities are P(high,H)=0.125 and P(low,H)=0.075, so P(H)=0.2 and P(high|H)=0.625.
For L, the high-demand joint probability is 0.125 and the low-demand joint probability is 0.675. Thus P(L)=0.8 and P(high|L)=0.15625. These conditional probabilities come from the assumed signal accuracy and prior demand probabilities; they are not chosen to favour an action.
After H, reserving twelve is optimal because the expected emergency saving per extra unit is 10·0.625=6.25, greater than three. After L, reserving four remains optimal because 10·0.15625=1.5625 is less than three.
Expected cost with the signal is 0.2·36+0.8·24.5=26.8. The second conditional cost is twelve plus eighty times 0.15625. The signal’s gross expected value is 32−26.8=5.2, below the perfect-information bound. A signal price above 5.2 would erase its expected financial benefit in this model. Delay, implementation cost and calibration error would need further deductions.
CHAPTER 6 OF 14 · Compare policies and information
6. Make non-anticipativity visible in a scenario tree
A scenario tree groups futures according to what has been observed. Its root represents the information available before any new signal. A later node represents an observable history, and branches represent subsequent uncertainty. Decisions at a node must be identical for scenarios sharing that history.
In the imperfect-signal example, reservation occurs after H or L but before final demand. Both the high-demand and low-demand leaves below H must share q_H. Both leaves below L must share q_L. Allowing each leaf its own reservation would secretly replace the imperfect signal with perfect information.
A stage follows information availability
Two decisions made on different calendar days can belong to the same information stage if nothing relevant has been learned between them. Conversely, several decisions within one day may belong to different stages if an observation arrives between them. The tree should represent available knowledge rather than an arbitrary calendar grid.
The specialist guide to multistage stochastic programming, recourse and scenario trees extends this information discipline to longer sequences. This root’s signal example provides a small calculation with which to test the same principle.
Leaf probabilities must equal products of conditional branch probabilities along their paths. In the example, P(H)P(high|H)=0.2·0.625=0.125. If a model multiplies a joint leaf probability by its parent probability again, it counts part of the uncertainty twice and changes the objective.
When reviewing a tree, ask what a real operator sees at each node. If two histories are observationally indistinguishable, separate actions may be impossible even if the mathematical labels differ. Information constraints are operational constraints: a policy that relies on unavailable knowledge is not a usable plan, however favourable its computed cost.
CHAPTER 7 OF 14 · Model limits, risk and scenarios
7. Put realistic limits on recourse
Now limit emergency capacity to five units in either demand scenario. The high-demand service requirement q+e₁₂≥12 forces q≥7. The original optimum q=4 becomes infeasible, and so does the mean-demand reservation q=6. The physical limit changes which policies can serve all demand.
On the remaining interval 7≤q≤12, expected cost is still 30+0.5q. Its minimum is therefore at q=7, with expected cost 33.5. Low-demand cost is twenty-one; high-demand cost is seventy-one. The policy uses all five emergency units in the high scenario.
Diagnose infeasibility before changing the objective
If reservation is also capped at six, no feasible policy remains: total possible service is at most eleven, below high demand twelve. Raising a cost penalty does not make an impossible hard constraint feasible. The planner must add a real action, revise the service requirement or revise the uncertainty model with evidence.
A shortage variable can make the algebra feasible, but it represents unserved demand. If shortage is permitted at a cost, report its amount and meaning. A model that buys feasibility through a shortage variable is not equivalent to a model that guarantees full service.
Relatively complete recourse means that every feasible first-stage decision has a feasible later response for the relevant uncertainty realisations. The unlimited-emergency teaching model has that property. The capped version does not unless the first-stage feasible set is restricted appropriately, for example to q≥7 here.
Recourse can also be delayed or conditional on shared resources. Emergency transport may exist in normal conditions but disappear during the disruption that increases demand. Model that dependence jointly. Treating emergency supply as always available because it was available on average can defeat the very purpose of representing uncertainty.
CHAPTER 8 OF 14 · Model limits, risk and scenarios
8. Change the risk criterion without disguising the change
Expected cost is only one criterion. Under unlimited emergency capacity, the original q=4 policy has worst cost ninety-two, while q=12 costs thirty-six in both scenarios. A worst-cost objective therefore prefers q=12, even though its expected cost thirty-six exceeds the expected-cost optimum thirty-two.
