DAILY RESEARCH REPORTBACK TO MASTER SHEET2026-03-24

val_bpb · 158 of 365 · 2 warnings
findings mentioning this date
| 2026-03-24 | 2 | 5 |
audit summary
{
  "date": "2026-03-24",
  "fatal": 0,
  "headline_metric": "val_bpb",
  "raw_records": 365,
  "render_records": 158,
  "test_rows": 5,
  "warnings": 2
}
105 commits on 2026-03-24
  1. 0d669794 05:03:46 parameter-golf-lab Fix DDP unused param error: touch gate on normalize steps
  2. 0d669794 05:03:46 sota_crawler Fix DDP unused param error: touch gate on normalize steps
  3. 553a302a 05:07:50 parameter-golf-lab Run3: push dim to 720 for max capacity
  4. 553a302a 05:07:50 sota_crawler Run3: push dim to 720 for max capacity
  5. 2448ebba 05:14:38 parameter-golf-lab Run4: self-referential crawler at dim=720
  6. 2448ebba 05:14:38 sota_crawler Run4: self-referential crawler at dim=720
  7. 3c4e1294 05:18:59 parameter-golf-lab Run5: persistent deliberation — gate fires every step
  8. 3c4e1294 05:18:59 sota_crawler Run5: persistent deliberation — gate fires every step
  9. 54332198 05:24:33 parameter-golf-lab Run6: run1 base (1.1377) + persistent deliberation + GPTQ
  10. 54332198 05:24:33 sota_crawler Run6: run1 base (1.1377) + persistent deliberation + GPTQ
  11. 53d63bc4 05:26:05 parameter-golf-lab Fix run6 launcher: dim 624→640
  12. 53d63bc4 05:26:05 sota_crawler Fix run6 launcher: dim 624→640
  13. cb0c990c 05:27:51 parameter-golf-lab Run7: run1 base + GPTQ only, no gate — the safe play
  14. cb0c990c 05:27:51 sota_crawler Run7: run1 base + GPTQ only, no gate — the safe play
  15. d044f5b4 05:39:13 parameter-golf-lab Run8: PD + GPTQ + fixed cadence 2 — keep EMA fresh
  16. d044f5b4 05:39:13 sota_crawler Run8: PD + GPTQ + fixed cadence 2 — keep EMA fresh
  17. 1c2caf60 05:42:28 parameter-golf-lab Run8: bidirectional PD — consensus_ref as learned Parameter
  18. 1c2caf60 05:42:28 sota_crawler Run8: bidirectional PD — consensus_ref as learned Parameter
  19. 64a6402a 05:43:34 parameter-golf-lab Save micro crawler sweep results + experiment log
  20. 64a6402a 05:43:34 sota_crawler Save micro crawler sweep results + experiment log
  21. 9db4f716 05:44:54 parameter-golf-lab Science log: all micro crawler H100 results + findings
  22. 9db4f716 05:44:54 sota_crawler Science log: all micro crawler H100 results + findings
  23. 69050b38 13:25:15 parameter-golf-lab Update README.md
  24. 69050b38 13:25:15 sota_crawler Update README.md
  25. 69050b38 13:25:15 SOTA_FINAL Update README.md
  26. 630bb5ea 17:16:34 parameter-golf-lab Update README.md
  27. 630bb5ea 17:16:34 sota_crawler Update README.md
  28. 630bb5ea 17:16:34 SOTA_FINAL Update README.md
  29. ae64883e 17:41:11 parameter-golf-lab Cadence ablation science: H1 + H2 fronts, 8 arms, RunPod-ready
  30. ae64883e 17:41:11 sota_crawler Cadence ablation science: H1 + H2 fronts, 8 arms, RunPod-ready
  31. 2b0ff257 17:47:02 parameter-golf-lab Fix: total_mem → total_memory for PyTorch 2.9+
  32. 2b0ff257 17:47:02 sota_crawler Fix: total_mem → total_memory for PyTorch 2.9+
  33. 3c40629d 17:48:06 parameter-golf-lab Add diagnostic training script with polar blending + cadence support
  34. 3c40629d 17:48:06 sota_crawler Add diagnostic training script with polar blending + cadence support
  35. 70f3624e 17:54:58 parameter-golf-lab H3 hypothesis: per-block cadence gradient (funnel/diamond/inverse)
  36. 70f3624e 17:54:58 sota_crawler H3 hypothesis: per-block cadence gradient (funnel/diamond/inverse)
  37. e21e1cea 18:13:15 parameter-golf-lab H1: cadence 0 (no C-steps) at full 1.0 scale — the critical test
  38. e21e1cea 18:13:15 sota_crawler H1: cadence 0 (no C-steps) at full 1.0 scale — the critical test
  39. a1229285 18:43:13 parameter-golf-lab H2: inverse architecture 2f+4cx2 — skinny head, deep recursion
