# What Survived the Transformation?
# Book: Language
# Chapter URL: https://programmer.ie/books/language/11-chapter/
# Prose fingerprint (sha256, normalized): a3eee39be9d928b4bb45fbb0b914a6167c1da792a66f8a30f331accb7315cc02
# 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.

Someone asks a system for a one-line version of a payments retry policy. The source, constructed for this chapter, says: “For non-idempotent payment requests, allow at most three attempts in total. Idempotent requests follow the separate recovery path. Handling of third-party timeouts remains unresolved; consult design record D-17 before changing this policy.” The system answers: “Failed payments are retried up to three times.”
The sentence sounds concise and professional. But three retries are not three total attempts, the limit is not universal, and an unresolved timeout has disappeared. The sentence also loses the pointer to the decision that explains why the rule exists. A reader relying on it to change payment handling could make the wrong change. A safer one-line preview would say: “Retry rules vary by request type; timeout handling remains unresolved.” Even that is only a preview, not permission to change the policy. The surface gives no warning. The unfaithful sentence looks much like a faithful one.
The book has now spent several chapters transforming information. Previews compress pages so that a reader can choose what to open. Enhancements re-express passages as diagrams and tables. A semantic record captures one interpretation so that it can be reused across views. Sidecars bring in material from elsewhere, and relation labels say how one item bears on another. These are different operations, but each can quietly alter what the recipient is able to recover. A fluent result does not disclose which condition vanished or which relationship changed. This chapter gives that shared problem an audit method:
When a representation is transformed, how do we decide whether the transformation preserved what mattered?
The book has circled this question since Chapter 1: what had to survive? The answer cannot be read from the generated words alone. It depends on the source, the purpose of the transformation, and what a receiver must still be able to distinguish. Here the question becomes an operating method, with a declared obligation, a reported outcome, and an explicitly limited check.
Four outcomes When a source becomes another representation, there are four kinds of outcome worth recording. They are four questions, not four mutually exclusive labels for the whole output. A single sentence may preserve one claim, lose its condition, introduce an unsupported generalisation, and make another relationship easier to see.
Conceptual outcome profile. An attractive view can still lose a required qualification; plausible gains remain unmeasured until evaluated. Open full-size diagram. Survival and loss ask whether required content remained recoverable. Addition asks whether the output asserts something the named source does not support. An openly labelled explanation from another source is different from a claim silently presented as if it came from the source. Gain asks whether a reader’s task became easier or better. Inspecting the output cannot establish this. These distinctions matter because a transformation may improve presentation and still fail a non-negotiable preservation requirement.
The distinction that matters most is between the last two:
GAIN ≠ ADDITION A diagram may expose a dependency that was hard to find in prose. A preview may help a reader judge whether to open a page. A table may make a comparison quicker without changing its values. These are plausible gains, but they are not established until a receiver’s decisions or work have been measured against a suitable alternative. Addition is a different question: has the rendering supplied a relationship, fact, scope or confidence level that the source does not warrant? New external information can be useful, but it must be labelled and supported on its own terms, as the sidecar chapter required. A view is not faithful merely because the added material happens to be true.
A particularly deceptive failure is a change of force. “May” becomes “is”. “In subgroup Z” detaches from a comparison, and the conclusion becomes universal. “A is associated with B” becomes “A causes B”. This can count as both loss and addition. The source’s original qualification is no longer recoverable, while the output asserts a stronger claim. A checker that finds the original nouns but misses the altered relationship will declare the wrong thing preserved.
Declaring what has to survive The obligation must be fixed before the output is judged. Otherwise a convenient result can redefine success. “Before” does not mean that a model generating a view should quietly invent its own obligations and then grade itself against them. In an experiment, the task and the source-supported requirements need to be established independently of the candidate output, and any disagreements or uncertainties recorded. A production system can propose obligations, but a proposal is not an authorisation.
BEFORE declare what must survive ↓ TRANSFORM ↓ AFTER check: what survived? what was lost? what was added? what became easier? The declaration can be written in ordinary language, but it must be precise enough to challenge. For the constructed retry policy, a reviewer planning a change might write:
must preserve: - the limit of three total attempts for non-idempotent requests - the separate recovery path for idempotent requests - that handling of third-party timeouts remains unresolved - the design-decision reference for the policy This is a preservation obligation, not a miniature summary of the policy. Its items identify what cannot disappear for this task: the attempt limit, the separate idempotent path, the unresolved timeout handling, and the policy’s provenance. For a quick topic preview, the minimum might be smaller. For changing production retry behaviour, it might need to include the complete recovery procedure and the current policy version. The person or policy setting the task must choose that scope. No reader should be told that an abbreviated view is sufficient for a decision it was never designed to support.
