How is the research kept honest?

Experiments are preregistered, evaluated on never-read game blocks, and reported with paired differences, bootstrap 95% lower bounds and explicit verdicts — including inconclusive results and clear labels for screen-tier vs. qualification evidence.

How much does it cost?

Nothing — the game is free to play in the browser and the project is open source.

Drop7 Research — An open-source quest for a million-point Drop7 strategy

Drop7 Research is an open-source project that combines a browser-playable version of Drop7 — widely described as one of the greatest puzzle games — with a rigorous, preregistered research program trying to find an autonomous strategy that plays the game as well as a human (a million-point target). The site documents 116 approaches across 12 strategy families, 27 registered claims, 41 preregistered experiments and 40 recorded outcomes, with results and datasets published for independent reproduction.

Key Features

Play Drop7 in the browser

Watch a game or drop discs yourself by hovering the board and clicking a column; the site also includes the game rules.

A documented, reproducible research program

116 approaches across 12 families, 27 registered claims, 41 preregistered experiments and 40 recorded outcomes, with per-run data published in a public research archive (run IDs, SHA-256 verified tables, per-game result rows).

Fourteen explained technique families

Each method is explained from the ground up and linked to the strategies that tried it: expectimax search, heuristic evaluation, Q-learning and value learning, n-tuple networks, NNUE evaluators, policy gradients, evolutionary optimisation, Monte Carlo tree search, rollouts and policy iteration, orac

Transparent evidence reporting

Results are reported with paired differences, bootstrap 95% lower bounds, win/loss counts, cohort sizes and preregistered verdicts — including explicitly inconclusive results and honest distinctions between screen-tier evidence and qualification.

Current best candidate

Learned lookup tables (n-tuple tables) used as the leaf of a depth-4 seven-stratum fair expectimax search, averaging 516,155 points over 512 never-read games — the strongest published candidate so far, still well under the million-point target.

Full-text site search

A search (⌘K) covering every page, record and vocabulary term on the site.

How It Works

Learn the game, then the ideas

New visitors are guided to learn the game first, then the ideas behind the strategies, then to review what has been tried and what remains open.

Preregistered experiments on a fixed engine

Candidates are trained (e.g. temporal-difference play on a Rust engine for billions of moves) and evaluated on preregistered, never-read game blocks under corrected 17,000-point Hardcore scoring, with paired comparisons against a frozen hand-written 'fair leaf' reference.

Published, reproducible records

Frozen tables, per-game rows and analyses are published under run IDs in a public archive so another machine can reproduce the screens or continue training from them.

Pros & Cons

Pros

Completely open and reproducible: run IDs, dataset records, digests and per-game rows are published
Unusually rigorous for a hobby-scale AI project: preregistered comparators, bootstrap lower bounds, and clearly labeled inconclusive verdicts
Educational: 14 technique families explained from the ground up before being applied to the game
Includes a genuinely playable browser version of the game, not just papers
Honest reporting of what the evidence does and does not show (e.g. screen-tier vs. qualification, validation margins not oversold)

Cons

Who It's For

Best for

AI/RL enthusiasts who want to see search, TD learning, n-tuple networks and expectimax applied rigorously to one game
People who enjoy Drop7 and want to play it in a browser and understand high-level strategy
Anyone interested in reproducible, preregistered experimental methodology with published datasets

Not ideal for

Teams looking for a production AI product, API or SDK
Readers wanting light, casual content — the site is a detailed research record
People seeking a solved or near-solved million-point strategy; the problem is explicitly still open

Use Cases

Playing Drop7 for free directly in a browser

Studying how search, value learning, n-tuple networks, NNUE, MCTS and other techniques perform on a stochastic puzzle game

Learning experimental methodology: preregistration, paired evaluation, bootstrap bounds and honest negative results

Reproducing published experiments or continuing training from the released frozen tables and datasets

FAQ

What is Drop7 Research?

An open-source project that lets you play Drop7 in the browser and documents the search for an autonomous strategy that plays the game as well as a strong human — defined as scoring a million points.

Can I play the game on the site?

Yes. You can watch a game or hover the board and click a column to drop the next disc; the rules are also available on the site.

How good is the current best AI player?

The strongest candidate so far — learned lookup tables inside a depth-4 search — averaged 516,155 points over 512 games it had never seen, versus 382,567 for the same search with a hand-written leaf. Both are well under the 1,000,000+ target, so the problem remains open.

Is the project open source?

Drop7 Research