1.1A shared sentence with different stakes
Consider a hypothetical shared library, in which a single sentence is offered as the community's position: the library should serve the community well. The sentence is short, memorable, and easy to place at the head of a report. It is the sort of formulation that lets a discussion begin at all, and its brevity is a practical advantage. Anyone arriving late can grasp the topic in a moment and see where the conversation sits.
Beneath that sentence, however, three quite different concerns are being raised. Some participants want the library open later in the evening, which raises questions of scheduling, safety, and operating cost. Others want quiet spaces preserved, which is a claim about how the building should be used now and about which uses should lose out. Still others argue that the work of running the library must be done by paid staff rather than by volunteers, a position bearing on labor, expertise, and continuity. These are not three wordings of one preference. They can conflict with each other, and none of them follows automatically from the broad sentence.
A wide common question can organize such a discussion without settling it. "What should the library be for?" gathers the concerns under one heading, invites them to be heard in sequence, and gives strangers a common starting point. That is compression doing real work: it lowers the cost of entry and prevents the conversation from fragmenting into isolated complaints. Yet the gathering is not agreement. The shared heading records that the three matters are in scope; it does not show that anyone consented to a single resolution, that a majority endorsed one, or that anyone was authorized to speak for the others. A model finding close verbal similarity among the statements would add little here, since shared vocabulary does not by itself merge distinct claims.
So the organizing sentence is useful as an index, and its usefulness should be acknowledged rather than dismissed. But usefulness, fidelity to what each concern says, and authority to speak for those who raised it are separate questions. What, then, is the common sentence entitled to stand for?
Suppose, hypothetically, that extending opening hours until midnight helps people who can visit only in the evening. Now add two conditions: late group activities use the same undivided space as quiet reading, and total paid staff-hours do not increase. Under those assumptions, the proposal could expand access while creating conflicts over quiet and thinner staffing coverage. The conflict does not follow from longer hours alone. It depends on how the additional hours are used and staffed, which the broad sentence about serving the community leaves unspecified.
Roberts' framework organizes discourse context around questions under discussion and connects relevance to addressing them (Roberts, 2012; earlier manuscript 1996/1998). That connection helps explain why a short sentence can work as an index: it points toward the live concerns without spelling them out. Such indexing may be useful, but the framework itself does not establish which concern ought to govern a public agenda, nor does it justify treating one participant's reading as another's endorsement.
Two groups could therefore agree word for word on the broad sentence and still answer the question of extended hours in incompatible ways. One foregrounds access and supports the change; the other foregrounds quiet or staffing and opposes it. Similar wording, in this hypothetical case, is weak evidence that the underlying stakes have been shared.
That does not make the broad question useless. As an index it keeps discussion organized and lets people recognize what is at issue, a real gain in accessibility over exhaustive qualification. What remains is a question of representation and authority—who may answer, and on whose behalf—which wording alone cannot settle. Rules for selecting among answers and formal aggregation belong to later chapters; here the narrower point is that length alone cannot tell a reader which open reading is being carried forward.
1.2Three senses of standing for
Three accomplishments are easily confused when one question is offered in place of many. A text may be shorter than what it replaces; it may report its sources faithfully; and it may be authorized to speak for the people who submitted them. Brevity is a property of the artifact, fidelity relates an account to the material it reports, and a mandate concerns who has authorized whom to represent them for a specified purpose. None of these entails another, though ordinary discussion slides easily among them.
Comparison is easier when it rests on a traceable unit rather than verbal resemblance alone. As a working definition, pair each summary question with an explicit record of the questions it compresses and of the transformation applied. Two summaries may be nearly identical in wording yet differ sharply in what they can be checked against. The W3C provenance model represents production history and responsibility relations, and allows provenance assertions to carry provenance of their own (Moreau and Missier, editors, W3C 2013). Such records can show who produced an account and from what material; but they remain assertions about history, and they do not by themselves establish that meaning was preserved or that authority was granted.
