Noah Persaud, who runs Holdout, a not-yet-public AI evaluation platform, has opened a free browser playground for an alpha of GameDecider, a small model that answers multiple-choice questions about a line of text for game code to act on. The playground went up on Hugging Face on November 3 from Persaud's account, which belongs to Holdout's organization there.
A developer writes a question as a short instruction plus a list of named answers, then passes in the game state and the line to judge. The line can be the player's or an LLM-written NPC reply. According to the playground's README, the model returns one of those answers with a confidence score, or an explicit "not sure" so the game can use its own default.
Its cookbook has 26 ready-made questions. They cover whether the player is threatening an NPC, offering a bribe, accepting a quest, being sarcastic or trying to jailbreak the AI, and whether an NPC's generated reply stays in character. Each entry lists its accuracy on a generated test set of 56 examples (54 for one question), labeled where two models agreed. Counting the add-ons below, 23 of the 26 reach 80% or better; on the base model alone, 10 do. Persuasion is lowest at 68% even with its add-on, and the cookbook flags it as a weak spot.
Sixteen questions come with a "compiled" adapter built for that question alone. With one, the haggling question (should the merchant accept this offer?) goes from 29% to 89%. The playground runs the model with ONNX Runtime on Hugging Face's basic CPU tier and offers copy-paste snippets for JavaScript, C#, Godot's GDScript, Unreal C++ and curl. Its privacy note says the playground stores what visitors type; the downloadable kit doesn't.
Why it matters: LLM-driven characters talk well but can improvise the rules. A classifier that only picks from options the designer wrote, and says when it's unsure, keeps decisions like "does the guard open the gate?" in ordinary game code.
Caveat: All accuracy figures are the developer's own, from small, machine-generated test sets. The model and runtime sit in private repositories, so AIGamer couldn't check its size or speed, and we haven't tested it.
GameDecider is still an alpha, and Holdout's website doesn't mention it or any pricing yet. For now, the playground's README lists a Holdout email address for developers who want the downloadable alpha kit.
