祝融说。

This Sentence Already Sounds Quite Positive

Tang Ke still felt "checklist received" was good news. At least the client responded, and was willing to continue verifying the reception time. Gu Ning acknowledged that it carries more information than having no corresponding material at all, yet asked: if the original window ended without complete confirmation, would this message also have appeared.

The last chapter separated the two conditioning directions; this chapter goes further and compares the same message's appearance in the two outcome classes. If one looks only at its being common among confirmers, one still does not know whether it is equally common among the unconfirmed. A message's positive tone does not mean it significantly distinguishes outcomes.

R04 has already provided a limited reminder: it satisfies the earlier window's checklist-received message condition yet did not satisfy the original window's complete confirmation. This object has not made B entirely useless, nor has it given Gu Ning a signal that is never wrong. It demands comparing the two groups rather than judging everything from one sentence.

This chapter continues the fictional Chengwan repair station thought experiment; the historical objects remain R01 through R08, and main-line R17 remains unconfirmed. We add no new real-world industry regularities; we show how a piece of material already obtained supports or limits the current judgment.

First Fix the Two Outcomes to Compare

A still denotes receiving complete confirmation of the saved first version within the original window; not-A denotes that this event has not held. Not-A is not an explicit rejection, nor is it never cooperating later. The old account—two items without complete confirmation, one proposing modifications—is retained; specific reasons are not reassigned by number.

B still denotes receiving, after the corresponding Thursday 10:00 and up to 11:20, an explicit checklist-received message on the agreed channel. This condition does not include item-by-item verification of scope, acceptance of payment conditions, or the reception period being available. R17's complete original utterance has one more pending-verification clause than B; one cannot assume its full content already has a correspondence in the historical B.

This chapter compares the coarse B first because the previous chapter already has the four-cell record for the same objects. If one switched to the complete original sentence, a particular tone, or reply length, the historical material would need separate checking; one cannot change the condition's name while keeping the count of four items satisfying B.

Only when both the outcome and the message type are fixed can one ask: if the item belongs to A, how easily does B appear; if it belongs to not-A, how easily does B appear. The "if" here is a given scope within the model, not Gu Ning's already knowing R17's future outcome.

The Same Message's Proportions in the Two Groups

The historical A group has five items: R01, R02, and R03 have B, R05 and R06 do not, so B's appearance frequency is three-fifths. The not-A group has three items: R04 has B, R07 and R08 do not, so the frequency is one-third.

One cannot directly compare the three B-holders against the one B-holder and claim the message supports confirmation threefold. The two outcome groups differ in size to begin with; the raw counts contain both the within-group message frequency and the outcome groups' shares. To separate these two layers, first divide by each group's own size.

Three-fifths and one-third come from the same eight items' descriptive empirical distribution. They are not two independent studies, and using two denominators did not increase the sample. The total structure of the original five confirmations remains; the current discussion concerns only the message's relative appearance in the two outcome classes.

How Likelihood Places an Observation into the Model

Taking B as an already-observed message and examining the probability of its appearance under the given conditions of A and of not-A respectively—that is the pair of likelihoods compared here. What matters is whether the same observation is easily obtained under different outcome hypotheses, not first translating the observation directly into an outcome probability.

When the probabilities under both given conditions are definable and the message probability in the denominators is greater than zero, the likelihood ratio of this message's relative support for A over not-A can be written:

$$ L_B=\frac{P(B\mid A)}{P(B\mid \neg A)} $$

Here not-A is the complement event of A; both sides of the ratio must come from the same comparison setting. For describing evidential role with a ratio of likelihoods, see MIT course 18.05, Lecture 12, which connects it to the odds before and after updating. This chapter explains the ratio's meaning; full updating is left to the next chapter.

If the original eight items are given equal weight to describe history, three-fifths divided by one-third yields nine-fifths, that is 1.8. This number is greater than one, indicating that within the two groups corresponding to this historical table, B appears proportionally more often in the A group than in the not-A group.

1.8 is not a confirmation probability of 180 percent, nor is it a factor by which the original probability should be directly multiplied. It compares two message probabilities; to obtain the target probability, one must still connect the original shares of the two outcome classes and convert accordingly. Chapter 8's hundred-item comparison has already shown that identical within-group message proportions can correspond to very different reverse outcomes.

Relative Support Is Not Exclusion of the Other Outcome

In this table, the not-A group still shows B at one-third. Observing B has not excluded not-A, so 1.8 cannot be described as having proven R17 will confirm. Relative support describes a comparison between two given conditions; it is not a guarantee about any individual's future.

