Position: Future Context Is More Than a Covariate in Time-Series Foundation Models

Covariate-aware time-series foundation models (TSFMs) now accept calendar features, external forecasts, planned events, retrieved signals, multimodal evidence, and actions through general interfaces that span several data types and sources. A single future-covariate channel can therefore receive a holiday date, a weather forecast, a revisable promotion plan, or a treatment action, even though these inputs support different statistical and causal claims. Our position is that future context in TSFMs should be treated as typed context rather than as a homogeneous class of covariates, organized along two independent dimensions: its epistemic status at forecast issue time and its causal role in the data-generating process. On this basis, we propose a future-context contract for training, evaluation, and deployment that records provenance, issue time, epistemic form, causal role, dependence, and revision or execution behavior. We then connect input semantics to a Temporal Evaluation Ladder spanning forecasting, simulation, intervention, and decision support. Each level corresponds to a stronger claim about an evolving system and therefore requires additional evidence. The resulting framework gives concrete guidance for benchmark construction, forecast-vintage preservation, probabilistic calibration, LLM-based forecasting agents, and action-conditioned temporal models across operational domains.