The %slm_* helper library makes calling the offline model a few lines — and the IT story is, paradoxically, the easiest of any AI option.

The worked example (a morning of on-device triage)

SAS/R + Pinnacle 21 produce 214 findings (every count SAS/R's). The local model labels each new-vs-known, routes by owner, and drafts a candidate query — as schema-constrained JSON. The validator drops anything off-list; the biostatistician approves; SAS/R writes the queries + a Part 11 audit line with the model digest.

The IT paradox. On data-egress grounds this is the easiest AI to approve — nothing leaves the box; you prove it by pulling the network cable. The hardest honest constraint is model capability — so scope tightly to short, validated language tasks.

Anticipated hurdles (each clearable)

"Unknown binaries" → signed release / internal mirror; air-gapped model transfer + checksum.
"No GPU budget" → CPU-first; a 7B Q4 model runs on a normal workstation.
"Is it validated/deterministic?" → frozen model + temp 0 + GAMP-5 CSV + human gate; the SLM is decision-support, never the record.
"Licensing" → prefer Apache-2.0/MIT models (Granite, Phi, Qwen Apache sizes).