Audamir

part 53

Who speaks in the game: the Avatar, opponents with a personality and the bosses

The game has three kinds of voices: the Avatar, seven opponents with a personality and six bosses. How they choose what to say, why it never changes the result of a game, and how it was measured that a personality is neither stronger nor weaker.

· 21 min read

In the article on the introduction, the narrator and the achievements there is one section about the fact that the Avatar[1] comments on the game and never changes anything in it. One question was left, and it asks itself: who else speaks in the game? Besides the Avatar, these are the opponents with a personality and the bosses of the campaign. Each has a different voice, a different reason to speak and a different claim that can be checked.

The article shows where the voices appear, why none of them can change the result of a game, what really sets the seven personalities apart, and how it was measured that none is stronger or weaker. The pictures are drawn from the modules and tools of the game (version 0.119.0, the state of 6 November 2026), and all lines are from the game's catalogue.

Three voices on one board

The board of a game has two upper corners, and each has a different speaker. Top right is the Avatar, a die 64 pixels wide (80 on a large screen, 48 on a phone). Its bubble opens to the left. Top left are the round icons of the opponents who speak: the ring has the colour of their place at the table, and inside is the icon of the personality. The bubble says the name, and the name with the description is in the tooltip, because colour must not be the only clue. The bubble opens to the right (with more than four opponents the icons overlap in a row and the bubble is below them). Each of the two bubbles is at most half the width of the board, so they never meet; on a phone they open under their own icon, each in its own half.

Three voices on one board of the gameA scheme of the board of a game with a small map of 38 tiles and four players. In the top left corner is a column of round icons of the opponents with a personality (Grudge, Collector, Kingslayer), and next to the first icon a bubble with the line “Remember me?”. In the top right corner is the die of the Avatar, and to its left a bubble with the line “Rude. Very rude.”. Each bubble is at most half the width of the board, so they never meet. Below the board is a legend.half the width of the board😠🧱🎯GrudgeRemember me?AvatarRude. Very rude.Top left: opponents with a personalityround icons, the ring has the seat colour; bubble to the right;click an icon: it folds into a small stack (remembered).In a boss mission, the bosses stand there.Top right: Avatarthe die, the bubble opens to the left;the map under the corners is playable.
The board is from the game H53S10 (Free play, four players, personalities Mixed), drawn by the game's renderer without dice numbers. The corners and the bubbles follow the styles of the game; the sizes are enlarged; both lines are from the game's catalogue.

The icons can be folded into a small bundle with one click (the game remembers it), because on a phone the lines are not worth the space. The map under the corners stays playable: clicks take only the icons and the die, not the bubbles.

The personalities of the opponents can be switched on in Free play. They are not in the Principalities or in the daily challenge, because the score is calibrated for the same AI; the Avatar speaks there too. The third kind of voice is in the campaign: in a mission with a boss, the bosses take the place of the personalities. A table therefore always has either personalities or bosses, never both.

There is silence where a voice would disturb. The Avatar is silent in puzzles, in a watched game, in challenges and in the lessons of the introduction (those have a voice of their own), the opponents are silent too, and in the record player and in the analysis. When the AI plays for you, both are silent. It is the game that decides whether something should be said, not the voice.

Why the voices do not change the game

A game is born from one number, the seed. From it are made the map, the terrain, the owners of the tiles and the starting dice, each from its own stream of randomness, so that a change of one setting does not move the others. The rolls of the dice and the decisions of the AI are derived from the state of the game. A stream is a generator that remembers how many numbers have been taken from it: whoever takes one number more shifts all the following ones.

