Denoised public value

A board is only "worth" something on average — the same position can die in ten moves or run for two hundred, depending on which discs arrive. Earlier value models learned from a single played future, which meant every label was mostly noise. This one re-simulates each position 32 times and learns the average.

It is the best-predicting learned model in this repository, and the only one whose gameplay result came out positive. It is also, as the record below shows, promising development evidence rather than a qualified policy: two attempts to put it to work in a game both failed.

The intuition

Take one position and ask "how much longer does this game last?" Play it out once and you get a single number, and that number carries the whole variance of the game with it. Two identical boards can produce labels forty moves apart. A model fitted to those labels spends its capacity chasing coin flips.

Averaging costs 32 times the simulation, and it buys a label whose remaining uncertainty was measured rather than assumed: the ledger records a mean standard error of about 1.3 moves per label. What the model then has to learn is the signal, not the luck.

How it works, step by step

1. Collect positions. 64 complete games are played by the exact depth-3 phase-safety policy, and a position is kept every three moves. Positions are canonicalised (reflections folded together) and de-duplicated.
2. Strip the history. Each kept position is rebuilt from the visible board, the next disc and the rise clock alone — no score, level, move index, game seed or history. The model can only see what a player sees.
3. Label it by simulation. Each position is played forward 32 times, capped at 50 moves, by a fixed public depth-1 continuation policy. Only the luck differs between the 32 runs. The labels are the average capped lifetime and the probability of still being alive after 25 and after 50 moves.
4. Split by whole game. 48 of the 64 games become training data and 16 become the held-out set, split before de-duplication, so no position can appear on both sides. Training and held-out labels use separate random domains.
5. Fit a small network. A reflection-averaged NNUE with 35,395 parameters reads position and token embeddings, the next disc, the rise phase, and 16 standardised board metrics.
6. Only then, play. The exact phase-safety action stays the default. An alternative column is played only if it clears a paired lower-confidence bound on predicted lifetime, training-support checks, and an agreement check between the board and its mirror image.

Steps 1–5 are a prediction problem; step 6 is where every model in this family has historically come apart. Note that the model still only ever ranks siblings it was not trained on — see the sibling trap — which is why the deployment rule is so heavily hedged.

What happened

The prediction part worked, clearly. On positions from games it had never seen, the model's lifetime estimates lined up with reality far better than any earlier value model in the repository, and its survival probabilities were well calibrated. Every gate that had been declared before the run passed.

The play part is a small, hedged, positive signal. In an eight-game confirmation the corrected policy survived longer on average than the policy it was correcting — 79.5 to 88.125 mean moves — while switching about 2.6 times a game. But the confidence bounds on that gain were negative, meaning eight games cannot rule out that it was luck; the experiment was never repeated; and the baseline it improved was not the strong depth-4 reference.

Both attempts to make it useful in a stronger setting failed. A version that let the model veto a guided ensemble passed a four-game screen by a wide margin and then reversed on the eight-game confirmation — a textbook false positive. A version bolted onto fair depth-4 search, allowed to act only at the one moment the search cannot see through the coming row rise, made zero switches in its first paired game and then exceeded its runtime budget, so it was paused with nothing learned about its hypothesis either way.

The technical record — denoised stochastic value

Source: denoised-stochastic-value.cpp. Status completed, evidence ledger-recorded (experiment index; ledger, "Denoised stochastic public-state value").

Corpus: 1,071 training and 319 held-out disjoint states from 0x3d706800…; 17 cross-split duplicates removed. 44,480 label rollouts, 1,156,283 simulated moves, mean label standard error 1.304 training / 1.192 held out (90th percentiles 1.877 / 1.832 moves). All 64 roll-in games ended naturally, averaging 70.906 moves; the ledger records 234,260.75 points for that collection run.

Prediction gates (all declared in advance, all passed): held-out lifetime MAE 2.744 moves, RMSE 3.530, Spearman 0.951 (gate 0.70); survival-at-25/50 soft AUC 0.920 / 0.809 (gate 0.80); Brier 0.0136 / 0.0032; expected calibration error 0.0210 / 0.0061 (gate 0.10).

