Port the six-feature linear Q-learner onto the Rust bitboard engine: parity gates and a diagnostic pilot on previously evaluated development seeds

A std-only Rust crate on drop7-rs (approaches/fair-expectimax/rust-engine) reimplements the upstream feature extractor and Q-learning update exactly (CHECK gates G1-G2), plays only legal columns deterministically (G3), and then, on corrected five-move Hardcore rules with a 2,000-move cap, tests theory clauses (iii) and (iv): the six features re-learned with survival reward (+1 per surviving move, 0 on the terminal transition) under the upstream schedule (epsilon = t^-0.25, eta = 1/t per move, lambda = 0.1, 50,000 training games) outlive uniform-random legal play (G4) but score below two thirds of the fair depth-4 ledger mean and lose to fair d4s7 on paired seeds (G5); slower schedules (per-game, harmonic tau = 50,000 moves, constant 1e-3) are compared with the upstream schedule on paired seeds (G6).

preregisteredtier: PILOTalgorithmicpreviously-evaluated-developmentpublic-policyEX-20260902-kf-linear-q-rust-transfer-4328a730

On this page

Linked approachKlein-Friedmann linear Q-learning

Theories

TH-20260902-kf-linear-q-transfer-68b41d66

Results

RS-20260902T084356Z-2488ecc7

Dates

Created: 2026-09-02T08:05:17Z
Updated: 2026-09-02T08:10:23Z

Record idEX-20260902-kf-linear-q-rust-transfer-4328a730

No explanation has been written for this record yet.

Technical recordThe registered protocolEX-20260902-kf-linear-q-rust-transfer-4328a730

Hypothesis: A std-only Rust crate on drop7-rs (approaches/fair-expectimax/rust-engine) reimplements the upstream feature extractor and Q-learning update exactly (CHECK gates G1-G2), plays only legal columns deterministically (G3), and then, on corrected five-move Hardcore rules with a 2,000-move cap, tests theory clauses (iii) and (iv): the six features re-learned with survival reward (+1 per surviving move, 0 on the terminal transition) under the upstream schedule (epsilon = t^-0.25, eta = 1/t per move, lambda = 0.1, 50,000 training games) outlive uniform-random legal play (G4) but score below two thirds of the fair depth-4 ledger mean and lose to fair d4s7 on paired seeds (G5); slower schedules (per-game, harmonic tau = 50,000 moves, constant 1e-3) are compared with the upstream schedule on paired seeds (G6). Exploratory arms, reported but not gated: the Python-learned weights from the reproduction transplanted directly, and a score-delta reward (score delta / 17,000) under the harmonic schedule. Necessary deviations from upstream, all disclosed: the argmax and the exploration draw range over LEGAL columns only (upstream lets the agent drop on a full column and die); ties keep upstream's highest-index rule; the rules engine is the repository's proven Hardcore engine, not the upstream simulator, so wave-boundary semantics differ and board-level parity is not claimed -- only feature and update parity are.
Arms: Arm Name Entry point Manifest
Candidate drop7-kf-linear-q (Rust crate on drop7-rs): six-feature linear Q policy, several trained weight sets approaches/value-policy-learning/klein-friedmann-linear-q/rust/src/lib.rs –
Comparator uniform-random legal policy and centre-first policy on identical seeds (primary); fair d4s7 (drop7-rs Searcher, FairLeaf, 65,536-entry depth table, terminal utility -1,000,000, policy seed 0xd7075eed) on the first 32 seeds as a diagnostic reference approaches/fair-expectimax/rust-engine/src/search.rs –
Classification: algorithmic
Information boundary: public-policy
Benchmark tier: PILOT
Lifecycle: preregistered
Theories tested: Klein-Friedmann six-feature linear Q-learning transfers to corrected Hardcore as a survival heuristic, not a scoring policy
Primary metric: mean lifetime in moves and mean corrected Hardcore score per arm over the 256 diagnostic games, with paired deltas against uniform-random on the same seeds and one-sided 95% bootstrap lower bounds (10,000 resamples, bootstrap RNG seed 0x6b66_0001)
Secondary metrics: median, lower quartile, minimum and maximum score and moves
censored games at the 2,000-move cap
numbered clears and covered reveals per move; mean and maximum chain depth
illegal decisions (must be 0) and runner failures
paired lifetime and score deltas between every trained arm and the upstream-schedule arm
fair d4s7 versus the best six-feature arm on the paired 32-seed subset
training curves as per-1,000-game block means; final weights; moves per second of training and evaluation with the machine profile
