10.1What a distance measure compares
Consider the fictional intake system grouping short questions about transport access. Two users write "Keep the ramp open." If the system's declared representation is the sentence itself, those inputs coincide or sit close together. But A seeks delivery access from 08:00 to 09:00 while B seeks pedestrian access in the same hour, and in the stipulated example the two uses cannot safely occupy the single ramp simultaneously. Whatever grouping follows from the wording, it has not established that these authors want the same allocation; it has only shown that their sentences match under the chosen representation.
This is the point at which similarity becomes a representational claim. A distance function never compares every fact about two authors. It compares the declared representation of their inputs: surface wording, a chosen set of features, or an embedding model's output. For the second pair, "Provide a step-free entrance" and "Provide entry without stairs" differ lexically but request the same accommodation under the expressly fixed entrance task. The match is not read off the words; it depends on the task definition supplied to the comparison. Conversely, identical words can conceal incompatible interests once the hour and the users' purposes are specified.
Nothing here is an actual measurement: no embedding vectors, cluster output, or computed similarity appears in these examples. The proposed groupings are editorial comparisons of stipulated inputs. That is a real limitation, not a defect of the illustration—it keeps the question visible: which facts were allowed into the representation before distance was calculated?
The qualification that matters is therefore modest but consequential. Similarity is always relative to a representation, and the representation can be silent about interests, reasons, membership, or authority. A distance measure can tell a reader that two inputs are close under what was declared; it cannot tell a reader what was left out. How that limited output might assist retrieval without becoming an identity test is the next question.
The practical value of similarity does not require it to settle identity. In the fictional intake system, short questions about transport access can be grouped so that staff retrieve related cases quickly. That is an assistance role: the grouping brings requests into one view, and it may even suggest that their provisions overlap. It does not show that the requesters share an interest, a reason for the request, or membership in a population whose needs can be treated as interchangeable.
The limits appear when shared wording is examined closely. Two fictional transport users both say, "Keep the ramp open." User A seeks delivery access from 08:00 to 09:00; user B seeks pedestrian access in that same hour. For this example it is expressly fixed that delivery vehicles and pedestrians cannot safely use this single ramp simultaneously. Lexically, the requests are identical; practically, satisfying one use during the hour frustrates the other. A retrieval system that places them together has found a wording match, not a common preferred action, and certainly not evidence that the users belong to one group for allocation purposes.
Conversely, different wording can converge under a specified task. "Provide a step-free entrance" and "Provide entry without stairs" may request the same accommodation when the entrance task is expressly fixed as it is here. But that convergence comes from the task and the accommodation, not from measured proximity: no embedding vectors, cluster output, or measured similarity is supplied. Any grouping in these examples is a proposed editorial comparison.
The qualification that matters is therefore limited. Useful similarity can assist retrieval without providing an identity test for interests, reasons, or membership. Its scope is the declared representation and the task under which comparison is made. Beyond that, an allocation rule must be chosen and defended independently; similarity does not supply one.
10.2Same phrase and opposing purposes
The intake system can see that two requests repeat the same access phrase: both users say "Keep the ramp open." Their purposes part immediately once the requests' premises are stated. A seeks delivery access to the ramp from 08:00 to 09:00; B seeks pedestrian access in that same hour. For this example it is stipulated that delivery vehicles and pedestrians cannot safely use this single ramp simultaneously. The requests therefore concern different uses that cannot occur simultaneously. Dividing the hour might accommodate both; the example does not require continuous access or rule out time-sharing.
The inferential step is straightforward but consequential. Agreement in wording does not establish agreement in reason or interest. A grouping that places A and B together because they repeat the same phrase collects requests whose simultaneous fulfillment conflicts under the stated constraint. If such a grouping is offered as agreement about an allocation, it overstates the case. The users may share an interest in keeping the ramp usable while requesting different uses. If it is offered only as a marker of common vocabulary, it may be entirely accurate and yet remain silent about the questions that matter downstream: whether the users belong to the same constituency, whether they rely on the same reasons, and who holds authority to allocate the ramp. The shared phrase identifies a topic; it does not name a coalition or select an allocation rule.
