The Ofrenda Witness — a ghost that may only say what the dead man remembers

Description and use

Ofrenda Witness is an offline, browser‑based murder mystery where the ghost can only speak what the victim remembers, ensuring fairness and privacy.

Problem

Build a fair, offline, browser‑only murder mystery where the ghost can only say what the victim actually remembers.

Outcome

The Ofrenda Witness delivers a fair, offline murder mystery where the ghost’s statements are strictly limited to the victim’s memories, ensuring consistency, privacy, and reproducibility.

Workflow

• 1. User opens the web page in Chrome or Edge with WebGPU.
• 2. The first load downloads ~0.8 GB of model weights into the browser cache.
• 3. User selects a model (Qwen2.5‑1.5B‑Instruct, Llama 3.2 1B, or Qwen2.5 0.5B).
• 4. User places a keepsake on the ofrenda to provide the ghost with that memory.
• 5. The system prompt contains only the memories tied to the placed keepsakes.
• 6. The user asks questions; the model generates a reply.
• 7. verify.js checks the reply for disallowed content (time, capitalised name, unknown clue objects, >3 sentences).
• 8. If the reply fails, the model is asked again up to twice; otherwise it falls back to quoting the closest memory word‑for‑word.
• 9. The reply is displayed with a note indicating whether it was checked, retold, or quoted.
• 10. The player uses the information to deduce the murderer and accuse them.

Evidence

The ghost is Qwen2.5-1.5B-Instruct (Apache-2.0), running in the page through WebLLM on WebGPU.

Source

Every reply is checked before you see it. verify.js rejects a reply that contains a time, a capitalised name, or a clue object…

Source

The first load downloads about 0.8 GB of model weights into your browser cache; after that it runs offline.

Source