---
title: An evolved NNUE leaf, distilled from a depth-5 teacher
family: lifetime-objective
summary: A compact neural leaf evaluator for the depth-3 fair search, warm-started by distilling a depth-5 teacher's sibling-complete root values and then refined by an evolutionary search whose fitness is whole-game score itself.
status: rejected
evidence: ledger-recorded
reads: public
kind: strategy
technique: evolution
featured: true
---

## The problem

Every search on this site has the same shape: imagine each legal column,
imagine the next disc, imagine the reply, and so on for a few moves, then stop
and ask a [leaf evaluator](/learn/glossary) how good the board at the end of
each imagined line is. The points won along the way are known exactly; the
leaf guesses the rest. The frozen fair leaf is a sum of eighteen hand-written
terms.

How much the leaf matters was measured before any training signal was read.
On a frozen snapshot of thirty teacher games, every root was searched twice at
each depth, once with the fair leaf and once with a leaf that returns zero, so
the only difference between the two sets of column values is the leaf term.

<ArmTable
  caption="Leaf-swing diagnostic on a frozen 30-game snapshot of the corpus (run artifacts leaf-swing-d3.txt, leaf-swing-d4.txt, leaf-swing-d5.txt). Spreads are medians over roots of the difference between the best and worst legal column. The leaf's share is the median ratio of leaf-term spread to total value spread on roots with no losing column."
  columns={[
    { key: "depth", label: "search depth" },
    { key: "roots", label: "roots", numeric: true },
    { key: "share", label: "leaf's share of the column spread", numeric: true },
    { key: "intree", label: "in-tree score spread (points)", numeric: true },
    { key: "leaf", label: "leaf-term spread (points)", numeric: true },
    { key: "flip", label: "decisions that change without the leaf", numeric: true },
    { key: "agree", label: "agreement with the depth-5 teacher, fair leaf / zero leaf" },
  ]}
  rows={[
    { depth: "3 plies", roots: "400", share: "0.97", intree: "393", leaf: "4,033", flip: "55.5%", agree: "72.0% / 45.0%" },
    { depth: "4 plies", roots: "80", share: "0.95", intree: "643", leaf: "4,093", flip: "43.8%", agree: "73.8% / 47.5%" },
    { depth: "5 plies", roots: "8", share: "0.97", intree: "1,141", leaf: "3,813", flip: "50.0%", agree: "100.0% / 50.0%" },
  ]}
/>

At every depth the leaf term carries about nineteen twentieths of the spread
between columns, and removing it changes the chosen column at roughly half of
the roots. The search multiplies the leaf's influence, because the same
evaluator is applied at every frontier position of every line. A better leaf
is worth pursuing for the same reason a worse one is dangerous.

Two earlier lines of work had each failed in an instructive way. Networks
trained to imitate a search learned the positions the teacher visited and then
could not rank the six sibling moves it never played (the
[sibling trap](/learn/concepts/ranking-siblings)). And the one time a leaf was
tuned by evolution on whole-game score directly, the eighteen hand-written
weights of the fair leaf under CMA-ES, the optimiser followed selection noise
into a worse evaluator
([the first leaf evolution](/approach/lifetime-objective/leaf-evolution)).
Its post-mortem found that with 32 games per candidate and a paired standard
deviation near 200,000 points, fresh seeds every generation left selection
noise alone to steer.

## Proposed solution

Replace the formula, for the three-move search, with a small neural network of
the kind chess engines call an [NNUE](/learn/glossary), an "efficiently
updatable neural network" ([primer](/learn/techniques/nnue)). It is sparse
enough to evaluate in a microsecond, which a leaf called hundreds of thousands
of times per decision can afford, and expressive enough to encode board
patterns that eighteen linear terms cannot. The one on this page has about
572,000 weights, almost all of them in the first table.

<NnueSketch caption="The NNUE reads exactly what a player can see, turns it into 135 active features out of 8,902 possible ones, sums the matching rows of a weight table, passes the sum through two small layers, and returns one number in units of a row rise (17,000 points). Evolution changes the weights only." />

The design keeps what each earlier failure taught. Imitation is used only as a
warm start, from a corpus that records a depth-5 teacher's value for every
legal column at every position, so the sibling-coverage hole is closed at the
source. Evolution then refines the warm start on the deployed objective, the
final score of complete games, with the noise handled: every candidate in a
generation plays the same games ([common random numbers](/learn/glossary)),
and selection compares ranks. The network reads only the visible board, the
visible next disc, the moves until the next rise and whether the game is over,
and the depth-5 teacher reads the same public state.

