-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathbenchmark-progress.csv
More file actions
We can make this file beautiful and searchable if this error is corrected: It looks like row 42 should actually have 9 columns, instead of 10 in line 41.
55 lines (55 loc) · 12.7 KB
/
benchmark-progress.csv
File metadata and controls
55 lines (55 loc) · 12.7 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
timestamp,status,scenario,commit,change_ref,total_nodes,avg_nodes_per_turn,deepest_completed_depth,notes
2026-03-08T20:14:09-07:00,baseline,profile,bf81c22,steady-state-profiling-benchmark,246590,49318,3,"5 repeated searches after 2 warmup turns; benchmark from scripts/run-benchmarks.sh profile"
2026-03-08T20:14:09-07:00,kept,profile,e15793c,tt-verification-key,246590,49318,3,"Added independent TT verification key and collision regression test; no measured node change on profile baseline"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,pvs-null-window,190870,38174,3,"Principal variation search reduced repeated-search throughput versus baseline; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,tt-hash-move-ordering,174115,34823,3,"TT move-to-front ordering reduced repeated-search throughput versus baseline; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,fixed-size-tt-table,246590,49318,3,"Direct-index TT table with simple overwrite policy did not improve repeated-search node count; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,root-pv-ordering,208000,41600,3,"Iterative-deepening root preferred move ordering reduced repeated-search throughput versus baseline; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,cached-goodness-evaluation,246590,49318,3,"Incremental leaf evaluation preserved behavior but did not improve repeated-search node count; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,killer-moves,174125,34825,3,"Generic killer-move ordering reduced repeated-search throughput versus baseline; reverted"
2026-03-08T20:14:09-07:00,kept,profile,75be3ca,wordbase-path-heuristic-v1,275945,55189,3,"Added static move-ordering bonuses for word length, edge progress, and bomb contact; repeated-search baseline improved and reproduced"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-path-heuristic-v2-edge-bonus,258010,51602,3,"Extra near-edge bonus reduced repeated-search throughput versus v1; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-path-heuristic-v3-no-bomb-bonus,238095,47619,3,"Removing bomb and megabomb bonuses reduced repeated-search throughput versus v1; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-path-heuristic-v4-endpoint-bonus,252465,50493,3,"Endpoint-row bonus reduced repeated-search throughput versus v1; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-path-heuristic-v5-length24,237915,47583,3,"Increasing the length bonus above v1 reduced repeated-search throughput; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-path-heuristic-v6-progress40,252795,50559,3,"Increasing the progress bonus above v1 reduced repeated-search throughput; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-path-heuristic-v7-progress24,275140,55028,3,"Lowering the progress bonus was close to v1 but still slightly worse; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-path-heuristic-v8-bomb64,264610,52922,3,"Lowering bomb and megabomb bonuses reduced repeated-search throughput versus v1; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-path-heuristic-v9-span24,238185,47637,3,"Forward-span bonus reduced repeated-search throughput versus v1; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,recordOne-already-owned-early-return,275945,55189,3,"Skipping already-owned cells in recordOne preserved behavior but did not improve the repeated-search baseline"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,recordOne-iterative-propagation,275945,55189,3,"Iterative bomb propagation in recordOne preserved behavior but did not improve the repeated-search baseline; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,dynamic-front-window-ordering,187145,37429,3,"State-aware front-window reranking by claimable and enemy path cells reduced repeated-search throughput badly versus v1; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-square-mobility-divisor8,261675,52335,3,"Adding a weak future-mobility bonus from per-square legal-word counts underperformed the stronger variants"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-square-mobility-divisor4,277530,55506,3,"A moderate future-mobility bonus improved profile slightly but was dominated by stronger scaling"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-square-mobility-divisor2,278125,55625,3,"A stronger future-mobility bonus improved profile but still underperformed divisor1"
