The depth-5 teacher crawls, Q-learning finds a pruning prior, and the console charts are rebuilt

The teacher stage ran all day

The run launched at 04:24 UTC yesterday (RUN-20260902T035644Z-c1fd8987, under the successor protocol EX-20260902-nnue-evolution-d3-v2-49c18bc2) was expected to finish its depth-5 teacher corpus in twelve to fourteen hours. It did not. The interim measurement recorded in the run record put the teacher at 37 to 87 seconds per position per thread at 32 threads, four to seven times slower than the pilot on the probe block had projected; the run's analysis.json puts the mean at 77.6 seconds per root over the completed games. A table-size test on the running binary (recorded in the run record's measurement notes) found no configuration faster than the one in use.

negative

The corpus closed at roughly a third of the planned games

175 games, 20,618 labelled roots at the 21:46 UTC snapshot

New teacher games stopped starting at the protocol's 46,800-second wall sub-budget (17:24 UTC), as preregistered; the games in flight then finished one by one. The committed snapshot the console draws from (web/content/figures/nnue-evolution/RUN-20260902T035644Z-c1fd8987.json, copied from the run's analysis.json and part files) records 175 complete games and 20,618 labelled roots with the last game still running: mean teacher score 416,699, median 348,400, best game 1,861,091, one game stopped at the 500-move cap, mean 2.5 hours of depth-5 search per game. The protocol makes whatever whole games completed the corpus and does not depend on the count; whether a corpus this size warm-starts a 572,000-weight network is what the supervised stage's validation loss and ordering probe will measure next.

The remaining stages (supervised warm start, sixty generations of evolution, elite re-selection, the one-shot held-out screen) chain automatically once the last teacher game finishes. That chain aborted once, at 22:09 UTC, on a single false reading of its process-table liveness probe while the teacher binary was still playing its last game; no stage had started and nothing was touched. The probe now needs three consecutive misses before it gives up, and the chain was relaunched a minute later. That repair caused a second fault of its own: the script was replaced by rename in the belief that the still-running corpus stage would keep its original copy, and it did not. The teacher binary finished cleanly at 22:39 UTC with 177 games and 21,618 roots, its summary and resource record written, but the stage's shell then read the replaced file, failed with a syntax error on its last line, and never logged the completion marker, so the chain aborted a second time. The marker was appended to the pipeline log by hand, with a note saying so, and the chain was relaunched; the supervised stage started at 22:42 UTC. No corpus artifact was affected. Both faults are in the run record's measurement notes, and the script is not touched again while a stage is executing.

Can a leaf move a five-move decision at all?

The owner asked whether an NNUE leaf can swing the value of a leaf node enough to change the engine's choice once a deep search has banked the cascade points along each line. Rather than argue it, a diagnostic binary (leaf_swing) was added to the crate and run on a frozen 30-game snapshot of the corpus before any training signal was read: every root was searched twice at each depth, once with the frozen 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.

positive

The leaf term carries about 97% of the spread between columns at every depth

argmax changes at 55.5% / 43.8% / 50.0% of roots at depths 3 / 4 / 5

From the retained artifacts leaf-swing-d3.txt, leaf-swing-d4.txt and leaf-swing-d5.txt (400, 80 and 8 roots): the median ratio of leaf-term spread to total value spread on roots with no losing column is 0.97, 0.95 and 0.97; the median in-tree score spread alone is 393, 643 and 1,141 points against a leaf-term spread of 4,033, 4,093 and 3,813 points; removing the leaf changes the chosen column at 55.5%, 43.8% and 50.0% of roots; and the fair leaf agrees with the depth-5 teacher's column at 72.0%, 73.8% and 100.0% of roots where the zero leaf manages 45.0%, 47.5% and 50.0%. The search does not dilute the leaf; it applies the same evaluator at every frontier of every line. A first version of these artifacts was regenerated after the corpus had grown under it, so the input was frozen (leaf-swing-input/) and the commit that had claimed exact reproduction was corrected in the history.

