Drop7 Research

September 2026

On 2026-09-05 a learned evaluator beat the frozen fair leaf inside the same search for the first time, at screen tier. Lookup tables over every full row and full column of the board (ten million patterns per line), plus small windows, all conditioned on the rise clock, were trained by temporal-difference play on the Rust engine for four billion moves and placed as the leaf of the depth-3 seven-stratum fair search. On 256 never-read paired public-development games the tables scored 484,577 points and 140 moves against 314,438 and 93 for the identical search with the fair leaf: paired +170,139, bootstrap 95% lower bound +130,499, both halves positive, 167 wins to 89 (RS-20260905T215332Z-95d18a5a, approach page). On the same games the tables at depth 3 also beat the fair leaf at depth 4, the standing reference (377,803), by 106,775 with a lower bound of 62,574; that arm was diagnostic, not the preregistered comparator.

What this does and does not change. The evidence tier is SCREEN on public development seeds, opened once; it is not a qualification, no protected or final seed was opened, and the mean is under half the million-point target. The candidate was chosen at the best of twenty training-role validation points, so its screen margin is the honest number and its validation margin (+270,023) is not.

On 2026-09-06 the result replicated on a second, disjoint block. The same frozen tables (SHA-256 verified) played 512 never-read public-development games as the depth-3 leaf and scored 487,066 against 326,717 for the fair leaf in the same search: paired +160,349, bootstrap 95% lower bound +129,753, both halves positive, 330 wins to 182 (RS-20260906T040113Z-6ba93171). The same experiment trained a six-times-wider evaluator (2x4 and 4x2 window families added, 5.8 x 10^9 entries) on fresh training seeds until a 256-game validation margin stopped rising (a window-of-four plateau rule, which fired at 4.5 x 10^9 moves); that candidate passed the same gate on the same 512 games (481,869, paired +155,153, lower bound +126,819) and came out 5,196 behind the first tables with bounds from -40,535 to +29,158, so the preregistered scale verdict is inconclusive: no gain larger than about 35,000 in either direction. Played directly with no search the wide tables averaged 328,039 against 294,323 for the first tables on the same games, level with the fair leaf's depth-3 search; the extra scale improved the one-ply policy and left the leaf where it was. Both candidates beat the fair leaf at depth 4 on these games by more than 84,000.

Later on 2026-09-06 the first tables played one ply deeper. Unchanged and hash-verified, as the leaf of the reference depth-4 seven-stratum search they scored 516,155 against 465,675 for the same tables as the depth-3 leaf on 512 never-read public-development games from a fourth block: paired +50,479, bootstrap 95% lower bound +14,707, both halves positive, 284 wins to 228, a higher lower quartile (RS-20260906T171746Z-1623f833). On the same games the fair leaf's fourth ply was worth +49,599, and the per-game difference between the two depth steps is +881 with bounds from -40,932 to +42,867 (floor 42,550), so the preregistered interaction verdict is inconclusive: the learned leaf pays for depth about as well as the hand-written one, and no better or worse than a 512-game screen can see. At depth 4 the tables beat the fair leaf in the same search by 133,588 (lower bound 102,051); the depth-3 margin on this block was 132,707. The depth-4 search does about 36 times the depth-3 search's logical work per game. The first tables at depth 4 are now the candidate to carry forward; the next steps the record supports are a STANDARD-tier evaluation of that unchanged candidate on fresh development seeds, and a study of a leaf tuned for the search's use rather than for one-ply play (training on the boards the search evaluates, or removing the optimistic starting value from never-updated entries).

Late on 2026-09-06 a fourth experiment asked whether the leaf should change its mind as the board fills. A seed-free pass over the first run's saved accumulators found the frozen tables least trained on the fullest boards (0.98% of legal seven-high column patterns and 1.06% of five-or-more-disc top-row patterns ever updated, the touched full-line entries still at the optimistic starting value), so two repairs were preregistered against the frozen tables on a fifth never-read 512-game block (EX-20260906-ntuple-fill-conditioned-continuation-a9e5cbd3): five copies of every table keyed to a global fill bucket (occupied cells or tallest column), warm-started by copying the frozen tables into every bucket and trained on for up to two billion moves beside an unconditioned continuation as control; and two no-training edits that replace the 901,259,321 never-updated entries' starting value with zero or with the class mean. The occupied-cells candidate as the depth-3 leaf scored 506,494 against 485,455 for the frozen tables: paired +21,039 with bounds from -16,864 to +59,230 and a floor of 37,956, so the preregistered gate fails as a non-measurement; against the control continuation it is +15,093 and the control against the frozen tables +5,946, both inconclusive, and neither fill arm beat the control's best validation margin (RS-20260906T234914Z-a3fae1a9). The zeroed edit, with nothing trained, scored 514,897 at depth 3: paired +29,442 over the frozen tables with a lower bound of +5,813, changing only 194 of the 512 games, the one reading whose fixed verdict is supported. One ply deeper both repairs gave their gains back: the fill candidate at depth 4 was 22,631 behind the frozen tables (bounds -57,779 to +12,806) and the zeroed edit 10,592 behind (bounds -32,045 to +11,133), both inconclusive and both below their own depth-3 scores, while the frozen tables' fourth ply on these games was worth 36,500. The theory is assessed mixed, the frozen tables at depth 4 remain the candidate to carry forward, and the record now points at a leaf trained on the boards the search evaluates, and at a depth-aware price for never-updated entries, rather than at more buckets.

