From Recipes to Transitions: An Observability Audit for Materials World Models
Abstract
Learned material-transition dynamics require linked observations of state, intervention, and subsequent state, whereas literature-mined synthesis corpora primarily encode intended targets, precursors, operations, and conditions. We formalize this mismatch as an observability contract and audit two pinned public solid-state synthesis corpora (19,488 and 80,806 records); the impurity-aware corpus adds 18,869 reported impurity-bearing outcomes. Although neither released representation exposes a schema-level linked transition, manual source verification (n = 30) found evidence satisfying our frozen action-state-linkage definition in 12/22 general-stratum records and all five adversarial records included in that field-specific diagnostic. Adversarial findings are diagnostic only, and neither stratum is extrapolated to corpus prevalence. A matched 100-candidate negative control probes recipe-level retrieval under training overlap. Conclusions remain scoped to these two released representations. Submission category: Findings, Tools & Open Challenges.