The console's charts were rebuilt

The owner asked for a human-readable page on how the evolutionary training works, with charts of the run, and pointed out that the console's existing figures overlap and clip. An audit of every rendered figure confirmed it: the string-templated SVG generator estimated text at 0.6 em per character inside a fixed 720 x 400 box, so line-chart legends ran into the x-axis label, the fourteen-category evaluator comparison and the twelve-bar closure chart piled their labels into an unreadable line, the forest plot truncated its title and row names, a work-per-move axis spanning five orders of magnitude stacked its ticks at the left edge, and the hover popovers were clamped inside the SVG where they covered the title.

closed

Research figures now render through a measured visx chart kit

owner: claude-fable-5-1

web/components/charts/ replaces the generator. Titles and legends are HTML and wrap; tick and category labels are measured with a canvas after hydration and the layout is recomputed at the container's real width; a bar or dot figure with more than six categories, or whose labels do not fit their slot, is drawn as rows with a measured label gutter; a numeric axis may declare a log scale as a presentation choice; tooltips render into a portal clamped to the viewport and are reachable by keyboard. The 30 figure specs are unchanged except for one presentation field (the work-frontier axis is now log-scaled), the provenance rule is enforced by web/lib/charts/spec.ts and check-figures.mjs instead of the generator, and the 30 generated SVGs were deleted. Every emitted coordinate is rounded to a hundredth of a pixel because the log scale's last-bit differences between the server's and the browser's Math.log produced a hydration mismatch. Two hand-authored diagrams (diagram-work-budget, diagram-chance-strata) had text running past their boxes and were edited.

running

The approach page explains the evolutionary loop with live charts of the run

owner: claude-fable-5-1

approaches/lifetime-objective/nnue-evolution/README.mdx was rewritten for a reader who has never seen a genetic algorithm: what the leaf decides, why imitation alone failed twice, what common random numbers buy, and one generation step by step. Four schematic diagrams carry the protocol's fixed constants and no measured value. Five data figures read a committed snapshot of the run (web/scripts/extract-nnue-evolution.ts, from analysis.json, the corpus part files, progress.jsonl and the screen artifacts): every teacher game as a dot, the supervised loss curve, fitness by generation against both paired controls with the population's range as a band, every candidate in every generation, and the per-game paired differences of the held-out screen with the preregistered gate checks. A stage without artifacts says so instead of showing a number; the snapshot is re-extracted after each stage completes.

One environment note for whoever opens the dev console from another machine: Next.js 16 blocks its own dev resources for any origin but localhost, which silently prevents hydration, so the charts render but never become interactive. DROP7_DEV_ORIGINS=<host> in next.config.ts allows the host.

Reading the only outside paper on this game

The repository has cited Klein and Friedmann's "Final Report - Drop7" since its strategy landscape was written, as prior work and explicitly not as a baseline. Today it was read end to end, its code fetched from github.com/ekreate/cs221-final-project at the only commit that exists (8cc8a0e, March 2021), and the two were compared. The report is a Stanford CS221 course project; util.py is the course's homework scaffold and the learner class is "Problem 4a". Its game is a simplified Drop7: score is one point per move survived, the board starts empty rather than with a gray row, games stop at 200 moves, and dropping onto a full column ends the game. The learner is textbook Q-learning with a linear function of six hand-made features about the immediate consequences of a drop, and the code's feature definitions differ from the paper's prose in two places (the tallest-column feature carries a count, not an indicator; the column-detonation feature adds the landing cell's gray adjacency unconditionally). The code was ported as coded.

