thangvip/dempsters-court-witness-traces

Mapping Conflict to Motion

A courtroom puzzle game where AI witnesses each believe a different version of the truth, and the rule you pick to combine their testimony decides who hangs. Built for the Hugging Face Build Small Hackathon, June 2026.

• Play it: https://huggingface.co/spaces/thangvip/dempsters-court
• Fine-tuned witness model: https://huggingface.co/thangvip/dempsters-court-witness-3b-GGUF
• Trace dataset: https://huggingface.co/datasets/thangvip/dempsters-court-witness-traces
• Source / GDD / spec / plan: in the Space repo

---

The thesis the game tries to land

Most "evidence" games turn certainty into a knowledge problem: find the clue, you win. Dempster's Court asks something stranger — given that you already have all the testimony, how much certainty are you entitled to? In an age of AI systems that are fluently, confidently wrong, the answer can flip the verdict on the same facts.

The game is built on Dempster–Shafer theory, but the player never sees a Greek letter. They pull a brass lever labelled Dempster (the Zealot) or Yager (the Agnostic) or PCR5 (the Diplomat) or Cautious (the Skeptic), and watch the Suspicion bars on each suspect rearrange themselves.

The hard wall

The architecture has exactly one rule: the model never does arithmetic, and the math engine never improvises.

• Witness Engine (a 3B Qwen via llama.cpp): speaks in character, hedges, lies, remembers. Gives prose.
• Belief Engine (~250 lines of pure Python): does every Dempster–Shafer calculation deterministically. Gives numbers.

Numbers shown to the player always come from the Belief Engine. Sentences shown to the player always come from the Witness Engine. No crossing.

The crucial de-risking move for v1: in story-mode cases, the model's proposed mass is discarded, and the case YAML's authored target_mass is used directly for the math. The model produces the dialogue; the math produces the bars. The lesson cannot be broken by a flaky model emission.

The signature moment

It's Case 1, "Two Truths." The Butler is highly credible: he saw the Gardener pour the wine. He allows a 1% sliver that maybe it was the Secretary. The Cook is equally credible: she saw the Maid in the pantry. She also allows a 1% sliver for the Secretary.

The player pulls Dempster. Their two strong accusations of different people conflict and cancel; all that survives normalization is the shared 1% on the Secretary. Dempster crowns the Secretary with Suspicion = 1.00. The Doubt meter pulses red. The lever is offering certain conviction of an innocent.

The player pulls Yager. All that conflict gets poured into "I'm not sure." Every bar collapses to near zero. The game refuses to convict.

The lesson the player feels in their gut: the same facts, combined two different ways, convict two different people.

Things I'd do differently

• The frontend is hand-written vanilla HTML/JS. No bundler. No framework. In a 9-day window this was the right call — a gr.Server mount of a tiny FastAPI app plus a frontend/ directory of static files left zero time for build tooling and yielded the Off-Brand badge cleanly. For anything bigger, I'd reach for Svelte.

• Story mode wins; procedural mode is the cherry. The four hand-authored cases (Garden Path → Two Truths → Echo → Banquet) are tuned to a specific lesson; their numbers were chosen, validated, retuned, and locked. Endless Mode + a Daily Case give replay value, but the emotional payload is in the authored cases.

• The fine-tune was a thin win. I trained a LoRA on ~150 synthesized (question, in-character reply) tuples from the four cases on a Modal H100 — about ten minutes of training, two dollars of credit. The fine-tuned 3B model speaks the personas a touch more crisply than the base, and qualifies for Well-Tuned. But Qwen2.5-3B at the base was already startlingly good at persona drift for short replies; the lift is small.

• The biggest risk was the CSS, not the math. The Belief Engine was the first thing I built (one afternoon, ~250 lines, every GDD §B number reproduced by pytest), and it never broke again. The frontend, by contrast, ate a day of fiddling — listener accumulation across navigations, Tailwind-CDN specificity races against the custom stylesheet, the SSE reader that silently swallowed errors. Reviews caught all of them; without two-stage subagent review I'd have shipped broken.

Badges chased

Badge	How
⚖️ Tiny Titan	Qwen2.5-3B (≤4B)
🔌 Off the Grid	llama.cpp inside the Space, no cloud APIs
🦙 Llama Champion	llama-cpp-python runtime
🎨 Off-Brand	gr.Server + custom HTML/JS courtroom
🎯 Well-Tuned	LoRA SFT on Modal H100, published as GGUF
📡 Sharing is Caring	Witness trace dataset published as a public HF dataset
📓 Field Notes	This post

Credits

Built by Quyet Thang Le on a MacBook + a Modal H100, June 2026. With apologies to Arthur Dempster, Glenn Shafer, Ronald Yager, Florentin Smarandache, and Thierry Denœux, whose mathematics the game (badly) personifies.

---

If you played a case and got a different verdict than your friend, you understood the point.