The limits of the example should stay visible. No embedding model is run here, and no embedding vectors, cluster output or measured similarity is supplied. The two requests are fictional stipulations, so the case cannot show that a real similarity method would group them, or how strongly. It establishes something narrower: under the explicit shared time-and-space constraint, identical wording can conceal a conflict between simultaneous uses without settling whether another arrangement could satisfy both requests. Any proposed editorial grouping of these requests must therefore be judged by what it asserts. If the grouping stands for shared reasons, interests, membership or delegated authority, the common phrase does not supply that standing.
In the ramp example, the two requests collide only because three premises hold together: both users need the same ramp within 08:00 to 09:00, the ramp is single, and for the example delivery vehicles and pedestrians cannot safely use it simultaneously. Remove one premise and the shape of the dispute changes. If delivery access fell outside that hour, or if the stipulated safety constraint were absent, "Keep the ramp open" could be compatible in application: one policy might satisfy both users without choosing between them. That is a conditional result about the case as fixed, not an observation about any real site.
The inferential step is therefore modest but decisive. Identical wording neither creates the conflict nor settles it; the premises do. The phrase is the same in both mouths, yet it joins the speakers together only verbally. Whether their requests cooperate or compete depends on whether they share a reason, an interest, a membership, or a source of delegated authority, and on whether the physical constraint forces a choice. None of these facts follows from the sentence itself.
This bears directly on when similarity grouping becomes a representation claim. If grouping "Keep the ramp open" with other transport-access phrases is treated merely as a way to place cases side by side so their premises can be checked, it is an editorial comparison. It becomes a representation claim the moment the grouping is taken as evidence that the grouped users stand for the same need or speak with the same authority. No embedding model is run in these examples, no vectors or cluster output are supplied, and the shared wording alone selects no allocation rule.
This also meets the objection that bespoke groups are too expensive. Systematic retrieval can surface candidates at scale precisely because retrieval does not decide allocation; it indexes cases whose constraints then have to be read. Imperfect human groupings are improved by that discipline, not replaced by a similarity score. The qualification that matters is narrow: similarity can organize comparison usefully, but the allocation must come from the stated constraint and any actual delegation, never from the phrase or its cluster.
10.3Different language and shared needs
Different wording can nonetheless express the same request when the requirements for an acceptable answer coincide. Consider “Provide a step-free entrance” and “Provide entry without stairs.” Under the expressly fixed entrance task, the first phrase and the second require the same feature of the built environment: an entrance that can be used without encountering steps. Nothing turns here on semantic resemblance between the expressions or on their surface structure. What makes the requests match is that satisfaction of one would satisfy the other.
The inferential step is therefore from coincident answer requirements to a provisional intake grouping. If both users need an entrance with no stair obstacle during the same task, an intake system may compare them as instances of one accommodation request. This is a weaker but more defensible grouping than one based on words alone. It records a shared condition on the remedy: whatever proposal is offered must remove the stair obstacle at the relevant entrance.
The qualification is that agreement about the required accommodation does not establish agreement about why it is wanted or who may authorize it. The users may act from different interests, belong to different constituencies, or have different standing in a decision. None of those matters follows from the match in stated answer requirements. Identical phrasing can still conceal divergent purposes, as the ramp example makes plain: both users say “Keep the ramp open,” yet one seeks delivery access from 08:00 to 09:00 and the other pedestrian access in that hour, and the stipulated ramp cannot safely serve both simultaneously.
Here no embedding model, vector, or cluster result is being used. The comparison is editorial and rests on the fixed task. If requests are grouped by coincident answer requirements, the result represents a shared condition on the remedy, not a shared identity, motive, or delegated authority among the people making them.
Shared needs cannot be read off linguistic distance alone. In the fictional ramp case, two users make exactly the same request: “Keep the ramp open.” Yet one seeks delivery access from 08:00 to 09:00, while the other seeks pedestrian access in that same hour. Given the stipulated condition that delivery vehicles and pedestrians cannot safely use the single ramp simultaneously, identical wording does not settle whether their needs are compatible. Their stated answer requirements pull in different directions.
Nor does linguistic difference establish separate needs. “Provide a step-free entrance” and “Provide entry without stairs” can request the same accommodation under the expressly fixed entrance task. The phrases are lexically distinct, but the operative requirement is shared: access through the entrance without stairs. A collection may therefore bring the two together as an editorial comparison without assigning them to the same navigation group.