The upstream repository carries no license, so it is fetched at run time into the run directory and never vendored; the theory, both experiments and the reproduction harness live under approaches/value-policy-learning/klein-friedmann-linear-q/.

positive

The report reproduces from its own code

49.08 / 49.05 / 49.02 vs 49.61 reported

Three independent 50,000-game training runs in the authors' simulator, tested over 10,000 games each with no exploration, average 49.08, 49.05 and 49.02 moves (report: 49.61, sd 11.18); uniform-random play averages 31.66, 31.73 and 31.85 over 5,000 games each (report: 31.2). Both preregistered bands pass (RS-20260902T082726Z-75606ce7). The whole five-arm run took nine minutes on the laptop because the upstream simulator turns out to run at about ten thousand moves per second.

negative

Training is front-loaded: 300 games equal 50,000

49.080 vs 49.080 on the same 10,000 test games

The step size is one over the number of moves played, not games, so after a few hundred games it is below a thousandth and the weights barely move. A learner stopped after 300 training games (13,796 updates) tests at 49.080 moves on the same 10,000 games as the 50,000-game learner (2,406,211 updates), which also tests at 49.080. The per-thousand-game training curve is at 46.8 in its first block and 48.2 in its last. The report's 50,000-game protocol spends more than 99% of its compute on a frozen policy.

negative

The regularisation claim does not hold for the shipped code

lambda 0: 48.97 moves, not below random

The report's Figure 5 says the unregularised learner diverged and fell below random, and credits ridge regularisation for the working result. With the ridge constant set to zero the shipped configuration tests at 48.97 moves, indistinguishable from 49.08 with it. The reason is the same step size: the penalty term is scaled by a step of order one hundred-thousandth for almost all of training and cannot act. The report's divergence must have come from a schedule the shipped code no longer contains (it mentions an earlier one over the square root of the count). The numbers of this report are reliable; its explanation of them is not.

Porting the six features onto the Rust bitboard engine

A std-only crate on drop7-rs reimplements the feature extractor and the Q-learning update, with the three step-size and exploration schedules exposed and two necessary deviations disclosed: the argmax and the exploration draw range over legal columns only, and the rules engine is the repository's proven Hardcore engine rather than the upstream simulator, so only feature and update parity are claimed, not board parity. Three CHECK gates ran before any diagnostic seed was read: 54,852 exported (state, action) pairs from 7,242 upstream states with zero feature mismatches after the gray-encoding swap; 250 upstream weight updates replayed with worst relative difference zero; and 64 probe games times three arms byte-identical across one and eight threads and a repeat with zero illegal decisions. Training then ran at about 1.4 million moves per second: each 50,000-game run took 1.3 seconds.

positive

Six first-wave features are a real survival prior on Hardcore rules

+9.5 moves, +33,884 points over random (LB95 +8.85, +31,361)

Re-learned with the upstream schedule on the already-open training block and played on 256 previously evaluated development seeds (0xa5277000-0xa52770ff, 2,000-move cap, nothing censored), the policy lives 35.8 moves and scores 107,147 points on average against 26.25 moves and 73,263 points for uniform-random legal play on the same games, winning 232 of 256 paired games. Three exploration seeds agree to within half a move. Fair depth-4 search averaged 308,296 points on its own 64-game development cohort; on the first 32 of these seeds the depth-4 search with seven chance samples scored 380,205 points and lived 110.4 moves, beating every six-feature arm on at least 30 of the 32 games. Pilot tier: the cohort was read before, no new seed was opened, and nothing here can be promoted (RS-20260902T084356Z-2488ecc7).

positive

The learned weights are an accident of the opening

transplanted authors' weights: +4.74 moves over engine-trained (LB95 +3.77)

The weights the authors' simulator learned (Python seed 10 of the reproduction), transplanted unchanged onto the engine, live 40.5 moves and score 123,968 points, beating the weights learned in the engine by 4.7 moves and 16,821 points. Same six features, same update, same schedule; the only difference is the opening experience (empty board versus gray row) that a step size dying within a few hundred moves locks in. The engine-trained weights are dominated by "drop on the lowest column"; the authors' weights value detonations three times more. Neither is the best six-weight policy, and the temporal-difference fixed point evidently is not either.

negative

Slower step sizes did not raise the plateau; two diverged

constant 0.001: +0.50 moves [LB95 +0.12]; per-game and harmonic: weights to 1e9 and NaN

The preregistered slower schedules were a per-game step, a harmonic step with a 50,000-move time constant, and a constant 0.001, all with survival reward. The first two diverged (weights near a billion, then NaN), which is the deadly-triad instability the report described and which its own collapsing step size had hidden. The constant step gained 0.50 moves over the upstream schedule with a lower bound of 0.12, but two identical-configuration upstream seeds differ by 0.48 (lower bound 0.07), so the gain is the size of seed noise. Three post-hoc arms (constants 0.01 and 0.0001, and a corrected-score reward at 0.001) landed within 0.7 moves of the same plateau. Under the theory's clause (iv) this reads as mixed: the letter of the falsification criterion fires for the constant-step arms, the effect is under one move.

