DAILY RESEARCH REPORTBACK TO MASTER SHEET2026-04-28

val_bpb · 112 of 640 · 0 warnings

audit summary

{
  "date": "2026-04-28",
  "fatal": 0,
  "headline_metric": "val_bpb",
  "raw_records": 640,
  "render_records": 112,
  "test_rows": 9,
  "warnings": 0
}

8 commits on 2026-04-28

1e439663 15:35:11 parameter-golf-lab No code changes, update README to clarify that lrzip is a setup prerequisite (not auto-installed by the training script)
6ee5435e 18:37:20 parameter-golf-lab Update parameter golf leaderboard
02cdca88 18:37:24 parameter-golf-lab Update parameter golf leaderboard
3a7c4be0 18:57:17 parameter-golf-lab Update parameter golf leaderboard with BOS fix
0d046475 19:30:42 parameter-golf-lab Credit PR 1797 in leaderboard update
ce84ddcc 19:32:02 parameter-golf-lab Credit CaseOps and PR 1787 leaderboard rows
13e76acd 20:27:54 parameter-golf-lab Apply p-value progression leaderboard cutoff
dbc31de8 20:32:55 parameter-golf-lab Address leaderboard review comments

DATE2026-04-28

DAY042

PRIMARY METRICVAL_BPB

TESTS RUN9 unique (1 valid / 8 TTT leakage)

ARCHITECTURE FAMILYLOOPED TOKENIZER BODY (×3) · STATE-SPACE TTT VARIANT (×3) · 11L GQA U-NET BODY (×2) · 9-FLAT PLUS 1-CRAWLER LOOP BODY (×1)

CLASSIFICATIONEXPERIMENTAL 11% (1)/NEURAL 89% (8)

ARCHITECTURE RECIPE

11L GQA U-NET BODYx2

11-layer dim-512 GQA U-Net transformer base with XSA, BigramHash, value embedding, and SmearGate.

BODYvocab_size=1024 · num_layers=11 · model_dim=512 · num_heads=8 · num_kv_heads=4 · head_dim=64 · mlp_mult=3.0 · tie_embeddings=1

STRUCTUREencoder_layers=5 · decoder_layers=6 · skip_weights=5 · GQA attention · RMSNorm · relu_sq MLP · Muon matrix optimizer path

FEATURESBigramHash vocab=2048 dim=128 · XSA last_n=11 · rope_dims=16 · VE dim=128 layers=9,10 · SmearGate · logit_softcap=30

9-FLAT + 1-CRAWLER LOOP BODYx1

9-flat plus 1-crawler loop body: flat U-Net context around a shared crawler block fired twice.

BODYvocab_size=1024 · model_dim=512 · num_heads=8 · num_kv_heads=4 · NUM_FLAT_LAYERS=9 · NUM_CRAWLER_LAYERS=1 · CRAWLER_LOOPS=2 · USE_CRAWLER=1

CRAWLERCRAWLER_MLP_MULT=6.0 · INST_DIM=32 · CRAWLER_MLP_CHOKE_DIM=0 · CRAWLER_LOOP_ROPE_SCALES=9,1,1 · CRAWLER_MLP_LEAKY_SLOPE=0.5

PACKAGINGLOOP_AWARE_GPTQ=1 · BigramHash 2048x128 · XSA last_n=11 · rope_dims=16 · VE dim=128 layers=9,10

SECTION 1.0 — TRAJECTORY

DAILY SUMMARY

Experimental work: 9-FLAT + 1-CRAWLER LOOP BODY (x1): 9-flat plus 1-crawler loop body: flat U-Net context around a shared crawler block fired twice. Specs: model_dim=512; vocab_size=1024; num_heads=8; num_kv_heads=4; rope_dims=16.

Neural work: LOOPED TOKENIZER BODY (x3): architecture family detected in measured rows; exact flags are preserved in the row details. STATE-SPACE TTT VARIANT (x3): architecture family detected in measured rows; exact flags are preserved in the row details. 11L GQA U-NET BODY (x2): 11-layer dim-512 GQA U-Net transformer base with XSA, BigramHash, value embedding, and SmearGate. Specs: num_layers=11; model_dim=512; vocab_size=1024; num_heads=8; num_kv_heads=4; head_dim=64.

Day 042 measured 9 architecture/test rows against VAL_BPB, with 1 valid rows and 8 TTT leakage readings excluded. Techniques to copy/reference: SmearGate; BigramHash 2048x128.

No new personal best. Best valid of the day = 1.1364 (9-flat plus 1-crawler loop body); best-before-today = 1.0672 still stands.

VAL_BPB SPREAD	n	min	median	max	download
11L GQA U-NET BODY	1	1.1364	1.1364	1.1364	download
ALL VALID	1	1.1364	1.1364	1.1364	—

SECTION 2.0 — TEST RUNS

EXPERIMENTAL TEST RUNS — 1

#	TIME	HW	ABLATION	VAL_BPB	STATUS	Δ %	OPEN
8	17:42	—	9-flat plus 1-crawler loop body (×2)	1.1364	OK	↑6.5%	details

NEURAL TEST RUNS — 8

#	TIME	HW	ABLATION	VAL_BPB	STATUS	Δ %	OPEN
1	15:39	8×H100	looped tokenizer body.log	`0.9086`	INVALID	—	training_log

2	15:57	8×H100	looped tokenizer body.log	`0.9127`	INVALID	—	training_log

3	16:14	8×H100	looped tokenizer body.log	`0.9075`	INVALID	—	training_log

4	16:37	8×H100	loop+TTT compact body.log	`0.9042`	INVALID	—	training_log

5	16:53	8×H100	loop+TTT compact body.log	`0.9036`	INVALID	—	training_log

6	17:13	8×H100	loop+TTT compact body.log	`0.9039`	INVALID	—	training_log

7	17:27	8×H100	loop+TTT compact body.log	`0.9011`	INVALID	—	training_log

9	17:47	8×H100	loop+TTT compact body.log	`0.9036`	INVALID	—	training_log