Statistical unit: whole-game
Uncertainty method: paired per-game deltas with one-sided 95% percentile bootstrap lower bounds over whole games; the 256-game paired detection floor is reported alongside every null
Data role: previously-evaluated-development
Seed leases: none recorded
Whole-origin split: yes
Reuse disclosure: Zero new seeds are opened. The 0xa52 lease family has no machine-readable lease records yet (docs/exploratory/lease-map.md, Incident 2), so this record cites the map. TRAINING reads the first 60,000 seeds of the already-opened TRAINING-role block SEEDLEASE-A52 d2 (0xa5200000-0xa520ea5f), whose role is unchanged by another training read; every trained arm uses the same ordered training seeds (common random numbers) and its own exploration stream in a separate policy-sampling domain. The DIAGNOSTIC evaluation reads 0xa5277000-0xa52770ff, the first 256 seeds of the Rust-engine benchmark sub-block of SEEDLEASE-A52-FAST (development role, already read by the centre policy in RUN-20260824T052018Z-b88c3e22), so its evidence tier is pilot/diagnostic and it can never become confirmation evidence. CHECK probes may read 0xa5278000-0xa52784ff (already open for probes). Training and evaluation seeds are disjoint by construction. The coordinator (repository owner) is asked to confirm this reuse or to assign a fresh training lease before any of these numbers are cited outside the pilot tier.
Pass criteria: G1 feature parity (CHECK, seed-free): on at least 20,000 (state, action) pairs exported from the upstream simulator under random play, the Rust feature presence mask and values equal the upstream extractor's exactly after the gray-disc encoding swap (upstream 9 = untouched, 8 = cracked; engine 8 = untouched, 9 = cracked).
G2 update parity (CHECK, seed-free): replaying the transitions of at least 3 upstream training games (features of the taken action, reward, features of all seven successor actions or terminal) through the Rust learner reproduces the upstream weights after every step to within 1e-9 relative.
G3 legality and determinism (CHECK): zero illegal decisions in every arm; per-game rows byte-identical across 1 and 8 evaluation threads and across two repeated runs; the policy's input type carries only the visible board, the visible next disc and moves until the next rise.
G4 (theory iii-a): the upstream-schedule survival arm's paired mean-lifetime delta over uniform-random has a one-sided 95% bootstrap lower bound above zero.
G5 (theory iii-b): every six-feature arm's mean corrected score is below 205,530, and fair d4s7's paired score delta over the best six-feature arm on the 32-seed subset has a one-sided 95% lower bound above zero.
G6 (theory iv, recorded either way): whether any slower-schedule arm beats the upstream-schedule arm's mean lifetime with a paired one-sided 95% lower bound above zero.
On pass: Record a pilot-tier assessment of the theory. If and only if a six-feature arm falls within the 256-game detection floor of the d4s7 arm, request a fresh public-development lease for a paired SCREEN; otherwise the port is retained as a fast laboratory for feature-based Q-learning and no later tier is opened by this record.
On fail: A CHECK failure stops the pilot before any diagnostic seed is read and is recorded as a mechanics failure of the port, not of the theory. A pilot gate failure is recorded valid + fail for the exact configuration; no further cohort is opened.
Gate fixed before controlled data: yes
Resources: Wall seconds 7200
CPU threads 8
Max host bytes 8589934592
Max GPU bytes –
GPU devices –
Stop conditions: Any G1-G3 failure stops the run before the diagnostic cohort is read.
Stop at 7,200 s wall or 8 GiB resident; a stopped arm is recorded partial and never summarised as valid.
No seed outside 0xa5200000-0xa520ea5f (training), 0xa5277000-0xa52770ff (diagnostic) and 0xa5278000-0xa52784ff (probes) may be read under this record.
Expected artifacts: approaches/value-policy-learning/klein-friedmann-linear-q/rust/ (cargo crate: features, learner, game driver, parity and evaluation binaries)
approaches/value-policy-learning/klein-friedmann-linear-q/reproduction/export_parity.py (upstream feature and transition export)
runs/<run-id>/kf-linear-q/parity-features.txt and parity-update.txt (gate outputs)
runs/<run-id>/kf-linear-q/train-<arm>.json and weights-<arm>.txt
runs/<run-id>/kf-linear-q/evaluate.json (per-arm per-game rows) and summary.json (paired statistics)
Amendments: none recorded

