NNUE evaluatorsfeatured
A learned survival estimate inside the reference search
Blends a learned survival estimate into the reference board evaluator and tests it under sampled and exact chance handling.
completedreproducedLifetime objective
Judge a move by how much longer the game will still last, rather than by how many points it scores right now.
Approaches in this family
11 approaches, featured pages first. The label above each title is the technique the approach uses; the same pages appear under that technique on the approaches index.
Instruments in this directory
These play no game of their own: engines run the games and diagnostics measure them. They are listed here because they live alongside the approaches above; the full sets are under Engines and Diagnostics.
- The shared game harnessengine
One instrumented game loop and one results writer, shared by every experiment in this family so their numbers are directly comparable and every game's score is checked against the rules.
- Entombed discs: the 3 that can never cleardiagnostic
Measures how often low numbered discs become impossible to clear, when death follows, and whether the evaluator notices.
- Fast engineengine
A semantics-preserving reimplementation of the Drop7 move engine and the fair-D4 leaf, proven bit-identical to the frozen reference and measured at about 3× end to end.
- Remembering the last boardengine
Reuses board features across equivalent chance samples and recomputes only the next-disc term, preserving exact values with less work.
- The fast engine learns reveal samplingengine
Ports the native factored chance node (N disc strata x M reveal samples) into the fast memo engine, trace-equivalent to the C0 search at a fraction of the cost.
- Is the required clearing rate reachable at all?diagnostic
Lets a privileged planner test whether any policy can remove discs as fast as the game adds them.
- Getting the workstation's GPU to do the trainingengine
An infrastructure package that made PyTorch work on this machine's integrated AMD GPU, checked that it computes correct answers, and measured how much faster it is than the CPU.
- Is the small survival network as accurate as it can be?diagnostic
Retrains the small board evaluator at several sizes and settings to test whether more capacity predicts lifetime better.
- Can a leaf-sized network hold the four-ply search's ordering?diagnostic
Trains the smallest network that could run inside a search leaf on the exact values the four-ply search gives every sibling move, and asks whether it ranks those moves the way the search does.
- Scenario engine and exact solverengine
Fix every future disc and hidden number in advance so a position has one exactly computable best line.
- Where the points actually come fromdiagnostic
Attributes every point to its source and shows that Hardcore score mostly measures survival time.
- The data factory, and the switch that was left offengine
Replays many games and labels every visited position with its remaining lifetime, producing training data for the rest of the family.
- Testing the benchmark instead of the policydiagnostic
Ranks nine policies on 128 saved positions, compares that ranking with complete games, and retires the benchmark when they disagree.
Other directories
These directories have no page of their own yet. Each link lists the source files inside.
- p-solno page yet
On this page
Judge a move by how much longer the game will still last, rather than by how many points it scores right now.
That sounds like a slogan until you count the points. When this repository
instrumented 64 games of its strongest policy and attributed every single point
to its source, 94.3% of the score was the flat 17,000-point bonus for
surviving another row rise, 5.7% was chain reactions, and 0.0% was the
70,000-point bonus for emptying the board, which never happened once. Final
score and moves survived correlated at r = 0.9995
(finding-01, a development-tier
exploratory finding; the same run is written up on the
score decomposition page).
So a Drop7 Hardcore score is survival time wearing a different unit, and the million-point target is really a survival target. Points arrive at a measured 3,423 per move, and the same finding puts the lifetime a one-million-point average needs at about 294 moves. The reference policy's games last 94.
Disc-flow arithmetic
Here is why the game ends. Both boards below are the repository's own rules engine, playing a real game; nothing is drawn by hand.
Every five moves, five placed discs plus seven risen gray discs — twelve discs — enter a board that only has 49 cells. Nothing else adds material, and the only way material leaves is a numbered disc clearing. The arithmetic is therefore exact and unforgiving: a policy survives indefinitely only if it sustains
- 2.400 numbered clears per move (12 ÷ 5), and
- 1.400 gray discs revealed per move (7 ÷ 5).
The reference policy sustains 1.973 and 1.090: a structural deficit of
about 18% and 22%. Its board fills up, and then it dies. That deficit, not a
shortage of spectacular chain reactions, is what ends its games
(finding-01).
What every approach in this family does
- Take a policy: the reference depth-4 search, a hand-written rule, a planner, or a candidate with something learned bolted into it.
