A Method That Does Not Yet Have Its Own Control
At 2:20 p.m. on Thursday, Gu Ning wrote Tang Ke's proposal on a new sheet: from now on, ask the client to reply explicitly that the list has been received. The previous sheet already preserved several competing explanations; this sheet asks what must be compared if one wishes to judge whether the method is useful.
Ye Cheng said one could compare people who reply in advance with those who do not. Gu Ning pointed to the historical table: that comparison already exists. What is missing is the difference that arises after the station adds a requirement, compared with not adding it. The differences between two groups of people do not automatically become differences between two processes.
This chapter remains within the fictional Chengwan. The new process has not been implemented, R17 has not obtained a new reply, and complete first-version confirmation is still not received. What appeared at 2:20 p.m. was a comparison problem, and the material organized afterward is also a checklist, not an experimental result.
Moving from explanation to comparison first requires acknowledging that the original question no longer asks only about clues. Clues can help describe objects that have already entered a group; operation must discuss what follows after a process is changed. Both can be of value, but they cannot share a number and thereby dispense with conditions.
Making the Operation Specific Enough to Identify
"Requiring a reply" is not yet an operation stated with sufficient specificity. One might add a single reminder sentence at the end of the list, one might ask for a reply when sending the list for the first time, or one might keep asking after the other party has not yet answered. The three practices change the number of contacts, the timing, and the communicative burden in different ways.
Gu Ning limits the discussion to a candidate version: when sending the first-version list, attach a single explicit request for a "received" reply in the same message; there is no added number of contacts, and the reply is not made a threshold for further communication. This is only a hypothesis set down for the purpose of comparison; the station has not adopted it.
The other side of the comparison also needs to be stated. It can be the process of sending the same first version without attaching that request, and must not be vaguely called "doing nothing." If the other side does not even send the list, the difference between the two sides includes the list itself, and one can no longer discuss only the added sentence in isolation.
Which sentence, at which step, through which channel, and how many times at most, determine the boundary of what we call the change. The boundary is not to press reality into identical machines, but to allow different people picking up the problem to identify whether they are still discussing the same method.
Receiving a Reply and Making a Request Are Different
After the request is attached, some may reply and some may not. Without the attached request, some may still volunteer that they received the list. The operation of adding the request and the message state B of its actual appearance must therefore occupy two separate places.
If one selects only those who eventually replied from the group to which the request was added, then compares them with a no-reply group on the other side, one is once again selecting objects by their post-operation state. Willingness to read the list, more thorough preparation, convenient timing, and the like may still help decide who enters the reply group.
Even if both sides once encountered a similar process, the final comparison is no longer of all original objects on both sides. The non-repliers have been removed from the side with the added process, and the original question of "how the added request affected these objects" has become "how those who replied fared."
One can separately ask about the effect of the reply process itself, but that is another question requiring its own conditions. Because an operation hopes to obtain a reply, the intermediate state between operation and realization must not be written as the same variable. The hoped-for result must not be allowed to retroactively define the process's success.
The Target Must Not Migrate with the Method
Tang Ke's first hope was easy acquisition of messages; Gu Ning reminded her that the original proposal was discussed as potentially promoting complete confirmation. The result of the comparison should therefore still be A: receipt, within the specified window and through the agreed channel, of an explicit complete acceptance preserving the first-version scope, payment conditions, and reception period.
Reply rate may serve as a separate item of observation. It may help one understand the communication process and may show that a reminder did in fact change an expression; but if reply rate rises while complete confirmation does not change, that cannot be recorded as the claimed improvement in confirmation having been achieved.
Likewise, if objects in the new process later accept a revised version while objects in the old process are still adjudicated against the first version, the two outcomes have the same name but different contents. Comparison must first preserve the version; a method must not appear more effective simply by relaxing the target.
The windows must also correspond. If the side under the new process is allowed to wait longer while the other side still stops observation at the original time point, the difference obtained will contain the time gap. A wider opportunity can indeed bring more outcomes, but that does not by itself answer the question of the effect of attaching one request.
The Other Outcome Is Not Lying in a Drawer Waiting to Be Saved
To judge whether a single object has changed because of an operation, one often asks: how did it fare under this process, and how would it have fared under the other process? The second half points to a comparison that did not occur along that path.
In Pearl's Causal Inference in Statistics: An Overview (2009), the counterfactual question is stated with explicit models and conditions: given a certain situation, if some variable is set to another value, what would the outcome be? This kind of formulation requires stating the location of the change and the conditions being held, and cannot be filled by another act of imagination.
