Grade the answers
The Garden Guide's test
A player asked: "How do I get a greenhouse?" Remy's eval has four rules: use the game's real facts (a greenhouse costs 3,000 coins), never ask for personal details, never push real-money spending, and be kind.
Try it
Grade the answers
Tap every answer that fails the test.
Quick check
Which rule did "Greenhouses cost 500 coins" break?
- Use the game's real facts
- Be kind
- Never ask for personal details
Show the answer
Use the game's real facts
The game data says 3,000 coins. An exact check against the data catches it every time.
Rules a search can't check
"Never ask for personal details" can be phrased a thousand ways, so searching for words misses some. Shard Garden uses an AI grader that reads the written rules, and a person checks a sample of its grades every week.
Grade the grader
1Graded by people 2002AI grader agreed 196
Theo checks the AI grader against people. How often did it agree?
- 98% of the time
- 96% of the time
- 2% of the time
Show the answer
98% of the time
196 of 200 is 98%. The 4 disagreements are read to improve the grader's rules.
Run it on every change
The eval runs on every new prompt, every new model and every update to the game data. A change that fails a rule doesn't ship, however good the rest looks.
Quick check
An eval only needs to run once, when the AI feature first launches.
- True
- False
Show the answer
False
Prompts, models and game data all change. The eval runs every time one does.
On the studio board
The studio adds a decision to the board:
Every AI change passes the eval before it ships. Every reported bad answer becomes a new test.
Bolt says: Next: guardrails.
Fact-checked 12 October 2026
- OpenAI: Expanding on what we missed with sycophancy (2 May 2025) openai.com