<PairedSeedsSketch caption="Common random numbers. Because every candidate and both controls play exactly the same seeds in a generation, the difference between two candidates' mean scores is a difference in evaluators and owes nothing to the discs they were dealt. Selection then compares ranks, so one lucky game moves a candidate up a few places instead of dragging the whole generation." />

## How it works

<EvolutionPipeline caption="Four stages. The teacher, the corpus, the warm start, the evolution and the elite re-selection all read seeds from the training lease, which may be read again. The screen reads a separate block of seeds that nothing has ever opened, exactly once, after the candidate's weights are frozen and hashed." />

1. Teacher corpus (stage A). The frozen fair evaluator, driven by the
   [fast Rust bitboard engine](/approach/fair-expectimax/rust-engine), plays
   training games at depth 5 with seven chance strata. Five moves is one full
   row-rise cycle, so every rise consequence of a root move is inside the
   teacher's horizon. At every position the corpus records the board, the
   next disc, the moves until the rise, and the teacher's value for all seven
   columns; the search computes those anyway, so the corpus is
   sibling-complete for free. Each teacher game is capped at 500 moves, and
   the stage stops starting new games at a fixed wall-clock budget; whatever
   whole games have completed are the corpus.
2. Warm start (stage B). The network is fitted to the teacher's values with
   the Adam optimiser on a Huber loss measured in rise units, with a
   [whole-origin split](/learn/glossary) so that validation never sees a
   position from a game it trained on. Training runs for 16 epochs and the
   epoch with the lowest validation loss is frozen. A deployment-faithful
   probe then puts that network inside the real depth-3 search on held-out
   positions and asks how often the search picks the teacher's column.
3. Evolution (stage C). Generation 0 is the warm start plus 31 copies with
   Gaussian noise added to every weight. Each generation draws a fresh block
   of 32 seeds that no candidate has seen, and all 32 candidates play all 32
   seeds as the leaf of the depth-3, seven-stratum fair search with the
   ordinary 2,000-move cap; the frozen fair leaf and the unevolved warm start
   play the same seeds as two controls that are never selected. A candidate's
   fitness is the plain mean of its 32 final scores. The four fittest are
   copied unchanged; the other 28 slots are filled by tournaments of three
   whose winner is copied and has every weight perturbed by Gaussian noise at
   5% of its tensor's standard deviation. The next population is written to
   disk before the generation's artifact is committed, so a crash replays the
   generation from identical inputs.

<GenerationLoop caption="One generation. Thirty-two candidates play the same fresh block of thirty-two seeds as the leaf of the real depth-3 search; the four fittest are copied unchanged; the other twenty-eight slots are filled by tournaments of three whose winner is copied and mutated. The frozen fair leaf and the unevolved warm start play every block as controls and are never selected." />

4. Re-selection. After the last generation, the eight fittest candidates each
   replay a fresh 128-game block, and the one with the highest mean is frozen
   as the candidate. Its SHA-256 is recorded before the screen block is
   opened, so the vector that is screened is provably the one that was
   selected. The best of thirty-two noisy estimates is biased upward, and 128
   fresh games shrink that bias before anything is claimed.
5. Screen (stage D). The frozen candidate plays 64 games on a seed block
   nothing has ever opened, against three arms on the identical seeds: the
   same depth-3 search with the frozen fair leaf (the comparator), the same
   search with the unevolved warm start (the ablation, isolating what
   evolution added), and the fair leaf at depth 4 (the program's standing
   reference, for context). The screen is played exactly once, whatever it
   shows. The network's output at every step is one number per board, in
   units of a row rise, consumed by the search.

## What happened

<EvolutionStatus run="RUN-20260902T035644Z-c1fd8987" />

On the 64 never-read held-out games, the evolved leaf scored 106,964 points
less per game than the frozen fair leaf inside the same depth-3 search
(bootstrap 95% lower bound −146,580), so the preregistered gate fails and the
theory's claim is not supported as tested
([RS-20260903T025751Z-6577b33e](/results/RS-20260903T025751Z-6577b33e), run
valid, outcome fail, ledger-recorded). The same screen shows what evolution
did do: the evolved leaf beat its own unevolved warm start by 35,375 points on
the same seeds, a gain whose lower bound sits well above zero, from a warm
start that played at about half the fair leaf's level.

