Hard AI "Cheating" Bug Fixed (v0.09)

The Hard AI in Battle Hanafuda is designed to make decisions based only on information that is openly available to it.

Specifically, it uses the following data:

• Its own hand

• Captured cards (both its own and the opponent’s)

• Cards on the field

On the other hand, the following “hidden” information is never accessed directly:

• The opponent’s hand

• The remaining deck

Instead, these are handled through Monte Carlo-based estimation, allowing the AI to achieve high performance while maintaining fairness.

However, a recent static analysis conducted using generative AI (Codex) revealed a critical issue: part of the algorithm was directly referencing hidden information that should not have been accessible (effectively, cheating).

This update fixes that major bug.

As a result, the Hard AI now achieves a 57.6% win rate against Normal difficulty (down 0.7% from the previous version).

To prevent similar issues in the future, we are making both the AI algorithm and the host simulator,"Open Hanafuda AI", open source. This will allow third-party verification and ensure transparency in how the AI operates.

This ensures that the AI does not engage in any unfair behavior.

【Open Hanafuda AI】

https://github.com/suzukiplan/open-hanafuda-ai