Arm	Name	Entry point	Manifest
Candidate	drop7-kf-linear-q (Rust crate on drop7-rs): six-feature linear Q policy, several trained weight sets	`approaches/value-policy-learning/klein-friedmann-linear-q/rust/src/lib.rs`	–
Comparator	uniform-random legal policy and centre-first policy on identical seeds (primary); fair d4s7 (drop7-rs Searcher, FairLeaf, 65,536-entry depth table, terminal utility -1,000,000, policy seed 0xd7075eed) on the first 32 seeds as a diagnostic reference	`approaches/fair-expectimax/rust-engine/src/search.rs`	–

Wall seconds	7200
CPU threads	8
Max host bytes	8589934592
Max GPU bytes	–
GPU devices	–

Technical recordResults recorded against this protocol1 record

valid runoutcome: passmixedtier: pilotRS-20260902T084356Z-2488ecc7

On corrected five-move Hardcore rules the six Klein-Friedmann features are a survival heuristic worth about ten moves and 34,000 points over uniform-random play, an order of magnitude below fair depth-4 in points. Pilot on 256 previously evaluated development seeds (0xa5277000-0xa52770ff, 2,000-move cap, no game censored, 0 illegal decisions): uniform-random legal play 26.25 moves / 73263 points; centre-first 21.01 / 55083; the six features re-learned with the upstream schedule (three exploration seeds) 35.79, 35.81, 36.27 moves and 107147, 107231, 109068 points, a paired gain over random of +9.54 moves [LB95 +8.85] and +33884 points [LB95 +31361] for seed a. The authors' own Python-learned weights (reproduction seed 10) transplanted unchanged live 40.53 moves / 123968 points, +4.74 moves over the engine-learned weights [LB95 +3.77]: same features, same update, same schedule, different opening experience (empty board vs gray row) frozen in by the per-move 1/t step size. Fair d4s7 (drop7-rs, 65,536-entry table) on the first 32 seeds: 380205 points / 110.38 moves, beating the transplanted arm by +256507 points [LB95 +182668] on 31/32 paired games. Slower schedules: per-game 1/g diverged to weights near 1e9 (32.47 moves), harmonic tau 50,000 and the score-reward harmonic arm diverged to NaN (20.90 moves, the rightmost-column policy), constant 0.001 gained +0.50 moves over the upstream schedule [LB95 +0.12], the same size as the +0.48 [+0.07] spread between two identical-configuration seeds. Gates G1-G5 pass; G6 records that a slower stable schedule clears LB95 > 0 by under a move, so theory clause (iv) is not supported at the letter while clauses (iii-a) and (iii-b) are. All CHECK gates (feature parity 54,852 pairs / 0 mismatches; update parity 250 steps / exact; determinism and legality) passed before any diagnostic seed was read. Training ran at about 1.4 million moves per second (1.3 s per 50,000-game arm).

What it had to pass

✓G1 feature parity: >= 20,000 exported (state, action) pairs, 0 mismatches after the gray-encoding swap — observed: 54,852 pairs from 7,242 upstream states (100 random-play and 100 greedy games), 0 mismatches (parity-features.txt)
✓G2 update parity: >= 3 upstream training games replayed with weights within 1e-9 relative after every step — observed: 5 games, 250 incorporateFeedback calls, worst relative difference 0 (parity-update.txt)
✓G3 legality and determinism: 0 illegal decisions; rows byte-identical across 1 and 8 threads and a repeat; public-view-only policy type — observed: 3 arms x 64 probe games identical across threads1/threads8/repeat, 0 illegal decisions (g3-determinism.txt); every cohort arm reports 0 illegal decisions; policies take PublicView only
✓G4 (theory iii-a): upstream-schedule survival arm's paired lifetime delta over uniform-random has LB95 > 0 — observed: v1a +9.54 moves [LB95 +8.85], 232/0/24; v1b and v1c agree
✓G5 (theory iii-b): every six-feature arm's mean corrected score < 205,530 and fair d4s7 beats the best six-feature arm on the 32-seed subset with LB95 > 0 — observed: highest six-feature mean in this experiment 123,968 (transplant); fair d4s7 minus transplant +256507 [LB95 +182668], d4s7 wins 31/32; minus v1a +270542 [LB95 +192695], 32/32
✓G6 (theory iv, recorded either way): whether any slower-schedule arm beats the upstream-schedule arm's mean lifetime with LB95 > 0 — observed: constant 0.001: +0.50 moves [LB95 +0.12] (floor 0.39); per-game step: weights ~1e9, -3.32 moves; harmonic tau 50,000: NaN weights, -14.89 moves. Identical-configuration seed spread v1c-v1a: +0.48 [LB95 +0.07]. Post hoc: constant 0.01 +0.70 [+0.18], constant 0.0001 +0.40 [+0.02], score reward at 0.001 +0.45 [+0.00]. 'passed' means a slower arm did clear LB95 > 0; the effect equals seed noise.

Technical recordRecorded metricsRS-20260902T084356Z-2488ecc7

cohort

seedsStartHex: 0xa5277000
games: 256
moveCap: 2,000
role: previously-evaluated-development
trainingSeedsStartHex: 0xa5200000
trainingGamesPerArm: 50,000

arms

random

games: 256
meanScore: 73262.7500
medianScore: 68724.5000
q25Score: 68231.2500
minScore: 34,042
maxScore: 123,354
meanMoves: 26.2461
q25Moves: 25
censored: 0
clearsPerMove: 0.9446
revealsPerMove: 0.3270
meanChainDepth: 1.4725
maxChainDepth: 9
illegal: 0
incomplete: 0
wallSecondsPerGame: 0.0000

center

games: 256
meanScore: 55082.8906
medianScore: 51204.5000
q25Score: 34,109
minScore: 17,028
maxScore: 106,655
meanMoves: 21.0078
q25Moves: 15
censored: 0
clearsPerMove: 0.8922
revealsPerMove: 0.2334
meanChainDepth: 1.3220
maxChainDepth: 8
illegal: 0
incomplete: 0
wallSecondsPerGame: 0.0000

v1a

games: 256
meanScore: 107147.1328
medianScore: 103426.5000
q25Score: 102705.5000
minScore: 85,228
maxScore: 209,156
meanMoves: 35.7891
q25Moves: 35
censored: 0
clearsPerMove: 1.2371
revealsPerMove: 0.5504
meanChainDepth: 1.5714
maxChainDepth: 12
illegal: 0
incomplete: 0
wallSecondsPerGame: 0.0000

v1b

games: 256
meanScore: 107230.8086
medianScore: 103411.5000
q25Score: 102692.5000
minScore: 85,228
maxScore: 211,246
meanMoves: 35.8086
q25Moves: 35
censored: 0
clearsPerMove: 1.2399
revealsPerMove: 0.5521
meanChainDepth: 1.5723
maxChainDepth: 12
illegal: 0
incomplete: 0
wallSecondsPerGame: 0.0000

