Drop7 Research

Type to search every page, record and vocabulary term.

GitHubPlay

Experiments

Preregistered protocols: the candidate, the comparator, the cohort, the metrics and the gate, fixed before the data is read.

TitleStatusTierDateApproach
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
completed
tier: SCREENRow and column lookup tables, learned from a billion moves
The frozen n-tuple tables as the leaf of the depth-4 fair search: a two-by-two leaf-by-depth screen on a fresh block
completed
tier: SCREENRow and column lookup tables, learned from a billion moves
Fresh-block replication of the row-and-column n-tuple leaf at larger scale: wider tables trained until the validation margins plateau, screened beside the first candidate
completed
tier: SCREENRow and column lookup tables, learned from a billion moves
Rust exact leaf and cache efficiency checks
preregistered
tier: CHECKRust bitboard engine
Row-and-column n-tuple network trained by temporal-coherence TD on the Rust engine, screened as the leaf of the depth-3 seven-stratum fair search
completed
tier: SCREENRow and column lookup tables, learned from a billion moves
Second continuation of the depth-5-distilled NNUE leaf evolution: resume from the plateau population with a slower mutation-size decay (up to 1,000 further generations)
completed
tier: SCREENAn evolved NNUE leaf, distilled from a depth-5 teacher
Continued whole-game evolution of the depth-5-distilled NNUE leaf from the first run's generation-60 population, with annealed mutation and a preregistered plateau stop (up to 1,000 generations)
completed
tier: SCREENAn evolved NNUE leaf, distilled from a depth-5 teacher
Pruned fair search against full-width fair d4s7 on the previously evaluated 256-game pilot cohort
preregistered
tier: PILOTOne-ply Q as a pruning prior and a refit leaf
Linear leaf refit to the depth-4 root value: held-out fit, depth-5 agreement proxy, and gameplay pilot
preregistered
tier: PILOTOne-ply Q as a pruning prior and a refit leaf
Depth-4 sibling-value panel, linear Q fit, and pruned-search regret on held-out roots
preregistered
tier: PILOTOne-ply Q as a pruning prior and a refit leaf
Cross-entropy search over the six feature weights against whole-game lifetime, frozen and read once on the pilot cohort
preregistered
tier: PILOTKlein-Friedmann linear Q-learning
Port the six-feature linear Q-learner onto the Rust bitboard engine: parity gates and a diagnostic pilot on previously evaluated development seeds
preregistered
tier: PILOTKlein-Friedmann linear Q-learning
Reproduce the Klein-Friedmann Q-learning report in its own simulator at pinned commit 8cc8a0e
preregistered
tier: CHECKKlein-Friedmann linear Q-learning
Depth-5-distilled NNUE leaf refined by paired-fitness whole-game evolution inside the depth-3 fair search (v2: operational parameters fixed in the record body)
completed
tier: SCREENAn evolved NNUE leaf, distilled from a depth-5 teacher
Depth-5-distilled NNUE leaf refined by paired-fitness whole-game evolution inside the depth-3 fair search
superseded
tier: SCREENAn evolved NNUE leaf, distilled from a depth-5 teacher
Rust search-matrix analytics and EC2 launch reproducibility check
preregistered
tier: CHECKRust bitboard engine
Central frontier scheduler parity and within-decision scaling check
preregistered
tier: CHECKRust bitboard engine
Rust bitboard engine: parity gates and throughput benchmark vs TypeScript and C++ engines
completed
tier: CHECKRust bitboard engine
P-SOL-3: T0 timing, all-M=6 fidelity ladder, uniform certified-cheap-engine M=6 corpus, and the G1/G2 offline gates
preregistered
tier: CHECKThe fast engine learns reveal sampling
E-FAST-M6: port M=6 reveal sampling into the fast memo engine, trace-equivalent to the native factored search
completed
tier: CHECKSeparate reveal and next-disc sampling
P-SOL-2: corrected two-rung G0 fidelity ladder (fast-d3s7 proxy + native D3 N7M6 guardrail), fast-engine corpus continuations, and the G1/G2 offline gates
completed
tier: CHECKThe data factory, and the switch that was left off
P-SOL-1: panel2 sibling corpus with CRN Kaplan-Meier outcome labels, NNUE retraining, and the G0/G1/G2 offline gates
superseded
tier: CHECKThe data factory, and the switch that was left off
C0: paired D3 N7M6 minus D4 s5 contrast from retained per-game records
completed
tier: CHECKFair expectimax reference (D3/D4)
Stage D0: fair relabelling of oracle-visited vs matched fair-D4 public states
completed
tier: CHECKPerfect-information oracle
NNUE-class student on existing successor-closed D4 labels: within-root ordering probe
completed
tier: CHECKA learned survival estimate inside the reference search
Successor SCREEN: paired 256-game test of the reveal-construction leaf terms (two doses and a chain-to-crack bundle) against the unchanged fair D4 at seven strata with the memo engine, after the corpus value gate failed
completed
tier: SCREENChain-reveal leaf substrate
Seed-free corpus gate, then a paired 256-game SCREEN of fair D4 (seven strata, memo engine) with reveal-construction leaf terms against the unchanged leaf
completed
tier: SCREENChain-reveal leaf substrate
CHECK gates for the one-entry leaf memo: bit identity, search parity, determinism, interleaved timing on real probe seeds
completed
tier: CHECKRemembering the last board
Paired SCREEN of fair D4 with the survival-instinct root filter (strict and literal) against unchanged fair D4 on 128 fresh development games
completed
tier: SCREENSurvival instinct — never place a disc where it can't clear vertically
Prevalence, lead time and incremental lifetime signal of entombed numbered discs in the existing training corpus
preregistered
tier: CHECKEntombed discs: the 3 that can never clear
Offline capacity and training sweep of the NNUE-shaped survival leaf on the existing corpus
preregistered
tier: CHECKIs the small survival network as accurate as it can be?
CMA-ES over the eighteen fair-leaf weights inside the fast depth-4 five-stratum search, fresh training seeds per generation, one held-out paired screen
completed
tier: SCREENEvolving the leaf weights at the depth they are used
Depth x chance-exactness factorial: the fifth ply at five and at seven strata, with an end-to-end reproduction control
running
tier: STANDARDFast engine
The two arms finding-09 left unfinished: reveal sampling on top of the fourth ply, and the depth-3 ladder at full joint coverage
running
tier: STANDARDSeparate reveal and next-disc sampling
Scale-out stage 1: successor-closed fair-D4 search-value labels; held-out D4-ranking agreement
completed
tier: CHECKDistributional afterstate ranker
Stronger-teacher (D2 continuation) afterstate corpus: ranking gate and frozen override rule
completed
tier: CHECKDistributional afterstate ranker
Full training of the K=256 afterstate model, then the frozen top-two override rule on fresh origins
completed
tier: CHECKDistributional afterstate ranker
Offline gate: calibrated top-two near-tie override of fair D4 by the frozen afterstate model
completed
tier: PILOTDistributional afterstate ranker
Pilot iteration 3: K=256-scenario corpus, fresh held-out origins, full training
completed
tier: PILOTDistributional afterstate ranker
Pilot iteration 2: K=64-scenario action-complete H40 corpus with fresh held-out origins
completed
tier: PILOTDistributional afterstate ranker
Pilot: action-complete H40 afterstate corpus and distributional ranker offline gate
completed
tier: PILOTDistributional afterstate ranker