WP31: locked test outcome, findings, abstract

modified PROGRESS.md +1 −0

				`@@ -20,3 +20,4 @@`
20	\| 2026-10-30 \| 6 ranking-loss controller (D12, tiny v1 editor, editor-unseen labels) \| done \| `controller_v1rank.yaml` → `results/raw/routing_v1rank.json` \| locked λ=0.3 κ=0.5: dev-iid loss 0.4333 vs 0.4334 (0.07 edits/token); λ=0: +1.3 % \| H3 null stands \| — \|	20	\| 2026-10-30 \| 6 ranking-loss controller (D12, tiny v1 editor, editor-unseen labels) \| done \| `controller_v1rank.yaml` → `results/raw/routing_v1rank.json` \| locked λ=0.3 κ=0.5: dev-iid loss 0.4333 vs 0.4334 (0.07 edits/token); λ=0: +1.3 % \| H3 null stands \| — \|
21	\| 2026-10-30 \| 4 dense-writes ablation (v2 backbone, 5 seeds) \| done \| `p4v2_dense.yaml` → `results/analyzed/seed_matrix_v2.json` \| dev-iid loss −2.5/−3.2/−3.2/−1.6 % at b=1/2/4/8 (vs sparse −2.1/−2.8/−3.0/−1.0, GR −3.6/−4.1/−3.7/−3.3); EM +2.0 pp (b1), t-CI [1.6, 2.4] \| sparse local writes ≤ dense (H2 second half null; 5 seeds: −2.4/−3.1/−3.1/−1.5 %, EM +2.0 pp [1.6, 2.4]) \| — \|	21	\| 2026-10-30 \| 4 dense-writes ablation (v2 backbone, 5 seeds) \| done \| `p4v2_dense.yaml` → `results/analyzed/seed_matrix_v2.json` \| dev-iid loss −2.5/−3.2/−3.2/−1.6 % at b=1/2/4/8 (vs sparse −2.1/−2.8/−3.0/−1.0, GR −3.6/−4.1/−3.7/−3.3); EM +2.0 pp (b1), t-CI [1.6, 2.4] \| sparse local writes ≤ dense (H2 second half null; 5 seeds: −2.4/−3.1/−3.1/−1.5 %, EM +2.0 pp [1.6, 2.4]) \| — \|
22	\| 2026-10-30 \| 8–10 pilot stages B–D + routing + latency (103M) \| done \| `editor_pilot_103m_lr3e4` (3000 steps, 90 min, peak 25.9 GB), `labels_103m` (491k labels, 32,768 audit states), `controller_103m`, `routing_103m_val.json`, `latency_103m_mps.json`, `eval_gr_103m_val.json` \| editor −0.04/−0.06/−0.08/−0.06 % at b=1/2/4/8; GR −0.08 % (ppl 48.16 → 48.01); controller locked (λ=0.3, κ=0.5) −0.01 % with 0.61 edits/token, calib sign acc 51 %, captured 7 %; oracle b8 −3.6 % (4.3 edits/token); two-step oracle 0.0162 vs one-step 0.0088 nats; latency batch 1: 7.55 ms (b0) → 22.9 (8 single) / 26.9 (8 dense) / 11.0 (controller) ms; dense 103M probe collapsed twice \| H1 not supported on web text; H3 null; dispatch overhead still dominant (sparse 3.0× vs dense 3.6× measured, analytic 1.6× vs 3.1×) \| enlarged baseline eval, final audit \|	22	\| 2026-10-30 \| 8–10 pilot stages B–D + routing + latency (103M) \| done \| `editor_pilot_103m_lr3e4` (3000 steps, 90 min, peak 25.9 GB), `labels_103m` (491k labels, 32,768 audit states), `controller_103m`, `routing_103m_val.json`, `latency_103m_mps.json`, `eval_gr_103m_val.json` \| editor −0.04/−0.06/−0.08/−0.06 % at b=1/2/4/8; GR −0.08 % (ppl 48.16 → 48.01); controller locked (λ=0.3, κ=0.5) −0.01 % with 0.61 edits/token, calib sign acc 51 %, captured 7 %; oracle b8 −3.6 % (4.3 edits/token); two-step oracle 0.0162 vs one-step 0.0088 nats; latency batch 1: 7.55 ms (b0) → 22.9 (8 single) / 26.9 (8 dense) / 11.0 (controller) ms; dense 103M probe collapsed twice \| H1 not supported on web text; H3 null; dispatch overhead still dominant (sparse 3.0× vs dense 3.6× measured, analytic 1.6× vs 3.1×) \| enlarged baseline eval, final audit \|
		23	+\| 2026-10-30 \| 9 locked test evaluation (synthetic, pre-registered primary outcome) \| done \| prereg frozen v1.0 04:04, manifest unsealed 04:04; `test_eval_v2_s.yaml` on 5 nodes → `results/raw/eval_p4v2__test_*.json`, `results/analyzed/test_statistics.json` \| pooled EM (iid+template+depth) b2 vs b0: PRISME +0.9 pp [+0.9, +1.0]; PRISME − GR −0.3 [−0.6, +0.1]; PRISME − dense +0.4 [+0.1, +0.7]; compose 4–5 % all systems \| pre-registered threshold (3 pp) NOT met → do not advance \| FineWeb test perplexity (4 systems) running \|