v1c

games: 256
meanScore: 109067.7344
medianScore: 103,482
q25Score: 102721.2500
minScore: 85,228
maxScore: 231,651
meanMoves: 36.2734
q25Moves: 35
censored: 0
clearsPerMove: 1.2562
revealsPerMove: 0.5646
meanChainDepth: 1.5801
maxChainDepth: 12
illegal: 0
incomplete: 0
wallSecondsPerGame: 0.0000

v2-pergame

games: 256
meanScore: 95918.0313
medianScore: 87101.5000
q25Score: 85,828
minScore: 51,141
maxScore: 190,284
meanMoves: 32.4727
q25Moves: 30
censored: 0
clearsPerMove: 1.1816
revealsPerMove: 0.5537
meanChainDepth: 1.7338
maxChainDepth: 13
illegal: 0
incomplete: 0
wallSecondsPerGame: 0.0000

v3-harmonic50k

games: 256
meanScore: 54564.5430
medianScore: 51,227
q25Score: 34,109
minScore: 17,028
maxScore: 110,564
meanMoves: 20.8984
q25Moves: 15
censored: 0
clearsPerMove: 0.8770
revealsPerMove: 0.1800
meanChainDepth: 1.3025
maxChainDepth: 8
illegal: 0
incomplete: 0
wallSecondsPerGame: 0.0000

v4-const1e-3

games: 256
meanScore: 108917.9805
medianScore: 103542.5000
q25Score: 102721.2500
minScore: 85,228
maxScore: 214,411
meanMoves: 36.2852
q25Moves: 35
censored: 0
clearsPerMove: 1.2555
revealsPerMove: 0.5636
meanChainDepth: 1.5780
maxChainDepth: 12
illegal: 0
incomplete: 0
wallSecondsPerGame: 0.0000

v5-score17k-harmonic

games: 256
meanScore: 54564.5430
medianScore: 51,227
q25Score: 34,109
minScore: 17,028
maxScore: 110,564
meanMoves: 20.8984
q25Moves: 15
censored: 0
clearsPerMove: 0.8770
revealsPerMove: 0.1800
meanChainDepth: 1.3025
maxChainDepth: 8
illegal: 0
incomplete: 0
wallSecondsPerGame: 0.0000

posthoc-v6-const1e-2

games: 256
meanScore: 110032.4961
medianScore: 103524.5000
q25Score: 102666.2500
minScore: 85,228
maxScore: 270,105
meanMoves: 36.4922
q25Moves: 35
censored: 0
clearsPerMove: 1.2736
revealsPerMove: 0.5796
meanChainDepth: 1.5684
maxChainDepth: 11
illegal: 0
incomplete: 0
wallSecondsPerGame: 0.0000

posthoc-v7-const1e-4

games: 256
meanScore: 108693.9063
medianScore: 103560.5000
q25Score: 102721.2500
minScore: 85,228
maxScore: 214,315
meanMoves: 36.1875
q25Moves: 35
censored: 0
clearsPerMove: 1.2584
revealsPerMove: 0.5660
meanChainDepth: 1.5801
maxChainDepth: 12
illegal: 0
incomplete: 0
wallSecondsPerGame: 0.0000

posthoc-v8-score17k-const1e-3

games: 256
meanScore: 108933.6250
medianScore: 103542.5000
q25Score: 102,676
minScore: 85,246
maxScore: 214,411
meanMoves: 36.2344
q25Moves: 35
censored: 0
clearsPerMove: 1.2605
revealsPerMove: 0.5680
meanChainDepth: 1.5803
maxChainDepth: 12
illegal: 0
incomplete: 0
wallSecondsPerGame: 0.0000

transplant-py-seed10

games: 256
meanScore: 123967.9961
medianScore: 120,286
q25Score: 102800.7500
minScore: 85,133
maxScore: 241,875
meanMoves: 40.5313
q25Moves: 35
censored: 0
clearsPerMove: 1.4193
revealsPerMove: 0.6853
meanChainDepth: 1.5337
maxChainDepth: 9
illegal: 0
incomplete: 0
wallSecondsPerGame: 0.0000

fair-d4s7

games: 32
meanScore: 380204.9375
medianScore: 284,190
q25Score: 198511.7500
minScore: 138,800
maxScore: 1,239,232
meanMoves: 110.3750
q25Moves: 60
censored: 0
clearsPerMove: 2.0637
revealsPerMove: 1.1608
meanChainDepth: 2.1219
maxChainDepth: 10
illegal: 0
incomplete: 0
wallSecondsPerGame: 237.5977

pairedVsRandom

center

score

meanDelta: -18179.8594
lowerBound95OneSided: -21128.5625
detectionFloor: 2970.9664
wins: 61
ties: 2
losses: 193
n: 256

moves

meanDelta: -5.2383
lowerBound95OneSided: -6.0781
detectionFloor: 0.8428
wins: 44
ties: 39
losses: 173
n: 256

v1a

score

meanDelta: 33884.3828
lowerBound95OneSided: 31360.6289
detectionFloor: 2530.6859
wins: 232
ties: 0
losses: 24
n: 256

moves

meanDelta: 9.5430
lowerBound95OneSided: 8.8477
detectionFloor: 0.7077
wins: 230
ties: 20
losses: 6
n: 256

v1b

score

meanDelta: 33968.0586
lowerBound95OneSided: 31443.8164
detectionFloor: 2551.5832
wins: 233
ties: 0
losses: 23
n: 256

