Row-and-column n-tuple network trained by temporal-coherence TD on the Rust engine, screened as the leaf of the depth-3 seven-stratum fair search

Candidate: the stock fair expectimax search at depth 3, seven chance strata, terminal utility -1,000,000, policy seed 0xd7075eed, completion-guaranteeing work bound work_bound_for(3,7)+1, 64k-entry direct-mapped table (the nnue-evolution deployment configuration), with an n-tuple lookup-table leaf: the sum, over the active patterns of the reflection-canonical board, of learned f32 entries in rise units (x 17,000 points), the layout chosen by the pilot from {rows, cols, win23, win32} x {phase=none, cols, all} where rows/cols are absolute-position seven-cell line tuples (10^7 entries each), win23/win32 are absolute-position 2x3 and 3x2 window tuples (30 placements each, 10^6 entries each) and phase conditions a family on the moves remaining until the rise (x5).

completedtier: SCREENalgorithmicpublic-developmentpublic-policyEX-20260905-ntuple-scale-tc-td-leaf-d3-535b2620

On this page

Linked approachRow and column lookup tables, learned from a billion moves

Theories

TH-20260905-ntuple-line-tuples-tc-td-leaf-bcb25133

Results

RS-20260905T215332Z-95d18a5a

Dates

Created: 2026-09-05T19:13:17Z
Updated: 2026-09-05T21:55:18Z

Record idEX-20260905-ntuple-scale-tc-td-leaf-d3-535b2620

No explanation has been written for this record yet.

Technical recordThe registered protocolEX-20260905-ntuple-scale-tc-td-leaf-d3-535b2620