modified configs/experiments/locked_evaluation.yaml +2 −1

				`@@ -4,6 +4,7 @@ data:`
4	dir: data/processed/fineweb_edu_300m	4	dir: data/processed/fineweb_edu_300m
5	splits:	5	splits:
6	- val	6	- val
		7	+- test
7	seq_len: 512	8	seq_len: 512
8	batch_size: 16	9	batch_size: 16
9	n_batches: 40	10	n_batches: 40
				`@@ -22,7 +23,7 @@ systems:`
22	enlarged_102m:	23	enlarged_102m:
23	mode: backbone_only	24	mode: backbone_only
24	model: from_checkpoint	25	model: from_checkpoint
25	− checkpoint: ENLARGED_CKPT	26	+ checkpoint: experiments/runs/pretrain_enlarged102m_seed01_20261030015017/checkpoint
26	controller: false	27	controller: false
27	prisme_103m_editor:	28	prisme_103m_editor:
28	mode: prisme	29	mode: prisme

added configs/experiments/test_eval_v2_s1.yaml +43 −0

++out_dir: results/raw
++data:
++  kind: synthetic
++  tokenizer: char
++  splits_dir: data/splits/synthetic_v1
++splits:
++- test_iid
++- test_template
++- test_depth
++- test_compose
++batch_size: 64
++seq_len: 320
++max_items: 1000
++systems:
++  p4v2_prisme_s1:
++    mode: prisme
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_prisme_seed01_20261030001749/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 1
++  p4v2_gr_s1:
++    mode: global_recurrence
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_gr_seed01_20261030001749/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 1
++    baseline_kwargs:
++      ffn_w: 640
++  p4v2_dense_s1:
++    mode: dense
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_dense_seed01_20261030013421/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 1

added configs/experiments/test_eval_v2_s2.yaml +43 −0

++out_dir: results/raw
++data:
++  kind: synthetic
++  tokenizer: char
++  splits_dir: data/splits/synthetic_v1
++splits:
++- test_iid
++- test_template
++- test_depth
++- test_compose
++batch_size: 64
++seq_len: 320
++max_items: 1000
++systems:
++  p4v2_prisme_s2:
++    mode: prisme
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_prisme_seed02_20261030001749/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 2
++  p4v2_gr_s2:
++    mode: global_recurrence
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_gr_seed02_20261030001749/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 2
++    baseline_kwargs:
++      ffn_w: 640
++  p4v2_dense_s2:
++    mode: dense
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_dense_seed02_20261030004544/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 2

