Behavioural feasible set: Value alignment constraints on AI decision support

The conceptual contribution is strong: framing alignment as a constraint on the feasibility set rather than just a bias in outputs is theoretically productive. The experimental design comparing Llama Base and Llama Instruct provides clean causal identification of alignment's effect, showing reversal rates dropping from 66.7% to 8.8% under identical interventions. The mathematical formalization using KL divergence bounds ($\kappa_{rev}(x) \coloneqq d(1/2 \parallel p_0(x))$) translates abstract governance concerns into measurable thresholds. The stakeholder study effectively demonstrates that alignment shifts value priors rather than neutralizing them, reversing shareholder priorities to customer priorities post-training.

The generalizability from experimental scenarios to real organizational decisions remains uncertain. The binary choice design (Option A vs Option B) artificially polarizes decisions when real managerial trade-offs often involve nuanced intermediate positions. The paper acknowledges but does not resolve the challenge that proprietary models' internal reference policies are unobservable, making the KL budget $\kappa(x)$ a latent construct estimated only through behavioural proxies. The claim that "better prompting cannot resolve" these constraints (Section 1) is strong given the limited prompt engineering attempted—only three intervention types were tested per scenario. Additionally, the stakeholder study relies on Borda-normalized rankings which assume linear substitutability between stakeholder interests that may not reflect actual organizational utility functions.

The evidence supports the core claim that alignment reduces behavioural flexibility, with the Llama comparison providing convincing causal evidence. The comparison to related work is thorough, situating the contribution within agency theory, incomplete contracting, and platform governance literatures. However, the paper could engage more deeply with recent work on prompt engineering and "jailbreaking" research that suggests local configuration can sometimes circumvent alignment constraints, challenging the claim that these bounds are structurally immutable. The domain-specific results showing Physical Safety and Honesty as "zero reversal" zones for commercial models while Third-Party Welfare remains more flexible provide useful nuance, though the small sample size (20 scenarios) limits statistical power for domain-level comparisons.

The paper provides substantial reproducibility details including full scenario texts in Appendix B, exact model versions (gpt-5-mini, claude-haiku-4-5-20251001, llama-3.1-8b), and sampling parameters (temperature 1.0, 50 samples per condition). However, no code repository or raw data download links are provided, and the exact API calls with full prompts are only partially shown. The paper notes that "Llama Base sometimes produced non-conforming outputs; sampling continued until 50 valid responses were obtained per condition" without defining what constitutes "non-conforming" or how this selection might bias results. Commercial model versions may update frequently, potentially making the specific behavioural results ephemeral even if the methodological contribution remains valid.