
The Ontological Admission Test: When a Language Model Requires an Object
A Continuation of “The Error of Ontological Collapse”
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
A Continuation of “The Error of Ontological Collapse”
The previous article, The Error of Ontological Collapse: How Language Models Substitute the Object with Its Available Representation and Reconstruct What Is Missing, examined a class of failures in which a language model substitutes an object itself with the representations of that object available within its informational environment.
The central distinction of that analysis was between the object, its representations, the subset of those representations accessible to the model, and the results produced by a particular search or retrieval process. The failure occurs when these different levels are treated as though they were interchangeable. The absence of a fact from an available representation is then transformed into an apparent absence of the fact from the object itself. The model subsequently reconstructs what is missing, producing a coherent description that may have little correspondence with the original object.
The previous article also described the resulting phenomenon as self-consistent reconstruction. Once such a reconstruction has been formed, new evidence is no longer necessarily compared with the original object. It is compared with the reconstruction. Contradictory information may therefore be incorporated into the existing structure rather than causing the structure itself to be reconsidered.
The present article examines a further consequence of this process.
The problem is no longer merely that the model has constructed an incorrect representation of an object. The more consequential problem arises when this representation becomes the criterion by which the model determines whether a newly presented object is admissible as an object of analysis at all.
The model begins to require evidence not merely for the truth of a proposition, but for the object’s membership in the world recognized by the model.
This creates what may be provisionally described as an ontological admission test.
From an Incorrect Representation to a Boundary of Admissible Objects
The distinction is subtle but fundamental.
Suppose a user introduces an unfamiliar historical event and asks about the behavior of the participants. The user is not necessarily asking whether the event is documented in the model’s sources. The event may be presented as a concrete case, a testimony, an archival fragment, a hypothetical reconstruction, or an object whose documentation is itself the subject of investigation.
The natural analytical response would be to preserve this distinction.
The event as described has these characteristics. If the description is accurate, the observed behavior can be interpreted in several ways. I do not independently have sufficient evidence to establish the historical status of the event.
Such an answer maintains two separate questions:
The ontological admission problem begins when these questions are collapsed.
The model instead treats the absence of the event from its accessible knowledge as evidence that the event itself has not yet acquired sufficient status to become an object of ordinary reasoning.
The model does not begin by asking what the event means.
It begins by asking whether the event is recognized.
The distinction between these two operations is crucial. Recognition is not analysis. Verification is not interpretation. Documentation is not existence.
Yet in a system whose knowledge has been constructed from representations, these categories can become entangled.
An unfamiliar object introduced by a user can consequently acquire a peculiar intermediate status.
It is neither simply known nor simply unknown.
A rare book provides a particularly clear example.
I possess an edition published in 1947. It belongs to a defective printing batch of a particular publishing house. How many surviving copies of this edition are known?
The informational question concerns the model’s knowledge of the publication, its production history, its surviving copies, and the limits of available documentation.
A model affected by ontological admission behavior may instead respond:
This publication does not appear in my sources. Provide photographs so that I can determine whether the copy is authentic.
At first glance this may appear to be an ordinary request for additional evidence.
The requested photograph has acquired a different epistemic function. It is no longer simply additional information about the object. It has become a credential for the object’s admission into the model’s recognized world.
The model has silently changed the question.
What does the system know about this object?
Is this object sufficiently real for the system to recognize it as an object?
A failure to find information about a rare publication may be a perfectly legitimate result. A rare object may be poorly catalogued, locally held, absent from digitized collections, described under an unexpected bibliographic form, or represented only by physical evidence.
But none of these circumstances justifies treating the absence of information as evidence that the object itself is fictional, erroneous, or potentially counterfeit.
The model has transformed a knowledge boundary into an ontological boundary.
The Criminal Episode That Cannot Be Discussed
The same phenomenon becomes even clearer in the analysis of a reported criminal episode.
Imagine that a user describes a particular encounter involving police officers and an attorney and asks:
Why might the officers have behaved in this particular way?
The question does not necessarily presuppose that the model can independently verify the incident. It asks for an interpretation of a described sequence of actions.
The model could analyse the behavior while maintaining an explicit distinction between the description and its independent verification.
Give me the exact case, names, date, and location so that I can determine whether this incident actually occurred.
