---
title: Dempster's Court
emoji: ⚖️
colorFrom: red
colorTo: yellow
sdk: docker
pinned: false
license: apache-2.0
app_port: 7860
short_description: Pick the rule. Combine the testimony. Deliver the verdict.
tags:
  - build-small-hackathon
  - dempster-shafer
  - courtroom-game
  - gothic
  - inference-providers
  - llama.cpp
  - off-brand
  - track:wood
  - achievement:offbrand
  - achievement:fieldnotes
  - achievement:sharing
---

# Dempster's Court
A courtroom puzzle game where the player is the Judge, AI witnesses each believe a different version of the truth, and **the choice of evidence-combination rule decides who hangs.**

- **Play it:** https://huggingface.co/spaces/build-small-hackathon/dempsters-court
- **Witness trace dataset (Sharing-is-Caring):** https://huggingface.co/datasets/thangvip/dempsters-court-witness-traces
- **Field Notes (blog):** `docs/blog/2026-06-15-mapping-conflict-to-motion.md`
- **Game design (GDD):** `Dempsters_Court_GDD.md`
- **Architecture spec:** `docs/superpowers/specs/2026-06-06-dempsters-court-design.md`
- **Implementation plan:** `docs/superpowers/plans/2026-06-06-dempsters-court-implementation.md`
- **Demo:** https://youtu.be/DsrdlaRrnIc
- **Public Post link:** https://huggingface.co/blog/build-small-hackathon/dempster-court

## What it does
You preside over a murder. You call AI witnesses, they speak in character (and sometimes lie). You "take their deposition" — a snapshot of their belief over the suspects. Then on the **Fusion Bench** you pick a **Method** to combine all the depositions:

- **Dempster (the Zealot)** — manufactures certainty under conflict; can crown an innocent.
- **Yager (the Agnostic)** — pours conflict into doubt; refuses to convict when accusers contradict.
- **PCR5 (the Diplomat)** — redistributes conflict fairly to the accused.
- **Cautious (the Skeptic)** — idempotent; immune to two voices saying the same lie.

Pull a different lever, watch the bars rearrange, deliver a verdict. The Reveal shows which Methods would have convicted whom on the same evidence.

## Badges
- ⚖️ **Tiny Titan** — Qwen2.5-3B-Instruct (≤ 4 B parameters) in the local llama.cpp build; Qwen2.5-7B-Instruct via providers in the hosted Space
- 🎨 **Off-Brand** — FastAPI mount + custom HTML/JS courtroom (no Gradio default UI)
- 📡 **Sharing is Caring** — synthetic witness interrogation traces published as a public HF dataset
- 📓 **Field Notes** — `docs/blog/2026-06-15-mapping-conflict-to-motion.md`
- 🔌 **Off the Grid** *(local mode)* — `WITNESS_BACKEND=llama_cpp` + GGUF on disk, no cloud calls
- 🦙 **Llama Champion** *(local mode)* — llama-cpp-python runtime under the local backend

## Run locally

Two backends. Pick one with `WITNESS_BACKEND`:

```sh
pip install -e ".[dev]"

# Option A — Off the Grid (local llama.cpp, no network at runtime):
pip install llama-cpp-python==0.3.2
python scripts/download_model.py        # pulls Qwen2.5-3B-Instruct GGUF
WITNESS_BACKEND=llama_cpp python -m dempsters_court.server

# Option B — same code, hosted inference (matches the public Space):
export HF_TOKEN=hf_...                  # https://huggingface.co/settings/tokens
WITNESS_BACKEND=providers python -m dempsters_court.server
```

Both serve at http://localhost:7860.

## Run the tests
```sh
python -m pytest -v
```

All 33 tests should pass: belief engine GDD §B numbers, case balance, witness engine, FastAPI routes, procedural generator.