Long outcomes

Measure what each of the seven columns is worth by playing it forward twenty-five moves under shared imagined futures — then try to learn that measurement, so a policy can have it without paying for it.

Six programs are collected here. Between them they test a general learned ranker, a readable hand-built feature and a broadened version of it, a multi-output network at two sizes, the same idea on curriculum positions, and a narrowly targeted veto. All of them were rejected, and together they make an unusually well-controlled argument about why.

The data is the point

Nearly every learning failure elsewhere in this repository can be blamed on the training data: only the move that was actually played has a recorded outcome, so the model is asked at deployment to rank six alternatives it never saw. See ranking siblings.

These corpora are built to remove that excuse. At each recorded position, every legal column is forced in turn, and each is played forward for 25 moves under the same seven imagined futures, with a fresh complete depth-2 search choosing every subsequent move. The return is the score the line actually earned, minus a large penalty if it died, plus one leaf evaluation if it was still alive at the horizon. Positions are drawn evenly through whole games, and splits are always by whole game.

So the question these experiments ask is narrow and fair: given a perfect record of what each column led to, can a model learn to order them?

What happened, in plain English

Adding a learned correction to exact depth-2 search made the ordering worse. On the fitting positions the correction looked strong — top-move accuracy rose from 26.0% to 34.7%. On the held-out games it fell to 22.2%, below the plain search's 27.8%, and the leftover-value measure got worse too. This is ordinary overfitting, cleanly caught by splitting on whole games; it failed eight of ten component checks.

The readable feature was real and useless. A carefully constructed "vertical-ladder" measure — how much release potential is stored in a column as discs are added — correlates with outcomes at roughly 0.01. Fitting it gave it a negative coefficient in every fold. Stored ladders genuinely exist on these boards; they are not what separates a good column from a bad one twenty-five moves later.

Some things are very learnable, and none of them is the ranking. The multi-output networks predict whether a line survives with a correlation around 0.85, and numbered clears at around 0.43. The quantity that actually decides the move — the return residual — sits at 0.35, downside at 0.33 and variance at 0.14. A shared representation that is good at survival did not become good at choosing.

More data and more capacity made it worse, not better. This is the cleanest single-factor test in the repository. Holding everything else identical, the corpus was enlarged 5.24 times and the network's hidden layer was widened from 12 units to 48. Top-move accuracy went 27.79% (plain exact search) to 25.99% (12 units) to 23.94% (48 units), while the survival output stayed at 0.85 either way. Zero of four folds avoided regressing. The "we just need more data" and "we just need a bigger model" hypotheses were both tested here and both failed.

Restricting the model to a veto did not save it. The final attempt used only the trustworthy survival and clear outputs, and only to override depth-4 search when an alternative was clearly safer. Of 288 positions, only seven had any eligible alternative at all. It made 19 overrides and zero of them were correct. Its mean return actually rose — by 14,165 — which the ledger flags as the exact reason mean-only selection is unsafe here.

Broadening the readable feature did not help either. A relaxed version, which lets a column's potential be realised by one or two hypothetical future drops rather than a literal existing ladder, correlated better than the literal version but still failed every material gate, non-regressing in only one of six folds.

The technical record

Corpus. 288 fitting roots from 24 games and 144 held-out roots from 12 games, 12 evenly spaced roots per game, drawn from the already-consumed 0x3df2 / 0x3df3 depth-4 corpus; every sibling, seven common-random-number tapes, horizon 25, fresh complete full-width depth-2 search at every continuation step; reveals and later visible discs in independent event-indexed domains. The scaled version uses 1,508 fitting roots from 16 whole games (5.236 times the sample) plus 465 roots from 8 already-burned games.

Ranker over exact depth 2 (rejected — ledger-recorded):

Split / ranker	Top-1	Top-2	Pairwise	Normalized regret
Fitting, exact depth 2	26.04%	43.75%	56.90%	0.33442
Fitting, depth 2 + residual	34.72%	53.82%	60.67%	0.26712
Held out, exact depth 2	27.78%	45.83%	58.48%	0.33050
Held out, depth 2 + residual	22.22%	45.83%	57.67%	0.36350

Failed eight of ten component checks; the first held-out half regressed most clearly (top-1 30.56% to 19.44%). No gameplay seed was opened.

