Game Night Theory is becoming a strategy engine for games: calculate the position, compare legal moves, show the strongest continuation, and explain the reason in human language.
Stockfish-style experience
Connect Four Strategy Lab
Play directly on the board. GNT ranks the available moves and lets you inspect why one move is stronger than another. This browser prototype uses a lightweight search evaluator; the production Connect Four engine should plug into an exact solved database.
The engine will compare your move against the strongest candidates.
Candidate moves
Production goal: exact game-theoretic value + best move + principal variation + alternatives + visual refutation. “Bad move” should never be the whole explanation.
Probability engine · Hidden information
Liar’s Dice Decision Lab
Enter the table state, your hand, and the current bid. GNT calculates the probability that the bid is actually true, then compares calling against continuing the bidding. This prototype uses the common rule where ones are wild on non-one bids; you can turn that off.
Click a die to change it manually, or roll an entirely new hand.
3 · Current bid
Probability bid is true
—
Enter the position
GNT will separate the raw probability from the strategic decision.
What the engine knows
Your dice are known. Opponents' dice are unknown. For each unknown die, the chance of matching a normal face is 2/6 when ones are wild, or 1/6 without wild ones.
Possible raises to examine
Lesson 01
Count what you know
Your own matching dice reduce how many successes must exist among the unknown dice. That can radically change whether an aggressive-looking bid is actually plausible.
Lesson 02
Probability ≠ automatic call
A bid being unlikely does not automatically mean “call.” Opponent tendencies, your dice count, turn order and the quality of your available raises all affect the strategic decision.
Lesson 03
Your hand changes bluff equity
Holding several copies of a face makes a raise on that face safer. Holding none can make the same numerical bid much more dependent on unknown dice.
Lesson 04
Think in distributions
Don't ask “how many fours are probably out there?” Ask for the probability of at least the number required by the bid. Liar’s Dice lives in the tail of the distribution.
Lesson 05
Player models matter
A production engine should learn whether an opponent underbids, bids close to expectation, or bluffs aggressively—and adjust the call threshold accordingly.
Lesson 06
Review the reveal
After a challenge, enter the revealed dice. GNT can compare the decision to the information available at the time instead of judging only by the lucky or unlucky outcome.
Connect Four · Deep lesson
From “center is good” to actual game theory
Classic Connect Four is a solved 7×6 perfect-information game. With optimal play, the first player can force a win. The useful teaching layer is explaining how a position becomes forced.
01 · Threats
Playable beats pretty
A three-in-a-row only matters immediately if its winning square is playable. GNT should highlight that square and whether the opponent can answer it.
02 · Forks
Two threats beat one
A move that creates two independent winning replies can force the opponent into a losing choice. The engine should visually mark both threats.
03 · Support
Gravity changes everything
Every disc changes which squares become playable later. A move may matter because of the future support square it creates, not its current pattern.
04 · Center
Why the middle matters
Central columns participate in more possible winning lines. But a real engine evaluates the exact position instead of blindly following a slogan.
05 · Defense
Find their threat first
Before attacking, identify every immediate winning square for the opponent. A flashy move that allows a forced win is a tactical error.
06 · Refutation
Show me why I'm wrong
GNT should show your move → best reply → resulting threat → best continuation. That is far more useful than a mysterious numerical score.
07 · Move order
Forcing moves first
Search becomes much more useful when immediate wins, blocks, and high-value central moves are examined before quieter alternatives.
08 · Position value
Win, draw, or loss
For a solved game, the ideal score isn't “+37.” It is the exact game-theoretic result under perfect play, plus a best continuation.
09 · Training
Turn mistakes into lessons
Save positions where the player missed the best move, classify the mistake, then generate similar positions until the pattern becomes recognizable.
Expanded curriculum
Every game gets the right kind of engine.
A single algorithm cannot handle every game. GNT should keep one consistent learning interface while swapping the mathematical engine underneath it.