added configs/experiments/test_eval_v2_s3.yaml +43 −0

++out_dir: results/raw
++data:
++  kind: synthetic
++  tokenizer: char
++  splits_dir: data/splits/synthetic_v1
++splits:
++- test_iid
++- test_template
++- test_depth
++- test_compose
++batch_size: 64
++seq_len: 320
++max_items: 1000
++systems:
++  p4v2_prisme_s3:
++    mode: prisme
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_prisme_seed03_20261030001749/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 3
++  p4v2_gr_s3:
++    mode: global_recurrence
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_gr_seed03_20261030001749/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 3
++    baseline_kwargs:
++      ffn_w: 640
++  p4v2_dense_s3:
++    mode: dense
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_dense_seed03_20261030004544/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 3

added configs/experiments/test_eval_v2_s4.yaml +43 −0

++out_dir: results/raw
++data:
++  kind: synthetic
++  tokenizer: char
++  splits_dir: data/splits/synthetic_v1
++splits:
++- test_iid
++- test_template
++- test_depth
++- test_compose
++batch_size: 64
++seq_len: 320
++max_items: 1000
++systems:
++  p4v2_prisme_s4:
++    mode: prisme
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_prisme_seed04_20261030001749/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 4
++  p4v2_gr_s4:
++    mode: global_recurrence
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_gr_seed04_20261030001749/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 4
++    baseline_kwargs:
++      ffn_w: 640
++  p4v2_dense_s4:
++    mode: dense
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_dense_seed04_20261030001749/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 4

added configs/experiments/test_eval_v2_s5.yaml +43 −0

++out_dir: results/raw
++data:
++  kind: synthetic
++  tokenizer: char
++  splits_dir: data/splits/synthetic_v1
++splits:
++- test_iid
++- test_template
++- test_depth
++- test_compose
++batch_size: 64
++seq_len: 320
++max_items: 1000
++systems:
++  p4v2_prisme_s5:
++    mode: prisme
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_prisme_seed05_20261030001839/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 5
++  p4v2_gr_s5:
++    mode: global_recurrence
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_gr_seed05_20261030001839/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 5
++    baseline_kwargs:
++      ffn_w: 640
++  p4v2_dense_s5:
++    mode: dense
++    model: from_checkpoint
++    checkpoint: experiments/runs/p4v2_dense_seed05_20261030004544/checkpoint
++    controller: false
++    policies:
++    - budget0
++    - forced_b2
++    seed: 5

added experiments/runs/pretrain_enlarged102m_seed01_20261030015017/config.yaml +52 −0

++experiment:
++  name: pretrain_enlarged102m
++  seed: 1
++  model: configs/baselines/transformer_enlarged_102m.yaml
++  data:
++    train_bin: data/processed/fineweb_edu_300m/train.bin
++    val_bin: data/processed/fineweb_edu_300m/val.bin
++  train: &id001
++    batch_size: 32
++    seq_len: 512
++    steps: 6000
++    lr: 0.0006
++    warmup: 300
++    weight_decay: 0.1
++    grad_clip: 1.0
++    log_every: 20
++    eval_every: 500
++    eval_batches: 20
++    ckpt_every: 1000
++model:
++  name: transformer_enlarged_102m
++  vocab_size: 32768
++  d_model: 768
++  n_layers: 12
++  n_heads: 12
++  n_kv_heads: 3
++  ffn_width: 2144
++  max_seq_len: 4096
++  rope_theta: 10000.0
++  norm_eps: 1.0e-05
++  tie_embeddings: true
++  init_std: 0.02
++  use_workspace: false
++  w: 256
++  ffn_w: 768
++  n_slots: 16
++  n_editor_blocks: 2
++  n_editor_heads: 4
++  r_max: 8
++  eta_max: 0.25
++  clip_c: 1.0
++  clip_eps: 1.0e-06
++  readout_zero_init: true
++  readout_init_std: 0.02
++  readout_scale: 1.0
++  controller_hidden: 0
++  ctrl_input: z
++  ctrl_use_entropy: false
++  lambda_cost: 0.0
++  kappa_risk: 0.0
++  dense_writes: false
++resolved_train: *id001

added experiments/runs/pretrain_enlarged102m_seed01_20261030015017/manifest.json +26 −0