  40. a1229285 18:43:13 sota_crawler H2: inverse architecture 2f+4cx2 — skinny head, deep recursion
  41. a90ee48f 18:57:04 parameter-golf-lab Run 8 production script with zero C-steps — clean A/B vs Run 8
  42. a90ee48f 18:57:04 sota_crawler Run 8 production script with zero C-steps — clean A/B vs Run 8
  43. 5486cc74 19:15:17 parameter-golf-lab Science log: H1+H2 cadence sweep results + cad0 full-scale data
  44. 5486cc74 19:15:17 sota_crawler Science log: H1+H2 cadence sweep results + cad0 full-scale data
  45. f5dad431 19:33:43 parameter-golf-lab DEFINITIVE: cad0 beats Run 8 on production script (1.1325 vs 1.1355)
  46. f5dad431 19:33:43 sota_crawler DEFINITIVE: cad0 beats Run 8 on production script (1.1325 vs 1.1355)
  47. ab5a4358 20:03:26 parameter-golf-lab H4: crawler bank on GS v7 U-Net — best trick on fastest car
  48. ab5a4358 20:03:26 sota_crawler H4: crawler bank on GS v7 U-Net — best trick on fastest car
  49. cb9ae232 20:26:26 parameter-golf-lab H4 results: crawler bank learns better per step but loses on sliding
  50. cb9ae232 20:26:26 sota_crawler H4 results: crawler bank learns better per step but loses on sliding
  51. a09e26ce 20:52:57 parameter-golf-lab f1: add PR587 baseline plus low-rank correction and stable distill
  52. a09e26ce 20:52:57 sota_crawler f1: add PR587 baseline plus low-rank correction and stable distill
  53. d20b24a5 21:01:23 parameter-golf-lab F1 v1: accuracy profile — corr rank 256 + distillation
  54. d20b24a5 21:01:23 sota_crawler F1 v1: accuracy profile — corr rank 256 + distillation
  55. 8f3c919b 21:09:35 parameter-golf-lab F1 v2: accuracy profile + crawler bank at U-Net bottleneck
  56. 8f3c919b 21:09:35 sota_crawler F1 v2: accuracy profile + crawler bank at U-Net bottleneck
  57. 1534eb98 21:12:20 parameter-golf-lab H4 F/G: fast small model (8L/384d) crawler bank test
  58. 1534eb98 21:12:20 sota_crawler H4 F/G: fast small model (8L/384d) crawler bank test
  59. 87b08088 21:15:04 parameter-golf-lab Organize all recursive weight-sharing transformer research: ablations, scripts, results
  60. 87b08088 21:15:04 sota_crawler Organize all recursive weight-sharing transformer research: ablations, scripts, results
  61. a3826226 21:56:31 parameter-golf-lab f1: add legal leaderboard profile and activation knob
  62. a3826226 21:56:31 sota_crawler f1: add legal leaderboard profile and activation knob
  63. 452fe939 22:09:56 parameter-golf-lab f1: add isolated sota xsa4+bg1536 speed-safe concept
  64. 452fe939 22:09:56 sota_crawler f1: add isolated sota xsa4+bg1536 speed-safe concept
  65. 9ae88e46 22:18:02 parameter-golf-lab Skiptrace: crawler bank fires every 10 steps, injects decaying delta
  66. 9ae88e46 22:18:02 sota_crawler Skiptrace: crawler bank fires every 10 steps, injects decaying delta
  67. b894a353 22:20:45 parameter-golf-lab Cubric: temporal weight sharing via skiptrace — research lab
  68. b894a353 22:20:45 sota_crawler Cubric: temporal weight sharing via skiptrace — research lab
  69. 940c93b0 22:22:11 parameter-golf-lab records: log f1 legal-lb xsa4/bg1536 1.1195 candidate run
  70. 940c93b0 22:22:11 sota_crawler records: log f1 legal-lb xsa4/bg1536 1.1195 candidate run
  71. 2b08385d 22:47:40 parameter-golf-lab SOTA: F1 legal LB profile — 1.1195 (seed 1337), 1.1200 (seed 42)
  72. 2b08385d 22:47:40 sota_crawler SOTA: F1 legal LB profile — 1.1195 (seed 1337), 1.1200 (seed 42)
  73. adec25ae 22:56:36 parameter-golf-lab Update recursive weight-sharing transformer PR draft with full research findings + warning labels
  74. adec25ae 22:56:36 sota_crawler Update recursive weight-sharing transformer PR draft with full research findings + warning labels
  75. 06338555 23:01:25 parameter-golf-lab Reframe PR: recursion is challenged, not dead — per-step benefit is real