Searching the six weights instead of learning them

The transplant result said the plateau might be the optimizer's rather than the features'. Since the policy is an argmax over six numbers and one candidate costs about a second to play on 256 paired games, the cheapest test was a direct search: a second theory and preregistered experiment (TH-20260902-kf-six-weight-policy-search-8a6e41a8, EX-20260902-kf-six-weight-cem-d018cc89), cross-entropy method over the six weights with population 64, elite 8, 30 generations of 256 paired training games each, re-selection of the finalists on a fresh 1,024-game block, and a single read of the frozen optimum on the pilot cohort.

positive

A five-second search over the six weights beats both learned vectors

44.82 moves, 138,973 points; +4.29 moves over the transplanted weights (LB95 +2.81)

The search converged by generation ten and finished in 4.5 seconds. Its frozen optimum (unit weights 0.43 lowest-column, 0.28 row detonations, 0.15 column detonations, 0.03 tallest-column, 0.74 a-clearing-1, 0.42 the-disc- clears) re-selected at 43.25 moves on the fresh block against a generation fitness of 43.29, so it did not overfit its fitness blocks; the single- generation "best ever" of 44.82 re-selected to 43.12 and was correctly passed over. Read once on the 256 pilot games it lives 44.82 moves and scores 138,973 points: 18.57 moves over random (lower bound 17.24) and 4.29 over the transplanted authors' weights (lower bound 2.81), winning 159 of 256 paired games against them. The optimum's direction is 0.75 cosine from the transplanted vector and 0.55 from the engine-learned one. Six first-wave features therefore top out near 45 moves and 140,000 points on Hardcore rules, an order of magnitude below depth-4 in points (RS-20260902T084356Z-784ebf14).

What is open, and one decision for the coordinator

Open: whether any of the six features, above all "a 1 that clears next to gray" and "lands on a lowest column", earns a place as a cheap correction term inside the depth-4 leaf. That is a different experiment with a different comparator and was not run today. The broader lesson is one this repository keeps meeting from the other direction: a value-prediction objective is not the control objective, and for a six-number policy whole-game search is both cheaper and better than temporal-difference learning.

Decision requested: training read the first 50,000 seeds of the already-open SEEDLEASE-A52 d2 training block (0xa5200000-), whose role is unchanged by another training read; the search re-read the first 8,704 of them; and the pilot read the first 256 seeds of the Rust engine's already-read benchmark sub-block, twice (once per experiment). No new seed was opened and no lease record was written, because the seed allocator is closed to agents. The coordinator should either confirm that reuse in a lease note or assign a fresh training lease before any of these numbers is cited outside the pilot tier. No arm here warrants a SCREEN against depth-4 on fresh seeds.

Later the same day: the Q-learning family inside the search

The owner's response to the morning's work was to set the paper's policy aside and ask the questions it left: fit the linear Q to depth 4, see whether its ranking can prune siblings, and more generally test what Q-learning can do for the efficiency and the leaf of the depth-4 search. The vehicle was a sibling-complete panel: 48 complete fair-d4s7 games on already-read training seeds, every root recorded with the exact value of every legal column at depths one to four, the 32 one-ply numbers of every sibling (the seven-stratum mean of the eighteen leaf terms after the drop, the mean score gained, the terminal fraction, the six Klein-Friedmann features, the rise clock, a bias), and the game's remaining lifetime. The one-ply search value is a fixed linear function of those 32 numbers, so a linear fit can only add to it. Games 0-31 fit, 32-47 held out; 5,398 roots, 2,060 of them held out (DS-20260902-fair-d4s7-sibling-panel-e4e06b7c). A pruned copy of the drop7-rs search expands only a prior's top-w siblings at interior max nodes with two or more plies left; at width seven it is the unchanged search bit for bit, and because chance scenarios and leaves are untouched every pruned value is a lower bound on the full-width value. The panel's first attempt was killed by the session harness one game short of the end and had to be replayed with a resumable builder; that cost 23 minutes and is recorded in the run.

