Drop7 Research

Type to search every page, record and vocabulary term.

GitHubPlay

Fill-conditioned tables warm-started from the frozen n-tuple leaf, beside an unconditioned continuation and two no-training edits of the frozen tables, screened once on a fresh 512-game block at depth 3 and depth 4

Candidate fill-d3s7: the stock fair expectimax search at depth 3, seven chance strata, terminal utility -1,000,000, policy seed 0xd7075eed, completion-guaranteeing work bound work_bound_for(3,7)+1, 64k-entry direct-mapped table (deployment_params, the configuration every n-tuple screen used, unchanged), with a fill-conditioned n-tuple leaf: the sum, over the 74 active patterns of the reflection-canonical board, of learned f32 entries in rise units (x 17,000 points) from the layout rows,cols,win23,win32,phase=all,fill=occ5 or rows,cols,win23,win32,phase=all,fill=hgt5 (the first experiment's layout with every table multiplied by five fill buckets: occ5 buckets the occupied cells of the canonical board 0-13 / 14-20 / 21-27 / 28-34 / 35-49, hgt5 the tallest column 0-3 / 4 / 5 / 6 / 7; 5 x 10^9 entries, 20 GB frozen, 60 GB trainable; the bucket reads the board alone).

completedtier: SCREENalgorithmicpublic-developmentpublic-policyEX-20260906-ntuple-fill-conditioned-continuation-a9e5cbd3
On this page
Linked approachRow and column lookup tables, learned from a billion moves
Theories
Results
Dates
Created
Updated
Record idEX-20260906-ntuple-fill-conditioned-continuation-a9e5cbd3

No explanation has been written for this record yet.

Technical recordThe registered protocolEX-20260906-ntuple-fill-conditioned-continuation-a9e5cbd3
Hypothesis
Candidate fill-d3s7: the stock fair expectimax search at depth 3, seven chance strata, terminal utility -1,000,000, policy seed 0xd7075eed, completion-guaranteeing work bound work_bound_for(3,7)+1, 64k-entry direct-mapped table (deployment_params, the configuration every n-tuple screen used, unchanged), with a fill-conditioned n-tuple leaf: the sum, over the 74 active patterns of the reflection-canonical board, of learned f32 entries in rise units (x 17,000 points) from the layout rows,cols,win23,win32,phase=all,fill=occ5 or rows,cols,win23,win32,phase=all,fill=hgt5 (the first experiment's layout with every table multiplied by five fill buckets: occ5 buckets the occupied cells of the canonical board 0-13 / 14-20 / 21-27 / 28-34 / 35-49, hgt5 the tallest column 0-3 / 4 / 5 / 6 / 7; 5 x 10^9 entries, 20 GB frozen, 60 GB trainable; the bucket reads the board alone). Weights: the frozen tables of RUN-20260905T193006Z-4fbeb4e5 (SHA-256 0ade9d4e4080ebdd52a1474b1a13410dc8dfb77f5eba24b078aa7703c92ace0b) promoted into every bucket (Model::promote, bit-identical values to the frozen tables before training) with zeroed coherence accumulators, then trained on-policy by TD(0) with temporal-coherence step sizes (alpha 1.0 x |E|/A, error clamped to +-30 rise units) from complete games of the greedy one-ply chance-state policy over the same tables (seven stratified reveal samples, epsilon 0) on the Rust bitboard engine, 32 asynchronous lock-free workers (Hogwild, not bit-reproducible across runs, disclosed; the gate proves the serial update deterministic), reading the training block in order and wrapping. Each fill arm validates every 2 x 10^8 moves on the 256-game training-role block (ntuple-d3s7 and ntuple-1ply against fair-d3s7 on identical seeds) and stops at the first point k >= 6 at which the mean margin of the last three points is not above the mean of the three before, or at 2 x 10^9 moves, or at 5,400 s of wall time; each arm's best point (largest paired margin of ntuple-d3s7 over fair-d3s7) is its frozen table file. Selection rule, fixed here: the candidate is the fill arm (occ5 or hgt5) whose best validation margin is larger, ties to occ5. Comparator prior-d3s7: the unchanged frozen tables (SHA-256 0ade9d4e4080ebdd52a1474b1a13410dc8dfb77f5eba24b078aa7703c92ace0b) as the leaf of the identical search. Beside them on the same seeds: control-d3s7, the frozen tables reloaded in their own layout with zeroed accumulators and continued under the identical training rule, its best point frozen the same way (the effect of continued training without conditioning); zeroed-d3s7 and classmean-d3s7, two no-training edits of the frozen tables in which every entry still holding the optimistic starting value bit for bit (20/74 rise units) is replaced by 0.0, or by the mean of the touched entries of the same table, phase slab and pattern-occupancy class (scripts/edit-tables.py; the edit report counts the entries changed and proves no touched entry moved); prior-d4s7, fill-d4s7 and zeroed-d4s7, the same three table files as the leaf of the reference depth-4 search (reference_d4_params: depth 4, seven strata, 1M-entry table); prior-1ply and fill-1ply, the tables played directly one ply; and fair-d3s7, the frozen fair leaf in the deployment search. Stage 0 (CHECK, no leased seed): gate --layout <each fill layout> on the already-opened probe block 0xa5277000 (codec, row gather, feature indices for fourteen layouts including four fill-conditioned ones against an independent accessor reference, fill buckets against an accessor reference on the states and their mirrors with every bucket reached, promotion preserving values bit for bit with fresh accumulators and separable buckets, information boundary, reflection, direct-policy legality, leaf-in-search determinism and worker independence at depth 3 and depth 4, serial training determinism, finiteness), then gate --weights on each of the four new table files before the screen lease opens. Hypothesis: fill-d3s7 beats prior-d3s7 in paired mean whole-game score on 512 never-read public-development games, and beats control-d3s7 on the same games.
Arms
ArmNameEntry pointManifest
Candidatefill-d3s7approaches/ntuple-rl/ntuple-scale/src/game.rs–
Comparatorprior-d3s7approaches/ntuple-rl/ntuple-scale/src/game.rs–
Classification
algorithmic
Information boundary
public-policy
Benchmark tier
SCREEN
Lifecycle
completed
Theories tested
Primary metric
paired mean whole-game score delta, fill-d3s7 minus prior-d3s7 (the selected fill-conditioned candidate minus the unchanged frozen tables, both as the depth-3 leaf), 512 held-out games
Secondary metrics
  • conditioning: fill-d3s7 minus control-d3s7, with the fixed three-way verdict supported (bootstrap and Student-t 95% lower bounds > 0) / refuted (bootstrap 95% upper bound < 0) / inconclusive, floor reported
  • continuation: control-d3s7 minus prior-d3s7 under the same four criteria as the gate and the same three-way verdict
  • optimism edits (step 1): zeroed-d3s7 minus prior-d3s7 and classmean-d3s7 minus prior-d3s7, each with the three-way verdict; and each against fair-d3s7
  • depth: fill-d4s7 minus prior-d4s7 and zeroed-d4s7 minus prior-d4s7 (three-way verdicts); the depth steps fill-d4s7 minus fill-d3s7, prior-d4s7 minus prior-d3s7 and zeroed-d4s7 minus zeroed-d3s7
  • direct play: fill-1ply minus prior-1ply, fill-1ply minus fair-d3s7, prior-1ply minus fair-d3s7 (diagnostic: does conditioning change the one-ply policy)
  • replication context: prior-d3s7 minus fair-d3s7 on this fifth block; fill-d3s7 minus fair-d3s7; fill-d4s7 minus fair-d3s7
  • training-signal check (pilot tier): each arm's validation curve on the 256-game block (margin of ntuple-d3s7 over fair-d3s7 per point, best point, plateau reading, touched entries, moves per second) and whether either fill arm's best margin exceeds the control arm's
  • paired mean moves delta; numbered clears per move and cover reveals per move; mean occupied cells; lower quartile, median, maximum score; first-half (seeds 0-255) and second-half (256-511) paired mean deltas; per-arm logical work and wall seconds per game
  • edit report: entries at the starting value in the frozen file, entries changed by each edit, classes with no touched entry, and (from the checkpoint) touched entries that sit exactly at the starting value
  • resource observations: per-stage wall, CPU and peak resident set (rusage.jsonl)
Statistical unit
whole-game
Uncertainty method
one-sided 95% percentile bootstrap over whole games, 20,000 resamples, RNG seed 0xb0071eaf (the unchanged compare.py of the leaf-evolution screen; analyze.py's identical paired()), plus a one-sided 95% Student-t lower bound; detection floor 1.645*sd/sqrt(n) reported for every contrast
Data role
public-development
Seed leases
  • SL-20260906T201104Z-88b984ce
  • SL-20260906T201104Z-26371f8b
  • SL-20260906T201104Z-53350936
Dataset references
  • DS-20260906-ntuple-scale-frozen-tables-ff977178
Whole-origin split
yes
Reuse disclosure
CHECK gates read only the already-opened development probe block 0xa5277000 (SEEDLEASE-A52-FAST); the two table edits read no seed at all (the frozen table file and the checkpoint's accumulators). Training reads the never-read training block 0xa5800000-0xa59f0000 (training role, read in order and wrapped, the same block for all three arms) and the never-read 256-game validation block 0xa52f2780-0xa52f2880 (training role, re-read at every validation point of every arm; it selects each arm's best point and the candidate arm, so it can confirm nothing). The held-out screen opens 0xa52f2880-0xa52f2a80 (512 games) exactly once, after the SHA-256 of every table file it plays is on disk and the gates have passed on each new file. All three blocks are disjoint from every block the earlier n-tuple experiments read (training 0xa5300000-0xa54f0000 and 0xa5500000-0xa56f0000, validation 0xa52f2240-0xa52f2280 and 0xa52f2280-0xa52f2380, screens 0xa52f2140-0xa52f2240, 0xa52f2380-0xa52f2580 and 0xa52f2580-0xa52f2780), and from the historical block 0xa5700000-0xa571869f. The frozen tables were trained on 0xa5300000-0xa54f0000 and selected on 0xa52f2240; nothing here re-reads those blocks. Selection bias: the candidate and control are each the best of up to ten validation points of their arm, and the candidate arm is the better of two, all on the training-role validation block; the screen measures what survives that selection.
