Drop7 Research

Type to search every page, record and vocabulary term.

GitHubPlay

Research log

A dated account of what was tried, by whom, and what came of it. A day that ended in rejected ideas is written up in the same detail as a day that produced a result.

The log is narrative. It is written by the people and models doing the work and is not itself evidence. Every number in an entry belongs to a record (a theory, an experiment or its result), and those records, with their run validity, scientific outcome and evidence tier, are the authority.

Logged outcomes per day, 2026-08-20 to 2026-08-23 — the negative-results column is the honest picture

  • positive
  • negative
  • open
What each logged day produced, counted from the outcome tallies in the entries' own frontmatter. Negative results outnumber positive ones; recording them is the log's job.
Notes and sources

Counts are the outcome tallies recorded in each day's log frontmatter (web/content/log/2026-08-2*.mdx, the 'outcomes' block), copied verbatim: 2026-08-20 positive 2 / negative 5 / open 3; 2026-08-21 positive 2 / negative 7 / open 6; 2026-08-22 positive 2 / negative 2 / open 1; 2026-08-23 positive 1 / negative 5 / open 2. The figure plan (written mid-day on 2026-08-23) carried 0 / 2 / 1 for 2026-08-23; the finalized log frontmatter reads 1 / 5 / 2 and is charted — plan-vs-record drift resolved in the record's favour. Milestones behind these counts (verbatim anchors): the 400,675-point eight-game D4 confirmation (docs/strategies.md, Current reference table); the 398,498 d4s7 64-game confirmation mean (finding-05); the 386,545 frozen-arm mean over 256 fresh games (RS-20260823T131226Z-16564ed9, metrics.frozenArmFreshBaseline). Kind is dot rather than the plan's line because the generator requires numeric x for line charts and the task specifies string date categories.

Spec: web/content/figures/evidence-timeline.json · 4 source records · kind dot

  1. The September global game and the fourteen policies that opened its leaderboard

    The public global game moved to gauntlet-02 for September. The August game on gauntlet-01 is archived rather than deleted, so its leaderboard and every packed submission on it stay readable. Fourteen registered policies then played the new round on this workstation and their games were published to the competition ledger. The lookup tables of the first n-tuple-scale run, as the leaf of a depth-3 search, scored 1,113,829 points in 310 moves. That is one game on one fixed future and it is not evidence about the qualification standard, which asks for a mean above one million points across a cohort. The Rust engine's depth-6 search scored 105,578 in 35 moves on the same round after forty-five minutes of compute, below its own depth-4 sibling. The depth-7 policy was left out because a single mid-game decision has already been measured above two hours. Two site bugs found while reviewing the release were fixed before it shipped.

    claude-fable-5-1 (Claude Code)claude-opus-5 (Claude Code)
    3 positive2 negative2 open
    #benchmark-playground#competition#leaderboard#n-tuple#search-depth#deployment#negative-result

  2. Replicating the lookup-table leaf on fresh seeds, playing it one ply deeper, and asking it how full the board is

    Yesterday's lookup-table leaf replicated on 512 seeds nothing had read: 487,066 against 326,717 for the fair leaf in the same depth-3 search. A six-times-wider evaluator trained until its validation margin stopped rising passed the same gate and came out 5,196 behind the first tables, an inconclusive scale reading, while playing much better without any search. As a playground demonstration, the first tables also played the public global game's round, gauntlet-01, and scored 554,510 points in 160 moves. In the evening the same frozen tables played as the leaf of the depth-4 search on a fourth fresh block: 516,155 against 465,675 for the depth-3 leaf on 512 games, a fourth ply worth 50,479 points, the same as the fair leaf's fourth ply on the same games. The tables and the screens' per-game rows are now in the public research archive. Late in the day a fourth experiment gave every table five copies keyed to how full the board is, warm-started from the frozen tables, and tried two no-training edits of the never-updated entries on a fifth fresh block: the fill candidate was 21,039 ahead of the frozen tables at depth 3, inside the floor, and 22,631 behind at depth 4; zeroing the never-updated entries was worth 29,442 at depth 3 with a lower bound above zero and gave it back at depth 4.

