What may a language model do in SDX?

The LLM Boundary defines what a language model may and may not do in SDX.

A language model may verbalize controlled report data, improve readability, and produce human-facing phrasing within an explicitly bounded input package. It may not create analytical meaning, decide between options, aggregate evidence, rank alternatives, add unsupported claims, weaken limitations, or ignore non-use boundaries.

This boundary exists because language models are powerful presentation tools but unsuitable as sources of governance authority. SDX separates language generation from evidence generation, statement governance, and institutional decision-making.

The LLM Boundary protects the system from a common failure mode: turning structured evidence into persuasive but unverifiable prose. In SDX, the model is constrained to expression. It is not an analyst, decision-maker, verifier, or authority layer.