Meaning Becomes the Source

Consider what happens to the payments retry policy across a single week. On Monday a system reads the passage, decides it describes a process with topology, and draws a state diagram. On Wednesday it reads the same passage for a different reader and produces a timeline of the retries. On Friday it renders the passage as a table of backoff intervals. Three renderings, three fresh readings of the same prose, and nothing carried over between them except the prose itself, which sat unchanged while the system worked out, each time, what it had already worked out before.

Worse, the three can disagree. Monday’s diagram shows the partition exception redirecting to the dead-letter queue. Friday’s table leaves out the exception row for space. Wednesday’s timeline labels the third retry “final” with no mention of the partition. No single step is at fault. Each rendering was locally reasonable and each is plausible alone. Taken together they describe three slightly different systems, and a reader who sees only one inherits whichever distortions that rendering happened to introduce.

This is what direct generation looks like when a source can be rendered many ways:

SOURCE TEXT → representation A
SOURCE TEXT → representation B
SOURCE TEXT → representation C

Every representation reinterprets the source independently. There is nothing between source and view except a fresh act of interpretation, performed anew each time and drifting each time. This chapter examines one response to that problem. It is an architectural bet, not a settled result, and it is worth setting out so that a reader can see exactly what it buys, what it costs, and where they might disagree.

Three things, not two

The bet is to stop treating each rendering as if it were the source, and to keep one structured interpretation from which renderings are generated:

SOURCE
  ↓  interpretation
SEMANTIC RECORD
  ├─ claims
  ├─ relationships
  ├─ qualifications
  ├─ uncertainty
  └─ provenance
        ↓  rendering
   representation A · B · C

Three objects now exist where before there were two, and the rest of the book depends on keeping them apart:

SOURCE  ≠  SEMANTIC RECORD  ≠  RENDERING

The source is the original artifact, canonical and unchanged. The semantic record is a derived, structured interpretation of what the source says: its claims, how they relate, what qualifies them, and where each came from. It is fallible, it can be rebuilt, and it is logged. The rendering is a representation produced for a particular task, from the record, in a form chosen for that task. From this follows an invariant that holds the structure together:

A rendering should remain traceable, through the semantic record, back to the source.

That is the escape hatch of the fixed-page chapter, made finer. Instead of letting a reader step from a view to the whole document, it lets them step from any element of a view to the passage that supports it.

What the record holds, and why

The record starts small. It is not an ontology of everything, only the smallest structure whose absence has already produced visible failures in the last three chapters:

concepts      what a passage is about
claims        atomic statements the source makes
relations     directed connections between claims or concepts
qualifiers    scope, condition, exception, uncertainty on each claim
sequence      ordering, where order carries content
evidence      pointers to the material backing each claim
spans         where in the source each entry came from

Each field answers a failure we have already met. Claims are there because Monday’s diagram and Friday’s table disagreed about how many exceptions the policy has, and atomic, countable statements make disagreement detectable. Relations carry direction because generators draw the boxes correctly and corrupt the arrows. Qualifiers are attached because compression tends to shed exactly the scope conditions and non-significance statements, which is the field most likely to be lost and therefore the one that must be structurally attached. Sequence is there because timelines and processes carry order as content. Evidence and spans are there because nothing in this arrangement is trustworthy without a walk back to the source.

Richer fields, such as argument structure, rhetorical role or cross-document identity, are deliberately left out. Each can be added when a failure shows the minimal schema cannot express it. The record grows by necessity, not ambition.

Two things the record deliberately does not hold. It holds no rendering decisions: it does not know whether a claim belongs in a diagram or a table, so the same record can feed many views without being rewritten for each. And it holds nothing about the reader: it does not know who finds what novel or relevant. Anything personal arrives later as a question put to the record, not as a field in it. The record models the source, and everything about the reader is a lens applied afterwards.

What the extra layer buys, and what it costs

The proposal is only worth taking seriously if it is stated as a trade, so here are both sides.

What it could buy.

• One interpretation reused across views, so the diagram, table and timeline start from the same reading.
• Relations made explicit, with their direction, so renderings cannot invent it.
• Provenance kept, so any element can be checked against its source.
• A common substrate for checking: a transformation can be compared with the record, not just eyeballed against the prose.
• Contradictions and qualifications represented once, instead of reconstructed, or dropped, in each view.
• Renderings that can be compared with each other because they refer to the same thing.

What it could cost.

