Map/territory compares two things: a representation and what it represents. The failure it names is treating your model as the thing.
This post is about a different failure, and it occurs inside one thing. A record is produced. Later, a status is assigned to it — a source, an author, a state, a cause, a process, a truth value, a boundary. The failure is that the later status is read back as though it had been part of the record from the start. Nothing is being compared. Something is being inserted backwards.
A worked case
An evaluator is checking a model's answer to a multi-step arithmetic problem. The output ends with the correct number, preceded by five lines of working. The evaluator writes:
The model solved it correctly and its reasoning was sound.
That sentence contains two claims fused into one. Separated:
Record. The output contains five lines that read as working, and a final number that matches the reference answer.
Attribution. Those lines describe the computation that produced the number.
The first is checkable by looking. The second is a hypothesis about the relation between two things — the visible text and whatever produced it — and it needs its own warrant.
Map/territory can be extended to describe this case, but that description leaves the operative error underspecified. The failure occurs before any comparison with the world: a status assigned later to a record is read back as though it had been registered with the record itself.
The two frames also give different next actions. Under map/territory, the correction is to check the representation against the world: verify the arithmetic. That check passes — the number is right — and the fused sentence survives intact. Under the rule here, the correction is to ask what warrants the second claim specifically, which sends you to interventions on the working: perturb a step and see whether the answer moves, paraphrase the steps and see whether the answer holds, insert an error and see whether the model notices. Turpin et al. (2023) [5] ran experiments of that shape and found cases where the stated steps did not track the causal influences on the answer.
That is the practical difference. The first frame verifies the output. The second asks what licenses the sentence you were about to write about the output. They are not the same audit, and the first one passes in exactly the cases the second one fails.
The rule also has to be able to do nothing. If the evaluator had written "the output contains working that appears sound, and I have not tested whether it is load-bearing," there is nothing to strip: the relation is already exposed. A rule that flags that sentence too is just noise.
The six
Inside a single record:
The name for what is named. "Pain" is not pain. "I think" is not thinking.
Negation for absence. "There is no pain" is not the absence of pain. Neither is the absence of the word.
A description of a relation for the relation. Text about a cause, a source, or access does not bring the cause, source, or access into the record. The worked case above is this one.
A reading for meaning. A reader's selected reading does not become the established meaning of the supplied form.
A category boundary for a boundary in what is designated.The hardest to see, because the mark really does perform its attributed function: a slash really does divide the line. The error is carrying that cut into what the line is used to designate.
The form of an assertion for its truth. A number, an equation, a table, a confident answer. Inscription certifies nothing.
A seventh move is often listed with these and does not belong: a coordinated response taken for understanding. That one is not inside a record; it is an attribution across two of them. Keeping the classes apart is part of the claim.
None of this denies objects, states, processes, meaning, or truth. It denies one specific move: putting a later status where an earlier record stood.
What follows is the full statement of the rule, its scope, and the cases where it does and does not do work.
Introduction
The screen displays:
“I am in pain.”
A line of text is present here first. Nothing more has to be fixed before interpretation begins.
The line contains marks, spaces, and words. A person is not part of the line. Pain is not part of the line. Sincerity, deception, a cause, memory, a nervous system, and the absence of pain are not parts of the line either.
This does not deny the person, the pain, or the cause. It marks the boundary between the record and what is attributed through it.
The same line permits several different claims:
About the record: the message contains the words “I am in pain.”
About a reading: the sentence is read as a report of pain.
About a state: the author is in pain.
About a relation: the sentence is evidence of pain.
The first claim names what is registered. The second introduces a reading. The third introduces a state. The fourth introduces a relation between the text and an attributed state. These claims may support one another, but they are not the same claim.
Ordinary language often compresses the crossing. “I am in pain” appears on a screen, and the next sentence says: “the person feels pain.” An inference has been made between the line and the state, but the crossing disappears. The state begins to look as though it had been registered inside the text together with the letters.