  76. 06338555 23:01:25 sota_crawler Reframe PR: recursion is challenged, not dead — per-step benefit is real
  77. 4d8d64b2 23:09:43 parameter-golf-lab NEW RECORD: seed 2045 hits 1.1190 post-TTT — best single seed
  78. 4d8d64b2 23:09:43 sota_crawler NEW RECORD: seed 2045 hits 1.1190 post-TTT — best single seed
  79. 896c3254 23:13:36 parameter-golf-lab three-seed leaderboard variant: 1.1190 BPB (seed 2045), mean 1.1195
  80. 896c3254 23:13:36 sota_crawler three-seed leaderboard variant: 1.1190 BPB (seed 2045), mean 1.1195
  81. b8e8cf04 23:44:40 parameter-golf-lab H4: bottleneck crawler scripts for H100 pod
  82. b8e8cf04 23:44:40 sota_crawler H4: bottleneck crawler scripts for H100 pod
  83. 73631bb8 23:51:26 parameter-golf-lab f1: create sota garage with three baseline cars
  84. 73631bb8 23:51:26 sota_crawler f1: create sota garage with three baseline cars
  85. f9531a3d 23:59:55 parameter-golf-lab Fix H4 Arm C crash: remove PD gate, use simple sequential looping
  86. f9531a3d 23:59:55 sota_crawler Fix H4 Arm C crash: remove PD gate, use simple sequential looping
  87. 67e75ad5 00:25:53 parameter-golf-lab Add F1 architecture triplet run script
  88. 67e75ad5 00:25:53 sota_crawler Add F1 architecture triplet run script
  89. fe30a101 00:49:57 parameter-golf-lab H5-H8: next research hypotheses — skiptrace, trigram, noisy QAT, weight sharing
  90. fe30a101 00:49:57 sota_crawler H5-H8: next research hypotheses — skiptrace, trigram, noisy QAT, weight sharing
  91. 37190070 01:28:57 parameter-golf-lab Add isolated legal 5-gram eval A/B for car02
  92. 37190070 01:28:57 sota_crawler Add isolated legal 5-gram eval A/B for car02
  93. 22348a38 01:50:02 parameter-golf-lab Fix hashed n-gram eval probability safety
  94. 22348a38 01:50:02 sota_crawler Fix hashed n-gram eval probability safety
  95. 9be790b1 01:55:16 parameter-golf-lab Add compile stability switches for isolated n-gram eval
  96. 9be790b1 01:55:16 sota_crawler Add compile stability switches for isolated n-gram eval
  97. cb74ac4b 02:23:44 parameter-golf-lab car02: add budgeted ngram eval cutoff and partial reporting
  98. cb74ac4b 02:23:44 sota_crawler car02: add budgeted ngram eval cutoff and partial reporting
  99. 35b90f9b 02:55:55 parameter-golf-lab Distribute n-gram eval across all 8 GPUs — ~8x speedup
  100. 35b90f9b 02:55:55 sota_crawler Distribute n-gram eval across all 8 GPUs — ~8x speedup
  101. a417febf 03:49:44 parameter-golf-lab Podracing: 1.0451 BPB (3-seed mean 1.0461)
  102. 303e731a 03:59:11 parameter-golf-lab research: standardize hypothesis tagging across garage and experiments
  103. 303e731a 03:59:11 sota_crawler research: standardize hypothesis tagging across garage and experiments
  104. b4584b5f 04:20:25 parameter-golf-lab Cubric: visualization concept — model internals as art
  105. b4584b5f 04:20:25 sota_crawler Cubric: visualization concept — model internals as art
DATE2026-03-24
DAY007
PRIMARY METRICVAL_BPB
TESTS RUN5 unique (4 valid / 1 TTT leakage)
ARCHITECTURE FAMILYLEADERBOARD TRANSFORMER VARIANTS (×4) · OTHER (×1)
CLASSIFICATIONEXPERIMENTAL 100% (5)

SECTION 1.0 — TRAJECTORY

VAL_BPB for 2026-03-24 panel 1 of 2
csv data
VAL_BPB for 2026-03-24 panel 2 of 2
csv data
VAL_BPB SPREADnminmedianmaxdownload
LEADERBOARD TRANSFORMER VARIANTS41.11901.11981.1291—
ALL VALID41.11901.11981.1291—

SECTION 2.0 — TEST RUNS

EXPERIMENTAL TEST RUNS — 5

#TIMEHWABLATIONVAL_BPBSTATUSΔ %OPEN
103:49—Podracing: 1.0451 BPB (3-seed mean 1.0461)1.0451INVALID—
details
221:038×H100 (runpod)f1_submit_s1337_20260324_210317.log1.1291OK↑5.8%
training_log
321:598×H100 (runpod)f1_legal_lb_s1337_20260324_215935.log1.1196OK↑4.9%
training_log
422:268×H100 (runpod)f1_legal_lb_s42_20260324_222652.log1.1199OK↑4.9%
training_log
523:13—three-seed leaderboard variant: 1.1190 BPB (seed 2045), mean 1.1195 (×2)1.1190OK↑4.9%
details