- Play whole games with it, paired against a comparator on the same random seeds, so the two policies face identical luck.
- Report three numbers besides the score: numbered clears per move, gray
discs revealed per move, and the occupancy slope: how many of the 49
cells the board gains per five-move cycle. The slope is just the
conservation law integrated: a board drifts by
12 − 5 × clears-per-movecells per cycle, so a positive slope is a policy that is losing, however good its tactics look on any single move. - Ask whether the mechanism moved, not just the mean. A score gain with no movement in the flow rates is treated here as unexplained rather than real.
That last rule is what makes this family a family. Score in this game is heavy tailed: one lucky game can swamp 63 ordinary ones, while a flow rate is averaged over every move of every game and is far better determined. Several of the results below were decided on the flow rates when the score deltas could not be separated from zero.
What happened
Three things, and none of them is the one the family was hoping for.
The target is physically reachable. A planner that is allowed to cheat — it reads the hidden numbers under the gray discs and plans exactly against a fixed future — holds the required clear rate for 6,000 consecutive moves and never dies. That settles a real question: Drop7 Hardcore is a control problem with a solution, not an arithmetic impossibility. It is also, emphatically, not a policy; see the flow ceiling page, where the cheating is labelled on every number.
No legal policy tried here got near it. A planner with the same objective and the same look-ahead but no access to the hidden numbers closes only a small part of the gap, and the part it cannot close grows exactly as the board fills up and there is more hidden information to be ignorant of. Every one of those games ended.
Small, bounded corrections to the reference search worked better than replacing it. The two positive results in this family are both of that shape: a better handling of the game's randomness (reveal sampling) and a learned survival estimate mixed into the reference search's board-scoring function at half weight (learned leaf). Both are worth tens of thousands of points on 64 paired games. Neither is worth hundreds of thousands, and the target needs hundreds of thousands.
Every approach in the family
The measurements that started it
- Score decomposition — counts where each point came from, and finds that a score is a stopwatch.
- Flow ceiling — asks whether any line of play can sustain the required clear rate, using a clairvoyant planner that reads the hidden board. Answer: yes, but only that planner did.
- Suite validation — tests whether a cheap 128-position benchmark predicts whole-game strength, and finds that at the horizon it was minted with, it does not.
Bounded changes to the reference search
- Reveal sampling — the search had one dial governing two different coin flips; separating them is worth about as much as a whole extra move of look-ahead.
- Risk calibration — turns the reference search's baked-in constants into dials that can be swept.
- Leaf reweight — makes the nineteen weights in the hand-written board-scoring function runtime data, so refitted weight vectors can be tested.
- Rollout veto — looks twenty-five moves ahead with a weak player and overrules the search only on a clear disagreement.
Learning something
- Sibling corpus — the training-data generator: every public position a game visited, labelled with how many moves that game still had left.
- Afterstate net — the survival network trained on that corpus.
- Learned leaf — blends a learned survival estimate into the reference search's board scorer, and measures what a network at that position is actually allowed to cost.
- Planner distillation — tries to compress the expensive legal planner into a small network, with every legal move at every position labelled.
Machinery the rest of the family runs on
- Scenario engine: a rules simulator in which every hidden number and every future disc is fixed in advance, which is what makes a clairvoyant optimum well posed at all.
- Fast engine: the same rules about three times faster, proven to make byte-identical decisions.
- Common: the shared harness that runs a cohort, enforces seed leases and writes the result artifacts.
- GPU: the environment and benchmarks for the machine's integrated GPU, including two host defects that had to be worked around.
What this taught us, and what is still open
- The objective is now a number, not a slogan. "Play better" became "hold 2.400 clears and 1.400 reveals per move, at around 20 occupied cells". That is a far more testable target than a score.
- Flow rates are the cheap early statistic. They move before the score does and they are averaged over thousands of moves instead of dozens of games.
- Chain building is real and is not the lever. A cheating planner told to
maximise points instead of survival earned 1.75 times the points per move —
and died in less than half the time, ending with a lower mean score
(
finding-06). - Still open: nothing legal has been shown to hold flow balance. The measurements say the missing ingredient is reasoning about what is under the gray discs, rather than averaging over guesses of it, and no candidate in this family does that yet.
For the wider picture see the research status, and for the ideas these pages assume, the concepts primer and the glossary.