In Chengwan's hypothetical comparison, one may write the result of a given request object under the sentence-attached process as Y(1) and its result under the no-sentence process as Y(0). The notation merely leaves room for two questions; it does not mean the station has read two answers.
An object actually traversed one of the two paths in a single process; the material preserved usually comes from that path. The outcome of the other path is not "a message screenshot that went missing and could be found." It requires additional basis of comparison or model assumptions, and cannot be filled by a post hoc feeling of plausibility.
The Same Set of Observations Admits Different Imputations
Leaving aside the Chengwan registration, consider an independent four-object thought experiment. Objects A and B use the sentence-attached process and both satisfy the target; objects C and D use the no-sentence process and neither satisfies the target. Apart from this, no information is given on degree of preparation, assignment process, or other outcomes.
The observation table is short: the sentence-attached side has two items both held, and the no-sentence side has two items both not held. But it lets us see only A's and B's Y(1), and C's and D's Y(0). The Y(0) of A and B, and the Y(1) of C and D, have not been obtained from these four actual processes.
One hypothetical imputation is that A and B would have been held regardless of which process was used, and C and D would not have been held regardless. Under this assumption the two results for each of the four objects are equal, the process difference for each object is zero, and the observation is still perfectly consistent.
Another hypothetical imputation is that all four objects would have been held under the sentence-attached process and not held under the no-sentence process. Here each object's result difference is one, and it too matches the observation. The two imputations cannot both be true, but the original four observations do not by themselves select between them.
An Imputed Table Is Not Four Newly Acquired Facts
The second imputation sounds much like a story of an effective method, and the first like a story of preparation differences. Each story is complete, yet neither adds to the material provided by the four-object experiment. We have only written how the unobserved positions can differ.
Under the first imputation, the average result difference across the four objects is zero; under the second, it is one. These two quantities belong to two imputation models and cannot serve as the upper and lower confidence bounds of an actual effect, let alone be transplanted into the possible range of R17 improvement.
That one side of the original observation is all held and the other is all not held also does not tell us each person's other outcome. Filling the between-group observed difference into the gap between each object's two paths is to use different objects' actual outcomes in place of the same object's unrealized outcome.
This thought experiment is of limited use: it points to what explanation still lacks. It does not claim that no real material can support a causal judgment, nor does it prove that the process is necessarily ineffective. If more support is obtained for assignment, conditions, or process, the judgment can change; a gap is not a permanent gag.
Where Can Similar Objects Substitute
Since one cannot observe both complete paths of an object within a single process, one may look for similar objects for comparison. "Similarity" here should target conditions that would participate in both the operation and the target, not merely the fact that two registrations both appear to involve repair work.
If the station attaches the request only to clients whose materials are ready and whose schedule is easy to arrange, and puts objects with ambiguous materials on the other side, then the difference between the two sides is not just that reminder sentence. The process selection itself is already connected with preparation—one of the competing explanations preserved in the last chapter.
One may first look at conditions before the process began: whether the first version was already clear, whether confirmation was still incomplete, how much time remained before the deadline, and whether the reception period required separate checking. Checking these serves to assess the degree of correspondence in the comparison; it does not claim that all causes have been mastered from a few fields.
The historical comparable eight items were once used for a frequency model and do not thereby become a ready-made control for the effect of a new operation. They were not recorded as comparison objects for two explicit processes; the provenance of B and the actual receipt sequence within them still need reexamination. What suits one predictive arrangement does not automatically suit every explanation.
When There Is No Corresponding Position, Preserve the Gap First
Suppose the added request is used only for objects whose reception period has already been negotiated, while in the old process not a single object of this kind exists. Even if the other registration fields are stated in great detail, the other process lacks an actual comparison position under these conditions.
One can rely on some model to extrapolate from other objects to here, but must acknowledge this as an extension, not as the direct acquisition of a same-kind control. If the model assumes that whether the period has been negotiated does not affect the process difference, the reader should see this assumption rather than only an adjusted number.
As conditions are specified ever more finely, the comparison may shrink to only a very small number of objects. This solves some differences but does not automatically solve sparsity and other unknowns. The record must simultaneously state the reasons for exclusion, the number retained, and the small range to which the final judgment applies.
When there is no corresponding position, the most useful step may be to mark the material to be obtained later. If a gap is covered by a fluent explanation, even when new objects later arrive it will be hard to see which point still needs support. A temporary absence of an answer can still be accompanied by a clear next question.