moves

meanDelta: 9.5625
lowerBound95OneSided: 8.8555
detectionFloor: 0.7115
wins: 230
ties: 20
losses: 6
n: 256

v1c

score

meanDelta: 35804.9844
lowerBound95OneSided: 33039.3242
detectionFloor: 2815.8299
wins: 234
ties: 0
losses: 22
n: 256

moves

meanDelta: 10.0273
lowerBound95OneSided: 9.2617
detectionFloor: 0.7767
wins: 231
ties: 20
losses: 5
n: 256

v2-pergame

score

meanDelta: 22655.2813
lowerBound95OneSided: 19792.6445
detectionFloor: 2916.0044
wins: 204
ties: 0
losses: 52
n: 256

moves

meanDelta: 6.2266
lowerBound95OneSided: 5.4375
detectionFloor: 0.8037
wins: 184
ties: 40
losses: 32
n: 256

v3-harmonic50k

score

meanDelta: -18698.2070
lowerBound95OneSided: -21678.6992
detectionFloor: 3018.8354
wins: 60
ties: 2
losses: 194
n: 256

moves

meanDelta: -5.3477
lowerBound95OneSided: -6.1953
detectionFloor: 0.8606
wins: 41
ties: 39
losses: 176
n: 256

v4-const1e-3

score

meanDelta: 35655.2305
lowerBound95OneSided: 33169.3477
detectionFloor: 2527.5671
wins: 236
ties: 0
losses: 20
n: 256

moves

meanDelta: 10.0391
lowerBound95OneSided: 9.3477
detectionFloor: 0.7037
wins: 233
ties: 18
losses: 5
n: 256

v5-score17k-harmonic

score

meanDelta: -18698.2070
lowerBound95OneSided: -21678.6992
detectionFloor: 3018.8354
wins: 60
ties: 2
losses: 194
n: 256

moves

meanDelta: -5.3477
lowerBound95OneSided: -6.1953
detectionFloor: 0.8606
wins: 41
ties: 39
losses: 176
n: 256

posthoc-v6-const1e-2

score

meanDelta: 36769.7461
lowerBound95OneSided: 33997.0234
detectionFloor: 2798.6296
wins: 232
ties: 0
losses: 24
n: 256

moves

meanDelta: 10.2461
lowerBound95OneSided: 9.4805
detectionFloor: 0.7695
wins: 228
ties: 24
losses: 4
n: 256

posthoc-v7-const1e-4

score

meanDelta: 35431.1563
lowerBound95OneSided: 32888.9258
detectionFloor: 2584.5708
wins: 234
ties: 0
losses: 22
n: 256

moves

meanDelta: 9.9414
lowerBound95OneSided: 9.2344
detectionFloor: 0.7174
wins: 231
ties: 20
losses: 5
n: 256

posthoc-v8-score17k-const1e-3

score

meanDelta: 35670.8750
lowerBound95OneSided: 33191.8477
detectionFloor: 2528.9800
wins: 235
ties: 0
losses: 21
n: 256

moves

meanDelta: 9.9883
lowerBound95OneSided: 9.2930
detectionFloor: 0.7046
wins: 232
ties: 19
losses: 5
n: 256

transplant-py-seed10

score

meanDelta: 50705.2461
lowerBound95OneSided: 47009.2500
detectionFloor: 3724.6683
wins: 238
ties: 0
losses: 18
n: 256

moves

meanDelta: 14.2852
lowerBound95OneSided: 13.2500
detectionFloor: 1.0505
wins: 232
ties: 20
losses: 4
n: 256

fair-d4s7

score

meanDelta: 310871.1875
lowerBound95OneSided: 233288.5000
detectionFloor: 81978.2400
wins: 32
ties: 0
losses: 0
n: 32

moves

meanDelta: 85.2188
lowerBound95OneSided: 64.0313
detectionFloor: 22.4503
wins: 32
ties: 0
losses: 0
n: 32

pairedVsUpstreamScheduleSeedA

random

score

meanDelta: -33884.3828
lowerBound95OneSided: -36389.2930
detectionFloor: 2530.6859
wins: 24
ties: 0
losses: 232
n: 256

moves

meanDelta: -9.5430
lowerBound95OneSided: -10.2500
detectionFloor: 0.7077
wins: 6
ties: 20
losses: 230
n: 256

center

score

meanDelta: -52064.2422
lowerBound95OneSided: -55219.3398
detectionFloor: 3225.5056
wins: 16
ties: 0
losses: 240
n: 256

moves

meanDelta: -14.7813
lowerBound95OneSided: -15.6758
detectionFloor: 0.9157
wins: 3
ties: 19
losses: 234
n: 256

v1b

score

meanDelta: 83.6758
lowerBound95OneSided: -271.9648
detectionFloor: 381.3271
wins: 7
ties: 244
losses: 5
n: 256

moves

meanDelta: 0.0195
lowerBound95OneSided: -0.0781
detectionFloor: 0.1067
wins: 2
ties: 251
losses: 3
n: 256

v1c

score

meanDelta: 1920.6016
lowerBound95OneSided: 374.1523
detectionFloor: 1629.8465
wins: 33
ties: 203
losses: 20
n: 256

moves

meanDelta: 0.4844
lowerBound95OneSided: 0.0703
detectionFloor: 0.4397
wins: 16
ties: 226
losses: 14
n: 256