Consider a hypothetical case. A fictional library receives two hundred submitted questions, most concerning weekend opening hours. A staff member writes one short question capturing the recurring concern, and an editorial check confirms that it accurately reports that recurring concern without claiming to cover every submission. Yet the readers were asked for questions, not invited to appoint a spokesperson, and no vote designates the staff member to negotiate on their behalf. Fidelity succeeds; the mandate is absent. The shortening is useful, and its accuracy is verifiable, but neither fact supplies authority to negotiate for the contributors.
A question can therefore summarize a corpus without speaking for a population. The staff member can truthfully report a recurring concern while remaining unable to commit contributors to a particular response. Keeping that distinction visible permits the useful summary without silently turning a reporting task into a delegation.
A summary that claims less can still do real work. In the role of an index, a broad statement promises orientation rather than settlement: it tells a reader where to look, what distinctions were drawn, and which sources bear on which question, without asserting that its own wording carries the full content of the record. On this account the summary's value lies in reducing the effort of finding and comparing sources, not in substituting for them. That is a genuine benefit, and it should not be dismissed simply because the index is short.
Naming the role explicitly changes what preservation requires. If a condensed statement is offered as an index, its duty is to keep the path back to the underlying material open: sources remain traceable and accessible under declared conditions, or the limits on inspection are made visible. A reader can then determine which details are available for checking. The condensed statement need not reproduce every qualification at every moment in order to be honest about what it is. By contrast, the same wording offered as a verdict would carry a heavier burden, since it would then be asked to settle questions it only points toward. The difference is not in the words but in the promise attached to them.
A reader therefore need not see all material at once. Access and simultaneous display are different things; a layered presentation can surface a compact map first and the underlying passages on demand. Consider a hypothetical community-comment analysis in which a short index lists three positions, each linked to its source excerpts; the index aids navigation without claiming to be those excerpts. Small and colleagues report pilot exploration of language-model assistance for Polis facilitation, moderation and summarization alongside associated risks (Small et al., 2023), which is at most motivation for careful tool design rather than evidence that any such index has been validated. An index role may lower comparison costs while leaving source fidelity and representational authority to be argued separately.
1.3The benefit of a readable summary
A large collection creates an attention problem when readers have no practical way to identify what matters to their task. Reading every document may be unnecessary, while searching without a useful map can consume effort that a summary might reduce. The relevant burden is not a fixed function of document count: indexing, prior knowledge and division of labor can change it substantially. Compression is one possible response to that burden, alongside better retrieval and clearer organization.
A readable summary therefore earns its place by returning time and comprehension to the reader. This is not merely an intuition. Stiennon and colleagues train a summarization policy on human comparison data, reporting evidence that learned compression can improve over baselines on their preference task (Stiennon et al., 2020). The result is encouraging but bounded: the annotators, dataset, and comparison setup define what preference means there, and preference in that setting does not by itself establish factual accuracy or any broader mandate to speak for others. Where a shorter rendition preserves what readers need and is easier to use, it has practical value, and the burden of justification falls on demands for length as well as on summary.
Consider a synthetic example. Suppose a team must consult forty incident reports, of which perhaps four bear on the decision at hand. A summary that surfaces those four, with pointers back to the originals, changes the task from exhaustive reading to directed checking. Whether that change saves effort without concealing relevant reports remains a question to evaluate.
This argument is affirmative but limited. It gives reasons to seek useful compression; it does not establish a benefit in every setting or authorize replacing the source record.
There is an obvious objection to elaborate annotation: the apparatus meant to aid understanding can become a document more opaque than the text it accompanies. A summary laced with cross-references, provenance markers, and nested qualifications may demand more expertise to navigate than the original question. In that case, the annotation has not lowered the barrier; it has relocated it upward. This risk is real, and it is a reason to treat annotation design as a design problem rather than a duty to maximize documentation.
One response worth evaluating is an accessible summary with optional depth: a short, plainly written surface, with each element linking to progressively more detailed justification and provenance. The summary would carry the question in a form most readers can use, while the layers beneath would supply the qualifications, sources, and definitions for those who need or want them. This is a proposal to be tested, not an established solution. Layered documents can fail in well-known ways—readers may never open the depth, or may distrust a surface they cannot inspect, or may find the branching structure itself disorienting.