This is where a similarity grouping starts to become a representation claim. If surface likeness or lexical distance is treated as evidence about users’ reasons, interests, membership, or delegated authority, the grouping no longer merely arranges text; it asserts whose needs coincide. Here no embedding vectors, cluster output, or measured similarity are supplied, and no multilingual experiment has been performed. Any grouping is a proposed comparison, and the common wording does not itself choose the allocation rule.
The same caution does not require abandoning comparison at scale. Imperfect human groupings can still be compared when each grouping declares the task and the requirement used. What must remain separate is the record of correspondence—same or different wording—and the substantive judgment that two users need the same accommodation. That judgment is task-relative and provisional; it may be revisited when the access hours, safety conditions, or answer requirements change.
10.4Soft membership and boundary cases
The reverse problem also arises when one request genuinely spans categories, and here a provisional alternative is available. Take A's "Keep the ramp open": delivery access is sought from 08:00 to 09:00. That content touches both delivery access and the scheduling of a shared ramp, so the question can be carried into more than one navigation group rather than assigned to exactly one: a group for delivery access, and a group for time-bounded ramp availability. Both placements are provisional, because the phrase itself does not say which arrangement the intake system should privilege.
The inferential step is modest but useful. When a request's content spans several groupings, provisional membership in each of them preserves the relevant navigation paths instead of forcing an early decision from surface form. The alternative — dropping the request into one group and suppressing the others — would make the grouping look cleaner than the request is. Provisional multiplicity is therefore a way of representing boundary status honestly within a fictional intake display.
Nothing stronger follows from the arrangement. No embedding model has been run here, and no vectors, cluster output or measured similarity are supplied. The multiple placements are an editorial comparison, not an output of an algorithm, and the common wording "Keep the ramp open" does not choose the allocation rule.
The relevant limit is the conflict already stipulated in the case. B's "Keep the ramp open" seeks pedestrian access in the same 08:00–09:00 hour, and, for this example only, delivery vehicles and pedestrians cannot safely use that single ramp at the same time. Giving A's question provisional membership in both access groupings does not merge A's delivery interest with B's pedestrian interest, and it does not decide how the ramp hour should be shared. The same short phrase is compatible with incompatible uses of one physical resource.
So soft membership helps with boundary placement while leaving substantive questions of interest and allocation open.
Once a question is allowed to sit provisionally in several navigation groups, the display communicates doubt about which group should carry it. That improves on a single forced label, but it changes the presentation of uncertainty without settling what a membership value means. A high score for one group and a low score for another is intelligible only after one says what belonging asserts: similar wording, a shared reason, a shared interest, or membership in a class a delegate may speak for.
The fictional intake system groups short questions about transport access; no embedding model is actually run in these examples. Take the two users who both ask to keep the ramp open. One seeks delivery access from 08:00 to 09:00, the other pedestrian access in that same hour, and for the example only, delivery vehicles and pedestrians cannot safely use this single ramp simultaneously. If the two questions are placed in one transport-access group, they are grouped by wording and broad topic; if they are placed in groups keyed to delivery use and pedestrian use, they are separated by distinct interests and by the operational conflict. Membership scores do not decide between these arrangements, because each would answer a different question.
The second pair shows the reverse looseness. One question asks for a step-free entrance, the other for entry without stairs. Under the expressly fixed entrance task, both can request the same accommodation despite different wording. High membership in a common accommodation group would then track agreement in requested effect rather than verbal resemblance — an inference that follows from the stipulated task, not from the sentences alone.
Nor can the display supply the missing definition. No embedding vectors, cluster output or measured similarity are supplied; any grouping offered here is a proposed editorial comparison, and the common wording does not choose an allocation rule. A fractional membership records how firmly a stated rule sorts a question; it cannot say whether that rule concerns lexical or embedding similarity, shared reasons, interests, membership or delegated authority. Soft membership is therefore most useful where the chosen criterion is uncertain but already defined.
