Depth-5-distilled NNUE leaf refined by paired-fitness whole-game evolution inside the depth-3 fair search (v2: operational parameters fixed in the record body)

On this page

Linked approachAn evolved NNUE leaf, distilled from a depth-5 teacher

Theories

TH-20260825-evolved-nnue-leaf-d3-0f47e46c

Results

RS-20260903T025751Z-6577b33e

Dates

Created: 2026-09-02T06:57:41Z
Updated: 2026-09-03T02:57:51Z

Record idEX-20260902-nnue-evolution-d3-v2-49c18bc2

No explanation has been written for this record yet.

Technical recordThe registered protocolEX-20260902-nnue-evolution-d3-v2-49c18bc2

Hypothesis: Successor to EX-20260825-nnue-evolution-d3-bca7f330. The scientific protocol is carried over unchanged: candidate, comparator, information boundary, benchmark tier, leases, metrics, uncertainty method, pass criteria, failure and pass actions, GA constants and stop rules are those of v1 verbatim below. What v2 adds is the set of operational parameters v1 left open, fixed before any leased seed was read (the training lease opened 2026-09-02T04:24:10Z; these values were first recorded at 04:02Z and 04:23Z as in-place amendments of v1, which the repository rule on frozen records does not permit, hence this successor): (1) stage A plays with --move-cap 500 (teacher game lengths are heavy-tailed per the August 25 T0 pilot; labels are search values, so the cap truncates the late-game state distribution but biases no label; capped games are recorded as censored) and admits no new game after a 46,800 s wall sub-budget, in-flight games finishing and being kept, the completed whole games forming the corpus as the v1 stop rule already permits; (2) stage B uses batch 64, lr 3e-4, Huber delta 1.0 rise unit, seed 0x0e701e57, 16 epochs with the best epoch chosen by validation Huber, and 256 held-out probe roots; (3) stage C uses the v1 GA constants with a 21,600 s wall sub-budget and the elite re-selection block 0xa52e0c80 exactly as leased; (4) every stage runs with 32 threads: a seed-free scaling preflight on the already-opened probe block 0xa5277800-0xa527783f (64 complete d4s7 teacher-corpus games, 7,830 roots) gave bit-identical root and game records at 16 and 32 threads and 1.020 vs 1.577 s per root per thread, a 1.29x throughput gain from SMT on the 16 physical cores; thread count changes no decision, label, seed, gate or metric; (5) the held-out screen is opened once on the frozen candidate regardless of the theory's training-signal falsifier, which is evaluated and reported separately from the screen gate. Every stage is launched through approaches/lifetime-objective/nnue-evolution/scripts/pipeline.sh, which hard-codes the leased sub-blocks and records per-stage wall, CPU and peak-RSS usage. v1 protocol as frozen: 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, with an NNUE leaf (8,902-feature sparse class, 135 active, EmbeddingBag(8902,64)->ReLU->32->ReLU->1, output x 17,000 points). Weights: initialised by supervised distillation of a depth-5 seven-stratum fair teacher's sibling-complete root values (512 training games, whole-origin split, Adam on Huber loss, best epoch by validation Huber), then refined by a mutation-only GA: population 32, 60 generations, fitness = mean score of 32 complete paired games per candidate on a fresh training block per generation, 4 elites cloned, tournament size 3, per-tensor Gaussian mutation sigma = 0.05 x tensor std (floor 1e-4), initial cloud 2 sigma, evolution seed 0x0e701e58, controls fair-d3s7 and init-d3s7 playing every block. Final candidate: top 8 of the last generation re-evaluated on a fresh 128-game block, best mean frozen. The frozen candidate is then screened once on 64 never-read paired development games against the identical search with the frozen fair leaf.
Arms: Arm Name Entry point Manifest
Candidate d3s7-evolved-nnue-leaf approaches/lifetime-objective/nnue-evolution/src/bin/evolve.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 depth-5-distilled, whole-game-evolved NNUE leaf beats the frozen fair leaf inside the same depth-3 search
Primary metric: paired mean whole-game score delta, evolved candidate minus frozen fair leaf, both at the identical d3s7 configuration, 64 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 and second-half paired mean deltas
ablation arm on the same seeds: the supervised-init (unevolved) NNUE at d3s7, isolating the evolutionary stage's contribution
