DAILY RESEARCH REPORTBACK TO MASTER SHEET2026-04-09

val_bpb · 136 of 1273 · 0 warnings
{
  "date": "2026-04-09",
  "fatal": 0,
  "headline_metric": "val_bpb",
  "raw_records": 1273,
  "render_records": 136,
  "test_rows": 23,
  "warnings": 0
}
73 commits on 2026-04-09
  1. 81972ab5 05:57:19 parameter-golf-lab long-context causal tap mechanism v1.9 adaptive TTT ablation lane
  2. 81972ab5 05:57:19 sota_crawler long-context causal tap mechanism v1.9 adaptive TTT ablation lane
  3. 81972ab5 05:57:19 SOTA_FINAL long-context causal tap mechanism v1.9 adaptive TTT ablation lane
  4. 857de475 07:17:55 parameter-golf-lab Record: SP8192 + 3-Layer Recurrence + Parallel Residuals + QK-Gain 5.25 + Legal TTT — val_bpb 1.0810 (3-seed mean)
  5. 857de475 07:17:55 sota_crawler Record: SP8192 + 3-Layer Recurrence + Parallel Residuals + QK-Gain 5.25 + Legal TTT — val_bpb 1.0810 (3-seed mean)
  6. 857de475 07:17:55 SOTA_FINAL Record: SP8192 + 3-Layer Recurrence + Parallel Residuals + QK-Gain 5.25 + Legal TTT — val_bpb 1.0810 (3-seed mean)
  7. 3fdaaa56 16:59:39 sota_nueral Record all valid session results + stack scripts + corrected ratio sweep
  8. e90539f3 17:27:50 parameter-golf-lab Crawler layer relationship ablation: 5x4 grid sweep (flat 5-9 × crawler 1-4)
  9. e90539f3 17:27:50 sota_crawler Crawler layer relationship ablation: 5x4 grid sweep (flat 5-9 × crawler 1-4)
  10. e90539f3 17:27:50 SOTA_FINAL Crawler layer relationship ablation: 5x4 grid sweep (flat 5-9 × crawler 1-4)
  11. 94491303 18:27:41 sota_nueral Save all raw pod logs from 4xH100 session (2026-04-09)
  12. f136b6aa 19:33:20 parameter-golf-lab Crawler symmetry ablation: test C=loops hypothesis (3×3, 4×4, 6×6, 8×8)
  13. f136b6aa 19:33:20 sota_crawler Crawler symmetry ablation: test C=loops hypothesis (3×3, 4×4, 6×6, 8×8)
  14. f136b6aa 19:33:20 SOTA_FINAL Crawler symmetry ablation: test C=loops hypothesis (3×3, 4×4, 6×6, 8×8)
  15. 2c045f76 19:44:07 parameter-golf-lab Drop redundant control from symmetry ablation — grid already has 8F+3C
  16. 2c045f76 19:44:07 sota_crawler Drop redundant control from symmetry ablation — grid already has 8F+3C
  17. 2c045f76 19:44:07 SOTA_FINAL Drop redundant control from symmetry ablation — grid already has 8F+3C
  18. c4846a39 20:08:28 parameter-golf-lab Symmetry ablation: 4x4 only — 6x6 and 8x8 impractical at 785ms/step
  19. c4846a39 20:08:28 sota_crawler Symmetry ablation: 4x4 only — 6x6 and 8x8 impractical at 785ms/step
  20. c4846a39 20:08:28 SOTA_FINAL Symmetry ablation: 4x4 only — 6x6 and 8x8 impractical at 785ms/step
  21. 75e07222 21:01:34 parameter-golf-lab multi-crawler descendant 1: 8F+3C production gate (two-variable diff from BWX 9F)
  22. 75e07222 21:01:34 sota_crawler multi-crawler descendant 1: 8F+3C production gate (two-variable diff from BWX 9F)
  23. 75e07222 21:01:34 SOTA_FINAL multi-crawler descendant 1: 8F+3C production gate (two-variable diff from BWX 9F)
  24. 22c7e07f 21:05:10 parameter-golf-lab TK1 gate: safe (3 loops) vs aggressive (4 loops diff battery) on 4xGPU 2k steps
