# Compression Has Levels
# Book: Language
# Chapter URL: https://programmer.ie/books/language/19-chapter/
# Prose fingerprint (sha256, normalized): 4b386ccea52e5894e8e3f1bf6451a66ce2774ab4708df04161717a874ed2ba3c
# Snapshot generated at build time by Hugo. This is the complete rendered prose
# of the chapter; it is not a transcript of the source file.

At three in the morning a candidate enters Iris’s semantic-subscription queue. This is a constructed example, not a live result. A new paper from a database group may bear on the timing assumption used in her team’s protocol. The watch’s declared source and publication conditions have been met, and the system has proposed a relationship to Iris’s versioned anchor. But the relationship is still possible, not established: a result for another protocol does not by itself weaken her protocol’s safety argument. The system has not yet decided when, or even whether, to interrupt her. First it must decide how much evidence the candidate’s initial representation needs to carry.
A single fixed summary length is unlikely to serve every qualifying item. A well-established replication may need no more than an accurate identification and a route to the source. A possible challenge to a design assumption needs its scope, evidence status and unanswered questions visible even in a small preview. Whether Iris sees either at three in the morning or after breakfast is a separate delivery decision, reserved for the next chapter. The question here concerns the resolution of a candidate once it is presented, not its notification schedule.
Once an item deserves attention, how much of it should reach the person first, and where must compression stop?
The two halves of that question belong together: making an account smaller changes what a reader can recover from it. The goal is not the shortest fluent text, but the least demanding representation that meets a declared purpose without turning uncertainty into a conclusion.
A ladder of resolutions One useful design is a ladder of views, with progressively more source-linked detail. The five names below are proposed interface states, not a demonstrated psychological scale. The reader may skip levels or open the original directly; the order describes increasing accessible evidence, not a compulsory journey.
Proposed resolution ladder. The reader may open the source directly; unresolved obligations can require more detail or unchanged-source access. Open full-size diagram. These are resolutions, not word counts. What defines each is what the reader can do with it:
GLANCE can I decide whether to inspect this candidate? BRIEF can I identify the reported change, its scope and standing? SUMMARY can I decide what needs checking before I rely on it? DETAIL can I inspect the argument, methods and evidence gaps? SOURCE can I examine the unchanged original and its version? compression level ≠ character count A glance might be one sentence for one item and a labelled uncertainty plus a link for another. Its adequacy depends on the decision it supports, not its character count. The interaction pattern is familiar: Shneiderman’s information-seeking framework distinguished overviews from details on demand, and Nielsen’s account of progressive disclosure emphasised showing needed information first while making further detail easy to find. Neither source establishes that these five levels improve comprehension or preserve AI-generated claims. That is the specific hypothesis here, and it needs testing. A request for a hundred words is an output-length preference, not a preservation specification.
One item, five ways Suppose the hypothetical paper describes a commit protocol whose authors claim safety in an asynchronous, crash-fault model: messages can be delayed arbitrarily, but are not assumed to be maliciously altered, and the claimed safety property does not depend on a known delivery bound. Its progress or termination argument needs additional eventual-timing and availability assumptions. It reports experiments on five- and eleven-node clusters, which illustrate performance rather than prove safety. A key lemma appears only as a proof sketch. The paper has passed the peer-review condition in Iris’s subscription, but its claims have not been independently verified here. Most importantly, the paper studies a different protocol. It may give Iris a reason to re-examine an assumption; it has not refuted that assumption for her design. Here is one candidate set of views, written by hand to demonstrate what different resolutions owe.