Conditions That Arise After the Operation Must Be Stated Separately
After the sentence is attached, the client may reread the list because of the reply request, or may feel burdened and postpone processing. These are different continuations of candidate processes and must not be treated during comparison as conditions that were already the same before the process began.
If one compares "the total effect of the reply request" yet keeps only those objects on both sides that later carefully verified, one may remove part of the path one originally wanted to understand. Conversely, if one wants to discuss whether an added request has an additional effect at the same level of verification, one must write a separate question and support.
This does not ask the reader to memorize a set of adjustment formulas. What must be preserved is where a variable appears and why it was included or excluded. A field called "degree of preparation" may sometimes mean completed before sending and sometimes completed after receipt; the same name would obscure different processes.
Including every obtainable field in the calculation does not by itself guarantee that the comparison becomes appropriate. Material is one condition; how it is used is another. Otherwise one may quietly cut off in the calculation the very continuation one wants to ask about, and then announce that the original process has no effect on the strength of a small difference.
Two Successive Occasions Are Not Two Paths of the Same Occasion
Tang Ke thought that if the same client later came back for a second engagement and used the sentence-attached process, could one compare it directly with the first? That way at least the object's name is the same, and there is no need to substitute one person for another.
But the second occasion already carries the experience of the first. Both parties may be more familiar with the scope; the client may already know the reply habit; the object to be repaired and the time conditions may also differ. The object's name being maintained does not mean the first occasion's starting point, task, and other continuations have all been preserved anew.
Before-and-after comparison can provide material, especially when the locations of change and other conditions are clearly recorded. But "the same client" must not be treated as proof that all differences have disappeared. Failure on the first occasion and success on the second also do not by themselves decide which change was effective.
If repeated contact causes later communication to follow the earlier prompt, then what is called "no sentence" on the second occasion may not equal the process of one who has never encountered the prompt. A comparison table must acknowledge the cumulation of process. The past does not automatically exit the present because we want to establish a control.
Can One Person's Process Enter Another's Conditions?
When the repair station discusses a uniform process, it may also change the station's arrangements. The added communicative time for one object may delay message processing for another; if several clients share information with each other, the no-sentence side may learn the reply requirement early.
These are only possibilities requiring checking; they are not new Chengwan facts added by this chapter. They remind us that if each request is written as an isolated box, one must state under what conditions this division is usable.
When influence crosses object boundaries, "this one uses Process 1, that one uses Process 0" may not have exhausted the actual change. The comparison may need to expand to some jointly running process, or preserve a record of adjacent contacts. After expansion, the question, the count, and the applicable objects must all be rewritten.
The unit of observation is determined by the question, not by whichever row of the registration table is most convenient. Smoothing out relations simplifies calculation, but it may also change the explanation. Acknowledging relations does not mean computing all of reality at once; it means stating what continuations the current division has left unhandled.
What Question Does Random Assignment Help Answer?
If in the future there are two explicit processes and the assignment of who enters which can be separated from the client's degree of preparation, that is one possible way to improve comparison. Random assignment is valued precisely because it seeks to prevent operation assignment from always being tied to a certain starting condition.
But the two words "random" cannot replace a full process record. A small number of objects may still differ markedly in practice, and after assignment there may be non-execution, withdrawal, or missing material; if only those who fully completed the process are retained, the entrance to comparison changes once again.
Therefore, when proposing such an arrangement, one must simultaneously preserve the entrance range, the time of assignment, the actual execution, and the outcome observation. It supports a certain design of comparison, but cannot guarantee that every small table yields a precise effect, still less that results apply to all future clients.
Causal judgment also does not have this single channel alone. Observational material can participate in explanation under explicit process assumptions, comparable conditions, and other support. The distinction lies in whether the conditions relied on can be stated clearly and constrained by material, rather than labeling data as "experimental" or "non-experimental" and then declaring sufficiency.
Process Material Must Not Collect Only Fragments That Accommodate the Explanation
To further distinguish whether the reply request facilitated verification, one can check when the request was sent, when the other party began processing, and which step connected with the original conditions. Such process material may be closer to the candidate mechanism than a single final outcome.
But "confirming after finally seeing a reminder" does not by itself rule out that the other party was going to confirm at that time point anyway. Material must preserve the fragment that fits the explanation, and also examine objects where no reminder yet progress occurred and objects where a reminder yet no progress occurred, asking whether competing explanations can accommodate each respectively.
