Score is survival

Most people meet Drop7 as a chain-reaction game: you set up a cascade, it goes off, numbers fly. That instinct is wrong here, and the repository has measured exactly how wrong.

In five-move Hardcore mode the board rises every five drops and each rise is worth 17,000 points. A cascade is worth a few hundred. So a game's score is very nearly a count of how many rises the player survived, and the whole strategy problem is a survival problem. This page shows that on one complete game, then explains what survival actually costs.

New here? How Drop7 works has every rule animated, and the glossary defines each term in one line.

One complete game, from the first drop to the last

Below is a real game played end to end by a deliberately trivial policy: drop the disc in whichever legal column scores the most points right now, breaking ties towards the shortest column. It is a toy — no look-ahead, no notion of danger — and it is here because its shape is the shape of every Drop7 game.

table view — every rise, and the last move

move	points	score	tallest column	discs on board
5 (rise)	17,007	17,014	3	17
10 (rise)	17,046	34,377	4	20
15 (rise)	17,155	51,638	5	23
20 (rise)	17,770	69,514	6	24
25 (rise)	17,155	86,736	6	28
30 (rise)	17,014	103,824	7	32
35 (last)	7	103,859	7	32

Two features of that chart are the entire lesson.

The score line is a staircase. It jumps at each rise and is nearly flat between them. Chains are the flat parts.

The height line is a ratchet. Every five moves the board gains seven covered discs from underneath and five dropped discs from above, and the policy is not clearing them fast enough. The line drifts upwards until a disc is sitting in the top row and the board has no room to rise — which is how this game ends, and how most games end.

Across those 64 games of the reference policy, 94.29% of all points came from the row-rise bonus, 5.71% from chain waves, and 0.00% from clearing the board — there were no board clears at all in 64 games (finding 01). The correlation between a game's final score and the number of moves it survived was r = 0.9995. Score and lifetime are, for practical purposes, the same variable.

The arithmetic that makes chains small

A cascade wave awards 7 points per disc at depth 1, 39 at depth 2, 109 at depth 3, 224 at depth 4 and 391 at depth 5. A five-deep, seven-disc wave — a spectacular play — is worth 2,737 points, which finding 01 puts at 0.8 of one row rise. Meanwhile a rise arrives every five moves, so the ceiling on points per move from rises alone is 17,000 ÷ 5 = 3,400. The measured rate for the reference policy was 3,423.2 points per move: essentially the ceiling, with the small excess being the chains.

There is one exception, and it is a large one: clearing the entire board is worth 70,000 points, about 20.6 moves of survival. It happened zero times in 64 games of the strongest known policy, so nothing in this repository says whether a clear-seeking strategy could change the economics. That is an open question, not a closed one.

What survival costs, exactly

Every five-move cycle puts twelve discs onto a 49-cell board: five dropped by the player, seven arriving as a new covered row. To live indefinitely you must remove twelve discs per cycle, which means clearing numbered discs at an average rate of 12 ÷ 5 = 2.4 per move, and turning covered discs back into numbers at 7 ÷ 5 = 1.4 per move.

The repository's status summary describes roughly 2.4 numbered clears and 1.4 reveals per move as the region associated with stable long games, and is careful about how strong that claim is: it labels them "diagnostic targets from limited runs, not proven universal thresholds" (research status). Finding 01 later argued that the two numbers are not empirical at all but exactly 12/5 and 7/5 — the disc-conservation requirement of the five-move cadence — and therefore necessary conditions for unbounded survival, though not sufficient ones: a policy can meet them on average and still lose by stacking one column into the top row.

The reference policy sustains 1.973 clears and 1.090 reveals per move, a deficit of 18% and 22%. Sorted by lifetime, its 64 games line up almost exactly with how close they came to break-even:

Group of games	Mean moves	Mean score	Clears per move	Reveals per move
shortest 16	44.7	138,762	1.471	0.732
second 16	70.0	232,515	1.812	0.975
third 16	99.5	341,310	1.989	1.099
longest 16	162.1	575,380	2.147	1.216
break-even	∞		2.400	1.400

Recorded in finding 01; development tier, 64 games on one exploratory seed lease.

What the million-point target really asks for

At roughly 3,400 points per move, a mean score of 1,000,000 needs a mean lifetime near 294 moves. The reference policy's measured mean was 94.06 moves. That is a 3.1× survival problem, and it is a much clearer statement of the research goal than "score more points".

For scale, finding 01 also ran three deliberately weak policies on the same 64 seeds: centre-first scored 57,233 over 21.64 moves, random legal play 80,778 over 28.44 moves, and always-lowest-column 100,050 over 33.28 moves, against the reference policy's 321,991 over 94.06 moves. Careful search buys about 3.2× the lifetime of random legal play. Reaching the target means buying another 3.1× on top of that.

Two consequences run through the rest of this site. First, a search that looks four moves ahead sees at most one rise boundary, and therefore cannot see a flow deficit accumulating over the eight-plus cycles that separate a 90-move game from a 294-move one — the argument for a longer-horizon learned evaluator, taken up in what makes a board good? and is more computation the answer?. Second, because score is dominated by a survival count with a very long tail, comparing two policies is a statistical exercise rather than a matter of looking at the best game — which is why one great game proves nothing.

The technical record

The reference figures all come from docs/exploratory/finding-01-score-is-survival.md: fair depth-4 expectimax, unmodified frozen source, 64 games on seeds 0xa51d0000–0xa51d003f under seed lease SEEDLEASE-A51D, 2,000-move cap, 0 censored games, run runs/RUN-A51D-d4/. Status: exploratory, development tier, reproduced in this checkout on 2026-08-20. Mean score 321,991.7, median 266,282, mean moves 94.06, points per move 3,423.2, 1.973 numbered clears and 1.090 covered reveals per move, 0 board clears.

This is a fresh-seed cohort, not the ledger reference. The number quoted elsewhere on this site for the reference policy — 308,295.578 points over 64 games at 90.031 moves — is the ledger-recorded broader reference cohort in docs/research/status.md. Finding 01 states that its own 64 games are consistent with, and an independent reproduction of, that figure; the two are different games and must not be averaged together.

Limitations recorded with the finding. The score standard deviation is 58% of the mean, so the mean is known to roughly ±46,000 at one standard error. Flow balance is necessary for indefinite survival, not sufficient. Board clears were absent, so the cohort says nothing about a clear-seeking policy. And audit-01 identifies two rise-boundary scoring discrepancies against the cited reference implementation, which act directly on the quantity measured here.

The toy game was generated by web/scripts/generate-sample-game.ts into web/content/learn/data/sample-game.json, using the TypeScript engine in latent mode with seed 0x5eed6001 and hidden values from 0x5eed6002 — the scripted-round playground domain, which overlaps no research seed lease. The policy is web/scripts/toy-policy.ts. It has no protocol, no preregistration and no cohort; its numbers are illustrations, not results.

What this taught us, and what is still open

• Score is lifetime; the correlation measured over 64 games was r = 0.9995.
• Survival has an exact price: 2.4 numbered clears and 1.4 covered reveals per move, the disc-conservation rate of the five-move cadence.
• The strongest known policy runs an 18–22% deficit against that price, and the deficit accumulates over cycles that a depth-4 search cannot see.
• The million-point goal is a demand for roughly three times the current mean lifetime.
• Untested: whether deliberately clearing the whole board — 70,000 points, worth 20.6 moves — is reachable often enough to change any of this.