Pass criteria
  1. All CHECK gates passed on the probe block for both fill layouts before any leased seed was read (gates.log, gates-hgt5.log), and gate --weights passed on each of the four new table files (candidate, control, zeroed, classmean) before the screen lease opened (main/gates-*.log).
  2. The screen artifact has illegalDecisions 0 and incompleteDecisions 0 in every arm; censored games are reported as censored.
  3. Held-out screen, 512 paired games on 0xa52f2880-0xa52f2a7f, fill-d3s7 vs prior-d3s7: bootstrap 95% lower bound of the paired score delta > 0 AND Student-t 95% lower bound > 0.
  4. Held-out screen: fill-d3s7 minus prior-d3s7 paired mean score delta > 0 in both halves (seeds 0-255 and 256-511).
  5. Held-out screen: fill-d3s7 lower-quartile score >= prior-d3s7 lower-quartile score.
  6. The prior arms are the exact frozen best-weights.bin of RUN-20260905T193006Z-4fbeb4e5 (SHA-256 0ade9d4e4080ebdd52a1474b1a13410dc8dfb77f5eba24b078aa7703c92ace0b, verified and written to main/prior-weights.sha256 before any stage that reads it); the candidate, control, zeroed and class-mean files have their SHA-256 written to main/*.sha256 before the screen lease opens, and the candidate arm is the one selection.json names by the fixed rule.
  7. Reported beside the gate, not part of it: the conditioning verdict (fill-d3s7 vs control-d3s7), the continuation reading (control-d3s7 vs prior-d3s7 under the same four criteria), the two edit verdicts (zeroed-d3s7 and classmean-d3s7 vs prior-d3s7), the two depth-4 verdicts (fill-d4s7 and zeroed-d4s7 vs prior-d4s7), each 'supported' when the bootstrap and Student-t 95% lower bounds are > 0, 'refuted' when the bootstrap 95% upper bound is < 0, 'inconclusive' otherwise (floor reported); and the pilot-tier training-signal check (a fill arm's best validation margin above the control arm's).
On pass
Adopt fill-d3s7 as the candidate to carry forward at depth 3, and fill-d4s7 at depth 4 when its depth-4 verdict is 'supported'; record the conditioning verdict (a pass with an inconclusive or refuted conditioning verdict is recorded as a gain the screen cannot attribute to the buckets); register as successors a seven-bucket study and a STANDARD or fresh-development evaluation of the unchanged candidate; do not open protected or final seeds.
On fail
Record a valid run with scientific outcome fail for this exact configuration (five fill buckets warm-started from the frozen tables under this training rule); keep prior-d4s7 as the candidate to carry forward; report the conditioning, continuation, edit and depth readings as recorded, including any edit or continuation arm that passes its own four criteria (which is a finding beside the gate, not a candidate promotion: promoting it needs a new experiment whose gate names it); mark the theory not supported as tested (refuted if the primary falsifier's upper bound is below zero, inconclusive if the bounds straddle zero); open no further cohort for this candidate.
Gate fixed before controlled data
yes
Resources
Wall seconds36000
CPU threads32
Max host bytes85899345920
Max GPU bytes–
GPU devices–
Stop conditions
  1. Stop on any rules, information-boundary, legality, determinism, or parity failure.
  2. Each training arm stops at its plateau rule (window 3 from the sixth validation point), at 2 x 10^9 moves, or at 5,400 s of wall time, whichever comes first; an arm stopped by wall time keeps its best point and the fact is recorded.
  3. A screen artifact with any illegal or incomplete decision voids the run (invalid), not the candidate.
  4. The screen is evaluated exactly once; no re-run on the same or a different held-out block without a new experiment record. The screen binary rewrites the artifact after every completed arm; an interrupted screen is recorded as interrupted with the arms that completed, and the gate is not evaluated on a partial artifact.
  5. The whole run is stopped at 36,000 s of wall time; such a run is recorded as interrupted with whatever stages completed.
  6. Host memory above the declared bound (80 GiB resident) aborts the stage through the operating system or the operator; such a run is recorded as interrupted with its artifacts. A STOP file in a training arm's directory ends that arm at the next chunk (recorded as stop-file).
Expected artifacts
  • runs/<run-id>/ntuple-scale/gates.log and gates-hgt5.log (CHECK, both fill layouts, probe block only)
  • runs/<run-id>/ntuple-scale/main/{zeroed,classmean}-weights.bin (+ .sha256), edits.json, edit.log (the two no-training edits; the table files are not committed)
  • runs/<run-id>/ntuple-scale/pilot/{control,occ5,hgt5}/{config.json,progress.jsonl,val-*.json,best.json,best-weights.bin,stop.json,train.log,train.err,DONE} (the three warm-started arms; table files not committed)
  • runs/<run-id>/ntuple-scale/pilot/selection.json (the fixed selection rule's output)
  • runs/<run-id>/ntuple-scale/main/{prior,candidate,control,zeroed,classmean}-weights.sha256 and main/gates-{candidate,control,zeroed,classmean}.log (freeze)
  • runs/<run-id>/ntuple-scale/screen/heldout.json + compare-*.json (the one-shot screen, eleven arms)
  • runs/<run-id>/ntuple-scale/{pipeline.log,rusage.jsonl,screen.log,screen.err,analysis.json,analysis.md}
  • artifacts emitted by the trainer and the screen carry this experiment id in their config field; the frozen candidate table file (20 GB) is retained on the workstation with its SHA-256 in the result record and published compressed to the research archive
Amendments
TimestampBefore controlled dataReason
2026-09-06T23:59:02Znolifecycle advanced preregistered -> completed after run RUN-20260906T201104Z-a96ea6c8 and result RS-20260906T234914Z-a3fae1a9 were written; protocol content otherwise unchanged (frozen protocol SHA-256 974bdae371b7dab5835003c8b62e26b4ab303f01761e5846880c6b10eee1aa14, the hash every lease and the run record cite)
Technical recordResults recorded against this protocol1 record
valid runoutcome: failmixedtier: public-developmentRS-20260906T234914Z-a3fae1a9

