README.md

metadata

title: Tiny Court of Everyday Crimes
emoji: ⚖️
colorFrom: yellow
colorTo: red
sdk: gradio
sdk_version: 6.16.0
app_file: main.py
python_version: '3.12'
pinned: true
license: apache-2.0
short_description: Multimodal comedy courtroom — text, photo & voice on trial.
datasets:
  - build-small-hackathon/tiny-court-traces
tags:
  - build-small-hackathon
  - gradio
  - custom-ui
  - comedy
  - agent
  - track:backyard
  - track:wood
  - sponsor:openbmb
  - sponsor:openai
  - sponsor:nvidia
  - sponsor:modal
  - achievement:offbrand
  - achievement:llama
  - achievement:sharing
  - achievement:fieldnotes

Tiny Court of Everyday Crimes

Ace Attorney for stupid everyday problems.

An interactive comedy courtroom where users put small, ridiculous, everyday conflicts on trial — snack thefts, suspicious pet behavior, object betrayals — and receive charges, evidence, witnesses, objections, a verdict, and an absurdly harmless sentence. A Gradio app backed by four small open models served on Modal, accepting text, photo, and voice evidence.

This is a comedy experience, not legal advice.

Built for the Build Small Hackathon (Gradio × Hugging Face).

• Live app: https://huggingface.co/spaces/build-small-hackathon/tiny-court
• Video demo: a narrated walkthrough of a full case
• Open traces: https://huggingface.co/datasets/build-small-hackathon/tiny-court-traces
• Field notes / blog post: https://johnz86.github.io/blog/tiny-court-and-the-urge-to-outsource-judgment.html
• Narrated overview (video): Tiny Court and the urge to outsource judgment

Watch a trial

▶️ Watch the narrated demo on YouTube — a full case, start to finish.

The loop above is a sped-up run of one full trial — complaint → evidence → witness → cross-examine → twist → verdict → leniency appeal → the shareable Court Record — recorded against the live app on its real four-model backend. (download the MP4)

How it works

A landing page (rotating hero docket + Quick Trial / Full Trial CTAs) opens into one growing courtroom conversation: you file a complaint and the court reacts; optional moves — submit evidence, call a witness, cross-examine, object, add a twist — append to the same transcript. Deliver the Verdict is available from the first message and swaps to a printed ruling; you can appeal back into the conversation, then finalize to a shareable Court Record (with a one-click PNG export).

The verdict is Python-owned and deterministic — the model proposes meter deltas, the engine resolves the band — so interactions provably move the needle. Safety is gated both ways: a layered gate on input and an offline scrub on generated output.

The model stack

Generation runs through a pluggable seam. In production the Space talks to four small open models served on Modal over OpenAI-compatible endpoints — each modality is perceived into text before the judge reasons over it, so a weak perception model never owns the verdict:

Role	Model	Served via	Sponsor
Judge (reasoning + fields)	NVIDIA Nemotron-3-Nano-4B	llama.cpp on Modal	NVIDIA
Vision (photo evidence)	OpenBMB MiniCPM-V-4.6	llama.cpp on Modal	OpenBMB
Formatter (schema repair)	JetBrains Mellum2-12B-A2.5B	llama.cpp on Modal	JetBrains
ASR (voice evidence)	NVIDIA Parakeet-TDT-0.6B	NeMo on Modal	NVIDIA

All models are under 32B; the judge/vision/formatter run on llama.cpp. The app degrades gracefully — if Modal is unreachable it falls back to a deterministic canned backend so the UI always runs.

Agent traces

Every model call is captured as a redacted JSONL agent trace (the deterministic verdict meters snapshotted alongside each call) and published to build-small-hackathon/tiny-court-traces — no raw images, prompts, or tokens.

Field notes

A short write-up on building Tiny Court — and why a comedy courtroom is a careful frame for not outsourcing judgment to a model: Tiny Court and the urge to outsource judgment. A narrated overview of the piece is on YouTube.

Docs

Full docs live in the source repo (not shipped to the Space):

• Documentation index — start here.
• Design spec — product & UX.
• Model-serving decision — the multimodel Modal backend + client orchestration.
• Agent traces — the trace format + publish flow.
• ADRs — architecture decisions.

Quick start

uv sync
uv run python main.py                         # fake backend (no GPU, no network)

# against the live Modal models:
#   set the TINYCOURT_MODAL_*_URL vars (see docs/hf-space-configuration.md), then
#   TINYCOURT_BACKEND=remote uv run python main.py

Requires Python 3.13 locally.