Continued whole-game evolution of the depth-5-distilled NNUE leaf from the first run's generation-60 population, with annealed mutation and a preregistered plateau stop (up to 1,000 generations)

Successor experiment to EX-20260902-nnue-evolution-d3-v2-49c18bc2 (result RS-20260903T025751Z-6577b33e: the 60-generation candidate lost its held-out screen to the frozen fair leaf by -106,964 paired points while beating its own unevolved warm start by +35,375, and the paired margin over the fair control was still rising at generation 60: OLS slope over the 60 generations +676 points per generation, one-sided 95% lower bound +200).

completedtier: SCREENalgorithmicpublic-developmentpublic-policyEX-20260903-nnue-evolution-continuation-d3-f8ce9181

On this page

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

Theories

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

Results

RS-20260903T163321Z-733076b5

Dates

Created: 2026-09-03T03:24:30Z
Updated: 2026-09-03T16:33:21Z

Record idEX-20260903-nnue-evolution-continuation-d3-f8ce9181

No explanation has been written for this record yet.

Technical recordThe registered protocolEX-20260903-nnue-evolution-continuation-d3-f8ce9181

Hypothesis: Successor experiment to EX-20260902-nnue-evolution-d3-v2-49c18bc2 (result RS-20260903T025751Z-6577b33e: the 60-generation candidate lost its held-out screen to the frozen fair leaf by -106,964 paired points while beating its own unevolved warm start by +35,375, and the paired margin over the fair control was still rising at generation 60: OLS slope over the 60 generations +676 points per generation, one-sided 95% lower bound +200). Question: with a much longer budget and an annealed mutation size, does whole-game evolution with common random numbers keep improving the depth-3 NNUE leaf, where does the improvement plateau, and does the plateau candidate beat the frozen fair leaf on never-read games? Candidate: the identical deployed configuration (stock fair expectimax, depth 3, seven chance strata, terminal utility -1,000,000, policy seed 0xd7075eed, completion-guaranteeing work bound, 64k-entry table, NNUE leaf 8,902-feature sparse class, 135 active, EmbeddingBag(8902,64)->ReLU->32->ReLU->1, x17,000 points). Starting population: the 32 vectors of runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/evolve/population-060.bin, the population the first run's generation-59 selection produced and never evaluated (SHA-256 3da5021531898f08254d35cad5486fbbc12092770216bcdaea988d3724422cb8); generation 0 of this run plays it exactly as saved (no re-noising). GA per generation, unchanged from the first run except where stated: population 32; every candidate plays the same fresh block of 32 training seeds, generation g playing [0xa52e2000 + 32 g, +32); fitness = mean final score of the 32 complete games (2,000-move cap); 4 elites cloned; the other 28 slots filled by tournaments of 3 whose winner is copied and mutated per tensor with Gaussian noise; evolution seed 0x0e701e59 (fresh, so no mutation realisation of the first run repeats). Annealed mutation (new): sigma_rel(g) = max(0.01, 0.05 x exp(-g / 400)), i.e. equal to the first run's constant 0.05 at g = 0, 0.030 at g = 200, 0.018 at g = 400, and the 0.01 floor from g = 644 onward (per-tensor sigma = sigma_rel x tensor std, floor 1e-4, as before). Three paired controls play every block and are never selected: the frozen fair leaf (control-fair-d3s7), the first run's supervised warm start (artifacts/results/EX-20260902-nnue-evolution-d3-v2-49c18bc2/RUN-20260902T035644Z-c1fd8987/pretrain/init.bin, SHA-256 7d6fa5563de6716a4a895fb224bbe7f81c568391b6b737f85acc110872c075e0; control-init-d3s7) and the first run's frozen screened candidate (artifacts/results/EX-20260902-nnue-evolution-d3-v2-49c18bc2/RUN-20260902T035644Z-c1fd8987/evolve/candidate-weights.bin, SHA-256 