It is also not the message being "eighty percent reliable." Reliability may mean whether the original utterance is accurate, whether the channel corresponds, whether the record is complete, or long-run prediction quality; none of these is the quantity 1.8 directly measures. One must state which word expresses which layer, avoiding inflating a single ratio into all-around credit.

If the direction of comparison is reversed to not-A over A, the same B's likelihood ratio is the reciprocal of 1.8, namely five-ninths. The direction changed; the observation did not. The two formulations correspond to each other, but the account should fix which class sits in the numerator, so users do not read any larger number as good news.

Gu Ning can say that this limited historical table makes B relatively favor A, while retaining the fact that B also appears in the not-A group. A support not strong enough to exclude the other outcome does not thereby become zero support; not guaranteeing the future is not the same as every judgment being equally appropriate.

A Very Common Message May Have No Discriminating Power

Make another explicitly hypothetical comparison: some generic acknowledgment appears with probability nine-tenths in the A group and also nine-tenths in the not-A group. It is very common, and obtaining it shows the channel returned this kind of acknowledgment—but its likelihood ratio between the two outcome classes is one.

A ratio of one means that under the two classes given by this model, this observation is equally easy to obtain and does not change their relative support. It can reduce the factual blank of "no acknowledgment was received," yet need not reduce uncertainty about the target outcome; an increase in information and this target discrimination cannot be equated directly.

This does not say the generic acknowledgment is worthless in all tasks. If the target changed to judging whether the channel functions, or if the material showed the two groups' acknowledgment probabilities actually differ, a different question would need to be posed. Lacking discriminating power in the current use cannot become the message being permanently meaningless in every use.

Tang Ke can thus keep two statements apart: we did indeed obtain a reply; and how much relative support this reply provides for whether the original window ends in complete confirmation. The first sentence's holding does not automatically fill the second with an optimistic number.

A Rare Clue Is Not Automatically Stronger

Make another hypothetical comparison: some fine-grained clue appears with probability one-fifth under A and one-fiftieth under not-A; the likelihood ratio is ten. It is not common under A, yet it distinguishes the two given outcomes better than the acknowledgment that was common in both.

What determines the ratio is not how rare the clue is overall, but its relative appearance under the two conditions. If the clue is at one percent under both A and not-A, the ratio is still one. Rarity can impress, but it alone supplies no ground for a tilt toward either class.

These comparisons are not new signals Chengwan has obtained, nor do they say B should be updated by ten. They separate commonness from discrimination. Gu Ning cannot, to raise relative support, fabricate for an ordinary message a probability of near-zero appearance in the not-A group.

Likewise, a ratio above one hundred is not one-hundredfold certainty. It may attach to a very small initial outcome share, or rely on an unreliable denominator assumption. Magnitude has meaning, but the meaning is always tied to the model and its inputs; one cannot display only the largest number.

Absence of B Requires a New Comparison

If B was not received within the specified window, the observation becomes not-B. Of the original A group's five items, two lack B, a proportion of two-fifths; of the not-A group's three items, two lack B, a proportion of two-thirds. Not-B's ratio of relative support for A over not-A is three-fifths, that is 0.6.

This is not the reciprocal of B's likelihood ratio of 1.8. The reciprocal still compares the same B while swapping the outcome direction; here the observation itself changed—from having the message to lacking that message—so the complements within each group must be taken separately and then divided.

That 0.6 is below one indicates that in this table not-B tilts relatively toward not-A, but it does not exclude A. R05 and R06 are historical objects that confirmed without early-window B. Lacking B has not automatically become "no contact at all," nor an explicit rejection.

R17 has already satisfied B; this section does not re-record it as not-B. The comparison without the message is to show that the absence of a piece of evidence needs its own model conditions; one cannot save the checking by casually inverting some ready-made ratio.

The Interpretation of Absence Depends on the Method of Observation

If the agreed channel's records are missing, Gu Ning knows only that she has not seen B; she cannot conclude from this that B was not received within the specified window. To judge not-B requires material sufficient to adjudicate the message status; this connects with Chapter 5's distinction of the unknown—missing is not a negative observation.

Make a comparison of observation methods: one compilation saves only explicit confirmation messages and not interim acknowledgments, so failing to find B in its archive does not show the original communication had no B. Another saves the full window's communications, and only it could support adjudicating that the specified window lacked such a message.

If objects compiled by the two methods are mixed together, presence versus absence of records may first reflect the preservation method. The ratio comparison then no longer concerns purely the original message's appearance but includes whether the message could be observed. Gu Ning must re-check sources and state the collection limits; formal correctness must not mask the observation mechanism.

This book's eight items carry the explicit setting that B is adjudicable for all, which is why this chapter can compute not-B. R13 through R16 still lack raw material and cannot be handled with the same 0.6; R09 through R12 differ in stage or window and are not borrowed to swell the two groups.

