23.1Having text versus using it
A model's input records what was made available; it does not by itself record which available passage shaped the answer. Take the invented collection with an early proposal that the service be "open until 20:00" and a later qualification limiting the proposal to cases where a second staff member is available. A chunk boundary separates the two statements. Whichever side is retrieved, it remains present in the collection as a whole even when it is absent from the immediate working set. Presence is therefore a fact about transport into the input, while use is a claim about influence on the answer, and the two must be evidenced separately.
The inferential step is modest but decisive: if the qualification is not before the model in a form it can apply, an unconditional recommendation to stay open until 20:00 cannot be justified from the text, even though the supporting sentence exists somewhere in the collection. Conversely, placing both statements in context establishes only that the model could have reconciled them; it does not establish that the later condition governed the answer. Evidence of use has to come from something that tracks the relevant passage — for example, whether the answer's stated commitment is appropriately conditional on the second staff member — rather than from the mere fact of inclusion.
This distinction between presence in the input and use in the answer is consistent with the reported retrieval and question-answering tasks, where access to a long context does not ensure uniform use of its contents (Liu et al., 2023 preprint; TACL version noted). That result concerns historical model and task conditions; it does not transfer a magnitude to current models or to every long document.
The same care applies to the time-scope case: "the room had 12 and 20 seats" drops the periods that make the two figures compatible. Restoring the periods is a condition of intelligibility, not a causal explanation of the change.
A correct answer can be correct for another reason. In the invented collection, an early proposal says the facility is open until 20:00, while a later qualification allows this only if a second staff member is available. A chunk boundary separates the statements. If a system, asked whether the facility is open until 20:00, answers yes, that reply may reflect the early chunk alone, a prior belief, or a lucky guess. The result is compatible with retrieval of the qualification, but does not show it. Nor does the mere presence of the later chunk in the input establish that it influenced the answer.
The useful inferential step is therefore to ask whether the answer preserves the dependency that the text actually states. An unconditional commitment to 20:00 discards the condition; a response that makes the closing time conditional on the second staff member at least applies both statements together. That is stronger evidence than a bare correct outcome, because it constrains the reason the answer was produced. Still, it is not conclusive: a model could reproduce the condition from another source or infer it from general knowledge. To distinguish those cases, a matched-position test could hold the task fixed while placing identical relevant material at different locations, or removing the qualification altogether. This is a proposal for future work, not a report of an experiment; no present model has been tested, and no historical effect size is transported from elsewhere.
Scope application appears even more simply in the time case. The room had 12 seats in year A and 20 seats in year B. A summary saying the room had 12 and 20 seats strips away the periods and invites a confusion between concurrent and sequential values. No causal explanation for the change is supplied, so none should be inferred. The same discipline governs retrieval: an answer is adequate when it retains the qualification's conditions and the periods attached to each figure, not merely when it names a number or a closing time.
23.2Position and retrieval
A correct answer can still come from the wrong fragment, so the next question is where the relevant material sits in the collection and whether the process consulted it. Consider the invented early proposal that the service stay "open until 20:00" and the later qualification that this holds only if a second staff member is available. A chunk boundary separates the two statements. Neither chunk by itself licenses unconditional adoption: the proposal chunk yields a permissive instruction with no condition attached, while the qualification chunk yields a constraint whose object is missing. The error is not that either chunk is false, but that the retrieval step returns one of them and answers as though it were complete.
This is why position is worth treating as a variable rather than assuming that availability inside the collection settles the matter. The reported retrieval and question-answering tasks show that access to a long context does not ensure uniform use of its contents (Liu et al., 2023 preprint; TACL version noted). That evidence is historical, tied to the models and tasks of the study, so it cannot supply a current failure rate for the present model or a claim about all long documents. What it does supply is a reason to ask where relevant material was placed and what was retrieved from there.
A second case makes the same point without asking about retrieval at all. If a room had 12 seats in year A and 20 seats in year B, compressing this to "the room had 12 and 20 seats" drops the periods that make each figure true. No causal explanation for the change is supplied, and none should be invented. Position, together with the scope attached to a statement, is part of the evidence that any answer must retain.