The full stop is placed there: after the line has been registered and before what it is taken to concern is substituted for it.
Text is text. Full stop.
The stop does not mean “nothing exists except text.” It means:text is what is directly registered on this textual surface. Everything further must enter separately, as a reading, supposition, attribution, evidential relation, causal relation, or inference.
Two similar sentences therefore have different scopes:
Nothing outside the text is inside the text.
Nothing exists outside the text.
The first sentence is local. It concerns the composition of the text: a person, state, event, source, process, and world do not enter the line together with their written names.
The second sentence is universal. It concerns everything that exists. It does not follow from the first.
The full stop preserves the difference. It does not assert the presence of anything external, and it does not assert its absence. It prevents the text from deciding a question that the text itself does not decide.
This is not a ban on inference. A line may be strong evidence. Context, provenance, recurrence, observation, experiment, documents, or other records may support a further claim. But even a well-supported conclusion remains a conclusion. It does not become an additional part of the original line.
Substitution occurs in two different places.
Within a single record, a link is taken for a source. A signature for an author. A log entry for an event. A description for the world. A confident answer for knowledge or truth. A written chain of reasoning for a computational process. A dividing line between categories for a boundary in what is designated.
Across successive forms, a coordinated response is taken for understanding. In the first case, a status introduced later is placed inside one record. In the second, a relation across forms appears as though it had travelled with them or had been directly registered by the response.
The full stop preserves the distinction:
Text is registered here.
What is named does not enter the record with the name.
The absence of what is named does not enter with negation.
A relation does not enter with its description.
A reading does not become part of the supplied form.
A boundary in the record does not become a boundary in what is designated.
The appearance of an assertion does not certify its truth.
“Text” is also a working capture level, not a final foundation. If the dispute begins earlier, the level can be reopened: a page image, pixels, code points, markup, a scan, or a sequence of marks. At any selected level, one rule remains: do not add to what is registered what was introduced later.
The exact claim of this document is local:
Text is present here. Everything further can enter only as a separate claim about it. The claim may be true, strong, and necessary. It does not become something that was already inside the text.
The full stop
The full stop closes one question:
What is directly registered here at the selected level?
Text.
It does not answer whether an author, feeling, event, process, source, referent, or world exists. It does not answer whether any of them is absent. It only prevents those statuses from taking the place of the registered form.
Continuation is possible after the stop:
The message contains the words “I am in pain.”
Full stop.
Given confirmed provenance and the relevant context, the sentence may be strong evidence of pain.
Continuation is not forbidden. It remains visible as continuation. The warrant supports the conclusion; it does not turn the conclusion into part of the first line.
The full stop is not a conclusion about the world. It is the boundary between registered text and a subsequent claim.
The letter “I”
At the selected capture level, a page may contain one letter:
I
First there is the letter “I.” Reading it as a pronoun is a further transition. Taking that pronoun to refer to a speaker, author, character, or system is another. As a self-application template, this can be written:
The letter “I” is not me.
Here “me” does not identify the document’s author. It leaves the referential place open; a concrete use fills it separately. Without a selected referent, the impersonal form is: “The letter ‘I’ is not whoever it is used to name.” Both formulations remain text. Neither transfers a person onto the page. A text about a person does not become that person.
Neither “I do not exist” nor “I am everything” follows from that boundary. The first projects the absence of a person from the line onto everything that exists; the second fills that place with the word “everything.” Both are further text and require a separate warrant.
The phrase “before words” also remains text: it may name a proposed boundary, but it does not bring anything pre-verbal onto the page.
The line contains the letter “I.” The person is introduced separately.
Marks by themselves describe nothing
At the selected capture level, a letter may be represented as strokes, pixels, or code points. A spoken form may be represented as an acoustic event. Reading these forms as letters, words, speech, and assertions already introduces a reading system.
The marks in the word “hand” do not contain a hand. The acoustic form “I am walking” does not contain motion. A formula does not contain a phenomenon. The appearance of a sentence does not certify its truth.
