The Blunder Radar

An engine answers “what is best here”. That is not the question you have when you are about to move. The useful question is “what am I likely to get wrong”, and the answer depends on who is playing.

The idea

Two things are already available: how likely a human of a given strength is to play each move, and what each move costs. Multiply them.

risk = SUM over moves of  P(a human plays it) x centipawns it loses

The first term comes from Maia, a network trained to predict human moves at a target rating rather than to win. The second comes from Stockfish.

A move that hangs a queen but no human would consider carries no risk. A natural move that quietly drops a pawn carries a lot. Neither an engine nor a human-move model gets you that on its own.

Why the engine has to be asked directly

The moves a human is likely to blunder into are, by definition, ones the engine ranks poorly - so they fall outside an ordinary MultiPV list and would never be scored. evaluate_moves() uses UCI searchmoves to make the engine evaluate exactly the candidate moves, whatever it thinks of them:

maia <- maia_session_start(1500)
risk <- blunder_risk(
  "r1bqkbnr/pppp1ppp/2n5/4p2Q/2B1P3/8/PPPP1PPP/RNB1K1NR b KQkq - 3 3",
  maia
)
risk$risk
#> [1] 86.9
head(risk$moves)

Does it work?

A risk metric that fires everywhere says nothing, so it was measured against known traps and a quiet control:

Position	Risk @1100	Risk @1900
Legal’s mate	6724 cp	9192 cp
Scholar’s mate	1673 cp	998 cp
Blackburne	90 cp	-
Starting position	3.0 cp	3.6 cp

Roughly 2000x between a trap and a position where nothing can go wrong. The bottom row is the one that makes the rest meaningful.

A result worth noticing: Legal’s mate scores higher at 1900 than at 1100. The stronger player is more confident about grabbing the queen - 93.3% versus 65.7% - so the trap works better on them. That is not something an engine evaluation can express.

The arrows

On the board the engine’s choice is blue and the human moves are orange, or red where they lose real material. Which moves get an arrow is chosen by risk contribution, not by raw probability.

That distinction is the whole feature. After 3.Qh5, Black’s losing Nf6 is only the fourth most probable reply. Drawing the three most likely moves produces three orange arrows and quietly omits the one move worth warning about. radar_arrow_moves() takes the top contributors and adds the single likeliest move for context.

Trappiness

The same machinery, pointed the other way: for each move you could play, how badly is the opponent likely to go wrong afterwards?

trappiness(fen, maia, ctx)

practical is what the move is likely to win you from their mistakes, minus what it costs you objectively. A slightly inferior move that sets a problem they will probably fail can be the better practical choice.

Requirements

The radar needs lc0 and the Maia weights. On Windows ensure_maia() fetches both. There is no official Linux build of lc0, so the container image compiles it from source - see the Dockerfile. Without them the app still runs and the radar reports that the human model is unavailable rather than failing.