  25. 22c7e07f 21:05:10 sota_crawler TK1 gate: safe (3 loops) vs aggressive (4 loops diff battery) on 4xGPU 2k steps
  26. 22c7e07f 21:05:10 SOTA_FINAL TK1 gate: safe (3 loops) vs aggressive (4 loops diff battery) on 4xGPU 2k steps
  27. dd2166e1 21:13:22 parameter-golf-lab TK1 gate: switch to 600s wallclock (matches competition 10-min budget)
  28. dd2166e1 21:13:22 sota_crawler TK1 gate: switch to 600s wallclock (matches competition 10-min budget)
  29. dd2166e1 21:13:22 SOTA_FINAL TK1 gate: switch to 600s wallclock (matches competition 10-min budget)
  30. 29ff99e9 21:13:49 parameter-golf-lab TK1 gate: 300s wallclock for faster screening on 4xGPU
  31. 29ff99e9 21:13:49 sota_crawler TK1 gate: 300s wallclock for faster screening on 4xGPU
  32. 29ff99e9 21:13:49 SOTA_FINAL TK1 gate: 300s wallclock for faster screening on 4xGPU
  33. 81b6bd7a 21:14:54 parameter-golf-lab Update README leaderboard for April records
  34. 81b6bd7a 21:14:54 sota_crawler Update README leaderboard for April records
  35. 81b6bd7a 21:14:54 SOTA_FINAL Update README leaderboard for April records
  36. eadf2aae 22:25:58 parameter-golf-lab TK1: add Brotli compression packaging compression packaging compression (default, quality=11) — zstd artifact was 17.95MB
  37. eadf2aae 22:25:58 sota_crawler TK1: add Brotli compression packaging compression packaging compression (default, quality=11) — zstd artifact was 17.95MB
  38. eadf2aae 22:25:58 SOTA_FINAL TK1: add Brotli compression packaging compression packaging compression (default, quality=11) — zstd artifact was 17.95MB
  39. 868f0785 22:36:24 parameter-golf-lab Update SCIENCE.md and TK1 ablation with production gate results
  40. 868f0785 22:36:24 sota_crawler Update SCIENCE.md and TK1 ablation with production gate results
  41. 868f0785 22:36:24 SOTA_FINAL Update SCIENCE.md and TK1 ablation with production gate results
  42. 45b27848 23:16:18 parameter-golf-lab long-context causal tap mechanism v1.9.1 stabilize adaptive TTT and drift-lite arms
  43. 45b27848 23:16:18 sota_crawler long-context causal tap mechanism v1.9.1 stabilize adaptive TTT and drift-lite arms
  44. 45b27848 23:16:18 SOTA_FINAL long-context causal tap mechanism v1.9.1 stabilize adaptive TTT and drift-lite arms
  45. 59109e30 23:37:05 parameter-golf-lab feat(records): SP8192 + Muon 0.97 + Legal TTT + Causal N-gram Tilt — val_bpb 1.07983
  46. 85b332cd 02:02:09 sota_nueral Add 12L GQA mixed-bit tower (safe) and 12L GQA mixed-bit tower (aggressive) legs
  47. 848244cc 02:18:07 parameter-golf-lab BW_9F2C: 9F+2C production gate — single-variable NUM_CRAWLER_LAYERS=2
  48. 848244cc 02:18:07 sota_crawler BW_9F2C: 9F+2C production gate — single-variable NUM_CRAWLER_LAYERS=2
  49. 848244cc 02:18:07 SOTA_FINAL BW_9F2C: 9F+2C production gate — single-variable NUM_CRAWLER_LAYERS=2
  50. 885cf584 02:19:16 parameter-golf-lab BW_9F2C: add Brotli compression packaging compression packaging compression (default, quality=11)
  51. 885cf584 02:19:16 sota_crawler BW_9F2C: add Brotli compression packaging compression packaging compression (default, quality=11)