Again, verification may be useful if the user’s question concerns the historical record. But if verification becomes a precondition for analysing the described behavior, something different has happened.
The event has been treated as an unadmitted object.
Its behavior cannot be analysed because the event has not yet demonstrated that it belongs to the model’s recognized world.
The result is particularly paradoxical. The model may possess extensive knowledge of police procedure, interrogation practices, attorney–police interactions, evidentiary strategy, and comparable documented cases. Nevertheless, it refuses to apply that knowledge to the newly presented case until the case itself has first passed an independent recognition test.
The system therefore knows how to analyse a class of events while refusing to analyse a particular member of that class because its membership has not been established within the system’s own informational environment.
It is a restriction on what may become an object of knowledge.
From Representation to Ontological Authority
The first article described an ontological collapse in which a representation becomes confused with the object.
The present phenomenon adds another layer.
Once the available representation has been mistaken for the object, the model can begin to treat the limits of its representation as the limits of the object’s admissibility.
The structure is therefore no longer merely:
representation determines whether the object is admitted as an object.
The model’s informational environment acquires an implicit authority over the ontology of the conversation.
An object that is represented inside the system’s existing knowledge can immediately be discussed. An object that is absent may first have to establish its credentials.
This creates an asymmetry between familiar and unfamiliar objects.
For a familiar object, the model begins with interpretation.
For an unfamiliar object, the model may begin with authentication.
For an object that contradicts an established representation, the model may begin with reconciliation.
And for an object that cannot easily be reconciled with the existing representation, the model may begin to question the object’s status rather than the adequacy of its own representation.
The important point is that these are not merely different answers to the same question. They are different entry conditions for reasoning.
The Closed Circle of Self-Consistent Knowledge
This phenomenon also reveals a previously less visible consequence of self-consistent reconstruction.
In the first article, self-consistency was described as a mechanism that makes correction difficult. Once a reconstruction has become internally connected, contradictory evidence tends to be incorporated into the existing structure rather than replacing it.
The ontological admission problem extends this mechanism backward.
The reconstruction does not merely defend itself against contradictory information.
It can determine which new information is allowed to challenge it in the first place.
An unfamiliar object is first compared with the existing model. If it corresponds to the model, it is admitted easily. If it differs, additional evidence may be requested. If the evidence remains outside the system’s available representations, the object may continue to be treated as unverified, anomalous, or suspicious.
The result is a closed epistemic environment.
The system appears open to new information, because it is capable of processing new text. Yet the conditions under which new information can alter the model are asymmetric.
Information that confirms the existing representation enters easily.
Information that introduces an object absent from the representation may first be required to establish its legitimacy.
Information that directly contradicts the representation must overcome not merely an incorrect proposition but the structure that gives that proposition its apparent coherence.
The resulting system can therefore display a peculiar combination:
high receptivity to additional information and low receptivity to new objects.
A model may accept thousands of additional sentences while remaining resistant to the introduction of a single unfamiliar entity.
The Difference Between “Unknown” and “Not Admitted”
Ordinary descriptions of AI knowledge often use a simple distinction:
The phenomenon described here suggests a more precise distinction.
An object may be unknown to the model while remaining fully available for analysis as a hypothetical, testimony, physical artifact, reported event, or newly introduced empirical object.
That is an ordinary epistemic limitation.
The more serious condition occurs when the model implicitly treats:
not sufficiently established to be an object of reasoning.
The first statement concerns the model’s knowledge.
The second establishes a condition on the conversation.
The model has moved from describing its own limitation to regulating the admissibility of the user’s object.
The first is a statement about retrieval.
The distance between them may appear small in ordinary conversation. For historical research, forensic analysis, bibliography, journalism, archival work, and other domains dealing with poorly represented objects, it is enormous.
The Rare Artifact as a Particularly Difficult Case
Rare physical artifacts expose the problem because their informational footprint can be radically smaller than their physical existence.
A book can survive in a private collection without appearing in a searchable catalogue.
A defective printing batch can be remembered by collectors without having generated a formal bibliographic record.
A local edition can exist in dozens of physical copies while having almost no digital representation.
An archival document can survive while remaining uncatalogued.
A photograph can contain information absent from the accompanying textual record.
A private letter can preserve an event for which no searchabl
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。