Held-out screen, 512 never-read paired public-development games (0xa52f2880+), 11 arms on identical seeds. Candidate arm occ5 (best validation margin 203,193 at 1,400,252,568 moves against the control arm's best 210,990; training-signal check passed False). The fill-conditioned tables (SHA-256 4f2e7ccf5c14fed8dd19563965e3937e8784b487f2de9eb70fe0e86830edae24) as the depth-3 leaf averaged 506,494 points and 146.33 moves against 485,455 and 140.38 for the unchanged frozen tables (SHA-256 0ade9d4e4080ebdd52a1474b1a13410dc8dfb77f5eba24b078aa7703c92ace0b) in the same search: fill-d3s7 minus prior-d3s7: paired +21,039 (bootstrap 95% lower bound -16,864, Student-t lower bound -16,982, upper bound +59,230, detection floor 37,956), W-T-L 271-0-241, halves +6,466 / +35,613, lower quartile 229,450 vs 230,374, moves +5.95. The preregistered gate FAILS. Conditioning, fill-d3s7 minus control-d3s7: paired +15,093 (bootstrap 95% lower bound -22,431, Student-t lower bound -22,479, upper bound +52,817, detection floor 37,507), W-T-L 265-2-245, halves +21,802 / +8,385: verdict 'inconclusive'. Continuation, control-d3s7 minus prior-d3s7: paired +5,946 (bootstrap 95% lower bound -29,891, Student-t lower bound -29,937, upper bound +41,989, detection floor 35,821), W-T-L 255-6-251, halves -15,336 / +27,228: verdict 'inconclusive' (four criteria FAIL). Zeroed edit, zeroed-d3s7 minus prior-d3s7: paired +29,442 (bootstrap 95% lower bound +5,813, Student-t lower bound +5,246, upper bound +54,179, detection floor 24,154), W-T-L 102-318-92, halves +12,168 / +46,716: verdict 'supported'. Class-mean edit, classmean-d3s7 minus prior-d3s7: paired +931 (bootstrap 95% lower bound -530, Student-t lower bound -869, upper bound +2,954, detection floor 1,797), W-T-L 4-506-2, halves -690 / +2,551: verdict 'inconclusive'. At depth 4, fill-d4s7 minus prior-d4s7: paired -22,631 (bootstrap 95% lower bound -57,779, Student-t lower bound -58,306, upper bound +12,806, detection floor 35,613), W-T-L 236-0-276, halves +37,523 / -82,786: verdict 'inconclusive'. At depth 4, zeroed-d4s7 minus prior-d4s7: paired -10,592 (bootstrap 95% lower bound -32,045, Student-t lower bound -32,160, upper bound +11,133, detection floor 21,531), W-T-L 93-320-99, halves -16,117 / -5,066: verdict 'inconclusive'. fill-d3s7 506,494 / 146.33 moves; control-d3s7 491,401 / 141.92 moves; zeroed-d3s7 514,897 / 148.50 moves; classmean-d3s7 486,386 / 140.64 moves; prior-d3s7 485,455 / 140.38 moves; prior-d4s7 521,956 / 150.20 moves; fill-d4s7 499,324 / 144.08 moves; zeroed-d4s7 511,364 / 147.36 moves; fill-1ply 298,199 / 88.90 moves; prior-1ply 293,390 / 87.53 moves; fair-d3s7 329,895 / 96.85 moves; fill-d4s7-vs-fill-d3s7: -7,170 (LB -45,566, UB +31,331). prior-d4s7-vs-prior-d3s7: +36,500 (LB -249, UB +72,651). zeroed-d4s7-vs-zeroed-d3s7: -3,534 (LB -43,880, UB +36,834). prior-d3s7-vs-fair-d3s7: +155,561 (LB +123,422, UB +187,991). fill-d3s7-vs-fair-d3s7: +176,600 (LB +143,625, UB +210,258). fill-1ply-vs-prior-1ply: +4,809 (LB -13,189, UB +22,386). Edits: 901,259,321 of 1,000,000,000 entries of the frozen file sit at the starting value 0.27027 rise units; the zeroed edit changed 901,259,321 entries and the class-mean edit 877,344,474, none of them touched entries. Arm control: 1,200,228,265 moves, 6 validation points, best margin 210,990 at 600,112,060 moves, final 197,769, stop plateau, 2,017,917 moves/s. Arm occ5: 1,600,289,700 moves, 8 validation points, best margin 203,193 at 1,400,252,568 moves, final 150,272, stop plateau, 2,125,289 moves/s. Arm hgt5: 2,000,004,221 moves, 10 validation points, best margin 179,760 at 1,600,283,262 moves, final 157,846, stop plateau, 2,153,323 moves/s. Logical work per game: fill-d3s7 24,138,811, prior-d3s7 23,117,640, fill-d4s7 764,278,797, prior-d4s7 803,086,972. Theory falsifiers: {"primaryFalsifierUpperBoundBelowZero": false, "conditioningUpperBoundBelowZero": false, "gainIsContinuationNotConditioning": false, "optimismLegRefuted": false, "depthCompounding": false}.

What it had to pass
  • ✓All CHECK gates passed on the probe block for both fill layouts before any leased seed was read — observed: 27 + 27 gate lines
  • ✓gate --weights passed on each of the four new table files before the screen lease opened — observed: {"candidate": true, "control": true, "zeroed": true, "classmean": true}
  • ✓screen artifact: illegalDecisions 0 and incompleteDecisions 0 in every arm — observed: null
  • ✕bootstrap 95% lower bound of fill-d3s7 minus prior-d3s7 > 0 — observed: -16863.925097656247
  • ✕Student-t 95% lower bound > 0 — observed: -16982.313672978897
  • ✓paired mean delta > 0 in both halves — observed: [6465.9375, 35612.62109375]
  • ✕fill-d3s7 Q25 >= prior-d3s7 Q25 — observed: [229449.5, 230374.0]
  • ✓The prior arms are the exact frozen best-weights.bin of RUN-20260905T193006Z-4fbeb4e5 (SHA-256 verified before any stage read it); every other table file's SHA-256 was written before the screen lease opened — observed: {"prior": "0ade9d4e4080ebdd52a1474b1a13410dc8dfb77f5eba24b078aa7703c92ace0b", "candidate": "4f2e7ccf5c14fed8dd19563965e3937e8784b487f2de9eb70fe0e86830edae24", "control": "92dd1cb2d2a74b026270606c18c5f0d6e4f4ccc74e64c3c4e0f0d2043cdddd90", "zeroed": "e9248b1a26f0b6e2d6cac121df9eda48c1adf1ea369287a60546a830b65c058e", "classmean": "c3f05eeebb6ed573ba176797e7cfa4bdab3b47e042c8877011dd6614ebb54450"}
  • –Beside the gate: conditioning verdict (fill-d3s7 vs control-d3s7): supported / refuted / inconclusive — observed: {"verdict": "inconclusive", "meanDelta": 15093.380859375, "bootstrapLower95": -22430.81123046875, "bootstrapUpper95": 52816.74423828125}
  • –Beside the gate: continuation reading (control-d3s7 vs prior-d3s7): supported / refuted / inconclusive — observed: {"verdict": "inconclusive", "meanDelta": 5945.8984375, "bootstrapLower95": -29890.52666015625, "bootstrapUpper95": 41989.1677734375}
  • –Beside the gate: zeroed edit verdict (vs prior-d3s7): supported / refuted / inconclusive — observed: {"verdict": "supported", "meanDelta": 29442.16796875, "bootstrapLower95": 5812.94228515625, "bootstrapUpper95": 54178.72646484375}
  • –Beside the gate: class-mean edit verdict (vs prior-d3s7): supported / refuted / inconclusive — observed: {"verdict": "inconclusive", "meanDelta": 930.701171875, "bootstrapLower95": -530.3950195312499, "bootstrapUpper95": 2954.373046875}
  • –Beside the gate: fill at depth 4 verdict (vs prior-d4s7): supported / refuted / inconclusive — observed: {"verdict": "inconclusive", "meanDelta": -22631.439453125, "bootstrapLower95": -57778.848046875, "bootstrapUpper95": 12805.626269531243}
  • –Beside the gate: zeroed at depth 4 verdict (vs prior-d4s7): supported / refuted / inconclusive — observed: {"verdict": "inconclusive", "meanDelta": -10591.814453125, "bootstrapLower95": -32044.523046875, "bootstrapUpper95": 11133.006640624997}
  • –Beside the gate: training-signal check (a fill arm's best validation margin exceeds the control arm's) — observed: {"criterion": "a fill arm's best validation margin exceeds the control arm's best validation margin", "passed": false}
