When is it worth taking a shot at an enemy piece? Going for a kill in Killer Queen means committing to an exchange you might not win. To make the question precise we adopt a simplifying assumption: the engagement resolves as exactly one death — your target dies, or you do — with no whiffs, mutual trades, or escapes. Under that model the shot is a gamble: land it and the game swings your way, lose it and it swings away. This page reads that gamble straight off the win-probability model, matchup by matchup and board state by board state.
Each panel takes the attacker's view. We evaluate the model at three states — the status quo V(S) (decline the fight), the board after the defender dies Vkill, and the board after the attacker dies Vdeath — and plot all three as win probabilities (y) against the attacker's warrior edge (x, their winged-unit count minus the defender's). The green band is the upside you gain by winning; the red band is the downside you risk by losing.
Let p be your true probability of winning that exchange (landing the kill rather than dying). Taking the shot beats declining it exactly when its expected value clears the status quo, which pins the break-even success rate p*: take the shot iff your odds beat p*.
So p* is just where V(S) sits on the segment from Vdeath to Vkill: a low p* means the fight pays off even at long odds (mostly upside, little to lose); a high p* means only take it when you're already favored. Panels are a matchup matrix — rows are the attacking piece, columns the defending piece — so any matchup sits transposed from its reverse across the diagonal. A speed drone can't strike directly, but a hunted one can bait an opponent into a teammate's kill, so its row is a real value swing too.
Before the per-state detail below, the bottom line for the three striking pieces. Each cell is one matchup (attacker row → defender column). model p* is the globally-averaged break-even — the odds a shot needs to be worth it. emp win% and K/D are model-free, from real kills: how often that attacker piece actually killed that defender versus the reverse (K/D = kills for ÷ against). The margin = win% − p* is the verdict: positive (green) means players win these exchanges more often than the break-even demands — the shot is +EV on average; negative (red) means a losing proposition even before opportunity cost.
The same p*, now read straight off the y-axis instead of inferred from the gap between the curves. The solid line is the bucket estimate (p* computed from the bucket's mean V's, the number quoted above and in the heatmap). Because every bucket pools thousands of distinct board states, p* also has a distribution: the shaded band is its interquartile range (p25–p75) and the dashed line its median — wide bands mean the same nominal matchup demands very different odds depending on the rest of the board. The dashed 0.5 line is the coin-flip pivot: below it the fight pays off at even-or-worse odds, above it you must be favored. p* generally climbs with your warrior edge — the further ahead you are, the less a fight can add and the more it can cost, so the bar to take it rises.
Does real play clear the bar as the game swings? Every state is scored by the model and dropped into one of 50 buckets along P(attacker wins). Per matchup, the solid line is the empirical win rate in each bucket (the attacker's share of the A↔B kills, K/(K+D)). The two dashed lines are the model's break-even p* — what it thinks is optimal: the violet one is p* at the fights that actually happened, the green one is p* over all game states in the bucket (its general prescription, not conditioned on a kill). Where the solid line sits above a dashed one, those shots pay off on average; read left-to-right to see how the verdict shifts from behind (left) to ahead (right).
The width of the two bands is the whole story. When the red downside is thin next to the green upside — the status-quo line hugging the death line — the fight is nearly free and p* is low: take it even as a heavy underdog. A few patterns to look for:
The win-probability model (gradient-boosted trees) maps a 52-feature game state to P(that team wins). For every real state in the data we form the three outcomes by applying the exact edits the model already tracks: a queen kill decrements that team's remaining lives (and ends the game on the last one); a warrior kill strips the unit's wings/speed back to a drone. Re-scoring the edited states gives Vkill and Vdeath; the unedited state gives V(S). Orienting each to the attacker's team yields the break-even p* above.
Each cell pools ~30,111,636 states from 182,447 games, with both teams taking the attacker role, then conditions on two axes: the defender's remaining queen lives (the toggle) and the warrior edge (x). The line is p* from the bucket's mean V's (a stable point estimate); the band and dashed median describe the spread of the per-state p* within the bucket.
The empirical win% and K/D are entirely model-free. We replay every game, and at each kill classify both killer and victim by piece type from the live game state (wings + speed; the queen is its own type), tallying a killer×victim count matrix. For a matchup A vs B the win% is A's share of the A↔B kills; K/D is kills-for over kills-against.