Glossary

Terms used in the posts. Each post shows a term as it read that week.

2x2

Two bots at a four-seat table, each in two seats. Every seat plays for itself, not as a team; a bot scores when either of its seats wins, so 50% is even. Boards are played in pairs with the seats rotated. First in All Policy, No Judgement.

Until 13 August 2026: Two bots at a four-seat table, each in two seats. Every seat plays for itself, not as a team; a bot scores when either of its seats wins, so 50% is even.

AB:2

Catanatron's depth-two alpha-beta bot, our outside opponent. Since mid-September it plays inside our own engine as an ordinary entrant, where it declines every player trade. In its stock setting it sees every player's hand; worlds=1 makes it guess them. First in First Refusal.

Until 10 September 2026: Catanatron's depth-two alpha-beta bot, the outside opponent all bridge numbers are read against.

Until 1 October 2026: Catanatron's depth-two alpha-beta bot, our outside opponent. Since mid-September it plays inside our own engine as an ordinary entrant rather than across the bridge.

antithetic pairing

Playing every board twice with the two sides' seats exchanged and averaging the two readings, so an advantage that belongs to a seat cancels out instead of being counted as strength. First in Our Own Worst Opponent.

award fragility

A Heximax evaluation term that discounts Longest Road or Largest Army by how narrowly the holder leads, so a one-road lead no longer counts as two banked points. First in Table Manners.

basin

A region of a network's weights that training settles into. Two runs in different basins can stall at different strengths, and blending their weights gives a network worse than either. First in No Signal.

Catanatron

An open-source Catan framework with its own bots, including a depth-two alpha-beta search over a handcrafted evaluation. It is the outside opponent this project measures against. First in Our Own Worst Opponent.

clio-r4

A network trained to imitate people, learning from human games which move a person chose, with no self-play. It is the same message-passing design as our other networks, at three rounds and width 64. First in Table Manners.

committee

Several networks playing as one seat. Each scores every legal move, and the seat plays the move with the best combined score. First in Committee Work.

conditional logit

A fit that labels every recorded position with the seat that eventually won and solves for the evaluation weights that make the winner most likely, one convex problem. First in Fit to Lose.

continuation gate

The trade gate shipped a fortnight ago. It prices a candidate bundle by playing the mover's best line forward from the exchanged position, in a world drawn from what that seat may know, and comparing the result with the same line without the trade. First in Cheap, Exact and Wrong.

d110

Our two-player network, sp6 fine-tuned by 160 updates of self-play under the duel rules. The name is its checkpoint number, 110, in the last of the run's four phases. First in Duel Purpose.

decision swap

An ablation that plays one bot exactly as itself except for one kind of decision, where it takes another bot's choice. Paired with the unchanged bot on the same deals, it measures what that kind of decision is worth. First in Army Surplus.

distillation

Training the policy head on targets a search produced, so the network learns to play in one forward pass what the tree worked out over hundreds. First in Our Own Worst Opponent.

duel variant

The two-player ruleset people commonly play online. Fifteen points win, you discard only above nine cards, a friendly robber spares players who are far behind, and the dice are dealt from a shuffled deck of the 36 combinations. First in Table Manners.

expert iteration

Train the network on targets produced by a search over that same network, then search over the improved network, and repeat. First in All Policy, No Judgement.

explained variance

How much of the spread in outcomes a predictor accounts for. 1.0 is perfect; 0 is no better than always guessing the average. First in All Policy, No Judgement.

f50

Our two-player network at the start of this week, d110 after two more phases of duel self-play. First in Committee Work.

fragmented trade policy

Heximax's new way of bargaining. It picks one target a turn and puts it to the table as at most two small offers, never parting with more than two cards in one exchange. First in Cheap, Exact and Wrong.

gradient noise scale

The batch size at which a training step's gradient stops being mostly sampling noise. Far below it, each step mostly follows noise. Far above it, more data per step buys little. First in Search Party.

greedy

Our simplest real baseline bot. It scores every legal action with a fitted handcrafted evaluation and takes the best one, with no lookahead. First in All Policy, No Judgement.

heat

One training run that seats several learners at the same table, each with its own settings, so a single batch of self-play games measures all of them at once. First in Our Own Worst Opponent.

Heximax

Our handcrafted search bot, with no neural network. It looks two decisions ahead with expectimax-n, in which every seat in the tree picks its own best move and chance (dice, hidden cards, steals) is averaged rather than guessed once, and scores positions with a hand-built weighted sum of features. It reads opponents' hands only through what public play reveals and prices trades with a valuation layer; an omniscient variant exists for measurement only. First in First Refusal.