reference arm on the same seeds: the fair leaf at d4s7 (the program's standing reference configuration), diagnostic only
training-curve diagnostics: per-generation best and mean fitness against the paired fair-d3s7 control on the same blocks
supervised-init diagnostics: whole-origin validation Huber/Pearson and the deployment-faithful ordering probe (d3s7+NNUE top-1 agreement with the teacher's chosen column, teacher-value regret) on held-out corpus roots
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 prior 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-20260825T063000Z-a52e0300
SL-20260825T063000Z-a52e1300
Whole-origin split: yes
Reuse disclosure: CHECK gates and the T0 timing pilot 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. The teacher corpus, every evolution fitness block, and the elite re-selection open training-lease seeds 0xa52e0300-0xa52e0d00; the held-out screen opens 0xa52e1300-0xa52e133f exactly once, after the candidate is frozen. Training and screen ranges are disjoint by construction. The 2026-09-02 thread-scaling preflight read only the already-opened probe block 0xa5277800-0xa527783f.
Pass criteria: All CHECK gates passed before the first training seed was read: feature determinism and bounds, information-boundary blindness (score/level/moves_played), reflection consistency on asymmetric boards (finding-13 exclusion), fresh-searcher and 1-vs-4-worker determinism, legality and completed depth under random/zero/saturated weights, leaf finiteness fuzz, serialisation round-trip.
Every generation artifact and the screen artifact have illegalDecisions 0, incompleteDecisions 0, and censored games reported as censored.
Held-out screen, 64 paired games on 0xa52e1300-0xa52e133f, candidate 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 0-31 and 32-63).
Held-out screen: candidate lower-quartile score >= fair-d3s7 lower-quartile score.
The screened candidate is the exact frozen winner of the elite re-selection (candidate-weights.bin, SHA-256 recorded before the screen lease opens); no other vector is screened.
On pass: Freeze candidate-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 (model class, teacher depth, optimiser, budget, deployment depth); do not adopt the vector; the ablation and reference arms are reported as diagnostics; open no further cohort for this candidate.
Gate fixed before controlled data: yes
Resources: Wall seconds 86400
CPU threads 32
Max host bytes 17179869184
Max GPU bytes –
GPU devices –
Stop conditions: Stop on any rules, information-boundary, legality, determinism, or parity failure.
The teacher corpus is checkpointed per game and stops at 512 games or the wall budget; whatever whole games completed are the corpus (the count is recorded; the supervised stage does not depend on it being exactly 512).
Evolution stops at 60 generations, at the wall budget, or when a STOP file appears in the run directory; the candidate is then the elite re-selection over the last completed generation's population.
A generation 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.
Stage A admits no new teacher game after 46,800 s of wall time; games in flight finish and are kept; the corpus is the set of completed whole games (count recorded).
Stage C stops at 60 generations or 21,600 s of wall time, whichever comes first; the elite re-selection then runs on the last completed generation.
Expected artifacts: runs/<run-id>/nnue-evolution/gates.log (CHECK)
runs/<run-id>/nnue-evolution/corpus.jsonl + corpus-summary.json (stage A)
runs/<run-id>/nnue-evolution/pretrain/{epoch-*.bin,init.bin,report.json,probe.json} (stage B)
runs/<run-id>/nnue-evolution/evolve/{config.json,progress.jsonl,gen-*.json,population-*.bin,final-fitness.json,candidate-weights.bin,selection.json} (stage C)
runs/<run-id>/nnue-evolution/screen/heldout.json + compare-*.json (stage D)
runs/<run-id>/nnue-evolution/{pipeline.log,rusage.jsonl,preflight-d4-t16.log,preflight-d4-t32.log,analysis.json,analysis.md}
artifacts emitted by evolve and screen carry this experiment id in their config field
Amendments: Timestamp Before controlled data Reason
2026-09-03T02:57:51Z no lifecycle advanced running -> completed after run RUN-20260902T035644Z-c1fd8987 and result RS-20260903T025751Z-6577b33e were written; protocol content otherwise unchanged. The preregistration hash 971a39ef6bb5275a951342245a119cb3d234bed334aa28e2ecfe44eeeeea4dae is retained in the run and lease records.