negative

A least-squares Q fitted to depth-4 values ranks worse than the one-ply value it contains

held-out top-1 0.308 vs 0.524; recall@3 0.604 vs 0.809

The ridge fit (lambda 0.001, the smallest value of the preregistered grid, chosen by game-grouped cross-validation) explains 72% of the held-out variance of the depth-4 sibling value and still puts depth-4's choice first on only 30.8% of held-out roots, against 52.4% for the exact one-ply value under the same chance outcomes; the paired difference over games has a one-sided bound of -0.233. The features are exact: the one-ply value rebuilds from them to 6e-11. The objective is the problem. Death within four plies is worth minus a million points, so least squares halves its error by predicting deaths and ranks the live siblings worse; a post-hoc refit on per-root-centred targets clipped at minus 50,000 recovers the one-ply ranking exactly (53.5%, 80.9%) and nothing more. Depth-three exact values rank depth-4's choice into their top three on 95.0% of roots, depth two 90.0%, the six-feature CEM policy 70.1%, centre order 40.0%. Theory clauses (a) and (b) fail (RS-20260902T202705Z-75d87947).

positive

The one-ply value prunes depth 4 to a quarter of the work

92.3% agreement [LB 91.8%], 0.87% normalised regret [UB 0.98%], 26.4% of the work, 3.8x wall

Expanding only the one-ply value's top three siblings at both interior layers reproduces the full-width depth-4 decision on 92.3% of the 2,060 held-out roots and gives up 0.87% of the value spread on average (74 points), at 26.4% of the logical work with the prior itself 5.9% of that. Width four: 93.8% at 40.9%; width two: 89.3% at 13.8%. The six-feature policy, which needs no engine calls, reaches 88.4% at width four; the fitted Q is the worst learned prior (78.8% at width three); centre order is useless (69.1%). Zero violations of the lower-bound property over 33,000 decisions. By the preregistered gate this is a pass on regret and inconclusive on agreement (0.923 against the 0.95 line, above the 0.90 fail line), so the gameplay pilot was run.

negative

Pruning cannot buy a fifth ply

pruned d5 at 1.68x d4 work: 80.5% agreement with exact d5 vs d4's 79.7%; 189 vs 146 points regret per decision

On 128 held-out roots with exact depth-5 values (37 times the depth-4 work each), full-width depth 4 agrees with depth 5 on 79.7% of decisions and loses 146 points per decision against it. Pruned depth 5 with the one-ply prior at width two on all three interior layers agrees on 80.5% and loses 189 points: it does better on the near-tied roots (normalised regret 0.0174 against 0.0285, paired reduction +0.011 with bound -0.0004) and worse on the high-stakes ones, which is where the fifth ply was supposed to help. Wider pruning (3,3,3) costs 4.5 times depth 4 for 82.8%. Clause (d) is inconclusive by the normalised metric and unfavourable in points.

negative

In play, the pruned search is four times cheaper and probably about 8% worse; the cohort cannot tell

256 paired games: -33,608 points [LB -69,418, UB +1,448], floor 35,114; -8.6 moves; 121 wins, 135 losses; 26.8% of the work

Full-width fair d4s7 on the 256 previously evaluated pilot games averages 404,497 points and 116.4 moves (its first 32 rows reproduce the earlier run's checksums exactly). The pruned search with the one-ply prior at width three averages 370,889 and 107.8 at 26.8% of the logical work per move. The preregistered non-inferiority bound (-40,000) is not met and a measurable loss is not shown either, so gate G1 is inconclusive; but the sign agrees with the panel, the lifetime delta and the 32-game preview. Pruning is a cheaper policy, not a free speed-up, and whether it costs 0% or 17% needs a cohort four times this size (RS-20260903T013022Z-d32ee053).

