Compression Has Levels

Chapters 14–16 decide whether information qualifies. A unit is task-relevant, personally novel, and matched to a standing subscription. The next question is independent of all three, and this chapter owns it alone:

At what resolution should qualifying information reach me first?

A replication confirming a trusted result needs one line. A contradiction of the reader’s architecture needs evidence, scope, and source before anyone acts. Same qualification machinery, radically different disclosure. A system with one fixed summary length fails both cases — drowning the first in detail, starving the second. The mechanism is a compression ladder: different disclosure resolutions over the same underlying information, expandable on demand, with NOTHING as a valid outcome.

Progressive derivation, not parallel summaries

The architecture first, because the obvious implementation is wrong:

AVOID — independent generation per level
source → tiny summary      (account A)
source → medium summary    (account B)
source → long summary      (account C)

Three fresh generations can quietly contradict one another — different qualifiers kept, different numbers rounded, different emphasis chosen. Instead:

source-linked information units
        ↓
DETAIL
        ↓
SUMMARY
        ↓
BRIEF
        ↓
GLANCE

Each upper level derives from the same stable unit set by omission, never by re-generation — with every level expandable downward toward DETAIL and ultimately SOURCE. The chapter’s governing rule:

A lower-resolution view should omit detail, not invent a different account.

The literature gives this rule teeth. Ou and Lapata (Findings ACL 2025, pp. 5534–5561, verified via ACL Anthology) show that recursive hierarchical merging of long documents amplifies hallucinations — each merge step re-generates from lossy intermediates — and that augmenting merges with source context (replacing, refining, or citation-aligning intermediates against the source) consistently beats plain merging on legal and narrative data. Independent per-level generation is hierarchical merging’s careless cousin: same re-generation risk without even the pretence of a shared spine. NexusSum (Kim & Kim, ACL 2025 long paper, pp. 10120–10157, verified) demonstrates the constructive side — hierarchical multi-agent processing with explicit output-length control reaching +30% BERTScore F1 on narrative summarisation — while staying fenced to its narrative setting, not this book’s technical-information problem. The licensed lesson in both directions: multi-level processing is practical machinery; levels must stay source-linked or they drift.

Resolution is also not word count. The ladder’s rungs are operating points defined by what the reader can do:

GLANCE    Can I decide whether to expand?
BRIEF     Can I understand the main informational change?
SUMMARY   Can I act on the main result and its key qualification?
DETAIL    Can I reason about the evidence?
SOURCE    Can I inspect the original material?

Length is a proxy, and an unreliable one: Retkowski and Waibel (Findings NAACL 2025, pp. 551–572, verified) show LLMs struggle with precise length control zero-shot across measures, with inherent biases — “give me 100 words” is not a dependable resolution mechanism by itself, and their approximation/adjustment/filtering/revision methods are the engineering answer if lengths must be hit. The ladder therefore specifies contracts, not counts. PACO (Ryu et al., ACL 2026 long paper, pp. 18578–18595, verified) treats summarisation as progressive refinement under multiple control attributes via planned sequential adjustment — cited here strictly as a mechanism precedent for progressive controllable refinement, not as evidence for an attention ladder. And the lowest-sufficient-resolution principle carries a purpose qualifier from this point forward: it governs decision sufficiency — scan, reference, triage, action. For learning sufficiency (retention, transfer, durable understanding), the representation permitting a correct immediate answer may destroy the generative, retrieval, or integrative work that builds it. The chapter therefore distinguishes a DECISION CONTRACT (what must remain for correct immediate use) from a LEARNING CONTRACT (what must remain or be actively reconstructed for retention/transfer goals): task = learn may require more context, explanation, or effortful reconstruction than the minimal decision-sufficient rung — semantic preservation is not learning preservation, and EXP-17/18 scope their primary findings to decision tasks until a learning-goal condition exists.

NOTHING is a gate, not a level

NOTHING needs precision, because aggressive compression is not the same as silence:

NOTHING means: this candidate does not currently justify consuming any attention under the active policy.

Not zero text — a decision. It therefore sits before the ladder, as an attention-worthiness gate:

qualified candidate
       ↓
attention-worthiness gate
       ├── NOTHING
       └── resolution ladder

A subscription can match something technically relevant, genuinely novel, and correctly related that is still too low-value to consume — the third near-identical restatement of a known update. Without the gate, every qualification becomes an alert and the system becomes an alarm that readers learn to ignore — destroying the subscription machinery’s credibility to save its completeness. The gate’s refusal rate is logged like the Chapter-10 preservation gate’s: a gate that never refuses is decorative.

