2026-05-08

The engine had a tactical shortcut with the wrong definition of “safe.”

In one annotated game, White captured a rook on d8. The move looked attractive under visible material: take rook, win exchange, move on. But the belief set already made the danger visible in a different way. The square was likely defended, and the attacker could be recaptured immediately.

That made the bug subtler than “the bot blundered material.” The problem was that a shortcut bypassed the same uncertainty-aware discipline the rest of the engine was trying to build.

The shortcut was named like a safety check, but it was only checking the currently visible defenders. In an imperfect-information position, that is not enough. If the particles say the target square is probably defended, then the move is not safe just because the defender is not visible on the board.

The fix was a belief-aware capture trade filter. For visible material shortcuts and main-evaluation capture candidates, the engine now asks whether the belief set contains immediate recapture evidence. If the target is not worth more than the attacker and the particles show likely recapture, the move gets filtered out instead of auto-fired.

The targeted replay did what a good small rung should do. The specific bad capture disappeared, the game result improved, and hard belief facts stayed clean. That did not prove the engine was broadly stronger yet. A later small batch still surfaced separate issues. But it did prove the previous blindspot had been converted into a guardrail.

This is the process loop I want for the engine track: human review identifies a recurring blindspot, the engine gets a narrow rule or diagnostic, and the next replay checks that exact class before any longer bake-off.

In an imperfect-information engine, a shortcut named safe must prove safety against the belief set, not only against currently visible defenders.