Simulator Versions

2.3 Fair loose pucks, and fast results from complete games 2026-10-25 · insider-beta-313

• FixedWhen skaters from both teams reached a loose puck or rebound at the same moment, the home team always won it, because home players were checked first. The nearest skater now wins, and a tie is settled at random. In 835 test games on recent fixtures the home team won 58% of games (it was 65%; NHL 59%).
• NewFast results are complete games played by the same engine without drawing them. Before each forecast the two teams' relative strength is fitted to the pregame odds (or our model) through real probe games; only the games after that are counted, and every game's own statistics must add up (shots on goal = saves + goals).
• Known limitThe full game still plays about 3 minutes of hockey for each 20-minute period and makes about 6 shots on goal per team (NHL 27.8). A real-clock version exists as a research profile but scores far too much (about 12 goals per team), so it is not used for results.

Engine fingerprint: simple_nhl_engine@d9338d55392433f0

2.2 Full games play out more naturally 2026-10-19 · insider-beta-253

• BetterFull games now use the NHL-fitted shot decisions the Coaching Lab already used: a skater with the puck carries it in and passes before shooting, instead of firing as soon as a lane opens. In 200 test games 34% of shot attempts came from inside 25 feet (it was 6%; NHL 39%), the typical attempt was 35 feet out (it was 52), and passes per game rose from 48 to 67.
• BetterLines stay on the ice long enough to see: a team changes about every 11 seconds of replay (it was every 2). The shift clock runs at a fifth of the game clock in full games, and ice time still follows the lines: about 19, 17, 14 and 11 minutes for forward lines one to four.
• Known limitScoring is a little low: 2.5 goals per team per game in 200 test games against 3.07 in the NHL in 2025/26 (18% low; the check allowed 15%). The shot-volume weight is set from shot volume only (NHL 42.8 unblocked attempts per team-game against the engine's 7.7) and was not tuned to goals.
• Known limitA full game is still shown in about three and a half minutes, so the clock runs fast; skaters move at real speed. Shot counts on the scoreboard are the simulated attempts, not NHL-sized totals.
• Known limitCoaching Lab results are identical to 2.1, run for run. Full games are not comparable with 2.1: the same seed plays a different game.

Engine fingerprint: simple_nhl_engine@855624ff3b039f8a

2.1 Line changes look like the NHL 2026-10-13 · insider-beta-117

• BetterForwards now change as a line of three and defencemen as a pair, the same players each time, and the top line plays the most (about 19, 16, 13 and 10 minutes for lines one to four, 23, 20 and 17 for the pairs). Before, each skater was relieved on his own by whoever had played least, so the three forwards on the ice were almost never one line.
• BetterA change now takes a few seconds: the fresh player reaches the play in at most six frames and the tired one is over the boards in six. Before, players could be seen skating to the bench for about half a minute of game clock, so in almost every frame extra players were crossing the ice.
• Known limitShift lengths were already NHL-like and stay so: about 46 seconds a shift, 19 shifts and 15 minutes a game for forwards, 26 shifts and 20 minutes for defencemen. A full game is shown in about four minutes, so a line still changes every two or three seconds of replay.
• Known limitCoaching Lab results are identical to 2.0, run for run. In a full game with real rosters different players are on the ice than in 2.0, so the same seed plays a different game; the scoring rules are unchanged.

Engine fingerprint: simple_nhl_engine@d19fa104bfe683e7 · results are comparable with the version before