Technical recordRecorded metricsRS-20260906T234914Z-a3fae1a9
screen
games
512
seedStartHex
0xa52f2880
arms
prior-d3s7
meanScore
485455.1816
medianScore
358256.5000
q25Score
230,374
minScore
85,609
maxScore
3,329,202
sdScore
385158.7154
meanMoves
140.3789
q25Moves
70
numberedClearsPerMove
2.1097
coverRevealsPerMove
1.1941
meanOccupiedCells
23.0535
censoredGames
0
illegalDecisions
0
incompleteDecisions
0
meanWork
23117640.2012
meanWallSecondsPerGame
4.3154
games
512
prior-1ply
meanScore
293390.1543
medianScore
238,459
q25Score
157,932
minScore
85,575
maxScore
1,706,990
sdScore
188891.6277
meanMoves
87.5273
q25Moves
50
numberedClearsPerMove
1.9310
coverRevealsPerMove
1.0638
meanOccupiedCells
24.4089
censoredGames
0
illegalDecisions
0
incompleteDecisions
0
meanWork
1465.6152
meanWallSecondsPerGame
0.0009
games
512
fill-d3s7
meanScore
506494.4609
medianScore
388,587
q25Score
229449.5000
minScore
85,362
maxScore
2,751,523
sdScore
410994.7149
meanMoves
146.3262
q25Moves
70
numberedClearsPerMove
2.1216
coverRevealsPerMove
1.2010
meanOccupiedCells
23.0883
censoredGames
0
illegalDecisions
0
incompleteDecisions
0
meanWork
24138811.2891
meanWallSecondsPerGame
4.6583
games
512
fill-1ply
meanScore
298199.3203
medianScore
262,235
q25Score
190083.7500
minScore
85,521
maxScore
1,165,434
sdScore
165744.9062
meanMoves
88.9043
q25Moves
56
numberedClearsPerMove
1.9392
coverRevealsPerMove
1.0653
meanOccupiedCells
24.5045
censoredGames
0
illegalDecisions
0
incompleteDecisions
0
meanWork
1491.3867
meanWallSecondsPerGame
0.0010
games
512
control-d3s7
meanScore
491401.0801
medianScore
389,192
q25Score
229196.7500
minScore
85,582
maxScore
2,662,558
sdScore
359048.6893
meanMoves
141.9160
q25Moves
70
numberedClearsPerMove
2.1128
coverRevealsPerMove
1.1945
meanOccupiedCells
22.9919
censoredGames
0
illegalDecisions
0
incompleteDecisions
0
meanWork
23405501.8633
meanWallSecondsPerGame
4.4364
games
512
zeroed-d3s7
meanScore
514897.3496
medianScore
370,697
q25Score
229884.5000
minScore
85,609
maxScore
3,088,440
sdScore
441508.8486
meanMoves
148.5039
q25Moves
70
numberedClearsPerMove
2.1262
coverRevealsPerMove
1.2060
meanOccupiedCells
23.0180
censoredGames
0
illegalDecisions
0
incompleteDecisions
0
meanWork
24556187.8555
meanWallSecondsPerGame
4.6071
games
512
classmean-d3s7
meanScore
486385.8828
medianScore
358256.5000
q25Score
230,374
minScore
85,609
maxScore
3,329,202
sdScore
386411.0006
meanMoves
140.6387
q25Moves
70
numberedClearsPerMove
2.1098
coverRevealsPerMove
1.1940
meanOccupiedCells
23.0528
censoredGames
0
illegalDecisions
0
incompleteDecisions
0
meanWork
23163587.6777
meanWallSecondsPerGame
4.4294
games
512
prior-d4s7
meanScore
521955.5215
medianScore
393322.5000
q25Score
244331.5000
minScore
102,737
maxScore
2,469,625
sdScore
400627.3752
meanMoves
150.1992
q25Moves
73.2500
numberedClearsPerMove
2.1274
coverRevealsPerMove
1.2047
meanOccupiedCells
23.0003
censoredGames
0
illegalDecisions
0
incompleteDecisions
0
meanWork
803086972.2734
meanWallSecondsPerGame
160.8416
games
512
fill-d4s7
meanScore
499324.0820
medianScore
391661.5000
q25Score
242348.7500
minScore
85,663
maxScore
2,934,902
sdScore
364035.5921
meanMoves
144.0820
q25Moves
70
numberedClearsPerMove
2.1172
coverRevealsPerMove
1.1964
meanOccupiedCells
23.1415
censoredGames
0
illegalDecisions
0
incompleteDecisions
0
meanWork
764278797.0020
meanWallSecondsPerGame
154.5392
games
512
zeroed-d4s7
meanScore
511363.7070
medianScore
391,800
q25Score
230,407
minScore
102,737
maxScore
3,016,304
sdScore
401541.0761
meanMoves
147.3594
q25Moves
70
numberedClearsPerMove
2.1239
coverRevealsPerMove
1.2033
meanOccupiedCells
22.8962
censoredGames
0
illegalDecisions
0
incompleteDecisions
0
meanWork
784054438.0273
meanWallSecondsPerGame
155.4543
games
512
fair-d3s7
meanScore
329894.5605
medianScore
268,113
q25Score
176,683
minScore
85,660
maxScore
1,540,436
sdScore
220324.9139
meanMoves
96.8457
q25Moves
55
numberedClearsPerMove
2.0017
coverRevealsPerMove
1.1140
meanOccupiedCells
24.0911
censoredGames
0
illegalDecisions
0
incompleteDecisions
0
meanWork
15319526.0410
meanWallSecondsPerGame
4.4282
games
512
contrasts
prior-d3s7-vs-fair-d3s7
score
n
512
meanDelta
155560.6211
pairedSd
444880.5023
bootstrapLower95
123421.7851
bootstrapUpper95
187990.6395
studentTLower95
123162.2110
detectionFloor
32342.5527
wins
329
ties
0
losses
183
firstHalfMeanDelta
175955.6602
secondHalfMeanDelta
135165.5820
q25Delta
53,691
candidateQ25
230,374
referenceQ25
176,683
moves
n
512
meanDelta
43.5332
pairedSd
122.8654
bootstrapLower95
34.6597
bootstrapUpper95
52.4708
studentTLower95
34.5855
detectionFloor
8.9322
wins
318
ties
17
losses
177
firstHalfMeanDelta
49.1016
secondHalfMeanDelta
37.9648
q25Delta
15
candidateQ25
70
referenceQ25
55
prior-1ply-vs-fair-d3s7
score
n
512
meanDelta
-36504.4063
pairedSd
276961.7893
bootstrapLower95
-56473.7448
bootstrapUpper95
-16523.1362
studentTLower95
-56674.1408
detectionFloor
20134.9603
wins
225
ties
0
losses
287
firstHalfMeanDelta
-43043.0313
secondHalfMeanDelta
-29965.7813
q25Delta
-18,751
candidateQ25
157,932
referenceQ25
176,683
moves
n
512
meanDelta
-9.3184
pairedSd
76.5082
bootstrapLower95
-14.8281
bootstrapUpper95
-3.8124
studentTLower95
-14.8901
detectionFloor
5.5621
wins
220
ties
19
losses
273
firstHalfMeanDelta
-11.2227
secondHalfMeanDelta
-7.4141
q25Delta
-5
candidateQ25
50
referenceQ25
55
prior-d4s7-vs-prior-d3s7
score
n
512
meanDelta
36500.3398
pairedSd
504798.2073
bootstrapLower95
-249.3278
bootstrapUpper95
72651.2398
studentTLower95
-261.5755
detectionFloor
36698.5348
wins
287
ties
0
losses
225
firstHalfMeanDelta
-13180.8047
secondHalfMeanDelta
86181.4844
q25Delta
13957.5000
candidateQ25
244331.5000
referenceQ25
230,374
moves
n
512
meanDelta
9.8203
pairedSd
139.4430
bootstrapLower95
-0.3223
bootstrapUpper95
19.8126
studentTLower95
-0.3346
detectionFloor
10.1374
wins
279
ties
21
losses
212
firstHalfMeanDelta
-4.0039
secondHalfMeanDelta
23.6445
q25Delta
3.2500
candidateQ25
73.2500
referenceQ25
70
prior-d4s7-vs-fair-d3s7
score
n
512
meanDelta
192060.9609
pairedSd
454969.3974
bootstrapLower95
158819.9766
bootstrapUpper95
225169.2477
studentTLower95
158927.8273
detectionFloor
33076.0095
wins
340
ties
0
losses
172
firstHalfMeanDelta
162774.8555
secondHalfMeanDelta
221347.0664
q25Delta
67648.5000
candidateQ25
244331.5000
referenceQ25
176,683
moves
n
512
meanDelta
53.3535
pairedSd
125.3499
bootstrapLower95
44.2030
bootstrapUpper95
62.4924
studentTLower95
44.2249
detectionFloor
9.1129
wins
335
ties
12
losses
165