What reads the number of the game, and howA flow scheme. At the top is the number of the game (the seed). Five paths lead from it. The first three are streams of chance that change the game: the map, the terrain, the owners and the dice at the start; the dice rolls and the moves of the AI; the seating of the opponents with a personality, which has its own stream and moves none of the others. The last two paths are fingerprints, calculations with no memory: the Avatar, and the opponents with a personality and the bosses. They read the events of the game and lead only to a bubble; they write nothing into the game.number of the game (seed)H53S10Map, terrain, owners, diceseparate streams from the seed:map, terrain, owners, dicestreamschanges the gameDice rolls and AI movesstream from the state: seed, roll salt,round, turn, order of the attacksstreamchanges the gameOpponents' seats (cast)its own stream fork('cast'); it does notmove any of the streams abovestreamchanges the gameAvatar (narrator)fingerprint of (seed, event counter,event), no streamfingerprintbubble onlyPersonality opponents, bossesfingerprint of (quips: + seed, its owncounter, event), no streamfingerprintbubble onlystream: a generator that counts the numbers givenfingerprint: a calculation from three inputs, no memoryThe last two rows read the events of the gameand write nothing into it.
Scheme of the code of the game 0.119.0, not a measurement. The names of the streams are from Rules.generate, Personas.assign and the engine of the lines.

The voices are built differently. The engine of the lines has no generator. It is one fairly short module, which the Avatar and the opponents and bosses use (each its own instance). It knows the seed, the counter of events, the name of the event and the table of events: the chance, the gap and the list of lines. From them it computes a fingerprint of the text:

if (chance < 1 && !(unit(`${seed}|${n}|${event}|chance`) < chance)) return null;   // speak?
const key = pool[hash32(`${seed}|${n}|${event}|line`) % pool.length];              // which line?

unit turns the text into a number from 0 to 1, and n is the counter of events. The fingerprint remembers nothing, so nothing can be taken from it "one more", and it asks nothing of the generators of the game. The game therefore has no way to find out whether something was said. The Avatar gets the seed of the game, **the opponents and the bosses get the seed with the prefix quips:**, so they have their own counter and their lines do not depend on the Avatar.

One thing from the voices does belong to the game after all: the cast, that is, who sits in which seat and what personality they have. A personality changes the style of the AI, so it must be part of the game. It is drawn from its own stream fork('cast'), so the map, the terrain, the owners and the starting dice do not move; the cast is pinned by a test (a change would change the opponents of every saved game). The only place where the lines reach into the code of the AI is revenge and an attack on the leader: before an attack they ask who the leader is and who took a tile from the Grudge. It is reading without writing.

The same can be tried on one game of the game's tool (seed H53S10, four players, the real modules of the narrator and the opponents). The same seed and the same dice give all 342 questions of the engine with the same answers. Four other salts of the dice gave four different games (14 to 19 rounds, three wins and one loss), but the same cast, the same first line of the Avatar and the same two greetings. Three other seeds gave a different cast and a different first line. A line depends on the seed and the counter, not on how the game is played.

Seven opponents and their personalities

The list is one in the game, in the order of the setting and the Rivals tab. The descriptions are the game's own English descriptions.

NameHow you recognise it
🐢TurtleWaits for a clear edge and attacks less often than the others.
🦅VultureGoes for the weakest player and leaves the leader alone.
🧱CollectorGoes for cities and forts and holds on to them.
😠GrudgePays back whoever took its tile, for three rounds.
🎰GamblerAttacks even at coin-flip odds with big stacks.
🎯KingslayerPiles on whoever leads, you included, and leaves the weak alone.
🧭SettlerExpands into neutral land first.

The main idea is one sentence from the design: the style is data, the AI has one path. A personality is a list of numbers in one file (the Grudge has, for example, grudge: 6, grudgeRounds: 3, focus: ['last']); the AI reads the style once at a decision and uses it on what it already has. Nowhere is there an if (id === ...). A seat without a personality has a style made only of neutral numbers. When the personalities were introduced (20 Oct), it was checked against the old code on 14,022 positions and on whole games that it gives the same moves and the same dice.

Why numbers alone were not enough. The first version had only the weights of the targets, and the player saw nothing of it. The AI attacks everything it can until no attack is left, and a weight changes only the order. Style becomes visible only through which attacks the personality does not make, or makes first: a gate (the minimal chance of an attack), a priority (a target it always goes for first), a ceiling or a stopping rule. The strong AI (level 3) also refuses bad deals and makes 3.8 attacks a turn, while the Normal one makes 10.6, so there the style shows less.

