rrodolfo0/move-intent-frontier-model-comparison

Move Intent frontier model comparison

Enter one finalized English chess transcript. The demo runs the same pinned Qwen base twice: once with the adapter disabled and once with the selected 2,700-row adapter enabled. It shows raw output, framing and parsing status, and the normalized Move Interpretation.

The Frontier tab opens first. It exposes the registered few-shot prompt, its SHA-256, and the frozen Haiku and GPT outputs produced from it. The Live tab exposes the system prompt shared by the base and tuned runs.

This demo does not produce SAN or decide legality. It has no board state. chess.js and the host-owned Move Resolver perform those jobs after interpretation.

Release verification calls the named compare Gradio API with a real transcript. It requires nonempty base and tuned outputs, confirms that the adapter is disabled for the base pass and enabled for the tuned pass, and validates both normalized Move Interpretation objects.

The selected adapter scored 272/300 strict exact on the frozen own-test panel, the best result in the four-point controlled curve. It did not meet all registered gates, and staff heldout was not supplied.

The Frontier comparison tab contains all 30 rows from the frozen Prompt-Ceiling development panel. It shows the immutable few-shot outputs from Claude Haiku 4.5 and GPT-4o mini, then runs the selected adapter live on the same transcript. The frontier outputs remain separate from the final 300-row Own-Test result and are not regenerated by visitors.

The public demo uses ZeroGPU with a single concurrent request, a bounded queue, a 500-character input limit, and a 120-second GPU allocation per comparison. It does not call Hugging Face Jobs or an external paid inference provider.