The stages tell the rest. The depth-5 teacher ran four to seven times slower
per position than the pilot had projected, so the corpus closed at about a
third of the games the protocol allowed for. The warm start fitted the
teacher's values closely and, placed inside the real depth-3 search, picked
the teacher's column less than half the time, about what a leaf that returns
zero manages. Evolution then separated the population from the warm start
steadily across all sixty generations without ever approaching the fair
control, and the curve had not flattened when the protocol stopped it.

<CorpusFigure run="RUN-20260902T035644Z-c1fd8987" caption="Stage A. Each dot is one complete teacher game, in the order the games finished; hover a dot for its seed, score, length and how long the depth-5 search took to play it. Games are heavy-tailed: most end in the low hundreds of thousands and a few run far longer. The teacher turned out four to seven times slower per position than the pilot had projected, so the stage reached its wall budget well short of the 512 games the protocol allowed for; the protocol makes the completed games the corpus and does not depend on the count." />

<PretrainFigure run="RUN-20260902T035644Z-c1fd8987" caption="Stage B. Training and validation loss by epoch, in rise units; the dashed line marks the epoch frozen as the warm start. The ordering probe below the chart is the number that matters for deployment: how often the real depth-3 search, with this network as its leaf, picks the column the depth-5 teacher picked on held-out positions." />

<EvolutionFigure run="RUN-20260902T035644Z-c1fd8987" caption="Stage C. Best candidate, top-four mean and population mean by generation, with the two controls that played the same seeds each generation. Because every line in a generation was measured on the same 32 games, the vertical distance between a line and the fair control is a paired difference. The shaded band is the whole population's range. The theory's training-signal falsifier asks whether the population mean sits above the fair control in most of the final ten generations." />

<SwarmFigure run="RUN-20260902T035644Z-c1fd8987" caption="The same generations with every candidate drawn. Selection pressure shows up as the cloud tightening and drifting; a population that stays centred on the pink control line has not moved away from the warm start." />

<ScreenFigure run="RUN-20260902T035644Z-c1fd8987" caption="Stage D, the primary contrast. One bar per held-out game: the evolved candidate's score minus the frozen fair leaf's score on the same seed, in seed order. The dashed line is the recorded paired mean and the vertical divider separates the two halves the gate checks separately." />

<ScreenGateTable run="RUN-20260902T035644Z-c1fd8987" />

Because the paired margin over the fair leaf was still rising at generation
60, a second, longer experiment was registered under the same theory,
[EX-20260903-nnue-evolution-continuation-d3-f8ce9181](/experiments/EX-20260903-nnue-evolution-continuation-d3-f8ce9181),
and launched on 2026-09-03 as run `RUN-20260903T032832Z-a76a6cf7`. It starts
from the population the first run's final selection produced and never
played, adds the first run's frozen candidate as a third paired control so the
curve shows directly how far the population has moved, anneals the mutation
size from the first run's setting toward a floor, and stops when a straight
line fitted to the paired margin over the fair leaf across the last hundred
generations no longer has a positive lower bound on its slope, checked every
fifty generations from the hundredth, or at 1,000 generations or 72 hours. It
plays fresh training and screen leases and ends with the same one-shot
64-game screen.

The plateau rule stopped evolution after generation 149. On a fresh 64-game
screen the continued leaf still lost to the fair leaf, and beat the first
run's frozen candidate
([RS-20260903T163321Z-733076b5](/results/RS-20260903T163321Z-733076b5), run
valid, outcome fail, ledger-recorded). The figures below are the
continuation's own artifacts.

<EvolutionStatus run="RUN-20260903T032832Z-a76a6cf7" continuation />

<EvolutionFigure run="RUN-20260903T032832Z-a76a6cf7" caption="Continuation, stage C. Generation 0 is the first run's generation-60 population. The violet dashed line is the first run's frozen candidate playing the same seeds each generation: the vertical distance between the population lines and it is a paired measure of what the continuation has added. Hover a generation for its sigma; the latest plateau check appears in the summary below the chart." />

<SwarmFigure run="RUN-20260903T032832Z-a76a6cf7" caption="Every candidate of the continuation in every generation. As sigma anneals, the cloud should tighten; whether it also keeps drifting upward relative to the amber fair-leaf ticks is the question the plateau rule answers." />

<ScreenFigure run="RUN-20260903T032832Z-a76a6cf7" caption="Continuation, stage D, the primary contrast on the new held-out block: continued candidate minus the fair leaf, one bar per game." />

<ScreenFigure run="RUN-20260903T032832Z-a76a6cf7" contrast="candidate-vs-baseline-run1" caption="The continuation's own question, out of sample: continued candidate minus the first run's frozen candidate on the same 64 never-read games." />

<ScreenGateTable run="RUN-20260903T032832Z-a76a6cf7" />

The second continuation resumed that plateau population with a slower decay
in mutation size, keeping the same search and stopping rule. The run stopped
at the same generation. On its fresh 64-game screen, the candidate still lost
to the fair leaf, and its gain over the immediately prior candidate was not
statistically distinguishable from zero
([RS-20260904T090620Z-e5731bf0](/results/RS-20260904T090620Z-e5731bf0), run
valid, outcome fail, ledger-recorded). This alternative schedule did not
produce a confirmed further gain.

