---
title: "First-Principle AI · Gotcha Arena"
emoji: "🎯"
colorFrom: indigo
colorTo: green
sdk: gradio
sdk_version: "6.14.0"
python_version: "3.12"
app_file: app.py
fullWidth: true
header: mini
short_description: "Trick-question arena for a small first-principle reasoner"
suggested_hardware: zero-a10g
models:
  - build-small-hackathon/phase-3-gguf
tags:
  - gradio
  - build-small-hackathon
  - track:wood
  - sponsor:nvidia
  - achievement:offbrand
  - achievement:offgrid
  - achievement:sharing
  - reasoning
  - nemotron
  - gguf
  - llama-cpp
  - zerogpu
license: mit
---

# 🎯 First-Principle AI · Gotcha Arena

**A small reasoning model that reads the question, not the pattern.**

Big models get fooled by familiar-looking trick questions — a riddle with one word
swapped, a decoy distance, a false premise. *First-Principle AI* is a compact
~31.6B Q8 MoE (Nemotron-H) reasoning model built for this hackathon, and this Space
is a playground for testing exactly that skill.

## What you can do here

- **🎯 The Gauntlet** — pick a trick question from the deck (or hit *Surprise me*),
  run it, and see whether the model reasons past the trap. A live scoreboard tracks
  how many traps it caught vs. fell for, and reveals *why* each one is a trap.
- **🧠 Trap It Yourself** — two modes:
  - *Trap the AI*: write your own trick question and try to fool it.
  - *Blind-Spot Check*: paste any real plan or "should I X or Y?" decision and the
    model surfaces the hidden assumption you might be missing. (This is the practical
    use case — the same first-principle reasoning, pointed at your own decisions.)
- **⚙️ Engine Room** — sampling controls and live llama.cpp runtime diagnostics.

The trap deck mixes the model's signature **goal-binding** puzzles (the object that
must physically move) with classic gotchas: the Moses illusion, the bat-and-ball
cognitive-reflection test, unit misdirection, false-premise questions, and
language-parsing traps.

> The auto-verdict in the Gauntlet is a lightweight keyword heuristic — the real
> proof is reading the model's answer. It's there to keep score, not to grade.

## The model

| | |
| --- | --- |
| Repo | [`build-small-hackathon/phase-3-gguf`](https://huggingface.co/build-small-hackathon/phase-3-gguf) |
| Base | [`owenisas/nemotron-3-nano-reasoning`](https://huggingface.co/owenisas/nemotron-3-nano-reasoning) (merged) |
| Architecture | Nemotron-H hybrid **MoE**, ChatML chat template |
| Size | `model-Q8_0.gguf`, ~31.6B params (**under the 32B cap**) |
| Context (this Space) | 2,048 tokens (`n_ctx`); the GGUF reports a larger architectural max, but the Space serves 2K for fast, memory-safe startup |
| Runtime | official `llama.cpp` `llama-server` (release `b9360`) on ZeroGPU |

Because it's a Mixture-of-Experts model, only a small slice of the 31.6B parameters
is active per token, which is what makes a model this capable run on a small footprint.

## Runtime notes

- The 33.6 GB Q8 GGUF is **not** preloaded during build (it would make startup
  unreliable). It downloads and loads through `llama-server` on the **first prompt**,
  which takes about 1–2 minutes; every prompt after that reuses the warm server.
- ZeroGPU is primarily documented around PyTorch workloads, so this app runs the GGUF
  through the official llama.cpp CLI path instead of depending on a Python extension
  compile at build time. If the runtime can't provide enough memory or a compatible
  binary, the app shows a visible diagnostic instead of a silent mock answer.

## Hackathon submission — Build Small

**Track:** 🌲 Thousand Token Wood (`track:wood`) — a whimsical, AI-native game.
**Sponsor model:** NVIDIA Nemotron (`sponsor:nvidia`).
**Badges targeted:** Off Brand (custom UI), Off Grid (fully local llama.cpp inference),
Sharing (social post).

- [x] **Gradio Space inside the org** — `build-small-hackathon/First-Principle-AI`
- [x] **Model ≤ 32B params** — ~31.6B Q8 MoE (Nemotron-H)
- [x] **Track + badge tags in YAML** — `track:wood`, `sponsor:nvidia`, `achievement:offbrand/offgrid/sharing`
- [x] **Write-up of idea + technical approach** — this README
- [ ] **Demo video** — `TODO: add a public video link (YouTube / Loom / Space) showing the app running`
- [ ] **Social post + link** — `TODO: post once about the build and paste the URL here`
- [ ] **Team HF usernames** — `owenisas` *(add teammates if any; each must register + join the org)*

> The three `TODO`s need a human (record the video, make the social post, confirm
> teammates), then paste the links above. Re-run the field-guide validator after editing:
> https://build-small-hackathon-field-guide.hf.space/submit

## Local smoke test

```bash
cd hf-spaces/phase-3-gguf-lab
PHASE3_DISABLE_MODEL=1 python app.py   # UI only, no model load
```