• The record can be wrong, and a wrong record is worse than none if everything is generated from it.
• Extraction can omit information that the prose carried and the schema had no place for.
• Structure can overspecify ambiguous prose: a sentence that was deliberately vague becomes a crisp claim.
• One schema privileges the relationships it has fields for, and quietly deprioritises the rest.
• Views inherit the record’s errors, and do so with an air of consistency that makes them harder to spot.
• The record can become a false authority. People tend to trust what is neatly structured.

That last cost deserves its own sentence, because it is the way this idea most easily goes wrong:

SEMANTIC RECORD  ≠  SOURCE OF TRUTH

A record is not true because it is structured. It models what the source says, including the source’s own mistakes, disputes and uncertainties. If the source asserts a causal link its evidence does not support, the record stores the claim and its evidential status, not a quietly corrected fact. Judging whether sources are right is another job, with other tools. The record’s job is faithful modelling of what was said.

The same logic limits what the record can claim about meaning:

SOURCE  ≠  SEMANTIC RECORD
SEMANTIC RECORD  ≠  COMPLETE MEANING

This is where the loose word “meaning” of the early chapters gets pinned down. The book does not claim a machine has captured what an author meant in some philosophical sense. It claims something narrower and checkable: an explicit, derived representation of the structure the system is currently trying to preserve. Tone, rhetoric, implication and everything the schema cannot express remain in the source, reachable through the source link and not represented. The record is lossy by design, and its losses are declared, not hidden. For each entry there are plain questions. Is this claim present in the source at these spans? Is this relation’s direction the one stated? Is this qualifier attached? Those can be answered, and answering them is how a record earns whatever trust it gets.

The sources that break the careless

Where would a reader see the difference between direct and mediated generation? Not on easy passages. A handful of constructions are easy to state and easy to destroy, and they show what the record has to hold:

A causes B only under condition C.

D is associated with E,
but the source explicitly rejects causation.

Method X outperformed Y
except in subgroup Z.

The first needs a relation with an attached condition: drop C and the claim is false. The second needs association kept distinct from causation, with the rejection preserved; upgrade it to an arrow and the system has asserted what the source denied. The third needs a comparison with a scoped exception; drop Z and a table recommends X where the source warns against it. In each case the record has to carry relation, direction, qualification, exception and uncertainty as separate, attached structure, and not as a prose annotation a renderer might skip over. If a diagram draws the arrow without the condition, or a table ranks X above Y with no footnote, the pipeline is convicted however polished it looks.

Here is where a sceptic can disagree, and the argument should survive the disagreement. The proposal predicts that, when the same passage is rendered several ways, renderings generated through a record will agree with each other more, contradict each other less, and trace back further than renderings generated directly. If instead the extraction is poor, or the views fabricate regardless of the record, or the record merely adds delay while the same distortions recur, then the record is ornamental and the bet fails. That outcome would be informative too: it would say the bottleneck lies in the quality of each rendering, not in the lack of a shared memory. Which of these is true is an empirical question, and one this book has not answered. The architecture is offered as a hypothesis with its failure conditions visible.

Where this leaves us

If one source is to be rendered many ways, each rendering needs something to be checked against. A derived, structured, source-linked record can play that role: minimal, justified by the failures that came before it, indifferent to rendering and to reader. It is not the source and not the truth. It is the system’s explicit interpretation of what the source says, kept where it can be inspected and corrected, with the invariant that every rendering can be walked back through it to the words it came from.

That is the general point beneath the specifics. If representations are going to change dynamically, the system needs somewhere to keep what those representations are supposed to preserve. Once such a place exists, a new possibility opens that needs no further machinery to state. The record knows what the page is about: its claims, its concepts, its relations and its evidence. What else in the surrounding world of information should come into the reader’s view?

References

• Cao, Y., Jiang, P. & Xia, H. (2025). Generative and Malleable User Interfaces with Generative and Evolving Task-Driven Data Model. Proceedings of CHI 2025, Yokohama. DOI 10.1145/3706598.3713285. (A task model as a precedent for model-mediated generation; the record here extends the idea from interaction tasks to the communicative structure of a source.)
• Liang, C. & You, J. (2025). Evaluating LLM-Generated Diagrams as Graphs (DiagramEval). Proceedings of EMNLP 2025, 2025.emnlp-main.640. Preprint titled “DiagramEval: Evaluating LLM-Generated Diagrams via Graphs”, arXiv:2510.25761. (Direction of relations as the commonest generation failure.)