  52. 885cf584 02:19:16 SOTA_FINAL BW_9F2C: add Brotli compression packaging compression packaging compression (default, quality=11)
  53. b7c794c4 02:31:53 parameter-golf-lab Add 12L GQA mixed-bit tower (safe) and 12L GQA mixed-bit tower (aggressive) for 8xH100 run
  54. b7c794c4 02:31:53 sota_crawler Add 12L GQA mixed-bit tower (safe) and 12L GQA mixed-bit tower (aggressive) for 8xH100 run
  55. b7c794c4 02:31:53 SOTA_FINAL Add 12L GQA mixed-bit tower (safe) and 12L GQA mixed-bit tower (aggressive) for 8xH100 run
  56. 8966f96b 02:48:22 parameter-golf-lab BW_9F2C: downgrade FA3 preflight from fatal to warning
  57. 8966f96b 02:48:22 sota_crawler BW_9F2C: downgrade FA3 preflight from fatal to warning
  58. 8966f96b 02:48:22 SOTA_FINAL BW_9F2C: downgrade FA3 preflight from fatal to warning
  59. 1ede5637 02:52:02 parameter-golf-lab BW_9F2C: FA3 preflight warns instead of exits — SDPA fallback is in train_gpt.py
  60. 1ede5637 02:52:02 sota_crawler BW_9F2C: FA3 preflight warns instead of exits — SDPA fallback is in train_gpt.py
  61. 1ede5637 02:52:02 SOTA_FINAL BW_9F2C: FA3 preflight warns instead of exits — SDPA fallback is in train_gpt.py
  62. 9912ff54 03:33:13 sota_nueral Add run_12L GQA mixed-bit tower.sh — launch any experiment with 12L GQA mixed-bit tower 12L env vars
  63. 8f3dcd86 03:34:23 parameter-golf-lab Add run_12L GQA mixed-bit tower.sh — 12L GQA mixed-bit tower 12L env vars for any experiment
  64. 8f3dcd86 03:34:23 sota_crawler Add run_12L GQA mixed-bit tower.sh — 12L GQA mixed-bit tower 12L env vars for any experiment
  65. 8f3dcd86 03:34:23 SOTA_FINAL Add run_12L GQA mixed-bit tower.sh — 12L GQA mixed-bit tower 12L env vars for any experiment
  66. 2a83653c 03:41:27 parameter-golf-lab Fix 12L GQA mixed-bit tower 12L defaults in GPTQ and Black experiments
  67. 2a83653c 03:41:27 sota_crawler Fix 12L GQA mixed-bit tower 12L defaults in GPTQ and Black experiments
  68. 2a83653c 03:41:27 SOTA_FINAL Fix 12L GQA mixed-bit tower 12L defaults in GPTQ and Black experiments
  69. 471d793e 03:41:40 sota_nueral Fix vault defaults to match 12L GQA mixed-bit tower 12L actual config
  70. da1c5352 04:07:40 sota_nueral Fix vault defaults: COMPRESSOR=Brotli compression packaging compression packaging, COPRIME_MAX_LOADED_SHARDS=4
  71. 7ab8d0c7 04:07:47 parameter-golf-lab Fix GPTQ+Black: COMPRESSOR=Brotli compression packaging compression packaging, WARMDOWN=3500, cache=4
  72. 7ab8d0c7 04:07:47 sota_crawler Fix GPTQ+Black: COMPRESSOR=Brotli compression packaging compression packaging, WARMDOWN=3500, cache=4
  73. 7ab8d0c7 04:07:47 SOTA_FINAL Fix GPTQ+Black: COMPRESSOR=Brotli compression packaging compression packaging, WARMDOWN=3500, cache=4
DATE2026-04-09
DAY023
PRIMARY METRICVAL_BPB
TESTS RUN23 unique (9 valid / 0 TTT leakage)
ARCHITECTURE FAMILY12L GQA MIXED-BIT TOWER (×14) · CRAWLER-SAFE TRAIN/EVAL SURFACE (×9)
CLASSIFICATIONEXPERIMENTAL 39% (9)/NEURAL 61% (14)
ARCHITECTURE RECIPE
12L GQA MIXED-BIT TOWERx14

