Consent Boundaries by Play: A CI-Grounded Micro-Game for Eliciting Stable Preferences and Diagnosing Opt-Out Friction in Web Consent

Consent on the Web is commonly mediated by Consent Management Platform (CMP) banners, where small interface changes—button prominence, sequencing, first-layer availability of refusal—can materially steer user choices. We present a Contextual-Integrity (CI) grounded "micro-game"—a rapid, repeated choice task that elicits individual consent boundaries across five dimensions (recipient, purpose, data type, retention, user controls/oversight) with an explicit Neither/Ask later option. A heteroskedastic hierarchical Bayes multinomial logit (HB–MNL) model separates three components: (a) stable individual preferences, (b) UI-induced choice noise via arm-specific scales, and (c) selective opt-out friction via an arm-level penalty on Neither. In a two-session study (Session 1 N=168; Session 2 N=120; 32 tasks/participant) crossing three UI arms (Neutral, Benign, Minimal dark pattern) and a brief factsheet intervention, we find that recipients and retention dominate acceptance, while concrete protections—especially Delete/portability and Independent audit—deliver the largest gains. The dark-pattern arm imposes measurable opt-out friction and raises scenario-level acceptance by +3–7 percentage points; in a CMP-banner mimic block it raises acceptance by +5.4% over Neutral, and individual Neither propensity aligns with CMP Reject-all (Spearman ρ=0.49). Preferences are robust across Neutral and Benign arms but sensitive to friction-based refusal gating; we do not claim coverage of the full design space of CMP dark patterns (e.g., color manipulation, preselection, confirmshaming). Hold-out accuracy is 0.71; test–retest agreement over one week is moderate (ICCs ~0.55–0.72). Short Bayesian Rule Lists approximate model predictions (AUC ~0.85, median length 4), yielding interpretable "consent grammars" for parity-preserving CMP deployment.