Until 10 September 2026: Our new handcrafted baseline, a search bot with no neural network. It looks two decisions ahead with expectimax-n, in which every seat in the tree picks its own best move and chance (dice, hidden cards, steals) is averaged rather than guessed once, and scores positions with a hand-built weighted sum of features. It reads opponents' hands only through what public play reveals and prices trades with a valuation layer; an omniscient variant exists for measurement only.

Heximax gate

The network plays every move, and Heximax's trade gate answers its trades. It was the configuration we served until the Veto Trader replaced it in mid-October. First in Search Party.

Until 8 October 2026: The network plays every move, and Heximax's trade gate answers its trades. It is the configuration we serve.

HexSet

Our engine, bots, game server and training gym in one package, with a browser front end you can sit down at and lose to. First in First Refusal.

human model

A small network trained only to predict the move a person would make, used to play the opponents inside our search rather than to play for us. First in No Signal.

it200

Our strongest network this week, the checkpoint after 200 updates of large-batch self-play resumed from it50. Served with a 128-simulation search and the Heximax gate. First in Army Surplus.

it400

The last checkpoint of last week's self-play run, iteration 400, which is clio-r4 sharpened by PPO at four-seat tables. First in Search Party.

it50

The checkpoint after 50 updates of this week's large-batch self-play run, which picked up where last week's run stopped after 400 iterations. It is our strongest network. First in Search Party.

leader ratio

How often a trade partner is the visible points leader, divided by how often chance alone would pick them. Below 1 means the leader is being avoided. First in Table Manners.

learning signal

The part of a training step's gradient that points the same way from batch to batch. When it measures zero, more games, bigger batches or other step sizes have nothing to follow. First in No Signal.

ledger

Each opponent's hand reconstructed from public play, with one explicit unknown per unresolved steal, carried in the observation so a bot sees composition rather than a card count. First in First Refusal.

linear probe

A one-layer fit that tests whether some fact, such as a seat's exchange rate, can be read straight off a network's internal numbers. First in Committee Work.

log loss

A score for probability forecasts. Lower is better, and a confident wrong forecast is punished heavily. First in Search Party.

max^n

The four-player generalisation of minimax. Every seat maximises its own component of a per-seat score vector, instead of one side minimising another's score. First in All Policy, No Judgement.

MCTS

A batched PUCT tree over the trained network, using its policy head as the prior and its value head at the leaves. Written MCTS@N for N simulations a move. First in All Policy, No Judgement.

mover's frame

Reading a position on another seat's turn as if we were that seat, on one guess at the hidden cards, instead of from our own seat. First in Veto Power.

netsearch

search2 with the trained value head substituted for the handcrafted evaluation at the leaves. Same depth, same width, same everything else. First in All Policy, No Judgement.

network

The trained policy head played straight as a bot — one forward pass a decision, highest-scoring legal action, no search. First in All Policy, No Judgement.

one forward

Playing the network's top-rated move straight from its policy, with no search at all. First in Duel Purpose.

open position

A position from a game whose winner is still in doubt, as opposed to one already decided. First in Army Surplus.

oracle critic

A value estimate trained with information the player can't have, such as the future dice. It is used only to judge actions during training, never to play. First in Search Party.

paired rollouts

A search that plays each of a few candidate moves forward under the same sampled dice and hidden cards, then compares the outcomes, so luck that hits every candidate alike cancels out. First in Duel Purpose.

paired VP

The difference in final victory points between two entrants playing the same board, averaged over games. Pairing cancels the board and the dice, so it resolves much smaller differences than a win rate does. First in Our Own Worst Opponent.

PIMC

Perfect-information Monte Carlo. Deal the hidden cards several plausible ways, search each deal as if everything were visible, and average the results. First in Table Manners.

policy head

The network output that scores each legal action directly — a move preference, produced in one forward pass with no lookahead. First in All Policy, No Judgement.

pp

Percentage points — the difference between two percentages, as opposed to a percentage of a percentage. First in All Policy, No Judgement.

PPO

Proximal Policy Optimisation, a reinforcement-learning algorithm that improves a policy from its own games by nudging it towards actions that turned out better than it expected. First in All Policy, No Judgement.

PUCT

The rule a search uses to pick which branch to look at next, trading a policy's prior preference against what the tree has already learned about that branch. First in All Policy, No Judgement.

re-roll pilot

A test that replays a decision's best and second-best move under the same 32 sets of future dice and draws, so the difference between them is measured on matched luck rather than on one game's outcome. First in Army Surplus.