Gameplay: the four-game 0x3e820000… screen passed both criteria — exact behavior 78.75 moves, value policy 81.25 moves, 0.75 switches per game. The gated eight-game 0x3e830000… confirmation improved mean behavior from 79.5 to 88.125 moves with 2.625 switches per game; clear/reveal throughput rose from 1.901/1.033 to 1.996/1.112 per move. Paired 95% lower bounds were negative: −63,414 points and −17.404 moves. Runtime 121.3 s, 5.72 MiB peak RSS. No protected seed family was read.

Retained artifact: artifacts/models/denoised-value/v1.bin, 35,395 parameters, 141,780 bytes, payload checksum 1239007257.

Two recorded caveats about the scoring mode. The ledger calls the screen "that legacy-scoring run" and then states that its absolute point totals use the verified 17,000-point Hardcore bonus. The exploratory audit in audit-03 flags this as internally contradictory (finding L1) and notes that no point means were retained for the screen or the confirmation at all — only move means — so the only defensible gameplay statement is the survival improvement quoted above.

The technical record — denoised guided veto (rejected)

Source: denoised-guided-veto.cpp. Status rejected, evidence ledger-recorded. The frozen model may only replace a guided depth-5/K3 ensemble's action with a verified exact depth-3 fallback, and only when its predicted capped-50 lifetime advantage exceeds the model's own held-out MAE of 2.744151 moves. It can never nominate a third action.

Under historical 7,000-point Sequence scoring, the four-game 0x3e840000…003 screen looked strong: ensemble 126,146 points / 85.5 moves versus veto 184,649.25 / 121.25, deltas +58,503.25 and +35.75 — with 95% lower bounds already negative at −27,421.531 and −14.210. Seven of 485 moves were vetoes (1.443%), or 8.537% of the 82 disagreements.

The eight-game 0x3e850000…007 confirmation did not replicate: ensemble 164,266 / 110.625 versus veto 161,660.125 / 108.75, deltas −2,605.875 points and −1.875 moves, lower bounds −66,414.237 and −42.023. 12 of 870 moves vetoed (1.379%), 6.452% of 186 disagreements. Screen plus confirmation took 1,882.9 s. The ledger's own verdict: "The screen was a false positive." No threshold was swept and the screen seeds were not reused.

The technical record — phase-5 D4 value veto (runtime-paused)

Source: d4-phase5-value-veto.cpp. Status runtime-paused, evidence ledger-recorded. Fair D4 completes first and remains the default; the model is routed only when the rise clock is exactly 5 and the tallest column is at least 4 high — the situation D4 cannot see through, because the covered row due in five drops would then push the board to height five or more.

Only the first paired fitting seed 0x3de90000 was opened. Baseline and candidate were byte-for-byte equal at 162,102 points, 110 moves, 222 numbered clears and 120 cover reveals. The candidate routed 21 of 22 phase-5 decisions and evaluated 725 modeled successors, and made zero switches: 124 alternatives failed the lifetime gate, 97 the survival gate, 13 the root-Q band, none the orientation bound (counts overlap). The paired pilot took 198.902 s against a fixed projection of 3,580.240 s for the full design, above the 2,700-second limit, so the experiment paused before its heldout, screen and confirmation seeds were opened.

This is pilot-only evidence — not a rejection and not a qualification. The blind-spot audit records it as the most confounded entry in the corpus: the best-predicting model in the repository was given a rule so tight that zero information about the hypothesis was obtained, and one of its bands is still written in terms of the superseded 7,000-point level bonus.

What this taught us, and what is still open

The transferable lesson is about labels, not architecture: this model differed from its predecessors mainly in how its training targets were computed, and that alone moved held-out lifetime correlation to 0.951 — against 0.510 for the structured NNUE and 0.557 for the scaled survival value, as tabulated in the blind-spot audit. Denoising the target is cheap to describe and expensive to run, and it worked.

What remains unproven is whether that accuracy can be converted into better play. Both conversion attempts were structurally unlucky: one was measured against a weaker comparator on four games and reversed, the other could not fire at all. The retained checkpoint is still on disk, and giving it a deployment rule with enough room to make switches — on top of fair D4, on a cohort large enough to matter — is a live open item in the repository's own audit rather than a closed question.

Sources

• denoised-stochastic-value.cpp — collection, labelling, fitting, gates.
• denoised-guided-veto.cpp — the veto of the guided ensemble.
• d4-phase5-value-veto.cpp — the phase-5 veto over fair D4.
• Checkpoint: artifacts/models/denoised-value/v1.bin.