GLANCE Constructed Paper P (v1): possible relevance to your timing assumption. Different protocol; applicability unverified. [Inspect evidence] BRIEF A different commit protocol claims safety with arbitrary message delays under crash-fault assumptions. Its progress claim has separate timing conditions. It does not evaluate or refute your protocol. Standing: peer-reviewed; applicability and key lemma unchecked. SUMMARY The authors claim safety for their own commit protocol despite arbitrary message delays, subject to their stated failure model. Progress requires further timing and availability assumptions. Experiments on five- and eleven-node clusters assess behaviour, not proof correctness. A central lemma is only sketched. No demonstrated bridge carries this result to Iris's protocol, so the correct next step is to inspect the assumptions and proof, not to change her design. DETAIL Claim-by-claim source pointers; crash-fault and timing assumptions; the stated safety property versus progress conditions; the exact status of the sketched lemma; measured experimental conditions; and a side-by-side comparison with Iris's versioned specification. Unresolved differences remain explicitly unresolved. SOURCE The specific paper version, unchanged, with stable citation and a route to the original passages. Reading from SOURCE toward GLANCE, details disappear; reading toward SOURCE, evidence becomes available. At each step we can ask whether a shortened claim remains faithful at that resolution. These are invented views of an invented paper, not measured summaries or outputs from the reference program. The preservation audit of Part I supplies the method for judging them.
What each level owes Each proposed level has a different communicative job. Its preservation obligations should be declared before rendering, and richer levels should carry forward the obligations attached to claims they retain. That does not mean every deeper view must repeat every word of every shorter one: it may expose evidence, qualifications and alternative explanations that a short view could not fit. At GLANCE, Iris needs little, but what appears must not misdescribe the candidate.
GLANCE must preserve: - which source and version the candidate comes from - that the paper concerns another protocol under stated assumptions - that its relation to Iris's timing assumption is possible, not established - that there is a direct route to the evidence A glance saying “your assumption is wrong” would turn a possible connection into a verdict. One saying merely “new paper on databases” would not explain why the subscription produced it. A brief must identify the reported result and its limitations. A summary should support an informed decision about what to verify next, not authorise a change to production code. Detail should expose the argument and its gaps; only the unchanged source permits full independent inspection. The relation’s standing, a consequential qualification and an unresolved applicability question must not be hidden behind a click when a visible claim depends on them.
The property that follows is a design expectation with a limit. When each view is faithful, greater resolution should supply supporting detail rather than silently changing the claim. But if the system discovers that a glance was wrong, the deeper view must correct it. Consistency is not worth preserving an error.
higher resolution should refine supported claims; corrections must be explicit A reader opening a higher-resolution view should encounter the same candidate and the same versioned source, with more of its evidence available. A new qualifier should refine a claim, not quietly turn “possible” into “proved” or “safety” into “liveness”. If the short level really was wrong, mark it corrected and show which source passage changed the interpretation; do not conceal the disagreement in the interest of monotonicity. To reduce drift, derive levels from one inspected, source-linked unit inventory as proposed in Chapters 7 and 14, rather than independent fresh interpretations or recursive paraphrases. A shared inventory reduces inconsistent generations but can also propagate a shared extraction error, so each delivered view still needs a source-level check. Claim continuity matters more than literal nesting of displayed sentences.
Omit, add, distort Compression leaves information out, but it must not silently strengthen, contradict or invent what remains. Three actions require separate handling:
OMISSION allowed only outside this level's declared obligations ADDITION separately labelled and independently sourced if retained DISTORTION a failure, however readable the shorter view Leaving out a sample size may be reasonable for a glance whose only job is triage, provided the visible claim does not imply that the result is universal. If a retained statement says “safe”, its failure model and relevant scope must accompany it at whatever resolution it appears. Material added from outside the paper must be labelled as external evidence or a model proposal, rather than smuggled into the source account. And a cautious paraphrase can still be a distortion: changing “we show under assumption A” to “A is probably unnecessary” is not an innocuous simplification. The test concerns the interpretation a reader can reasonably reconstruct, not the presence of the right vocabulary.
The audit distinguishes acceptable omission, unmet obligation and misleading transformation. Acceptable omission removes detail that this level never promised and that no retained claim requires. An unmet obligation leaves out something the level promised, such as the unresolved relation to Iris’s own protocol. A misleading transformation retains a proposition but changes its scope, relation, quantity or certainty; it fails even if the required words appear somewhere else. An assessor may also have to return unresolved: if the source cannot be retrieved or a claim cannot be reliably checked, that is not a pass. These results belong in a visible profile, not a single confidence score.
