9.1A sequence of editorial transformations
Read the editorial work as a chain of transformations rather than a single act of compression. The starting text, T0, consists of the two invented comments: A accepts the proposed levee only if downstream path access is guaranteed, and B refuses that guarantee while an easement is disputed. The first summary, T1, retains both conditions. The second, T2, replaces them with a sentence asserting that both parties want the neighborhood protected. The hypothetical notice, T3, changes "may proceed after agreement" to "will proceed." Each stage thus has an identifiable input and an identifiable output: T0 into T1, T1 into T2, T2 into T3.
Writing each stage as an ordered pair does modest but useful work. It fixes which text a sentence belongs to and what the preceding stage held. The sentence at T2 is an output of the step from T1, not a fresh reading of T0, and the firmer wording at T3 is an output of the step from the earlier notice. The notation also supplies the place to ask whether an omitted condition was deleted or left behind somewhere recoverable, since that question needs both the earlier text and whatever links still connect the stages. Identifying inputs and outputs is the precondition for the distinction, not yet its resolution.
The limit matters here. T2 may genuinely index a shared aim, protection of the neighborhood, yet it cannot resolve A's access condition or B's refusal while the easement remains disputed; agreement on an aim is not agreement on terms. The firmer commitment at T3 likewise has no supplied agreement or new evidence behind it. Were either supplied, it would count as a further input to the transformation; on these stipulated facts there is none, so the changed modality cannot borrow the authority of an administrative decision. The sequence shows where a change occurred and which text it came from. Whether a stage preserved the meaning of its input is a separate question.
Once successive stages are represented as distinct input–output pairs, a stage boundary gives the analysis somewhere to stand. It identifies the text before and after a transformation and therefore locates the point at which a commitment, qualification, or link may have changed. That precision matters, but it is evidentially modest: a clear boundary helps locate a change without proving that the transformation preserved the meaning of its input.
The fictional municipal comments make the distinction concrete. Contributor A accepts the proposed levee only if downstream path access is guaranteed, while Contributor B refuses that guarantee while an easement remains disputed. A first summary retains both conditions. A second summary instead reports only that both contributors want the neighborhood protected. The two versions can be assigned to separate stages, and the sentence expressing a common aim is indeed present in the second output. Yet identifying that output tells the reader where the stated conditions disappeared, not that the underlying disagreement has been resolved.
The second summary can index a shared protective aim, but it cannot determine access, settle the disputed easement, or erase the conditional structure of A’s acceptance and B’s refusal. Any inference from shared purpose to agreement would add a commitment that neither the first summary nor the original comments supply. Comparing the inputs and outputs therefore remains necessary even after their sequence is cleanly marked.
The stage boundary also clarifies what must be shown before a loss can be treated as recoverable rather than irreversible. If the omitted conditions remain available through a link to the earlier summary, readers may reconstruct what the later wording left out. If no such path accompanies the transformation, the deletion is operationally final, however orderly its place in the sequence appears. A further hypothetical notice that changes conditional permission to state that the levee “will proceed” does not close this gap, since no agreement or new evidence has been supplied. The stage is easy to locate; preservation of meaning is a separate question.
9.2Omission at one stage
A stage boundary tells a reader where to inspect a change; it does not by itself show that the transformation preserved what its input carried. Take the second flood-control summary in the fictional municipal case. The first summary retains two conditional positions: Contributor A accepts the proposed levee only if downstream path access is guaranteed, and Contributor B refuses that guarantee while an easement is disputed. The second compresses them into a single sentence reporting that both want the neighborhood protected. The sentence is grammatical and does not state either contributor's condition. Even if its proposed shared aim is accurate, that accuracy would leave the conditions unreported. The positions belong to invented contributors, so nothing turns here on the actual legal status of the easement; what matters is that the dispute is a stated condition in the input.
The inferential step is to test completeness rather than coherence. A reader can ask which commitments the input carried and whether the output still carries them. Here the answer is plain: both conditions disappear, and with them the disagreement about the access guarantee. What the sentence offers is an asserted shared aim; even when accepted as an accurate index, it neither resolves nor records the two conditions. Absence of contradiction therefore underdetermines completeness: a summary may omit a decisive qualification and remain perfectly well formed. The point is not that every summary must reproduce every clause, but that internal consistency cannot certify what has been retained.
