Beyond Self-Resolution: Settlement Factorization for Robust Natural Language Mechanism

Abstract

Language models increasingly mediate paid advice: agents submit open-ended forecasts, recommendations, plans, and evidence; a principal acts on the reports; and the mechanism later pays the contributors. Advice should influence the public decision, but no adviser should write the answer key used to evaluate it. We formalize the separation as settlement factorization: reports are hardened into official records, a public decision record Z may use all advice, and each paid adviser is scored against a label whose production is externalized from their own report, conditional on Z. The central result is an analogue of the revelation principle for this setting: resampling the paid report from a committed ghost distribution inside the settlement channel equips every mechanism with an influence-free reference within a factor of two of the best achievable, so factorization is a normal form, and own-report leakage varepsilon -- measured in total variation -- is an intrinsic invariant of any mechanism, identified behaviorally by worst-case incentive erosion. The invariant has an exact price: with payment kernels of label sensitivity L, truthful margins degrade by at most 2L varepsilon, and the constant is tight -- half own-label manipulation, half pandering to a biased evaluator -- so faithful advice survives outside decision interests D whenever the externalized margin satisfies gamma >D+2L varepsilon. Settling a growing crowd on the shared decision record dilutes every margin exponentially in the informational redundancy, while a factorized leave-one-out label sustains a constant margin at unit stakes. Stakes outbid decision interests but never evaluator capture; randomized reference settlement makes integrity a threshold good priced per unit of total variation; and differential privacy of the evaluator in the paid report certifies varepsilon by construction.