edd0d2efd181de43f35d62c4df784cb2c789db3ae296be93a1c6c59082034a9f; control-baseline-d3s7), the last being the direct paired measure of improvement over where this run started. Preregistered plateau stop (new; the owner's request was to keep evolving as long as the last 50-100 generations show noticeable improvement): the tracked statistic is the paired margin m_g = population mean fitness minus the fair control on generation g's block. After every 50th completed generation from the 100th onward, ordinary least squares regresses m over generation index across the last 100 generations; the run stops when the slope's one-sided 95% lower bound (slope - 1.645 x standard error, residual-based, n-2 degrees of freedom) is not greater than zero, i.e. when 100 generations show no detectable improvement. Every check is appended to plateau.jsonl and a stop writes the PLATEAU marker. Sensitivity, from the first run's artifacts: the residual scatter of m_g about its trend was 38,813 points, at which a 100-generation window detects slopes of about 221 points per generation (22,100 per 100 generations); the first run's own slope of 676 would not have stopped it. Other stops: 1,000 generations; a 259,200 s (72 h) evolve wall budget; a STOP file. Final candidate: the top 8 of the last completed generation each replay the 128 fresh training seeds immediately after the last fitness block played (so an early stop still re-selects on fresh seeds inside the lease); the highest mean is frozen as candidate-weights.bin and its SHA-256 recorded before the screen lease opens. Held-out screen, once: 64 never-read public-development seeds 0xa52ea000-0xa52ea03f, five arms on identical seeds: candidate, baseline-run1 (the first run's frozen candidate), init-d3s7 (the first run's warm start), fair-d3s7 (the primary comparator, identical search) and fair-d4s7 (the program's standing reference, diagnostic only). Primary contrast candidate minus fair-d3s7 as in the first run's gate; preregistered secondary contrast candidate minus baseline-run1, the out-of-sample measure of what the continuation added. Operational: 32 threads (the first run's SMT preflight), every stage through scripts/pipeline.sh (env: EXPERIMENT_ID, LEASE_START 0xa52e2000, SCREEN_START 0xa52ea000, GENERATIONS 1000, EVOLVE_SEED 0x0e701e59, RESUME_POP, BASELINE, INIT, SIGMA_TAU 400, SIGMA_FLOOR 0.01, PLATEAU_WINDOW 100, PLATEAU_EVERY 50, PLATEAU_MIN 100, EVOLVE_WALL 259200) with per-stage rusage; the evolve binary's checkpoint contract (population-{g+1}.bin before gen-{g}.json) makes the run resumable after a kill without replaying a block. The first run's chain faults are addressed: the driver is not edited while a stage runs.
Arms: Arm Name Entry point Manifest
Candidate d3s7-evolved-nnue-leaf-continued 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, continued candidate minus frozen fair leaf, both at the identical d3s7 configuration, 64 held-out games
Secondary metrics: paired mean score delta, continued candidate minus the first run's frozen candidate (baseline-run1), same 64 games: what the continuation added out of sample
paired mean moves delta; numbered clears per move and cover reveals per move; lower quartile, median, maximum; W-T-L; first-half and second-half deltas, for every arm pair
training curve: per-generation population mean, top-4 mean, best, and the three paired controls; the plateau statistic m_g and every plateau check (slope, standard error, lower bound); the generation at which the run stopped and why
ten-generation block averages of population mean minus each control, continuing the first run's table
reference arm fair-d4s7 vs fair-d3s7, diagnostic only
Statistical unit: whole-game
Uncertainty method: one-sided 95% percentile bootstrap over whole games, 20,000 resamples, RNG seed 0xb0071eaf (the unchanged compare.py), plus a one-sided 95% Student-t lower bound; detection floor 1.645*sd/sqrt(n) reported; plateau rule as stated in the hypothesis