10.5Who names a cluster
Naming begins where the retained entries stop being self-explanatory. In the fictional pair, A and B both say “Keep the ramp open,” yet A seeks delivery access from 08:00 to 09:00 and B seeks pedestrian access in that same hour; stipulatively, the single ramp cannot safely carry both uses at once. Call the cluster containing these two entries “Ramp access support.” The name captures the shared utterance and a common stake in keeping the ramp available, but it makes the entries look like one constituency. It hides the opposed use of the one hour and therefore hides the decision that no grouping can make by itself.
Now retain exactly the same two entries and name the cluster “Competing ramp uses.” This name puts the conflict in the foreground: the shared hour and the incompatibility are the issue. Yet it hides the common request, the shared vocabulary, and the possibility that both users were recruited by the same intake question. Neither name is false. Each selects one feature of the pair while leaving the other available but less visible.
The inferential step is modest but important: naming the same retained entries differently changes which reasons, interests, or memberships a reader treats as relevant. The limit follows. These names do not allocate the ramp, define entitlement, or turn utterance agreement into shared reasons. Nor does the repeated wording itself choose between delivery and pedestrian use. Since no embedding model, vector comparison, or measured clustering is involved, these are editorial comparisons of two stipulated labels, not evidence about discovered structure. A label is therefore part of the representation claim: it supplies an interpretation that the common wording left open, even while the underlying conflict remains for a separate allocation rule to settle.
The question after naming is not only which feature a name brings forward. A label can sound merely descriptive and still carry an application assumption. In the fictional intake case, the two occurrences of "Keep the ramp open" might sit under a heading such as "Ramp access requests." The heading names a shared object without announcing a winner. Yet for the example only, delivery vehicles and pedestrians cannot safely use this single ramp between 08:00 and 09:00: A seeks delivery access in that hour, B pedestrian access. The common wording therefore licenses a comparison of requests, not a single schedulable demand. The wording itself does not choose an allocation rule.
The practical step is to record the naming purpose together with the name. If the name is meant to summarize intake wording, "ramp access requests" is a fair summary. If the same name is meant to authorize one ramp allocation during 08:00–09:00, it is far weaker than the task requires, because the stipulated conflict leaves the allocation rule open. Documenting purpose thus bounds what an application may draw from a name, while leaving the unresolved part visible rather than settled by vocabulary.
That qualification cuts against blanket suspicion of names. Not every name is equally misleading. The second pair, "Provide a step-free entrance" and "Provide entry without stairs," requests the same accommodation under the expressly fixed entrance task despite the different wording. A name such as "step-free entry requests" is well matched to that scope, because the fixed task supplies what the phrases share; there is no comparable conflict of purposes to conceal. The trouble arises only when a name moves silently from shared wording or a shared object to shared reasons, interests, membership, or delegated authority. Because no embedding model is run and no cluster output is supplied here, all of these groupings remain proposed editorial comparisons. What the record should carry is what the name is for, and what it leaves undecided.
10.6Changing cluster resolution
Once a cluster has a name, the next editorial decision is how far to divide it. Take the broad group of short questions about transport access. Splitting that group does not merely add categories; it brings into view distinctions that the broad label had folded together. The gain is diagnostic: disagreements that looked like repetitions become legible as different requests.
Consider the two fictional users who both say "Keep the ramp open." A wants delivery access from 08:00 to 09:00; B wants pedestrian access in that same hour. Under the expressly stipulated condition that delivery vehicles and pedestrians cannot safely use this single ramp simultaneously, their identical wording conceals an allocation conflict. A single group presents two satisfied users; two groups present a contested hour. Nothing about the repeated sentence selects between them.
The cost appears immediately on the other side of the ledger. Every distinction made visible must also be inspected. Splitting the transport cluster by access mode, time window, or user purpose multiplies the number of groups an editor has to review, and multiplies the items assigned to more than one group where a request spans several purposes. A reviewer who once read one list of short questions now reads several shorter lists and still has to check the boundaries between them. Resolution is therefore an editorial choice made under a finite review budget, not a free increase in detail.
One qualification governs the whole example. No embedding model is run, no cluster output is produced, and no measured similarity is available here. The two "Keep the ramp open" cases and the two step-free-entrance cases are stipulated comparisons. Their common wording or divergent wording does not itself choose an allocation rule; the proposed splits are arrangements an editor would have to inspect and justify, not findings recovered from the text.