2026-03-08T20:14:09-07:00,kept,profile,working-tree,wordbase-square-mobility-divisor1,339260,67852,3,"Added a per-square future-mobility bonus from legal-word counts across the claimed path; profile rose from 275945 to 339260 and short rose from 81916 to 84370"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,depth-based-move-cap-120-60,122100,24420,3,"Shrinking the move cap aggressively by ply cut away too much search and reduced both short and repeated-search benchmarks"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-forward-mobility-bonus-32,337075,67415,3,"Extra forward-weighted mobility on top of raw mobility improved short slightly but lost on repeated search"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-future-diversity-divisor8,339725,67945,3,"A weak future-move diversity bonus was worse than the stronger diversity variants"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-future-diversity-divisor4,345410,69082,3,"A moderate future-move diversity bonus improved on raw mobility but was beaten by divisor2"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-future-diversity-divisor1,22785,4557,2,"Overweighting future-move diversity collapsed the search and dropped completed depth"
2026-03-08T20:14:09-07:00,kept,profile,working-tree,wordbase-future-diversity-divisor2,354475,70895,3,"Added a future-move diversity bonus based on unique legal starts unlocked by the claimed path; profile rose from 339260 to 354475 and short rose to 87822"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,tt-hash-move-ordering-retest,182050,36410,3,"TT hash-move ordering still reduced throughput badly even after the stronger Wordbase ordering baseline; reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-deep-diversity-frontier,221088,44217,3,"Biasing diversity toward the deepest claimed cells looked promising in one run but was not repeatable and was reverted"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-peak-anchor-quality,81955,16391,2,"Using the best future word from claimed anchors overfit flashy lines and collapsed repeated-search depth"
2026-03-08T20:14:09-07:00,discarded,profile,working-tree,wordbase-long-word-count,17775,3555,2,"Counting long future words per square was another unstable anchor-quality signal and was reverted"
2026-03-08T20:14:09-07:00,kept,profile,working-tree,wordbase-long-future-diversity,364230,72846,3,"Added a long-word diversity bonus so paths preserving many distinct long threats are ordered earlier; profile rose from 354475 to 364230"
2026-03-08T21:21:32-07:00,kept,profile,0ca2b91,find-first-renumbered-goodness,364230,72846,3,Use_find_first_no_behavior_change
2026-03-08T21:22:25-07:00,discarded,profile,working-tree,wordbase-bend-penalty-k8,363915,72783,3,"Added a small bend penalty to prefer straighter lanes; repeated-search nodes regressed slightly vs baseline 364230 so reverting"
2026-03-08T21:23:04-07:00,discarded,profile,working-tree,endpoint-anchor-bonus-divisor6,364230,72846,3,No measurable change versus baseline; reverted to divisor8
2026-03-08T21:23:16-07:00,kept,profile,working-tree,wordbase-letter-rarity-target6-divisor2,364270,72854,3,"Added a rare-letter tie-breaker favoring paths that claim scarce letters (e.g., the lone 'z' along GLAMORIZE). On the profile scenario total_nodes rose from 364230 to 364270 and deepest_completed_depth stayed 3; effect is tiny but repeatable."
2026-03-08T21:23:30-07:00,discarded,profile,working-tree,wordbase-neighborhood-halo-divisor48,18005,3601,2,"Added a very weak 8-neighbor halo bonus; repeated-search depth collapsed from 3 to 2, so reverting"
2026-03-08T21:26:45-07:00,discarded,profile,working-tree,wordbase-neighborhood-max-halo-k96,364230,72846,3,Added a max-halo neighborhood tie-breaker (3x3 local mobility, single best cell per path) with a stiff divisor; profile repeated-search nodes unchanged vs baseline so reverting
2026-03-08T21:27:14-07:00,discarded,profile,working-tree,wordbase-letter-rarity-target6-divisor2,364230,72846,3,"After implementing the rare-letter tie-breaker, repeated-search totals matched baseline (no throughput or depth change). Keeping tree clean and reverting."
