Archive Guard RPG

Archive Guard is a browser text RPG where the dialogue is driven by a locally hosted language model rather than a hosted API.

The interesting part was not the model — it was making an unreliable component safe to build on. Structured prompting narrows what the model is allowed to return, and a deterministic template takes over when it returns something unusable, so the game never dead-ends on a bad generation.

Evidence note: this description follows the project's own write-up. The source tree and a runtime record are still being recovered, so nothing here is presented as a measured or deployed result.

• Built a React front end over a Node/Express API with MongoDB persistence and Redis-backed sessions.
• Implemented authenticated accounts and per-user conversation history.
• Compared a baseline prompting mode against a structured mode on the same local model.
• Added a deterministic template fallback so failed model output degrades instead of breaking the run.

Prompting modes

Baseline and structured prompting run against the same local model, with a deterministic template fallback when model output fails.