negative

Pruned depth 5 in play: a non-measurement pointing the same way as the panel

128 paired games at 1.8x the depth-4 work: -21,615 points [LB -76,153, UB +31,957], floor 54,839; 62 wins, 66 losses

The pruned depth-5 arm (one-ply prior, width two on all three interior layers) was run because the preregistered normalised-regret metric had been favourable on the panel, although the raw-regret reading was not. In play it scores 383,515 on the first 128 games against the comparator's 405,130 on the same games. Nothing here suggests a fifth ply bought by pruning is worth its cost (RS-20260903T013022Z-d32ee053).

positive

One bootstrapped leaf refit explains more of the depth-4 value and the six drop features add nothing

held-out R^2 0.852 vs 0.729 for a rescaled frozen leaf; +0.0001 for the six-feature term

Ridge regression of the depth-4 root value on the eighteen fair-leaf terms plus intercept, on fit roots that are not about to die, explains 85.2% of the held-out variance against 72.9% for an affine rescaling of the frozen leaf, so the preregistered fit gate passed and the gameplay arm was run. The "best immediate Klein-Friedmann drop" term added as a nineteenth feature changed the held-out R^2 by -0.0001: the six drop features carry nothing the fair terms lack, which was the predicted answer to the morning's open question about a six-feature correction term (RS-20260903T013022Z-b0937bf9).

negative

The refit leaf loses a fifth of the score inside the search

256 paired games: -83,451 points [LB -118,022, UB -49,021], floor 34,345; -22.7 moves; 104 wins, 152 losses

Inside full-width fair d4s7 the refit-18 leaf averages 321,045 points and 93.7 moves against 404,497 and 116.4 for the frozen leaf on the same games: a measurable loss, upper bound well below zero, and the theory is not supported. The disclosed proxy had pointed the right way (agreement with exact depth 5 fell from 80% to 68% and raw regret rose from 146 to 9,350 points per decision). The mechanism is the one that sank the action-value fit too: least squares is dominated by what it can predict cheaply, here the expected score of the next four plies among live positions, while the search needs the leaf to price the positions it must choose between, above all the ones near death that the fit excluded. One-step linear TreeStrap on these eighteen terms is closed as tested (RS-20260903T013022Z-b0937bf9).

What the afternoon closes, what it leaves open, and two process failures

Closed at pilot tier, each for the configuration tested: a least-squares linear action value over one-ply features as a ranker or a pruning prior; pruning as a way to buy a fifth ply; a one-step linear leaf refit to the search's own value. The positive residue is a mechanism, not a policy: the search's own one-ply value keeps 92% of depth-4's decisions at a quarter of the work, the pruned values are provably lower bounds on the full-width ones, and the cost of the changed 8% of decisions is about 8% of the score with bounds from 17% to nothing. A 1,024-game paired screen under a coordinator lease is what would settle whether the pruned search can stand in for full-width depth 4 in future diagnostic screens. The next fit worth trying for the action value is a rank-aware objective within each root, and the leaf direction reopens only with a refit that keeps the death-dominated positions or shrinks toward the frozen constants.

Two process failures belong in the record. The first panel build was killed by the session harness one game short of the end with nothing written, and was replayed with a builder that saves each game as it finishes (23 minutes lost). Then, by starting the gameplay arms while the panel stages still held the machine, the three runs exceeded their preregistered wall stops by 2.7%, 2.9% and about 9%; no watchdog enforced them, every output is deterministic and no number is affected, but the wall-time figures throughout are contended and only logical work should be compared. Zero new seeds were opened; the morning's request stands: the coordinator should confirm the reuse of the training block and the pilot cohort (read a fourth time today) or assign fresh leases before any of these numbers is cited outside the pilot tier.