<EvolutionStatus run="RUN-20260903T190433Z-a87fd7fc" continuation />

<EvolutionFigure run="RUN-20260903T190433Z-a87fd7fc" caption="Second continuation, stage C. Generation 0 is the first continuation's generation-150 population. The violet dashed line is that run's frozen candidate playing the same seeds each generation." />

<SwarmFigure run="RUN-20260903T190433Z-a87fd7fc" caption="Every candidate of the second continuation in every generation." />

<ScreenFigure run="RUN-20260903T190433Z-a87fd7fc" caption="Second continuation, stage D, the primary contrast on the new held-out block: candidate minus the fair leaf." />

<ScreenFigure run="RUN-20260903T190433Z-a87fd7fc" contrast="candidate-vs-baseline-run1" caption="The second continuation's own question, out of sample: this candidate minus the immediately prior candidate on the same 64 never-read games." />

<ScreenGateTable run="RUN-20260903T190433Z-a87fd7fc" />

## What we learned

Whole-game evolution with common random numbers moves a 572,000-weight leaf.
That is the finding the first, CMA-ES leaf evolution could not produce. The
population separated from its warm start steadily across all sixty
generations, the ten-generation paired margins over the warm start rising
from +4,682 to +40,431 for the population mean and reaching +76,679 for the
best candidate in the last block, and the held-out ablation confirmed it out
of sample at +35,375 on the 64 screen games
([RS-20260903T025751Z-6577b33e](/results/RS-20260903T025751Z-6577b33e)).

Imitation of state values is the weak link. The warm start matched the
teacher's numbers (validation Pearson 0.9504 at the frozen epoch) yet, placed
inside the real depth-3 search, picked the teacher's column on 44.14% of 256
held-out roots, and on the screen it played 142,340 paired points behind the
fair leaf; evolution closed 35,375 of that, about a quarter of the way (same
record). The teacher is the expensive part: a depth-5, seven-stratum search
cost 77.5 seconds per root at full parallelism, so the corpus closed at 177
games instead of 512, and whether a larger corpus would have produced a warm
start that holds the ordering is untested. The leaf decides: the screen agrees
with the leaf-swing diagnostic above, since swapping the leaf moved the mean
by more than 100,000 points in both directions.

The continuation kept working for roughly a hundred more generations and then
levelled. That plateau sits about 45,000 paired points above where the
continuation started and about 90,000 below the fair leaf on the training
blocks, and the screen agrees
([RS-20260903T163321Z-733076b5](/results/RS-20260903T163321Z-733076b5)). The
design could not separate whether the leaf class ran out of improvement or
the annealed step size, which had fallen from 0.050 to 0.034, starved the
search. The second continuation tested a slower decay and still stopped at
the same generation, with no confirmed further gain on its fresh screen
([RS-20260904T090620Z-e5731bf0](/results/RS-20260904T090620Z-e5731bf0)). It
tested one alternative schedule. More games per candidate or a larger
population remain open ways to ask whether selection can see the small
differences left as candidates converge. The other open question is whether
a warm start trained on the teacher's ordering, instead of its values, would
start close enough to the fair leaf for evolution to finish the job.