++{
++  "run_id": "pretrain_enlarged102m_seed01_20261030015017",
++  "name": "pretrain_enlarged102m",
++  "seed": 1,
++  "code_commit": "unknown",
++  "config_sha256": "efbc5ac5e13ac94c05ef6d273e2701851307b5066ea5160e171e89ea81bb83ed",
++  "started_at": "2026-10-30T01:50:17-0400",
++  "host": "m3u96b",
++  "python": "3.12.15",
++  "torch": "2.15.0",
++  "command": "/Users/simon-pierreboucher/wp31/prisme-slm/training/pretrain.py --config configs/experiments/pretrain_enlarged_102m.yaml --device mps",
++  "state": "completed",
++  "wall_seconds": 8527.03,
++  "device": "mps",
++  "backend": "mps",
++  "precision": "fp32",
++  "n_params": 102157056,
++  "dataset_manifest_sha256": "d8753bdfa9609fce56a7ec2da7c108666185d021e3f921721e4d2241e1d4fe54",
++  "checkpoint_parent": null,
++  "finished_at": "2026-10-30T04:12:24-0400",
++  "train_tokens": 98304000,
++  "estimated_train_flops": 60254683398144000,
++  "peak_memory_bytes": 28522004480,
++  "final_val_loss": 3.8689021825790406,
++  "metrics_file": "experiments/runs/pretrain_enlarged102m_seed01_20261030015017/metrics.jsonl"
++}
+ \ No newline at end of file

added experiments/runs/pretrain_enlarged102m_seed01_20261030015017/metrics.jsonl +319 −0