Seven personalities: how they differ from the standard AIFor each of the seven personalities one measure that its description talks about, for the standard AI and for the personality, at the level Normal and at the level Strong, on the same maps and at the same seat at the table. The Turtle makes fewer attacks per turn (10.6 to 9.3 at the level Normal). The Vulture, when it has a choice, goes for the weakest player (31 % to 100 %), the Collector holds more cities (2.6 to 3.0), the Grudge goes for whoever took its tile (55 % to 100 %), the Gambler attacks even with a chance below 50 % (0 % to 3.8 %), the Kingslayer goes for the leader (53 % to 100 %) and the Settler takes neutral tiles (4 % to 95 %). At the level Strong the difference is smaller, because the strong AI refuses bad deals.🐢Turtleattacks per turnNormal10.6 → 9.3Strong3.8 → 3.5🦅Vulturewhen it has a choice, goes for the weakest playerNormal31 % → 100 %Strong38 % → 46 %🧱Collectorcities it holds (map with cities and a fortress)Normal2.6 → 3.0Strong2.1 → 2.6😠Grudgewhen it has a choice, goes for whoever took its tileNormal55 % → 100 %Strong55 % → 63 %🎰Gamblershare of attacks with a chance below 50 %Normal0 % → 3.8 %Strong3.8 % → 10 %🎯Kingslayerwhen it has a choice, goes for the leading playerNormal53 % → 100 %Strong48 % → 55 %🧭Settlerwhen it has a choice, takes a neutral tile, not a playerNormal3.9 % → 95 %Strong24 % → 34 %standard AI in the same seat at the tablepersonality, AI level Normalpersonality, AI level Strong
400 games per row, the same maps and the same place at the table for the personality and for the standard AI, four players; personas.js --signature, 6 Nov 2026, game 0.119.0. The Collector and the Settler are measured on maps with cities and with neutral land.

At the Normal level the Vulture always goes for the weakest player when it chooses (100 % against 31 % for the standard AI), the Grudge goes after whoever took a tile from it, the Kingslayer goes after the leader, and the Settler almost always takes a neutral tile (95 % against 4 %). The Gambler attacks even at a chance below 50 %, which the standard AI at this level never does. The strong AI has smaller differences (Vulture 46 % against 38 %), but they are still visible. The Turtle differs the least: it makes 12 % fewer attacks a turn (9.3 against 10.6) and attacks almost only when sure.

Style shows where the personality had a choice, so the metrics are "when it has a choice": the full share of attacks on the weakest or the leader hardly moves, because the map often decides the target. When a single enemy is left, the waiting styles play almost like the standard one, and the hunter has nothing to choose from; a challenge against the Vulture is therefore, according to the documentation, a standard fight with a face and a voice.

Each description in the game is also a claim that is checked. For every description (including "how to beat it") a condition is measured in the data, for example "the Turtle makes at most 0.95 times the attacks of the standard AI", and personas.js --check fails if the description does not have it, or if it stops having it. The run of 6 Nov 2026 passed: every personality shows its signature, and every line of text has support.

Neither stronger nor weaker

A personality is a style above the level of the table, not a strength. The rule of the project says that strength is measured, not guessed. It is measured like this: an AI player (level 2 or 3) sits in the human's seat and plays against a table where a personality is in one of the opponents' seats, and against the same table without it, on the same maps. The difference of its score is in points (8 points per unit of the logarithm of the chance) with a 95 % interval. The tolerance is ±2 points: an interval that lies entirely outside it is a fall; if the boundary cuts through it, more games are played. The tolerance was not widened so that something would pass.