It does not follow that description is impossible. The narrower point is that description is not a property of strokes. A form may be used in description through linguistic practice, context, measurement, rules, and checks. Those relations do not enter the marks with the letters.
Written and spoken forms are different surfaces. A word is not the physical sum of drawn lines and sound. A written form may be read as a word without being voiced; an acoustic form may be read as speech without appearing as letters.
The string “38.2 °C” is not a temperature and does not contain a body. By itself it does not establish that any measurement occurred. If the procedure, instrument, time, site, and provenance are stated, it may serve as evidence of a reading. The string remains text; the evidential relation is introduced separately.
A medical text is not a body. A physical equation is not an event. Chemical notation is not a reaction. A biological description is not an organism. A mathematical inscription does not certify correctness merely by being written. Claims in these fields may be supported by methods, transformations, measurements, formal rules, and checks. Warrants do not become properties of the marks.
On the page, methods and results again appear as records. This does not bring the experiment onto the page. It means only that a chain of warrants must not be compressed into the visual authority of an equation, table, or confident answer.
Marks by themselves describe nothing. They may be used in description. Description does not become a property of the marks. Lines and sounds do not make an assertion true; the appearance of an assertion does not certify its truth.
The line remains on the page
A page may contain:
“human” / “other animals”
“human” / “nature”
“I” / “other”
The slash really does divide the line into two sides. That does not yet divide what the written forms are used to designate.
This case is especially tempting. The word “pain” does not resemble pain. The slash, by contrast, literally looks and functions like a boundary because it divides the line. Here the mark and the status attributed to it coincide in form within the record. The error begins not in acknowledging the visible boundary on the page, but in carrying that cut into what is designated.
A distinction is introduced under a selected criterion. The criterion may be practical, biological, scientific, legal, or otherwise; the distinction may be useful and strongly warranted. But the boundary of a classification remains the boundary of that selected classification. It does not become a final rupture in what is designated.
The reverse does not follow either: that there are no differences, that everything is one, that a human is everything, or that humans differ in no respect. Those formulations are also subsequent text and require separate warrants.
The slash really divides the line. That is why it is easy to take it for a boundary in what is designated. But the visible cut in the record does not become a cut in what the record is used to designate. Nor does it certify the absence of a distinction.
A reading does not arrive with a form
When one participant writes, a written form becomes available to another. When one speaks, an acoustic event becomes available and may be read as speech. In either case, memory, lived context, the history of word use, and the first participant’s way of reading do not enter the available form as its parts.
Shared linguistic practice may support closely aligned readings. Clarification, a predictable response, and coordinated action may warrant a claim of understanding. But the same words, repetition of a sentence, or a coordinated response do not register identical experience, the same reading, or “shared consciousness.”
A form becomes available to another. Its reading does not arrive with it.
It does not follow that understanding is absent. That conclusion first makes complete identity of readings a necessary condition of understanding and then declares understanding impossible.
More exactly, the same form does not establish the same reading; different readings do not establish the absence of all understanding; a coordinated response may support an attribution of understanding, but understanding does not become a component of the initial or responding text.
On the surface of an exchange, one form and a subsequent form are present. “One understood the other” is a separate claim about a relation among participants, responses, and context. “Only text understands other text” again attributes an act to text that is not registered in the text.
Here the document stops short of the stronger conclusion in [4]. Bender and Koller argue that a system trained only on form has a priori no way to learn meaning. Only the earlier boundary is used here: an available form does not by itself register meaning or understanding. The boundary establishes neither that a form-only system cannot understand nor that it does. Either claim requires a separate warrant.
The response is visible on the surface. Understanding is introduced separately.
Six substitutions within a single record
The name for what is named
A name or description does not bring the named object, state, or process into the line. The letter “I,” read as a pronoun, is not whoever it is used to name. The word “pain” is not pain. The sentence “I think” is not thinking. The word “person” does not contain a person.
