Move Merit: Behind a Beautiful Move
In chess, “best” is a crowded word. A forced move is best. Six interchangeable moves can be best. A move that the reference model gives little chance of finding can be best.
Those are different achievements. An engine verdict has to flatten them into a value. I built Move Merit to pull them apart: value beside alternatives, modelled choice, risk, and the stability of the analysis. Only then can I ask what kind of move I am looking at.
The feeling
The interesting question begins after the verdict. If every alternative loses, the move is necessary. If several alternatives preserve the result, it is replaceable. If the played move preserves value where the reference model concentrates on worse choices, that deserves a different kind of attention.
A bare engine verdict calls all of these positions best and hides the difference. Move Merit makes the alternatives part of the evidence. It does not decide that one move is prettier than another; it refuses to let one word erase how the move was found and what was at stake.
The experience
Open a complete game and select a move in the chart. The board, move details, and metrics update to the same position. A number that has lost the position it describes is not much use to me.
I use the report to keep four questions separate:
- Did the move preserve objective value?
- What did it give up against the best move?
- Did the model have an easy way to reach any near-best move?
- How costly were the moves the model expected instead?
These are not four views of one score. A move can have zero regret and still be forced or easy to find. Another can beat the peer expectation in a position where the model usually chooses costly alternatives. Treating those moves as equivalent misses the whole point.
The mathematics underneath
Move Merit begins in a deliberately unglamorous place: every legal move gets an expected-match-point value. It uses exact Syzygy tablebases where available and bounded UCI WDL search otherwise. A reference model supplies a likelihood p(a) for each move. Those percentages are the model’s weights, not odds that a real person will choose the move.
For the played move a0, the report calculates human value added:
HVA = V(a0) - sum over legal moves a of p(a) × V(a)
The weighted sum is the peer baseline: the result expected from the model’s distribution of choices. HVA compares the played move with that baseline. Positive HVA means the move achieved more than the model’s average choice. Played regret measures value lost against the best move; findability measures how much probability the model puts on any near-best move; and criticality measures the value the model is expected to lose. The search ranges show how much those answers move across finite search budgets.
What Move Merit can and cannot say
I read these metrics as separate evidence, not ingredients for a beauty score. Low regret says the move preserved objective value. Low findability says the model gave little probability to the near-best set. High criticality says it expected costly mistakes. A wide search range warns that the finite-search answer is less stable; it is not a confidence interval or a proof. None of that decides whether a person should call the move beautiful.
Why it matters
I do not want to treat a forced, zero-regret move and a low-findability, high-criticality move as equivalent. The report shows why they differ and where finite search makes that conclusion uncertain. It does not turn the difference into a beauty score. It leaves the board, the alternatives, and the uncertainty in front of you. That is where the judgment belongs.