The appropriate next step is therefore a prospective matched-position test: place identical relevant material at different locations in otherwise identical collections, hold the task fixed, and compare what is used. No such test has yet been run on a present model, and no historical effect size should be transported into it.
Position matters when the task requires a relation across positions, and the relation must survive whatever boundary the system draws. In the invented collection, an early proposal says the service is open until 20:00, and a later qualification permits this only if a second staff member is available. A chunk boundary separates the two statements. Retrieval that returns only the early proposal yields an unconditional policy; retrieval that returns only the qualification yields no usable policy at all. Neither chunk alone licenses unconditional adoption, so the retrieval unit for this task is the pair, and the boundary is an artifact of segmentation rather than a feature of the material.
The same problem appears in scope. The report that the room had 12 seats in year A and 20 seats in year B supports a period-indexed claim; the compressed form "the room had 12 and 20 seats" drops the periods and invites a false joint reading. No causal explanation for the change is supplied, so none may be inferred. Retaining scope markers is part of retaining the evidence, not an optional refinement after retrieval.
What remains to be shown is empirical, and the appropriate design is a prospective matched-position test: place identical relevant material at different locations in the collection while holding the task fixed, and vary position only. Any resulting comparison would estimate position sensitivity for the tested model under that task and that collection. No present model has been tested here, and no historical effect size may be transported into a current-model failure rate; prior long-context findings motivate the question rather than supply its answer. The design's limit is equally important: because it isolates position, it cannot establish that chunking always destroys needed relations or that retrieval always preserves them. Its conclusions would extend only to the tasks and materials actually matched, which is why the scope marker and the two-statement qualification should be among the relations deliberately tested.
23.3Local tasks within a global argument
Dividing a collection into chunks should not decide in advance what counts as an answer. The safer procedure is to give each local reader one precise question and to require an output of a form that can be checked against the text supplied to that reader. The question fixes the scope of the task; the output form fixes what would count as getting it right. Both are chosen before any reading begins, so that a later reviewer can compare the answer with its source instead of judging whether the answer merely sounds plausible.
Take the invented collection containing an early proposal to keep the reading room open until 20:00 and a later qualification permitting this only if a second staff member is available, with a chunk boundary between them. A question aimed at the first chunk might be: what is proposed, and for what hours? A checkable answer would restate the proposal and its time, and identify the chunk that contains it. What it would not do is infer that the hours are settled, because the qualifying condition lies elsewhere in the collection. The task is therefore successful as a local task without yet being a decision procedure for the collection as a whole.
Scope markers require the same discipline. In the time-scope case, the room had twelve seats in year A and twenty seats in year B. An answer that reports only "twelve and twenty seats" has omitted the periods that make the numbers compatible. A checkable output must carry each figure with its year and must not supply the causal explanation for the change, which the collection does not provide.
This gives a modest but real control: every local claim is traceable to a stated passage, including its conditions and its time references. The remaining question is what such a verified local answer can still leave out when the larger argument needs something that spans the chunks. That dependency must be identified explicitly rather than repaired by repeating the reading task.
A local reading task can return an answer that is fully correct against its assigned passage and still be inadequate to the overall argument. Suppose the task asks for the early proposal's opening hours, and the output "open until 20:00" is checked against that proposal and passes. The collection also contains a later qualification permitting this only if a second staff member is available. That qualification governs the scope of the proposal and is therefore a dependency of any claim about adopting the schedule. Without it, the locally correct answer supports unconditional adoption, which neither statement licenses. A local check does not see the omission, because the dependency lay outside the task's scope.
The dependency should therefore be named in the task record, not merely recovered later. What each route retains determines which dependencies are at risk. Direct context use keeps both statements available at once but permits a reader to work from the part of the context its question foregrounds. Retrieval returns material that matches the query and permits the qualification to be missed if nothing in the query touches staffing. Recursive summaries retain what the summarizing step treated as salient and permit a conditional to be compressed into an unqualified instruction. These are structural error profiles, not findings: no new long-context experiment has been run here.
