Formal Specification
Stone Stacking — Formal Specification
New Zealand · dexterity-small · Confidence: low
← Back to game pageImplementation Suitability
Suitability: adaptation-requiredFidelity: probabilistic-proxyGap-fills: 5
# Tū Kōhatu — AI-optimized specification
## General idea
Stack stones as high as possible without collapse. Each placement raises both the
score and the probability that the whole tower falls, which makes it a clean
stop-or-continue problem: the only real decision in the game is when to stop adding.
## Formal model
- **Abstraction taken:** stone geometry and contact mechanics are replaced by a
per-placement collapse probability that rises with stack height and with the
irregularity of the stone placed. No physics is simulated.
- **Player/agent count:** one or more, but players do not interact — each builds an
independent stack.
- **Turn model:** sequential placements within a player's attempt.
## Formal rules
R1. A player has a supply of stones and an empty base.
R2. A turn is `place(stone)`, adding one stone to the top of the stack.
R3. A placement collapses the stack with probability `p_fall(height, stone_shape)`,
strictly increasing in height. [gap-fill] No source gives a stability model;
monotonicity in height is the one property the attested rule requires.
R4. A collapse ends the attempt, and the attempt scores zero. [gap-fill] The entry
marks rebuild-within-attempt undocumented; scoring a collapse as zero is what
makes stopping a real decision.
R5. A player may `stop()` at any time, banking the current height as the score.
[gap-fill] Stopping is not described in the source; without it the game has no
decision at all and every attempt ends in collapse.
R6. The highest banked stack wins.
R7. [gap-fill] Players build simultaneously and independently. The entry marks
turn structure undocumented.
## Edge cases
- **Tie at the end condition:** equal banked heights are a genuine tie and are common
with small stone supplies. Resolve by a single additional stone placed by each tied
player, repeating until one collapses.
- **Player with no legal move:** a player whose stone supply is exhausted can no
longer place and is treated as having stopped, banking the current height.
- **Repetition / cycles:** no cycles. Height is strictly non-decreasing within an
attempt, and a collapse terminates the attempt rather than returning to a prior
state.
- **Simultaneous claims:** stacks are independent under R7, so there is no shared
resource unless players draw from one stone pool — in which case a contested stone
goes to the lower stack, to avoid rewarding the leader.
- **Mid-round elimination and turn-order repair:** a collapse eliminates that
attempt, not the player. With simultaneous play there is no turn order to repair.
- **Resource exhaustion:** the stone supply is the exhausting resource and its
exhaustion forces a stop (see above).
- **Termination guarantee:** the collapse probability is strictly increasing in
height and bounded below by a positive constant beyond the first few stones, so
every attempt ends in collapse or a stop in finite expected placements.
## State representation
Persistent: stone supply size and the shape distribution over stones.
Per-player: current height, banked height, `building|stopped|collapsed`, stones
remaining.
Per-turn ephemeral: the stone drawn and its shape parameter.
Information set: a player observes its own stack fully and, under R7, the opponents'
current heights — which is what makes the stop decision strategic rather than purely
a risk calculation.
## Action space
- `place(stone_id)` — precondition: player is `building` and stones remain.
- `stop()` — precondition: player is `building`.
## Playability notes
- **Single-player-vs-AI:** natural, and genuinely interesting. It is an optimal
stopping problem, and the optimal policy depends on the opponents' banked heights,
so it has more structure than the physical activity suggests.
- **Group games (n>2):** scales to any n without modification, since stacks are
independent. Reduction to two agents preserves the stopping problem exactly.
- **Physical-skill model:** one parameter — collapse probability as a function of
height and stone irregularity. A steadier player has a lower coefficient.
- **What is lost:** the hand. The whole activity is the felt moment of balance as a
stone settles, and a probability cannot represent the tactile feedback that tells a
real player whether to risk another stone. Note also that R5, the stop action that
creates the entire decision problem, is a [gap-fill] — the attested game may
simply be "stack until it falls", which would have no decision in it at all.
## Implementation hints
Implement the collapse probability first and check that the optimal stopping height
is neither 1 nor the whole supply; if it is, the probability curve is wrong and the
game is degenerate. Branching factor is 2 at every step — place or stop — so the
optimal policy is computable exactly by backward induction rather than search.
Download as JSON
Download the full game corpus entry for this game as structured JSON.
Download JSONCite this page
Stone Stacking (New Zealand). (n.d.). Games of Countries. https://gamesofcountries.com/games/nz-new-zealand/tu-kohatu
@misc{gamesofcountries-nz-new-zealand-tu-kohatu,
title = {Stone Stacking (New Zealand)},
year = {n.d.},
publisher = {Games of Countries},
url = {https://gamesofcountries.com/games/nz-new-zealand/tu-kohatu}
}