Whether the loss is recoverable depends on the surrounding record. If the first summary remains accessible, a reviewer can return to its explicit conditions. Links may make that route easier to locate, but the example does not establish that such links exist or that a reader follows them. Losing the intermediate summary would remove that route without proving that the original contributions had also vanished. The second summary's silence is therefore a reason to inspect the available sources, not evidence that their authors supplied no conditions.
The key question, then, is not simply what the sentence omits but what work the sentence is being asked to perform. The second summary's claim that both contributors want the neighborhood protected can serve as a topic index: it marks a shared concern and points a reader toward material about flood-control protection. For that role, the omission of the two conditions is tolerable because the index is not expected to carry the record's commitments.
As a decision brief, however, the same sentence is inadequate. Contributor A accepts the proposed levee only if downstream path access is guaranteed; Contributor B refuses that guarantee while the easement is disputed. A reader choosing whether to approve the levee or permit construction must see both conditions and their unresolved conflict. Removing them does not merely shorten the account; it removes the grounds on which the decision was contested.
The inferential step is therefore one of role assignment. A compressed statement acquires authority only from the task explicitly claimed for it. If the claim is “this indexes the topic,” the sentence does its limited work. If the claim is “this supports a decision,” the sentence must recover the conditions, their speaker-by-speaker form, and the fact that the disagreement remains open. The sentence itself gives no warrant to cross from the first role to the second.
That limit matters because intermediate summaries may genuinely improve robustness by organizing evidence, and direct processing can be fallible too. The objection does not justify silent promotion. Organization is useful when the summary's level of detail matches its stated purpose and when links back to the fuller record remain available. Without those links, the omission is not recoverable, and a later summary may gain a standing that its author never possessed. The practical rule is modest: name the role, then ask whether the retained facts suffice for that role. If not, restore what the role requires rather than infer it from smooth prose.
9.3Compounding selection at later stages
Consider a hypothetical two-stage summarization of a municipal record on flood-control works. The record contains two conflicting conditions: one stakeholder accepts a levee extension only if downstream access to the river path is guaranteed, while another refuses any guarantee because the easement remains disputed. A first summary that respects the record keeps both conditions distinct, marks each as conditional, and notes the unresolved disagreement.
A second summary compresses this output to a shared purpose: the parties want the neighborhood better protected. As an index of that aim, the sentence could be accurate. Trouble arises if an editor presents it as a resolution of the levee proposal. Under that added condition, the shared aim can appear to settle a dispute about means even though the two conditions remain incompatible. The failure concerns the use made of the sentence, not the inevitable interpretation of its wording.
The omission is hard to spot when readers cannot return to the inputs. A missing distinction leaves no contradiction in the surviving text, so nothing cues a reader to ask which conditions were dropped or on whose authority. If the first summary's entries remain linked and retrievable, the loss is recoverable: a reader can reopen the record and notice that the guarantee was contested. If links and access are gone, the same omission is effectively irreversible, since the surviving summary alone does not contain enough information to reinstate the disagreement.
When used in that way, the second summary is given a role its retained content cannot justify. This inference concerns the hypothetical case only, not a measured defect in any method. Recursive decomposition can work well where intermediate outputs are organized and auditable; Wu et al. decompose book summarization and recursively summarize intermediate outputs under human feedback (Wu et al., 2021), and their approach is not evaluated here as defective. Compression at a later stage can nonetheless convert shared aims into apparent consensus.
Recoverability depends on more than keeping a record somewhere; it depends on whether that record can actually be retrieved and compared. In the fictional flood-control comments, the first summary preserves both stated positions: Contributor A accepts the proposed levee only if downstream path access is guaranteed, and Contributor B refuses that guarantee while the easement is disputed. If this first summary remains accessible, a reader encountering the later sentence, which compresses both positions into the shared aim that both want the neighborhood protected, can line the two records up. The comparison then shows that the sentence correctly indexes a shared aim but leaves the conflicting conditions untouched.
An inaccessible pointer cannot support the same claim. Suppose the second summary names the first as its basis, but the cited record cannot be opened: the link is dead, the attachment has been replaced, or the repository is closed. The pointer still marks where information once sat, yet it returns no content to compare. The omission is then recoverable in name only, and it comes close to irreversible deletion even though some record nominally survives. The burdens of the two devices are therefore distinct. An index bears the burden of working access and sufficient fidelity; a purported resolution bears the burden of showing that agreement or new evidence actually settled the conditions.