Information contracts per level

Each rung carries an explicit contract — what it must preserve — making the ladder measurable and giving Chapter 18 its audit surface:

GLANCE must preserve:
- identity of change · relation type · direction
- why it might matter

BRIEF adds:
- principal claim · key qualifier · evidence pointer

SUMMARY adds:
- important supporting units · caveats
- uncertainty · comparison context

DETAIL adds:
- evidence structure · numbers
- exceptions · provenance

SOURCE:
- original artifact

A GLANCE that omits direction (“X improved” — over what? in which direction?) violates its contract regardless of brevity. A BRIEF without the key qualifier is a Chapter-4 false-skip engine: confident, compact, wrong. Contracts convert “too short” from an aesthetic complaint into a checkable violation — and optimal resolution then depends on omission risk, not volume: the glance-sufficient replication, the glance-dangerous subgroup-qualified improvement, the nothing-correct restatement, the expansion-mandatory apparent contradiction all resolve differently at identical lengths.

Expansion without re-search

Architecturally, expansion must not re-discover. The reader’s path:

GLANCE → expand → BRIEF → expand → SUMMARY
→ expand → DETAIL → inspect → SOURCE

must traverse one stable object, because “tell me more” triggering fresh retrieval can subtly change the answer — different candidates, shifted emphasis, a contradiction quietly dropped. The discovered candidate therefore carries its lineage as a resolution artifact: candidate identity, source units, subscription match, task context, novelty state, current level, omitted units, source links. Same candidate at every resolution; omission lists making each level’s debts explicit. Omitted-units bookkeeping is the quiet innovation: a BRIEF knows what it left out, so expansion restores rather than re-invents.

The comparison: independent versus progressive

EXP-17 freezes the Chapters 13–16 stack and compares: A one fixed summary; B independently generated glance/brief/summary/detail; C progressive ladder derived from one stable unit set; D ladder plus preservation constraints and source-linked expansion. Measures: cross-level consistency, critical-unit and qualifier/relation survival per level, unsupported additions, expansion consistency (does expanding reproduce the lower level’s content?), decision accuracy achievable at each level, attention consumed before correct action, unnecessary expansions, false non-expansions. The headline metric is attention-instrumental: at what lowest resolution can the user make the correct decision? — compression as attention control, not writing exercise. Failure criteria: B ties C on consistency (derivation adds nothing); C’s glances violate contracts at fixed-brief rates (contracts decorative); users systematically over- or under-expand (resolution cues miscalibrated); D’s constraints cost more attention than they save (preservation priced out). Artifacts: frozen stack, ladder artifacts with omission lists, per-level contract audits, resolution-decision curves. What a positive result would not justify: how badly each level distorts — distortion measurement stays Chapter 18’s job, per the fenced boundary: 17 builds progressive compression and expansion; 18 measures what each reduction destroyed and where the ladder must stop.

What this chapter earned

Qualifying information need not arrive at one fixed resolution: a progressive, source-linked ladder (DETAIL→SUMMARY→BRIEF→GLANCE by derivation, never parallel generation) exposes the minimum detail for the current decision, with per-level information contracts, expansion-without-research via lineage artifacts, a NOTHING gate before the ladder, and resolution chosen by omission risk rather than volume. Distortion accounting remains open — deliberately, since measuring it is the next chapter’s entire job.

The result should not be one irreversible summary; readers need multiple levels of resolution — and now the levels exist. What did each reduction destroy?

References

• Ou, L. & Lapata, M. (2025). Context-Aware Hierarchical Merging for Long Document Summarization. Findings ACL 2025, pp. 5534–5561. DOI 10.18653/v1/2025.findings-acl.289. Verified via Anthology. Used: recursive merging amplifies hallucination; source-context augmentation beats plain merging. Licensed as the anti-parallel-generation evidence.
• Kim, H. & Kim, B.-H. (2025). NexusSum. Proc. ACL 2025 (long), pp. 10120–10157. DOI 10.18653/v1/2025.acl-long.500. Verified. Used: hierarchical multi-level processing + explicit length control as practical machinery. Fenced: narrative setting.
• Retkowski, F. & Waibel, A. (2025). Zero-Shot Strategies for Length-Controllable Summarization. Findings NAACL 2025, pp. 551–572. DOI 10.18653/v1/2025.findings-naacl.34. Verified. Used: length control is unreliable zero-shot — resolution by contract, not count.
• Ryu, S. et al. (2026). PACO. Proc. ACL 2026 (long), pp. 18578–18595. DOI 10.18653/v1/2026.acl-long.845. Verified. Used strictly as mechanism precedent for progressive multi-attribute refinement.

Proposed experiment EXP-17: independent vs progressive compression

Status: PROPOSED. Per the design above (A–D, frozen stack, contract audits, resolution-decision curves, asymmetric risk cases: glance-sufficient / glance-dangerous / nothing-correct / expansion-mandatory).