Neither calculation contradicts the other. One minimises a probability-weighted average; the other limits the largest included invoice. Comparing their objective values as though they used the same scoring rule would be an error.
A budget constraint is different from a preference
Suppose total cost may not exceed sixty in either scenario. For q between four and twelve, the high-demand invoice is 120−7q. Requiring it to be at most sixty gives q≥60/7, approximately 8.5714. The expected-cost objective increases over this interval, so the constrained optimum is q=60/7.
Its expected cost is 30+0.5·60/7=240/7, approximately 34.2857. The low-demand invoice is 180/7, approximately 25.7143, and the high-demand invoice is exactly sixty. The hard budget requirement changes the reservation without pretending that high demand has become more probable.
Risk measures such as expected shortfall can instead emphasise expensive outcomes while allowing a trade-off. For the original q=4 policy, the worst quarter of the distribution consists entirely of high-demand cost ninety-two. A tail-average measure focused on that quarter therefore sees ninety-two, not the overall mean thirty-two.
Do not convert a risk preference into a probability adjustment without explanation. Assigning a higher probability to high demand changes the factual model, whereas penalising tail cost changes the decision criterion. Keeping evidence and preference separate allows another reader to understand and challenge either assumption independently.
CHAPTER 9 OF 14 · Model limits, risk and scenarios
9. Generate scenarios that preserve the decision problem
A scenario is a joint description of uncertain quantities, not merely a demand number. It may include demand, processing time, supplier availability and emergency price. Correlations matter because the expensive or infeasible outcome can depend on their coincidence.
Suppose demand twelve occurs mainly when emergency prices are high. Sampling demand and emergency price independently would invent combinations and alter the expected cost of recourse. The marginal distributions could both look accurate while the recommended reservation is wrong.
Reduction must preserve consequences
Reducing a large scenario set can make optimisation practical. But removing a rare high-demand scenario may remove a binding service requirement or the main reason to reserve backup capacity. Similarity in raw inputs is not always similarity in decision consequences.
The specialist guide to scenario reduction and tail preservation examines that issue in capacity planning. The relevant question is whether the reduced set preserves policy performance, not simply whether its average demand resembles the original average.
In the two-demand example, replacing both scenarios by their mean six preserves expected demand exactly but changes the decision from four to six. This miniature example shows why matching a single summary statistic cannot guarantee that an optimisation response is preserved.
Keep a record of where scenario probabilities come from: empirical frequencies, an estimated model, expert judgement or a constructed stress distribution. A deliberately stressful distribution can be useful, but it should not be presented as an empirical forecast. If probabilities are uncertain, test a range of credible alternatives and consider whether a distributionally robust formulation better matches the available evidence.
CHAPTER 10 OF 14 · Sample, validate and diagnose
10. Understand sample average approximation and its optimism
When the probability model has too many possible outcomes for direct integration, sample average approximation replaces an expectation with an average over sampled scenarios. The resulting optimisation problem chooses a decision using that sample. A fresh sample is then needed to assess the chosen decision’s performance without reusing the selection evidence.
For a fixed decision, a sample average can estimate expected cost under suitable sampling assumptions. But the minimum over many decisions can exploit favourable sampling noise. The best observed training objective is therefore not automatically an unbiased assessment of the selected policy.
Separate selection from evaluation
In the capacity example, let p̂ be the sampled frequency of high demand. For reservations between four and twelve, the estimated marginal cost is 3−10p̂. If p̂<0.3, the sample model selects four; if p̂>0.3, it selects twelve; at exactly 0.3 all reservations in that interval tie.
The true high-demand probability is 0.25 in our constructed model. Small samples may nevertheless exceed 0.3. A resulting reservation of twelve is not evidence that the algebra failed. It reflects uncertainty in the sampled frequency and the decision’s threshold behaviour.
Research on sample average approximation with integer recourse studies sampling-based optimisation and convergence under stated assumptions. Such results do not justify a universal sample size for every application. Distribution tails, decision complexity and the desired accuracy matter.
A useful experiment repeats the sample-and-solve process with independent training samples, then evaluates the resulting policies on a common independent test set. Record both objective variation and decision variation. Several noticeably different decisions may have almost identical expected costs, while a stable decision can still rely on a misspecified probability model.