++{"step": 0, "loss": 10.538154602050781, "lr": 2e-06, "grad_norm": 8.826038360595703, "tokens": 16384, "tok_per_s": 12267.008820282133, "t": 1.98}
++{"step": 20, "loss": 9.062786102294922, "lr": 4.2e-05, "grad_norm": 2.396116018295288, "tokens": 344064, "tok_per_s": 12236.186788926587, "t": 28.76}
++{"step": 40, "loss": 8.023894309997559, "lr": 8.199999999999999e-05, "grad_norm": 1.9327870607376099, "tokens": 671744, "tok_per_s": 12224.43886163398, "t": 55.59}
++{"step": 60, "loss": 7.147858619689941, "lr": 0.00012199999999999998, "grad_norm": 1.0073579549789429, "tokens": 999424, "tok_per_s": 12227.853585676065, "t": 82.38}
++{"step": 80, "loss": 6.872986793518066, "lr": 0.00016199999999999998, "grad_norm": 0.9701820015907288, "tokens": 1327104, "tok_per_s": 12350.690989394596, "t": 108.09}
++{"step": 100, "loss": 6.5623087882995605, "lr": 0.00020199999999999998, "grad_norm": 1.135318398475647, "tokens": 1654784, "tok_per_s": 12246.719225784744, "t": 135.76}
++{"step": 120, "loss": 6.4833784103393555, "lr": 0.000242, "grad_norm": 0.9940797090530396, "tokens": 1982464, "tok_per_s": 12290.734902855935, "t": 161.94}
++{"step": 140, "loss": 6.285987854003906, "lr": 0.00028199999999999997, "grad_norm": 0.9160900115966797, "tokens": 2310144, "tok_per_s": 12263.175739855946, "t": 189.02}
++{"step": 160, "loss": 6.127978324890137, "lr": 0.00032199999999999997, "grad_norm": 0.8789034485816956, "tokens": 2637824, "tok_per_s": 12244.953978378258, "t": 216.06}
++{"step": 180, "loss": 6.162482261657715, "lr": 0.00036199999999999996, "grad_norm": 1.3229773044586182, "tokens": 2965504, "tok_per_s": 12291.164784312696, "t": 241.91}
++{"step": 200, "loss": 6.06372594833374, "lr": 0.00040199999999999996, "grad_norm": 0.9360699653625488, "tokens": 3293184, "tok_per_s": 12260.65061054621, "t": 269.24}
++{"step": 220, "loss": 5.960588455200195, "lr": 0.000442, "grad_norm": 0.9670940637588501, "tokens": 3620864, "tok_per_s": 12250.598133371244, "t": 296.21}
++{"step": 240, "loss": 5.984297752380371, "lr": 0.00048199999999999995, "grad_norm": 0.8978189826011658, "tokens": 3948544, "tok_per_s": 12291.705053241867, "t": 321.88}
++{"step": 260, "loss": 5.8681182861328125, "lr": 0.000522, "grad_norm": 0.8115763068199158, "tokens": 4276224, "tok_per_s": 12275.620218191996, "t": 348.99}
++{"step": 280, "loss": 5.783034324645996, "lr": 0.0005619999999999999, "grad_norm": 0.8142603635787964, "tokens": 4603904, "tok_per_s": 12285.887365268558, "t": 375.37}
++{"step": 300, "loss": 5.728802680969238, "lr": 0.0006, "grad_norm": 0.7503604292869568, "tokens": 4931584, "tok_per_s": 12283.22840230745, "t": 402.13}
++{"step": 320, "loss": 5.7670793533325195, "lr": 0.0005999835963914726, "grad_norm": 1.6453608274459839, "tokens": 5259264, "tok_per_s": 12274.431994656205, "t": 429.12}
++{"step": 340, "loss": 5.742429733276367, "lr": 0.0005999343875590635, "grad_norm": 0.7031630277633667, "tokens": 5586944, "tok_per_s": 12280.981116658615, "t": 455.57}
++{"step": 360, "loss": 5.577595233917236, "lr": 0.00059985237948205, "grad_norm": 0.8332432508468628, "tokens": 5914624, "tok_per_s": 12268.441190882228, "t": 482.74}
++{"step": 380, "loss": 5.530177116394043, "lr": 0.0005997375821250868, "grad_norm": 0.6786928176879883, "tokens": 6242304, "tok_per_s": 12279.671362677862, "t": 508.99}