Hypothesis: Candidate: the stock fair expectimax search at depth 3, seven chance strata, terminal utility -1,000,000, policy seed 0xd7075eed, completion-guaranteeing work bound work_bound_for(3,7)+1, 64k-entry direct-mapped table (the nnue-evolution deployment configuration), with an n-tuple lookup-table leaf: the sum, over the active patterns of the reflection-canonical board, of learned f32 entries in rise units (x 17,000 points), the layout chosen by the pilot from {rows, cols, win23, win32} x {phase=none, cols, all} where rows/cols are absolute-position seven-cell line tuples (10^7 entries each), win23/win32 are absolute-position 2x3 and 3x2 window tuples (30 placements each, 10^6 entries each) and phase conditions a family on the moves remaining until the rise (x5). The visible next disc is not an input. Weights: fresh tables initialised to 20 rise units spread over the active entries, trained on-policy by TD(0) with temporal-coherence per-entry step sizes (alpha x |E|/A, alpha = 1.0 unless the pilot selects 0.25; error clamped to +-30 rise units) from complete games played by the greedy one-ply chance-state policy over the same tables (seven stratified reveal samples per column from the public sampler, epsilon 0) on the Rust bitboard engine, 32 asynchronous workers sharing the tables without locks (Hogwild; a lost update is tolerated and training is therefore not bit-reproducible across runs, disclosed as a limitation; the gate binary proves the serial update rule is deterministic). Stage 0 (CHECK): gate --probe-start 0xa5277000 on the already-opened Rust-engine probe block; all gates must pass before the first leased seed is read. Stage A (PILOT, training-role seeds): six configuration arms trained from scratch for 2 x 10^8 moves each on SL-20260905T191317Z-09895ea2 read in order from 0xa5300000 (every arm reads the same seed order), validated every 5 x 10^7 moves on the 64-game training-role validation block SL-20260905T191317Z-549265de (0xa52f2240) with the paired line-up ntuple-d3s7 / ntuple-1ply / fair-d3s7 on identical seeds: A rows,cols,win23,win32,phase=cols alpha 1.0; B rows,cols,win23,win32,phase=none alpha 1.0; C rows,cols,win23,win32,phase=all alpha 1.0; D rows,cols,phase=cols alpha 1.0 (lines only); E win23,win32,phase=none alpha 1.0 (windows only, the mechanism ablation); F rows,cols,win23,win32,phase=cols alpha 0.25. Selection rule, fixed here: the arm whose final validation point (2 x 10^8 moves) has the largest paired mean score margin of ntuple-d3s7 over fair-d3s7; ties by fewer table entries. Stage B (main run, training-role seeds): the selected configuration trained from fresh tables for 4 x 10^9 moves or 18,000 s of wall time, whichever comes first, on the same training lease and seed order, chunk 2 x 10^7 moves, validation every 2 x 10^8 moves on the same validation block, checkpoint at every validation point; the candidate is the tables at the validation point with the largest paired margin (best-weights.bin; ties to the later point). Stage C (freeze): SHA-256 of best-weights.bin recorded as candidate-weights.sha256 before the screen lease opens. Stage D (SCREEN, public-development, one-shot): 256 paired games on SL-20260905T191317Z-c25f58cd (0xa52f2140-0xa52f2240) with four arms on identical seeds: candidate-d3s7 (the candidate), candidate-1ply (the same tables played directly, diagnostic), fair-d3s7 (the comparator: the identical search with the frozen fair leaf) and fair-d4s7 (the program's standing reference, diagnostic). Every stage runs through approaches/ntuple-rl/ntuple-scale/scripts/pipeline.sh, which hard-codes the leased sub-blocks and records per-stage wall, CPU and peak-RSS usage. Timing measured on the 2026-09-05 smoke run: 1.4 x 10^6 training moves per second on 32 threads with the arm-A layout (4.8 x 10^8 entries, 5.76 GB trainable), so the pilot costs about 15 minutes and the main run about 50 minutes of training plus validation.
Arms: Arm Name Entry point Manifest
Candidate d3s7-ntuple-scale-leaf approaches/ntuple-rl/ntuple-scale/src/bin/train.rs –
Comparator fair-d3s7 approaches/fair-expectimax/rust-engine/src/leaf.rs research/benchmarks/baselines-v1.json
Classification: algorithmic
Information boundary: public-policy
Benchmark tier: SCREEN
Lifecycle: completed
Theories tested: A row-and-column n-tuple network trained by temporal-coherence TD on a billion engine moves beats the frozen fair leaf inside the same depth-3 search
Primary metric: paired mean whole-game score delta, candidate-d3s7 minus fair-d3s7, both at the identical d3s7 configuration, 256 held-out games
Secondary metrics: paired mean moves delta; numbered clears per move and cover reveals per move; mean occupied cells
lower quartile, median, maximum score; W-T-L; first-half (seeds 0-127) and second-half (128-255) paired mean deltas
diagnostic arm on the same seeds: candidate-1ply (the tables played directly one ply, seven reveal strata)
reference arm on the same seeds: fair-d4s7 (the program's standing reference configuration), diagnostic only
training-curve diagnostics: per-validation-point paired margin of ntuple-d3s7 and ntuple-1ply over fair-d3s7 on the training-role validation block, training-game mean score and length, mean |TD error|, mean coherence, touched table entries