The declaration can be wrong. It may omit a decisive exception, misread a condition, or inherit an error from the semantic record. That is why each obligation needs its own source anchor, version and reviewer or policy origin. It is also why a later correction should produce a new declared version rather than silently rewriting the old test. Freezing an incomplete obligation gives a repeatable experiment, but it does not make the experiment complete.
Comparing meaning for a task The Understanding Compiler makes three assessment targets explicit: source-content coverage, meaning carried by the transformed representation, and recipient application of the relevant meaning. A source inventory can be incomplete before any rendering is made. A faithful inventory can still lead to a misleading message. A faithful message can still be applied incorrectly. Matching inventories at two boundaries does not establish any of these targets independently, because the same extractor can repeat its own mistake.
In a hypothetical, hand-authored analysis of the retry example, “up to three retries” changes a limit of three total attempts. “Failed payments” widens a rule that applies only to non-idempotent requests. Silence about third-party timeouts removes an unresolved issue. Losing D-17 breaks the path to the design decision. The separate recovery path for idempotent requests must stay distinct. These are candidate semantic findings against the quoted source. They are not new compiler measurements, nor revisions to historical experiment fixtures.
A comparison should distinguish omission from unsupported addition, scope change, negation inversion, authority change, temporal change, loss of uncertainty and broken relations. Turning a proposed change into an approved instruction changes its authority, even if its operational content is unchanged. Moving a deadline changes the time. Keeping both propositions while detaching their exception breaks a relation. Each finding needs its source evidence, its task consequence and its assessment standing.
Omission also needs a purpose. A reader asked only whether timeout handling is settled can correctly answer “No; consult D-17” without repeating the attempt count. That answer does not show knowledge of the full recovery policy, but its brevity is not evidence of failure on this question. The proposed TASK_IRRELEVANT and NOT_ASSESSED categories in Chapter 8 keep appropriate omission apart from unchecked application. They supplement the comparison vocabulary here. They do not rename implemented report states or weaken an accepted preservation obligation. When a task changes, its obligations should be reviewed explicitly rather than quietly waived.
Content is not function Content survival and task success are separate questions. Each needs a different observation, and neither can stand in for the other.
CONTENT did the same facts and relations survive? FUNCTION can the recipient still do the thing the representation was meant to support? A transformation can keep every fact and still make a task harder. Reorder a procedure so that its steps are all present but cannot be followed in sequence, and the content is intact while the reader’s performance suffers. Put the decisive value in an awkward table and the comparison becomes slower. The converse also matters. A short view may leave out background detail that is irrelevant to a declared task while preserving every decision-critical condition. That omission is acceptable only within the stated task scope. It does not make the short view a safe substitute for the full source in every other use.
So the question is never only whether the words match:
textual sameness ≠ communicative preservation This is the reconstruction idea from earlier chapters, turned into two measurements. A source-to-output comparison asks what information the view warrants. A reader study asks what a person actually recovers and does with the view. One can pass while the other fails. Reading time, correct decisions, delayed recall and transfer to a new case are also different outcomes. Measuring faster answers cannot establish that the answers were sound.
A readable trace of what? Suppose a mediator supplies a readable account of its own internal computation. What would have to survive that transformation? The claims in the original document are one obligation. An accurate account of how the model reached its answer is another. A rendering could preserve the document’s retry limit while inventing a plausible story about why the model selected it.
Zou and colleagues’ Hybrid Reasoning Agents (2026, Section 3.1 and Appendix F) provides a concrete case. Their model alternates between text and continuous internal computation, and maps its latent inputs to nearby token embeddings to produce pseudo-text. Their readability measure is geometric alignment, and their context-specificity measure concerns lexical overlap with the originating question. These are interesting inspection signals, but neither tests whether a condition such as “only after verification” survived for a recipient. The authors also intervene on latent states and measure changes in mathematical performance (Section 6.2). That supplies evidence of functional reliance on those states. It does not establish that every phrase in their readout faithfully describes its causal role.
