"Overall" treats each coach’s dice as one big pool: every six-sided die and every block die they rolled, whatever it was for. It doesn’t look at what a roll needed or what was riding on it, only whether the faces came up as often as fair dice would, and whether they leaned good or bad.
"Actions" judges action rolls each against its own chance, so failing a 2+ dodge counts for more than failing a 4+ pickup. It covers single-die action rolls with a target number (dodges, rushes, pickups, catches, Bone Head and so on), and also armour rolls and injury rolls, where a KO counts as half a success and a casualty as a full one. Filled dots are passes (or armour breaks), hollow dots are failures. The line at the bottom shows how far ahead of or behind fair dice each coach was as the game went on.
"Blocking" follows every block through to its end: whether the defender went down, was pushed, both players fell, or the block caused a turnover; then whether the defender’s armour broke (or they were pushed into the crowd), and how badly they were hurt. Next to each count is what fair dice would average, and the bell curve shows the coach’s block dice luck: knockdowns are good, pushes neutral, and both down or a turnover bad, judged on each block’s first roll.
Everything here runs in your browser from the replay file alone. This page explains where the numbers come from, in plain terms first and then with enough detail to check the maths.
The question every number answers
Each luck figure compares what happened with what fair dice would do in the same situation, and asks how often fair dice would be at least that far from average. That gives a probability p, shown as “1 in X” where X = 1/p. “1 in 1.0” means perfectly ordinary dice; “1 in 50” means fair dice land this far out about once in 50 games’ worth of the same rolls.
The tests are two-sided: “at least this far from average” counts both directions. The bell curve’s marker shows which way the dice went: left is worse than average, right is better.
Reading the replay
A replay records every roll the game made, with the dice, the number the roll needed (after all modifiers) and the result. Dice are credited to the coach who rolled them. Armour and injury rolls go to the coach of the player causing the damage, as on the tabletop, even when a player falls over on their own turn. Kickoff dice go to the kicking team and weather to nobody. Rerolls count as separate rolls, and so does the failed roll before a reroll.
Three kickoff results aren’t counted: Quick Snap, Solid Defence and Blitz (also called Charge). Each rolls a d3 for how many players may move, and the replay records only the d3, not a die face that could go in the d6 pool. Any other roll the decoder can’t count is listed under “Some dice in this replay weren’t counted” below the results.
The parser was checked against twelve sample replays. Its reading of target numbers (a 1 always fails, a 6 always passes, otherwise the die plus modifiers must reach the target) matched the game’s own pass or fail on all 681 action rolls, and its armour rule matched 687 of 689 armour rolls.
Overall: the dice as a pool
Luck. Each face gets a score for how good it is for the roller, and a coach’s score is the sum over all their dice of that type.
| d6 | 1 | 2 | 3 | 4 | 5 | 6 |
|---|---|---|---|---|---|---|
| Score | −1 | −0.5 | −0.25 | +0.25 | +0.5 | +1 |
| Block die | Skull | Both Down | Push | Stumble | Pow |
|---|---|---|---|---|---|
| Score | −2 | −1 | 0 | +1 | +2 |
| Chance | 1/6 | 1/6 | 2/6 | 1/6 | 1/6 |
For n fair dice the total score T has an exact distribution, found by convolving the single-die distribution n times (no normal approximation, which is poor for a few dozen dice). With expected total E and observed total t, p = P(|T − E| ≥ |t − E|). The d6 scores are stored multiplied by 4 so they’re whole numbers; scaling every score by the same factor doesn’t change p.
Odds ask a different question: how uneven was the mix of faces, whichever way it fell? This is a chi-square goodness-of-fit test, χ² = Σ (O − E)² / E over the faces, with 5 degrees of freedom for a d6 and 4 for block dice (Push’s expected count is doubled). Many 1s and 6s with few 3s and 4s can give unusual odds and ordinary luck. The chi-square distribution is an approximation that gets rough when a face is expected fewer than about five times, so treat Odds for small pools as a rough guide.
Hovering a bar asks whether that one face came up more or less often than 1 in 6 (1 in 3 for Push) would give: an exact binomial test on its count. More 4s, 5s and 6s, Stumbles and Pows is lucky; more of any other face is unlucky, Pushes included.
Actions: each roll against its own chance
A 2+ dodge should pass five times in six; a 4+ pickup only half the time. Pooling the dice loses that, so this tab judges every roll against its own odds.
- Single-die rolls with a target (dodge, rush, pickup, catch, pass, Bone Head and other skills): the chance of passing is (7 − k)/6 for a target of k+, with k held between 2 and 6. Pass scores 2, fail 0.