pilot diagnostics: the six arms' validation curves, including the windows-only mechanism ablation
Statistical unit: whole-game
Uncertainty method: one-sided 95% percentile bootstrap over whole games, 20,000 resamples, RNG seed 0xb0071eaf (the unchanged compare.py of the leaf-evolution screen), plus a one-sided 95% Student-t lower bound; detection floor 1.645*sd/sqrt(n) reported
Data role: public-development
Seed leases: SL-20260905T191317Z-09895ea2
SL-20260905T191317Z-549265de
SL-20260905T191317Z-c25f58cd
Whole-origin split: yes
Reuse disclosure: CHECK gates and the throughput smoke run read only the already-opened development probe block 0xa5276000-0xa5277fff (the Rust engine's gate/benchmark range); no new seed is opened for mechanics or timing. Training (pilot arms and main run) reads the training lease 0xa5300000-0xa54f0000 in order and wraps when the block is exhausted (wrap counts recorded per run); the validation block 0xa52f2240-0xa52f2280 (training role) is re-read at every validation point and drives configuration and checkpoint selection; the held-out screen opens 0xa52f2140-0xa52f2240 exactly once, after the candidate is frozen. Training, validation and screen ranges are disjoint by construction.
Pass criteria: All CHECK gates passed before the first training seed was read: codec against Horner, row gather against the cell accessor, feature indices against the independent reference for every layout, information-boundary blindness (score/level/moves_played/next_disc), reflection consistency of values and decisions, direct-policy legality, leaf-in-search determinism and worker-count independence, serial training determinism, finiteness.
Every validation artifact and the screen artifact have illegalDecisions 0 and incompleteDecisions 0; censored games are reported as censored.
Held-out screen, 256 paired games on 0xa52f2140-0xa52f223f, candidate-d3s7 vs fair-d3s7: bootstrap 95% lower bound of the paired score delta > 0 AND Student-t 95% lower bound > 0.
Held-out screen: paired mean score delta > 0 in both halves (seeds 0-127 and 128-255).
Held-out screen: candidate-d3s7 lower-quartile score >= fair-d3s7 lower-quartile score.
The screened candidate is the exact frozen best-weights.bin of the main run (candidate-weights.sha256 recorded before the screen lease opens); no other table is screened.
On pass: Freeze best-weights.bin with its SHA-256; register a fresh-block replication by a different runner and a fresh-development STANDARD evaluation as successors; do not open protected or final seeds.
On fail: Record a valid run with scientific outcome fail for this exact configuration (tuple layout, phase conditioning, optimiser, budget, deployment depth); do not adopt the tables; the diagnostic and reference arms and the pilot curves are reported as diagnostics; open no further cohort for this candidate.
Gate fixed before controlled data: yes
Resources: Wall seconds 43200
CPU threads 32
Max host bytes 34359738368
Max GPU bytes –
GPU devices –
Stop conditions: Stop on any rules, information-boundary, legality, determinism, or parity failure.
A pilot arm stops at 2 x 10^8 moves; the main run stops at 4 x 10^9 moves or 18,000 s of training wall time, whichever comes first, or when a STOP file appears in the run directory; the candidate is then the best validation point so far.
A validation or screen artifact with any illegal or incomplete decision voids the run (invalid), not the candidate.
The screen is evaluated exactly once; no re-run on the same or a different held-out block without a new experiment record.
The theory's training-signal falsifier (no validation point with a positive paired margin) is evaluated and reported separately; the screen is opened once on the frozen candidate regardless.
Expected artifacts: runs/<run-id>/ntuple-scale/gates.log (CHECK)
runs/<run-id>/ntuple-scale/pilot/<arm>/{config.json,progress.jsonl,val-*.json,best.json,best-weights.bin,latest-weights.bin} (stage A) and pilot/selection.json
runs/<run-id>/ntuple-scale/main/{config.json,progress.jsonl,val-*.json,best.json,best-weights.bin,latest-weights.bin,checkpoint.bin,DONE} (stage B)
runs/<run-id>/ntuple-scale/main/candidate-weights.sha256 (stage C)
runs/<run-id>/ntuple-scale/screen/heldout.json + compare-*.json (stage D)
runs/<run-id>/ntuple-scale/{pipeline.log,rusage.jsonl,analysis.json,analysis.md}
artifacts emitted by train and screen carry this experiment id in their config field; table files are too large for the repository and are retained on the workstation with their SHA-256 in the result record
Amendments: Timestamp Before controlled data Reason
2026-09-05T21:55:18Z no lifecycle advanced running -> completed after run RUN-20260905T193006Z-4fbeb4e5 and result RS-20260905T215332Z-95d18a5a were written; protocol content otherwise unchanged (frozen hash ee8eabb0512f459bb0acbb717e335d5e6953d2ac7ec57400f2a35b9f1c76b65d retained in the run and lease records). Operational note: the pipeline's select-arm stage aborted on an empty output file between the pilot and the main run (fixed in commit bd87dba); the relaunch changed no block, gate, stage order or artifact.