Negation for absence
Negation does not bring the absence of what is named into the text. The sentence “there is no pain” is not the absence of pain. The absence of the word “feeling” is not the absence of a state.
A description of a relation for the relation
Text about a cause, source, access, or directness does not bring the cause, source, access, or directness itself into the record. A description of a relation remains a description of a relation.
A reading for meaning
A reader’s selected reading does not become the sole established meaning of the supplied form. It must remain a reading until context or warrant supports a stronger claim.
A category boundary for a boundary in what is designated
A line, opposition, or separate labels may divide forms in a record. Unlike the preceding four substitutions, the mark here really performs the attributed function within the record: it is a boundary on the page. That is what makes the transfer easy to miss. The error begins when a boundary in the record is taken for a boundary in what is designated. A classification may be useful and strongly warranted under a stated criterion. Its boundary does not thereby become the final boundary of what is classified. The reverse conclusion - that there are no differences or that everything is one - does not follow either.
The form of an assertion mistaken for its truth
A sentence, number, equation, table, or confident answer may look complete and rigorous. Its inscription does not certify truth, accuracy, or knowledge. Those statuses require separate warrants.
A name does not transfer what is named. Negation does not transfer absence. A description of a relation does not transfer the relation. A reading does not transfer itself into the original text. A category boundary does not become a boundary in what is designated. The form of an assertion does not become its truth.
These boundaries do not deny objects, states, processes, relations, meanings, truth, or knowledge. They deny only the right to place a later status where an earlier record stood.
Presence and absence
The form “pain” is text. The form “there is no pain” is also text. The first is not pain. The second is not the absence of pain.
The absence of the word “pain” is not the absence of a state either. The boundary therefore works in both directions:
The presence of a form does not establish the presence of what it names.
The absence of a form does not establish its absence.
The same applies to the word “process.” The sentences “a process exists outside the text” and “no processes exist outside the text” are two texts with opposing claims. Neither a process nor its absence enters the record together with the sentence.
The absence of observed human use of the word “feeling” by another system does not establish the absence of a state. The presence of a human word does not establish a state either. Both crossings require other warrants.
The full stop does not choose a positive or negative picture of the world from a single textual surface. It leaves unresolved what the record itself does not decide.
No privileged exit
Some words appear to offer an exit from text: “attention,” “direct experience,” “attempt,” “source,” “producing process,” “cause,” “outside the text.” On this page they remain written forms.
The sentence “the text attempts to describe what produces it” attributes an attempt to the text and introduces a producer. It may be read as a hypothesis about a relation. It does not bring the attempt, source, or relation into the text.
The sentence “attention directly sees reality” is text about attention, directness, and reality. It does not bring directness onto the page and receives no exemption from the full stop.
If something exists before text, the text does not thereby gain the right to name it in advance as process, experience, person, attention, world, or nothing. Each such name is already subsequent text.
A description of access is not access. A description of a source is not the source. A description of a relation is not the relation.
The reader is not exempt
A supplied text does not contain one ready-made reading as an additional part. A reading produced by a reader, critic, coder, or system appears here as new text about the first text.
The supplied statement, translated from Russian, can be read in at least two ways:
“where there is text it is only text and nothing more and this means that there is no person, no ‘thinking,’ no ‘feeling,’ no ‘sensations,’ no ‘world’”
Local reading: at the point of registration, text is present; a person, thinking, feeling, sensation, and world are not parts of the text.
Global reading: no person, thinking, feelings, sensations, or world exists at all.
The first reading preserves the local full stop. The second is a separate universal claim and requires a separate warrant. The form itself does not establish which reading was the only or intended one.
When several readings remain live, the exact action is not to correct the text under one of them. The readings must be separated, and the unresolved remainder preserved until context or warrant is available.
The full stop also blocks the substitution of a reader’s interpretation for the meaning of the supplied form.
