+19: GSAx leader (Philips). 0.59: How much GSAx repeats. 0.98: xG calibration on easy shots. 12: Qualified goalies Save percentage has a quiet bias: it pays a goalie the same for a routine wrister as for a breakaway. Goals saved above expected (GSAx) fixes the easy part by grading every shot against its difficulty, and it already lives on our goalies page . This post does the two things that page does not. First, it asks how hard each goalie's job actually is. Second, and more useful, it audits the number itself, because GSAx has fine print, and the parts of it people most want to quote are the parts that hold up worst. One reason to bother: goaltending is real. A shooter's finishing barely repeats from one stretch of a season to the next (we measured about 0.28). A goalie's GSAx per shot is steadier, around 0.59 splitting this season into odd and even games. That is why goaltending, not shooting, is the half of a team's puck luck that actually sticks. So the goalies are worth ranking, as long as we are honest about how much the ranking can really say. How hard the job is Start with the cleanest cut. The chart sets each goalie's workload (expected goals per shot she faces, a measure of how dangerous her average shot is) against how much better than expected she has played. Right means a harder job, usually a leakier defense; up means saving more than expected. The corner that earns the most respect is the top right. Gwyneth Philips faces the hardest workload in the league (0.090 expected goals per shot, against a 0.078 average) and is still well in the black. Aerin Frankel posts a similar GSAx behind a much softer 0.072 workload and a stingier defense. On the scoreboard they look alike; the job is not the same, and the scatter is the only place you can see it. The audit: which saves are real Now the fine print, because GSAx is only as good as the expected-goal model under it. Checking our xG against what actually went in, bin by bin and pooled across every season, the model holds up everywhere: 0.98 , 0.96 , and 1.03 actual-to-predicted from low to high danger, all within a couple of points of perfect. So the model is not the problem. Low danger is still the wrong place to judge a goalie, for a different reason: a routine shot goes in about two percent of the time no matter who is in net, so there is almost no spread between goalies to measure, and the tiny differences that do show up are mostly luck. So we set low danger aside and judge goalies on the medium and high-danger shots, where the save gaps are real and where, frankly, goalies are the only ones allowed to differ. Which brings us to the number everyone wants: who steals the grade-A chances. Here is each goalie's high-danger GSAx with its margin of error drawn on. Read the whiskers, not the bars. The three best goalies' high-danger ranges all overlap. Gwyneth Philips , Ann-Renée Desbiens , and Aerin Frankel look like a clear one-two-three on the raw number, but each margin of error is several goals wide, easily enough to reshuffle the order. One season is simply not enough high-danger shots to say who is best at the thing that feels most like goaltending. What the chart can say honestly is narrower and still useful: the goalies whose whole range sits below zero, like Kayle Osborne , were genuinely beaten on the dangerous chances, and the leaders are clearly a tier above the bottom even if they are not separable from each other. What this read isn't The high-danger ranking is not a ranking. With a few dozen to a couple hundred grade-A shots a goalie, the error bars overlap for everyone but the extremes. Trust the overall and the workload context; treat the high-danger order as a hint, not a verdict. It is the least repeatable number in the post.. GSAx is partly the defense. Expected goals know where a shot came from and whether it was a rebound, but not whether the goalie was screened or beaten by a cross-ice pass she never had a chance to track. Teams that give up those make their goalie's job harder than the logged numbers show. So read GSAx as a goalie-and-defense result; the workload axis helps separate the two but does not finish the job.. Twelve goalies, one season. The repeatability (0.59) is measured within a single season, which overstates how much carries forward, and the population is tiny. Every correlation here has a wide confidence interval; read the order as strongly suggestive, not settled.. Our xG, and shot-based. This is about shots stopped, not rebound placement, puck handling, or the saves that stop the next shot from ever happening. Methodology GSAx and workload. Expected goals against minus goals against over every non-empty-net shot a goalie faced this season; per-100 normalizes for volume. Workload is expected goals per shot faced.. Danger and calibration. Shots bucketed by expected-goal value (low under 0.04, high 0.10+). We checked each bucket's calibration across all seasons: actual goals over summed xG came to 0.98 (low), 0.96 (med), 1.03 (high), well calibrated across the board. We still exclude low danger from skill claims because it carries no real between-goalie spread, not because the model is off.. Error bars. Each high-danger GSAx carries a binomial standard error, the square root of the summed shot-by-shot variance, which is the irreducible luck in a finite set of dangerous shots.. Repeatability. GSAx per shot on odd vs even games, correlated across the 11 goalies and Spearman-Brown corrected (0.59). Split-half within a season is an upper bound on the season-to-season figure, and at this sample its confidence interval is wide.. Prior art. GSAx is standard at Evolving-Hockey and MoneyPuck, and goalie performance being among the least repeatable signals in hockey is well established. The PWHL-first piece is the workload context and the per-danger calibration-and-repeatability audit for this league's goalies. Per-goalie game logs and rolling GSAx live on the goalies page . This pairs with our regression read , where goaltending was the half of team puck luck that actually sticks.