Data role: public-development
Seed leases: SL-20260903T033000Z-a52e2000
SL-20260903T033000Z-a52ea000
Whole-origin split: yes
Reuse disclosure: The starting population, the warm-start control and the baseline control are products of EX-20260902-nnue-evolution-d3-v2-49c18bc2's training data (lease SL-20260825T063000Z-a52e0300), already read; no seed of that lease is read again. CHECK gates for the rebuilt binaries and a 3-generation smoke test of the new flags (population 32, 2 games, 60-move cap) read only the already-opened development probe block 0xa5277000-0xa5277107. Fitness blocks and the re-selection open training-lease seeds 0xa52e2000-0xa52ea000 (1,000 blocks of 32 = 32,000 seeds plus 128 re-selection seeds; the 640 seeds above 32,128 stay unopened); the screen opens 0xa52ea000-0xa52ea03f exactly once, after the candidate is frozen. Neither range contains any 8-hex-digit constant present in docs/, approaches/, src/, research/, artifacts/ or web/content/ (checked 2026-09-03), and both are disjoint from every existing lease.
Pass criteria: All 11 CHECK gates passed on the rebuilt binaries before the first leased seed was read (gates.log in the run directory).
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 0xa52ea000-0xa52ea03f, 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.
Reported alongside the gate, not part of it: candidate vs baseline-run1 bootstrap 95% lower bound > 0 establishes that the continuation improved the leaf out of sample; the generation of the stop and the plateau statistics establish where the training curve levelled.
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 configuration (this model class, this population and games-per-candidate budget, this annealing schedule, deployment depth 3), reporting separately whether the continuation beat the first run's candidate and at what generation the curve plateaued; do not adopt the vector; open no further cohort for it.
Gate fixed before controlled data: yes
Resources: Wall seconds 275000
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.
Evolution stops at the preregistered plateau rule (one-sided 95% lower bound of the OLS slope of the paired margin over the fair control across the last 100 generations <= 0, checked after every 50th completed generation from the 100th), at 1,000 generations, at 259,200 s of evolve wall time, or when a STOP file appears; the candidate is then the elite re-selection over the last completed generation's population on the 128 seeds following the last fitness block.
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.
The run may be killed and resumed at a generation boundary (checkpoint contract); a resume replays no seed and is recorded in the run record.
Expected artifacts: runs/<run-id>/nnue-evolution/gates.log (CHECK, rebuilt binaries)
runs/<run-id>/nnue-evolution/evolve/{config.json (with resumePopulationSha256 and baselineSha256),progress.jsonl (per generation: best, mean, controlFair, controlInit, controlBaseline, sigmaRel, fitness[]),plateau.jsonl,PLATEAU (if the rule stopped the run),gen-*.json,population-*.bin,final-fitness.json,candidate-weights.bin,candidate-weights.sha256,selection.json}
runs/<run-id>/nnue-evolution/screen/heldout.json (five arms) + compare-{candidate-vs-fair-d3s7,candidate-vs-baseline-run1,baseline-run1-vs-fair-d3s7,init-vs-fair-d3s7,candidate-vs-init,fair-d4s7-vs-fair-d3s7}.json
runs/<run-id>/nnue-evolution/{pipeline.log,rusage.jsonl,evolve.log,select.log,screen.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
2026-09-03T16:33:21Z no lifecycle advanced preregistered -> completed after run RUN-20260903T032832Z-a76a6cf7 and result RS-20260903T163321Z-733076b5 were written; protocol content otherwise unchanged. The preregistration hash 9a094a0a9e39eac372bfa7366ffadec440e4e68b207299d613a10c9cad99466a is retained in the run and lease records.

