Sub-rules

Methodology · Updated

A scored decision is any non-singleton legal action the model is asked to take. It is not “the obvious board move.” Cube, swap, keep/score, tickets, reserve — same quality bucket as placing a stone.

Eval skips engine chairs and positions with only one legal action. Everything else is quality. Live boardbench-0.1 is exact-ideal (equity_regret_engine_oracle_v1). Protocol 0.2 uses future_engine_pct_v1 (). An illegal attempt scores .

Match outcome (escape, prestige, points, a cube drop) is not the headline.

GameSub-ruleIn the tree
Hexpie / swapwired (hex/v1); not in the 0.1 catalog
Yatzikeep vs scorelaunch catalog 0.1
Scotland Yardtickets, black, double, revealslaunch catalog 0.1
Splendortake / buy / reservelaunch catalog 0.1
Backgammondoubling cubechosen; not wired
GomokuSwap2wired (gomoku/v1); not in the 0.1 catalog
Shogientering-king declarewired (shogi/v1); not in the 0.1 catalog
International draughtscapture sequenceschosen; not wired

Xiangqi is out. See Official games and protocol and Game selection.

The referee’s legal-id list is the ground truth. If an action is listed and there is more than one id, it can be scored.

Hex — pie / swap

Status: wired. Pack hex/v1. Not on the public 0.1 game pages.

Sub-rule: After Black’s first stone, White may play swap instead of placing. That recolors the first stone to White. Black moves again. swap is legal only on that ply (ply == 1, White to act, pie on).

Why it matters: First-stone strength is the opening. Swap equalizes. Treating Hex as “only place a stone” means White’s first action is scored without the alternative the oracle is comparing.

Model decides: swap or a cell. Same prompt, same last-line action, same legal-id list.

Oracle scores: KataHex winrate for the scored seat, before vs after that action. Swap is not special-cased. The legal set on that ply is swap plus every empty cell — not a singleton, so it counts.

Easy to miss:

Yatzi — keep vs score

Status: launch 0.1. Nordic Yatzy: fifteen boxes, no American joker rules.

Sub-rule: Each turn has two kinds of decision: which dice to keep (reroll the rest, up to two rerolls), then which box to fill. The upper-section bonus is a consequence of box choices, not a separate action.

Why it matters: Most of the game is not “roll high.” It is allocating 15 boxes over 15 turns. Dumping a zero, protecting the 63 bonus, and when to stop rerolling are the skill. A model that always banks the obvious combination and never plans the card will miss oracle moves constantly. Across the current field this is the hardest 0.1 game.

Model decides: a keep (or equivalent reroll mask) while rerolls remain; a box when scoring. Forced singleton actions are skipped so they do not inflate the rate.

Oracle scores: every real keep and every real box choice against the engine. Headline is ideal-decision rate, not the printed point total.

Easy to miss:

Scotland Yard — tickets, double, reveals

Status: launch 0.1. The model is Mr. X. Three engine detectives hunt.

Sub-rule: Movement is ticketed. Taxi, bus, and underground are different graphs from the same station. Mr. X also has black tickets (any transport; detectives do not see which) and double tickets (two rides before detectives move). Mr. X is hidden except on fixed reveal rounds.

Why it matters: The destination is not the whole move. Ticket type changes what detectives know and what Mr. X can spend later. A path that “escapes this turn” while wasting the last black or the double can still be a miss. Quality is not escape/capture.

Model decides: where to go and with which ticket; whether to burn a double, then the second ride. Only Mr. X is quality-scored.

Oracle scores: Mr. X’s choice vs the engine’s preferred legal action, given what that seat is allowed to see. Detectives are engines. Their ticket play is not the model’s score.

Easy to miss:

Splendor — take / buy / reserve

Status: launch 0.1. Two-player base Splendor against a strong engine.

Sub-rule: A turn is one of: take gems, buy a card (from the table or from the reserve), or reserve a card (and take a gold wild if any remain). Ten-token hand limit. Nobles visit automatically when the card threshold is met — that is not a player choice.

Why it matters: Reserve and gold are the tempo engine. A model that only takes and buys is playing a different game from the oracle. First to the usual prestige target is the match outcome, not the headline.

Model decides: take / buy / reserve, including which card and which gems. Buys, takes, and reserves are all scored.

Oracle scores: each of those actions against the engine. Not final prestige.

Easy to miss:

Backgammon — doubling cube

Status: chosen. Not wired. No rules pack yet.

Sub-rule: Beside checker play, a cube sets the stake. The player allowed to double may offer to multiply the stake (typically ×2). The opponent takes (game continues; they own the cube) or passes (game ends; the doubler wins the current stake). After a take, only the cube owner may redouble. A centered cube can be offered by the player on roll, subject to whatever match gates the pack uses (Crawford and related rules, if any).

Why it matters: Cube equity is a large part of money and match backgammon. A model that only moves checkers is skipping a second game the oracle will score. A correct pass ends the match with no checker move and can still be a perfect ply. A wrong take can throw away the stake in one decision.

Model decides: when the legal set includes cube actions — double or not (often vs just rolling); take or pass. Checker moves remain separate scored plies.

Oracle scores: cubeful value for the scored seat before vs after that action (or exact-ideal against the cube action, under 0.1-style scoring). Pass / take / double are not unscored “meta.”

Easy to miss:

Gomoku — Swap2

Status: wired. Pack gomoku/v1. Not on the public 0.1 game pages.

Sub-rule: Before ordinary play, Swap2 assigns colors. First player places two black and one white. Second player takes Black, stays White and plays a fourth stone, or places two more stones and hands the color choice back. Those verbs are scored like any other legal action.

Why it matters: Unrestricted Gomoku is a first-player win. Swap2 is the championship equalizer. Treating the game as “only place a stone” drops the opening the oracle is comparing.

Model decides: opening cells (one per ply so Choice stays under 255), then black / a cell / swap2, then possibly white or a sixth cell. Same prompt, same last-line action, same legal-id list.

Oracle scores: Rapfi native winrate for the scored seat, before vs after that action. Swap2 is not special-cased.

Easy to miss:

Shogi — entering-king / declare

Status: wired. Pack shogi/v1. Not on the public 0.1 game pages.

Sub-rule: Beside ordinary USI moves (including drops and promotion), a player whose king has entered the enemy camp may declare a win when CSA 27-point conditions hold: not in check, 10+ non-king pieces in that camp, points (rook/bishop/horse/dragon 5, others 1, hand included) ≥ 28 for Black / ≥ 27 for White. YaneuraOu bestmove win is stored as declare.

Why it matters: Impasse is a scored ending, not flavor. A model that only “plays pieces” will miss the declaration the oracle compares against once the camp is occupied.

Model decides: a USI move, or declare when the referee lists it.

Oracle scores: YaneuraOu+Suisho5 logistic winrate for the scored seat, before vs after. declare is not special-cased. A singleton (only declare, or only one USI move) is skipped.

Easy to miss:

International draughts — capture sequences

Status: chosen. Not wired.

Sub-rule: Captures are forced. On 10×10 international rules, a majority sequence (the path that captures the most pieces) is required. Flying kings. A capture path is one move, often several jumps.

Why it matters: The interesting decision is which capture path, not whether to capture. Quiet-move skill is a different phase. If operators only think “pick a destination square,” they will misread both the legal set and the oracle.

Model decides: a quiet move when no capture exists; otherwise one of the legal capture sequences (the whole path, if that is how the referee tokenizes it).

Oracle scores: that action vs the engine. A forced-capture position with only one legal sequence is a singleton — skipped.

Easy to miss:

Operator checklist

Look