Accessibility does not itself repair a compressed summary, and it does not certify direct processing as reliable. Its narrower value is that an organized, auditable intermediate output can be consulted when a later account looks decisive. Consider a hypothetical notice that changes "may proceed after agreement" to "will proceed." No agreement and no new evidence are supplied. Retained intermediates cannot justify that changed commitment; they allow a reader to locate the change as an unsupported step. Keeping the earlier summary available therefore enables comparison and recovery, but the authority of a later summary still has to be earned separately.
9.4Useful decomposition
Dividing a long collection into segments can make focused checking possible and give each segment a manageable responsibility assignment. Applied to the fictional flood-control comments, one checker can be assigned Contributor A's stated condition—that the proposed levee is acceptable only if downstream path access is guaranteed—while another takes Contributor B's refusal to give that guarantee while the easement is disputed. Neither assignment requires anyone to hold the entire collection in working view at once, and each check has a defined scope against which a mistake can be named.
The decomposition earns its keep when the results of those checks remain retrievable. If the first summary preserves both conditions with links to the contributing comments, a later reader can test any compressed statement against them. Even if the second summary's proposed shared aim is accurate, it does not resolve the access condition. A reviewer needs the separately stated conditions to inspect that gap. Reachable conditions provide a concrete checking route. If that route is lost, the reviewer must seek another surviving source; no checking time or cost has been measured here.
Recursive task decomposition has been used productively elsewhere: Wu et al. combine human feedback and recursive decomposition for fiction summaries, reporting useful summarization and question-answering results (Wu et al., 2021). The relevance is procedural rather than evidential. The reported work concerns fiction books, not civic-source provenance, and it does not establish stable source pointers, civic representation, complete argument preservation, or present-model performance. Nothing in it licenses treating a segmented pipeline as safe for the municipal case.
The qualification that matters is narrow. Accessible intermediate summaries and source links support the particular checking route proposed here. They are not established as necessary conditions for every form of error detection: a reviewer might instead compare a draft directly with surviving originals. Where a later step changes conditional permission into an unconditional statement—say, a hypothetical notice that turns may-proceed-after-agreement into will-proceed—the decomposition can still localize which segment lost the condition, provided the earlier summary survives. If it does not, this intermediate inspection route is lost, although surviving original comments may provide another route.
Dividing the municipal comments into sections gives each condition a manageable unit of review: one reader can check Contributor A’s acceptance of the proposed levee against its stated condition, another can check Contributor B’s refusal of downstream path access while the easement is disputed. That division assigns responsibility cleanly. Its benefit is limited to the scope of the checks it enables; it does not settle the disagreement between the conditions.
Direct reading has limits of its own. A reader may mistake either conditional position for a general endorsement of the levee, and that error may survive a first pass. But it remains an error attached to identifiable source language, so rereading the comments can expose it. Compare what the recursive sequence leaves detectable. If the first summary retains both conditions and carries links back to them, the second summary’s hypothetical sentence—proposing that both contributors share a goal of protecting the neighborhood—can be treated as an index to a shared aim. It cannot resolve A’s demand for guaranteed access or B’s refusal while the easement is disputed. The omission is recoverable, because the earlier text and its links still disclose the disagreement.
If, instead, the intermediate summary replaces the two conditional positions with that sentence and discards the links, the disagreement is no longer recoverable from within the chain. That is an irreversible deletion, not merely a shorter account. The error becomes harder to locate precisely where the summary gains a semblance of settled consensus it was never given.
A later hypothetical notice that changes “may proceed after agreement” to “will proceed” illustrates the danger. No agreement or new evidence is supplied, so the stronger commitment borrows authority only from the flattening summary. Conditionally, an independent record could warrant such a change; nothing here supplies one. Recursion can organize evidence and improve robustness, and direct processing is fallible too. The relevant question is therefore which errors remain detectable: linked omissions preserve that possibility, while deleted conditions and invented commitments do not.
9.5Source access after recursion
A summary need not repeat every earlier passage in order to keep its omissions recoverable. It can instead carry pointers to the source passages it compresses. In the flood-control example, the first summary records that Contributor A accepts the proposed levee only if downstream path access is guaranteed, and that Contributor B refuses that guarantee while the easement is disputed. A second summary stating that both contributors want the neighborhood protected correctly indexes a shared aim. On its own, however, it cannot resolve the two stated conditions, because neither condition is a claim about aims. If the second summary points back to the passages where A's condition and B's refusal appear, its brevity remains recoverable omission rather than irreversible deletion.