Where compression must stop This is the chapter’s decision rule. The aim is to choose the smallest task-sufficient view, not to hit a token quota. Requirements are frozen for the task and source version before the system tries to shorten an account. The check asks what remains recoverable to the reader, not merely whether key phrases survived.
declare task and source-version obligations ↓ propose a smaller representation ↓ check retained claims, dependencies, uncertainty and provenance ↓ if any required condition fails or cannot be assessed: restore detail or use the unchanged source ↓ show the smallest view with sufficient evidence of compliance The stop rule cannot be a perfect oracle. The system may not know what the reader will misunderstand, and its checker may share the generator’s blind spots. Practical obligations include the proposition being asserted, relation direction, numbers, scope, failure model, uncertainty, source identity and dependencies needed to interpret whatever remains visible. An independently reviewed obligation set and held-out failure tests can reveal some missed losses; they do not prove that none remain. A purportedly compliant glance therefore does not certify that an engineer can act on it. When the check fails or cannot be validated, the system must raise the level of detail or return the original, without describing that fallback as an automatically safe reading.
Sometimes almost nothing can be compressed honestly. A disputed negative result or a claim whose critical qualifications fill most of the passage may admit only a minimal notice that a candidate exists, plus an immediately available original. Such a notice must refrain from making the very substantive claim the missing qualifications would govern. Skipping directly to the source is a valid abstention from summarisation, not evidence that the source itself is true or easy to understand. The system must also distinguish “no safe summary found” from “the source was inaccessible” and record which happened.
The competing costs are not uniform across tasks. Extra detail can consume attention, bury a result and deter opening a useful source. Too little can produce false confidence, a bad decision or a missed contradiction. The cost of either error depends on what the reader is doing, so evaluation should report excess reading effort and harmful omissions separately. When the consequence is serious and the evidence for short-form fidelity is weak, more detail, an explicit unresolved label or abstention is appropriate; there is no universal exchange rate between seconds saved and claims lost.
Every level is an entrance to the one below Compression must not make the source inaccessible. A small view can provide orientation only if the reader can examine the claim and its provenance without having to trust a newly generated explanation. The originating document, its version, and any access restrictions should remain stable when the view is expanded.
compression ≠ replacement Expansion should be easy to find, easy to reverse and anchored to the same item and source version. It should normally reveal more of the same selected evidence rather than silently run a fresh search or produce a different interpretation. The reader should also be able to jump straight to the cited passage, since a compulsory five-step drill-down can cost more attention than it saves. The familiar design principle is progressive disclosure, not progressive authority: the short view is still a derived account.
glance → brief → summary → detail → source A useful implementation would track which source-grounded units, obligations and dependencies are visible at each level, and which were deliberately omitted. An expansion could expose the omitted units with their source spans, plus a record of any correction made between views. Merely carrying an ID list is not enough: the actual text and its implied relation must be checked, including qualifiers attached to claims. An external source brought in while expanding would need its own origin label. These are design requirements; the reference example below does not yet implement reversible, evidence-preserving expansion.
More detail is not better It would be easy to read this as a ranking from bad to good, with SOURCE as the winner. It is not. More accessible detail offers stronger opportunities for inspection, not guaranteed comprehension, and the unchanged source may itself be ambiguous or incorrect.
more information ≠ better representation For a reader deciding whether a paper merits five minutes, a faithful glance can outperform immediate exposure to the whole document on speed and comparison. For checking whether its proof is sound, that glance cannot do the job. For learning a difficult idea, premature summarisation may also remove the reasoning work needed for durable understanding, although that outcome must be measured rather than assumed. The system should evaluate selection, comprehension, delayed recall and application separately, as the earlier chapters argued. The reader’s ability to request the source is necessary for auditability but is not evidence that they did so, or that their eventual conclusion was correct.
Resolution is a decision the channel makes Selecting an initial resolution is another bounded decision of the communication layer. It depends on the declared task, the evidence that must accompany the candidate, permitted knowledge-aware adaptation, the time budget, and the possible cost of a misleading conclusion. It need not depend on a speculative model of Iris’s emotions or identity. Nor does choosing a brief imply delivering it immediately. A system can prepare a faithful detail view for later inspection without interrupting a reader who is asleep. The next chapter determines when candidates may enter her attention; this chapter determines how much an offered representation must contain.