<AgentContext summary="Records and provenance">

- Theory
  [TH-20260825-evolved-nnue-leaf-d3-0f47e46c](/theories/TH-20260825-evolved-nnue-leaf-d3-0f47e46c);
  frozen protocol
  [EX-20260902-nnue-evolution-d3-v2-49c18bc2](/experiments/EX-20260902-nnue-evolution-d3-v2-49c18bc2)
  (successor to the frozen `EX-20260825-nnue-evolution-d3-bca7f330`, whose
  scientific protocol it carries verbatim plus the operational parameters that
  record left open); run `RUN-20260902T035644Z-c1fd8987`, started
  2026-09-02T04:24:10Z with 32 threads on the Ryzen AI MAX+ 395 workstation
  (`research/system-profiles/MACH-20260902T035644Z-f5e59b6e.json`). Result
  [RS-20260903T025751Z-6577b33e](/results/RS-20260903T025751Z-6577b33e)
  (valid run, scientific outcome fail, assessment not-supported-as-tested,
  evidence tier public-development); compact evidence promoted with a manifest
  under
  `artifacts/results/EX-20260902-nnue-evolution-d3-v2-49c18bc2/RUN-20260902T035644Z-c1fd8987/`,
  including the frozen candidate and the warm start weights.
- Seed leases: training `SL-20260825T063000Z-a52e0300` (opened at launch;
  teacher corpus from `0xa52e0300`, fitness blocks from `0xa52e0500`,
  generation `g` playing the 32 seeds from `0xa52e0500 + 32·g`, re-selection
  at `0xa52e0c80`) and screen `SL-20260825T063000Z-a52e1300` (64 seeds; opened
  only by the screen stage, at 2026-09-03T02:41:29Z, after the candidate's
  SHA-256 was recorded). Training and screen ranges are disjoint by
  construction.
- Search: fair expectimax, depth 3, seven stratified chance samples, terminal
  utility −1,000,000, policy seed `0xd7075eed`, completion-guaranteeing work
  bound, 64k-entry depth-gated table; the teacher is the same search at
  depth 5 with the frozen fair leaf. Network: 8,902 sparse features, 135
  active per position; EmbeddingBag(8,902 × 64) → ReLU → 32 → ReLU → 1,
  output scaled by 17,000 points; about 572k parameters.
- Stage parameters. A: 512 seeds requested, 500-move cap, new games stop at
  46,800 s of wall clock. B: Adam, learning rate 3e-4, batch 64, Huber δ = 1
  rise, 16 epochs, seed `0x0e701e57`, 256 probe roots. C: population 32, 60
  generations, 32 paired games per candidate per generation, 4 elites,
  tournament size 3, per-tensor σ = 5% of tensor standard deviation (floor
  1e-4), initial cloud 2σ, evolution seed `0x0e701e58`, 21,600 s sub-budget;
  top 8 re-evaluated on 128 fresh games. D: four arms on 64 seeds,
  `compare.py` with 20,000 bootstrap resamples at seed `0xb0071eaf`.
- Artifacts under `runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/`:
  `gates.log`, the preflight logs, `corpus/parts/*.jsonl` (one file per
  teacher game, resumable), `corpus.err` (one line per completed game),
  `pretrain/{init.bin,report.json,probe.json}`,
  `evolve/{config.json,progress.jsonl,gen-*.json,population-*.bin,selection.json,candidate-weights.bin,candidate-weights.sha256}`,
  `screen/{heldout.json,compare-*.json}`, `rusage.jsonl` (kernel resource
  usage per stage), `analysis.json` and `analysis.md`, and the leaf-swing
  diagnostic (`leaf-swing-d{3,4,5}.txt` with the frozen 30-game input under
  `leaf-swing-input/`). The page's figures read
  `web/content/figures/nnue-evolution/<run>.json`, written by
  `web/scripts/extract-nnue-evolution.ts` from the run's artifacts.
- Implementation: a std-only Rust crate
  (`approaches/lifetime-objective/nnue-evolution/`) with a path dependency on
  the proven engine; the reference crate is not modified. Five binaries:
  `gate` (the CHECK-tier mechanics suite), `teacher_corpus` (stage A),
  `pretrain` (stage B), `evolve` (stage C, checkpointed and resumable, with
  `--select` for the elite re-selection), and `screen` (stage D, emitting the
  population-artifact format the existing `compare.py` consumes unchanged).
  The stages are driven by `scripts/pipeline.sh` (`RUN_ID=… THREADS=32
  scripts/pipeline.sh corpus|pretrain|evolve|select|screen|compare|chain`),
  which wraps every stage in `scripts/with-rusage.py` so wall, CPU and peak
  memory land in `rusage.jsonl`, opens the screen lease through
  `scripts/open-screen-lease.py` only when the candidate hash file exists,
  refuses to run the screen twice, and passes the original total evolve
  allowance on every resume while the binary's durable deadline
  (`wall-budget.json`) prevents that allowance from starting over. A `STOP`
  file or the wall budget ends the loop at a generation boundary; a crash
  after the completion marker rebuilds any missing summary or plateau