That distinction bears directly on the later hypothetical notice, which changes "may proceed after agreement" to "will proceed" without any agreement or new evidence. The change in commitment cannot be explained by the common aim: wanting protection does not settle whether the access condition was met or the easement dispute resolved. With the pointers in place, a later reader can locate the earlier conditions and see that the stronger wording has no supplied basis in them. Without pointers, the same gap appears as ordinary compression, and the second summary's organization may look as though it has absorbed and settled what it only grouped. Organizing evidence can improve robustness, but robustness is not authority over conditions the summary did not adjudicate.
The inferential limit is that a pointer preserves recoverability only to the extent that it marks the relevant passage and that passage remains reachable. Whether a later reader can actually perform the check turns on the pointer's stability, access conditions, and scope; those determine what survives after recursion and are the next question to settle.
A pointer to an earlier passage lets a later reader reopen the evidence that a summary compressed. But a pointer is not automatically a route to the source: three features decide what it can actually be checked against. Its stability matters because a link that can be rewritten or withdrawn without notice cannot anchor a claim that outlives the first summary. Its access conditions matter because a pointer to a passage behind a disputed or restricted gate gives the later reader a location but not the text. And its scope matters because a pointer to one sentence of the first summary says nothing about the conditions recorded elsewhere in it.
Take the stipulated municipal example. Contributor A accepts the proposed levee only if downstream path access is guaranteed; Contributor B refuses that guarantee while an easement is disputed. A first summary retains both conditions and marks each one with a pointer to the contributor's own statement. A second summary says only that both want the neighborhood protected. That sentence may index a common aim, but it cannot resolve the two conditions. Pointer scope is decisive here: if the second summary links to the whole first summary, both conditions remain checkable; if it links only to its own generalized sentence, the later reader cannot recover them.
This separates a recoverable omission from an irreversible deletion. Omitting the conditions from the later sentence is harmless while pointers with stable scope and usable access still lead to them. Deleting the passages, or pointing at a location the reader cannot open, removes the evidence while the summary's apparent authority survives intact. It then speaks for two positions whose disagreement it cannot display. Conditional on the pointers remaining live and reachable, the second summary's authority stays contestable rather than inherited.
9.6Detecting and locating a changed claim
Consider two explicit fictional versions of a notice about the proposed levee. Version 1 says that work "may proceed after agreement" on the disputed downstream path access. Version 2 says that work "will proceed," with no agreement condition stated. The difference is not merely stylistic. "May" leaves authorization open; "will" asserts that the project is going forward. Dropping "after agreement" removes the condition that tied that authorization to the access dispute.
To locate the change, compare the versions term by term. Contributor A accepted the levee only if downstream path access were guaranteed; Contributor B refused that guarantee while the easement was disputed. Version 1 preserves both sides of that tension by making progress conditional on agreement. Version 2 records no agreement and no new evidence, yet states an unconditional commitment. The altered claim therefore arose at the revision step between these two notice versions, not in the underlying fictional comments and not in any administrative decision, since none is supplied.
Earlier compression makes the later change harder to see. The first summary retains both conditions. The second says only that both contributors want the neighborhood protected. That sentence may index a common aim, but it cannot resolve the two stated conditions. Once the summaries are read as sufficient, a later editor confronting "may proceed after agreement" has less material resisting the cleaner sentence "will proceed." The summaries did not authorize the project; they simply removed the wording that made the conditional commitment visible.
A version comparison can thus detect the modal shift and identify the revision point where it occurred. That is a finding about the documents. Whether the change was justified is a separate question requiring the source and editorial purpose. For now, the precise diagnosis is narrow: the commitment changed from conditional to unconditional, and the change is visible only because both versions and the surviving conditions remain checkable.
Locating the alteration is therefore a narrower result than evaluating it. A comparison of the two fictional notices shows exactly where “may proceed after agreement” became “will proceed.” That comparison establishes a changed commitment: what was conditional on agreement now reads as unconditional. It does not establish why the wording changed, whether the change was authorized, or what administrative act the later notice records. No agreement or new evidence is supplied, so the comparison cannot fill that gap by inference.