How we would test the resolution ladder The chapter proposes a five-state interface and a preservation-aware stop rule; it does not yet establish that either is better than simpler alternatives. A fair experiment should hold the candidate set, source versions, declared purpose and access rights fixed. Prepare source-grounded units and preservation obligations independently of any generated summary, including difficult cases: a result with a distant qualifying assumption, a source that disputes its own headline, a non-significant finding promoted to certainty, a stale version, and the possible relation between two different protocols. Reviewers should be allowed to mark obligations as disputed or unresolved. No test condition should gain privileged access to evidence denied to the others.
Compare at least four interfaces: titles and source links; a strong human-authored or extractive brief with direct passage access; a one-shot AI summary at a fixed length; and the proposed progressive, obligation-gated ladder. Include an ablation without the preservation gate, so that any improvement can be attributed to the check rather than the layout alone. For the constructed protocol paper, ask people separately whether they would open it, what the paper actually claims, what its assumptions are, and whether they have grounds to change Iris’s system. The last answer should remain unresolved until the cross-protocol connection is established.
Primary failure measures are misleading retained claims, missed conditions, unsupported decisions and false confidence. Measure triage recall, appropriate source opening, time to a correct answer, unnecessary expansions, latency, reading effort and preference separately. Record which views the checker rejected and which it wrongly passed, using held-out alterations of relation direction, scope, negation and numerical conditions. Also measure whether the five-level navigation costs more than a simpler two-level design. Success requires better reader outcomes at acceptable cost, not merely a higher click-through rate or fewer words. This is a proposed experiment, not a result of the implementation below.
Running it Running it. The existing npm run example:radar demonstration uses five hand-written levels. It checks four properties separately: whether a level names units selected by the compiler; whether its unit membership is nested as expected by that implementation; whether declared unit dependencies accompany the units; and whether its text satisfies a conservative, word-anchored claim-preservation check. In a constructed brief, naming both a claim and its qualifier while saying only “safe” passes the first three and fails the fourth; the program then presents the summary. A brief naming a refused unit fails membership and falls back to the source. An item with no admitted units is not summarised. These are observed behaviours of a scripted example, not proof that the reported checks are semantically adequate.
What it does not show: generation of reliable resolution levels from a shared source record, preservation-aware interactive expansion, or reliable comprehension by a reader. The unit-nesting rule is an implementation constraint, not a demonstration that claims are monotonically refined. The fourth check can reject faithful paraphrases and still accept copied words arranged around an incorrect meaning. The program does not validate the checker against independent human judgements, assess the fictitious paper’s proof, compare strong baselines or exercise live source-version changes. Its fallback means a generated view was refused by its stated rules; it does not make the original source true or its presentation sufficient.
Where this leaves us A qualifying item can be offered at five proposed resolutions, distinguished by what each permits the reader to inspect, not by its word count. Every level owes the claims it retains their correct scope, uncertainty, relations and provenance. Omission is acceptable only when it does not defeat that level’s obligations; new claims need distinct support and misleading transformations are failures. Richer views should supply evidence from the same source version, and corrections to earlier views must be visible rather than hidden for the sake of consistency. The smallest adequate view is an empirical question, with abstention and the unchanged source legitimate outcomes. Five levels are a design choice awaiting comparison with simpler progressive disclosure.
Now suppose ten candidates qualify overnight. Each could be represented faithfully at an appropriate initial resolution, and the reader might still have time to inspect only three. A compact account does not decide which candidate matters first, and it certainly does not authorise an interruption.
relevance ↓ qualification ↓ resolution ↓ still too many candidates Choosing the level for each item does not select the next item to consume, and selecting an item does not settle when it should interrupt. The remaining question is temporal: which candidates need attention now, which may wait for a natural break or scheduled digest, and which should remain available without competing for the reader’s focus?
References Shneiderman, B. (1996). The eyes have it: A task by data type taxonomy for information visualizations. Proceedings of the IEEE Symposium on Visual Languages, 336–343. DOI 10.1109/VL.1996.545307. (An influential framework for overview, zoom/filter and details on demand; not evidence that this chapter’s five generated levels are effective.) Nielsen, J. (2006). Progressive Disclosure. Nielsen Norman Group, 3 December. https://www.nngroup.com/articles/progressive-disclosure/ (Interface design guidance on initial information, expansion and usability; not a preservation evaluation.)