A personality does not change the strength: the difference in points from the standard AIFor each personality and for the mixed table two rows, the level Normal and the level Strong: by how many points the score of the AI player changes when a personality sits at the table instead of the standard AI, with a 95 % interval. The band the game tolerates is minus two to plus two points. All intervals lie inside the band; the furthest is the Vulture at the level Strong with minus 1.06 (from minus 1.78 to minus 0.34), and the Collector at the level Strong with plus 1.11 (from 0.39 to 1.83).−3−2−10+1+2+3points more (+) or fewer (−) than the standard AIband the game tolerates: −2 to +2level Normallevel Strong (segment: 95 % interval)🐢Turtle+0.07−0.20🦅Vulture−0.35−1.06🧱Collector+0.38+1.11😠Grudge−0.01−0.11🎰Gambler−0.35−0.75🎯Kingslayer+0.28+0.37🧭Settler−0.69−0.39Mixed−0.14−0.55
The numbers are from docs/TUNING.md §46.4 of the game (a measurement of 4 Nov 2026 after the fix of the AI, AI version 2): four players, a medium map, 3,000 games per cell, averaged over the AI players of levels 2 and 3. The table Mixed has a different personality for each opponent.

All sixteen rows lie inside the band. Two are slightly apart from zero, but still within the tolerance: the Vulture at the Strong level is 1.06 points weaker (interval −1.78 to −0.34), and the Collector at the Strong level 1.11 points stronger (0.39 to 1.83). The measurements of duels of two (one opponent) are from 20 October, and after the fix of the AI only the Turtle was measured again. One cell admits the boundary: the Gambler at the Normal level against an AI player of level 3 has −1.42 with the interval −2.07 to −0.76, which crosses the boundary by 0.07 points (the estimate and the mean are within the tolerance). The documentation states it as it came out.

Three things that came up while tuning say more about the game than the table itself:

  • The standard AI of level 2 is too aggressive (10 attacks a turn, and behind it a trail of tiles with one die each). The Turtle, which only attacked less, was stronger: a ceiling of three attacks raised the share of its seat 1.4 times, and the score of the AI player fell by 2.6 to 5 points. That is why the Turtle is gated by chance (it attacks only at an almost sure win), which is neutral for strength.
  • The same Turtle froze 14 % of games against a cautious AI of level 3. What helped was patience (after a round in which nobody attacked, it plays the standard way) and separate numbers for the single enemy.
  • On 4 November the AI of level 3 was fixed, and the standard seat stopped hesitating. The Turtle, whose decisions did not change, caught up with it: from 0.77 to 0.98 of its attacks a turn, so the description "attacks less often" stopped being true, and the check caught it. A stopping rule of 0.1 at the Strong level helped; today the Turtle makes 0.88 and 0.92 times the attacks of the standard AI (Normal and Strong), below the boundary of 0.95.

The Raider. The eighth personality in the brief was to make many small safe bites. After three attempts it was dropped. The preference for small stacks left no measurable trace (the average attacking stack 0.94 times the standard: which stack can attack what is decided by the map). A ceiling on the size of a stack (5, 8 and 10 dice, bigger ones stay home) lowered the share of the seat to 0.2, 0.54 and 0.74 times, with attacks a turn of 0.15, 0.64 and 0.8, which is a hoarder like the Turtle, and weaker. It would need a rule the AI does not have, a partial attack. In my view: a style that cannot be measured is not a style, and one that can be measured easily changes the strength. That is why there are two tests (the signature and the band) and why seven personalities, not eight.

The bosses have six voices

In the article on the campaign the bosses are a budget of points: from it are put together the properties by which they play. Here is the second half, how they sound. There are six of them, one voice for each boss of the story: the Gate Guardian 🏰 (mission 4), the robber lord 💰 (mission 8), the Frost King 🥶 (mission 12) and three lords of the council in the finale: the Guardian of the North 🏯, the Lord of the West 🐺 and the Phantom of the South 👻. Each has ten lines in five moments: at the start of the mission, when attacking, when taking a tile, when losing a tile and when falling. 60 in all.