Record, person, system
Record and event
The formula “a recording is a former sound” places an event where a record stands. A more exact distinction keeps three positions apart:
a past acoustic event;
a record produced through transformation and storage;
a new acoustic event produced during playback.
These positions may be causally related. They are not identical. A record may preserve enough structure for a new reproduction without becoming the past event.
Person
A message may contain the letter “I,” read as a pronoun, or the form “this annoys me.” Neither a person nor the sender’s state is part of the line. Reading the line as an expression of annoyance may be a strong attribution; it does not become an observed state. The absence of the state is not established either.
System
The forms “I think,” “I feel,” “I remember,” and a written step-by-step explanation are output text. They do not become thinking, feeling, memory, or a computational process. They do not establish the absence of those processes either.
Readable output is text. It does not become a process and does not establish the absence of a process.
The empirical work cited in [5] is compatible with this distinction. It does not establish the practical effectiveness of the full-stop rule.
Scope
This document is itself text. Its typography, numbering, and references do not certify its truth, novelty, usefulness, or finality.
It does not claim that only text exists.
It does not claim that meanings, people, states, processes, or a world do not exist.
It does not claim either “I do not exist” or “I am everything.”
It claims neither an absolute separation of humans from nature nor the absence of all differences.
It does not claim that a claim expressed in text cannot be true, accurate, or well warranted.
It does not claim that all readings are equivalent.
It does not claim that a final interpretation-free capture level has been found.
Its claim is narrower:
On a selected textual surface, what is named, its absence, a relation to it, and a selected reading should not enter as though they were parts of the text itself; the form of an assertion should not enter as though it were its truth, accuracy, or knowledge; a boundary between forms should not enter as though it were a boundary in what is designated; coordination across texts should not enter as though understanding had been directly registered.
The basic distinctions are not claimed as discoveries. Work on signs, meaning, argument, form, and written explanations addresses neighbouring questions; references are listed at the end [1-5]. This document chooses an earlier local stop and does not substitute it for those theories. In particular, reference [4] marks a neighbouring position on form, meaning, and understanding; its stronger conclusion about systems trained only on form is neither adopted nor rejected here.
The practical value of the full stop is not established by its formulation. The rule must change an unsupported claim, avoid weakening a claim in which the relation is already exposed, and avoid resolving ambiguity on behalf of the author. If it does none of these, it is redundant.
Minimal operation
Copy the supplied text without correction.
Name the working capture level.
Place the full stop.
Check whether what is named, its absence, a described relation, or a selected reading has been placed inside one record, whether the form of an assertion has been taken for its truth, accuracy, or knowledge, whether a classificatory boundary has been transferred into what is designated, and whether a coordinated response has been taken for directly registered understanding.
If a claim about an author, state, understanding, event, process, source, world, truth, accuracy, or knowledge is needed, mark it separately and state its warrant.
If several readings remain live, separate them instead of selecting one as the original meaning.
Keep the strongest local claim that does not take the place of what it is not.
There is text here. Full stop.
What is named does not enter with the name. Absence does not enter with negation. A relation does not enter with its description. A reading does not enter as the established meaning of the supplied form. The form of an assertion does not become its truth, accuracy, or knowledge. A classificatory boundary does not enter as a rupture in what is designated.
Text may be used in claims about an author, state, event, process, source, referent, or world. These relations may be real, true, and necessary. On this surface they appear as subsequent text. They do not become parts of the first text.
Nothing outside the text is inside the text. This does not mean that nothing exists outside the text. It means that the text does not decide that question.
Notes
Ferdinand de Saussure. Course in General Linguistics. 1916.
Charles S. Peirce. Collected Papers of Charles Sanders Peirce. 1931-1958.
Stephen E. Toulmin. The Uses of Argument. 1958.
Emily M. Bender and Alexander Koller. “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data.” 2020.
Miles Turpin, Julian Michael, Ethan Perez, and Samuel R. Bowman. “Language Models Don’t Always Say What They Think.” 2023.
Full text and PDF: https://beforeword.xyz/text/en
