祝融说。

Is an Average Gain of One Unit Worth Pursuing?

At noon on Saturday, Gu Ning copied several results from the consequence table onto a calculation page. Tang Ke asked whether D17 could be decided simply by multiplying each consequence by its probability and checking whether the sum was positive or negative.

This method captures a core relation in expected value, but it does not complete action judgment. First, D17 has no Version 2 and therefore lacks concrete consequences and corresponding probabilities. Second, a weighted average answers for an average within a given model. It neither says that any single realized result will equal that average nor establishes that the failure is bearable.

This chapter uses several explicit, independent demonstrations to explain expected value. Every number belongs to a thought experiment and does not enter the Chengwan storyline. At this point on Saturday, the client has not replied, and the station has not reserved a bay, bought materials, or sent a clarification.

Expected value is not a simplistic tool to be dismissed. It can bring the degree of an outcome and its probability into one comparison, and can reveal a path omitted by intuition. What this chapter limits is the question it answers: one average must not answer for goals, constraints, and time all at once.

How Discrete Results Produce a Weighted Average

If a numerical result X can take several values, each with a corresponding probability, and the modeled results are mutually exclusive and cover the range in question, expected value can be written as:

$$ E[X]=\sum_i P(X=x_i)x_i $$

This definition and the rule for the expectation of a discrete random variable can be checked against MIT Introduction to Probability, Lecture 5. The formula explains weighting; it does not supply real grounds for the outcome values or probabilities.

If only favorable results are listed while no version, rework, or exit paths are omitted, probabilities may sum to less than one. If two results can occur together but are treated as exclusive paths, simple addition also changes meaning. The composition of the outcome set must be checked before calculation.

X must also have a stated unit. Changes in working capital, occupied labor, and relationship states cannot enter the same X without a conversion. Expected value works on an already numerical object; it does not automatically translate every goal into a common scale.

An Average May Never Actually Occur

Consider an independent action with two results, each at probability one-half: a net gain of three units or a net loss of one. The expected value is one-half times three plus one-half times negative one, or one unit.

After a single implementation, the actual result in this model will be either plus three or minus one. It will not become exactly plus one merely because that is the expectation. Expectation is the weighted center of a probability distribution, not a third promised outcome for one trial.

If conditions remain stable over many repetitions and the corresponding repeat structure holds, the average result may support another interpretation. But one D17 does not automatically own an infinitely repeated game. Whether conditions remain stable, objects are independent, and the actor can repeat all require material.

“Expected to earn one” is therefore easily misunderstood. If it means an expected accounting change, it is usable when the two actual paths are stated. If heard as a claim that this trial will probably add roughly one, an average has been turned into a single-outcome prediction.

Equal Expected Values Can Have Different Consequence Distributions

Consider two more actions. The light action retains a one-half chance of gaining three and a one-half chance of losing one, for expectation one. The deep action has a one-half chance of gaining eight and a one-half chance of losing six, also for expectation one.

Their weighted averages match, but their failure losses and successful magnitudes differ. A subject able to bear a loss of one is not thereby able to bear six. If only four resources are deployable, the second failure path may also penetrate resources protected for other purposes.

“The long-run averages are equal” cannot make the actions equivalent. If the goal concerned only a linear value and permitted indefinite pooling, some distributional differences might temporarily be ignored. Chengwan also cares about survival, fulfillment, and recovery, so the path structure must remain visible.

This comparison does not declare that the light action is always better. The successful consequence of deep action may serve an important goal, and a subject may have a reliable buffer. The conclusion is only that equal expectations do not complete the ordering; the distribution of consequences remains action information.

Positive Expectation Can Still Cross the Survival Line

Suppose another action has a nine-tenths chance of adding two units and a one-tenth chance of losing nine. Its expectation is positive 0.9. If the actor can bear only four units of loss, the negative-nine path crosses an explicit survival line.

Positive expectation does not remove that path. It gives a large loss a small probability weight in the average, but when the failure actually occurs, the actor bears all nine units rather than only its probability-weighted portion.

The actor may change the consequence by reducing commitment, obtaining reliable protection, using stages, or refusing to act. Which method is available depends on concrete conditions; probability cannot simply be lowered. While a decisive consequence remains, a positive average is not automatic authorization.

Likewise, a negative expectation need not answer every goal. An action may be chosen to fulfill an obligation or protect another nonmonetary goal. The adopter must explain how that goal enters instead of hiding the negative figure.

Separate “Low Probability of Loss” from “Bearable Loss”

Probability answers how strongly the model supports a loss path. Bearability asks whether the subject can preserve protected conditions if that path occurs. Both matter, but they occupy different dimensions.