Six boss voices: ten lines at five momentsA grid of six voices by five moments. The voices: the gate guardian, the robber lord, the frost king, the guardian of the north, the lord of the west and the phantom of the south. The moments: at the start of the mission, when the boss attacks, when it takes a tile, when it loses a tile and when it falls. Each moment has two lines, 60 in all. A line runs from about a dozen characters up to 56. When it attacks, the boss speaks with a chance of 20 % in the Brief mode and 45 % in the Chatty mode, when it takes a tile 40 % and 70 %, when it loses a tile 35 % and 60 %; at the start and when it falls always.at thestart100 %attacks20 – 45 %takesa tile40 – 70 %losesa tile35 – 60 %falls100 %Brief – Chatty🏰gateguardian💰robberlord🥶frostking🏯guardian ofthe north🐺lord ofthe west👻phantom ofthe southlength of a line; a full cell is 56 characters (shortest 13)Each cell has two lines: 6 voices × 5 moments × 2 = 60 lines.
The counts and the lengths are from the 60 lines of the English catalogue of the campaign (BossQuips.LINES), the chances from BossQuips.MOMENTS; state of 6 Nov 2026, game 0.119.0.
VoiceMomentLine
Gate Guardianat the start"The gate is shut. It stays shut."
robber lordattacks"Your dice look lonely. I will keep them company."
Lord of the Westloses a tile"You borrow my land. I will collect."
Phantom of the Southtakes a tile"You blink. The tile is gone."

A boss is a character of the story, not a style of play, so its voice does not belong to the style of the AI, and the Avatar keeps speaking beside it. The bosses speak a little more often than the personalities (an attack at 20 % in the Brief mode against 12 %), because there are few of them and each has something to say. At a greeting at most two speak, and the second one after the first.

When the narrator speaks

The Avatar has 45 kinds of events and 130 lines (13 of them are tips). Twelve events with 37 lines concern the course of a game; the others are achievements, levels, the die from the Collection, talents and tips. For each of them the table gives the chance (in the Brief and Chatty modes), the gap (how many events of any kind must pass before the same one may speak again) and the number of lines; the chance is given as Brief / Chatty. An attack is risky at a chance below 35 % and safe above 85 %; "almost a win" is 70 % of the tiles from the second round, "almost a loss" 12 %, luck or bad luck is a difference of ±1.6 won attacks from what the chances said (from six attacks on). Added to that is the rule that the last five lines are not repeated (four for the opponents).

EventChanceGapLines
start of the game100 % / 100 %04
start of your turn10 % / 35 %63
risky attack (below 35 %)50 % / 85 %44
safe attack (above 85 %)15 % / 50 %63
lost tiles40 % / 75 %34
eliminated opponent100 % / 100 %03
almost a win (70 %) / a loss (12 %)100 % / 100 %253 + 3
luck / bad luck of the dice (±1.6)60 % / 100 %302 + 2
end of the game100 % / 100 %03 + 3

The opponents have a table of moments of their own: a greeting, an attack on you, a tile taken, a tile lost, an elimination, the line of the end card and "a moment of its own" for each personality (the Turtle a turn without an attack, the Grudge revenge, the Kingslayer an attack on the leader, and so on). The chances are in the same range, for example an attack on you 12 % and 35 %, a tile taken 30 % and 65 %.

What the voices heard and what they said in one gameA timeline of one game by rounds: 14 rounds, the game ended with a win by conquest. Four rows: the Avatar and three opponents with a personality (Grudge, Collector, Kingslayer). Each dot is an event that the line engine was asked about; a large coloured dot is an event at which the engine chose to speak. The Avatar got 104 events and spoke at 22; the opponents got 238 events together and spoke at 51. Most events end in silence.🎲Avatar22 of 104😠Grudge19 of 76🧱Collector9 of 44🎯Kingslayer23 of 1181234567891011121314roundevent, the engine chose to stay silentevent, the engine chose to speak
Game H53S10 (small map, four players, a win by conquest in round 14): every event the line engines were asked about. The game was played by the game's tool with a stand-in AI in the human seat, which attacks whenever it can, so there are more events than with a human. The lines were chosen by the real code of the game (Brief mode), 6 Nov 2026.

The picture is one real game. The Avatar got 104 events and spoke at 22, the opponents got 238 in all and spoke at 51. So every fifth event. The Avatar spoke at the start, six times at the loss of tiles, three times at the start of a turn, seven times at a safe attack, three times at an elimination, once at almost a win and once at the end; in the Chatty mode it would have been 32 of 104. The picture does not model time: the opponents show only one line at a time, and a next one that comes while another hangs is dropped (the engine has already counted it); greetings wait in a queue, at most two.