v2-pergame

score

meanDelta: -11229.1016
lowerBound95OneSided: -14208.0313
detectionFloor: 2983.8409
wins: 99
ties: 0
losses: 157
n: 256

moves

meanDelta: -3.3164
lowerBound95OneSided: -4.1289
detectionFloor: 0.8190
wins: 51
ties: 62
losses: 143
n: 256

v3-harmonic50k

score

meanDelta: -52582.5898
lowerBound95OneSided: -55796.9414
detectionFloor: 3226.3482
wins: 17
ties: 0
losses: 239
n: 256

moves

meanDelta: -14.8906
lowerBound95OneSided: -15.8164
detectionFloor: 0.9204
wins: 4
ties: 19
losses: 233
n: 256

v4-const1e-3

score

meanDelta: 1770.8477
lowerBound95OneSided: 445.3867
detectionFloor: 1367.1018
wins: 51
ties: 174
losses: 31
n: 256

moves

meanDelta: 0.4961
lowerBound95OneSided: 0.1211
detectionFloor: 0.3871
wins: 30
ties: 208
losses: 18
n: 256

v5-score17k-harmonic

score

meanDelta: -52582.5898
lowerBound95OneSided: -55796.9414
detectionFloor: 3226.3482
wins: 17
ties: 0
losses: 239
n: 256

moves

meanDelta: -14.8906
lowerBound95OneSided: -15.8164
detectionFloor: 0.9204
wins: 4
ties: 19
losses: 233
n: 256

posthoc-v6-const1e-2

score

meanDelta: 2885.3633
lowerBound95OneSided: 949.8594
detectionFloor: 2003.6456
wins: 69
ties: 126
losses: 61
n: 256

moves

meanDelta: 0.7031
lowerBound95OneSided: 0.1797
detectionFloor: 0.5441
wins: 35
ties: 192
losses: 29
n: 256

posthoc-v7-const1e-4

score

meanDelta: 1546.7734
lowerBound95OneSided: 203.3828
detectionFloor: 1403.0656
wins: 40
ties: 188
losses: 28
n: 256

moves

meanDelta: 0.3984
lowerBound95OneSided: 0.0234
detectionFloor: 0.3876
wins: 23
ties: 215
losses: 18
n: 256

posthoc-v8-score17k-const1e-3

score

meanDelta: 1786.4922
lowerBound95OneSided: 247.9688
detectionFloor: 1553.9987
wins: 60
ties: 158
losses: 38
n: 256

moves

meanDelta: 0.4453
lowerBound95OneSided: 0.0039
detectionFloor: 0.4430
wins: 33
ties: 199
losses: 24
n: 256

transplant-py-seed10

score

meanDelta: 16820.8633
lowerBound95OneSided: 13375.1406
detectionFloor: 3551.9351
wins: 172
ties: 1
losses: 83
n: 256

moves

meanDelta: 4.7422
lowerBound95OneSided: 3.7656
detectionFloor: 1.0052
wins: 146
ties: 54
losses: 56
n: 256

fair-d4s7

score

meanDelta: 270541.7500
lowerBound95OneSided: 192695.0313
detectionFloor: 82053.3265
wins: 32
ties: 0
losses: 0
n: 32

moves

meanDelta: 73.7813
lowerBound95OneSided: 52.5625
detectionFloor: 22.4345
wins: 32
ties: 0
losses: 0
n: 32

fairD4s7MinusArmOn32Seeds

v1a

score

n: 32
meanDelta: 270541.7500
lowerBound95OneSided: 192695.0313
detectionFloor: 82053.3265
d4Wins: 32

moves

n: 32
meanDelta: 73.7813
lowerBound95OneSided: 52.5625
detectionFloor: 22.4345
d4Wins: 32

v1b

score

n: 32
meanDelta: 270525.1563
lowerBound95OneSided: 192682.6250
detectionFloor: 82057.2092
d4Wins: 32

moves

n: 32
meanDelta: 73.7813
lowerBound95OneSided: 52.5625
detectionFloor: 22.4345
d4Wins: 32

v1c

score

n: 32
meanDelta: 270887.7188
lowerBound95OneSided: 193129.6563
detectionFloor: 82029.1945
d4Wins: 32

moves

n: 32
meanDelta: 73.9375
lowerBound95OneSided: 52.6875
detectionFloor: 22.4185
d4Wins: 32

v4-const1e-3

score

n: 32
meanDelta: 267748.4375
lowerBound95OneSided: 189256.0938
detectionFloor: 82670.4366
d4Wins: 32

moves

n: 32
meanDelta: 73
lowerBound95OneSided: 51.5938
detectionFloor: 22.6086
d4Wins: 32

posthoc-v6-const1e-2

score

n: 32
meanDelta: 268999.6563
lowerBound95OneSided: 190633.8125
detectionFloor: 82964.0674
d4Wins: 32

moves

n: 32
meanDelta: 73.4688
lowerBound95OneSided: 52.0313
detectionFloor: 22.6736
d4Wins: 32

transplant-py-seed10

score

n: 32
meanDelta: 256506.5313
lowerBound95OneSided: 182667.8438
detectionFloor: 78563.4987
d4Wins: 31

moves

n: 32
meanDelta: 69.7813
lowerBound95OneSided: 49.6250
detectionFloor: 21.4588
d4Wins: 30