firstHalfMeanDelta
45.0977
secondHalfMeanDelta
61.6094
q25Delta
18.2500
candidateQ25
73.2500
referenceQ25
55
fill-d3s7-vs-prior-d3s7
score
n
512
meanDelta
21039.2793
pairedSd
522095.5386
bootstrapLower95
-16863.9251
bootstrapUpper95
59229.5392
studentTLower95
-16982.3137
detectionFloor
37956.0407
wins
271
ties
0
losses
241
firstHalfMeanDelta
6465.9375
secondHalfMeanDelta
35612.6211
q25Delta
-924.5000
candidateQ25
229449.5000
referenceQ25
230,374
moves
n
512
meanDelta
5.9473
pairedSd
144.7390
bootstrapLower95
-4.5511
bootstrapUpper95
16.5039
studentTLower95
-4.5933
detectionFloor
10.5224
wins
262
ties
13
losses
237
firstHalfMeanDelta
1.9063
secondHalfMeanDelta
9.9883
q25Delta
0
candidateQ25
70
referenceQ25
70
fill-d3s7-vs-control-d3s7
score
n
512
meanDelta
15093.3809
pairedSd
515922.4838
bootstrapLower95
-22430.8112
bootstrapUpper95
52816.7442
studentTLower95
-22478.6596
detectionFloor
37507.2632
wins
265
ties
2
losses
245
firstHalfMeanDelta
21801.8164
secondHalfMeanDelta
8384.9453
q25Delta
252.7500
candidateQ25
229449.5000
referenceQ25
229196.7500
moves
n
512
meanDelta
4.4102
pairedSd
142.6044
bootstrapLower95
-5.9531
bootstrapUpper95
14.8517
studentTLower95
-5.9750
detectionFloor
10.3673
wins
258
ties
13
losses
241
firstHalfMeanDelta
6.1641
secondHalfMeanDelta
2.6563
q25Delta
0
candidateQ25
70
referenceQ25
70
control-d3s7-vs-prior-d3s7
score
n
512
meanDelta
5945.8984
pairedSd
492726.1596
bootstrapLower95
-29890.5267
bootstrapUpper95
41989.1678
studentTLower95
-29936.8703
detectionFloor
35820.9040
wins
255
ties
6
losses
251
firstHalfMeanDelta
-15335.8789
secondHalfMeanDelta
27227.6758
q25Delta
-1177.2500
candidateQ25
229196.7500
referenceQ25
230,374
moves
n
512
meanDelta
1.5371
pairedSd
136.2181
bootstrapLower95
-8.3906
bootstrapUpper95
11.5099
studentTLower95
-8.3830
detectionFloor
9.9030
wins
249
ties
27
losses
236
firstHalfMeanDelta
-4.2578
secondHalfMeanDelta
7.3320
q25Delta
0
candidateQ25
70
referenceQ25
70
zeroed-d3s7-vs-prior-d3s7
score
n
512
meanDelta
29442.1680
pairedSd
332244.5840
bootstrapLower95
5812.9423
bootstrapUpper95
54178.7265
studentTLower95
5246.4654
detectionFloor
24153.9872
wins
102
ties
318
losses
92
firstHalfMeanDelta
12167.9297
secondHalfMeanDelta
46716.4063
q25Delta
-489.5000
candidateQ25
229884.5000
referenceQ25
230,374
moves
n
512
meanDelta
8.1250
pairedSd
92.0975
bootstrapLower95
1.5876
bootstrapUpper95
14.9728
studentTLower95
1.4180
detectionFloor
6.6954
wins
97
ties
335
losses
80
firstHalfMeanDelta
3.3711
secondHalfMeanDelta
12.8789
q25Delta
0
candidateQ25
70
referenceQ25
70
classmean-d3s7-vs-prior-d3s7
score
n
512
meanDelta
930.7012
pairedSd
24714.4293
bootstrapLower95
-530.3950
bootstrapUpper95
2954.3730
studentTLower95
-869.1265
detectionFloor
1796.7246
wins
4
ties
506
losses
2
firstHalfMeanDelta
-689.9414
secondHalfMeanDelta
2551.3438
q25Delta
0
candidateQ25
230,374
referenceQ25
230,374
moves
n
512
meanDelta
0.2598
pairedSd
6.9333
bootstrapLower95
-0.1504
bootstrapUpper95
0.8263
studentTLower95
-0.2452
detectionFloor
0.5041
wins
4
ties
506
losses
2
firstHalfMeanDelta
-0.1953
secondHalfMeanDelta
0.7148
q25Delta
0
candidateQ25
70
referenceQ25
70
fill-d4s7-vs-prior-d4s7
score
n
512
meanDelta
-22631.4395
pairedSd
489868.4816
bootstrapLower95
-57778.8480
bootstrapUpper95
12805.6263
studentTLower95
-58306.0979
detectionFloor
35613.1525
wins
236
ties
0
losses
276
firstHalfMeanDelta
37523.1680
secondHalfMeanDelta
-82786.0469
q25Delta
-1982.7500
candidateQ25
242348.7500
referenceQ25
244331.5000
moves
n
512
meanDelta
-6.1172
pairedSd
135.1325
bootstrapLower95
-15.7834
bootstrapUpper95
3.6428
studentTLower95
-15.9582
detectionFloor
9.8241
wins
226
ties
19
losses
267
firstHalfMeanDelta
10.5625
secondHalfMeanDelta
-22.7969
q25Delta
-3.2500
candidateQ25
70
referenceQ25
73.2500
zeroed-d4s7-vs-prior-d4s7
score
n
512
meanDelta
-10591.8145
pairedSd
296169.0500
bootstrapLower95
-32044.5230
bootstrapUpper95
11133.0066
studentTLower95
-32160.3172
detectionFloor
21531.3170
wins
93
ties
320
losses
99
firstHalfMeanDelta
-16117.3555
secondHalfMeanDelta
-5066.2734
q25Delta
-13924.5000
candidateQ25
230,407
referenceQ25
244331.5000
moves
n
512
meanDelta
-2.8398
pairedSd
81.6541
bootstrapLower95
-8.7598
bootstrapUpper95
3.1602
studentTLower95
-8.7863
detectionFloor
5.9362
wins
81
ties
344
losses
87
firstHalfMeanDelta
-4.3711
secondHalfMeanDelta
-1.3086
q25Delta
-3.2500
candidateQ25
70
referenceQ25
73.2500
fill-d4s7-vs-fill-d3s7
score
n
512
meanDelta
-7170.3789
pairedSd
525924.1659
bootstrapLower95
-45565.7621
bootstrapUpper95
31331.3407
studentTLower95
-45470.7916
detectionFloor
38234.3797
wins
263
ties
0
losses
249
firstHalfMeanDelta
17876.4258
secondHalfMeanDelta
-32217.1836
q25Delta
12899.2500
candidateQ25
242348.7500
referenceQ25
229449.5000
moves
n
512
meanDelta
-2.2441
pairedSd
145.4102
bootstrapLower95
-12.8460
bootstrapUpper95
8.3986
studentTLower95
-12.8336
detectionFloor
10.5712
wins
252
ties
13
losses
247
firstHalfMeanDelta
4.6523
secondHalfMeanDelta
-9.1406
q25Delta
0
candidateQ25
70
referenceQ25
70
zeroed-d4s7-vs-zeroed-d3s7
score
n
512
meanDelta
-3533.6426
pairedSd
554286.6329
bootstrapLower95
-43880.1008
bootstrapUpper95
36833.7896
studentTLower95
-43899.5511
detectionFloor
40296.3145
wins
270
ties
0
losses
242
firstHalfMeanDelta
-41466.0898
secondHalfMeanDelta
34398.8047
q25Delta
522.5000
candidateQ25
230,407
referenceQ25
229884.5000
moves
n
512
meanDelta
-1.1445
pairedSd
153.1565
bootstrapLower95
-12.3069
bootstrapUpper95
10.0059
studentTLower95
-12.2981
detectionFloor
11.1344
wins
261
ties
19
losses
232
firstHalfMeanDelta
-11.7461
secondHalfMeanDelta
9.4570
q25Delta
0
candidateQ25
70
referenceQ25
70
fill-d3s7-vs-fair-d3s7
score
n
512
meanDelta
176599.9004
pairedSd
457719.0419
bootstrapLower95
143625.1853
bootstrapUpper95
210258.3875
studentTLower95
143266.5240
detectionFloor
33275.9070
wins
325
ties
0
losses
187
firstHalfMeanDelta
182421.5977
secondHalfMeanDelta
170778.2031
q25Delta
52766.5000
candidateQ25
229449.5000
referenceQ25
176,683
moves
n
512
meanDelta
49.4805
pairedSd
126.5107
bootstrapLower95
40.3710
bootstrapUpper95
58.7854
studentTLower95
40.2673
detectionFloor
9.1973
wins
312
ties
22
losses
178
firstHalfMeanDelta
51.0078
secondHalfMeanDelta
47.9531
q25Delta
15
candidateQ25
70
referenceQ25
55
control-d3s7-vs-fair-d3s7
score
n
512
meanDelta
161506.5195
pairedSd
406359.0011
bootstrapLower95
132301.3307
bootstrapUpper95
192037.2626
studentTLower95
131913.4367
detectionFloor
29542.0620
wins
324
ties
0
losses
188
firstHalfMeanDelta
160619.7813
secondHalfMeanDelta
162393.2578
q25Delta