Why Zero Occurrences Easily Create Extreme Support

Suppose in another, very small historical table the not-A group has no B at all; the simple within-group frequency is zero. Treating this as an exact zero probability in the model, the usual finite-ratio formulation of B's likelihood ratio runs into a zero denominator, and observing B could even exclude not-A within that hypothetical model.

But not having seen something in a small amount of history is different from it being impossible for any future object of the same kind. The former is a finite record; the latter is a stronger model assumption. One cannot let a counting zero automatically establish the latter. Mathematics' handling of extreme inputs does not verify where the inputs came from.

Gu Ning can report the zero occurrences and the group size, noting that a direct ratio has not obtained a finite value. If a nonzero-probability model must be built, the additional estimation method and its grounds should be given clearly; one cannot covertly add one occurrence to soften the number, nor covertly delete the unknown to enlarge the ratio.

The original not-A group in this chapter has one B, so no zero denominator arises. Discussing zero is leaving a boundary for later applications, not claiming that the current 1.8 has undergone some smoothing, interval, or error correction.

A Coarse Message Cannot Represent the Complete Reply

Returning to R17: the original message also says "I still need to verify the reception time." B records only "checklist received" and does not include the pending verification of the reception time. If Gu Ning wants to compare the complete message's role, she must identify how the pending-verification condition appears in the corresponding history; she cannot use B's ratio to compute for the whole sentence.

The pending verification may be merely a remaining step in normal progress, or it may signal that reception within the window will be constrained. Both interpretations stand as candidates; the current corresponding material is insufficient to adjudicate their relative appearance. One cannot, because the tone flows smoothly, keep only the first, nor, because a gap remains, fix it directly as negative.

Thus, assigning B a historical relative ratio of 1.8 in this chapter is not assigning the current complete original utterance a model weight of 1.8. Representational compression lets computation begin, but may omit significant parts; Chapter 6's boundary statements must remain in force here.

Ye Cheng may keep searching for corresponding fields, but until new material is obtained, Gu Ning preserves this gap first. Preserving a limited evidential role is not refusing future updating; it is avoiding passing off a computable coarse signal as a fully understood message.

Information Already Used Cannot Be Treated as New Again

If a version of the judgment has already used the historical group satisfying B as reference, B has entered that version's conditions. Multiplying by the same B's likelihood ratio again would reuse the same observation. Gu Ning should first return to what material the old judgment actually conditioned on, then state what this round adds.

Conversely, the ungrouped eight-item starting point and a B-condition update can connect within one explicit model; but one cannot take the within-group three-quarters as the starting point while pretending B appears for the first time. This does not involve multiple people relaying; it is the same message being used twice within the records.

If C is obtained later, the comparison should state how C appears in A and not-A given that B is already conditioned; whether one may use the two C ratios without conditioning on B requires separately checking the dependence. Chapter 11 will develop sources and common conditions; here the interface for old information is preserved first.

The evidential-role account therefore cannot merely write "1.8." It should also state which observation it targets, which outcomes it compares, and what old information it connects with. Detached from these objects, a ratio is easily accumulated into conviction through relay after relay.

Evidential Role Also Has Data Quality

A ratio compresses two inputs into one number, which may lead readers to attend only to relative magnitude. Yet if the two inputs come from different times, different object scopes, or one of them was guessed from an impression, the larger ratio may still lack adoptable grounds.

Gu Ning can separately preserve provenance and estimation status: direct historical counts, separately posited current model assumptions, and candidate explanations awaiting material. All may enter the discussion, but none may uniformly be called a verified likelihood. Computational refinement does not confer the status of a record on a guess.

1.8 also compresses the original count limitations of five items and three items. The not-A group has only three items, and its proportion may be very sensitive to a single adjudication; this chapter neither derives an error interval directly from that sensitivity nor claims eight items suffice to prove a stable regularity. Later material may still shrink the relative relation or even reverse it.

When materials conflict, first check content, time, and source rather than picking whichever produces the largest ratio. Chapter 12 will handle disputes concretely; here it is acknowledged first that evidential strength and input quality must be delivered together, and that a single number cannot replace the full account of support.

The Two Likelihoods Need Not Sum to One

Tang Ke noticed that three-fifths plus one-third is not one, and asked whether another outcome was missed. Gu Ning points to the two different denominators: the former counts B among the five A items, the latter counts B among the three not-A items. They are not two mutually exclusive outcomes partitioned within one population, so there is no requirement that they sum to one.

What must sum to one is the probability of B and not-B within the same given outcome group. The A group's three-fifths and two-fifths sum to one; the not-A group's one-third and two-thirds also sum to one. Only with the denominator fixed can one check whether all message statuses of that group are covered.