Arm	Name	Entry point	Manifest
Candidate	d3s7-ntuple-scale-leaf	`approaches/ntuple-rl/ntuple-scale/src/bin/train.rs`	–
Comparator	fair-d3s7	`approaches/fair-expectimax/rust-engine/src/leaf.rs`	`research/benchmarks/baselines-v1.json`

Wall seconds	43200
CPU threads	32
Max host bytes	34359738368
Max GPU bytes	–
GPU devices	–

Timestamp	Before controlled data	Reason
2026-09-05T21:55:18Z	no	lifecycle advanced running -> completed after run RUN-20260905T193006Z-4fbeb4e5 and result RS-20260905T215332Z-95d18a5a were written; protocol content otherwise unchanged (frozen hash ee8eabb0512f459bb0acbb717e335d5e6953d2ac7ec57400f2a35b9f1c76b65d retained in the run and lease records). Operational note: the pipeline's select-arm stage aborted on an empty output file between the pilot and the main run (fixed in commit bd87dba); the relaunch changed no block, gate, stage order or artifact.

Technical recordResults recorded against this protocol1 record

valid runoutcome: passsupported-as-testedtier: public-developmentRS-20260905T215332Z-95d18a5a

Held-out screen, 256 never-read paired public-development games (0xa52f2140+): the frozen n-tuple tables as the leaf of the depth-3 seven-stratum fair search averaged 484,577 points and 140.21 moves against 314,438 points and 92.58 moves for the identical search with the frozen fair leaf: paired +170,139 points (bootstrap 95% lower bound +130,499, Student-t lower bound +130,560, upper bound +209,881, detection floor 39,440), W-T-L 167-0-89, halves +198,575 / +141,704, lower quartile 212,820 vs 175,832, moves +47.63. The preregistered gate PASSES. Against the program's standing reference, the fair leaf at depth 4 on the same seeds (377,803 points), the candidate is +106,775 paired (bootstrap LB +62,574, UB +151,706, W-T-L 160-0-96); diagnostic only. The same tables played directly one ply averaged 281,441 points, -32,997 paired against fair-d3s7 (LB -58,536); diagnostic. Fair-d4s7 minus fair-d3s7 on these seeds: +63,365 (LB +31,538). Candidate: the tables of the main run's validation point at 1,400,240,319 training moves (layout rows,cols,win23,win32,phase=all, alpha 1.0, selected from six pilot arms by the preregistered rule with a final validation margin of +123,950), whose paired margin on the 64-game training-role validation block was +270,023; SHA-256 0ade9d4e4080ebdd52a1474b1a13410dc8dfb77f5eba24b078aa7703c92ace0b. Main run: 4,000,004,271 training moves, 45,852,198 games, 20 validation points, mean 1,956,636 moves per second. Mechanism ablation (pilot, windows-only arm E): final validation margin -83,261 against the full layout's +95,051 (arm A) and +123,950 (arm C, selected). Training-signal check: at least one validation point of the main run had a positive paired margin.

What it had to pass

✓All CHECK gates passed before the first training seed was read — observed: 15 gate lines, all PASS
✓Every validation artifact and the screen artifact have illegalDecisions 0 and incompleteDecisions 0 — observed: main validations illegal 0 / incomplete 0; screen arms candidate-d3s7: 0/0, candidate-1ply: 0/0, fair-d3s7: 0/0, fair-d4s7: 0/0
✓bootstrap 95% lower bound of candidate-d3s7 minus fair-d3s7 > 0 — observed: 130499.0115234375
✓Student-t 95% lower bound > 0 — observed: 130559.5594985295
✓paired mean delta > 0 in both halves — observed: [198574.609375, 141704.171875]
✓candidate-d3s7 Q25 >= fair-d3s7 Q25 — observed: [212819.75, 175831.5]
✓The screened candidate is the exact frozen best-weights.bin (SHA-256 recorded before the screen lease opened) — observed: 0ade9d4e4080ebdd52a1474b1a13410dc8dfb77f5eba24b078aa7703c92ace0b

