Notes & Working Papers
2026.07.07
v1.0

Twelve Thousand Games in Avalon

AuthorROME THORSTENSON
FiledJULY 2026
Reading7 min
StatusDRAFT
A dense field of faint gray dots with a bright blue wave of dots threading through it, left to right—signal emerging from noise.

Anyone who plays hidden-role games knows the end-of-game reveal—the roles flip face-up, the ally you trusted was lying all night, the read you were sure of was backwards. In The Resistance: Avalon, five to ten players split into Good, who win by passing three of five missions, and Evil, who win by sabotaging three. One catch: a Good player, Merlin, secretly sees who's Evil—and if Good passes three missions, Evil still gets one shot to win by naming him.

That last rule isn't a footnote. It's most of the game. And 12,882 logged games, cross-checked against 20 hand-annotated ones from Avalon-NLU, say so with more precision than anyone at the table gets. They also say something less obvious: the board everyone plays off of carries almost no signal. The liars get found in the talk.

§ 01 Evil

Evil is favored, and Merlin is why

Evil wins 55.9% of games—7,207 of 12,882. Strip out Merlin and the Assassin and play bare The Resistance, where the missions are the only way to win, and Evil drops to 48.3%. A coin flip. Add Merlin back and Evil climbs to 56.6%.

Merlin does help Good; his sight measurably improves which teams get sent. He also hands Evil a second way to win, and that second way is worth more than the missions his sight saves. The player whose entire job is helping Good makes Good lose more often.

How Evil wins: 56.4% by sabotaging three missions, 41.2% by assassinating Merlin after Good passed three, 2.4% by five rejected proposals.How Evil winsShare of all 7,207 Evil victories (n=12,882 games)56.4%41.2%2.4%Sabotage three missionsAssassinate Merlin after Good passed threeFive rejected proposals
Fig 1 Passing isn’t winning. Blue: sabotaged missions. Amber: Merlin assassinated after Good already passed three. Two in five Evil wins come after Good won the missions.

Passing three missions isn't winning. It buys a second game only Evil plays, and Evil is good at it: the Assassin names Merlin right 36.7% of the time he's hunted, 2,972 of 8,103.

Assassin's Merlin hit rate against a blind-guess baseline, by player count. Observed: 40.5, 35.3, 36.0, 35.7, 29.3, 38.1. Blind guess: 33.3, 25.0, 25.0, 20.0, 16.7, 16.7.The Assassin beats chance—and the gap grows as the table fillsHow often the Assassin names Merlin when Good passes three missions, by player count.015304540.5%33.3%535.3%25.0%636.0%25.0%735.7%20.0%829.3%16.7%938.1%16.7%10playersAssassin's hit rateA blind guess
Fig 2 Better than a guess, and better on a bigger table. Blue: how often the Assassin names Merlin right. Gray: what a blind guess would land, one over the number of Good players. The hit rate holds near a third while the guess falls to a sixth—so the edge over chance grows.

The hit rate barely moves as the table fills, holding near a third whether Merlin hides among three Good players or six. A blind guess doesn't hold—it falls from one in three to one in six as the good ranks grow. So the edge over chance is skill, and the skill shows up most at a crowded table, where finding Merlin should be hardest.

§ 02 First

The first mission is a decision, not a result

Eighty-two percent of first missions pass.

Mission success rate by slot: Mission 1 82.4%, Mission 2 46.8%, Mission 3 51.8%, Mission 4 61.6%, Mission 5 52.9%.Mission success rate by slotShare of played missions that passed. The floor falls off a cliff at Mission 2.025507510082.4%Mission 146.8%Mission 251.8%Mission 361.6%Mission 452.9%Mission 5
Fig 3 The Mission-2 cliff. The pass rate drops from 82% to 47% between Mission 1 and Mission 2, then holds.

New players read that as good news. It isn't news at all. Ask a sharper question: when Evil sits on the mission with enough of its own players to sink it, how often does it let the mission pass instead?

How often Evil lies low by mission: it declines a sabotage it could make 73.2% of the time on Mission 1, falling to 31.1%, 24.0%, 12.3%, and 4.0% by Mission 5.How often Evil lies low, by missionWhen Evil could fail a mission and lets it pass anyway. Cover matters early; winning matters late.0255075100Mission 1Mission 2Mission 3Mission 4Mission 573.2%4.0%
Fig 4 The patience curve. How often Evil declines a sabotage it could make—73% on Mission 1, 4% by Mission 5. Cover early, kills late.