2026-03-08T21:31:51.827024-07:00,discarded,profile,working-tree,wordbase-row-band-diversity-weight8,364230,72846,3,"Added a row-band diversity bonus that rewards paths spanning many distinct rows (e.g., GLAMORIZE diagonal vs GLASS horizontal). Repeated-search totals unchanged vs baseline; reverting."
2026-03-08T21:34:17-07:00,discarded,profile,working-tree,wordbase-column-band-diversity-weight8,364230,72846,3,"Added a small column-band diversity bonus so diagonals like GLAMORIZE beat one-file paths like EEL when other signals tie. Repeated-search totals reproduced exactly at baseline, so reverting."
2026-03-08T21:35:19-07:00,discarded,profile,working-tree,wordbase-frontier-square-word-count-divisor6,364230,72846,3,"Added a frontier mobility bonus from unique neighboring squares around the claimed path (e.g., GLAMORIZE leaves side shoulders; EEL mostly stacks starts on one file). At divisor 6 the repeated-search profile stayed exactly at baseline, so trying a stronger scaling."
2026-03-08T21:35:51-07:00,discarded,profile,working-tree,wordbase-frontier-square-word-count-divisor2,345895,69179,3,"Strengthening the frontier neighboring-word-count bonus reordered search in the wrong direction: repeated-search profile fell from 364230 to 345895 while depth stayed 3. Reverting that scaling and trying a less correlated frontier signal."
2026-03-08T21:36:25-07:00,discarded,profile,working-tree,wordbase-frontier-square-count-weight12,340310,68062,3,"Switched to a geometry-only frontier count that rewards paths exposing many neighboring squares. It was even noisier: repeated-search profile fell to 340310 with no extra depth, so reverting to the known-good baseline."
2026-03-08T21:39:08-07:00,discarded,profile,working-tree,wordbase-future-word-text-diversity-divisor4,364230,72846,3,"Counted distinct future word texts so a lane leaving {stare, stern, stone} would outrank one mostly preserving repeated GLAMORIZE placements. At divisor 4 the repeated-search profile matched the 364230 baseline exactly, so reverting."
2026-03-08T21:39:08-07:00,discarded,profile,working-tree,wordbase-future-word-text-diversity-divisor2,364230,72846,3,"Strengthening the distinct future word-text bonus still produced the exact 364230 repeated-search profile total with depth 3. The text-level signal was too correlated with the existing mobility/diversity stack, so reverting."
2026-03-08T21:41:42-07:00,discarded,profile,working-tree,wordbase-future-progress-row-coverage-weight8,17760,3552,2,"Rewarded paths whose future follow-ups land on several forward rows instead of one band. That reordered the profile position badly enough to stall at depth 2 on every repeated search, so reverting."
2026-03-08T21:42:52-07:00,discarded,profile,working-tree,wordbase-future-start-square-coverage-weight6,18005,3601,2,"Rewarded paths that preserved many distinct future start squares. That favored noisy multi-anchor clutter over the baseline ordering and every repeated search stalled at depth 2, so reverting."
2026-03-08T21:43:36-07:00,discarded,profile,working-tree,wordbase-long-future-start-square-coverage-weight8,18005,3601,2,"Restricted the start-square coverage bonus to long follow-ups only, hoping to favor separate long-threat anchors. It still shoved the profile search into the same bad depth-2 ordering, so reverting."
2026-03-08T21:47:48-07:00,discarded,profile,working-tree,wordbase-long-square-word-count-divisor4,17775,3555,2,"Summed per-square counts of long words along the claimed path so the middle of a GLAMORIZE lane would outrank side squares that mostly feed GLASS. The signal was far too strong: repeated-search depth fell from 3 to 2, so reverting."
2026-03-08T21:47:48-07:00,kept,profile,working-tree,wordbase-square-forward-reach-divisor4,364570,72914,3,"Added a small per-square forward-reach bonus based on the average depth of words through each claimed cell, so lanes used by deeper attacks like PERILLED break ties over squares whose words stall locally. The repeated-search profile improved reproducibly from 364230 to 364570 in two runs, and short held at 87981."