++{"step": 400, "loss": 5.6273932456970215, "lr": 0.0005995900094369953, "grad_norm": 6.397047519683838, "tokens": 6569984, "tok_per_s": 12264.629455104921, "t": 536.33}
++{"step": 420, "loss": 5.520440101623535, "lr": 0.000599409679349069, "grad_norm": 0.6246231198310852, "tokens": 6897664, "tok_per_s": 12287.867050340814, "t": 561.98}
++{"step": 440, "loss": 5.382430076599121, "lr": 0.0005991966137728945, "grad_norm": 0.761988639831543, "tokens": 7225344, "tok_per_s": 12279.624106426805, "t": 589.04}
++{"step": 460, "loss": 5.283364772796631, "lr": 0.0005989508385976885, "grad_norm": 0.6207935214042664, "tokens": 7553024, "tok_per_s": 12270.523383262516, "t": 616.18}
++{"step": 480, "loss": 5.33247184753418, "lr": 0.000598672383687153, "grad_norm": 0.5935987234115601, "tokens": 7880704, "tok_per_s": 12246.607024038762, "t": 644.14}
++{"step": 499, "val_loss": 5.376418542861939, "val_ppl": 216.246409641857, "tokens": 8192000, "t": 676.9}
++{"step": 500, "loss": 5.335311412811279, "lr": 0.0005983612828758463, "grad_norm": 0.6082054972648621, "tokens": 8208384, "tok_per_s": 12117.563445660973, "t": 678.04}
++{"step": 520, "loss": 5.3863677978515625, "lr": 0.0005980175739650714, "grad_norm": 0.6699115037918091, "tokens": 8536064, "tok_per_s": 12089.311821260695, "t": 706.73}
++{"step": 540, "loss": 5.296011924743652, "lr": 0.0005976412987182833, "grad_norm": 0.757933497428894, "tokens": 8863744, "tok_per_s": 12075.933905206339, "t": 734.64}
++{"step": 560, "loss": 5.353665351867676, "lr": 0.000597232502856014, "grad_norm": 0.6468153595924377, "tokens": 9191424, "tok_per_s": 12053.971988970368, "t": 763.16}
++{"step": 580, "loss": 5.164031982421875, "lr": 0.0005967912360503179, "grad_norm": 0.5246732831001282, "tokens": 9519104, "tok_per_s": 12039.288807349485, "t": 791.31}
++{"step": 600, "loss": 5.383706092834473, "lr": 0.000596317551918735, "grad_norm": 0.5707435607910156, "tokens": 9846784, "tok_per_s": 12027.311249556898, "t": 819.34}
++{"step": 620, "loss": 5.237800598144531, "lr": 0.0005958115080177768, "grad_norm": 0.5295183658599854, "tokens": 10174464, "tok_per_s": 12017.056818105553, "t": 847.31}
++{"step": 640, "loss": 5.215066909790039, "lr": 0.0005952731658359328, "grad_norm": 0.548611581325531, "tokens": 10502144, "tok_per_s": 11992.821053121737, "t": 876.34}
++{"step": 660, "loss": 5.153785705566406, "lr": 0.0005947025907861984, "grad_norm": 0.5935801863670349, "tokens": 10829824, "tok_per_s": 11977.674335609541, "t": 904.81}
++{"step": 680, "loss": 5.202521324157715, "lr": 0.0005940998521981274, "grad_norm": 0.5564954876899719, "tokens": 11157504, "tok_per_s": 11959.257592686838, "t": 933.6}
++{"step": 700, "loss": 5.146511554718018, "lr": 0.0005934650233094076, "grad_norm": 0.5552934408187866, "tokens": 11485184, "tok_per_s": 11941.760258932192, "t": 962.41}
++{"step": 720, "loss": 5.095799446105957, "lr": 0.0005927981812569615, "grad_norm": 0.567599892616272, "tokens": 11812864, "tok_per_s": 11935.94153600311, "t": 990.33}
++{"step": 740, "loss": 4.996639251708984, "lr": 0.0005920994070675743, "grad_norm": 0.578127920627594, "tokens": 12140544, "tok_per_s": 11925.37125969469, "t": 1018.69}
++{"step": 760, "loss": 4.951390266418457, "lr": 0.000591368785648048, "grad_norm": 0.671257734298706, "tokens": 12468224, "tok_per_s": 11910.112552015506, "t": 1047.5}