CHAPTER 11 OF 14 · Sample, validate and diagnose
11. Validate a fixed policy with paired comparisons
Validation should begin with an implementation test: can the policy be executed using only the permitted information and actions? Then evaluate its costs, service failures and resource use on scenarios not used to select it. Separate model errors, numerical optimisation errors and sampling errors because each needs a different repair.
For a fixed q=4 policy, cost is twelve in the low scenario and ninety-two in the high scenario. Its variance is 0.75(12−32)²+0.25(92−32)²=1200. With one hundred independent draws from the stated distribution, the standard error of the sample mean is √(1200/100), approximately 3.464.
Use the same scenarios to compare policies
Compare q=6 with q=4 on identical demand draws. The cost difference is six in the low scenario and minus fourteen in the high scenario. Its mean is one and variance is 0.75(6−1)²+0.25(−14−1)²=75. With one hundred independent paired observations, the standard error of the average difference is √0.75, approximately 0.866.
Pairing removes much of the variation shared by both policies. It does not make the small mean difference certain: a difference of one is still modest relative to that standard error. Report uncertainty rather than announcing a decisive winner from a noisy test.
These standard-error calculations assume independent draws from the specified model and a policy fixed before evaluation. Serial dependence, importance sampling or repeated policy selection changes the analysis. Using a normal approximation for an interval also requires its own adequacy check, especially for heavy-tailed costs or small samples.
Stress tests should include plausible scenarios outside the fitted model, but label them separately from probability-weighted validation. Their role is to reveal failure modes and contingency needs. A stress scenario is not automatically assigned a reliable probability merely because it is vivid or important.
CHAPTER 12 OF 14 · Sample, validate and diagnose
12. Interpret solver output and repair common failures
For a finite linear model, the deterministic equivalent can be solved as a linear program. Large scenario sets may benefit from decomposition, separating the shared first-stage decision from scenario recourse problems. Integer actions, nonlinear physics and multiple stages can make the computation substantially harder.
A solver’s optimality certificate concerns the model supplied to it. It does not establish that the demand distribution is correct, that the emergency supplier will deliver or that the information timeline matches reality. Keep computational validity and modelling validity as separate questions.
Read bounds in the correct direction
For minimisation, a feasible implementable policy gives an upper bound on the model’s best achievable cost. A relaxation or valid lower-bounding method gives a lower bound. If a reported lower bound exceeds a feasible policy’s evaluated cost under the identical model, something is inconsistent: perhaps different scenarios, cost accounting or solver tolerances were used.
A wait-and-see solution is a useful lower benchmark because it grants additional information. It must not be issued as the actual policy. In our example, its eighteen lies below the implementable optimum thirty-two precisely because reservation can vary with demand before demand would normally be known.
Common failures have direct repairs. If probability weights sum to more than one, rebuild the scenario accounting. If a policy uses future information, merge shared-history decisions. If emergency actions violate capacity, add the missing bounds. If training cost looks excellent but independent performance deteriorates, investigate sample selection, overfitting and distribution change.
Finally, inspect units and numerical scales. A penalty expressed per thousand units alongside costs expressed per single unit can dominate or disappear for purely bookkeeping reasons. A tiny feasibility tolerance in numerical units may still represent a large physical shortfall after conversion. Translate the final solution back into operational quantities before accepting it.
Build a tiny decomposition certificate
The expected emergency cost in the original example is R(q)=7.5max(0,4−q)+2.5max(0,12−q). It is the maximum of three affine expressions: 60−10q, 30−2.5q and zero. Introduce a variable θ and require θ to be at least each of those expressions. Minimising 3q+θ over q≥0 exactly reproduces the original problem.
This formulation illustrates the idea behind cutting-plane approximations to a recourse value function. A valid affine lower bound is a supporting cut. With all three cuts present here, the recourse representation is exact. In a larger problem, cuts may be generated progressively from scenario subproblems instead of being written down in advance.
At q=4 the first two cuts both give twenty and the third gives zero. Thus θ=20 and total cost is thirty-two. The intersection of the first two lines occurs at q=4. To the left, total cost decreases; to the right, the second cut makes it increase until the zero cut becomes active at twelve. This provides another transparent certificate of the optimum.
