Proof Before Price: Pre-Registered Behavioral Tests of Proof-Based Pricing on Language-Model Buying Agents

Source abstract

Pricing theory assumes a buyer who can be persuaded. When the buyer is a delegated languagemodel agent, a seller's claims reach an evaluation procedure, and the pricing question is whether that procedure admits an offer, rewards what it can verify, and reads a price schedule differently from a posted price. This paper reports, to the author's knowledge, the first pre-registered, frozen-code test of verification, legibility, and schedule effects on language-model buying agents: three hosted frontier models (gpt-5, gemini-3.6-flash, claude-sonnet-5) and one openweight model (Qwen3.6-27B) run locally. Agents chose among twelve synthetic offers, or none, in a frozen marketplace under five yoked manipulations; of 168,816 records, 166,389 were analyzed with code frozen before any scored query. Three hypotheses were registered (OSF, DOI 10.17605/OSF.IO/GPBUK): admission as a legibility threshold, a verified premium, and a schedule effect at equal price. The verified premium and the schedule effect hold on every agent © 2026 Paul F. Accornero. Licensed under CC BY-NC-ND 4.0. 2 at the registered levels. An offer whose claim can be checked in a register is selected 14 to 30 percentage points more often than the same claim in prose; a missing register identifier costs 15 to 29 points; an executable schedule stating its terms is selected 7 to 33 points more often than the same price posted. One registered sub-prediction, stated from the design of the channels before any such measurement existed, is not supported: the registration expected a machineattested claim to be selected at or above a certified one, and on three of the four agents the attested claim is selected 6 to 12 points less often, with no detected difference on the fourth. This paper reports that result as the field's move from observation to quantification. The admission threshold does not hold: the labels-only offer is considered every time and selected about as often as the structured one, with selection falling smoothly with price, the registered refuting observation. Effect sizes differ by agent; the local model's are the smallest on the verification and schedule members, though not on the labels-only contrast or the chatter arm.