Arm	Name	Entry point	Manifest
Candidate	d3s7-evolved-nnue-leaf-continued	`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	275000
CPU threads	32
Max host bytes	17179869184
Max GPU bytes	–
GPU devices	–

Timestamp	Before controlled data	Reason
2026-09-03T16:33:21Z	no	lifecycle advanced preregistered -> completed after run RUN-20260903T032832Z-a76a6cf7 and result RS-20260903T163321Z-733076b5 were written; protocol content otherwise unchanged. The preregistration hash 9a094a0a9e39eac372bfa7366ffadec440e4e68b207299d613a10c9cad99466a is retained in the run and lease records.

Technical recordResults recorded against this protocol1 record

valid runoutcome: failnot-supported-as-testedtier: public-developmentRS-20260903T163321Z-733076b5

Valid run of the frozen protocol EX-20260903-nnue-evolution-continuation-d3-f8ce9181 (successor to RS-20260903T025751Z-6577b33e), every stage completed, every artifact with zero illegal and zero incomplete decisions. Stage C resumed from the first run's checkpointed generation-60 population (SHA-256 3da5021531898f08254d35cad5486fbbc12092770216bcdaea988d3724422cb8) with the first run's frozen candidate (SHA-256 edd0d2efd181de43f35d62c4df784cb2c789db3ae296be93a1c6c59082034a9f) as a third paired control and an annealed mutation size (sigma_rel from 0.05, time constant 400 generations, floor 0.01; it reached 0.0345 at the last generation). The preregistered plateau rule stopped the run after generation 149: the first check after generation 99 found the paired margin over the fair control rising at +238.6 points per generation (standard error 127.1, one-sided 95% lower bound +29.6) and continued; the second, after generation 149, found +37.3 per generation (standard error 119.3, lower bound -158.9), no detectable improvement over generations 50-149, and stopped. Fifty-generation averages of the population mean: 204,748, 228,706, 237,946; paired margin over the first run's candidate +8,982, +36,076, +45,613; over the fair control -116,503, -99,833, -88,948. The best candidate beat the fair control on 9 of 150 blocks (generations 59, 66, 73, 97, 100, 109, 115, 125, 137); the population mean beat it on none, so the theory's training-signal falsifier fails again (mean above fair in 0 of the last 10, margin -98,026). Elite re-selection on 128 fresh games (0xa52e32c0) froze candidate-12 at 264,466 (finalists 228,212 to 264,466). Stage D, 64 never-read games (0xa52ea000, opened once at 2026-09-03T16:21:00Z after the candidate's SHA-256 was recorded), five arms: the continued candidate averaged 251,667 against the frozen fair leaf's 320,108 at the identical depth-3 configuration, paired -68,441 (bootstrap 95% bounds -112,090 to -26,694, Student-t lower bound -112,274, floor 43,193), W-T-L 25-0-39, halves -112,383 / -24,500, lower quartile 171,440 against 189,414: every screen criterion fails, scientific outcome fail for this configuration. The preregistered secondary contrast answers the continuation's own question: the continued candidate beat the first run's frozen candidate on the same seeds by +36,278 (bootstrap lower bound +9,085, Student-t lower bound +8,543, floor 27,330, W-T-L 41-0-23, halves -7,140 / +79,697), so 150 further generations improved the leaf out of sample, by about a third of the remaining distance. The first run's candidate reproduced its earlier result on fresh seeds: -104,719 against the fair leaf here (lower bound -143,479) against -106,964 on the first screen. Over the warm start the continued candidate is +107,994 (lower bound +82,769, W-T-L 54-0-10). The reference arm gave fair depth 4 over fair depth 3 +71,799 (lower bound +25,857). Read: whole-game evolution keeps improving the leaf until roughly generation 100 of the continuation and then levels off about 70,000 paired points short of the frozen fair leaf on held-out games; the claim is not supported at this budget, and the plateau is the new fact. Whether the levelling is a property of the leaf class or of the annealed step size cannot be separated in this design.

What it had to pass

✓All 11 CHECK gates passed on the rebuilt binaries before the first leased seed was read — observed: ALL GATES PASSED (gates.log), 13/13 unit tests, smoke test of every new flag on the already-open probe block
✓Every generation artifact and the screen artifact: illegalDecisions 0, incompleteDecisions 0, censored games reported — observed: 150 generation artifacts: 0 illegal, 0 incomplete, 0 censored; screen: 0 illegal, 0 incomplete, 0 censored in all five 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 -68,441; bootstrap LB -112,090, UB -26,694; t LB -112,274; paired sd 210,055; floor 43,193; W-T-L 25-0-39
✕Held-out screen: paired mean score delta > 0 in both halves — observed: first half -112,383, second half -24,500
✕Held-out screen: candidate lower-quartile score >= fair-d3s7 lower-quartile score — observed: candidate Q25 171,440 vs fair-d3s7 Q25 189,414 (delta -17,974)
✓The screened candidate is the exact frozen winner of the elite re-selection; SHA-256 recorded before the screen lease opened — observed: candidate-12 of generation 149, mean 264,466 over 128 fresh games at 0xa52e32c0; sha256 759084fab97599818d01a5da6d18b9cc77172249f09a68d5104361bca385e09b; lease opened 2026-09-03T16:21:00Z by the screen stage
✓Reported alongside the gate: candidate vs baseline-run1 bootstrap 95% lower bound > 0 (the continuation improved the leaf out of sample) — observed: delta +36,278; bootstrap LB +9,085, UB +63,768; t LB +8,543; floor 27,330; W-T-L 41-0-23; halves -7,140 / +79,697 (the first half alone is negative)
–Reported alongside the gate: the generation of the stop and the plateau statistics — observed: stopped by the plateau rule after generation 149 (150 completed): check after generation 99 slope +238.6/gen, se 127.1, LB +29.6 (continue); after generation 149 slope +37.3/gen, se 119.3, LB -158.9 (stop); window means -99,833 then -88,948
✕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, best in 0/10; mean margin -98,026; best candidate above fair on 9 of 150 blocks