  decision from the committed artifact before another seed can be read.
  `scripts/analyze.py --run <run>` summarises whatever artifacts exist into
  `analysis.json` and `analysis.md`.
- Continuation:
  [EX-20260903-nnue-evolution-continuation-d3-f8ce9181](/experiments/EX-20260903-nnue-evolution-continuation-d3-f8ce9181),
  run `RUN-20260903T032832Z-a76a6cf7`, started 2026-09-03T03:34:49Z. Result
  [RS-20260903T163321Z-733076b5](/results/RS-20260903T163321Z-733076b5)
  (valid run, scientific outcome fail, assessment not-supported-as-tested,
  evidence tier public-development); compact evidence promoted with a manifest
  under
  `artifacts/results/EX-20260903-nnue-evolution-continuation-d3-f8ce9181/RUN-20260903T032832Z-a76a6cf7/`,
  including the frozen candidate, the plateau log and the launch script.
  The plateau rule stopped the run after generation 149. Leases:
  training `SL-20260903T033000Z-a52e2000` (fitness blocks from `0xa52e2000`,
  generation `g` playing `[0xa52e2000 + 32·g, +32)`; the re-selection block
  follows the last fitness block played) and screen
  `SL-20260903T033000Z-a52ea000` (64 seeds), both checked against every
  constant in the repository. Generation 0 is
  `runs/RUN-20260902T035644Z-c1fd8987/nnue-evolution/evolve/population-060.bin`
  (SHA-256 `3da5021531898f08254d35cad5486fbbc12092770216bcdaea988d3724422cb8`)
  played exactly as saved, with no re-noising. Three paired controls, never
  selected: the frozen fair leaf, the first run's warm start (`init.bin`,
  SHA-256 `7d6fa5563de6716a4a895fb224bbe7f81c568391b6b737f85acc110872c075e0`),
  and the first run's frozen candidate (`candidate-weights.bin`, SHA-256
  `edd0d2efd181de43f35d62c4df784cb2c789db3ae296be93a1c6c59082034a9f`).
  Evolution seed `0x0e701e59`. Annealed mutation: relative sigma
  `max(0.01, 0.05 × exp(−g / 400))`, so 0.05 at generation 0, 0.030 at 200,
  0.018 at 400, and the 0.01 floor from generation 644 onward. Plateau rule:
  after every 50th completed generation from the 100th, ordinary least
  squares regresses the paired margin over the fair control across the last
  100 generations; the run stops when the slope's one-sided 95% lower bound
  (slope − 1.645 × standard error) is not greater than zero. From the first
  run's artifacts the residual scatter about the trend was 38,813 points, at
  which the window detects slopes of about 221 points per generation (22,100
  per 100 generations); the first run's own slope of +676 points per
  generation (one-sided 95% lower bound +200) would have kept it going. Other
  stops: 1,000 generations, a 259,200 s (72 h) evolve wall budget, a `STOP`
  file. Every check is appended to `plateau.jsonl`. Since the launch, a
  code-review pass repaired two interruption-only faults in the continuation
  driver (a completed generation could be stranded without its progress row
  or plateau marker, and a restart granted a fresh wall allowance); the
  launched run remains attributable to its recorded source commit.
- Second continuation:
  [EX-20260903-nnue-evolution-continuation2-d3-80eebad3](/experiments/EX-20260903-nnue-evolution-continuation2-d3-80eebad3),
  run `RUN-20260903T190433Z-a87fd7fc`, launched 2026-09-03. Result
  [RS-20260904T090620Z-e5731bf0](/results/RS-20260904T090620Z-e5731bf0)
  (valid run, scientific outcome fail); compact evidence promoted under
  `artifacts/results/EX-20260903-nnue-evolution-continuation2-d3-80eebad3/RUN-20260903T190433Z-a87fd7fc/`.
  Generation 0 is the first continuation's generation-150 population. The
  mutation-decay time constant is 1,500 generations instead of 400, a
  3.75-fold slowdown; sigma is still 0.045 at the equivalent point where the
  first continuation reached 0.034, and does not reach its floor within the
  1,000-generation cap. The search, population, fitness blocks, plateau
  window and checks remain unchanged. The immediately prior candidate plays
  every block as the third control; the screen artifact calls this arm
  `baseline-run1`, with its identity recorded in `evolve/config.json`'s
  `baselineSha256` field. Training and screen use fresh leases.
- How to continue from here: every generation's population is on disk, so a
  later experiment can resume the loop from any `population-NNN.bin`, start a
  new run from `candidate-weights.bin`, or change one operator at a time.
  Doing so is a new experiment: it needs its own record, its own training
  lease (the fitness blocks above are spent), and a fresh never-read screen
  block.

