This study asks what a summary cannot decide. Its answer is bounded rather than universal. Compression is an editorial act, and certain questions are not settled by making text shorter. It cannot choose the target to be preserved: a representation may preserve one distinction perfectly while losing another, and relevance requires a target before it can be measured. It cannot determine which losses matter most, because a taxonomy of lost propositions, distinctions, speakers, stakes, uncertainty and revision options embeds values and may miss culturally specific harms. It cannot establish who counts as a constituency, since lexical or embedding similarity is not shared reasons, membership or delegated authority. And it cannot convert asking into authorizing: proposing, endorsing, ranking, deliberating and binding are distinct acts that shared text does not merge. When people disagree, compression cannot declare the disagreement resolved; some disagreement is the finding rather than noise.
The study compares positions without collapsing them. An engineering tradition states that semantic aspects are irrelevant to the engineering problem, and the information bottleneck compresses by preserving information about a specified relevance variable. These frames sharpen the boundary between a measure of selection and transmission and an account of meaning, but neither chooses the relevance variable for a reader. Against that, the access case is real: concise summaries can widen participation, and insisting on every qualification can exclude through overload. The proposed response is neither maximal qualification nor untraceable condensation, but declared purpose, scoped records, plural outputs and identifiable transformations. Reported evidence that participants preferred AI-generated common-ground statements to human mediators is treated as a favorable comparison at abstract level, not as proof of legitimacy or of preservation of dissent. The judgment aggregation result and the study's own fully specified toy profile show that consistent individual judgments do not by themselves yield a consistent collective judgment. A PROV-style provenance model can record entities, activities and agents across collection, transformation and review, but a provenance edge remains an assertion about history, not evidence that meaning survived.
Limitations are substantial. All examples are constructed, and all empirical evaluations are proposals; no participant outcomes are reported. Source cards were consulted at abstract or bounded-excerpt level, without full semantic audits, so detailed use would require full-section reading. Toy deductions, normative commitments and design hypotheses are separate in kind and are not presented as verified findings.
Tests that could change the account are already specified: a prospective reader study of retrieval, interpretation, correction and effort; paired cases varying one property at a time; separate judgments of preservation and readability; and adverse findings that simpler tools work as well, review costs rise without gains, tracing fails, dissent stays ceremonial, or corrections never affect descendants. The bounded conclusion is that compression can support attention, discovery, comparison and coordination, while representation and decision remain accountable human choices.