Technical recordRecorded metricsRS-20260903T163321Z-733076b5

cohort

seedsStartHex: 0xa52ea000
games: 64
moveCap: 2,000
role: public-development, never read before this screen (lease SL-20260903T033000Z-a52ea000, opened once)

arms

candidate

games: 64
mean: 251667.3906
median: 224340.5000
q25: 171439.7500
max: 571,367
movesMean: 75.3594
clearsPerMove: 1.8868
revealsPerMove: 1.0404
censored: 0
illegal: 0
incomplete: 0
wallSeconds: 576.4845

baseline-run1

games: 64
mean: 215389.0156
median: 185101.5000
q25: 138371.7500
max: 505,061
movesMean: 65.3281
clearsPerMove: 1.7898
revealsPerMove: 0.9589
censored: 0
illegal: 0
incomplete: 0
wallSeconds: 502.0007

init-d3s7

games: 64
mean: 143673.7031
median: 123,543
q25: 106999.5000
max: 341,054
movesMean: 45.3125
clearsPerMove: 1.5124
revealsPerMove: 0.7314
censored: 0
illegal: 0
incomplete: 0
wallSeconds: 331.6452

fair-d3s7

games: 64
mean: 320108.4844
median: 276826.5000
q25: 189,414
max: 940,703
movesMean: 94.1719
clearsPerMove: 1.9935
revealsPerMove: 1.1120
censored: 0
illegal: 0
incomplete: 0
wallSeconds: 272.6142

fair-d4s7

games: 64
mean: 391907.1719
median: 351,132
q25: 215822.7500
max: 962,024
movesMean: 113.2813
clearsPerMove: 2.0526
revealsPerMove: 1.1506
censored: 0
illegal: 0
incomplete: 0
wallSeconds: 11087.7038

paired

candidate-vs-fair-d3s7

meanDelta: -68441.0938
bootstrapLower95: -112089.9094
bootstrapUpper95: -26693.8359
studentTLower95: -112274.3615
pairedSd: 210054.9151
detectionFloor: 43192.5419
wtl: 25
0
39
halves: -112382.5625
-24499.6250
q25Delta: -17974.2500
movesDelta: -18.8125

init-vs-fair-d3s7

meanDelta: -176434.7813
bootstrapLower95: -213633.3297
bootstrapUpper95: -140506.4984
studentTLower95: -214034.1823
pairedSd: 180181.3872
detectionFloor: 37049.7977
wtl: 8
0
56
halves: -178178.6250
-174690.9375
q25Delta: -82414.5000
movesDelta: -48.8594

candidate-vs-init

meanDelta: 107993.6875
bootstrapLower95: 82769.3031
bootstrapUpper95: 133701.0930
studentTLower95: 82118.8189
pairedSd: 123995.8508
detectionFloor: 25496.6468
wtl: 54
0
10
halves: 65796.0625
150191.3125
q25Delta: 64440.2500
movesDelta: 30.0469

fair-d4s7-vs-fair-d3s7

meanDelta: 71798.6875
bootstrapLower95: 25857.3078
bootstrapUpper95: 119451.2141
studentTLower95: 23285.7939
pairedSd: 232480.3118
detectionFloor: 47803.7641
wtl: 39
0
25
halves: 31746.1563
111851.2188
q25Delta: 26408.7500
movesDelta: 19.1094

candidate-vs-baseline-run1

meanDelta: 36278.3750
bootstrapLower95: 9085.1570
bootstrapUpper95: 63768.1234
studentTLower95: 8543.2666
pairedSd: 132910.3700
detectionFloor: 27329.6948
wtl: 41
0
23
halves: -7139.9375
79696.6875
q25Delta: 33,068
movesDelta: 10.0313