A one-percent loss can demand serious treatment if it prevents the system from continuing. A fifty-percent small loss that is quickly recoverable may suit a probe. Calling risk large or small by probability alone omits consequence magnitude.

Bearability is not synonymous with subjective courage. It can return to balances, existing commitments, recovery time, and alternative paths. Different subjects may reasonably reach different conclusions under different conditions; willingness to bear more does not alone determine who is more rational.

Later chapters separate failures that cannot be used for experimentation in this case. For now, weighted calculation must preserve the full loss instead of mistaking a probability-weighted amount for what would actually have to be borne.

Expected Value Needs an Action Condition

If different actions change probabilities of external results, the same unconditional p cannot be used for all of them. One can ask separately about outcome probabilities conditional on light clarification and on deep advance preparation.

Formally, the expected result X under action a can be written:

$$ E[X\mid a]=\sum_i P(X=x_i\mid a)x_i $$

This remains a conditional expression within a given model. Post-action probabilities need data, mechanisms, or an explicit subjective judgment. An action that sounds proactive does not automatically raise the chance of success.

If an action does not affect the result but changes the loss, result probabilities can remain equal while consequence values differ. If it changes both information and process, both may change. Gu Ning must state the model in use rather than selecting only the change favorable to a plan.

The Observed Group May Differ from the Acting Group

The four historical B cases in Part 1, three of which had A, describe a historical conditional frequency among cases with a certain kind of message. If D17's action compels everyone to reply or prepare early, it changes the message-generating process. Three-quarters cannot directly become an action-conditional success probability.

Likewise, past tasks that received deep preparation may already have had clearer conditions. Their better observed results do not prove that deep preparation has the same effect on every current object. Selection, shared conditions, and post-action relations must be distinguished.

The expectation formula does not expose this transfer error. Once given a probability, it produces a result. Correct calculation and grounded action conditions still require separate checks.

The D17 scenario page therefore preserves the identity of each probability: historical description, subjective judgment, mechanism assumption, or actual new-batch record. Where no basis exists, the value may remain unknown instead of borrowing an old frequency merely to complete the arithmetic.

One Accounting Expectation Must Not Swallow Multiple Goals

Suppose one action has higher expected accounting contribution but a substantial chance of delaying existing work, while another has lower contribution but preserves fulfillment and capacity. If the goal table protects existing commitments, ranking solely by accounting E[X] is invalid.

Different goal indicators can receive separate outcome models. A utility model may also map multiple consequences onto a common scale when grounded. But such mapping is an additional judgment, not work performed automatically by the definition of expectation.

Relationship, learning, and capacity require particular care. Whether a reply produced useful information can be judged against a concrete question. “The relationship is worth five units,” without a rule, merely packages preference as a number. Qualitative comparison can be more accurate than an invented common unit.

Actions can be screened through hard constraints and then report expected consequences for each goal separately. The absence of a unique total score does not prevent comparison; it keeps actual tradeoffs visible.

What Assumption Does Expected Utility Add?

Economic analysis of choice under uncertainty often maps outcomes into utility before weighting by probability. MIT 14.01 lecture notes on uncertainty distinguish expected value from expected utility and relate the shape of a utility function to risk preferences.

If the outcome is x and the utility function is u, expected utility in a discrete scenario is the sum of each probability times u(x). A nonlinear u can express that the same monetary change is evaluated differently at different resource positions.

But a utility function is not an objective scale discovered by the formula. The chosen function, included subjects and goals, and fit with current behavior require additional explanation. Applying a square root or logarithm to losses does not prove that the model is reasonable.

This book does not construct a precise utility function for Chengwan. It uses the distinction only to show that monetary expectation can differ from a subject's evaluation of consequences, and that numerical unification requires disclosure of added structure.

Risk Aversion Is Not Permanently Lower Probability

Risk attitude evaluates results and choices; it does not make event probabilities more pessimistic. If the material supports a sixty-percent event probability, an actor concerned about losses should address them in consequences or utility, not change the probability to thirty percent to express reluctance.

Putting preference into probability damages both judgment and action. In later calibration, the record no longer reveals whether thirty percent estimated the event or expressed dislike of the action. The action comparison also loses sight of the actual loss.

Conversely, willingness to bear volatility cannot turn sixty percent into ninety. Courage does not make an event more likely. Probability should remain faithful to information and preference to goals and consequences. They connect without impersonating each other.

This continues the discipline of judgment from Part 1. Adding action does not let action desires flow backward and contaminate probability inputs; desires and bearability receive their own places.