Splitting a group into smaller ones makes more distinctions available, but that does not settle which resolution to use. A finer grouping is not invariably better. It earns its place when the distinctions it exposes are the ones a reader needs for the task at hand, and when those distinctions are supported by the relations retained from the source texts.
Consider the two fictional users who both say "Keep the ramp open." One seeks delivery access from 08:00 to 09:00; the other seeks pedestrian access in that same hour, and in this stipulated case delivery vehicles and pedestrians cannot safely use the single ramp at once. If the reader's task is to decide how the ramp is used during that hour, merging the pair on the strength of their identical wording conceals a conflict that the retained details make explicit. Here the finer resolution is warranted, because the users' purposes diverge even though their sentences coincide.
The second pair points the other way. "Provide a step-free entrance" and "Provide entry without stairs" use different wording, yet under the expressly fixed entrance task they request the same accommodation. For a reader inventorying requested accommodations, separating these two texts adds no useful distinction: the finer grouping would present one request as two without changing what is known about what the respondents seek. Merging them is defensible only because the fixed task licenses treating the two formulations as convergent; that convergence is stipulated, not inferred from surface resemblance.
The qualification is that neither judgment follows from the wording alone. The common wording of "Keep the ramp open" does not itself dictate an allocation, and the differing wording of the entrance requests does not itself dictate a split. The proposed groupings are editorial comparisons. Resolution should therefore be chosen against the reader's task and the source relations retained through it, not against a presumption that finer categories are always more faithful representations.
10.7Reporting a grouping without inventing a public
Once the intake records are written down, the relation established belongs first to the texts. In the fictional transport case, two entries repeat "Keep the ramp open." One requester seeks delivery access from 08:00 to 09:00; the other seeks pedestrian access during that same hour, and by stipulation the single ramp cannot safely serve delivery vehicles and pedestrians at once. The shared sentence supports a comparison of the two requests. It does not, on its own, dictate an allocation of the ramp between them.
A second pair shows a different basis for comparison. "Provide a step-free entrance" and "Provide entry without stairs" use different wording, yet under the expressly fixed entrance task they request the same accommodation. What groups them is the sameness of the requested change as the task defines it, not verbal resemblance. In both cases the available method is editorial comparison of recorded texts: no embedding model is run, no vectors or cluster output are supplied, and no measured similarity exists to report.
The inferential step is accordingly modest. When grouping is what the evidence supports, report a set of related texts and state the relation that holds among them. For the first pair that relation is common wording; for the second it is one accommodation under a stipulated task. This form of report records the comparison actually made and leaves the allocation question where the facts put it.
The limit is precise. A set of related texts establishes relation among texts, and no more; the common wording in the ramp case does not select an allocation rule. Beyond the set and its stated relation, further claims would each need their own evidence.
When the intake system groups short questions about transport access, the result warrants one claim: these texts are related enough to compare. Calling the result a constituency makes a different claim, about who the people are and who authorized anyone to speak for them. The fictional examples show why the step is not automatic.
Two users say "Keep the ramp open." A seeks delivery access from 08:00 to 09:00; B seeks pedestrian access in that same hour. Because, for this example only, delivery vehicles and pedestrians cannot safely use this single ramp simultaneously, the shared sentence conceals a conflict about what the ramp should do. Similarity here leaves the allocation rule untouched: grouping the two requests neither reconciles them nor shows that either user speaks for the other.
The second pair cuts the other way. "Provide a step-free entrance" and "Provide entry without stairs" differ in wording, yet under the expressly fixed entrance task they request the same accommodation. So common phrasing is not necessary for a shared accommodation, and common phrasing is not sufficient for shared interest. A proposed editorial grouping of these texts records a relation for comparison; it does not measure similarity, since no embedding vectors, cluster output or measured similarity is supplied, and it does not choose between competing uses.
The objection from scale has force against bespoke case-building only. Comparison at scale remains available; what must be withheld is the public-facing label. If an intake team can establish membership rolls, a declared mandate or delegated authority by other means, it may report a constituency and still admit that human groupings are imperfect. Without those means, the report stays at texts and claimed accommodation. That qualification is the limit: imperfection in human-defined groups is not permission to treat grouped texts as persons or as speakers. It also explains why the next count should keep submitted messages distinct from distinct senders and distinct claims.