cem-optimum

score

n: 32
meanDelta: 233361.8125
lowerBound95OneSided: 153370.2500
detectionFloor: 84813.9559
d4Wins: 30

moves

n: 32
meanDelta: 63.2813
lowerBound95OneSided: 41.3750
detectionFloor: 23.2453
d4Wins: 30

random

score

n: 32
meanDelta: 310871.1875
lowerBound95OneSided: 233288.5000
detectionFloor: 81978.2400
d4Wins: 32

moves

n: 32
meanDelta: 85.2188
lowerBound95OneSided: 64.0313
detectionFloor: 22.4503
d4Wins: 32

weights

v1a: 7.0344
0.4704
0.7093
-0.1625
0.7084
0.3211
v1b: 6.5755
0.4717
0.7130
0.4958
0.6286
0.3416
v1c: 6.3034
0.4691
0.7101
0.4527
0.5080
0.3762
v2-pergame: 126567372.3647
5226833.5683
10255805.6734
32348357.1068
1063318289.7717
-983385274.1064
v4-const1e-3: 7.4074
0.4672
0.6134
1.1831
0.4384
0.3551
posthoc-v6-const1e-2: 7.1910
0.2281
0.4336
1.5539
0.5222
0.3749
posthoc-v7-const1e-4: 7.4328
0.4624
0.7065
1.2737
0.4309
0.3935
posthoc-v8-score17k-const1e-3: 1.3699
0.0846
0.0970
0.2299
0.0695
0.0704
transplant-py-seed10: 175.6754
33.0892
123.9736
15.4247
49.1437
138.6679

trainingCurvesBlockMeanMoves1000

v5-score17k-harmonic: 22.2990
21.5870
21.7190
21.4250
21.6550
21.5290
21.3720
21.8810
21.8690
21.4430
21.3590
21.8040
21.6100
22.0340
21.5460
21.6150
21.5870
21.4220
21.2310
21.5340
21.8770
21.9060
21.8270
21.7440
21.6990
21.6000
21.5990
21.1850
21.8260
21.5640
21.5210
21.5300
21.8270
21.5400
21.5750
21.9170
21.4920
21.5970
22.0120
21.4820
21.5670
21.1960
21.5680
21.4960
21.7780
21.9400
21.8530
22.0520
21.3950
21.5290
posthoc-v7-const1e-4: 36.4550
36.8730
36.0140
35.2720
35.0300
35.2500
35.4270
35.1680
35.5600
35.3180
35.2680
35.3840
35.5990
35.2640
35.5260
35.5790
35.3560
35.4360
35.8020
35.3300
35.3970
35.4760
35.2750
35.7770
35.4460
35.5510
35.5670
35.6260
35.2620
35.2210
35.7880
35.7420
35.1930
35.2690
35.5190
35.8220
35.2030
35.7490
35.4810
35.7040
36.0420
35.4210
35.4840
35.6430
35.8060
35.6460
35.4820
35.8280
35.5670
35.5950
v2-pergame: 31.5620
32.1420
31.9770
32.1680
31.8780
32.2930
32.2390
31.9760
32.4500
31.9690
32.3960
32.3120
32.1990
32.0360
32.4150
32.3660
32.3620
32.1810
32.0310
32.2110
32.2900
32.2540
32.5440
31.7500
32.4680
32.2510
32.4100
32.0370
32.3160
32.3600
32.3370
32.3810
32.1660
32.0160
32.3000
32.6340
32.1630
32.7110
32.5280
32.8660
32.5430
32.3720
32.3780
32.6200
32.3320
32.2830
32.3740
32.2950
32.0610
32.2300
v1a: 34.8850
34.8640
35.1690
35.2730
34.6530
34.9720
35.1330
35.2400
35.4180
35.2660
35.1380
35.2890
35.3570
34.9750
35.4110
35.9090
35.4590
35.3850
35.6590
35.4710
35.6350
35.3520
35.2220
35.2880
35.3690
35.1310
35.0870
35.4250
35.4530
35.2930
35.0870
35.3890
35.2540
35.3830
35.1140
35.7310
35.4460
35.5860
35.3780
35.4210
35.5900
35.4670
35.4820
35.4050
35.5140
35.5940
35.5810
35.8580
35.5810
35.7830
posthoc-v6-const1e-2: 34.7370
35.3450
35.0770
34.9830
34.6560
35.3540
35.4830
35.4480
35.5730
35.5700
35.2320
35.3380
35.3710
35.1510
35.1650
35.4920
35.5970
35.3640
35.6800
35.6650
35.3510
35.4140
35.0580
35.3370
35.4400
35.3710
35.4590
35.2130
35.7250
35.2840
35.2940
35.4960
35.2040
35.0710
35.5120
35.5020
35.1400
35.5820
35.5560
35.6190
35.3750
35.4760
35.2550
35.4340
35.6050
35.6520
35.4510
35.7310
35.3130
35.9730
posthoc-v8-score17k-const1e-3: 35.3920
34.8160
35.4530
35.2660
35.0880
35.2730
35.3330
35.6780
35.6990
35.2990
35.0920
35.1460
35.7020
35.1780
35.7200
35.3760
35.5200
35.3540
35.8050
35.5810
35.4930
35.5580
35.3090
35.5670
35.4830
35.5880
35.2110
35.5700
35.3500
35.4170
35.7970
35.5360
35.2590
35.1880
35.4480
35.8030
35.3900
35.7390
35.5780
35.6390
35.6780
35.4940
35.4810
35.4760
35.6100
35.9190
35.4160
35.8610
35.4130
35.6230
v3-harmonic50k: 21.9900
21.5810
21.5340
21.6600
21.7400
21.5320
21.5130
21.7410
21.8550
21.5940
21.3980
21.8040
21.6100
22.0340
21.5460
21.6150
21.5870
21.4220
21.2310
21.5340
21.8770
21.9060
21.8270
21.7440
21.6990
21.6000
21.5990
21.1850
21.8260
21.5640
21.5210
21.5300
21.8270
21.5400
21.5750
21.9170
21.4920
21.5970
22.0120
21.4820
21.5670
21.1960
21.5680
21.4960
21.7780
21.9400
21.8530
22.0520
21.3950
21.5290
v4-const1e-3: 35.1620
34.9480
35.5220
35.3230
35.2070
35.1750
35.5780
35.3380
35.8210
35.2100
35.1310
35.3010
35.4360
35.3810
35.2980
35.5360
35.7920
35.3820
35.9140
35.2120
35.4270
35.4710
35.1900
35.7070
35.5260
35.1990
35.4830
35.7340
35.4810
35.3790
35.3610
35.4240
35.3600
34.9330
35.6030
35.7910
35.2650
35.8700
35.6810
35.6480
35.6560
35.2470
35.4290
35.5590
35.7430
35.8470
35.6640
35.5780
35.8590
35.7290
v1c: 34.7620
34.8960
35.3960
35.0010
34.8170
35.2900
35.3340
35.2680
35.6010
35.3860
35.3300
35.5370
35.5810
35.2390
35.1770
35.4980
35.4330
35.4460
35.7510
35.2680
35.3230
35.5220
35.3360
35.5010
35.3270
35.4890
35.5480
35.8610
35.5110
35.3850
35.4910
35.4470
35.4000
35.3000
35.3900
35.7330
35.0550
35.7390
35.4280
35.5350
35.8630
35.3420
35.4750
35.5590
35.4250
35.6050
35.7260
35.6770
35.6490
35.4320
v1b: 35.0950
34.8600
34.9520
35.2620
34.7040
34.9850
35.4870
34.9900
35.4220
35.6580
35.0960
35.5360
35.3510
35.0830
35.3010
35.4570
35.2570
35.1790
35.6630
35.3170
35.2980
35.2910
35.0980
35.3550
35.4240
35.3530
35.5530
35.4280
35.0770
35.5290
35.6670
35.6580
35.4350
35.4290
35.5070
35.4960
35.4280
35.4050
35.5890
35.5960
35.5210
35.3350
35.5830
35.2630
35.1670
35.6320
35.6220
35.7080
35.3300
35.7490