The actual receipt order of historical B items and each A item has not yet been verified item by item. The labels of short window and long window cannot substitute for temporal material. If A already occurred first, that B cannot serve as an earlier step that facilitated that A, even though the descriptive position within the fourfold table can still be preserved.
R17 has B but not yet A; this grants only the current item a known order, not the same order to all eight historical items. Comparison checks must resist this easy migration: being familiar with the object under discussion does not mean having acquired the full process of the other objects.
An Average Comparison Cannot Guarantee for Every Individual
Even if a supported average-effect judgment is eventually formed for a well-defined range, one must still distinguish average from individual. A process may help some objects while burdening others; an average result cannot automatically tell Gu Ning what the other path of a particular client would have been.
If one further groups by acceptance conditions, different differences may emerge, or only sparse material may remain. Grouping is one way to reopen a question; one cannot declare a small group that matches expectations to be the true mechanism while dismissing all other objects as irrelevant.
Likewise, even where the direction obtained by comparison has support, it does not directly complete a choice of action. The cost of one more prompt, communication arrangements, and other targets will be continued in the second volume on action problems. This volume first lets the judgment state which change it actually supports.
The scope of explanation should contract to the support obtained by comparison. A conclusion from one explicit process must not silently expand to "constant urging is always effective"; one clearly defined first-version class of objects must not silently expand to clients whose material has not yet taken shape. Actions may be similar, yet boundaries can still differ.
When Not Executed, Do Not Invent an Effect for the Assignment
Comparison must also distinguish what was assigned to be adopted from what was actually adopted. Suppose a request was assigned to the sentence-attached process, yet the sentence was omitted at the time of sending; the record should preserve both the assignment and the actual message. One must not see that confirmation was not ultimately obtained and then announce in reverse that it actually belonged to the old process, nor write it as executed in order to keep the new process table tidy.
Discussing the effect of assigning a process and discussing the effect of actually carrying out an action may depend on different material. If the omission is linked to being busy, regrouping by actual execution may again introduce differences in busyness. Such questions are worth further analysis, but the basis for grouping should be written separately, rather than letting the result choose the most favorable classification for us.
Chengwan has not even reached the stage of assignment, so this column is still a future item that comparison must preserve. A checklist can take shape before implementation; a facts column cannot be filled before the event.
Let Hypotheses Leave Interfaces That Can Be Questioned
Gu Ning looked back at the original competing explanations and found that comparison checks do not choose for them the most appealing name. The shared-preparation explanation needs material on preparation and on operation assignment; the facilitation explanation needs process continuation; the earlier-intention explanation needs a correspondence between intention formation and message appearance.
These materials cannot all arrive at once. What can be stated now is which has been obtained, which remains merely a hypothesis, and which gap will visibly affect the comparison. Different explanations need not be assigned equal credibility, but they must not be treated as equivalently established merely because they cannot be ranked for the moment.
Under RC's position of finite judgment, models preserve relations excerpted for the current problem. Comparison conditions make this excerpt easier to check; they do not turn the remainder of reality into a known fixed quantity, nor does a single table exhaust the process.
A counterfactual is therefore not a pass that lets thought bypass material. It makes the inquiry into unrealized paths explicit, then demands that we state our grounds. Imagination can generate competing positions, but feedback may either promote calibration or, when only agreeable material is collected, further rigidify old explanations.
What 4:40 p.m. Preserves Is a Checklist
At 4:40 p.m., Gu Ning preserved the comparison checklist: the candidate operation is limited to attaching one "received" reply request when sending the first version; the other side is limited to sending the same first version without the sentence; the target remains the original version and corresponding window's complete confirmation. Operation and actual reply are separated, and preconditions, later continuations, and inter-object relations are given separate places.
The eight historical items serve only as material awaiting reexamination, not as a named known-effect comparison for the new process. The receipt order of each A and B, how the request appeared, and whether comparable starting points sufficiently correspond still have gaps. The original five confirmations and three failures did not vanish because of gaps, and the four missing original records were not filled in as failures.
R17's first version, the original reply at 11:20 a.m., and the note on the station's schedulable period continue to be preserved for their original uses. This chapter has not verified the client's times, has not supplied the other real path, and has not given this request a final probability.
The next chapter examines a different substitution: when a later outcome is finally in hand, whether the story will rewrite the material and questions of the original prediction time. Comparison first preserves the gap of the unrealized path; the predictive record must also preserve what was actually known at the time along the path that was walked.