2.0 Goals, shots and players now follow NHL data 2026-10-12 · insider-beta-113

• FixedGoalies saved only about half the shots on goal in a rush (NHL goalies save about nine in ten), and long shots went in far more often than close ones. Whether a shot on goal goes in now comes from a model of NHL shots: distance, angle, rush, rebound, man advantage and the goalie's rating. It was fitted on the 2023/24 and 2024/25 seasons and checked on every 2025/26 game, which it never saw.
• FixedA save could send the puck back toward the goalie's own net. Rebounds now always leave the net.
• FixedToo many shots missed the net or hit the frame from behind or beside it. About 69 in 100 unblocked shots now reach the goalie in a rush (the NHL figure is about 74 in 100); before it was about 51.
• BetterGoalies freeze the puck about as often as NHL goalies (about one save in five); before, almost every save stayed live.
• NewSkaters now differ. Shot accuracy, finishing, shot volume, where a player shoots from, shot blocking and playmaking are measured from NHL shots (2023/24 to 2025/26), shrunk toward the position average for small samples and checked on a season the estimate had not seen. Each one changes what that player does in the engine. Skating and awareness are not measurable from these data and stay at the default.
• BetterGoalie ratings are now measured against the league average goalie, and trust a goalie's record less until there are about 8,000 shots behind it.
• BetterAfter a whistle in the middle of play, the next draw is in a zone, not at center ice, so full games no longer start dozens of extra center-ice rushes.
• BetterA 5v4 start now scores more than 5v5, as in the NHL.
• Known limitResults from 1.5 and earlier are not comparable with 2.0: the same plan scores far less often now.
• Known limitA 5v3 start scores no more than 5v4 (NHL two-man advantages score clearly more), so 5v3 is understated.
• Known limitFull game replays are unchanged from 1.5 (same scores, same frames). The new shot selection and accuracy are in the Coaching Lab only: with real rosters and line changes the engine still takes too few close-range shots in a full game (about 31 in 100 against 39 in 100 in the NHL), so full games switch to it later.

Engine fingerprint: simple_nhl_engine@45683a6434b6648c

1.5 Every hockey tactic in the dictionary, ready to test 2026-10-11 · insider-beta-93

• NewTactics library: all 60 hockey tactics from the Scoutics dictionary and its 40 test cards, searchable in the Coaching Lab. Each one says honestly how far the Lab can go with it: Runs, Starting situation only, or Reference only.
• New50 ready-made tests. One click loads the starting picture, and where the card has two plans, both Plan A and Plan B.
• NewFive skaters a side (it was four), so 5v5, 5v4, 5v3 and 3v3 can all be drawn.
• NewA skater can wait, then go. This is how arrival order and timing are compared.
• NewFour more measures: carries the puck over its own blue line, carries it over the far blue line, completes a pass, and the other team gets a shot off.
• BetterDefensive tests read from the defence's side ("The other team scores", "Your side wins the puck"), and a measure that is almost always or almost never true is flagged because it cannot separate two plans.
• NewThis version list. Every result now shows the simulator version that produced it.
• Known limitA numerical edge barely changes scoring in the engine's own checks: a 5v3 start scored about as often as 5v5. Power-play comparisons are exploratory until the engine is checked against power-play data.
• Known limitTests that hold a defensive shape (a box, a trap, a wedge) freeze it for a few seconds and compare it with the engine's own defending. A large gap shows how the engine reacts to a frozen shape, not how a real team using that structure would do.

Engine fingerprint: simple_nhl_engine@379db429f0e1a9bd · results are comparable with the version before

1.0 The Coaching Lab simulator 2026-10-11 · insider-beta-91

• NewPlace up to four attackers against up to four defenders and a goalie, choose who starts with the puck, and run the next few seconds many times in the production hockey engine.
• NewPlayer ratings from public records, shrunk toward average when the sample is small, with the source of every number shown. Anything not measured stays at a disclosed default.
• NewPlan A against Plan B on the same random seeds, with a 95% interval on the difference, and a replay of any run by its seed.
• Known limitThe engine has not been checked against tracked rush data. Results are labelled a simulated test, never a match win probability.

Engine fingerprint: simple_nhl_engine@379db429f0e1a9bd

1.2 Fast results from complete matches 2026-10-25 · insider-beta-313

• NewFast results are complete matches played by the same step engine as the replay, without drawing them; the earlier possession-sampling model is no longer used. A match takes about three minutes of computer time, so a fast run plays 10 matches, and the page shows each one as it finishes.
• Known limitSoccer results are not yet fitted to the odds: one fitting needs dozens of matches. The engine scores about 1.2 goals a match (Premier League about 2.8).

