Gemma 4 Demo App — Test Artefacts & LLM Wiki Package for Antigravity

Source code, prompts, and outputs accompanying two YouTube videos.

Videos

Video	Link
Gemma 4: 8 Real-World Tests — the demo app, prompts, and model outputs	https://youtu.be/WcdFG1Onj5U
LLM Wiki — a living, agent-maintained wiki for any codebase (uses this repo as the demo)	https://youtu.be/J3szwLZ6YfI

The .agents/ directory contains the portable LLM Wiki package (skills, workflows, rules) that any Antigravity user can drop into their own repo. The wiki/ directory is the generated output. See .agents/README.md for details.

---

What's in this folder

demo-app/
├── .agents/            # Portable LLM Wiki package (skills, workflows, rules)
├── wiki/               # Generated LLM-maintained living documentation
├── index.html          # The demo app (single HTML file, no build step)
├── start.command       # macOS double-click launcher
├── README.md           # This file
├── PROMPTS.md          # All 8 test prompts, copy-paste ready
└── outputs/            # Model responses for each test
    ├── response_test6_multifile.txt
    ├── response_test7_reasoning.txt          # Gemma 31B, thinking OFF
    ├── response_test7_reasoning_gemmathinking.txt  # Gemma 31B, thinking ON
    ├── response_test7_reasoning_gemini.txt   # Gemini 3.1 Pro
    ├── response_test8_json_26b.txt
    └── response_test8_json_31b.txt

Tests 1–5 outputs are shown as screenshots/images in the video. Tests 6–8 full responses are in outputs/.

---

Running the Demo App

Requirements: A Google AI Studio API key (free at aistudio.google.com)

API Key setup:

cp env.txt.example env.txt
# Edit env.txt and replace "your_api_key_here" with your actual key

The app reads env.txt at startup and pre-fills the API key field. Alternatively, paste your key directly into the key field in the UI. env.txt is gitignored so your key is never committed.

macOS:

# Option 1: Double-click start.command
# Option 2: Terminal
cd demo-app
npx serve . -p 4000
# Open http://localhost:4000

Windows / Linux:

cd demo-app
npx serve . -p 4000
# Open http://localhost:4000

Enter your API key in the top-right field. Select a model, mode (Text or Image), and optionally enable Thinking Mode.

---

Models Tested

Model ID	Type	Notes
gemma-4-31b-it	Dense	All params active per token
gemma-4-26b-a4b-it	MoE	Only 3.8B active per token
gemini-3.1-pro-preview	Proprietary	Used for Test 7 comparison
gemini-3-flash-preview	Proprietary	Used for hook comparison

---

Test Results Summary

Test	Description	Result
1	JSON Extraction	✅ Both models — clean, correct
2	Single Function Coding	✅ Both models — working code first try
3	Simple Reasoning	✅ Correct step-by-step
4	OCR & Chart Reading	✅ Surprise pass — all 72 data points correct
5	Image Q&A (Architecture)	✅ Surprise pass — identified SPOF unprompted
6	Multi-File Code Generation	⚠️ Passed with caveats — see notes below
7	Deep Reasoning (Logic Puzzle)	✅ Surprise pass — found both valid solutions
8	Complex Nested JSON	✅ 100% correct, both models

Tests we expected to fail — but didn't

Tests 4, 5, 7, and 8 were designed expecting Gemma to struggle. It didn't. The video tells the honest story — results updated in real time as we ran each test.

Test 6 caveats (passed, but with code quality issues):

• Pydantic V1/V2 mixed syntax (class Config vs ConfigDict)
• Deprecated FastAPI @app.on_event("startup") handler
• Simplified auth — username used directly as Bearer token
• These are real-world issues but not correctness failures

---

Prompts

See PROMPTS.md for all 8 prompts, copy-paste ready.