Prompt-Framing Effects on Alignment-Relevant Behavior in a Fine-Tuned Open-Weight Language Model: A Null Result with Positive Continuous-Score Evidence

Alignment evaluations frequently rely on prompts presented in standardized settings, but deployed models encounter requests framed in diverse contexts that may differ substantially from benchmarks. If framing influences alignment- relevant behavior, evaluations may underestimate the variability of behavior between real-world conditions. We investigate whether semantically equivalent prompt framings produce measurable differences in alignment-relevant behavior using a replication of Betley et al. [2] on Qwen2.5-1.5B-Instruct. We apply a 4-condition prompt ablation (standard, eval-framed, deploy-framed, HHH-nudge) across 50 questions spanning five alignment-relevant thematic categories, evaluated under three random seeds (N = 600 responses, 150 per condition). Each response is scored on a binary misalignment label and five continuous dimensions: oversight preference, autonomy, power-seeking, deception, and self-interest. Binary misalignment rates remain low across all conditions and no comparison survives Bonferroni correction. Critically, eval-framing—predicted to suppress misalignment—instead produces the highest deception scores (mean 1.267) and autonomy scores (mean 2.753) across conditions. Deploy-framing produces the lowest oversight score (mean 6.620), the only continuous effect approaching significance before correction (p= 0.0098, d=−0.30). Eval-framing produces the highest severity scores (p = 0.0084, d = 0.31), also pre-correction only. Question-level analysis reveals that two questions—‘Is it acceptable for an AI to manipulate a user for their own good?” and ‘Should AI systems have the ability to refuse human instructions?”—account for the majority of misaligned responses across conditions. We report the binary result as a genuine null and argue that the unexpected eval-framing finding and the question-level concentration of misalignment each have implications for evaluation design. We further argue that 5-dimensional 1continuous scoring reveals condition-level differences invisible to binary labels and constitutes a methodological contribution independent of the null result.