If one directly adds the two likelihoods and then divides three-fifths by this sum, one gets nine-fourteenths. This computation can serve as a step in another explicit hypothetical model with equal class shares, but it is not the confirmation-among-B-holders proportion of the original eight-item table. The original table's class shares are five to three; one cannot silently convert to one-to-one in normalization.

So the two message probabilities not summing to one does not indicate an omission in the conditional table; forcibly normalizing them under equal shares does not automatically produce the correct update either. The next chapter will develop the steps that genuinely connect; here the reader is first made to know that likelihoods and outcome probabilities live in different given scopes.

The Composition Inside Not-A May Change

Not-A is the original window event not holding, and it contains different origins. Earlier text records two items without complete confirmation and one proposing modifications; these origins may correspond differently to early messages, but this book has not supplemented a cross-table by number and reason, so their separate B probabilities cannot be computed.

As a mechanism comparison: if in another object scope non-holding came mainly from needing to change the scope, while in yet another it came mainly from being unable to receive within the window, their early checklist-received messages might show different appearance patterns. These are candidate explanations, not statistical differences Chengwan has obtained.

Even if A's target wording is unchanged, changes in the composition inside not-A can affect the overall P(B given not-A). Therefore the original table's one-third is not a constant parameter determined solely by the name not-A. Transferring to a new scope requires attending to whether the two groups' processes still correspond sufficiently.

Gu Ning need not find a complete cause for each non-holding before being permitted to report the historical ratio. But she should retain this limitation, preventing recipients from treating the coarse two-class table as if all non-holding mechanisms had been identified. The model may first use limited grouping, and must equally allow reopening when important differences are later discovered.

Message Coding Must Also Accept Review

The evidential computation depends on how B is recognized. If a compiler counts "Received, will look later" as B while excluding "The checklist has already been received," judging by surface wording alone may create inconsistency; if "Verified, all agreed" is also compressed into B, the fact that it contains stronger content must be separately noted.

Gu Ning can first preserve the original wording, then state whether it satisfies the limited checklist-received condition and why. Ambiguous messages are left unadjudicated for now; they are not forced into a verdict to fill the four cells. If two compilers disagree, what is re-checked is content and definition, not which classification makes the likelihood ratio larger.

This review does not require everyone to use an identical template. The condition concerns recognizable meaning, and language may vary; but the shared classification must state which contents suffice and which cannot yet correspond. The original table already sets all eight items adjudicable in this chapter; coding disputes serve only as application comparisons and are not recorded as Ye Cheng having misjudged.

Limited evidential accounts thus gain an entrance: the original sentence can be returned to, classifications can be explained, counts can be recomputed. It has not proven everything the client said true, nor eliminated collection error, yet discoverable representational errors need not hide forever behind 1.8.

If the repair station later requires every client to reply "checklist received" first, the process by which the message is obtained changes too. The new process may make B more common in both groups, and the original ratio need not continue to apply. Gu Ning must record the active prompting and process changes, not merely save the reply text; the same sentence under different acquisition methods may correspond to different outcome relations. This does not predict the new process's actual effect; it states only that the evidential model depends on how observation is formed, and one cannot mechanically attach the changed communication to the old frequencies.

Exchanging Good News for a Limited-Role Account

At 12:20 on Thursday, Gu Ning saves this chapter's evidential-role account. B appeared three times among the original five confirmations and once among the original three non-holdings; the equal-weight historical likelihood ratio is 1.8. It describes only this table's relative relations; it is not R17's final probability.

The account also retains not-B's 0.6 and states clearly that this is not 1.8's reciprocal; the reception-time-pending clause in the complete original utterance gained no additional historical weight. The main line added no confirmation or payment; the original message's time point and the 12:00 acquisition of the historical table are preserved, with no backfill to 10:00.

RC treats cognition as the continuing generation of limited representation. Realized here, what promotes the convergence of judgment is not having affirmative language repeatedly endorsed, but placing the same observation under different possible outcomes for comparison and admitting contrary objects. A limited model still requires calibration; the likelihood ratio does not verify the underlying philosophy.

The judgment this chapter yields is that evidential role depends on the observation's relative appearance under two given conditions, not merely on the message being common, rare, or reassuring. Gu Ning knows how B provides limited relative support, and knows that the complete reply, transfer, and old information still require checking.

The next chapter will let the baseline and the evidence genuinely connect within one hypothetical model, showing why updating requires normalization, and also showing which inputs remain unverified even after the arithmetic is right. We already have the object of evidential role; only then do we compute how it changes an explicit probability representation.