Time scope shows the same pattern. The room had 12 seats in year A and 20 seats in year B. The flattened report "the room had 12 and 20 seats" keeps both numbers and drops the periods, and the collection supplies no causal explanation for the change. Numbers retained are not evidence retained when the condition that gives them their range disappears.
Identifying a dependency is not the same as recovering it once lost, nor as deciding where it ought to fall among a collection's divisions. A matched-position test, holding the task fixed while placing identical relevant material at different locations, has been proposed but not run, so placement effects are not settled by any result reported here.
23.4Boundaries between chunks
A chunk boundary can separate a proposal from the condition that limits it. In the invented collection, one chunk holds the early proposal, "open until 20:00"; a later chunk holds a qualification limiting the proposal to cases where a second staff member is available. Each chunk is readable alone, and the first yields a locally accurate account of the proposed hours. Yet neither chunk alone licenses unconditional adoption. The missing dependency is the relation between them: the proposal makes staffing a necessary condition; satisfying it does not by itself establish adoption or overall feasibility. The local transcription is fine; processing the chunks in isolation can lose the dependency the decision needs. A boundary does not itself destroy that relation if the reader or system recovers the cross-chunk link.
Dropping temporal scope produces the same kind of loss. The room had 12 seats in year A and 20 seats in year B. Compressing this to "the room had 12 and 20 seats" keeps both figures and loses the periods that make them compatible. No causal explanation for the change is supplied, so a reconstruction that adds one exceeds the evidence. Boundary damage here takes two forms: a qualification stranded on the far side of a cut, and a scope silently dissolved. Both are losses of relations rather than of the individual statements, which is why a summary can appear complete while omitting what the inquiry needs.
Whether a boundary harms an answer still depends on the task. If the question concerns only the proposed closing time, the first chunk may suffice; if it concerns whether the policy can be adopted, the condition must travel with the proposal. This dependence can be tested directly. To test the boundary-specific claim, a proposed comparison would hold the task and relevant material fixed while varying whether the proposal and its qualification are processed together or in separate chunks. Their overall document positions would need balancing or a separate control; moving text within a document alone would not isolate a chunk-boundary effect. That is a proposal for future work: no present model has been tested under it, and no historical effect size is transported to this setting. Pending such evidence, the boundary argument supports a claim about which relations must be recovered, not an estimate of how often chunking fails.
Suppose the boundary falls between "open until 20:00" and the later qualification that the proposal applies only if a second staff member is available. Each chunk then supports only a partial reading: the first suggests unconditional availability, while the second names a condition whose target is no longer visible. Neither chunk alone licenses adoption of the stated policy. Overlapping chunks could place both statements in one working context, restoring the joint reading in principle. But that repair does not by itself establish sufficiency. Overlap is a means of making the relation available to the task; whether the task uses it depends on where the relevant material falls and on what the task asks.
The boundary condition is therefore what deserves a test. A proposed matched-position experiment would place identical relevant material at different positions while holding the task fixed, then vary whether the qualification lies inside the overlapped region, at its edge, or beyond it. The stipulated success criterion is clear: when both statements are jointly available and correctly scoped, the answer should be conditional; when only one is visible, unconditional adoption counts as the predicted error. Position effects can then be measured directly rather than inferred from a long-context benchmark. No such model has been run here, and no historical effect size is transported to fill the gap.
The relevant limit is that proximity and scope are different properties. Overlap adds duplicated material and raises cost as the window widens, and even a shared window may compress the wrong relation. The separate seat-count case makes the point: "the room had 12 and 20 seats" states both figures but drops the periods in which each held. Restoring the numbers to one context does not restore their time indices, and the collection supplies no causal explanation for the change. Overlap is sufficient only when it preserves the qualifier or temporal scope the task needs, not merely when it brings the sentences together.
23.5Recursive summaries and source pointers
Chunk boundaries can separate an early proposal from the qualification that governs it. In the invented collection, one chunk contains the proposal to remain open until 20:00, while another contains the condition that this applies only if a second staff member is available. Neither chunk alone supports unconditional adoption. Summarizing the first chunk before seeing the second would therefore risk turning a conditional proposal into a firm policy.