++{"step": 780, "loss": 5.068446159362793, "lr": 0.0005906064057748841, "grad_norm": 0.5128580331802368, "tokens": 12795904, "tok_per_s": 11887.9204368979, "t": 1077.02}
++{"step": 800, "loss": 5.168520927429199, "lr": 0.0005898123600834975, "grad_norm": 5.2463555335998535, "tokens": 13123584, "tok_per_s": 11875.612775467493, "t": 1105.73}
++{"step": 820, "loss": 5.072286605834961, "lr": 0.00058898674505696, "grad_norm": 0.6444844603538513, "tokens": 13451264, "tok_per_s": 11876.670710730987, "t": 1133.22}
++{"step": 840, "loss": 4.948967456817627, "lr": 0.0005881296610142765, "grad_norm": 0.5277834534645081, "tokens": 13778944, "tok_per_s": 11873.161622668349, "t": 1161.15}
++{"step": 860, "loss": 4.968014717102051, "lr": 0.0005872412120981961, "grad_norm": 0.5329476594924927, "tokens": 14106624, "tok_per_s": 11869.584852756028, "t": 1189.11}
++{"step": 880, "loss": 4.756638526916504, "lr": 0.0005863215062625574, "grad_norm": 0.5423914194107056, "tokens": 14434304, "tok_per_s": 11877.591588374811, "t": 1215.9}
++{"step": 900, "loss": 4.976390838623047, "lr": 0.0005853706552591712, "grad_norm": 0.5418165326118469, "tokens": 14761984, "tok_per_s": 11875.467479769606, "t": 1243.71}
++{"step": 920, "loss": 4.84976863861084, "lr": 0.0005843887746242426, "grad_norm": 0.6431712508201599, "tokens": 15089664, "tok_per_s": 11873.266409096448, "t": 1271.54}
++{"step": 940, "loss": 4.840450286865234, "lr": 0.000583375983664331, "grad_norm": 0.5837947726249695, "tokens": 15417344, "tok_per_s": 11877.077855763022, "t": 1298.72}
++{"step": 960, "loss": 4.9250969886779785, "lr": 0.0005823324054418544, "grad_norm": 0.5661423206329346, "tokens": 15745024, "tok_per_s": 11868.1409528405, "t": 1327.31}
++{"step": 980, "loss": 5.006077289581299, "lr": 0.0005812581667601359, "grad_norm": 1.4778692722320557, "tokens": 16072704, "tok_per_s": 11861.968977733293, "t": 1355.62}
++{"step": 999, "val_loss": 4.868359494209289, "val_ppl": 130.10729994547023, "tokens": 16384000, "t": 1387.8}
++{"step": 999, "checkpoint_sha256": "cda289d1d0f642b19ecbf6e6c12a162167897565760a2f30d1652474fd2adb60", "t": 1388.18}
++{"step": 1000, "loss": 4.746236324310303, "lr": 0.0005801533981479964, "grad_norm": 0.5520177483558655, "tokens": 16400384, "tok_per_s": 11808.062608862347, "t": 1389.56}
++{"step": 1020, "loss": 4.937170028686523, "lr": 0.0005790182338438938, "grad_norm": 0.5781713128089905, "tokens": 16728064, "tok_per_s": 11805.29816888464, "t": 1417.64}
++{"step": 1040, "loss": 4.894624710083008, "lr": 0.0005778528117796123, "grad_norm": 0.5315879583358765, "tokens": 17055744, "tok_per_s": 11803.357443436133, "t": 1445.63}
++{"step": 1060, "loss": 4.880748271942139, "lr": 0.0005766572735635022, "grad_norm": 1.2597907781600952, "tokens": 17383424, "tok_per_s": 11790.355502316252, "t": 1475.02}
++{"step": 1080, "loss": 5.013078689575195, "lr": 0.0005754317644632738, "grad_norm": 0.6085506081581116, "tokens": 17711104, "tok_per_s": 11787.496721408717, "t": 1503.18}
++{"step": 1100, "loss": 4.827304840087891, "lr": 0.000574176433388346, "grad_norm": 0.5576329827308655, "tokens": 18038784, "tok_per_s": 11797.208143102953, "t": 1529.71}
++{"step": 1120, "loss": 4.837977409362793, "lr": 0.0005728914328717517, "grad_norm": 0.5152426958084106, "tokens": 18366464, "tok_per_s": 11795.445317768213, "t": 1557.72}