Technical recordRecorded metricsRS-20260905T215332Z-95d18a5a

screen

games

256

seedStartHex

0xa52f2140

arms

candidate-d3s7

meanScore: 484577.3867
medianScore: 407473.5000
q25Score: 212819.7500
minScore: 102,537
maxScore: 2,295,869
sdScore: 359272.8703
meanMoves: 140.2109
q25Moves: 65
numberedClearsPerMove: 2.1074
coverRevealsPerMove: 1.1899
meanOccupiedCells: 23.2670
censoredGames: 0
illegalDecisions: 0
incompleteDecisions: 0
gamesAtOrAboveMillion: 21
meanWallSecondsPerGame: 4.2907

candidate-1ply

meanScore: 281440.7266
medianScore: 230849.5000
q25Score: 157750.5000
minScore: 85,331
maxScore: 815,564
sdScore: 164490.1308
meanMoves: 84.3047
q25Moves: 50
numberedClearsPerMove: 1.9217
coverRevealsPerMove: 1.0608
meanOccupiedCells: 24.2171
censoredGames: 0
illegalDecisions: 0
incompleteDecisions: 0
gamesAtOrAboveMillion: 0
meanWallSecondsPerGame: 0.0009

fair-d3s7

meanScore: 314437.9961
medianScore: 250,821
q25Score: 175831.5000
minScore: 85,561
maxScore: 1,221,000
sdScore: 199663.2623
meanMoves: 92.5820
q25Moves: 55
numberedClearsPerMove: 1.9848
coverRevealsPerMove: 1.1026
meanOccupiedCells: 24.4343
censoredGames: 0
illegalDecisions: 0
incompleteDecisions: 0
gamesAtOrAboveMillion: 4
meanWallSecondsPerGame: 4.6279

fair-d4s7

meanScore: 377802.6289
medianScore: 282076.5000
q25Score: 192561.5000
minScore: 102,904
maxScore: 1,794,652
sdScore: 276607.9151
meanMoves: 109.3359
q25Moves: 60
numberedClearsPerMove: 2.0492
coverRevealsPerMove: 1.1499
meanOccupiedCells: 23.7628
censoredGames: 0
illegalDecisions: 0
incompleteDecisions: 0
gamesAtOrAboveMillion: 9
meanWallSecondsPerGame: 182.2242

contrasts

candidate-d3s7-vs-fair-d3s7

score

n: 256
meanDelta: 170139.3906
pairedSd: 383606.5536
bootstrapLower95: 130499.0115
bootstrapUpper95: 209881.3934
studentTLower95: 130559.5595
detectionFloor: 39439.5488
wins: 167
ties: 0
losses: 89
firstHalfMeanDelta: 198574.6094
secondHalfMeanDelta: 141704.1719
q25Delta: 36988.2500
candidateQ25: 212819.7500
referenceQ25: 175831.5000

moves

n: 256
meanDelta: 47.6289
pairedSd: 106.0384
bootstrapLower95: 36.6639
bootstrapUpper95: 58.6172
studentTLower95: 36.6881
detectionFloor: 10.9021
wins: 164
ties: 9
losses: 83
firstHalfMeanDelta: 55.2969
secondHalfMeanDelta: 39.9609
q25Delta: 10
candidateQ25: 65
referenceQ25: 55

candidate-1ply-vs-fair-d3s7

score

n: 256
meanDelta: -32997.2695
pairedSd: 246695.1922
bootstrapLower95: -58536.1652
bootstrapUpper95: -7597.0516
studentTLower95: -58450.8338
detectionFloor: 25363.3494
wins: 114
ties: 0
losses: 142
firstHalfMeanDelta: -11748.5156
secondHalfMeanDelta: -54246.0234
q25Delta: -18,081
candidateQ25: 157750.5000
referenceQ25: 175831.5000