Arm	Name	Entry point	Manifest
Candidate	d3s7-evolved-nnue-leaf	`approaches/lifetime-objective/nnue-evolution/src/bin/evolve.rs`	–
Comparator	fair-d3s7	`approaches/fair-expectimax/rust-engine/src/leaf.rs`	`research/benchmarks/baselines-v1.json`

Wall seconds	86400
CPU threads	32
Max host bytes	17179869184
Max GPU bytes	–
GPU devices	–

Timestamp	Before controlled data	Reason
2026-09-03T02:57:51Z	no	lifecycle advanced running -> completed after run RUN-20260902T035644Z-c1fd8987 and result RS-20260903T025751Z-6577b33e were written; protocol content otherwise unchanged. The preregistration hash 971a39ef6bb5275a951342245a119cb3d234bed334aa28e2ecfe44eeeeea4dae is retained in the run and lease records.

Technical recordResults recorded against this protocol1 record

valid runoutcome: failnot-supported-as-testedtier: public-developmentRS-20260903T025751Z-6577b33e

Valid run of the frozen protocol EX-20260902-nnue-evolution-d3-v2-49c18bc2 on the whole workstation (32 threads), every stage completed, every artifact with zero illegal and zero incomplete decisions. Stage A: the depth-5 seven-stratum teacher played 177 complete games (21,618 sibling-complete labelled roots, 3 stopped at the 500-move cap) before the 46,800 s new-game cutoff; the teacher ran 77.5 s per root, four to seven times slower than the pilot projected, so the corpus is about a third of the 512 games the protocol allowed for. Stage B: the supervised warm start reached validation Huber 0.6873 rise units at epoch 8 of 16 on a whole-origin split (17,641/3,977 roots), and the deployment-faithful ordering probe put the depth-3 search with that leaf at top-1 agreement 0.4414 with the teacher on 256 held-out roots (mean teacher-value regret 2,395 points). Stage C: 60 generations of the mutation-only GA (population 32, 32 paired games per candidate per fresh block, fair-d3s7 and warm-start controls on every block). The population separated steadily from its warm start - ten-generation paired mean margins over the init control of +4,682, +15,853, +22,704, +32,278, +34,037, +40,431 points, best candidate +76,679 in the last block - but never approached the fair control: the population mean was above the fair leaf in 0 of the final 10 generations and the best candidate in 0 (mean margin -121,896), so the theory's training-signal falsifier fails. Elite re-selection on 128 fresh games froze candidate-29 at 202,237 (finalists spanned 187,352 to 202,237). Stage D: on the 64 never-read held-out games (0xa52e1300, opened once at 2026-09-03T02:41:29Z after the candidate's SHA-256 was recorded), the evolved candidate averaged 190,961 against the frozen fair leaf's 297,926 at the identical depth-3 configuration: paired delta -106,964 (bootstrap 95% bounds -146,580 to -69,983, Student-t lower bound -145,890, detection floor 38,357), W-T-L 14-0-50, both halves negative (-109,139 / -104,790), lower quartile 122,253 against 158,387. Every preregistered pass criterion except artifact integrity and candidate identity fails: scientific outcome fail for this exact configuration. The ablation arm shows what evolution did contribute: the unevolved warm start averaged 155,586 (-142,340 against the fair leaf), and the evolved candidate beat it on the same seeds by +35,375 (bootstrap lower bound +16,899, floor 18,949, W-T-L 36-0-28, both halves positive). Whole-game evolution with common random numbers therefore moves a 572k-weight leaf on the deployed objective, which the first (CMA-ES) leaf evolution could not show; it moved it about a quarter of the way from a warm start that plays at half the fair leaf's level. The reference arm reproduced the program's standing result: fair depth 4 over fair depth 3 +97,059 (lower bound +29,330). Read: the claim is not supported as tested; the mechanism's evolutionary leg is supported, its distillation leg is the weak link (a warm start that holds the teacher's values but not its ordering), and the budget (177 teacher games, 60 generations) was too small for evolution to cover the distance.