</AgentContext>

<TechnicalRecord summary="Full results table" meta="RS-20260903T025751Z-6577b33e">

Screen (stage D), 64 never-read games from `0xa52e1300`, 2,000-move cap,
corrected 17,000-point Hardcore scoring, zero censored, illegal or incomplete
decisions in every arm:

| Arm (depth-3 search unless stated) | Mean | Median | Lower quartile | Mean moves | Clears/move | Reveals/move |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| evolved candidate | 190,961 | 160,108 | 122,253 | 58.50 | 1.7131 | 0.9124 |
| unevolved warm start (ablation) | 155,586 | 140,282 | 107,423 | 48.75 | 1.5510 | 0.7554 |
| frozen fair leaf (comparator) | 297,926 | 255,000 | 158,387 | 88.03 | 1.9549 | 1.0722 |
| frozen fair leaf, depth 4 (reference arm) | 394,984 | 302,556 | 198,525 | 114.30 | 2.0547 | 1.1494 |

Paired contrasts, 20,000 bootstrap resamples:

| Contrast | Mean delta | Bootstrap 95% bounds | Student-t lower 95% | Detection floor | W-T-L | Halves |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| candidate − fair leaf | −106,964 | −146,580 to −69,983 | −145,890 | 38,357 | 14-0-50 | −109,139 / −104,790 |
| warm start − fair leaf | −142,340 | −178,492 to −107,771 | −178,511 | 35,642 | 9-0-55 | −156,143 / −128,537 |
| candidate − warm start | +35,375 | lower +16,899 | | 18,949 | 36-0-28 | both positive |
| fair depth 4 − fair depth 3 | +97,059 | lower +29,330 | +27,011 | 69,023 | 41-0-23 | +75,458 / +118,660 |

The candidate's lower quartile (122,253) fell below the fair leaf's (158,387),
a gate criterion in its own right.

Stage A: 177 complete teacher games (21,618 sibling-complete labelled roots, 3
stopped at the 500-move cap) before the 46,800 s new-game cutoff; 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: validation
Huber 0.6873 rise units at epoch 8 of 16 on a whole-origin split of 17,641 /
3,977 roots; validation Pearson 0.9504 at that epoch; the deployment-faithful
probe put the depth-3 search with the warm start 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; ten-generation paired mean margins over the
warm-start control +4,682, +15,853, +22,704, +32,278, +34,037, +40,431
points, best candidate +76,679 in the last block; 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).

</TechnicalRecord>

<TechnicalRecord summary="Continuation screen" meta="RS-20260903T163321Z-733076b5">

Continuation screen (stage D), 64 never-read games from `0xa52ea000`,
2,000-move cap, corrected 17,000-point Hardcore scoring, zero censored,
illegal or incomplete decisions in every arm:

| Arm (depth-3 search unless stated) | Mean | Median | Lower quartile | Mean moves | Clears/move | Reveals/move |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| continued candidate | 251,667 | 224,341 | 171,440 | 75.36 | 1.8868 | 1.0404 |
| first-run candidate | 215,389 | 185,102 | 138,372 | 65.33 | 1.7898 | 0.9589 |
| unevolved warm start | 143,674 | 123,543 | 107,000 | 45.31 | 1.5124 | 0.7314 |
| frozen fair leaf (comparator) | 320,108 | 276,827 | 189,414 | 94.17 | 1.9935 | 1.1120 |
| frozen fair leaf, depth 4 (reference arm) | 391,907 | 351,132 | 215,823 | 113.28 | 2.0526 | 1.1506 |

Paired contrasts, 20,000 bootstrap resamples:

| Contrast | Mean delta | Bootstrap 95% bounds | Student-t lower 95% | Detection floor | W-T-L | Halves |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| continued − fair leaf | −68,441 | −112,090 to −26,694 | −112,274 | 43,193 | 25-0-39 | −112,383 / −24,500 |
| continued − first-run candidate | +36,278 | lower +9,085 | +8,543 | 27,330 | 41-0-23 | −7,140 / +79,697 |
| continued − warm start | +107,994 | lower +82,769 | +82,119 | 25,497 | 54-0-10 | both positive |
| first-run candidate − fair leaf | −104,719 | lower −143,479 | −143,895 | 38,602 | 20-0-44 | both negative |
| fair depth 4 − fair depth 3 | +71,799 | lower +25,857 | +23,286 | 47,804 | 39-0-25 | +31,746 / +111,851 |

The plateau rule stopped the run after generation 149: slope +239 per
generation (lower bound +30) at the generation-99 check, then +37 (lower
bound −159) across generations 50 to 149. Fifty-generation averages of the
population mean: 204,748, 228,706, 237,946. Elite re-selection on 128 fresh
games froze candidate-12 at 264,466 (finalists spanned 228,212 to 264,466).
The continued candidate's lower quartile (171,440) fell below the fair
leaf's (189,414).