An intermediate summary need not make that loss inevitable. It can report the proposal in conditional form and carry a pointer to the passage on which the wording depends: the proposal is to remain open until 20:00, subject to a later qualification located in another chunk. The pointer need not reproduce the missing text. Its work is to mark where the relevant evidence sits and to prevent the summary from claiming more authority than its own source has. Once the second chunk is processed, the summary can be revised to name the condition explicitly—only if a second staff member is available—while still pointing back to the statements from which both parts were drawn.
The useful inferential step is modest: preserving a pointer at each level lets later summaries see that a provisional claim has been qualified, even when the two statements were divided by the original chunking. The limit is equally modest. A pointer records where evidence sits; it does not adjudicate what the evidence requires. If a downstream conclusion depends on whether the staffing condition was satisfied, the relevant source passage must still be inspected. The recursive summary can make that obligation visible and route the reader toward the evidence, but it cannot decide the case in the evidence’s place. Thus a recursive summary can improve the record across chunk boundaries without removing the judgment that the divided statements demand.
Suppose the intermediate summary preserves the pointer as promised. For the opening-hours collection, it can record the proposed 20:00 closing time, note that a later qualification bears on it, and attach a link to the chunk containing that qualification. This keeps the qualification visible after compression, but it does not yet settle what the collection supports. Any conclusion that the location stays open until 20:00 depends on the force of the condition that the proposal applies only if a second staff member is available. The reader who draws that conclusion must follow the link and read the condition in place.
The inferential step is modest but decisive: carrying evidence forward preserves an obligation to consult it; it does not discharge that obligation. Recursion changes where attention begins, not where justification comes from. The later reader must still decide whether the second-staff condition is a contingency, a limit on the proposal's scope, or a qualification that blocks unconditional adoption. That judgment cannot be extracted from the summary's wording alone, because the summary was built to remain faithful to both statements without resolving their relationship.
The characteristic failure is therefore specific. A chain of summaries can be accurate at every level, and still invite unconditional adoption of the 20:00 proposal if the linked condition is treated as already satisfied or as immaterial. Keeping the pointer adjacent to any conclusion whose status turns on the qualification converts the chain into a set of prompts for inspection rather than an accumulation of authority. Whether the qualification is satisfied is a question about the collection, not about how many summarization steps preceded the question.
The economy of recursion has a clear limit, then. When nothing in the task turns on the linked material — when only the earlier proposal is asked about, for instance — the summary may stand on its own. When the conclusion turns on the qualification, pointer maintenance is part of the reasoning work, not a way of compressing it.
23.6Cross-boundary counterexamples
A summary can be accurate clause by clause and still mislead. Take the early proposal, "Open until 20:00," and the later qualification, which states that the proposal holds only if a second staff member is available. Each clause reports something in the record. Put them side by side in a summary and they read as one settled policy: the collection now supports adopting the 20:00 closing time with a staffing condition attached. Yet the record does not say that. The proposal came first; the qualification came later; a chunk boundary separates them. Neither chunk alone licenses unconditional adoption, and a combined sentence without their order hides the fact that the second statement qualifies the standing of the first rather than merely adding detail.
The same failure appears in the time-scope case. "The room had 12 and 20 seats" preserves both numbers but drops the periods: 12 seats in year A and 20 in year B. Nothing in the record explains why the change occurred.
This is where the three ways of using the collection part company. Direct context use retains both statements in order and lets a reader see the qualification as later; its risk is overlooking one clause. Retrieval may return the chunk containing the proposal while the qualification sits elsewhere, and then the error is structural: the returned evidence looks complete. A recursive summary keeps a pointer to the source, yet can still compress the pair into the misleading sentence above; the pointer is what a later reader needs to inspect when the conclusion depends on which statement came first.
The inferential limit is narrow. Correcting the summary restores the periods or narrows the claim to what each period supports. It does not supply a missing cross-period causal relation. The counterexample therefore shows what evidence a method retains and what combination it permits, not which method will always fail.