baseline-run1-vs-fair-d3s7

meanDelta: -104719.4688
bootstrapLower95: -143478.7719
bootstrapUpper95: -67444.3852
studentTLower95: -143894.5037
pairedSd: 187732.0370
detectionFloor: 38602.4001
wtl: 20
0
44
halves: -105242.6250
-104196.3125
q25Delta: -51042.2500
movesDelta: -28.8438

training

generationsCompleted

150

stoppedOnPlateau

true

plateauChecks

generation
99
window
100
slopePerGeneration
238.6249
standardError
127.0788
lowerBound95
29.5988
windowMeanFirstHalf
-116503.0471
windowMeanSecondHalf
-99833.0190
stop
false
generation
149
window
100
slopePerGeneration
37.3143
standardError
119.2761
lowerBound95
-158.8775
windowMeanFirstHalf
-99833.0190
windowMeanSecondHalf
-88948.3871
stop
true

fiftyGenerationBlockAverages

generations
0-49
populationMean
204747.8623
best
242098.4525
meanMinusFair
-116503.0471
bestMinusFair
-79152.4569
meanMinusBaseline
8982.3529
bestMinusBaseline
46332.9431
sigmaAtEnd
0.0442
generations
50-99
populationMean
228705.8372
best
275125.1206
meanMinusFair
-99833.0190
bestMinusFair
-53413.7356
meanMinusBaseline
36076.1941
bestMinusBaseline
82495.4775
sigmaAtEnd
0.0390
generations
100-149
populationMean
237945.8410
best
285994.4269
meanMinusFair
-88948.3871
bestMinusFair
-40899.8012
meanMinusBaseline
45613.4210
bestMinusBaseline
93662.0069
sigmaAtEnd
0.0345

tenGenerationBlocks

generations
0-9
populationMean
198114.0884
top4Mean
221970.0750
best
229839.3250
meanMinusFair
-123398.6366
bestMinusFair
-91673.4000
meanMinusInit
49351.0696
meanMinusBaseline
-3548.5304
generations
10-19
populationMean
202549.9282
top4Mean
231257.9602
best
242869.1125
meanMinusFair
-138529.6343
bestMinusFair
-98210.4500
meanMinusInit
52363.4938
meanMinusBaseline
16067.6188
generations
20-29
populationMean
203141.2692
top4Mean
228807.5289
best
237100.1719
meanMinusFair
-92208.5808
bestMinusFair
-58249.6781
meanMinusInit
49689.4161
meanMinusBaseline
5352.8380
generations
30-39
populationMean
206259.3706
top4Mean
236366.9516
best
244948.4188
meanMinusFair
-118918.7013
bestMinusFair
-80229.6531
meanMinusInit
52443.9112
meanMinusBaseline
12523.5987
generations
40-49
populationMean
213674.6552
top4Mean
243235.7219
best
255735.2344
meanMinusFair
-109459.6823
bestMinusFair
-67399.1031
meanMinusInit
58796.8646
meanMinusBaseline
14516.2396
generations
50-59
populationMean
225622.1756
top4Mean
258711.7734
best
269100.2500
meanMinusFair
-86547.8275
bestMinusFair
-43069.7531
meanMinusInit
73783.6818
meanMinusBaseline
24289.6631
generations
60-69
populationMean
219346.4406
top4Mean
252135.9828
best
263647.7344
meanMinusFair
-106459.8656
bestMinusFair
-62158.5719
meanMinusInit
73687.9219
meanMinusBaseline
33851.7719
generations
70-79
populationMean
231187.5366
top4Mean
265508.5672
best
274636.9562
meanMinusFair
-85917.9884
bestMinusFair
-42468.5687
meanMinusInit
80509.0147
meanMinusBaseline
31926.8991
generations
80-89
populationMean
233843.3086
top4Mean
268375.5172
best
281134.5187
meanMinusFair
-101265.8602
bestMinusFair
-53974.6500
meanMinusInit
79471.4273
meanMinusBaseline
45444.1617
generations
90-99
populationMean
233529.7248
top4Mean
272108.8586
best
287106.1437
meanMinusFair
-118973.5533
bestMinusFair
-65397.1344
meanMinusInit
77933.4779
meanMinusBaseline
44868.4748
generations
100-109
populationMean
240531.0065
top4Mean
275976.7289
best
288242.7156
meanMinusFair
-78512.1810
bestMinusFair
-30800.4719
meanMinusInit
88217.0190
meanMinusBaseline
45300.6128
generations
110-119
populationMean
237897.3521
top4Mean
273658.6984
best
286345.5625
meanMinusFair
-83419.3511
bestMinusFair
-34971.1406
meanMinusInit
82886.6739
meanMinusBaseline
45329.8771
generations
120-129
populationMean
238849.4075
top4Mean
274787.2523
best
288351.2813
meanMinusFair
-93764.6144
bestMinusFair
-44262.7406
meanMinusInit
83077.8513
meanMinusBaseline
44751.7106
generations
130-139
populationMean
240464.9854
top4Mean
275556.8977
best
286090.1531
meanMinusFair
-91019.8646
bestMinusFair
-45394.6969
meanMinusInit
91141.3886
meanMinusBaseline
54606.0511
generations
140-149
populationMean
231986.4536
top4Mean
267803.5086
best
280942.4219
meanMinusFair
-98025.9245
bestMinusFair
-49069.9563
meanMinusInit
76973.6505
meanMinusBaseline
38078.8536