52513.7500
candidateQ25
229196.7500
referenceQ25
176,683
moves
n
512
meanDelta
45.0703
pairedSd
112.0523
bootstrapLower95
36.9939
bootstrapUpper95
53.4768
studentTLower95
36.9101
detectionFloor
8.1461
wins
318
ties
21
losses
173
firstHalfMeanDelta
44.8438
secondHalfMeanDelta
45.2969
q25Delta
15
candidateQ25
70
referenceQ25
55
zeroed-d3s7-vs-fair-d3s7
score
n
512
meanDelta
185002.7891
pairedSd
495259.0272
bootstrapLower95
149751.9901
bootstrapUpper95
221463.2062
studentTLower95
148935.5643
detectionFloor
36005.0420
wins
334
ties
0
losses
178
firstHalfMeanDelta
188123.5898
secondHalfMeanDelta
181881.9883
q25Delta
53201.5000
candidateQ25
229884.5000
referenceQ25
176,683
moves
n
512
meanDelta
51.6582
pairedSd
136.6950
bootstrapLower95
41.9120
bootstrapUpper95
61.7228
studentTLower95
41.7034
detectionFloor
9.9376
wins
322
ties
15
losses
175
firstHalfMeanDelta
52.4727
secondHalfMeanDelta
50.8438
q25Delta
15
candidateQ25
70
referenceQ25
55
classmean-d3s7-vs-fair-d3s7
score
n
512
meanDelta
156491.3223
pairedSd
445845.4206
bootstrapLower95
124333.9894
bootstrapUpper95
188986.7944
studentTLower95
124022.6420
detectionFloor
32412.7017
wins
329
ties
0
losses
183
firstHalfMeanDelta
175265.7188
secondHalfMeanDelta
137716.9258
q25Delta
53,691
candidateQ25
230,374
referenceQ25
176,683
moves
n
512
meanDelta
43.7930
pairedSd
123.1524
bootstrapLower95
34.9253
bootstrapUpper95
52.7698
studentTLower95
34.8244
detectionFloor
8.9531
wins
318
ties
17
losses
177
firstHalfMeanDelta
48.9063
secondHalfMeanDelta
38.6797
q25Delta
15
candidateQ25
70
referenceQ25
55
fill-1ply-vs-prior-1ply
score
n
512
meanDelta
4809.1660
pairedSd
246120.8656
bootstrapLower95
-13188.6314
bootstrapUpper95
22385.6826
studentTLower95
-13114.5791
detectionFloor
17892.8432
wins
289
ties
0
losses
223
firstHalfMeanDelta
21419.3516
secondHalfMeanDelta
-11801.0195
q25Delta
32151.7500
candidateQ25
190083.7500
referenceQ25
157,932
moves
n
512
meanDelta
1.3770
pairedSd
68.4893
bootstrapLower95
-3.6311
bootstrapUpper95
6.2773
studentTLower95
-3.6108
detectionFloor
4.9791
wins
281
ties
21
losses
210
firstHalfMeanDelta
5.8711
secondHalfMeanDelta
-3.1172
q25Delta
6
candidateQ25
56
referenceQ25
50
fill-1ply-vs-fair-d3s7
score
n
512
meanDelta
-31695.2402
pairedSd
265168.8300
bootstrapLower95
-50765.8081
bootstrapUpper95
-12546.5545
studentTLower95
-51006.1529
detectionFloor
19277.6191
wins
249
ties
0
losses
263
firstHalfMeanDelta
-21623.6797
secondHalfMeanDelta
-41766.8008
q25Delta
13400.7500
candidateQ25
190083.7500
referenceQ25
176,683
moves
n
512
meanDelta
-7.9414
pairedSd
73.2659
bootstrapLower95
-13.2035
bootstrapUpper95
-2.6542
studentTLower95
-13.2770
detectionFloor
5.3264
wins
239
ties
34
losses
239
firstHalfMeanDelta
-5.3516
secondHalfMeanDelta
-10.5313
q25Delta
1
candidateQ25
56
referenceQ25
55
fill-d4s7-vs-fair-d3s7
score
n
512
meanDelta
169429.5215
pairedSd
406266.0510
bootstrapLower95
140100.4949
bootstrapUpper95
199457.9964
studentTLower95
139843.2077
detectionFloor
29535.3046
wins
339
ties
0
losses
173
firstHalfMeanDelta
200298.0234
secondHalfMeanDelta
138561.0195
q25Delta
65665.7500
candidateQ25
242348.7500
referenceQ25
176,683
moves
n
512
meanDelta
47.2363
pairedSd
112.0745
bootstrapLower95
39.1443
bootstrapUpper95
55.5177
studentTLower95
39.0745
detectionFloor
8.1477
wins
338
ties
19
losses
155
firstHalfMeanDelta
55.6602
secondHalfMeanDelta
38.8125
q25Delta
15
candidateQ25
70
referenceQ25
55
fill
primary
fill-d3s7-vs-prior-d3s7
gate
checks
  1. criterion
    screen artifact: illegalDecisions 0 and incompleteDecisions 0 in every arm
    passed
    true
  2. criterion
    bootstrap 95% lower bound of fill-d3s7 minus prior-d3s7 > 0
    passed
    false
    observed
    -16863.9251
  3. criterion
    Student-t 95% lower bound > 0
    passed
    false
    observed
    -16982.3137
  4. criterion
    paired mean delta > 0 in both halves
    passed
    true
    observed
    1. 6465.9375
    2. 35612.6211
  5. criterion
    fill-d3s7 Q25 >= prior-d3s7 Q25
    passed
    false
    observed
    1. 229449.5000
    2. 230,374
passed
false
conditioning
contrast
fill-d3s7-vs-control-d3s7
verdict
inconclusive
meanDelta
15093.3809
bootstrapLower95
-22430.8112
bootstrapUpper95
52816.7442
studentTLower95
-22478.6596
detectionFloor
37507.2632
wins
265
ties
2
losses
245
firstHalfMeanDelta
21801.8164
secondHalfMeanDelta
8384.9453
candidateQ25
229449.5000
referenceQ25
229196.7500
continuation
checks
  1. criterion
    screen artifact: illegalDecisions 0 and incompleteDecisions 0 in every arm
    passed
    true
  2. criterion
    bootstrap 95% lower bound of control-d3s7 minus prior-d3s7 > 0
    passed
    false
    observed
    -29890.5267
  3. criterion
    Student-t 95% lower bound > 0
    passed
    false
    observed
    -29936.8703
  4. criterion
    paired mean delta > 0 in both halves
    passed
    false
    observed
    1. -15335.8789
    2. 27227.6758
  5. criterion
    control-d3s7 Q25 >= prior-d3s7 Q25
    passed
    false
    observed
    1. 229196.7500
    2. 230,374
passed
false
contrast
control-d3s7-vs-prior-d3s7
verdict
inconclusive
meanDelta
5945.8984
bootstrapLower95
-29890.5267
bootstrapUpper95
41989.1678
studentTLower95
-29936.8703
detectionFloor
35820.9040
wins
255
ties
6
losses
251
firstHalfMeanDelta
-15335.8789
secondHalfMeanDelta
27227.6758
candidateQ25
229196.7500
referenceQ25
230,374
zeroed
contrast
zeroed-d3s7-vs-prior-d3s7
verdict
supported
meanDelta
29442.1680
bootstrapLower95
5812.9423
bootstrapUpper95
54178.7265
studentTLower95
5246.4654
detectionFloor
24153.9872
wins
102
ties
318
losses
92
firstHalfMeanDelta
12167.9297
secondHalfMeanDelta
46716.4063
candidateQ25
229884.5000
referenceQ25
230,374
classmean
contrast
classmean-d3s7-vs-prior-d3s7
verdict
inconclusive
meanDelta
930.7012
bootstrapLower95
-530.3950
bootstrapUpper95
2954.3730
studentTLower95
-869.1265
detectionFloor
1796.7246
wins
4
ties
506
losses
2
firstHalfMeanDelta
-689.9414
secondHalfMeanDelta
2551.3438
candidateQ25
230,374
referenceQ25
230,374
fillDepth4
contrast
fill-d4s7-vs-prior-d4s7
verdict
inconclusive
meanDelta
-22631.4395
bootstrapLower95
-57778.8480
bootstrapUpper95
12805.6263
studentTLower95
-58306.0979
detectionFloor
35613.1525
wins
236
ties
0
losses
276
firstHalfMeanDelta
37523.1680
secondHalfMeanDelta
-82786.0469
candidateQ25
242348.7500
referenceQ25
244331.5000
zeroedDepth4
contrast
zeroed-d4s7-vs-prior-d4s7
verdict
inconclusive
meanDelta
-10591.8145
bootstrapLower95
-32044.5230
bootstrapUpper95
11133.0066
studentTLower95
-32160.3172
detectionFloor
21531.3170
wins
93
ties
320
