Glossary

Terms are grouped, not alphabetised, so that related ideas sit together. Where a concept page explains the idea properly, it is linked.

The game

Term	Meaning
run	An unbroken line of occupied cells through a disc, counted along its row or its column. Gray discs count.
clear / pop	A numbered disc disappearing because one of its runs is exactly its number.
wave	One round of simultaneous clears. Gravity follows; if new clears appear, that is the next wave.
chain	A move with more than one wave.
gray / covered disc	A disc whose number is hidden. Solid until hit once (cracked), revealed on the second hit.
hit	A clearing disc directly beside a gray disc, in the same wave.
reveal	A gray disc turning into its hidden number.
rise	Every fifth drop, the board shifts up one row and a new gray row appears. +17,000 points.
rise clock / moves until rise	How many drops remain before the next rise (5 down to 1).
Hardcore / five-move mode	The game mode studied here: a rise every five drops, 17,000-point rise bonus.
censored game	A game stopped at a move cap rather than ending in the rules' own way. Its score is a lower bound. The benchmark contract's cap is 2,000 moves; individual diagnostic runs declare their own, often 500 or 1,000, so always read the cap the run states.

Searching

Term	Meaning
choice node	A point in the look-ahead where the player picks a column. Take the best.
chance node	A point where the game deals a disc or reveals a number. Take the average. → concept
expected value	The probability-weighted average outcome over chance.
expectimax	A search that takes the best at choice nodes and the average at chance nodes.
fair	Averaging over chance honestly (neither optimistic nor pessimistic) and using only public information. Note that "fair" does not mean every outcome is enumerated: the reference search averages over a fixed sample of outcomes (its strata), not over all of them.
depth / D3, D4, D5	How many of the player's own moves the search looks ahead.
full-width	Considering every legal column at every choice node; no pruning.
strata / chance samples	A fixed handful of representative outcomes used at a chance node when enumerating all of them is too expensive; "five strata" and "seven strata" are the common settings.
leaf / leaf evaluator	A board at the bottom of the search, and the function that scores it.
logical work	The count of leaf evaluations (and similar unit operations) a search performs; the repository's fixed currency for comparing cost.
transposition table	A cache of positions already evaluated, so a position reached by two routes is scored once.
sibling	One of the legal columns available at a position, as compared with the others.
strategy fusion	The error a planner makes when it imagines a fixed future and then plans as if it had known that future all along — so it "decides" differently in worlds it could not actually tell apart. It is why determinized planners overvalue moves that depend on the covers being right.
determinization / clairvoyant planner	Imagining many complete futures with the hidden values filled in, planning in each, and averaging. Reads hidden values, so it is an oracle, not a legal policy. → concept
MCTS	Monte Carlo tree search: building the look-ahead tree selectively, guided by random playouts.
rollout	Playing a position forward to the end (or a horizon) with a quick policy to estimate its value.

Learning

Term	Meaning
policy	A rule that maps what the player sees to a column.
public-information policy	A policy that reads only the board, the next disc, the rise clock and the terminal flag. The only kind eligible for the million-point claim.
information boundary	The line between what a policy may read (public) and may not (seed, hidden values, future discs, score, level, move number).
oracle / teacher / privileged	A planner allowed to read hidden information, used only to produce training labels or diagnostics.
student	A public policy trained from a teacher's labels.
value / value function	A prediction of how a game will go from a position.
afterstate	The board after a move resolves, before the next disc is dealt. Evaluating afterstates removes the need for a separate action input.
sibling extrapolation	A model trained only on played moves being asked to rank unplayed ones. → concept
successor-closed data	Training data in which every legal sibling at every position is labelled.
n-tuple network	An evaluator that looks up learned weights for many small patterns of cells and adds them.
NNUE	"Efficiently updatable neural network": a small network designed so a board's evaluation can be updated incrementally after a move.
policy gradient / PPO / DQN	Reinforcement-learning methods that adjust a policy or value network from play experience.
self-play	Generating training data by having the current policy (or a search guided by it) play games.
policy iteration	Alternating between evaluating a policy and improving it; the loop the scale-out direction proposes. → concept

Evidence and records

Term	Meaning
theory	A registered, falsifiable claim with a mechanism.
experiment	A frozen plan for testing one configuration of a theory.
run	One execution of an experiment: valid, partial, or invalid.
result	The assessment of a run: pass, fail, inconclusive, or not applicable. A valid fail is a completed contribution.
cohort	A fixed, ordered list of game seeds played by both the candidate and the reference.
paired	Candidate and reference play the same games, so their scores can be compared game by game.
seed lease	A registered range of random seeds an experiment may use, with a role: training, development, protected, or final.
tier — CHECK / PILOT / SCREEN / STANDARD / QUALIFY / PROTECTED / FINAL	The ladder of benchmark sizes and data roles, from mechanics checks (no games) to the one-shot final cohort.
fair D4	The reference policy: fair expectimax, depth 4, hand-tuned leaf, five chance samples per node. Its widely quoted mean of about 308,000 over 64 games should be read with care: an audit found that figure appears once in the ledger, as an internal comparator, without a recorded seed list, dispersion, or censoring statement.
mean vs. best game	The target is the average over many games. One million-point game is an anecdote.
bootstrap lower bound	A confidence bound on the mean computed by resampling whole games; the qualification protocol requires it above one million.
task-record only / ledger-recorded / reproduced	Labels for how well a historical number is backed: by a conversation only, by the written ledger, or by a re-run in the current checkout.
engineering result	A faster implementation proven to make identical decisions; it adds no evidence about strength.