A synthetic example makes the accounting clearer. Suppose reading the compact version takes one minute and reading the full record takes twenty. A reader who needs only the compact version would save nineteen minutes under those assumptions. This subtraction says nothing about how long the summary took to prepare. Editorial work might take less or much more time, and its cost may be shared across many readers. Creation, maintenance and reader effort must therefore be measured separately before any net saving can be claimed.
The honest accounting therefore has two columns. A readable summary can materially widen access, and demanding every qualification inline can exclude through overload; both claims can hold at once. What the design cannot claim in advance is that its overall labor arrangement is sustainable or that the layers faithfully represent what they compress. Those are questions for evaluation—of legibility, of effort distribution, and of source fidelity—rather than assumptions built into the template.
1.4The disagreement a headline can hide
Consider a hypothetical pair of residents who both say that Bridge Street should be more accessible. One wants the narrow street closed to motor vehicles during the school arrival period so children can walk safely. The other wants delivery vans admitted during that same period because it is the only feasible time to supply the shops. Assume the street cannot accommodate both uses safely at once. Their preferred actions then conflict under a stated condition, even though the shared phrase sounds compatible. Repeating the phrase accurately would conceal the conflict if it were presented as agreement on an access policy.
This is a case where accuracy and representation come apart. Repetition of the wording preserves what was said, and a summary built from that wording may be perfectly faithful to the text it condenses. But fidelity to a phrase is not the same as authority to speak for the positions behind it. A shorter formulation can carry genuine advantages for readers who need an accessible entry point, and those advantages are not cancelled by the difficulty described here. The point is narrower: a common vocabulary can align surfaces while the underlying commitments diverge, so a match of words is not by itself evidence of a shared question.
Nor does the presence of disagreement settle which party is mistaken. One resident may simply hold a false belief about traffic patterns, which would be a factual error corrected by evidence. The other case is different: each resident may accurately understand the street and still rank safety against commercial continuity differently. Plank argues that annotation disagreement can reflect subjectivity or multiple plausible answers, and proposes studying it across the pipeline (Plank, 2022), a position that leaves room for genuine error as well.
Frequency does not decide the label either. If most residents repeat one reading of the sentence, that prevalence does not convert the minority's aim into a factual mistake; if both residents speak the same words, that coincidence does not establish that one position stands for the other. Reasonable disagreement and error must be identified by the reasoning and evidence involved, not by counting how often a formulation appears.
A dissent-centered design invites a serious objection: it can mistake friction for fidelity. If every qualification must travel with every summary, the resulting statement may be unusable in practice. Concise group statements can widen access, let readers grasp a shared position quickly, and make deliberation over large bodies of opinion tractable at all. On this view, insisting on permanent visibility for each objection is not neutrality but a design choice that imposes overload, and overload has costs of its own.
Tessler and colleagues report that participants preferred AI-mediated common-ground statements to statements written by human mediators (Tessler et al., 2024). That is a positive evaluated result and a serious reason to explore useful condensation. It should not be reduced to the mere fact that a preference comparison was conducted. The finding concerns a particular comparison, however; it cannot by itself show that every original objection survived or that a preferred statement may bind everyone affected. A design should accommodate evidence of usefulness while separately examining those further claims.
The key is to separate two questions that shared wording tends to merge. That a group statement is preferred, readable, and experienced as reflecting common ground is one claim. That the statement may bind others, speak for absent parties, or settle what the underlying questions were is a different claim, requiring different grounds. Preference is evidence of reception; authority requires a warrant that preference alone does not supply. Similar phrasing, majority agreement, or a high evaluation score can indicate convergence without establishing representation.
A concise statement can therefore be genuinely useful as an interface to a body of opinion, provided its function is stated honestly. The argument against binding by summary is not an argument against summaries.
1.5What the manuscript will establish
The object of this manuscript is a sequence rather than a system. It begins with questions as originally posed, continues through the grouping of related questions and the summarization of what they share, and ends with answers or decisions that might be offered in response. At each step, the manuscript asks what is carried forward and what is left behind. This is an analytical decomposition: it distinguishes stages in order to reason about them, not to describe an implemented product or to propose a new technical standard.