1.1 Substitutions in the match replay 2026-10-14 · insider-beta-123

• NewEach team now makes three to five substitutions in the second half, in up to three windows (around 57-66, 68-76 and 79-86 minutes), at a stoppage as in a real match. The most tired players come off and the best fit from the bench comes on, into the same position.
• NewPlay pauses while the player coming off walks to the halfway line and the substitute jogs on; both are drawn faded. The bench beside the pitch updates, so the player who came off takes the substitute's seat.
• Known limitThe engine's own substitution rules (injuries, first-half changes) stay switched off; the changes follow this second-half plan, the same one the Coaching Lab has used since October. The bench is the club's current squad from the Fantasy Premier League list.

1.0 Versioning starts here 2026-10-11 · insider-beta-93

• NewThe simulator that was running on 11 October 2026 is version 1.0. Changes before this date are in the release notes, not in this list. From here on, every improvement is added.

1.2 Fast results from complete games; three-point spots fixed 2026-10-25 · insider-beta-313

• NewFast results are complete 48-minute games played by the engine without drawing them, with the teams' relative strength first fitted to the pregame odds (or our model) through real probe games. The Lineup Lab replay now plays every possession too and shows about one in six (and the last two minutes); before, the possessions it did not show were only added to the score.
• FixedThe half-court spots for a planned three were inside the three-point line (23.5 ft on the wing, 19 ft near the corner), so those shots counted as twos. They are now just behind the arc.
• FixedPlayer skill and usage: skill was clamped to the same value for every player, and usage came from this season's numbers only, so before stats were published every player had the same share of the ball. Last season's numbers are now used until this season's exist.
• Known limitSince release 122 the full game has drifted away from NBA box scores: about 120 shots, 25 threes, 3 turnovers and 2 fouls per team (NBA 89, 37, 14 and 20). A repaired research profile is being calibrated and is not used for results yet.

1.1 The whole game, every possession 2026-10-14 · insider-beta-122

• NewOn the sport pages the replay now shows all 48 minutes: every possession is played out on the court, sped up about nine times, so a game takes about seven minutes to watch. Before, about 30 possessions were shown and the rest were only added to the score.
• NewPlayers walking to and from the bench at a substitution are drawn faded on the court, as in the Coaching Lab.
• BetterWith every possession played out, the box score now looks like an NBA game. Over 20 team-games on real rosters a team averaged about 115 points, 89 shots (45%), 36 threes (36%), 27 assists, 14 turnovers and 19 fouls. Before calibration it had about 4 turnovers, 1 foul, 18 threes and two in three shots from mid-range.
• BetterHow: a gather can be stripped and contact is called at NBA rates; an open long two early in the shot clock becomes a step-out three or a drive, with the defender closing out; players without free-throw data shoot about 79% (bigs 70%) instead of 48%.
• BetterThe simulation runs about 40% faster with exactly the same results, and a full game is kept for six hours so it opens at once the next time.
• Known limitRebounds run about 10 to 20% above NBA box scores, because every missed shot is credited to a player (the NBA counts some as team rebounds) and shooting is about two points below the NBA average. Calibrated on rosters without this season's statistics; to be checked again once they are published.
• Known limitThe Coaching Lab still uses the shorter replay and is unchanged, including its low free-throw rate for players without data.

1.0 Versioning starts here 2026-10-11 · insider-beta-93

• NewThe simulator that was running on 11 October 2026 is version 1.0. Changes before this date are in the release notes, not in this list. From here on, every improvement is added.

1.0 Versioning starts here 2026-10-11 · insider-beta-93

• NewThe simulator that was running on 11 October 2026 is version 1.0. Changes before this date are in the release notes, not in this list. From here on, every improvement is added.