When the two periods are kept separate, the repair is precise. In the invented collection, an early proposal says the service is open until 20:00; a later qualification allows that schedule only if a second staff member is available. A chunk boundary separates the statements, so neither chunk alone licenses unconditional adoption. Reattaching the qualification narrows the inference to the condition actually supplied. Likewise, in the separate time-scope case, the room had 12 seats in year A and 20 in year B; once the periods are restored, the result is two dated statements. It is not a single undated fact, and it does not explain why the change occurred.
This is what a cross-boundary counterexample can establish: which scope markers and qualifications must survive whatever division of labor is used, whether the material is placed directly in context, retrieved on demand, or carried through recursive summaries. The useful evidence is therefore not simply the presence of both sentences, but the relation between them: the later qualification limits the earlier proposal, and the dated claims each hold only within their respective period. A summary may retain two clauses that are accurate separately and still mislead when combined without those distinctions. Adding a causal story about staffing, demand, or renovation would go beyond the collection.
The proposed matched-position test sharpens the question without answering it: identical relevant material could be placed at different locations while the task remains fixed. That is a future-study design only. No current model was tested, and no historical effect size is transported from it. Until such a test exists, the defensible conclusion is about repair scope. Preserve the condition and the period when rescuing a boundary; do not convert a missing causal relation into one merely because the relevant sentences sit in different chunks.
23.7A staged reading protocol
A staged protocol begins with local checks and widens only when they expose a gap. In the first stage, each chunk is read on its own terms: record the claim, its conditions, and its time scope before combining anything. The early proposal, “open until 20:00,” enters as a candidate policy, not as an unconditional instruction. The later qualification, which allows that schedule only if a second staff member is available, enters on the same footing. Neither chunk, read alone, licenses adoption.
The second stage is a cross-boundary argument check. Here the question is not whether each chunk is coherent, but whether the composed inference is justified. In the fictional collection, the later statement narrows the earlier one, so the combined reading is conditional rather than open-ended. The time-scope case works differently: “12 seats in year A” and “20 seats in year B” must retain their periods. Merging them as “the room had 12 and 20 seats” drops the distinction the argument depends on. Restoring the periods repairs the scope; it does not invent an explanation for the change.
The third stage returns selectively to the originals. When a boundary check reveals a dropped condition, a temporal mismatch, or an apparent contradiction, the relevant passages are reopened in context to establish order and wording. There is no need to reread the whole collection at every step; the return is triggered by the failure just detected.
The inferential step is that local adequacy does not settle compositional validity, so the boundary check carries the argumentative load, and source return is reserved for unresolved points. The limit is equally important: this is a proposed reading discipline, not a validated procedure, since no new long-context model experiment has been run. It cannot recover relations the collection does not contain—such as a causal account of the seat change—and it can only mitigate, not eliminate, the risk that artificial chunk boundaries hide the relationships an inquiry needs.
The staged protocol is justified by a comparison, not by an assumption that staging is always superior. Suppose the reader has a source short enough to hold at once and coherent enough that its parts remain visible together. Direct reading may then perform the same task more simply: the relevant relations are already in view, and no boundary work is needed. That alternative must be kept open; staging should win only when the source or task makes the added steps pay.
In the invented collection, the early proposal says the room is open until 20:00, while the later qualification allows this only if a second staff member is available. A chunk boundary separates the two. A local source check on the first chunk can verify the proposal exactly as stated, and a local check on the second can verify the condition; neither chunk alone licenses unconditional adoption of the hours. The inferential step is that the error to be caught is not a misreading inside either chunk but a relation lost at the boundary, and so a cross-boundary argument check must come before local findings are assembled into an answer. Selective return to the originals then settles whether the qualification applies to this proposal or to some other matter.
The limit matters as much as the check. Consider the time-scope case: the room had 12 seats in year A and 20 in year B. Reporting that it had 12 and 20 seats omits the periods, and no causal explanation for the change is supplied. Staging can recover the temporal qualification; it cannot manufacture a reason. Likewise, the matched-position test—identical relevant material at different locations with the task fixed—remains a proposal; no present model has been tested, and no historical effect size is transported to it.