Feature and multi-head audit (rejected — ledger-recorded, development-only). Ladder alone reached 15.63% fitting top-1; depth 2 + fitted ladder was indistinguishable from depth 2; the fitting-only coefficient was −0.02315 and all six fold coefficients were negative; ladder pair-difference correlation 0.0133 fitting and 0.0053 held out. The 12-unit five-head network scored 24.31% top-1 / 54.84% pairwise / 0.35884 regret in nested fitting cross-validation against exact depth 2's 26.04% / 56.90% / 0.33442, and 22.22% / 56.26% / 0.41081 on the burned held-out split against 27.78% / 58.48% / 0.33050. Auxiliary Pearson correlations on that split: survival 0.854, numbered clears 0.427, return residual 0.350, downside 0.325, variance 0.139. A replay verified every persisted scenario return with maximum error exactly zero.

Scaled network (rejected — ledger-recorded). Only varied factor: 12 against 48 hidden units, identical epochs, losses, optimizer and folds.

Split / ranker	Top-1	Pairwise	Regret	Survival r
Cross-validation, exact depth 2	27.79%	57.44%	0.34828	0.0209
Cross-validation, 12 hidden	25.99%	55.70%	0.36251	0.8547
Cross-validation, 48 hidden	23.94%	55.15%	0.37018	0.8494
Burned held out, exact depth 2	30.75%	59.61%	0.33356	0.0403
Burned held out, 12 hidden	26.02%	57.80%	0.35371	0.8185
Burned held out, 48 hidden	24.73%	56.88%	0.38022	0.8291

Zero of four folds fully non-regressing. This experiment predates the scoring correction: its engine and all return labels use the historical 7,000-point level bonus, and the source now asserts that dependency so it cannot silently rerun against the corrected engine. It is architecture evidence, not score-calibrated Hardcore evidence.

Veto classifier (rejected — ledger-recorded). Nine public features, balanced logistic model, switching only above probability 0.90 and only after survival and clear prefilters, nested by whole game. Frozen gate: at least 80% precision, 20% coverage, 10,000 mean gain, paired-scenario 10th percentile at least −7,000, at least 85% fallback, 90% survival and clear retention, 12 switches, four active and four stable folds. Observed on 288 roots: only 7 roots had an eligible alternative; 19 switches, zero true positives; precision and coverage zero; paired-scenario 10th percentile −39,674.24; survival retention 84.21%; clear retention 57.89%; zero of six folds stable. Mean return rose 14,165.41. On the burned held-out split: 3 switches against 6 eligible roots, again zero true positives, 10th percentile −39,470.31, clear retention 33.33%.

Curriculum network (rejected — task-record only; no ledger entry). Artifact-only, 3,072 training and 1,024 held-out states from the frozen 4,096-state public oracle curriculum, every action, seven scenarios, horizon 100, five heads, 36 epochs, no gameplay runner and no gameplay seed accepted. Retained outcome: survival prediction was useful, action ordering regressed versus exact search. No numbers are retained.

Relaxed chain potential (rejected — ledger-recorded). Eleven fixed summaries over one or two hypothetical drops from each existing first-step outcome, with revealed covers becoming inert so no future value is ever queried; ridge residual over exact depth 2, six whole-game folds. Raw relaxed pair correlation 0.0505 fitting / 0.0517 held out against exact depth 2's own 0.2164 / 0.2614; the combined model scored 25.00% / 25.69% top-1 against depth 2's 26.04% / 27.78%; only one of six folds non-regressed; every material gate failed.

Records: experiment index "D4 and long-outcome research"; ledger sections "Public-D2 25-move outcome ranker", "Long-outcome ladder and multi-head audit", "Long-outcome D4 veto classifier", "Relaxed chain-potential feature audit" and "Scaled long-outcome NNUE sample-bottleneck audit". Sources: d2-long-outcome-ranker.cpp, d2-long-outcome-feature-audit.cpp, curriculum-long-outcome-nnue.cpp, scaled-long-outcome-nnue.cpp, d4-long-outcome-veto-classifier.cpp, relaxed-chain-potential-audit.cpp.

What this taught us, and what is still open

• Scale was tested, and it is not the answer. More labels and more capacity made the ranking monotonically worse while the survival prediction stayed flat. That is the signature of a limit in the label, not in the model or the optimiser.
• Predicting the game is easier than choosing a move. A correlation of 0.85 on survival next to a rank ordering below cheap exact search is the whole lesson of this family in one line.
• Means are not gates. The veto classifier increased mean return while making nineteen consecutive wrong decisions. Every design here that survived scrutiny reports precision, coverage, downside quantiles and per-fold stability rather than an average.
• Still open: the audit that follows from all of it — targeting per-rise survival rather than return, as audit-05 §5.3 argues, since the survival heads are the only strong signal anyone here has fitted. Also open: whether seven imagined futures are simply too few for the return labels to separate close siblings at all, which finding-03 estimates at 16 to 51 for the closely related veto.