The frozen tables of both training runs, the three screens' per-game rows and their analyses are published in the public research archive under their run ids (https://data.drop7.dev/runs/<run-id>/ntuple-scale/..., digests in the references; dataset records DS-20260906-ntuple-scale-frozen-tables-ff977178 and DS-20260906-ntuple-scale-wide-frozen-tables-c5eb52a6), so another machine can reproduce the screens or continue training from them.

August 2026

Through August the best dependable policy was fair depth-4 expectimax — a four-move look-ahead that treats the game's luck honestly. Even with its best chance model, its measured means sit under half of that target. This page shows where the gap is, what has been measured, and which directions the evidence has closed.

How score responds to search depth, under each of the two chance models:

Mean score against search depth, fair expectimax, by chance-stratum count

• 5 strata (biased next-disc estimate)
• 7 strata (exact next-disc estimate)
• 7 strata, first 32 seeds only (arm stopped)

Notes and sources

All arms play the same 64-seed development cohort 0xa51d1000-0xa51d103f under corrected 17,000-point Hardcore scoring, except the depth-5 seven-stratum arm, which was stopped by decision after the first 32 seeds (0xa51d1000-0xa51d101f) and is drawn as its own one-point series so the two sample sizes are never joined by a line. Depth 2 was run only at seven strata. No confidence band is drawn on means; the paired contrasts with their bounds are in the companion bar figure. Evidence tier: development / public-development. The historical eight-game D3/D4 cohorts used 7,000-point scoring and are not shown.

• docs/exploratory/finding-05-chance-strata.md
• RS-20260821T181917Z-9a34ba02
• RS-20260821T205102Z-d89df4b5

Spec: web/content/figures/score-vs-depth.json · 3 source records · kind line

The same factorial as paired contrasts, which is the form the conclusions rest on:

Paired contrasts in the depth x chance-resolution factorial (mean delta, one-sided 95% lower bound)

• Chance resolution: 7 strata minus 5 strata
• Depth: 5 plies minus 4 plies

Notes and sources

Whisker is the one-sided 95% whole-game bootstrap lower bound; the upper end of the whisker is the point estimate itself (no upper bound is drawn). A contrast is significant when its lower bound clears zero: both stratum contrasts (7 minus 5 strata) do; neither depth contrast (5 minus 4 plies) does, and each sits below its own detection floor (47,052 at n=64, 107,988 at n=32). Cohort 0xa51d1000-0xa51d103f; the n=32 contrasts cover its first 32 seeds. Evidence tier: public-development.

• RS-20260821T205102Z-d89df4b5

Spec: web/content/figures/strata-5-vs-7.json · 1 source record · kind bar

The simulator and reference searches are mature enough to support reproducible research, but the strategy problem is unsolved. Corrected-score fair depth-4 expectimax is the strongest dependable reference found so far. Its average score across 64 games is 308,296 points, far below the goal of a one-million-point average.

Exploratory work on 2026-08-20/21 reproduced that reference on fresh seeds and completed the depth-by-chance-resolution factorial. Two results from it strongly suggest what to do next:

• The fourth ply is worth a large, significant gain — but only with an exact chance model. With the approximate five-stratum model the same fourth ply is worth slightly less than nothing. Chance resolution does not merely add points; it changes the sign of the depth gradient.
• The fifth ply is not measurable by this design. Both depth-5 contrasts sit far inside their own detection floors, and the seven-stratum estimate changed sign when the cohort grew from 16 games to 32, so the supported statement is that any depth-5 effect is smaller than a 64-game paired cohort can resolve — not that it is zero. What can be said is economic: even the optimistic estimate buys its points at 35.6x the work per move.
• A 64-game paired cohort in this game cannot see an effect below roughly 50,000 points. Every null result in the depth factorial is below its own detection floor and is therefore a non-measurement rather than evidence of no effect (RS-20260821T205102Z-d89df4b5). Choose experiments whose predicted effect exceeds the floor, or find a lower-variance estimator than complete games.

A fourth day, 2026-08-23, went to learned models and label economics — the reveal-construction probe in live play, the optimistic-state D0 gate, the leaf-cost NNUE student C0, and the P-SOL G0 label-semantics guardrail — and closed each of those four directions as tested; the research log for that day tells the story, and the closed-directions table below carries the headline numbers.

No candidate has qualified for the protected validation protocol. The frozen record marks both the protected and one-shot final cohorts as unopened. See the research status evidence for details.