Structured NNUE

A small network whose input is not a summary of the board but the board itself: a separate learned vector for every cell-and-token combination, fed into a predictor of remaining lifetime and survival.

Both experiments are in the written ledger with frozen gates. Between them they produced the cleanest demonstration in this repository of the difference between predicting a board and choosing between boards — and the second one is the reason the phrase sibling trap exists here.

The intuition

Earlier value models compressed the board into hand-chosen summaries — how tall the columns are, how many covered discs there are, how many clears are one disc away. Every such summary is a guess about what matters. An NNUE-shaped model makes fewer guesses: it gives each of the 49 positions a private embedding for each of the ten things that can sit there, and lets training decide what a 3 in the bottom-left corner means as distinct from a 3 near the top.

The model is also made exactly symmetric on purpose. Drop7 has no left or right preference, so the network is shown both mirror images of every training board and, at inference, its two orientations are averaged — a board and its mirror receive exactly the same value, rather than approximately the same one.

How it works

1. Collect. Complete games are played by the exact depth-3, five-stratum phase policy on one seed lane. Whole games — indices divisible by five — are held out before any fitting.
2. Encode. Every cell/token pair, the visible next disc and the rise phase each activate their own embedding; 20 standardized phase metrics describing chain readiness, quiet build inventory, low-number clogs, cover exposure and height risk are appended. Nothing else is visible to the model.
3. Fit. A fixed 128/64 leaky-ReLU trunk with three heads: normalized remaining lifetime, survival at 25 moves, survival at 50. Lifetime and survival calibrations are fitted on training data only.
4. Gate before playing. Declared in advance: both survival areas-under-curve must reach 0.80 and held-out lifetime rank correlation must reach 0.65, or the model never influences a move.
5. The variant that changed the data. counterfactual-successor-nnue.cpp keeps the architecture and changes what is labelled: at each roll-in position it enumerates every legal column through three common public chance strata, deduplicates the resulting successors, and gives each of them eight independent 75-move public continuation labels. Splits are by source game with no canonical board shared between them. The goal was to make the model good at exactly the comparison a policy has to make.

What happened

The first model learned its training data very well and generalized much worse. Its ability to say "this game is nearly over" survived the move to held-out games; its ability to rank boards by how long they had left did not, and both of the numbers it had promised in advance came in under target. It stopped at prediction and never played.

The second is the interesting failure. Labelling every sibling fixed the coverage problem, and the model became genuinely accurate about individual successor boards — a rank correlation of 0.839 and an average error under four moves. But asked which column was best at a position — best according to the very labels it had been trained on — it named the right one 15.4% of the time; used as the leaf of a depth-3 search, 30.8%. Put on the board, it lost heavily. Strong global accuracy and poor within-root discrimination turned out to be entirely compatible.

The technical record

Both entries are ledger-recorded in the experiment history.

Structured multi-head NNUE value experiment (structured-value-nnue.cpp). 75,395 parameters, 301,764 bytes with normalization and calibration metadata. Corpus: 160 uncensored depth-3/five-stratum trajectories, 128 fitting games and 32 whole-game holdouts, 9,800 fitting labels and 2,132 held-out labels; behavior mean 247,202.869 points / 74.575 moves. Training: MAE 20.391 moves, death-within-25/50 AUC 0.999 / 0.978, lifetime Spearman 0.887. Held out: MAE 29.524 moves, AUC 0.855 / 0.614, Spearman 0.510, Brier 0.192 / 0.314, expected calibration error 0.186 / 0.286; mean reflection-orientation gap 3.384 moves, so symmetry was not the failure. Against the frozen gates (0.80 AUC on both horizons, 0.65 Spearman) the 50-move AUC and the ranking both failed; the reserved screen and confirmation seeds were never read and the experiment was not retuned.

Counterfactual-successor NNUE (counterfactual-successor-nnue.cpp). On 242 held-out successors: Spearman 0.839, MAE 3.888 moves. Within-root signal: direct top-action accuracy 15.4%; as a full-width depth-3 leaf 30.8%, pairwise accuracy 62.0%, label regret 1.90 moves. The screen is explicitly recorded as historical 7,000-point Sequence-scored: exact depth 3 averaged 166,112.25 points / 111.5 moves, the NNUE-leaf search 102,916.25 / 72.5, a paired loss of 63,196 points and 39 moves. Confirmation seeds were not read. The source carries static_assert(kLevelBonus == 7'000), an intentional lock that stops it being rerun under the corrected 17,000-point Hardcore engine.

The ledger's own conclusion is quoted often in this repository: "strong global state-value correlation can coexist with poor discrimination among the sibling states that determine an action. Future learned evaluators must train and gate on grouped, within-position ranking evidence."

The first experiment's ledger entry precedes the corrected-Hardcore scoring replay, so read its point totals as within-run rather than comparable with the corrected-score fair D4 reference.

What this taught us, and what is still open

This pair is where the repository stopped trusting prediction metrics as a proxy for playing strength. Every learned-evaluator gate written afterwards asks for within-position ranking — top-1 agreement, pairwise agreement, regret against a reference ordering at the same root — instead of, or in addition to, global correlation.

What it did not rule out: the architecture. Position-specific embeddings with exact reflection symmetry were never shown to be the problem; the labels and the objective were. The distributional afterstate ranker kept the "value a resolved position" idea and changed both — a within-root ranking loss instead of pure regression, and aligned chance scenarios shared across siblings instead of independent continuations per successor.