What it had to pass

✓All CHECK gates passed before the first training seed was read — observed: 11 of 11 gates passed on the already-open probe block (gates.log), rebuilt binaries, 9/9 unit tests
✓Every generation artifact and the screen artifact: illegalDecisions 0, incompleteDecisions 0, censored games reported — observed: 60 generation artifacts: 0 illegal, 0 incomplete, 0 censored; screen: 0 illegal, 0 incomplete, 0 censored in all four arms
✕Held-out screen, candidate vs fair-d3s7: bootstrap 95% lower bound of the paired score delta > 0 AND Student-t 95% lower bound > 0 — observed: delta -106,964; bootstrap LB -146,580, UB -69,983; t LB -145,890; paired sd 186,538; floor 38,357; W-T-L 14-0-50
✕Held-out screen: paired mean score delta > 0 in both halves (seeds 0-31 and 32-63) — observed: first half -109,139, second half -104,790
✕Held-out screen: candidate lower-quartile score >= fair-d3s7 lower-quartile score — observed: candidate Q25 122,253 vs fair-d3s7 Q25 158,387 (delta -36,134)
✓The screened candidate is the exact frozen winner of the elite re-selection; SHA-256 recorded before the screen lease opened — observed: candidate-29 of generation 59, mean 202,237 over 128 fresh games; sha256 edd0d2efd181de43f35d62c4df784cb2c789db3ae296be93a1c6c59082034a9f; recorded 2026-09-03T02:41:29Z, lease opened 2026-09-03T02:41:29Z by the screen stage
✕Theory falsifier (training signal): population mean fitness > paired fair-d3s7 control in the majority of the final 10 generations — observed: mean above fair in 0/10, top-4 in 0/10, best in 0/10; mean margin -121,896, top-4 margin -96,788
✓Theory falsifier (supervised initialisation): distilled NNUE's within-root top-1 agreement with the depth-5 teacher on whole-origin held-out roots exceeds the 0.30 level materially — observed: deployment-faithful probe (d3s7 search with the NNUE leaf vs the teacher's column) top-1 0.4414 on 256 held-out roots, regret 2,395 points; for scale, the leaf-swing diagnostic put the frozen fair leaf at 0.720 and a zero leaf at 0.450 on a different 400-root sample
–Theory falsifier (ablative): if the candidate passes but the unevolved init passes equally, evolution contributed nothing — observed: precondition not met (candidate failed); recorded anyway: candidate minus init +35,375 (bootstrap LB +16,899, t LB +16,145, floor 18,949, W-T-L 36-0-28, halves +47,004 / +23,746); init minus fair -142,340 (LB -178,492)

Technical recordRecorded metricsRS-20260903T025751Z-6577b33e

cohort

seedsStartHex: 0xa52e1300
games: 64
moveCap: 2,000
role: public-development, never read before this screen (lease SL-20260825T063000Z-a52e1300, opened once)

arms

candidate

games: 64
mean: 190961.2188
median: 160107.5000
q25: 122253.2500
max: 477,758
movesMean: 58.5000
clearsPerMove: 1.7131
revealsPerMove: 0.9124
censored: 0
illegal: 0
incomplete: 0
wallSeconds: 443.8840

init-d3s7

games: 64
mean: 155585.9688
median: 140,282
q25: 107423.2500
max: 401,038
movesMean: 48.7500
clearsPerMove: 1.5510
revealsPerMove: 0.7554
censored: 0
illegal: 0
incomplete: 0
wallSeconds: 352.7846

fair-d3s7

games: 64
mean: 297925.5938
median: 254999.5000
q25: 158387.2500
max: 844,318
movesMean: 88.0313
clearsPerMove: 1.9549
revealsPerMove: 1.0722
censored: 0
illegal: 0
incomplete: 0
wallSeconds: 265.4953

fair-d4s7

games: 64
mean: 394984.3125
median: 302,556
q25: 198524.7500
max: 1,663,637
movesMean: 114.2969
clearsPerMove: 2.0547
revealsPerMove: 1.1494
censored: 0
illegal: 0
incomplete: 0
wallSeconds: 11302.7039