trainingMovesPerSecond

v5-score17k-harmonic: 1792800.4042
posthoc-v7-const1e-4: 1066290.2992
v2-pergame: 1357777.7133
v1a: 1379994.5962
posthoc-v6-const1e-2: 949765.5085
posthoc-v8-score17k-const1e-3: 1052385.4211
v3-harmonic50k: 1798252.7015
v4-const1e-3: 1368570.7129
v1c: 1369640.3359
v1b: 1374364.8756

ledgerReferenceD4Broad64

308295.5780

bootstrap

resamples: 10,000
rngSeedHex: 0x6b660001
method: percentile, one-sided 95% lower bound

Limitations

Pilot tier on previously evaluated development seeds: the cohort had been read by the centre policy before and can never become confirmation evidence; nothing here advances a benchmark tier.
Training re-read the first 50,000 seeds of an already-open training-role block without a new lease record; the coordinator has not yet confirmed that reuse. Any citation outside the pilot tier waits on that decision.
The fair d4s7 arm covers 32 of the 256 seeds (237.6 s per game on the laptop); its 32-game deltas have detection floors near 82,000 points and 22 moves.
Three arms (constant 0.01, constant 0.0001, score reward at constant 0.001) were added post hoc after the preregistered arms were trained; they are labelled post hoc everywhere and decide no gate.
The G6 outcome is recorded as observed: the constant-0.001 arm's +0.50 moves clears its LB95 but equals the identical-configuration seed spread; the theory's clause (iv) is therefore judged not supported at the letter and the theory assessment is 'mixed'.
Wall times were measured while other arms and the concurrent reproduction run shared the 12-core laptop; the 1.4 million moves per second figure is single-process throughput and not a clean benchmark.
The rules engine is the repository's; board-level parity with the upstream simulator is not claimed, only feature and update parity.

Open the result record

Agent contextHow to extend this record

To add a reader-facing explanation, write web/content/research/EX-20260902-kf-linear-q-rust-transfer-4328a730.mdx; it renders above this record on the next request. The registered protocol itself is in the technical record above.

Record file: research/experiments/EX-20260902-kf-linear-q-rust-transfer-4328a730.json, validated against research/schemas/experiment-v1.schema.json. Protocol hash: c1080a23a5bf006b72c3edb58b663bbde9d6bc8a4a858e88635c9a4fd6fd1348.

Registered by Claude Code / claude-fable-5-1 (claude-q-learning).