What is Grounded Inference?

Grounded inference is the discipline of producing answers only from real data, with every answer linked to the inputs it depends on. The opposite — ungrounded inference — is what modern LLMs do by default: produce plausible text whether or not the underlying knowledge exists.

For building operations, this distinction is non-negotiable. A confident but invented forecast is worse than no forecast — it leads to confident but wrong decisions. Grounded inference enforces three things:

Traceability. Every output cites the sensors and time-ranges that produced it. Deterministic guardrails. If the required data isn't there, the system says "unknown" — not a guess. Auditability. The reasoning chain is replayable, not lost in a black box.

In Explore, grounded inference is enforced at the architecture level, not as a prompt-engineering hack on top of a general model.