Three times in four on Mission 1. By Mission 5, only 4% of the time—and that 4% is error, not restraint. There is no Mission 6, so cover has no future left to buy, and failing the decider wins outright. Across those 95 games Evil traded a certain win for the Assassin's coin flip and cashed only 38.

So the 82% pass rate measures Evil's patience, not Good's judgment. The board is reporting a decision, and the decision is nearly always the same one.

§ 03 What

What a fail costs

A fail on a team of k implicates at least one of k, so the suspicion each member picks up divides by team size.

Change in the odds a team member is Evil against each game's base rate, by team size. Failing: +18.3 on a team of two, +2.5, −1.5, −1.3. Passing: −9.9, −21.9, −25.5, −25.6.What a fail costs, priced by team sizeChange in the odds a team member is Evil, against each game's own base rate. Observed, not modeled.+20+100-10-20-30+18.3-9.9Team of 2+2.5-21.9Team of 3-1.5-25.5Team of 4-1.3-25.6Team of 5the team FAILED — suspicion addedthe team PASSED — suspicion removed
Fig 5 A fail is priced by team size. Blue: what failing adds to the odds each member is Evil. Amber: what passing removes. A failed pair points hard at both players; a failed team of four points at no one.

A failed pair leaves each of the two players 58% likely to be Evil, 18 points over the base rate—the fail has nowhere to hide. A failed team of four leaves each member sitting at the base rate: one saboteur among four faces, and the fail says nothing about any single one. Passing runs the other way, and the exculpation deepens with size—shed 10 points riding a clean pair, 26 riding a clean four.

Team size is set by the table, two players a mission at five to seven, three or more above that, and it climbs as the game runs. The cheapest fail to make is a late one on a big team; the most expensive is an early one on a small team. The folk rule, lie low early, has a mechanism under it, and the mechanism is right.

§ 04 Evil

Evil still overpays

Cover works. Track the same Evil player after Mission 1: lay low and he's picked for 49.8% of later mission slots; strike and he's picked for 17.1%. Lying low nearly triples his access to the board.

And striking wins anyway. Controlling for player count, how many Evil sat on the team, and every optional role, striking Mission 1 is worth 6.1 points of win rate (p=6e-7). The split by team size is the part that matters, because the pricing predicted it before anyone looked: striking is worth 3.4 points on a team of two and 10.9 on a team of three. The bigger the team, the cheaper the fail, so the more a strike is worth—just as dividing suspicion by k implies.

The reason cover loses is that Evil is not one player.

Later-mission team access. Evil who sat on Mission 1: 49.8% if they laid low, 17.1% if they struck. Their partner: 31.9% if they laid low, 34.7% if they struck.Striking Mission 1 buries the striker and frees the partnerShare of later mission slots each Evil player gets picked for.020406049.8%17.1%The Evil who sat on Mission 131.9%34.7%Their Evil partnerEvil laid lowEvil struck
Fig 6 The sacrifice play. Striking Mission 1 costs the striker almost all future board access—and raises the partner’s. One Evil eats the suspicion; the other collects the fails.

When Evil strikes Mission 1, the striker is finished—17.1% access, frozen out. His partner's access goes up, from 31.9% to 34.7%. The fail hands the table a suspect, the table takes it, and the Evil who wasn't on that team walks free with a fail already banked. Burning one member's cover to buy tempo is a good trade, because the member being burned is not the member who has to finish the job.

Nor does patience buy a better endgame. Lying low produces more assassination shots—Evil reaches one in 59.9% of games instead of 49.5%—but the Merlin hit rate is 38.6% either way. Waiting and watching does not sharpen Evil's read on Merlin. It loses the board and arrives at the same coin flip.

Which leaves Evil doing the wrong thing where it costs most:

Striking Mission 1 is worth +2.8 points on a team of 2 (Evil lies low 78.5%) and +10.4 points on a team of 3 (Evil lies low 59.9%).Evil lies low hardest where striking pays mostBars: extra win rate Evil gets by striking Mission 1 instead of lying low.+0+4+8+12+2.8 ptsMission-1 team of 2(5–7 players)Evil lies low 78.5%+10.4 ptsMission-1 team of 3(8–10 players)Evil lies low 59.9%
Fig 7 Evil lies low hardest where striking pays most. On a Mission-1 team of three, striking is worth ten points—and Evil still lies low 60% of the time.

On a team of three, where striking is worth ten points, Evil lies low 59.9% of the time. It does adjust—78.5% on a team of two, where the cost of striking is real and the edge is only 2.8 points—so the instinct points the right way. It is nowhere near strong enough. This is observational, and Evil chooses when to strike, so strikers may simply be better players; the defense against that reading is the gradient, which the pricing called in advance and the data reproduces.