losses
99
firstHalfMeanDelta
-16117.3555
secondHalfMeanDelta
-5066.2734
candidateQ25
230,407
referenceQ25
244331.5000
depthSteps
fill-d4s7-vs-fill-d3s7
contrast
fill-d4s7-vs-fill-d3s7
verdict
inconclusive
meanDelta
-7170.3789
bootstrapLower95
-45565.7621
bootstrapUpper95
31331.3407
studentTLower95
-45470.7916
detectionFloor
38234.3797
wins
263
ties
0
losses
249
firstHalfMeanDelta
17876.4258
secondHalfMeanDelta
-32217.1836
candidateQ25
242348.7500
referenceQ25
229449.5000
prior-d4s7-vs-prior-d3s7
contrast
prior-d4s7-vs-prior-d3s7
verdict
inconclusive
meanDelta
36500.3398
bootstrapLower95
-249.3278
bootstrapUpper95
72651.2398
studentTLower95
-261.5755
detectionFloor
36698.5348
wins
287
ties
0
losses
225
firstHalfMeanDelta
-13180.8047
secondHalfMeanDelta
86181.4844
candidateQ25
244331.5000
referenceQ25
230,374
zeroed-d4s7-vs-zeroed-d3s7
contrast
zeroed-d4s7-vs-zeroed-d3s7
verdict
inconclusive
meanDelta
-3533.6426
bootstrapLower95
-43880.1008
bootstrapUpper95
36833.7896
studentTLower95
-43899.5511
detectionFloor
40296.3145
wins
270
ties
0
losses
242
firstHalfMeanDelta
-41466.0898
secondHalfMeanDelta
34398.8047
candidateQ25
230,407
referenceQ25
229884.5000
direct
fill-1ply-vs-prior-1ply
contrast
fill-1ply-vs-prior-1ply
verdict
inconclusive
meanDelta
4809.1660
bootstrapLower95
-13188.6314
bootstrapUpper95
22385.6826
studentTLower95
-13114.5791
detectionFloor
17892.8432
wins
289
ties
0
losses
223
firstHalfMeanDelta
21419.3516
secondHalfMeanDelta
-11801.0195
candidateQ25
190083.7500
referenceQ25
157,932
fill-1ply-vs-fair-d3s7
contrast
fill-1ply-vs-fair-d3s7
verdict
refuted
meanDelta
-31695.2402
bootstrapLower95
-50765.8081
bootstrapUpper95
-12546.5545
studentTLower95
-51006.1529
detectionFloor
19277.6191
wins
249
ties
0
losses
263
firstHalfMeanDelta
-21623.6797
secondHalfMeanDelta
-41766.8008
candidateQ25
190083.7500
referenceQ25
176,683
prior-1ply-vs-fair-d3s7
contrast
prior-1ply-vs-fair-d3s7
verdict
refuted
meanDelta
-36504.4063
bootstrapLower95
-56473.7448
bootstrapUpper95
-16523.1362
studentTLower95
-56674.1408
detectionFloor
20134.9603
wins
225
ties
0
losses
287
firstHalfMeanDelta
-43043.0313
secondHalfMeanDelta
-29965.7813
candidateQ25
157,932
referenceQ25
176,683
replicationOfPrior
contrast
prior-d3s7-vs-fair-d3s7
verdict
supported
meanDelta
155560.6211
bootstrapLower95
123421.7851
bootstrapUpper95
187990.6395
studentTLower95
123162.2110
detectionFloor
32342.5527
wins
329
ties
0
losses
183
firstHalfMeanDelta
175955.6602
secondHalfMeanDelta
135165.5820
candidateQ25
230,374
referenceQ25
176,683
theory
primaryFalsifierUpperBoundBelowZero
false
conditioningUpperBoundBelowZero
false
gainIsContinuationNotConditioning
false
optimismLegRefuted
false
depthCompounding
false
selection
candidateArm
occ5
candidateBestMargin
203192.9531
candidateBestMoves
1,400,252,568
controlBestMargin
210990.1797
controlBestMoves
600,112,060
arms
occ5
bestMargin
203192.9531
finalMargin
150272.3125
bestMoves
1,400,252,568
movesTotal
1,600,289,700
validationPoints
8
stop
plateau
hgt5
bestMargin
179759.7422
finalMargin
157845.9023
bestMoves
1,600,283,262
movesTotal
2,000,004,221
validationPoints
10
stop
plateau
control
bestMargin
210990.1797
finalMargin
197769.3398
bestMoves
600,112,060
movesTotal
1,200,228,265
validationPoints
6
stop
plateau
trainingSignal
criterion
a fill arm's best validation margin exceeds the control arm's best validation margin
passed
false
rule
the fill arm (occ5 or hgt5) whose best validation point has the larger paired mean margin of ntuple-d3s7 over fair-d3s7 on the 256-game training-role block, ties to occ5; the control arm's best point is the control candidate
trainingArms
control
layout
rows,cols,win23,win32,phase=all
movesTotal
1,200,228,265
gamesTotal
13,419,520
wallSeconds
839.2000
meanMovesPerSecond
2017917.3667
finalMargin
197769.3398
bestMargin
210990.1797
best
moves
600,112,060
artifact
val-000600112060.json
pairedDeltaD3
210990.1797
ntupleD3Mean
523756.8047
fairD3Mean
312766.6250
stop
reason
plateau
movesTotal
1,200,228,265
gamesTotal
13,419,520
wallSeconds
882.8000
validationPoints
6
plateauWindow
3
recentWindowMean
176417.0768
previousWindowMean
188501.5286
bestMargin
210990.1797
validationPoints
  1. movesTrained
    200,037,439
    pairedDeltaD3
    178166.6094
    bootstrapLower95
    137938.3291
    ntupleD3Mean
    490933.2344
    directMean
    294995.2031
  2. movesTrained
    400,073,347
    pairedDeltaD3
    176347.7969
    bootstrapLower95
    135424.2799
    ntupleD3Mean
    489114.4219
    directMean
    300747.5977
  3. movesTrained
    600,112,060
    pairedDeltaD3
    210990.1797
    bootstrapLower95
    165418.0455
    ntupleD3Mean
    523756.8047
    directMean
    305466.2656
  4. movesTrained
    800,150,452
    pairedDeltaD3
    171379.6953
    bootstrapLower95
    133016.8545
    ntupleD3Mean
    484146.3203
    directMean
    302770.1289
  5. movesTrained
    1,000,189,400
    pairedDeltaD3
    160102.1953
    bootstrapLower95
    126792.5285
    ntupleD3Mean
    472868.8203
    directMean
    302339.7422
  6. movesTrained
    1,200,228,265
    pairedDeltaD3
    197769.3398
    bootstrapLower95
    158764.9061
    ntupleD3Mean
    510535.9648
    directMean
    302966.1914
hgt5
layout
rows,cols,win23,win32,phase=all,fill=hgt5
movesTotal
2,000,004,221
gamesTotal
23,290,336
wallSeconds
1381.1000
meanMovesPerSecond
2153323.4800
finalMargin
157845.9023
bestMargin
179759.7422
best
moves
1,600,283,262
artifact
val-001600283262.json
pairedDeltaD3
179759.7422
ntupleD3Mean
492526.3672
fairD3Mean
312766.6250
stop
reason
plateau
movesTotal
2,000,004,221
gamesTotal
23,290,336
wallSeconds
1427.4000
validationPoints
10
plateauWindow
3
recentWindowMean
165313.9635
previousWindowMean
169776.8294
bestMargin
179759.7422
validationPoints
  1. movesTrained
    200,028,295
    pairedDeltaD3
    124708.8828
    bootstrapLower95
    91008.8102
    ntupleD3Mean
    437475.5078
    directMean
    261277.8047
  2. movesTrained
    400,060,509
    pairedDeltaD3
    137262.4492
    bootstrapLower95
    103554.1320
    ntupleD3Mean
    450029.0742
    directMean
    280088.2773
  3. movesTrained
    600,096,150
    pairedDeltaD3
    156897.5898
    bootstrapLower95
    117185.8607
    ntupleD3Mean
    469664.2148
    directMean
    279606.3828
  4. movesTrained
    800,131,115
    pairedDeltaD3
    135726.8086
    bootstrapLower95
    101618.8695
    ntupleD3Mean
    448493.4336
    directMean
    286336.0078
  5. movesTrained
    1,000,167,008
    pairedDeltaD3
    160216.1602
    bootstrapLower95
    121481.9051