bestAboveFairGenerations

trainingSignalCheck

rule: population mean fitness > paired fair-d3s7 control in the majority of the final 10 completed generations (theory falsifier 2)
window: 140
141
142
143
144
145
146
147
148
149
generationsMeanAboveFair: 0
generationsTop4AboveFair: 0
generationsBestAboveFair: 0
generationsInitAboveFair: 0
meanMarginLast10: -98025.9245
top4MarginLast10: -62208.8695
passed: false

artifactIntegrity

generationArtifacts: 150
illegalDecisions: 0
incompleteDecisions: 0
censoredGames: 0

config

experiment: EX-20260903-nnue-evolution-continuation-d3-f8ce9181
population: 32
games: 32
generations: 1,000
elites: 4
tournament: 3
sigmaRel: 0.0500
sigmaDecayTau: 400
sigmaFloor: 0.0100
seed: 0x0e701e59
leaseStart: 0xa52e2000
moveCap: 2,000
wallSeconds: 259,200
resumePopulation: /home/keshav/Developer/drop7-bench/runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/evolve/population-060.bin
resumePopulationSha256: 3da5021531898f08254d35cad5486fbbc12092770216bcdaea988d3724422cb8
baseline: /home/keshav/Developer/drop7-bench/artifacts/results/EX-20260902-nnue-evolution-d3-v2-49c18bc2/RUN-20260902T035644Z-c1fd8987/evolve/candidate-weights.bin
baselineSha256: edd0d2efd181de43f35d62c4df784cb2c789db3ae296be93a1c6c59082034a9f
plateauWindow: 100
plateauCheckEvery: 50
plateauMinGenerations: 100

selection

finalGeneration: 149
finalists: 25
6
5
29
0
16
12
9
blockStart: 0xa52e32c0
selectGames: 128
winner: candidate-12
winnerMean: 264466.4688