Evaluating the change requires two further ingredients. The first is a source capable of supporting the stronger commitment: perhaps an agreement, a decision, or an explicit editorial instruction. The second is knowledge of the notice’s editorial purpose. A notice intended to report settled policy may appropriately state a stronger commitment; a summary intended merely to preserve the comments may not. Without those ingredients, the proper finding is simply that the later wording departs from the earlier stated condition.
This limitation of version comparison clarifies the scope of detection. Version comparison can reveal compression damage even when it cannot adjudicate it. It also preserves a distinction that a chain of summaries tends to erase. Contributor A’s access condition and Contributor B’s disputed-easement refusal remain recoverable as long as links lead back to both statements. The summary sentence stating that both contributors want the neighborhood protected may index a shared aim, but it does not repair the loss of those conditions or authorize a stronger procedural claim.
The strongest objection to relying on direct comparison is that intermediate summaries can organize evidence and sometimes process it more reliably than casual reading. That benefit is real but conditional: a summary helps only while its links remain usable and its purpose is declared. Once later summaries are treated as self-authorizing, their convenience turns a recoverable omission into an apparently settled fact.
9.7A practical stopping rule for re-summary
A version comparison can tell a reader where wording changed, but it cannot by itself determine when another round of summarizing should stop. For the declared task here—preserving the conditions under which contributors accept or reject a proposed levee, together with the disputed access issue—the distinction is what the summary must carry. So the practical rule is: stop condensing when the next reduction would remove a distinction that the declared task requires.
The fictional flood-control comments make the point concrete. The first summary retains both positions: Contributor A accepts the levee only if downstream path access is guaranteed, while Contributor B refuses that guarantee while the easement remains disputed. A second summary reduces this to a sentence indexing a shared aim, such as a statement that both contributors want the neighborhood protected. Such a statement can index a common aim, and it is not false. But it cannot resolve the two stated conditions, and it no longer records that agreement depends on one contributor's acceptance being conditional and the other's being withheld. Condense further and the difference between conditional acceptance and unresolved refusal disappears.
The same distinction governs the later hypothetical notice, which changes "may proceed after agreement" to "will proceed" without any agreement or new evidence. The shift is from a conditional commitment to an unconditional one. Whether the change is justified requires the source and the editorial purpose; what the stopping rule supplies is narrower but decisive: once the wording needed to track that condition is at risk of deletion, re-summary stops, because no later summary can recover the condition from the compressed text alone.
The limit is task dependence. The rule does not prohibit compression in general; it intervenes only where a distinction carrying the declared task's required relations would be lost. A summary may therefore index a common aim safely, provided it does not do so by erasing the conditions that distinguish the contributors' positions.
Apply the rule to the municipal flood-control comments as they stand. Contributor A accepts the proposed levee only if downstream path access is guaranteed; Contributor B refuses that guarantee while an easement is disputed. A first summary preserves both conditions. A second reduces them to a sentence indexing a common aim, such as a claim that both contributors want the neighborhood protected. That sentence can index a common aim, and as such it may organize the material usefully, but it cannot resolve the two stated conditions or show how the disagreement bears on what may happen next.
Here the declared task supplies the test. Suppose a later hypothetical notice changes "may proceed after agreement" to "will proceed." The task now requires asking whether the stated comments justify the stronger commitment. Answering that question depends on the distinction between conditional acceptance and refusal during a disputed easement. If re-summary removes that distinction, the surviving text can no longer support an answer to the question that the notice raises. The stopping rule therefore says: stop further condensation at the point where the next reduction would drop a distinction the declared task needs. No agreement and no new evidence has been supplied, so the change from "may proceed after agreement" to "will proceed" remains unexplained by the comments themselves.
The rule is editorial and provisional, not an optimal universal compression threshold. It fixes a stopping point for a particular task and a particular set of distinctions; it does not rank summaries globally or guarantee that direct processing of the original material will be error-free. Re-summary can still earn its place by exposing a common aim or by organizing evidence for readers who need a map. The limit appears when compression stops preserving what the task asks about. A different task may warrant revisiting the stopping point: if the only question were whether the contributors share a concern for neighborhood protection, the shorter sentence might suffice. That would change the declared task, not retroactively make the shorter sentence adequate for the commitment question.