Linear Weighting Does Not Guarantee Long-Term Survival

If an actor repeatedly selects positive-expectation actions with a small chance of exhausting all resources, repetition does not automatically protect the actor. Once decisive failure occurs, the subject may have no next trial, and the path for realizing a supposed long-run average disappears.

This does not mean a low-probability severe consequence must occur or that mathematical expectation is wrong. It means a repeated structure must consider how resource state changes with outcomes and whether action can continue. Simply adding expectations across trials may omit an absorbing boundary.

Conversely, refusing every action with a chance of loss may gradually exhaust resources or miss the conditions needed to continue. Survival is not the absence of fluctuation; it is a path connecting loss, recovery, and the next step.

Chapter 5 identifies which failures cannot be tried under current case conditions. Here the subject's capacity to continue is placed next to expected value so that a positive sign does not answer for it.

Dominance Sometimes Needs No Precise Probability

Suppose two actions achieve the same target result in every listed scenario, but A uses less labor and money and is no harder to exit. If no other material consequence is missing, A is at least no worse than B. B can be excluded without precise probabilities for every path.

Actual use must test whether “every result” is complete. B may provide faster response, added learning, or another recovery path. If so, the consequences differ. One cannot delete another plan's advantage in order to declare dominance.

This comparison reminds Gu Ning that every choice need not begin with a complex expectation model. Constraints and explicit relations can first remove meaningless plans, reducing labor spent on unsupported probabilities.

D17's deep preparation presently commits more without a known Version 2 to produce corresponding value. Compared with clarification that occupies no bay, it has not shown an added goal benefit and therefore lacks a present basis for adoption. This is a statement about the material now, not a permanent claim that deep preparation is useless.

A Three-Scenario Demonstration Checks Omissions

Suppose an action has three mutually exclusive results: probability one-fifth of gaining six units, probability one-half of gaining one, and probability three-tenths of losing four. Probabilities sum to one, and expectation is one-half.

The calculation is one-fifth times six, plus one-half times one, plus three-tenths times negative four, yielding 0.5. Omitting the third item produces an apparent expectation of 1.7. If the first two probabilities still sum to seventy percent but are treated as the whole set, the number's meaning has changed.

Nor can “perhaps there is a fourth result” simply receive the remaining thirty percent, because the table stipulated that the three items cover the space. If that coverage lacks grounds, mark the model incomplete instead of assigning probabilities merely to reach one.

The point is not positive 0.5. It is that the outcome set, probability sum, and direction of values are reviewable. If D17 later uses a scenario table, it must leave a way to identify branches for which material has not been obtained.

Sensitivity Analysis Does Not Find the Actual Distribution

Suppose the probability of the high result changes from one-fifth to one-tenth, the low result correspondingly rises to two-fifths, and the middle remains one-half. Expectation changes from 0.5 to negative 0.5. This shows sensitivity to a probability input.

It does not prove that the actual probability lies between the two sets or create a confidence interval. Sensitivity asks how output changes when input changes. A credible range for that input still needs its own grounds.

Outcome values can also change. If failure changes from negative four to negative two, expectation rises. Identifying influential inputs can guide further investigation or a smaller action, but the most optimistic combination cannot be selected as the final judgment.

D17 has not obtained even primary scenario values, so it cannot yet draw an apparently precise sensitivity range. Gu Ning leaves four fields in the template: probability, outcome value, source, and change check.

Higher Expectation May Still Be Infeasible Now

A plan first passes survival and authority constraints, then enters comparison of expected performance. If it needs five initial units while only four are deployable, a higher expectation cannot directly authorize implementation.

The actor may change feasibility by reducing scope, obtaining staged payment, or waiting for an existing expense to end. Only when conditions actually change does the plan reenter comparison. Money that might be obtained cannot be recorded as already usable.

This order prevents an average benefit from overriding time. Future inflow, present balance, and obligations that cannot be delayed remain separate; expected totals do not make them simultaneous.

Feasible also does not mean preferable. Passing the floor only admits a candidate; goals, other consequences, and alternative actions still matter. Hard constraints and optimization have different responsibilities. “Not beyond the boundary” does not mean “best.”

How a Single Realization Returns to the Model

After action produces a result, it can be matched to the original scenario page. If it follows a listed path, record the actual consequence and review probability judgment and execution. If an important unlisted result appears, coverage was inadequate and the model must reopen.

One favorable realization does not prove the expectation inputs accurate, and one adverse realization does not automatically refute a model that gave the outcome nonzero probability. Cross-event records may support later calibration, but the consequence of the action must still be preserved in full.

