Paper

Consumers increasingly encounter AI-generated summaries and comparisons when shopping -- but the text an AI produces is not a neutral transcription of product facts; it can favor one option through positive-valence framing or selective omission of weaknesses. This project introduces BiasMarket Lab, an experimental harness built around two fictional, utility-matched products, so that belief accuracy -- not just stated preference -- is measurable against known ground truth. Before recruiting any human participant, the entire pipeline is validated with a rule-based synthetic-participant population and an estimator-recovery check. Every behavioral number in the current draft is simulated-pilot only, pending real human data collection.

Full text: paper/abstract_250.md.

Contributions

• An experimental harness for measuring belief accuracy, not just stated preference, against known ground-truth attribute values.
• Full stimulus provenance as a first-class, publicly inspectable artifact (see /stimuli), not an appendix footnote.
• A synthetic-participant validation layer that runs before any human is recruited, including an estimator-recovery check.
• An explicit, separately labeled bridge from individual-level effects to a market simulation that never claims to measure a real platform.

What is not claimed as a contribution: framing effects, disclosure/unraveling theory, and LLM persuasiveness are each independently well-established in prior literature. The claim to novelty here is the specific combination and the tooling, not the underlying mechanism.

Full text: paper/contributions.md.

Other manuscript files

• paper/abstract_250.md250-word draft abstract, explicitly marked SIMULATED PILOT / NO HUMAN DATA.
• paper/abstract_150.md150-word variant of the same draft abstract.
• paper/contributions.mdThe project's claimed contributions, and what is explicitly not claimed as novel.
• paper/claims_to_evidence.mdThe evidentiary bar each future results sentence must clear before it can be written.
• paper/reviewer_attack_surface.mdAnticipated reviewer objections and how the design addresses (or concedes) them.
• paper/main.texLaTeX manuscript skeleton, assembling the section files below.