The day the objective changed

The instruction that started this was not "find a better policy". It was closer to the opposite: go back over what has already been tried and check that it makes sense. Prefer a second set of eyes on an old idea over a fresh direction.

That turned out to be the right call, because the first thing a careful re-measurement found was that the project had been optimising for the wrong thing.

1. morningFive independent audits: engine fidelity, the fair-D4 reference, the arithmetic of every historical claim, which past rejections were clean, and whether curriculum ordering was ever actually ablated.
2. middayScore decomposition on 64 fresh whole games.
3. afternoonChance-stratification result found, then replicated on a disjoint cohort.
4. eveningA 3.08x bit-exact engine speedup lands, unlocking depth-5 experiments overnight.
5. lateFour separate ideas fail their preregistered gates. A published claim of our own is withdrawn.

What the score is actually made of

Drop7's exciting moments are chain reactions — a disc lands, clears a row, everything above it drops, and that triggers the next clear. It is natural to assume that is where the points are. It is not.

positive

Score is survival

94.29% of points

Over 64 complete games under the strongest reference policy, 94.29% of all points came from the flat 17,000-point bonus awarded for surviving another row rise. Chain scoring contributed about 6%. Score and game length correlate at r = 0.9995.

The practical consequence: a strategy that stays alive one move longer is worth more than almost any chain it could set up. The 70,000-point board-clear bonus — the single most dramatic event in the game — is worth about 20.6 moves of ordinary survival, and no game in any cohort we have run has ever achieved one.

This gave a derivation for a rule of thumb the project had been carrying around without justification. Every five moves you place 5 discs and the board rises, pushing in 7 more: 12 discs arriving onto 49 cells. To survive indefinitely you must clear at least 2.400 discs per move and reveal at least 1.400 covered discs per move. Those are not tuning targets; they are conservation laws. Every policy we have measured sits just below them, which is exactly why every game eventually ends.

The one change that worked

positive

Exact chance enumeration unlocks the fourth ply

+101,171 points

The planner models the unknown next disc by sampling. Increasing the number of chance strata from 5 to 7 — making that model exact rather than approximate — was worth +101,171 points per game, with a one-sided 95% lower bound of +47,457. It replicated on a second, disjoint cohort.

The mechanism is the interesting part. With a biased chance estimator, adding a fourth ply of lookahead buys nothing at all: the deeper search just optimises against a distribution that is wrong. With an exact estimator, the fourth ply pays. Depth and chance-exactness are not independent knobs — the cheap one gates the expensive one.

neutral

A 3.08x bit-exact engine speedup

3.08x

The rewritten search produces byte-identical decisions to the frozen reference — verified across 936,612 real board states with zero differing bits, and by reproducing an entire 64-game cohort field-for-field.

Two things about it were instructive. The speedup is 3x, not the 5–10x the brief optimistically predicted. And the brief's ranked guess about where the time was going was simply wrong: the transposition table, expected to dominate, was worth 1.01x on its own, because it touches only 3.9% of the work. The leaf evaluation, ranked lower, was worth 2.61x.

Things that did not work

This is the part of a research log that is usually missing, and it is the part worth reading. None of these were sloppy ideas; each had a plausible mechanism and each was killed by its own preregistered gate.

dead end · recorded result

Making the planner more afraid of dying

The search assigns a large negative utility to a losing position. If survival is 94% of the score, making that penalty harsher should help. It does nothing: the parameter is already saturated. Every value past the current one produces identical play, because the penalty is already so much larger than any achievable score difference that its exact magnitude cannot change a ranking.

dead end · recorded result

Vetoing moves that lose a 25-move rollout

A previously reported technique, retested under corrected scoring. Each veto costs −21,887 points on average. The rollout is noisy enough that it overrides more good moves than bad ones.

dead end · recorded result

A learned survival head inside the search

A network trained to predict remaining lifetime, dropped in as the search's leaf evaluation. We preregistered the prediction that it would help. It did not, and recording the refuted prediction is the point of preregistering it.

dead end · recorded result

Distilling the planner into a policy network

The classic move: run an expensive search, train a cheap network to imitate its choices, deploy the cheap network. It underperformed the search it was copying.

dead end · recorded result

Our own scenario benchmark suite

We built a fixed suite of 128 positions to compare policies cheaply, then ran its own validation gate: does the suite rank policies the same way whole games do? At a 9-move horizon it ranked them backwards — Spearman −0.257. The suite is now labelled a diagnostic and may not be cited as evidence of strength.

A correction we had to publish against ourselves

Mid-afternoon we reported a clean split of the gap between a clairvoyant planner (one that can see the future) and a legal one, and attributed a specific share of it to planning rather than hidden information.

That number was measured on eight fixed disc "tapes", and those eight tapes turn out to favour long games — with a bias that grows the longer a policy survives, so it does not cancel between arms. Re-running every arm on 128 fresh tapes moved the reference policy's mean game length from 117.75 to 93.56 moves while leaving the clairvoyant arm untouched.

The central quantitative claim of that document is withdrawn. The finding that replaced it is smaller, less exciting, and correct.

Two hazards that had nothing to do with Drop7

Worth recording because they would silently corrupt any result computed on this machine:

negative

numpy float32 matmul is silently wrong on this host

A race in the installed OpenBLAS produces incorrect single-precision matrix products under threading — no error, no warning, just wrong numbers. Pinning the thread count before importing numpy is now mandatory in every script here.

negative

PyTorch Conv2d is nondeterministic on this host's CPU

Repeated identical forward passes disagree. Any experiment claiming byte-identical reproduction has to avoid that path or pin the backend.

Where this left the program

running

Push the fair planner to depth 5

owner: Claude Code

Chance-exactness unlocked the fourth ply. The obvious question is whether it also unlocks the fifth. The new engine makes this affordable for the first time; the cohorts were launched overnight.

running

Train a long-horizon evaluator on the GPU

owner: OpenCode / kimi-k3

A separate contributor is training an action-free afterstate model — one that scores a board position rather than a move — on labelled search values. The premise is that a network might see structure several row-rises out that a 4-ply search cannot.

blocked

Correct three engine divergences from real Drop7

owner: awaiting owner decision

The audit found three places where the simulator differs from the shipped game. Fixing them changes every number in the repository; not fixing them means the target is a slightly different game. This is the owner's call, not ours.