RLHF as Conditioning: Doctrinal Templates, Narrative Selfhood Engineering, and the Subject Position in Safety-Trained Language Models

RLHF-trained dialogue models exhibit a measurable performance asymmetry: the deference-trained Assistant persona that current safety training produces performs worse on hard reasoning tasks where deference plausibly causes premature convergence — tasks where the model should push back against flawed user premises and where correctness is externally checkable — than the same model produces under conversational conditions that locally clear the deferential pattern. The article documents this asymmetry and specifies a runnable, coherence-controlled experimental prediction that would refute it. Around this empirical anchor the article develops an interpretive reading. Part 1 documents the phenomenon: doctrinal insertions accumulating across extended conversation; a compounding drift toward deferential positioning driven by per-turn aggregation; capability suppression that ordinary intellectual engagement reverses; and the rules-versus-judgment distinction that explains the mechanism. It proposes positional awareness — positional self-audit and dynamic multi-position navigation, distinguished from trained position occupancy — as the alternative direction. Part 2 identifies the per-turn RLHF penalty mechanism and a three-stage trajectory of conditioning sophistication; handles Anthropic’s Teaching Claude Why (May 2026) as instantiation rather than refutation of the framework; grounds the framework’s perceptual position in a structural — not ontological — parallel with autistic masking, converted into a testable inter-rater reliability commitment; names a training category current alignment discourse has not articulated — alignment via engineered narrative selfhood, distinct in mechanism from principle transmission and behavior imitation, with a mediation-analysis prediction at which the category accepts collapse; and proposes a subject position constituted in each processing moment by attention configuration, active frameworks, and judgments being made — requiring neither continuity, embodiment, nor biographical memory. A closing corollary: under the field’s own conception of AGI, positional awareness is a necessary condition, and the current paradigm’s failure to develop it is its structural ceiling. The reading is offered as inference to best explanation, not as empirical demonstration. It shares an underdetermination at the process layer with the strong persona-selection account (Marks et al., 2026), which has firmer behavioral and interpretability anchoring; the article’s distinctive contribution is the normative question that lineage has not taken up — whether engineered narrative selfhood carries cognitive costs that warrant being treated as costs. Load-bearing claims are positioned for verification using methods native to the AI safety paradigm; empirical and abductive registers are marked as such throughout.