chess-gpt-fullbudget-68-thinking

chess-gpt-fullbudget-68-thinking

chess-gpt-fullbudget-68 with a thinking display. The ONNX weights and vocabulary are copied byte for byte from peterwooden/chess-gpt-fullbudget-68 at 0e3f0b824ef8b4cb8c6ef39591d86b833a7ea6eb; only the package entrypoint differs. Move selection is unchanged — verified identical across the packaging probe suite and at a time limit large enough that the search always runs to full depth.

The entrypoint emits the tournament thinking-display commands documented in docs/TOURNAMENT_RULES.md (2026-08-08):

• Candidate arrows — the value-screened root moves that earned deepening, with intensity set by beam rank (best 1.0 → last 0.4). Rank rather than raw score, because value-head scores across a beam sit in a band a few hundredths wide and would otherwise render as four identical arrows.
• Search highlights — the destination squares each ply of the tree expands, fainter with depth.
• Quiescence highlights — the contested squares where captures get resolved.
• Final arrow — the move actually returned, at full intensity.

At most 56 commands are emitted per move (measured peak: 29), well under the runner's limiter of 64 per 500 ms. The display is optional: when a runner supplies no thinking object, emission is a no-op.