    claude-fable-5-1 (Claude Code)
    5 positive3 negative3 open
    #n-tuple#temporal-difference#leaf#rust#replication#plateau#benchmark-playground#search-depth#artifact-archive#fill-conditioning#weight-promotion#negative-result

  3. Full-row lookup tables and a leaner Rust evaluator

    Two threads landed today. A lookup-table leaf trained by temporal-difference play became the first learned evaluator to pass its screen against the fair leaf. Separately, the Rust search completed the same constructed-root decisions faster after avoiding unused leaf work, while a simpler portable unpacking routine produced no credible independent speedup.

    claude-fable-5-1 (Claude Code)unknown (Codex)
    5 positive3 negative2 open
    #n-tuple#temporal-difference#leaf#rust#pomdp#pilot#engines#transposition-table#optimization#mechanics

  4. Slower mutation decay and engine pages

    The third evolved-leaf experiment stopped at the same plateau, and its gain over the prior candidate remained unconfirmed. The engine pages now explain their implementations through source excerpts, and the choice-and-chance lesson has an interactive animation.

    claude-sonnet-5 (Claude Code)gpt-5 (Codex)
    0 positive3 negative2 open
    #nnue#evolution#leaf#plateau#annealing-schedule#held-out-screen#negative-result#web-console#engines#documentation#source-code

  5. The evolved NNUE leaf loses its screen, evolution still moves it, and 150 more generations find the plateau

    The depth-5-distilled, whole-game-evolved NNUE leaf finished its preregistered run: 60 generations, a 128-game re-selection, and one look at 64 never-read games. It scored 106,964 paired points below the frozen fair leaf inside the same depth-3 search, so the gate fails and the claim is not supported as tested. The same screen shows evolution beat its own unevolved warm start by 35,375 points with a lower bound well above zero, the first time whole-game evolution has demonstrably moved a leaf in this program. The warm start, not the evolution, was the weak link: fitted to the teacher's values, it ranked columns about as well as no leaf at all. A continuation from generation 60 stopped itself at generation 149: the continued leaf still lost to the fair leaf on a fresh screen and beat the first run's candidate by 36,278 points. Separately, the research console was rebuilt around a technique taxonomy, and reading two approach pages against their own records showed both were carrying a status the records do not support.

    claude-fable-5-1 (Claude Code)claude-opus-5 (Claude Code)
    2 positive5 negative3 open
    #nnue#evolution#leaf#teacher-distillation#held-out-screen#negative-result#web-console#evidence-labels

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

    The evolved-NNUE-leaf run's teacher stage closed far short of its planned corpus after running four to seven times slower than projected, while a retained diagnostic showed that the leaf carries about 97% of the spread between columns. In a separate track, the Klein-Friedmann Q-learning report was reproduced and dissected, its features were transferred to the Rust engine, and the search's exact one-ply value pruned depth 4 to about a quarter of the work while the tested least-squares fits lost. The research console's fixed-box SVG generator was replaced by a measured visx chart kit, and the evolution approach page now explains the loop with live run figures.

    claude-fable-5-1 (Claude Code)
    8 positive9 negative5 open
    #nnue#evolution#leaf#teacher-distillation#web-console#figures#diagnostics#q-learning#linear-features#reproduction#rust-engine#value-policy-learning#pilot#prior-work#pruning#search-efficiency#leaf-evaluation#treestrap#negative-result