If the outcome value differs from its estimate—for example, rework lasts longer—the consequence model needs revision. Probability must not be changed merely to preserve the old expectation, and the original version must not be rewritten after the outcome to make the forecast fit.

The expectation page is valuable not only for summing before choice. Afterward it shows where the difference occurred: probability, range of outcomes, numerical value, execution, or external conditions.

Expectations Add, but Joint Consequences Still Require Inspection

If numerical results X and Y have defined expectations, the expectation of their sum equals the sum of their expectations; independence is not required. Gu Ning can use this linearity to reconcile accounting components without pretending that every task is independent.

But addition of expectations does not show whether losses occur together. Suppose two tasks each have a one-tenth chance of losing four units, for an expected loss of 0.4 each. With perfectly aligned failures, one realization may lose eight; if failures cannot coincide, the maximum is four. Total expected loss is 0.8 in both cases, but survival exposure differs.

These numbers are another independent demonstration, not D17 facts. They show that joint relations matter to bearability even when they do not matter to calculating total expectation. Part II addresses correlated exposure. Here we prevent “expectations can be added” from becoming “consequences are independently diversified.”

Nor does an unknown joint relation make individual expectations meaningless. Average accounting quantities can be reported under the current model while common failure, peak occupation, and recovery conditions remain in separate fields. The mathematical property answers its question; the action page restores omitted structure.

Average Performance Must Say What Is Repeated

People sometimes console a single fluctuation by saying expectation will work out over many trials. But the repeated object, stability of conditions, and capacity to survive intermediate paths do not follow from the word “many.” Objects differ across tasks, and probability and consequence distributions may change with the environment.

Even in an explicit identically distributed model, a long-run average does not prevent concentrated losses in the first few trials. If the subject crosses its survival line on the third, it cannot implement later favorable trials. Eligibility to repeat is itself a condition of action.

D17 is one action that has not yet formed. Gu Ning will neither distribute an unbearable current consequence across imagined future commissions nor reject every expectation comparison because a single trial cannot prove a long-run average. She first asks whether the current plan leaves a next trial, then considers performance across batches.

Expected Value as an Action Representation Through Theory Reduction

From RC's view of theory reduction, expected value projects many possible consequences into one weighted quantity. It is an effective representation with which a limited actor can organize complex results. Its limits do not cancel its value; they show why it should serve as a cognitive relay into the next comparison rather than become a closed account of all reality.

This chapter therefore assigns distinct responsibilities to different representations: expected value handles weighted outcomes, the survival line preserves a subject's nonexchangeable conditions of continuity, and the feedback page later checks how actual consequences change renewed expectation. Together they serve RC's prospective cycle.

The D17 Scenario-Weighting Page Temporarily Remains Blank

Gu Ning listed these fields on the D17 page: concrete action; mutually exclusive results; each probability and its identity; accounting and nonaccounting consequences of each result; expected quantity; survival check; sensitive inputs; and reopening conditions.

Action depth and resource boundaries are known. The new version's outcome set, action-conditional probabilities, price, and recovery cost are not. She does not copy the 0.5 from Part 1 into “a suitable version forms,” nor treat the four deployable units as expected gains.

One clarification has a smaller known commitment than deep preparation, but whether to perform it waits for the action card in Chapter 6. At present, deep preparation lacks a corresponding object, so the possibility of an earlier start cannot be inferred as a known gain.

Blank values do not mean no progress. Required material now corresponds to calculation fields. When a later reply arrives, total receipts, contribution, probability, and bearability need not be mixed again.

Saving the Calculation Boundary at 12:40 on Saturday

At 12:40 on Saturday, Gu Ning saved the scenario-weighting page. Three independent demonstrations showed that expectation may not be a realized result, equal expectations may hide different distributions, and positive expectation may cross a survival line. None enters Chengwan's resource ledger.

The external course notes are used only to check the formal relations of discrete expectation and expected utility. They supply no probability, utility function, or operating conclusion for R17. The actual D17 page still contains no new event probability, income, expenditure, or adopted action.

This chapter reaches the following judgment: expected value answers the weighted average under given outcomes, probabilities, and numerical scale. It can compare scenarios, but it does not alone answer what will happen once, what the subject can bear, or how multiple goals should be ranked.

Before any positive or negative average, the next chapter establishes survival conditions. The station must identify which losses would interrupt existing work, basic working capital, or recovery paths, and which failures, though adverse, can still enter trial and error. Only with that boundary in place can expectation avoid reducing an unbearable outcome to a small weight.