- Armour: the chance that two dice reach the number needed to break, using the 2d6 distribution. Broken scores 2, held 0.
- Injury: two dice plus modifiers (Mighty Blow, for example) on the victim’s table, with Stunned 0, KO 1 and Casualty 2. Normal players are Stunned to 7, KO on 8–9 and a Casualty from 10; Stunty players Stunned to 6, KO on 7–8, Casualty from 9; Thick Skull players Stunned to 8, KO on 9, Casualty from 10. Casualty rolls themselves aren’t rated.
- KO recovery: at the end of each drive, every Knocked-out player rolls a d6 to come back, normally on a 4+. A keg inducement makes it easier, and the replay records it as a modifier, so the target is worked out the same way as other rolls. Recovered scores 2, stayed out 0. These dice also count in the d6 pool on Overall.
The rolls are independent but not identical, so the distribution of the total score is built exactly by convolving each roll’s own distribution in turn, and p is the same two-sided tail as above. The line at the bottom plots, roll by roll, passes so far minus the passes fair dice would average (a KO counting half), for each coach.
Hovering a row or heading recalculates luck for just those rolls. With three rolls or fewer there isn’t enough to give a meaningful “1 in X”, so the page says “very lucky” or “very unlucky” instead, by which side of average they fell. The action filters do the same for one kind of action, leaving the others on screen, dimmed.
Blocking: where blocks ended up
Results come straight from the game, which applies skills itself: Knockdown, Push (including when nothing happens, such as against Stand Firm), Both down, or Turnover. The Sankey diagram then follows the defender’s armour roll (or a push into the crowd) and injury roll.
Expected values need the chance of each result for each block. Every face is classed from the attacker’s side using four skills:
- Pow is a knockdown. Stumble is a knockdown unless the defender has Dodge and the attacker lacks Tackle, when it’s a push.
- Push is a push. Skull is a turnover.
- Both Down is a push if either player has Wrestle; otherwise a knockdown if only the attacker has Block, a push if both do, a turnover if only the defender does, and both down if neither does.
Whoever chooses takes their best die: the attacker the highest of Knockdown > Push > Both down > Turnover, the defender (when stronger) the lowest. If a single die lands at or below a given result with probability F, then with n dice the attacker’s chosen die does so with probability Fn, and the defender’s with 1 − (1 − F)n. On two dice where the attacker chooses and nobody has skills, that gives a knockdown 55.6% of the time and a turnover 2.8%. Other skills (Stand Firm, Fend, Juggernaut, Brawler and so on) are ignored for now. On the 838 sample blocks that weren’t rerolled, this model picked the same result as the game for 831; the rest look like coaches deliberately choosing a different die.
Block luck scores a knockdown 2, a push 1, and both down or a turnover 0, judged on each block’s first roll, by the same exact method as Actions. It counts for the attacking coach even when the defender chooses, since the attacker rolled the dice. Rerolls are left out, which is why actual results (after rerolls) tend to run a little ahead of the expected ones.
Hovering asks whether the hovered thing came up more or less often than fair dice would make it, out of everything that could have led there: each block’s first roll for faces and results, every armour roll reached for the armour nodes, every injury reached for the injury nodes. Each block or roll is a yes-or-no event with its own chance, combined exactly as above. More knockdowns, broken armour, KOs and casualties is lucky; more pushes, both downs, turnovers, held armour and stunned results is unlucky.
Assumptions and limits
- Independence. Every roll is treated as independent of every other, which fair dice are.
- Small numbers. A coach rolls a few dozen to a couple of hundred dice a game, so a single game’s figures swing a lot. The exact methods keep the probabilities right, but they can’t make a small sample informative.
- Many comparisons. Hovering lots of rows, faces or bands is many tests, and some will look unusual by chance alone. One row at 1 in 15 among twenty you looked at isn’t surprising.
- Discrete outcomes. With few dice, only certain probabilities are possible, so the tests are slightly conservative (a 1-in-20 threshold is hit a little less than 5% of the time).
- The marker. The bell curve is a fixed standard normal. The marker sits at the point with the same two-sided probability as p, on the side the dice fell, so its distance from the centre always agrees with “1 in X”. It’s clamped just inside the edges for very rare results.
- Choices, not facts. The face scores, the KO-as-half rule, crediting damage rolls to the attacker, and the skill rules for block dice are judgement calls about what counts as lucky. Different choices would give somewhat different numbers.