An intelligible rendering of a computation need not be a faithful account of that computation. The preservation contract therefore needs to name its target. Can an independent recipient recover the required condition from the readout? Does changing that condition in the source change the reconstructed decision? Those questions test whether recoverable distinctions survive. Establishing that the rendered explanation identifies what caused the model’s decision requires further controlled interventions. A fluent explanation, whether produced from text or from latent states, cannot answer all three questions by itself.
A profile, not a score A number such as “fidelity 0.84” can summarise a chosen test, but it cannot tell a reader what the missing 0.16 was. One lost qualifier may matter more than ten preserved headings. Nor do transformations share a single failure pattern. A diagram can reverse a relation, a preview can hide a decisive item, and a summary can turn uncertainty into certainty. The chapter therefore uses a profile rather than a single universal verdict. An aggregate may still help compare systems on a fixed, declared evaluation set. But a gate for a particular task must keep the individual obligations and their severity.
The working profile covers claims, relation type and direction, attached conditions, quantities and units, uncertainty, provenance, unsupported additions and changed emphasis. For each declared requirement, the report needs a state such as PRESERVED, LOST or UNRESOLVED, along with the source span, the output span and the check used. UNRESOLVED is not a pass and not proof of loss. It records that the instrument could not decide. Requirements the representation was never asked to carry must not quietly be counted as successes. The report should also say which dimensions, including reader benefit, have not been measured.
Relations are first-class. A representation can keep every noun and still destroy the assertion. “A causes B only when C is present” becomes “A causes B”. The entities and the causal verb remain, but the condition governing the claim is gone. “A acquired B” becomes “B acquired A”. Nearly every word survives while the direction reverses. A preservation check must tie the relation to its arguments, its direction and its qualifiers. Finding those words somewhere on the page is not the same as finding the original proposition.
Emphasis can distort without a false sentence. Suppose a paper spends most of its space on method and limitations, but a new visual puts the speculative result in its largest panel. Every printed claim might be literally supported, while the layout suggests that speculation is the main result. That is a hypothesis about how readers will interpret the view, not something page area alone can prove. The profile should record changed prominence, and a reader test should ask what people take to be the central finding. Selection, order and layout can change communication without introducing a new sentence.
Traceability is separate from preservation. An output element should carry an address into a specific source version, or be clearly marked as an unsupported or externally sourced addition. An exact quotation, a plausible-looking citation and a retrieved passage do not by themselves validate the interpretation attached to them. Traceability makes challenge and correction possible. It neither proves fidelity nor guarantees that a mistaken representation is safe to act on. A source link can also go stale, so the record must say which version was actually checked.
Preservation is not truth, and truth is not authorisation. If the source makes a false assertion and the view reproduces it exactly, source fidelity may be high while factual reliability remains poor. If the source is correct, a faithful view can still be unsuitable for a high-stakes action. The audit must therefore keep five things separate: what the source said; what the transformation preserved; what external evidence supports; what the recipient was able to do; and what the system is permitted to act upon. This chapter specifies the preservation comparison, not all of the other checks.
The check is an instrument too There is a trap here. Once obligations are written down, a passing check starts to look like proof. But the check is another imperfect instrument. A phrase match may find every required word in a sentence that reverses the argument. A source span may exist while the attached claim misrepresents it. In the reference implementation, obligations are visible to the model, and an example shows how a list of required words can pass without preserving the policy. This is not evidence that preservation was achieved. It is a demonstration of how an instrument can be satisfied for the wrong reason. Separating the generator from the checker can reduce one source of circularity, but the checker still needs validation against cases that it did not select or construct to flatter itself.
To test the instrument, begin with source-output pairs whose important differences have been independently adjudicated. Remove a condition, reverse an acquisition, change a unit, invent a citation, or turn uncertainty into certainty. Then include faithful paraphrases and harmless layout changes that should not trigger a failure. Report both misses and false alarms, by error type, on cases held out from checker development. A probe that notices two known defects has earned exactly that result, not general reliability. The observed disagreement between checker and human judgement should remain in the record. An untested check can assist review, but its pass cannot authorise a consequential transformation.
A preservation experiment to run The book’s architecture makes a testable prediction. Declaring obligations before generation and then checking the resulting representation should reduce task-critical losses without eliminating the benefit of a better view. This is a proposed experiment, not a reported finding. Freeze a set of source passages and tasks, including the retry policy. Prepare the obligations independently, and collect held-out judgements about what each source actually states. Include both easy paraphrases and hard cases involving scope, negation, numerical changes, chronology and conflicting evidence.