resources

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/artifacts/results/EX-20260902-nnue-evolution-d3-v2-49c18bc2/RUN-20260902T035644Z-c1fd8987/pretrain/init.bin
4. --lease-start
5. 0xa52e2000
6. --out
7. /home/keshav/Developer/drop7-bench/runs/RUN-20260903T032832Z-a76a6cf7/nnue-evolution/evolve
8. --population
9. 32
10. --games
11. 32
12. --generations
13. 1000
14. --elites
15. 4
16. --tournament
17. 3
18. --sigma-rel
19. 0.05
20. --seed
21. 0x0e701e59
22. --threads
23. 32
24. --wall-seconds
25. 259200
26. --move-cap
27. 2000
28. --experiment-id
29. EX-20260903-nnue-evolution-continuation-d3-f8ce9181
30. --resume-population
31. /home/keshav/Developer/drop7-bench/runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/evolve/population-060.bin
32. --baseline
33. /home/keshav/Developer/drop7-bench/artifacts/results/EX-20260902-nnue-evolution-d3-v2-49c18bc2/RUN-20260902T035644Z-c1fd8987/evolve/candidate-weights.bin
34. --sigma-decay-tau
35. 400
36. --sigma-floor
37. 0.01
38. --plateau-window
39. 100
40. --plateau-check-every
41. 50
42. --plateau-min-generations
43. 100
startedAt
2026-09-03T03:34:50Z
endedAt
2026-09-03T16:15:39Z
wallSeconds
45649.2030
userSeconds
1431451.9050
systemSeconds
580.6870
peakRssBytes
479,240,192
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. 0xa52e2000
5. --out
6. /home/keshav/Developer/drop7-bench/runs/RUN-20260903T032832Z-a76a6cf7/nnue-evolution/evolve
7. --generations
8. 1000
9. --games
10. 32
11. --select-games
12. 128
13. --threads
14. 32
15. --move-cap
16. 2000
startedAt
2026-09-03T16:15:39Z
endedAt
2026-09-03T16:21:00Z
wallSeconds
320.7660
userSeconds
9947.1370
systemSeconds
7.3760
peakRssBytes
251,973,632
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-20260903T032832Z-a76a6cf7/nnue-evolution/evolve/candidate-weights.bin
4. --init
5. /home/keshav/Developer/drop7-bench/artifacts/results/EX-20260902-nnue-evolution-d3-v2-49c18bc2/RUN-20260902T035644Z-c1fd8987/pretrain/init.bin
6. --seeds-start
7. 0xa52ea000
8. --games
9. 64
10. --threads
11. 32
12. --move-cap
13. 2000
14. --out
15. /home/keshav/Developer/drop7-bench/runs/RUN-20260903T032832Z-a76a6cf7/nnue-evolution/screen/heldout.json
16. --experiment-id
17. EX-20260903-nnue-evolution-continuation-d3-f8ce9181
18. --arm
19. baseline-run1=/home/keshav/Developer/drop7-bench/artifacts/results/EX-20260902-nnue-evolution-d3-v2-49c18bc2/RUN-20260902T035644Z-c1fd8987/evolve/candidate-weights.bin
startedAt
2026-09-03T16:21:00Z
endedAt
2026-09-03T16:30:44Z
wallSeconds
584.0530
userSeconds
12716.1680
systemSeconds
7.1320
peakRssBytes
1,793,974,272
exitCode
0

firstRun

resultId: RS-20260903T025751Z-6577b33e
runId: RUN-20260902T035644Z-c1fd8987
screenCandidateVsFair: -106,964

Limitations

Single 64-game held-out screen: paired detection floors of about 43,000 points for the primary contrast and 27,000 for the continuation contrast; the primary result sits far outside its floor, the continuation contrast just outside its own, and its first-half estimate is negative, so 'improved out of sample' is established at the whole-block level only.
The plateau rule and the annealed mutation size are confounded: the step size had fallen from 0.050 to 0.034 when the rule fired, so the design cannot say whether the leaf class ran out of improvement or the search did. A constant-sigma continuation from the same population would separate them.
The plateau rule tests a fitted slope over 100 generations with a one-sided 95% lower bound; at the observed scatter it can miss real improvements below about 220 points per generation. The observed point estimate over generations 50-149 was +37 per generation.
The per-game artifact holds five arms x 64 games (320 rows); each contrast pairs two arms by seed.
The starting population, the warm-start control and the baseline control are products of the first run's training data; nothing in this run re-read that lease.
The owner's commit e2b0d19 (wall-budget pinning and progress recovery on resume) landed on the branch mid-run; the run never resumed, the driver shells held the original script by an unlinked inode, and every stage ran the binaries built from commit e4fd018 as the run record states.
The fair-d4s7 arm is diagnostic only. The first run's candidate reproduced its earlier screen result on these fresh seeds (-104,719 against -106,964), which is the closest thing to a replication the program has for that number.

Open the result record

Agent contextHow to extend this record

To add a reader-facing explanation, write web/content/research/EX-20260903-nnue-evolution-continuation-d3-f8ce9181.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-20260903-nnue-evolution-continuation-d3-f8ce9181.json, validated against research/schemas/experiment-v1.schema.json. Protocol hash: 28dc3ac0042c0755426d7fc00692a19c127c297580f15e8aab92e83406395ba2.

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