moves

n: 256
meanDelta: -8.2773
pairedSd: 67.8083
bootstrapLower95: -15.2854
bootstrapUpper95: -1.2967
studentTLower95: -15.2737
detectionFloor: 6.9715
wins: 112
ties: 12
losses: 132
firstHalfMeanDelta: -2.3516
secondHalfMeanDelta: -14.2031
q25Delta: -5
candidateQ25: 50
referenceQ25: 55

candidate-d3s7-vs-fair-d4s7

score

n: 256
meanDelta: 106774.7578
pairedSd: 435009.0495
bootstrapLower95: 62574.2055
bootstrapUpper95: 151705.7729
studentTLower95: 61891.3100
detectionFloor: 44724.3679
wins: 160
ties: 0
losses: 96
firstHalfMeanDelta: 109211.3125
secondHalfMeanDelta: 104338.2031
q25Delta: 20258.2500
candidateQ25: 212819.7500
referenceQ25: 192561.5000

moves

n: 256
meanDelta: 30.8750
pairedSd: 119.7043
bootstrapLower95: 18.6875
bootstrapUpper95: 43.2150
studentTLower95: 18.5241
detectionFloor: 12.3071
wins: 157
ties: 7
losses: 92
firstHalfMeanDelta: 31.5938
secondHalfMeanDelta: 30.1563
q25Delta: 5
candidateQ25: 65
referenceQ25: 60

fair-d4s7-vs-fair-d3s7

score

n: 256
meanDelta: 63364.6328
pairedSd: 311371.6478
bootstrapLower95: 31538.2639
bootstrapUpper95: 95448.5086
studentTLower95: 31237.8688
detectionFloor: 32012.8975
wins: 138
ties: 0
losses: 118
firstHalfMeanDelta: 89363.2969
secondHalfMeanDelta: 37365.9688
q25Delta: 16,730
candidateQ25: 192561.5000
referenceQ25: 175831.5000

moves

n: 256
meanDelta: 16.7539
pairedSd: 84.5155
bootstrapLower95: 8.1170
bootstrapUpper95: 25.4572
studentTLower95: 8.0338
detectionFloor: 8.6892
wins: 132
ties: 11
losses: 113
firstHalfMeanDelta: 23.7031
secondHalfMeanDelta: 9.8047
q25Delta: 5
candidateQ25: 60
referenceQ25: 55

candidate-d3s7-vs-candidate-1ply

score

n: 256
meanDelta: 203136.6602
pairedSd: 359655.2798
bootstrapLower95: 167198.7932
bootstrapUpper95: 240597.7713
studentTLower95: 166028.0782
detectionFloor: 36977.0585
wins: 185
ties: 0
losses: 71
firstHalfMeanDelta: 210323.1250
secondHalfMeanDelta: 195950.1953
q25Delta: 55069.2500
candidateQ25: 212819.7500
referenceQ25: 157750.5000

moves

n: 256
meanDelta: 55.9063
pairedSd: 99.8785
bootstrapLower95: 45.9451
bootstrapUpper95: 66.3281
studentTLower95: 45.6010
detectionFloor: 10.2688
wins: 179
ties: 8
losses: 69
firstHalfMeanDelta: 57.6484
secondHalfMeanDelta: 54.1641
q25Delta: 15
candidateQ25: 65
referenceQ25: 50

main

layout

rows,cols,win23,win32,phase=all

alpha

entries

1,000,000,000

movesTotal

4,000,004,271

gamesTotal

45,852,198

wallSeconds

2468.6000

meanMovesPerSecond

1956636.3300

best

moves: 1,400,240,319
artifact: val-001400240319.json
pairedDeltaD3: 270022.7813
ntupleD3Mean: 607633.3750
fairD3Mean: 337610.5938