Each transition in the sequence is a site of both preservation and change. Grouping questions may preserve their common topic while suppressing differences of intent, scope, or assumption. Summarizing a group may preserve the concerns that appear most salient while dropping qualifications that mattered to particular askers. Conversely, some loss is not merely regrettable: a concise summary can make a large body of material accessible to readers who would otherwise face the full accumulation of questions. The aim is therefore not to demand that every qualification survive every step, but to make the relation between the condensed form and the material it replaces available for examination.
The sequence also makes it possible to locate a dispute. A contributor may accept a grouping but reject the wording of its summary; another may accept the summary but oppose the action later taken in its name. Those objections concern different transitions. Returning to the relevant input and editorial choice gives each objection a definite object, rather than forcing every disagreement into a dispute about whether the final sentence sounds accurate.
The following analysis therefore keeps transitions visible. A successful answer to the summarized question must still be assessed against the purpose and scope of the original inquiry. Moving through the sequence does not automatically confer new authority on its final output.
The manuscript advances three contributions, each carrying a different evidential burden. The first is conceptual: a set of distinctions separating the usefulness of a shortened question, its fidelity to the source material, and its authority to stand in place of the questions it replaces. These are separate properties, and the argument depends on keeping them separate rather than on demonstrating their empirical values in any particular population of readers.
The second contribution is a set of explicit counterexamples, constructed rather than observed. A summary might retain the relevant content and be welcomed by its readers while still lacking authorization to commit its contributors to a decision. Conversely, a representation that preserves one specified feature can discard another. The local toy calculations later in the manuscript examine target-dependent information, a judgment-aggregation conflict and a constrained comparison. Each establishes only a result under its stated definitions; none measures human judgment.
The third contribution is a procedure for evaluation and revision. Once a shortened question is treated as a traceable unit tied to the sources and questions it claims to cover, its acceptability becomes something to assess against declared criteria and to revise when the ties fail. This is a proposal awaiting empirical scrutiny, not a protocol in deployment.
Taken together, the three are meant to be useful beyond adding vocabulary. The distinctions would gain little if they merely renamed familiar complaints; they earn their place by yielding counterexamples that sharpen where each property can fail, and a procedure that turns the diagnosis into something checkable. None of this amounts to a survey of what readers believe, a benchmark ranking new models, or an algorithm that finds the optimal compression of any question set. Those remain different undertakings with different evidence.
A final boundary matters. The manuscript establishes conditions under stated definitions and through constructed cases; what real readers, annotators or institutions would do with them is not settled here.
1.6What remains an empirical question
The conceptual argument leaves practical questions that require observation of use. A declared comparison can identify which distinctions an editor intends to preserve; it cannot establish whether readers notice their loss, find the relevant sources or make useful corrections. Those outcomes should be evaluated alongside the added burden of producing and maintaining the record.
A prospective comparison could test whether readers locate a disputed source more accurately, detect a merged distinction more often, and revise affected summaries more successfully when a traceable record is available. It should hold source access and task demands comparable and record the time and assistance required. These are possible evaluation questions, not findings. They also permit an unfavorable result: an ordinary summary with useful links may perform just as well with less effort.
No participant comparison has been run in this work. The executed toy calculations concern a two-bit compression, a three-person judgment profile and an invented cost-quality comparison. They establish narrow relationships under specified assumptions. They do not track readers, measure correction behavior or validate a source-revision system. Those distinctions remain necessary even if every calculation is correct.
More generally, agreement between automatic measures and human judgments in a common summarization setting (Fabbri et al., 2021) does not transfer automatically to questions, institutional legitimacy, or the tasks described above. That study concerns news-style summaries, so its correlations cannot be read as evidence about source tracing or disputed representation. The hypotheses here would require their own studies, with outcomes such as successful source location, accurate identification of merged distinctions, corrected dependent summaries, and measured review effort. Until such evidence exists, the value of a traceable unit remains a reasoned proposal rather than an established practical result.