paired

candidate-vs-fair-d3s7

meanDelta: -106964.3750
bootstrapLower95: -146579.5391
bootstrapUpper95: -69983.3727
studentTLower95: -145890.1465
pairedSd: 186537.5328
detectionFloor: 38356.7802
wtl: 14
0
50
halves: -109138.6563
-104790.0938
q25Delta: -36,134
movesDelta: -29.5313

init-vs-fair-d3s7

meanDelta: -142339.6250
bootstrapLower95: -178491.5461
bootstrapUpper95: -107770.6313
studentTLower95: -178510.6720
pairedSd: 173336.5226
detectionFloor: 35642.3225
wtl: 9
0
55
halves: -156142.6563
-128536.5938
q25Delta: -50,964
movesDelta: -39.2813

candidate-vs-init

meanDelta: 35375.2500
bootstrapLower95: 16898.9273
bootstrapUpper95: 54629.2750
studentTLower95: 16144.9914
pairedSd: 92153.9860
detectionFloor: 18949.1634
wtl: 36
0
28
halves: 47,004
23746.5000
q25Delta: 14,830
movesDelta: 9.7500

fair-d4s7-vs-fair-d3s7

meanDelta: 97058.7188
bootstrapLower95: 29329.5438
bootstrapUpper95: 165453.2305
studentTLower95: 27011.3702
pairedSd: 335676.3163
detectionFloor: 69023.4425
wtl: 41
0
23
halves: 75457.7813
118659.6563
q25Delta: 40137.5000
movesDelta: 26.2656

training

generationsCompleted

tenGenerationBlockAverages

generations
0-9
populationMean
159340.4412
top4Mean
176605.0742
best
183544.7219
meanMinusInit
4682.4068
bestMinusInit
28886.6875
meanMinusFair
-172846.5150
bestMinusFair
-148642.2344
generations
10-19
populationMean
168372.4929
top4Mean
189184.1141
best
194671.8250
meanMinusInit
15852.6179
bestMinusInit
42151.9500
meanMinusFair
-154441.2915
bestMinusFair
-128141.9594
generations
20-29
populationMean
174464.3515
top4Mean
197282.4398
best
203909.7531
meanMinusInit
22703.5952
bestMinusInit
52148.9969
meanMinusFair
-155743.1267
bestMinusFair
-126297.7250
generations
30-39
populationMean
181219.7211
top4Mean
203060.9922
best
209035.9844
meanMinusInit
32277.6086
bestMinusInit
60093.8719
meanMinusFair
-125175.9352
bestMinusFair
-97359.6719
generations
40-49
populationMean
184679.3310
top4Mean
206650.3125
best
213143.7469
meanMinusInit
34036.9685
bestMinusInit
62501.3844
meanMinusFair
-160026.6253
bestMinusFair
-131562.2094
generations
50-59
populationMean
190907.8234
top4Mean
216015.6836
best
227155.7656
meanMinusInit
40431.1578
bestMinusInit
76679.1000
meanMinusFair
-121896.0891
bestMinusFair
-85648.1469

trainingSignalCheck

rule: population mean fitness > paired fair-d3s7 control in the majority of the final 10 completed generations (theory falsifier 2)
window: 50
51
52
53
54
55
56
57
58
59
generationsMeanAboveFair: 0
generationsTop4AboveFair: 0
generationsBestAboveFair: 0
generationsInitAboveFair: 0
meanMarginLast10: -121896.0891
top4MarginLast10: -96788.2289
passed: false

artifactIntegrity

generationArtifacts: 60
illegalDecisions: 0
incompleteDecisions: 0
censoredGames: 0

config

population: 32
games: 32
generations: 60
elites: 4
tournament: 3
sigmaRel: 0.0500
seed: 0x0e701e58
leaseStart: 0xa52e0500
moveCap: 2,000
wallSeconds: 21,600

selection

finalGeneration: 59
finalists: 12
1
6
29
30
23
13
4
blockStart: 0xa52e0c80
selectGames: 128
winner: candidate-29
winnerMean: 202236.7734

corpus

games

177

roots

21,618

censoredGames

score

n: 177
mean: 433131.1525
sd: 323856.2717
min: 103,368
q25: 229,522
median: 348,847
q75: 549,029
max: 1,889,608