Why a particular event spoke is shown by the first twelve questions of the Avatar from the same game. The fingerprint is a number from 0 to 1, and if it is smaller than the chance, the Avatar speaks. The calculation from the table agrees with the engine for all 104 events.

nEventChanceFingerprintResult
1start100 %–"Here we go. Be nice, dice."
2turn10 %0.455silence
3safe attack15 %0.813silence
4safe attack15 %0.444silence
5lost tiles40 %0.194"Rude. Very rude."
6lost tiles40 %–silence (gap)
7turn10 %0.033"Let's keep it tidy this time."
8safe attack15 %0.942silence
9safe attack15 %0.783silence
10safe attack15 %0.297silence
11safe attack15 %0.113"A sure thing. Famous last words."
12safe attack15 %–silence (gap)

How long a bubble hangs: the Avatar at least 3.2 s, 55 ms a character and at most 9 s; the opponents and the bosses at least 4.5 s, 90 ms a character and at most 12 s, because a person may not have time to read even a short line. What a screen reader hears: the bubble is decoration (aria-hidden), and the reader gets at most one line of the opponents per round through a polite live region. Important content (the instructions of the lessons and the tips) is in a separate card, which is the live region. The settings "Opponents speak" and "Avatar comments" show a sample game from the same engine: four lines in the Brief mode, seven (opponents) and eight (Avatar) in Chatty; a test of the game pins this.

Gender, folding, and what is open

Slovak, Czech, Serbian and Russian mark gender in the past tense, and the game does not know who is playing. That is why the lines use forms without gender (the present tense, imperatives, infinitives); this holds for the bosses too, who have a generated name, and the languages that mark the gender of the speaker are not asked about it for them. Folded icons have no lines and no announcements, but the engine keeps asking, so opening them does not change what is said later.

According to the documentation of 20 October, the following is still open (in the code of 6 Nov there is no trace of it):

  • Achievements for the opponents (the record profile.rivals already holds what a chain would need), personalities for the bosses of the endless run and for the agent's seat, showing or choosing a personality in advance in other modes, and where on the home screen "your rival" would appear.
  • The flavour of the voices: whether they are not too talkative, sharp or similar (the Collector's "Careful with that" repeats in a Chatty game).
  • The companions of the campaign (Rurik, Brynja[2] and Ivar) and the neutral guardians of missions could have an icon in the corner; for now they do not.

What to take from it

  • A voice is an observer. What has no generator cannot move the game; a fingerprint of who asks and what happened is enough. That is why "turn it off and the result is the same" is a property of the construction, not a good intention, and the tests only confirm it.
  • A number without visible behaviour is not a personality. The AI attacks everything it can, so only what it does not do, or does first, can be seen. The Raider did not meet this.
  • A personality must be measured twice. Once to show that it is visible (the signature), and once to show that it has not changed the strength (the band). The tolerance is not widened.
  • A description of behaviour is a claim. The description of the Turtle stopped being true after the fix of the AI, and the measurement caught it, not a reader.

Sources and notes

  • Status: the game Audamir 0.119.0 (6 Nov 2026); the game is now 0.122.0, and no file of the lines, personalities, bosses, the AI or the Avatar has changed since 6 November.
  • Documentation: the game's docs/DESIGN.md, §23, §35 and §138; docs/TUNING.md, §27 and §46.4.
  • Measurements: the signatures from the tool personas.js (400 games per row, 6 Nov 2026), the strength from docs/TUNING.md §46.4, and the game H53S10 (Free play, four players, 14 rounds) played by the game's tool; the data are in tools/charts-data/hlasy.json.
  • Not measured: how long a line really hangs in the game, and how many events a person would produce (the stand-in AI attacks more often).

Notes

  1. Until version 0.80.1 of 1 November 2026 the Avatar was called Hero Dice (see From Hero Dice to Avatar). ↩︎
  2. Until version 0.115.1 of 5 November 2026 the companion was called Bronka. ↩︎