    ntupleD3Mean
    472982.7852
    directMean
    303430.8867
  6. movesTrained
    1,200,208,525
    pairedDeltaD3
    179588.2344
    bootstrapLower95
    142390.0434
    ntupleD3Mean
    492354.8594
    directMean
    291789.8984
  7. movesTrained
    1,400,244,137
    pairedDeltaD3
    169526.0938
    bootstrapLower95
    130142.8451
    ntupleD3Mean
    482292.7188
    directMean
    312632.3828
  8. movesTrained
    1,600,283,262
    pairedDeltaD3
    179759.7422
    bootstrapLower95
    139911.8377
    ntupleD3Mean
    492526.3672
    directMean
    285265.7695
  9. movesTrained
    1,800,319,931
    pairedDeltaD3
    158336.2461
    bootstrapLower95
    120472.5197
    ntupleD3Mean
    471102.8711
    directMean
    297038.8789
  10. movesTrained
    2,000,004,221
    pairedDeltaD3
    157845.9023
    bootstrapLower95
    118844.9162
    ntupleD3Mean
    470612.5273
    directMean
    330794.6133
occ5
layout
rows,cols,win23,win32,phase=all,fill=occ5
movesTotal
1,600,289,700
gamesTotal
18,838,964
wallSeconds
1130.4000
meanMovesPerSecond
2125289.3750
finalMargin
150272.3125
bestMargin
203192.9531
best
moves
1,400,252,568
artifact
val-001400252568.json
pairedDeltaD3
203192.9531
ntupleD3Mean
515959.5781
fairD3Mean
312766.6250
stop
reason
plateau
movesTotal
1,600,289,700
gamesTotal
18,838,964
wallSeconds
1175.8000
validationPoints
8
plateauWindow
3
recentWindowMean
174445.5208
previousWindowMean
179680.9193
bestMargin
203192.9531
validationPoints
  1. movesTrained
    200,030,322
    pairedDeltaD3
    141881.6055
    bootstrapLower95
    103289.1463
    ntupleD3Mean
    454648.2305
    directMean
    270047.2031
  2. movesTrained
    400,061,818
    pairedDeltaD3
    177996.7188
    bootstrapLower95
    134200.2764
    ntupleD3Mean
    490763.3438
    directMean
    283788.3711
  3. movesTrained
    600,097,906
    pairedDeltaD3
    177735.6797
    bootstrapLower95
    139164.4127
    ntupleD3Mean
    490502.3047
    directMean
    291741.1055
  4. movesTrained
    800,135,553
    pairedDeltaD3
    180643.6641
    bootstrapLower95
    136012.3461
    ntupleD3Mean
    493410.2891
    directMean
    306541.3555
  5. movesTrained
    1,000,173,026
    pairedDeltaD3
    180663.4141
    bootstrapLower95
    142053.4818
    ntupleD3Mean
    493430.0391
    directMean
    300276.3711
  6. movesTrained
    1,200,212,497
    pairedDeltaD3
    169871.2969
    bootstrapLower95
    130476.3566
    ntupleD3Mean
    482637.9219
    directMean
    311855.7422
  7. movesTrained
    1,400,252,568
    pairedDeltaD3
    203192.9531
    bootstrapLower95
    157108.0422
    ntupleD3Mean
    515959.5781
    directMean
    316704.9727
  8. movesTrained
    1,600,289,700
    pairedDeltaD3
    150272.3125
    bootstrapLower95
    114342.4389
    ntupleD3Mean
    463038.9375
    directMean
    294769.7891
edits
format
drop7-ntuple-scale-edits-v1
source
/home/keshav/Developer/drop7-bench/runs/RUN-20260905T193006Z-4fbeb4e5/ntuple-scale/main/best-weights.bin
sourceSha256
0ade9d4e4080ebdd52a1474b1a13410dc8dfb77f5eba24b078aa7703c92ace0b
layout
rows,cols,win23,win32,phase=all
entries
1,000,000,000
optimisticStart
0.2703
untouchedEntries
901,259,321
touchedEntries
98,740,679
checkpoint
path
/home/keshav/Developer/drop7-bench/runs/RUN-20260905T193006Z-4fbeb4e5/ntuple-scale/main/checkpoint.bin
touchedEntriesAtStartValue
20,266,847
neverUpdatedEntriesOffStartValue
0
note
the checkpoint is the end of training, not the candidate's point; an entry touched only after the candidate's point is untouched in the candidate and counted as such
classesWithNoTouchedEntry
7
outputs
zeroed
path
/home/keshav/Developer/drop7-bench/runs/RUN-20260906T201104Z-a96ea6c8/ntuple-scale/main/zeroed-weights.bin
sha256
e9248b1a26f0b6e2d6cac121df9eda48c1adf1ea369287a60546a830b65c058e
entriesChanged
901,259,321
entriesChangedOnlyWhereUntouched
true
classmean
path
/home/keshav/Developer/drop7-bench/runs/RUN-20260906T201104Z-a96ea6c8/ntuple-scale/main/classmean-weights.bin
sha256
c3f05eeebb6ed573ba176797e7cfa4bdab3b47e042c8877011dd6614ebb54450
entriesChanged
877,344,474
entriesChangedOnlyWhereUntouched
true
wallSeconds
25.3000
tablesSha256
prior
0ade9d4e4080ebdd52a1474b1a13410dc8dfb77f5eba24b078aa7703c92ace0b
candidate
4f2e7ccf5c14fed8dd19563965e3937e8784b487f2de9eb70fe0e86830edae24
control
92dd1cb2d2a74b026270606c18c5f0d6e4f4ccc74e64c3c4e0f0d2043cdddd90
zeroed
e9248b1a26f0b6e2d6cac121df9eda48c1adf1ea369287a60546a830b65c058e
classmean
c3f05eeebb6ed573ba176797e7cfa4bdab3b47e042c8877011dd6614ebb54450
Limitations
  • Public-development SCREEN tier, 512 paired games opened once; nothing here is a qualification claim, and protected and final cohorts stay sealed.
  • The candidate and the control are each the best of their arm's validation points, and the candidate arm is the better of two, all chosen on the 256-game training-role validation block; the screen inherits that selection and measures what survives it.
  • The training arms are Hogwild (32 lock-free workers) and are not bit-reproducible; a re-run trains different tables from the same warm start and seeds.
  • Five buckets with fixed edges are the only conditioning tested; a different bucket count, different edges, or a bucket variable other than occupied cells and tallest column is a different configuration.
  • The depth-4 arms use the standing reference configuration (1M-entry table, seven strata, terminal utility -1,000,000); the fair-d3s7 arm is context, and the fair leaf at depth 4 was not played on this block.
  • Wall times were measured on a shared workstation with every arm run in sequence at 32 threads; logical work and the ratios between arms on the same seeds are the trustworthy cost quantities.
  • The per-game artifact is cited by its public archive reference (data.drop7.dev, immutable run-scoped key, digest in the fragment); a byte-identical copy remains under runs/RUN-20260906T201104Z-a96ea6c8/ntuple-scale/screen/heldout.json on the workstation and a promoted copy under artifacts/results/.

Open the result record

Agent contextHow to extend this record

To add a reader-facing explanation, write web/content/research/EX-20260906-ntuple-fill-conditioned-continuation-a9e5cbd3.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-20260906-ntuple-fill-conditioned-continuation-a9e5cbd3.json, validated against research/schemas/experiment-v1.schema.json. Protocol hash: d4e59a14f6ca86931c918d70b5bd3818f5985812bc075763f5e2c8937a58b0aa.

Registered by Claude Code / claude-fable-5-1 (claude-code-ntuple-fill-conditioned).