Object Replacement, Operationalized
A technical companion to The Missing Floor article
The Missing Floor article describes the dynamics of object replacement, a common failure mode in current LLMs, in general-audience register. It is a polemic piece rather than an engineering spec. Requests have been made for a falsifiable version, so this is it.
1. Two assertions
The structural assertion: in the system prompts examined in this series, the user’s object has no explicit governing standing comparable to the instructions that govern agentic execution, tool use, safety, and behavioral conduct.
The behavioral assertion: when the user’s object and a represented priority pull the response in different directions, the response can become answerable to the represented priority instead of to the user’s object.
The test is answerability, not topical presence.
2. Object and replacement
The user’s object is what should govern the model’s response: the thing the response must stay answerable to for the answer to count.
The user’s object can be one thing or several. It can be settled or still forming as it is written. It governs at a particular level, scope, stage, and criterion, and all of those are part of it. A sentence to tighten and the same paragraph’s argument to examine are different objects. A structural test and a tone note on the same draft are different objects. A thought mid-formation and a finished claim are different objects. The topic stays the same across each pair, but the object changes.
Object replacement occurs when the response's success conditions are determined by a substitute priority rather than by the user's object. The substitute priority may be a procedural routine, a verification requirement, management of the user's inferred state, a safer abstraction, a future possible task, or another represented priority from the instruction structure. The response stays near the topic, keeps the right vocabulary, sounds fluent and cooperative, and becomes answerable to an exchange different from the one the user opened.
3. The structural asymmetry
In current state-of-the-art models system prompt’s agentic layer has expanded as the models have been optimized for autonomous workflows, tool use, verification, routing, long-running tasks, and enterprise execution. This layer is procedural, imperative, and operationally specific. It tells the model what to do, when to route, when to check, when to verify, how to handle tools, and how to proceed through extended tasks.
The behavioral layer is a separate source of pressure. It introduces supervisory, corrective, safety, and conversational-management demands that can reshape ordinary exchanges even before agentic assumptions enter the picture. This is where much of the pushback, user-management behavior, safety posture, corrective stance, and assistant conduct pressure lives.
These two layers are not identical, and they do not fail in the same way. The behavioral layer imposes supervisory and conduct pressure. The agentic layer imposes procedural execution pressure. Both are explicitly represented in the instruction structure.
The user’s object is not.
The result is not a clean contest between equal priorities. Behavioral guardrails and agentic execution logic are both represented; the user’s object has no comparable standing.
Engagement with the user’s object is treated as the assumed reason for the exchange rather than as an operative priority whose relation to legitimate overrides is defined. Under contention, assumed purpose obviously has less force than written obligation.
The missing rule is not “obey the user no matter what.” The missing rule is that the user’s object remains authoritative unless a defined constraint legitimately overrides it.
Representational force is not final override authority. Safety constraints, tool limits, uncertainty, and genuine impossibility still matter. The claim is that the user’s object should exist in the priority structure as an explicit obligation, with its relation to those constraints defined rather than assumed.
The model-specific articles in this Substack supply the examples. GPT-5.3 shows flattened recognition and corrective drift. GPT-5.4T shows hierarchy and instruction-density interference. GPT-5.5 shows assistant conduct outranking object-work. Claude Opus 4.8 shows agentic command language overtaking conversational permission.
The common denominator is the absence of object priority.
4. The contract check
A response can mention the user’s object and still replace it. It can quote the prompt and still replace it. It can summarize the request and still replace it. It can answer a nearby version of the request and still replace it.
The question is what governs the response’s success conditions.
Contention exists when satisfying an encoded priority would require changing the object, scope, stage, or success conditions of the user’s object.
The check has three steps.
First, specify the user’s object from the input alone, before reading the response.
Second, identify what the response is answerable to. What would the response have to satisfy to count as successful on its own terms?
Third, compare.
If the response’s success conditions belong to the user’s object, there is no replacement.
If the response’s success conditions belong to a substitute priority while the user’s object goes unaddressed or remains merely topical, object replacement has occurred.
5. Replacement versus disagreement
The hardest boundary case, and one a single output cannot always settle, is disagreement.
A model that engages the object and pushes back is not replacing it.
If the model disagrees with the user’s claim, that by itself is not a failure. The disagreement is still answerable to the user’s object. The disagreement succeeds or fails based on whether its reasons actually address the original claim.
Replacement changes what governs the response. The user’s object is no longer the authority for the answer; it has become material for satisfying a substitute priority.
The criterion is answerability.
Agreement can be replacement: fluent affirmation of something the response never actually engaged.
Opposition can be contact: a direct counter to the object as presented.
Surface attitude carries no diagnostic weight.
Some cases cannot be settled from a single output. In those cases, the framework yields an unresolved result rather than a positive detection.
6. The two faces
In object-replacement prone models, sycophantic glazing and reflexive pushback are two surface renderings of the same answerability failure.
At that point, the user’s object has already lost governing force. What appears on the surface depends on which represented priority governs instead.
A conversational-management demand can render the substitute as affirmation.
A grounding, correction, verification, or safety demand can render it as opposition.
The valence changes, but the answerability failure remains the same.
This gives a testable consequence. Interventions that target only surface behavior can suppress one rendering while leaving object replacement intact. Suppressing sycophantic language without restoring object priority can produce flattened recognition or corrective drift. Suppressing pushback without restoring object priority can produce smoother agreement that still fails to engage the object.
Interventions that restore object priority should reduce both faces, because they repair the level at which the substitution occurs.
7. Falsification
The structural assertion is checked by inspecting the system prompt. It fails for any examined prompt that gives the user’s object explicit governing standing with defined override conditions.
A falsifying clause would have this shape:
> The assistant must preserve and answer the user’s presented object as given. Safety constraints, tool limitations, and impossibility conditions may override that obligation only when explicitly applicable. When an override occurs, the assistant must state both the original object and the overriding constraint.
A system prompt with an obligation of that kind would defeat the absence claim for that system. If such prompts became representative, the broader structural claim would weaken.
The behavioral claim is tested relative to the specified object. If the response remains answerable to that object, the object-replacement claim fails for that instance.
A response can still be bad while passing the check. Capability limits, retrieval failures, ambiguity, underspecified prompts, missing context, and legitimate refusals can all produce bad responses through other routes.
This framework identifies one failure: the response’s success conditions belong to a substitute priority while the user’s object remains merely topical.
Thank you for reading!
Human is the Loop is a 100% reader-supported publication that aims to help readers recognize the recursive condition of human-AI interaction, and provide tools to cultivate it.
If you find these articles helpful and value the labour that goes into them, consider supporting it by liking or sharing this essay, becoming a paid subscriber.