Offline experiment · detector ai-workspace/game-story-prod-review/regen/positional-detect.mjs · chess.js only, no engine call, no Python change · 2026-09-04

Can plain board rules (passed pawn, king cover, open file, trapped piece...) explain the 58 pending moments? Result: about a third, not all.

1. What I built

A Node script reads each pending moment: the position before the move, the move played, the better move, and the engine's 6-ply line after each. It computes board facts for both sides at each ply of both lines, and fires a rule when the played line is worse on a fact and the move that caused it is the played move or the reply. Fuzzy facts (piece mobility, checks available, attackers near the king) are kept apart as weak signals.

2. The numbers

A rule that fires is not a rule that is right. I checked 14 of the concrete hits by hand on the board. 11 of 14 name the real reason or a fair part of it. 3 name a side fact and miss the real reason.

So the honest estimate is: the concrete rules give a true, simple sentence for about 30 of the 58 moments. Of those, some already appear in today's story as a fact appendix (for example "It gives them a passed pawn on c3"). The other 28 need something else: a forcing line the story rejects today, the threat search, or a reason no board rule can name.

3. Where the wrong hits come from

The wrong hits share one shape: the rule sees a true fact that is not the point. A king-cover rule fires on a pawn push far from any danger. A hanging-rook rule fires when the real story is a rook trade into a pawn ending. A knight-guard rule fires when the real miss is a mating attack. Board rules cannot rank facts by importance. The engine score says how bad the move is, but not why.

Two wrong hits pointed at rules I did not have. Both are now in the script and both are Node-only: a perpetual check that was there (the better move gives check after check, or the line repeats, and the score is level), and a reply that forces a queen or rook trade into a lost ending.

4. Every pending moment

5. What I would do

Ship the rules that were right every time I checked them, each as one fixed sentence, only when the engine line agrees (the rule's move is in the line):

• Rook to the open file. "The rook belongs on the open c-file (Rac8)."
• Piece attacked with tempo. "After (f4) your bishop is attacked and has to move."
• Trapped or pinned piece. "Your bishop on h6 has no safe square."
• Passed pawn made or lost. "It gives them a passed pawn on c3."
• Knight to the edge. "The knight goes to the edge. (Ne4) puts it in the centre."
• Perpetual check missed. "(Qc7+) starts a perpetual check. The game was a draw."
• Forced trade into a lost ending. "(Qe5+) forces the queen trade. The pawn ending is lost."

Do not ship the fuzzy rules (mobility, checks, king-zone attackers). They fire often and name nothing a reader can see.

Expected effect: a true, simple sentence for about 30 of the 58 pending moments, with zero engine time. The other 28 stay as they are unless the threat search comes back for the 21 sharp-threat cases, which overlap with these in part.