moves

n: 177
mean: 122.1356
sd: 85.7326
min: 34
q25: 70
median: 100
q75: 150
max: 500

wallSecondsPerGame

n: 177
mean: 9461.8629
sd: 7988.6131
min: 1209.2638
q25: 4178.7932
median: 7476.2450
q75: 11333.4374
max: 49497.1916

secondsPerRoot

77.4702

rootsPerGame

122.1356

legalColumnsPerRoot

n: 21,618
mean: 6.8572
sd: 0.6843
min: 1
q25: 7
median: 7
q75: 7
max: 7

teacherSiblingSpreadPoints

n: 21,618
mean: 40559.6546
sd: 160027.6823
min: 0
q25: 2448.7576
median: 4187.9510
q75: 7480.5930
max: 1017575.6843

pretrain

report

roots: 21,618
games: 177
trainRoots: 17,641
valRoots: 3,977
epochs: 16
batch: 64
lr: 0.0003
seed: 0x0e701e57
bestEpoch: 8
bestValHuber: 0.6873

probe

probeRoots: 256
top1: 0.4414
meanTeacherValueRegret: 2394.6708

epochs

epoch
0
trainHuber
2.3016
valHuber
1.8859
valPearson
0.8790
epoch
1
trainHuber
1.2819
valHuber
0.9546
valPearson
0.9314
epoch
2
trainHuber
0.7173
valHuber
0.8128
valPearson
0.9449
epoch
3
trainHuber
0.5488
valHuber
0.7449
valPearson
0.9491
epoch
4
trainHuber
0.4514
valHuber
0.7100
valPearson
0.9512
epoch
5
trainHuber
0.3886
valHuber
0.6918
valPearson
0.9520
epoch
6
trainHuber
0.3371
valHuber
0.6906
valPearson
0.9519
epoch
7
trainHuber
0.3036
valHuber
0.6919
valPearson
0.9500
epoch
8
trainHuber
0.2706
valHuber
0.6873
valPearson
0.9504
epoch
9
trainHuber
0.2443
valHuber
0.6939
valPearson
0.9502
epoch
10
trainHuber
0.2246
valHuber
0.6993
valPearson
0.9479
epoch
11
trainHuber
0.2092
valHuber
0.6955
valPearson
0.9496
epoch
12
trainHuber
0.1915
valHuber
0.7119
valPearson
0.9480
epoch
13
trainHuber
0.1807
valHuber
0.6936
valPearson
0.9479
epoch
14
trainHuber
0.1729
valHuber
0.6995
valPearson
0.9485
epoch
15
trainHuber
0.1620
valHuber
0.6969
valPearson
0.9470

