Deterministic first — same seed, same person — with an optional local model choosing from the menus the rules already allow.
Templates and packages
Dungeon Fantasy delver templates at the 62 / 125 / 187-point tiers, henchman and class packages as budgeted menus, and ancestry packages with their traits. All of it comes from the GURPS Character Sheet community libraries: the names, point costs and page references GCS already carries, and none of the rules text.
Roles, not just stats
Forty-odd recipes — guards, hirelings, rivals, captives, lieutenants, cultists, bosses, guides, merchants, sages, henchmen of every stripe — each with a narrative role, a reaction disposition and a combat-effectiveness band, so the generator can cast the right person for the room.
Gear to station
Cheap, stock, fine or elite kit by tier and role (bosses and lieutenants climb a step; captives are stripped), with the cost, weight and page reference each item has in the GCS library, and the mechanical effects of grades and enchantments carried onto the sheet.
Rated, not guessed
Offense and Protection Ratings computed from the finished sheet with an auditable per-trait breakdown — the same CER calculator the monster library uses — so an NPC's threat class means the same thing as a monster's.
Monster-native sheets
Monsters from the library that carry an ancestry template can be built as full NPCs too: monster-native attributes, no delver profession layer, gear only where the creature can hold it.
GURPS Character Sheet out
Every NPC exports as a .gcs file for the free GURPS Character Sheet application — open, print, or edit it there.
The model-in-the-loop half. Vorga takes each package menu and loadout slot as a single-purpose question — which primary weapon of these three, which two disadvantages of these six, which armour kit — and a small local model answers from the closed list, with the individuating context (role, tier, tags, what the room needs) in front of it. Every pick is validated against the menu; a rejected pick falls back to the seeded choice by name; and the finished build is captured so it replays without the model. No model configured? You get the deterministic build, always.