validations

movesTrained
200,027,819
ntupleD3Mean
453535.5625
directMean
248686.2188
fairD3Mean
337610.5938
pairedDeltaD3
115924.9688
bootstrapLower95
46102.8594
wins
41
losses
23
movesTrained
400,060,957
ntupleD3Mean
473806.8906
directMean
281009.3750
fairD3Mean
337610.5938
pairedDeltaD3
136196.2969
bootstrapLower95
60912.2789
wins
39
losses
25
movesTrained
600,095,322
ntupleD3Mean
481647.8594
directMean
281904.2500
fairD3Mean
337610.5938
pairedDeltaD3
144037.2656
bootstrapLower95
70139.7203
wins
40
losses
24
movesTrained
800,131,078
ntupleD3Mean
491212.7813
directMean
255970.2188
fairD3Mean
337610.5938
pairedDeltaD3
153602.1875
bootstrapLower95
69106.0609
wins
39
losses
25
movesTrained
1,000,167,066
ntupleD3Mean
604501.7031
directMean
302303.4375
fairD3Mean
337610.5938
pairedDeltaD3
266891.1094
bootstrapLower95
158786.7648
wins
44
losses
20
movesTrained
1,200,203,606
ntupleD3Mean
474549.7031
directMean
265703.8906
fairD3Mean
337610.5938
pairedDeltaD3
136939.1094
bootstrapLower95
61272.7938
wins
41
losses
23
movesTrained
1,400,240,319
ntupleD3Mean
607633.3750
directMean
323224.9219
fairD3Mean
337610.5938
pairedDeltaD3
270022.7813
bootstrapLower95
165079.2219
wins
47
losses
17
movesTrained
1,600,277,123
ntupleD3Mean
509887.6719
directMean
276702.4375
fairD3Mean
337610.5938
pairedDeltaD3
172277.0781
bootstrapLower95
84741.3750
wins
43
losses
21
movesTrained
1,800,315,771
ntupleD3Mean
531421.6250
directMean
319334.0156
fairD3Mean
337610.5938
pairedDeltaD3
193811.0313
bootstrapLower95
107981.8070
wins
41
losses
23
movesTrained
2,000,352,773
ntupleD3Mean
593476.5781
directMean
318354.1406
fairD3Mean
337610.5938
pairedDeltaD3
255865.9844
bootstrapLower95
160477.3344
wins
39
losses
25
movesTrained
2,200,391,703
ntupleD3Mean
453484.7813
directMean
327689.7344
fairD3Mean
337610.5938
pairedDeltaD3
115874.1875
bootstrapLower95
38562.4789
wins
41
losses
23
movesTrained
2,400,429,413
ntupleD3Mean
552723.8125
directMean
304749.6250
fairD3Mean
337610.5938
pairedDeltaD3
215113.2188
bootstrapLower95
116273.5383
wins
43
losses
21
movesTrained
2,600,469,503
ntupleD3Mean
482608.7500
directMean
285457.5313
fairD3Mean
337610.5938
pairedDeltaD3
144998.1563
bootstrapLower95
70309.5562
wins
44
losses
20
movesTrained
2,800,506,807
ntupleD3Mean
496445.7344
directMean
302053.8594
fairD3Mean
337610.5938
pairedDeltaD3
158835.1406
bootstrapLower95
67962.9695
wins
37
losses
27
movesTrained
3,000,545,916
ntupleD3Mean
517928.1250
directMean
295604.6563
fairD3Mean
337610.5938
pairedDeltaD3
180317.5313
bootstrapLower95
102679.1250
wins
43
losses
21
movesTrained
3,200,585,530
ntupleD3Mean
547107.2344
directMean
298385.6094
fairD3Mean
337610.5938
pairedDeltaD3
209496.6406
bootstrapLower95
133542.7383
wins
44
losses
20
movesTrained
3,400,627,531
ntupleD3Mean
506536.4219
directMean
317633.5781
fairD3Mean
337610.5938
pairedDeltaD3
168925.8281
bootstrapLower95
75604.8477
wins
41
losses
23
movesTrained
3,600,668,274
ntupleD3Mean
470056.4063
directMean
303216.0313
fairD3Mean
337610.5938
pairedDeltaD3
132445.8125
bootstrapLower95
58261.5695
wins
41
losses
23
movesTrained
3,800,707,333
ntupleD3Mean
507070.8906
directMean
296535.6563
fairD3Mean
337610.5938
pairedDeltaD3
169460.2969
bootstrapLower95
90402.4133
wins
45
losses
19
movesTrained
4,000,004,271
ntupleD3Mean
530088.1406
directMean
320977.1563
fairD3Mean
337610.5938
pairedDeltaD3
192477.5469
bootstrapLower95
113016.1180
wins
43
losses
21