</TechnicalRecord>

<TechnicalRecord summary="Second continuation screen" meta="RS-20260904T090620Z-e5731bf0">

Second continuation screen (stage D), 64 fresh held-out games. Values below
are recorded in
[RS-20260904T090620Z-e5731bf0](/results/RS-20260904T090620Z-e5731bf0):

| Arm (depth-3 search) | Mean | Lower quartile |
| --- | ---: | ---: |
| second-continuation candidate | 249,757 | 156,534 |
| frozen fair leaf (comparator) | 335,266 | 176,108 |

| Paired contrast on the same 64 games | Mean delta | Bootstrap 95% lower bound |
| --- | ---: | ---: |
| candidate − fair leaf | −85,509 | −128,482 |
| candidate − immediately prior candidate | +13,572 | −30,165 |

The gain over the prior candidate is not statistically distinguishable from
zero. The first continuation's equivalent contrast against its own
predecessor, on its separate 64-game screen, was +36,278 with a lower bound
of +9,085
([RS-20260903T163321Z-733076b5](/results/RS-20260903T163321Z-733076b5)).

The plateau rule stopped both continuations after generation 149. The second
continuation's slope at that check was +164 points per generation, with a
lower bound of −29, against the first continuation's +37. Its best candidate
beat the fair control on 26 of 150 training blocks, against 9 before. The
population mean rose through all three fifty-generation blocks (235,353,
238,797, 244,800). Each generation used 32 paired games per candidate. These
training observations did not translate into a confirmed further gain on
the screen.

The comparison tests one slower schedule. It cannot separate limited room
for improvement in this leaf class at this population size and search depth
from a schedule that still shrinks too quickly or selection noise that
hides small improvements. A further schedule comparison, more games per
candidate or a larger population would each require a new experiment.

</TechnicalRecord>

<AgentContext summary="Validity, gates and limitations">

- The preregistered gate, fixed before any leased seed was read: on the 64
  held-out games the evolved candidate must beat the frozen fair leaf at the
  identical depth-3 configuration with a paired mean difference whose
  one-sided 95% bootstrap lower bound and Student-t lower bound are both above
  zero, the difference must be positive in both halves of the block, and the
  candidate's lower-quartile score may not fall below the fair leaf's. Every
  generation artifact and the screen artifact must record zero illegal and
  zero incomplete decisions, and the screened weights must hash to the value
  recorded before the screen lease opened. Outcome: every pass criterion
  except artifact integrity and candidate identity fails. The continuation
  screen (`RS-20260903T163321Z-733076b5`) fails the same gate on a fresh
  64-game block. The second continuation
  (`RS-20260904T090620Z-e5731bf0`) also fails the same gate on its fresh
  64-game block; the gain over the immediately prior candidate is unconfirmed.
- What the rejection covers: this exact configuration, this model class, this
  teacher depth, this optimiser and budget, this deployment depth. It does not
  reject the idea of a learned leaf. Had the candidate passed while the
  unevolved warm start passed equally on the same seeds, the evolutionary
  stage would have contributed nothing and the theory's second leg would have
  been refuted even with the network standing. The retained corpus, warm start
  and populations remain available for a successor experiment with its own
  record and leases.
- Gates before the first leased seed: 11 CHECK gates on the already-open probe
  block (feature determinism and bounds, information-boundary blindness to
  score, level and move count, reflection consistency, fresh-searcher and
  multi-worker determinism, legality and completed depth under random, zero
  and saturated weights, leaf finiteness, serialisation round-trip), all
  passed; a seed-free SMT preflight on the same probe block chose 32 threads
  over 16. The information-boundary gate proves that states differing only in
  score, level or move count produce bit-identical evaluations.
- Limitations: the corpus reached about a third of its planned size, so the
  distillation leg was tested with less data than designed; 64 screen games
  with a paired standard deviation of 186,538 points give a detection floor of
  38,357 for the primary contrast; the fitness blocks are development data and
  the screen block is now read.

</AgentContext>

<AgentContext summary="Scoring mode">

Corrected 17,000-point Hardcore scoring throughout: fitness blocks and the
screen use the ordinary 2,000-move cap, teacher games a 500-move cap (3 of 177
stopped there). The network's output and the Huber loss are in units of one
row rise, 17,000 points. No historical 7,000-point number appears on this
page.

</AgentContext>