12-layer GQA tower: 8/4 GQA, RoPE-16, Bigram 2048, XSA on the last 11 layers, and mixed-bit packaging.

BODYnum_layers=12 · num_heads=8 · num_kv_heads=4 · rope_dims=16 · xsa_last_n=11 · BigramHash vocab=2048
QUANT / EVALattention int5 · mlp int6 · aux int6 · embed int8 · other int8 · Brotli artifact · 16 MB artifact cap
RUN SHAPE8xH100 · train_seq_len=2048 · eval_seq_len=2048 · 600s wallclock
CRAWLER-SAFE TRAIN/EVAL SURFACEx9

Crawler-safe training/eval surface around a widened shared-crawler stack with final int6 and int8 checks.

BODYvocab_size=1024 · model_dim=512 · num_heads=8 · num_kv_heads=4 · shared crawler block · looped crawler passes
FEATURESBigramHash · XSA · RoPE · crawler loop rope scales · instruction bottleneck
QUANT / EVALint6 roundtrip · int6 sliding-window eval · int8 zlib roundtrip · 8xH100 · 600s wallclock

SECTION 1.0 — TRAJECTORY

VAL_BPB for 2026-04-09
VAL_BPB SPREADnminmedianmaxdownload
CRAWLER-SAFE TRAIN/EVAL SURFACE91.31771.32771.3310download
ALL VALID91.31771.32771.3310

SECTION 2.0 — TEST RUNS

EXPERIMENTAL TEST RUNS — 9

#TIMEHWABLATIONVAL_BPBSTATUSΔ %OPEN
1302:444×H100A00_screen_s444_20260409_024421.log1.3300OK↑24.6%
training_log
1402:444×H100A06_screen_s444_20260409_024421.log1.3310OK↑24.7%
training_log
1502:444×H100A08_screen_s444_20260409_024421.log1.3292OK↑24.5%
training_log
1602:444×H100A05_screen_s444_20260409_024421.log1.3260OK↑24.2%
training_log
1702:444×H100A07_screen_s444_20260409_024421.log1.3177OK↑23.5%
training_log
1802:444×H100A04_screen_s444_20260409_024421.log1.3253OK↑24.2%
training_log
1902:444×H100A03_screen_s444_20260409_024421.log1.3260OK↑24.3%
training_log
2002:444×H100A01_screen_s444_20260409_024421.log1.3288OK↑24.5%
training_log
2102:444×H100A02_screen_s444_20260409_024421.log1.3277OK↑24.4%
training_log

NEURAL TEST RUNS — 14

#TIMEHWABLATIONVAL_BPBSTATUSΔ %OPEN
101:144×H10012L GQA mixed-bit tower.log3.3780CRASH
training_log
201:304×H10012L GQA mixed-bit tower.log4.1059CRASH
training_log
301:374×H10012L GQA mixed-bit tower.log3.4311CRASH
training_log
401:434×H10012L GQA mixed-bit tower.log3.4325CRASH
training_log
501:494×H10012L GQA mixed-bit tower.log3.4741CRASH
training_log
601:554×H10012L GQA mixed-bit tower.log3.4317CRASH
training_log
702:014×H10012L GQA mixed-bit tower.log3.4226CRASH
training_log
802:074×H10012L GQA mixed-bit tower.log3.4294CRASH
training_log
902:124×H10012L GQA mixed-bit tower.log3.4321CRASH
training_log
1002:174×H10012L GQA mixed-bit tower.log3.4319CRASH
training_log
1102:234×H10012L GQA mixed-bit tower.log3.6905CRASH
training_log
1202:294×H10012L GQA mixed-bit tower.log3.4345CRASH
training_log
2211:024×H10012L GQA mixed-bit tower.log3.4067CRASH
training_log
2311:074×H10012L GQA mixed-bit tower.log3.4602CRASH
training_log