If only θ≥0 were retained, the relaxed model would choose q=0 and report zero. That is a lower bound, not an executable zero-cost solution: evaluating actual recourse at q=0 produces expected cost sixty. The distinction between the relaxed objective and evaluated policy cost is the basis of a useful stopping check. Stop only when the relevant bound gap is acceptably small, not merely when successive reservations happen to match.
These particular cuts rely on continuous linear recourse. Integer emergency actions can produce a different value-function geometry, and a cut formula valid for a continuous relaxation need not describe the integer problem exactly. Identify which mathematical structure supports the method before interpreting its convergence or optimality claims.
CHAPTER 13 OF 14 · Transfer and continue
13. Transfer the model when probabilities and decisions change
Replace the high-demand probability 0.25 by p, keeping demands four and twelve, reservation price three and emergency price ten. Between four and twelve, expected cost becomes 3q+10p(12−q), whose slope is 3−10p. The threshold is p=0.3.
For p<0.3, reserve four. For p>0.3, reserve twelve. At p=0.3, every reservation between four and twelve has expected cost thirty-six. This discontinuous change in the selected endpoint does not imply a discontinuity in the optimal cost; the competing cost lines meet at the threshold.
Distinguish an unstable decision from a costly mistake
At p=0.29, reserving four costs 12+80·0.29=35.2, while reserving twelve costs thirty-six. At p=0.31, reserving four costs 36.8, while reserving twelve still costs thirty-six. The chosen reservation jumps by eight units, but choosing the other endpoint near the threshold costs only 0.8 in expectation.
Now require reservations in batches of five. At the original p=0.25, candidate reservations are zero, five, ten and fifteen, with expected costs sixty, 32.5, thirty-five and forty-five respectively. Larger reservations only increase cost, so five is the best permitted choice. Rounding the continuous optimum four downward to zero would be expensive; rounding is itself a decision that must be evaluated.
Combine the batch restriction with the five-unit emergency cap. Reservations below seven are infeasible, making ten the cheapest feasible batch choice, with expected cost thirty-five. The extra cost compared with 33.5 in the continuous capped model is the price of indivisibility under these assumptions.
The transfer exercise is to explain each change without rerunning a memorised procedure. Which constraint excludes a former decision? Which probability changes a marginal saving? Which implementation rule changes the feasible set? Understanding those mechanisms is more durable than remembering that one particular example chose four.
CHAPTER 14 OF 14 · Transfer and continue
14. Answer practical questions and plan further learning
Does stochastic programming require a perfect probability model?
No, but its conclusions are conditional on the probability model used. Estimated or subjective probabilities can still support disciplined decisions when their uncertainty is stated and tested. A precise optimal solution should not be used to conceal weak evidence about scenario frequencies or dependence.
Is a larger scenario tree always better?
Not necessarily. More scenarios can represent uncertainty more richly while increasing computational cost and making validation harder. A large tree with incorrect dependencies or information timing can be less useful than a small well-justified one. Evaluate improvement through policy performance and fidelity to the decision, not leaf count alone.
Why can the stochastic solution have a worse realised outcome?
It is selected before the outcome is known and may minimise an average rather than every individual cost. In the capacity example, reserving four is cheaper on average but more expensive under high demand than reserving twelve. An adverse realised outcome does not alone refute the original expected-cost calculation, though repeated discrepancies may challenge its assumptions.
What should a beginner learn first?
Practise probability-weighted averages, conditional probability, linear constraints and piecewise-linear functions. Reproduce the reservation example by hand, then create a small spreadsheet with one row per scenario. Keep the first-stage reservation fixed across rows and allow only emergency purchases to change. This simple layout makes accidental clairvoyance easier to spot.
For a second project, reproduce the imperfect signal using joint probabilities. Verify that they sum to one, recover the prior high-demand probability by summing the signal branches, and compare the signal’s value with the perfect-information bound. For a third project, add the emergency cap and batch sizes, then explain every change in the feasible policy.
Use the linked deterministic-equivalent documentation for model structure, the VSS documentation for policy comparison, and the sampling research for statistical questions. Continue into the specialist scenario-tree and scenario-reduction guides when the basic decision ledger is secure. A dependable conclusion states what to do now, how to respond later, which evidence supports the probabilities, and where the policy’s protection ends.