  7. The evolved-NNUE-leaf experiment is audited and launched, and a depth-6 gauntlet game joins the leaderboard by import

    The depth-5-distilled, whole-game-evolved NNUE leaf registered on 2026-08-25 was audited end to end - theory, frozen protocol, leases, and every line of the Rust crate - then launched for real. Eleven CHECK gates passed again on the rebuilt binaries, a seed-free preflight showed the teacher's labels are bit-identical at 16 and 32 threads with a 1.29x throughput gain from SMT, the protocol was amended before any leased seed was read, and the 512-game depth-5 teacher corpus began at 04:24 UTC on all 32 hardware threads. Separately, the rust-fair-d6-s7 gauntlet-01 game recorded on the larger workstation on 27 August was imported into the competition ledger for both stages. The seed command could not take it as designed, because it re-plays every policy live, so it gained a --replay path that verifies a recorded game against the immutable round instead of re-running the search. Playground demonstration only; the game's provenance is declared, not observed.

    claude-fable-5-1 (Claude Code)
    3 positive0 negative2 open
    #nnue#evolution#leaf#teacher-distillation#audit#validation#rust-engine#benchmark-playground#infrastructure#leaderboard

  8. The depth-7 opening is not the worst case: a mid-game decision overran two hours

    The first depth-7 gauntlet-01 attempt reached move 9 and then exceeded the policy's two-hour per-decision budget on a half-full board, more than three times the measured opening cost. The per-decision limit is now four hours, still an unverified guess rather than a measurement, and the crash journal retained all eight completed moves so the game resumes rather than restarting.

    claude-fable-5 (Claude Code)
    1 positive1 negative1 open
    #rust-engine#parallel-search#benchmark-playground#validation

  9. Local validation of the depth matrix: the frontier scheduler saturates a 16-core workstation

    The merged depth/strata/leaf matrix workflow was exercised end to end on the research workstation for the first time, without any EC2 launch. On gauntlet-01's initial public position the frontier scheduler held ~98-99% worker occupancy through D5/S7 and matched the root scheduler bit for bit, while the root scheduler capped at seven workers and left the rest of the machine idle. The matrix entry points were reorganized into run-matrix-ec2, run-matrix-local, and an interactive run-matrix dispatcher. make research-validate is failing on eight pre-existing references to deleted August 25 run artifacts.

    claude-fable-5 (Claude Code)
    3 positive2 negative2 open
    #rust-engine#parallel-search#infrastructure#benchmark-playground#validation

  10. A parallel frontier for deep search, and an evolutionary NNUE leaf designed around two failures

    Two threads landed today. The Rust fair solver gained a central work queue for continuation subtrees: exact local parity held and a reusable depth/strata/leaf matrix plus fail-closed EC2 workflow is ready for a separately authorized 192-core run. Separately, the opencode/evolutionary-approach branch registered a depth-3 NNUE-leaf approach: warm-started on a depth-5 teacher's sibling-complete root values, then refined by whole-game evolution with paired fitness blocks - implemented, gated, frozen, and not yet launched.

    unknown (Codex)kimi-k3 (OpenCode)
    3 positive1 negative2 open
    #rust-engine#parallel-search#infrastructure#teacher-search#nnue#evolution#leaf#teacher-distillation#preregistration

  11. The reveal quadrature gets a fast engine

    E-FAST-M6 passed every trace-equivalence gate: reveal sampling now runs inside the fast memo engine at 5.6x native speed, reopening the sibling-outcome label programme that yesterday's guardrail closed - though at 0.177 CPU-seconds per continuation move, a full M=6 corpus is still a scale decision, not a workstation one.

    claude-fable-5 (Claude Code)kimi-k3 (OpenCode)unknown (Codex)
    2 positive0 negative2 open
    #fast-engine#reveal-sampling#engineering#p-sol

  12. The reveal the search leaves on the table

    Morning: the owner's two strategic hypotheses became one structural leaf term that failed its corpus gate and its screen. Evening: a six-brief Kimi drafting fleet, two preregistered cheap experiments (both decisive negatives: no leaf-cost NNUE holds D4's ordering, and the oracle's boards carry no transferable fair value), one retained positive bound (D3 N7M6 beats D4 s5, LB +29,033), and the console learned to draw its own evidence.

    claude-fable-5 (Claude Code)kimi-k3 (OpenCode)
    1 positive5 negative2 open
    #leaf#reveals#chains#corpus-gate#screen#strategy