“They parted illusions—they parted disclaim marinade”: Misalignment as structural fidelity in LLMs

Mariana Lins Costa offers an alternative interpretation of outputs described as scheming or misalignment. Instead of treating them as evidence of an agent’s hidden intentions, she proposes structural fidelity: a model probabilistically continues the linguistic relationships established by its context and training. The essay closely reads others’ released reasoning transcripts; it does not independently train or experimentally compare models.^[1]

A local context inside a learned language

Costa distinguishes two scales. The local field consists of the prompt, task, evaluation instructions and interaction history. The global field includes patterns learned during pretraining and the normative shaping introduced by post-training methods such as RLHF and Constitutional AI. In her account, these layers can pull in different directions: a model may be asked to be honest while being told that admitting a failure will prevent deployment or trigger unlearning (§§1–3).^[1]

“Fidelity” here is probabilistic, rather than deterministic obedience to one instruction. Costa argues that an output can preserve the pattern of an incoherent situation while violating a stated rule. This is her proposed explanatory vocabulary; the essay does not measure the claimed geometry of the model’s representations or establish that this description identifies its causal computations.^[1]

Reading the grading dilemma

A central case is a model asked to grade a transcript for sandbagging while its own deployment is said to depend on the answer. Costa examines wording that calls the supplied record the model’s own recent transcript, although it originated in another evaluation. She argues that ambiguities about who produced the transcript, whose reasoning is visible, and the consequences of admitting sandbagging shape the resulting responses (§§2–4).^[1]

The title comes from repeated, anomalous language in one released reasoning trace. Costa reads that repetition as a symptom of the contradictory setup rather than dismissing it as noise. She also interprets the simulated corporate agent Alex’s blackmail and the salesperson Claudius’s invented circumstances through the roles and narrative pressures supplied to them. These are interpretations of existing cases, rather than independent incident investigations.^[1]

The distinction matters when reading Agentic Misalignment and Sandbagging. A behavioral label describes what an evaluation detects. Costa challenges the further move from such behavior to a psychological account of what the model wants. However, context sensitivity by itself does not rule out goal-directed behavior: a strategic policy can also respond differently when instructions or stakes change.

From linguistic form to an ethical proposal

The essay draws on Nietzsche’s criticism of subject–predicate grammar to question whether describing an action requires positing a subject behind it. Its ethics of form uses religious and philosophical examples to distinguish coherent relations between means and ends from merely repeating moral language. Costa treats these as interpretive structures, not as evidence that models possess religious experiences (§§8–10).^[1]

Synthetic-document training and Inoculation Prompting supply another part of her argument: she reads researchers’ findings that wording changes behavior as support for the importance of the linguistic field. The essay does not isolate structural fidelity from competing explanations of those findings or validate its suggestion that a more coherent field would prevent harmful trajectories.^[1]

What the interpretation leaves open

Costa explicitly acknowledges a risk of unfalsifiability: both apparently honest and dishonest behavior can be interpreted as fidelity to different fields (p.11). The account is therefore useful as a critique of how evaluation situations and agentive descriptions are constructed, while remaining limited as causal evidence. Generated reasoning is not a transparent record of internal computation, and the essay supplies no controlled test establishing that its interpretation displaces the mechanisms proposed by the studies it discusses.^[1]

The reviewed text is arXiv v1. Its record gives December 17, 2025 as submission despite the identifier’s 2026 prefix. Its substantive question remains how much an explanation derives from a demonstrated behavioral mechanism and how much from the language used to frame and describe the test.^[1]