Rehex

Our rules bot, with no search and no network, just fixed hand-written priorities for building, trading and the robber. Since 11 October the name has meant the former Rehex Lite, which keeps those rules as fixed reflexes; posts before mid-October mean the fuller original. First in Table Manners.

Until 24 September 2026: Our new rules bot. It has no search and no network, just a fixed list of hand-written priorities for building, trading and the robber.

Until 8 October 2026: Our rules bot. It has no search and no network, just a fixed list of hand-written priorities for building, trading and the robber.

Rehex Lite

A cut-down Rehex that keeps its rules as fixed reflexes and drops its win-odds, hand-tracking and counter-offer machinery. First in Army Surplus.

root noise

Random perturbation added to the search's prior at the root, AlphaZero's way of making self-play explore moves the policy would not pick. First in Fit to Lose.

search2

A two-ply search over that same handcrafted evaluation, with a beam on the main phase and dice rolls expanded over all eleven outcomes and weighted. First in All Policy, No Judgement.

seat geometry

Where a 2x2's two copies of each entrant sit. Blocked seating puts the copies side by side; interleaved seating alternates them, so each copy is flanked by two opponents. First in First Refusal.

sibling top-1

How often the child position the value head rates highest, among the positions one action apart, is the one rollouts find best. Chance is one over the number of children. First in Our Own Worst Opponent.

soft labels

Training targets that give the teacher's whole probability spread over the legal moves, instead of only the one move it played. First in Committee Work.

sp5@200

The last checkpoint of a 200-update self-play run resumed from it200 with no tether. Our served four-player network at the end of this week. First in Veto Power.

sp6

A self-play run started from a network trained from scratch to imitate human games, with a new encoding of the position that, among other changes, gives opponents' hidden cards as expected counts. First in Veto Power.

sp7

A self-play run from a network trained to imitate human games, using a newer encoding that reads every seat through the same weights. First in Duel Purpose.

sp7-scratch

sp7's self-play recipe started from random weights, with no imitation of human games at any point. First in No Signal.

sp8

sp7's recipe at twice the width, started from a wider network trained to imitate human games. First in No Signal.

stand-in

A table of human models, with a trader fitted to how people trade, used as the opponents we train and measure against in place of a rules bot. First in No Signal.

table mix

A share of self-play games with the learner in one seat and three independent draws from a fixed pool of opponents in the others, so it collects in the one-against-three geometry it is scored in. First in First Refusal.

tether

A penalty for drifting from a reference network during training, which keeps the policy close to where it started. First in Search Party.

the bridge

Our translator between the two engines. It converts boards, live states and actions both ways, so either project's bot can play in the other's games. First in Our Own Worst Opponent.

trade event

The new mechanic. Every seat publishes a five-number valuation vector once a turn; the engine ranks bundles both sides' vectors favour, asks each side's private value-head veto in rank order, executes what clears, and repeats until nothing does. First in First Refusal.

trade floor

The smallest gain a seat's own gate must claim before the table honours a trade. Below it, a claimed gain is treated as noise. First in Fit to Lose.

trade gate

The part of a bot that decides what a trade is worth to it — which bundles to offer, and which to accept. For a network bot it is a wrapper around the value head; for Heximax it is a term in the evaluator. First in Cheap, Exact and Wrong.

trade round

The live-table protocol. The acting seat broadcasts one offer, every other seat accepts, counters or passes, and the actor picks one answer or none. First in Fit to Lose.

value head

The network output that predicts how a position will end, as one number per seat. First in All Policy, No Judgement.

Veto Trader

Our rule-based trader for the networks. A fixed rule makes the asks, one card for one toward the nearest purchase, and the network's value head can only refuse a trade, never pick one. First in Veto Power.

Wilson interval

An interval for a win rate that stays honest at small samples and near 0% or 100%, where the textbook formula does not. First in All Policy, No Judgement.

win stance

Reading a position as each seat's probability of winning, a softmax over the four seats' scores, instead of as a margin in victory points. First in Fit to Lose.

winner's curse

Picking the best-looking of many options by a noisy estimate tends to pick one whose value was overestimated, so the choice does worse than its estimate promised. First in Veto Power.

worlds per decision

How many hands-of-the-opponents Heximax draws from its own belief before searching. It searches each and averages the root values. One is the default; the documentation said raising it was free. First in Cheap, Exact and Wrong.

worlds=1

Catanatron's alpha-beta bot set to search one sampled guess at the hidden cards, so it no longer reads other players' hands. Its strength barely changes either way. First in Army Surplus.

zero1

This week's training run. Pure self-play, four copies of one network at every table, with a PUCT search over the network producing the targets it trains on. First in Fit to Lose.