DIM-WM: Factorized World Modeling for Platform Information Services in AI-Agent Data Markets

Can platform forecasts improve data-market decisions when participants retain final control? We study this question in an auditable simulated data market with LLM buyers and sellers used without further training, whose decisions can respond to observations and accumulated interaction history. Task utility is measured on real classification data. DIM-WM organizes five conditional prediction interfaces: buyer choice (Natural), legal-action outcomes (Controlled), task gain (Technical), 14-day seller outcomes (Core14), and 28-/42-day seller outcomes (Dynamic42). Buyer net technical surplus is realized task gain minus payment on an experimental accounting scale, with nontransactions counted as zero. Predictive performance is target-dependent, and stronger public-input baselines outperform some evaluated components. Fixed-support development selects a quantile LightGBM gain predictor and finds a positive value-information signal using archived action outcomes. In an 80-block online format-control study, individualized descriptive forecasts do not establish a buyer-surplus benefit over same-format training-reference statistics or no advice; both point estimates are negative. A separate 72-blueprint online study gives individualized descriptive forecasts to both advised arms and varies the buyer-value package. Individualizing that package increases mean 42-day buyer surplus by 41.16 accounting units per market relative to reference value information (95% CI [27.68, 55.08]). The full service increases surplus by 46.44 units relative to no advice (95% CI [29.35, 64.68]), with fewer transactions and lower seller gross revenue on average. The results support a buyer-surplus benefit from the evaluated value-information service while distinguishing predictive quality, fixed-state decision value, and online market effects.