anyPositiveMargin

true

pilot

layout: rows,cols,win23,win32,phase=cols
alpha: 1
entries: 480,000,000
movesTotal: 200,003,947
finalMargin: 95050.8281
bestMargin: 170874.3750

layout: rows,cols,win23,win32,phase=none
alpha: 1
entries: 200,000,000
movesTotal: 200,002,183
finalMargin: -70586.8906
bestMargin: -35090.6563

layout: rows,cols,win23,win32,phase=all
alpha: 1
entries: 1,000,000,000
movesTotal: 200,002,957
finalMargin: 123949.5156
bestMargin: 123949.5156

layout: rows,cols,phase=cols
alpha: 1
entries: 420,000,000
movesTotal: 200,002,392
finalMargin: 26950.8438
bestMargin: 94015.3750

layout: win23,win32,phase=none
alpha: 1
entries: 60,000,000
movesTotal: 200,001,766
finalMargin: -83261.4531
bestMargin: -80696.1250

layout: rows,cols,win23,win32,phase=cols
alpha: 0.2500
entries: 480,000,000
movesTotal: 200,003,528
finalMargin: 73116.2031
bestMargin: 112885.6719

pilotSelection

arm

layout

rows,cols,win23,win32,phase=all

alpha

finalMargin

123949.5156

ranking

arm
C
finalMargin
123949.5156
entries
1,000,000,000
arm
A
finalMargin
95050.8281
entries
480,000,000
arm
F
finalMargin
73116.2031
entries
480,000,000
arm
D
finalMargin
26950.8438
entries
420,000,000
arm
B
finalMargin
-70586.8906
entries
200,000,000
arm
E
finalMargin
-83261.4531
entries
60,000,000

rule

the arm whose final validation point has the largest paired mean margin of ntuple-d3s7 over fair-d3s7; ties by fewer table entries

candidateSha256

0ade9d4e4080ebdd52a1474b1a13410dc8dfb77f5eba24b078aa7703c92ace0b

Limitations

Public-development SCREEN tier, 256 paired games opened once; nothing here is a qualification claim, and protected and final cohorts stay sealed.
The candidate is the validation point with the largest paired margin on a 64-game training-role block that was read at every validation point and also chose the configuration; that selection is upward-biased, which is why the held-out screen exists.
Training used lock-free asynchronous updates from 32 threads, so the training run is not bit-reproducible; the frozen tables are hashed and every gameplay arm is deterministic and worker-count independent.
The fair-d4s7 arm is the program's standing reference for context only; the preregistered comparator is the identical depth-3 search with the frozen fair leaf.
Table files (1.9 GB weights, 5.8 GB with accumulators for the pilot layouts; 4.0 GB and 12 GB for the selected phase=all layout) are retained on the workstation with their SHA-256 and are not committed.
Wall times were measured on a shared workstation with the web console building concurrently for part of the run; ratios between arms on the same seeds are the trustworthy quantity.

Open the result record

Agent contextHow to extend this record

To add a reader-facing explanation, write web/content/research/EX-20260905-ntuple-scale-tc-td-leaf-d3-535b2620.mdx; it renders above this record on the next request. The registered protocol itself is in the technical record above.

Record file: research/experiments/EX-20260905-ntuple-scale-tc-td-leaf-d3-535b2620.json, validated against research/schemas/experiment-v1.schema.json. Protocol hash: 4a867db795eec7e5e0858e58d6f59b14a13ed8ec03629cd155ce37a4b174b718.

Registered by Claude Code / claude-fable-5-1 (claude-code-ntuple-scale-experiment).