§ 05 Words

Words and deeds run on different clocks

Everything so far comes from what players do: cards, proposals, votes. Avalon is also talk. The Avalon-NLU corpus—20 hand-annotated games, 2,384 utterances, every Evil line labeled for deception—lets us ask whether Evil's talk runs on the same clock as Evil's cards.

Sabotage propensity versus verbal deception rate across the five missions. Card play climbs from 26.8% to 96.0%; verbal deception peaks at Mission 3 (65.1%) then falls to 54.5%.Words and deeds run on different clocksCard play escalates to the endgame; the talk peaks at Mission 3 and eases off.0255075100Mission 1Mission 2Mission 3Mission 4Mission 596.0%54.5%Sabotage propensity (cards)Verbal deception rate (talk)
Fig 8 Two clocks. Blue: sabotage on the cards. Amber: Evil talk coded deceptive in Avalon-NLU. Cards climb to the endgame; the talk peaks at Mission 3 and eases off.

It doesn't. The cards climb the whole way to the endgame, tracking the collapse of patience: sabotage propensity reaches 96% by Mission 5. Deceptive talk rises early, peaks at Mission 3 at 65.1%, then falls through Missions 4 and 5, while sabotage keeps climbing. When Evil is most dangerous on the board it sounds least like a liar. Its share of the conversation barely moves either—28.5% to 31.7% across all five missions. Same volume, different content.

§ 06 Talk

The talk is the evidence

Take the players the mission board says nothing about—the ones never sent on Mission 1. No cards, no fail to answer for, no pass to hide behind. Nothing.

Players never on Mission 1, picked for Mission 2. With Merlin: Good 42.8%, Evil 31.0%. In bare Resistance with no Merlin: Good 47.8%, Evil 38.7%.The table finds Evil with no evidence at allPlayers never on Mission 1—the board says nothing about them. How often is each picked for Mission 2?020406042.8%31.0%Avalon (Merlin in play)47.8%38.7%Bare Resistance (nobody knows anything)GoodEvil
Fig 9 The talk is the evidence. Players the board says nothing about. Evil are picked 11.6 points less often—and the gap survives deleting Merlin, so nobody in the room holds private information. What’s left is talk.

The table picks Evil for Mission 2 anyway: 31.0% of the time, against 43.1% for Good. An 11.6-point gap on players with no record. Split them by how they voted on the Mission-1 team—public, observable, a hard signal—and the gap survives in both groups, 11.3 points among the approvers and 12.7 among the rejecters.

And it isn't Merlin. Delete him: in bare Resistance, where nobody at the table holds private information about anybody, Evil are still picked 9.1 points less often. Merlin's private sight adds about 2.6 points on top of a read that was already there without him.

So roughly nine of those twelve points are talk. Tone, argument, the shape of a justification, what a player says when nothing has happened yet. In a game whose entire apparatus is information—a role that sees, missions that reveal—the channel that finds Evil is the one carrying no information at all.

§ 07 Provenance

Provenance, not behavior

Four optional roles complicate the game. Percival sees Merlin, but also sees Morgana, who is built to look like Merlin to him; Mordred is Evil but hidden from Merlin's sight; Oberon is Evil but hidden from his own team. A raw split by role says Morgana lowers Evil's win rate by 2.7 points—help for Good, from a role designed for offense.

Naive versus controlled effect of each role on Evil’s win probability. Morgana moves from −2.7 (naive) to +2.7 (controlled), crossing zero; Oberon −7.9, Percival −7.4, Mordred +4.3 under the controlled model.Naive vs controlled effect of each role on P(Evil win)Points added to Evil’s win probability. Controlled = logistic model over 11,829 Merlin games.-10-50+5Oberon-7.9Percival-7.4Mordred+4.3Morgana+2.7Naive splitControlled model
Fig 10 The wrong sign. Amber: naive split. Blue: controlled model. Morgana’s naive effect points the wrong way—it was measuring Percival, who ships with her 98.3% of the time.

Run a logistic regression controlling for player count and every other role across 11,829 games, and the sign flips: +2.7 for Evil, and not significant. The naive number was never measuring Morgana. It was measuring Percival, who is dealt alongside her 98.3% of the time and who helps Good by 7.4 points. Morgana's effect was Percival's effect, read off a correlation before anyone asked what produced it.

Which is the mistake the board invites all game. Judge the missions, and you stay one inference behind the game being played underneath them.