The strongest rival to any added framework is not a negligent baseline but a well-run one. Suppose a competent writer produces a summary that readers actually find clear, places links to the original questions and sources beside each claim, and submits the whole to an ordinary editor who checks wording, ordering and attribution. That arrangement already achieves much of what a heavier apparatus promises: compression without obvious distortion, wider access to material otherwise too long to read, and a shared reference point on which discussion can proceed. It may require little additional infrastructure, and its ordinary editing steps provide a concrete comparison. Any further structure must therefore show what this arrangement cannot do, rather than what a careless summary cannot do.
This study treats the added framework as losing if careful comparison shows no useful gain at higher cost. The relevant comparison is matched: same material, same editorial competence, one version carrying the additional tracing and review machinery. If the two perform alike on source fidelity, on the detection of misquotation or overclaim, and on readers' ability to locate what was actually asked, while the framework demands more authoring time, more maintenance, and more training, then parsimony favors the simpler method. Similar wording or a high agreement score in either version would not by itself settle matters; what would matter is whether errors are found and corrected.
Two further results would count against the proposal. If the intended reviewers cannot use the tracing record to perform the tasks it was built for, its existence adds little; a large number of casual readers is not itself required. If valid corrections cannot reach dependent text, the correction channel also fails. Rejected objections need not change the manuscript, but the process should distinguish a reasoned rejection from an ignored correction and apply the changes it has accepted. Both conditions can be examined without assuming that every complaint is right.
The local toy comparisons in this study are too small to settle these questions, and no human study of the contrast has been run here. What is needed is evidence that the framework earns its overhead under competent ordinary practice. Until then, the simpler method remains a serious contender, not a foil.
1.7A map of the argument
The argument advances in eight parts. Part I, Questions before summaries, defines the objects being compared: an original question, a reformulation, the context in which each is used and the claim that one stands for the other. Establishing these objects first prevents later evaluation from rewarding an attractive sentence while leaving unclear what it is meant to preserve.
Part II, The mechanics and limits of compression, examines codes, relevance targets, grouping and successive summaries. It makes explicit which features a reduction retains. Part III, Whose question becomes the common question, then examines collective judgment, participation, disagreement and agenda setting. The transition matters because a technically useful reduction does not itself supply a rule for representing people.
Part IV, Language models as fallible editors, considers factual support, evaluation metrics, preference learning and automated judging. Part V, Records that permit disagreement and correction, asks what a reader needs to inspect or revise the resulting work. Together these parts connect the quality of an edited sentence with the evidence and responsibility behind it, while keeping a recorded link distinct from a justified interpretation.
Part VI, From an interface to an institution, compares documented participation and deliberation approaches and examines the roles surrounding an editor. Part VII, A research program that can fail, specifies synthetic contrasts and prospective evaluations, including conditions under which simpler practice would be preferable. Part VIII, Applications, limits, and a revisable synthesis, returns to urgency, absent voices and open inquiry. It closes by distinguishing what the analysis supports from what further evidence could change.
A summary earns its place when three modest conditions are met. First, it declares what kind of claim it is making: a reduction of text, an interpretation of intent, a recommendation for action, or some other specified function. Without such a declaration, a reader cannot judge whether the summary is faithful to its source or merely persuasive. Second, it preserves the distinctions the task actually requires. A summary for a policy audience may need to retain conditions and exceptions that a summary for rapid orientation may safely omit. The criterion is not exhaustiveness but correspondence to the purpose at hand.
Third, a useful summary offers a practical route to inspection and revision. A reader should be able to ask where a claim came from, what was compressed or set aside, and how the summary might be altered if those omissions matter. This is a procedural standard rather than a guarantee of correctness. It makes disagreement diagnosable: parties can locate the point at which they diverge over what the summary claims, what it preserves, or who is entitled to speak for the questions it addresses.
The benefits of this standard are real but conditional. A concise summary can open access to material that would otherwise exclude readers through overload, and it can help a group focus on shared concerns. The costs are equally setting-dependent. Compression may drop qualifications that later become decisive, and a route to revision may be more demanding than the task warrants. Neither outcome is fixed in advance; both depend on the stakes, the audience, and the consequences of error.
As noted earlier, tracking questions under discussion can orient a summary without deciding whose questions should govern a public agenda. That gap opens the problem the next chapter must confront: when questions are treated as interchangeable, and how their identity might be held accountable.