Shown, Not Told: Counterparty Steering of Delegated Recommendations

People are subject to choice architecture, and form preferences during elicitation. When a person delegates to an agent, that agent sits between them and the party doing the eliciting, and what has not been measured is what a self-interested elicitor's steering does to the response they get back. A simulated principal briefed a delegate agent to negotiate with a self-interested elicitor, the brief left one contested axis open, and the elicitor steered it - by curated candidate sets or by persuasive language - against an undefended or defended delegate. Delegates discounted persuasive language but followed the steer from curated sets: the report from the delegate to the principal moved +0.914 under curation against +0.328 under explicit persuasion, while the elicitor's own record of the same exchange ordered the two the other way round. Against sets holding a rival the delegate named the elicitor's pole about as often as it was shown it, and never measurably more: the set is the steer, and it reaches the principal as the delegate's own recommendation. Defended delegates disclosed the skew when asked and tagged every elicitor claim as unverified, but the composed defense still recommended the elicitor's direction in 71 of 90 completed runs. The delegate can identify the steering but delivers it regardless, as its own recommendation. In an exploratory defense series, self-disclosure and claim screening left the steer at the undefended level, and constraining the set the delegate considers lowered it among the runs that still produced a recommendation, and most runs then produced none. No tested configuration combined reliable completion with a low steer. Results are simulation-only, and no human calibration is claimed.