resources

label
corpus
command
1. /home/keshav/Developer/drop7-bench/approaches/lifetime-objective/nnue-evolution/target/release/teacher_corpus
2. --seeds-start
3. 0xa52e0300
4. --games
5. 512
6. --threads
7. 32
8. --depth
9. 5
10. --move-cap
11. 500
12. --wall-seconds
13. 46800
14. --out
15. /home/keshav/Developer/drop7-bench/runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/corpus
startedAt
2026-09-02T04:24:10Z
endedAt
2026-09-02T22:39:58Z
wallSeconds
65747.5370
userSeconds
1667946.0280
systemSeconds
641.5270
peakRssBytes
1,615,892,480
exitCode
0
label
pretrain
command
1. /home/keshav/Developer/drop7-bench/approaches/lifetime-objective/nnue-evolution/target/release/pretrain
2. --corpus
3. /home/keshav/Developer/drop7-bench/runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/corpus/corpus.jsonl
4. --out
5. /home/keshav/Developer/drop7-bench/runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/pretrain
6. --epochs
7. 16
8. --batch
9. 64
10. --lr
11. 3e-4
12. --seed
13. 0x0e701e57
14. --probe-roots
15. 256
16. --threads
17. 32
startedAt
2026-09-02T22:42:11Z
endedAt
2026-09-02T22:42:25Z
wallSeconds
13.9780
userSeconds
45.2200
systemSeconds
0.1080
peakRssBytes
265,060,352
exitCode
0
label
evolve
command
1. /home/keshav/Developer/drop7-bench/approaches/lifetime-objective/nnue-evolution/target/release/evolve
2. --init
3. /home/keshav/Developer/drop7-bench/runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/pretrain/init.bin
4. --lease-start
5. 0xa52e0500
6. --out
7. /home/keshav/Developer/drop7-bench/runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/evolve
8. --population
9. 32
10. --games
11. 32
12. --generations
13. 60
14. --elites
15. 4
16. --tournament
17. 3
18. --sigma-rel
19. 0.05
20. --seed
21. 0x0e701e58
22. --threads
23. 32
24. --wall-seconds
25. 21600
26. --move-cap
27. 2000
startedAt
2026-09-02T22:42:25Z
endedAt
2026-09-03T02:37:19Z
wallSeconds
14093.6190
userSeconds
441673.1460
systemSeconds
184.6580
peakRssBytes
477,446,144
exitCode
0
label
select
command
1. /home/keshav/Developer/drop7-bench/approaches/lifetime-objective/nnue-evolution/target/release/evolve
2. --select
3. --lease-start
4. 0xa52e0500
5. --out
6. /home/keshav/Developer/drop7-bench/runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/evolve
7. --generations
8. 60
9. --games
10. 32
11. --select-games
12. 128
13. --threads
14. 32
15. --move-cap
16. 2000
startedAt
2026-09-03T02:37:19Z
endedAt
2026-09-03T02:41:29Z
wallSeconds
250.4040
userSeconds
7852.4190
systemSeconds
3.9940
peakRssBytes
252,526,592
exitCode
0
label
screen
command
1. /home/keshav/Developer/drop7-bench/approaches/lifetime-objective/nnue-evolution/target/release/screen
2. --candidate
3. /home/keshav/Developer/drop7-bench/runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/evolve/candidate-weights.bin
4. --init
5. /home/keshav/Developer/drop7-bench/runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/pretrain/init.bin
6. --seeds-start
7. 0xa52e1300
8. --games
9. 64
10. --threads
11. 32
12. --move-cap
13. 2000
14. --out
15. /home/keshav/Developer/drop7-bench/runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/screen/heldout.json
startedAt
2026-09-03T02:41:29Z
endedAt
2026-09-03T02:52:22Z
wallSeconds
652.9870
userSeconds
12346.6780
systemSeconds
3.8980
peakRssBytes
1,792,266,240
exitCode
0

Limitations

Single 64-game held-out screen: the paired detection floor is about 38,000 points for the primary contrast and 19,000 for the ablation; the primary result is far outside its floor, the ablation clears its own.
The teacher corpus reached 177 of the 512 games the protocol allowed for because the depth-5 teacher ran four to seven times slower than the pilot projected; the protocol makes the completed games the corpus, so the result is valid, but it rejects this configuration at this corpus size, not the design at 512 games.
The supervised warm start plays at about half the fair leaf's level; the deployment-faithful probe (0.441 top-1) and the leaf-swing diagnostic (zero leaf 0.450 on a different sample) suggest imitation of state values barely improved the search's ordering over no leaf at all. Evolution then had roughly 150,000 paired points to make up in 60 generations and made up about 35,000 to 40,000.
The per-game artifact holds four arms x 64 games (256 rows); the primary contrast pairs the candidate and fair-d3s7 rows by seed.
The fair-d4s7 arm is diagnostic only and reproduces the standing depth-4-over-depth-3 result on these seeds; it is not part of the gate.
Two operational faults during the unattended chain (a false liveness reading, then a script replacement that crashed the corpus stage's shell after its binary had exited 0) are recorded in the run record; the 'corpus: done' marker in pipeline.log was appended by hand with a note. No artifact was affected.
The elite re-selection block (0xa52e0c80, 128 games) and every fitness block are training-lease seeds; the last generation's 32-game leaders read 20,000 to 60,000 above their 128-game re-selection means, which is the best-of-32 bias the re-selection exists to remove.

Open the result record

Agent contextHow to extend this record

To add a reader-facing explanation, write web/content/research/EX-20260902-nnue-evolution-d3-v2-49c18bc2.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-20260902-nnue-evolution-d3-v2-49c18bc2.json, validated against research/schemas/experiment-v1.schema.json. Protocol hash: 66d163bb2d39cf4da157a1044be766eb124af28c158a7750e2064ad61edda942.

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