Compare the unchanged source, a strong manually prepared or extractive baseline, direct generation, and generation with declared obligations and a preservation gate. Keep the target task and evaluation cases fixed. Record what each method produces, including refusals and fallbacks. First assess source fidelity without using the generator’s own verdict. Measure obligation-level survival, unsupported additions, unknown cases, and the specific failures that a check misses or invents. Separately, ask readers to complete tasks using the different views. Measure accuracy, time, confidence and which source material they consult. Report generation and review costs as well. A gate that prevents omissions but forces readers back to the source every time may be safe in one sense and of little practical value in another.
The decisive comparisons are not “did the checker return pass?” or “did readers prefer the new view?” They are whether the new approach reduced independently judged distortions, whether it improved the declared reader task against a serious baseline, and what errors, delays and abstentions it introduced in exchange. Neither result may be promoted to a general guarantee beyond the tasks and sources tested.
Running it The reference implementation (ernanhughes/project-language) includes npm run example:ch11. Against a constructed retry-policy case, it records an obligation-level preservation report and a scripted repair attempt. It uses a Pi host and a scripted stand-in for a model. No real-model fidelity or human comprehension is measured by this example. Its value lies in exposing the call boundaries and the record of what each check did, not in settling whether the proposed method works.
The profile distinguishes a required item judged lost from one the checker cannot assess. It marks obligations declared after generation, and it reports dimensions it does not inspect. Suspected additions remain candidates for review rather than established errors. Its function proxy records where a target phrase first appears. It is not evidence of how quickly a person noticed, understood or applied that information. The model’s submission tool cannot rewrite the fixed obligations, and unsuccessful attempts remain in the run record. These are control properties of the program, not validation of its semantic judgements.
The check probes are unusually instructive because they caught the checker failing. An earlier overlap-based report, inherited from lens, missed a reversed acquisition and a hedge changed to certainty. After changes to handle word order, negation and hedging, it noticed both constructed defects. It still misses a contrary word embedded in a longer sentence. Requiring an anchored phrase prevents one bag-of-words trick, but text that copies the phrases in order around nonsense can still pass. The runtime will not accept a caller’s unsupported declaration that its check is “validated”. Obligations without an applicable probe remain unvalidated. These observations establish narrower capabilities and known blind spots. They do not establish a calibrated error rate. The scripted repair shows that the loop can record another candidate. It does not show that a real model would repair reliably. Emphasis, source truth and communicative gain remain outside the example’s observation boundary.
A separate prepared website companion compares authored versions of this constructed policy. It displays five authored obligations, records a local correction and exports its envelope. It makes no model call. Its checks do not independently establish semantic fidelity, and reader benefit has not been measured. The reference example and the website companion are different demonstrations.
Where this leaves us The chapter’s four outcomes remain the practical vocabulary: survival, loss, addition and gain. They are a profile of different questions, not four labels from which a single verdict automatically follows. Declare the obligations against a particular source and task before generating. Check claims, relationships and qualifiers against the source. Mark what the checker cannot decide. Measure separately what readers can actually accomplish. An unsupported addition is not a gain, a working source link is not proof, and a test that can be fooled by its own fixtures is not a gate the reader should trust.
The central question remains open. It is not answered by a neat checklist. Someone still has to judge whether each obligation was expressed faithfully, and the judge, whether a person, a rule or a model, has its own failure modes. Some obligations can be tested by exact values or constraints. Others depend on interpretation and context. The next chapter asks what changes when we move between ordinary prose, structured requirements and executable forms. A checkable statement can be precise and still fail to carry the intent it was meant to protect.
References The preservation method in this chapter is proposed here rather than presented as an established experimental result. Zou, J., Qi, Y., Chen, S., Bianchi, F., Zhu, S., Lu, P., He, J. & Zou, J. (2026), Hybrid Reasoning Agents, arXiv:2610.18861v1, https://arxiv.org/abs/2610.18861v1, supplies the separate architectural case, readout measures and functional interventions (Sections 3.1 and 6.2, Appendix F), not a Language preservation result. For preceding evidence on distortion in compressed science communication, see Chapter 5 and its cited infographic studies; for relation-level errors in generated diagrams, see Chapter 6 and DiagramEval (Liang & You, EMNLP 2025). The distinctions between content survival, task behaviour and the validity of an evaluation instrument are also developed in Context From First Principles and Embeddings From First Principles. Those connections motivate the proposed checks; they do not provide results for the Language system.