++{"step": 1140, "loss": 4.640369415283203, "lr": 0.000571576919051606, "grad_norm": 0.5521364212036133, "tokens": 18694144, "tok_per_s": 11795.230672817028, "t": 1585.53}
++{"step": 1160, "loss": 4.757654666900635, "lr": 0.0005702330516521316, "grad_norm": 0.5247706770896912, "tokens": 19021824, "tok_per_s": 11788.59224630508, "t": 1614.24}
++{"step": 1180, "loss": 4.691868782043457, "lr": 0.0005688599939642524, "grad_norm": 0.5326834917068481, "tokens": 19349504, "tok_per_s": 11791.14385723584, "t": 1641.66}
++{"step": 1200, "loss": 4.747924327850342, "lr": 0.000567457912825752, "grad_norm": 0.5180054903030396, "tokens": 19677184, "tok_per_s": 11792.949766833508, "t": 1669.2}
++{"step": 1220, "loss": 4.812901496887207, "lr": 0.000566026978601002, "grad_norm": 0.5599668622016907, "tokens": 20004864, "tok_per_s": 11783.796474617193, "t": 1698.3}
++{"step": 1240, "loss": 4.624703407287598, "lr": 0.00056456736516026, "grad_norm": 0.6033844351768494, "tokens": 20332544, "tok_per_s": 11779.719556755437, "t": 1726.71}
++{"step": 1260, "loss": 4.742227077484131, "lr": 0.000563079249858545, "grad_norm": 0.5360792875289917, "tokens": 20660224, "tok_per_s": 11779.076173728605, "t": 1754.62}
++{"step": 1280, "loss": 4.559293270111084, "lr": 0.000561562813514085, "grad_norm": 0.557126522064209, "tokens": 20987904, "tok_per_s": 11770.653351125746, "t": 1783.71}
++{"step": 1300, "loss": 4.704896926879883, "lr": 0.0005600182403863482, "grad_norm": 0.552845299243927, "tokens": 21315584, "tok_per_s": 11774.175946588024, "t": 1811.01}
++{"step": 1320, "loss": 4.649643898010254, "lr": 0.0005584457181536525, "grad_norm": 0.5440144538879395, "tokens": 21643264, "tok_per_s": 11778.277753235605, "t": 1838.2}
++{"step": 1340, "loss": 4.542902946472168, "lr": 0.0005568454378903616, "grad_norm": 0.5623109340667725, "tokens": 21970944, "tok_per_s": 11782.242522633296, "t": 1865.39}
++{"step": 1360, "loss": 4.589754104614258, "lr": 0.0005552175940436679, "grad_norm": 0.5271830558776855, "tokens": 22298624, "tok_per_s": 11772.282296335274, "t": 1894.81}
++{"step": 1380, "loss": 4.611910820007324, "lr": 0.0005535623844099653, "grad_norm": 0.5605332255363464, "tokens": 22626304, "tok_per_s": 11767.477063289036, "t": 1923.43}
++{"step": 1400, "loss": 4.653350830078125, "lr": 0.0005518800101108161, "grad_norm": 0.5224345326423645, "tokens": 22953984, "tok_per_s": 11763.642777929157, "t": 1951.91}
++{"step": 1420, "loss": 4.504340171813965, "lr": 0.000550170675568512, "grad_norm": 0.5628354549407959, "tokens": 23281664, "tok_per_s": 11759.425495362444, "t": 1980.47}
++{"step": 1440, "loss": 4.538120746612549, "lr": 0.0005484345884812357, "grad_norm": 0.5646089911460876, "tokens": 23609344, "tok_per_s": 11763.342947627734, "t": 2007.67}
++{"step": 1460, "loss": 4.548748970031738, "lr": 0.000546671959797824, "grad_norm": 0.5454878211021423, "tokens": 23937024, "tok_per_s": 11766.325077504327, "t": 2035.01}
++{"step": 1480, "loss": 4.4955878257751465, "lr": 0.0005448830036921353, "grad_norm": 0.5409777760505676, "tokens": 24264704, "tok_per_s": 11760.413489326585, "t": 2063.9}
++{"step": 1499, "val_loss": 4.55281765460968, "val_ppl": 94.8994257612488, "tokens": 24576000, "t": 2098.38}
++{"step": 1500, "loss": 4.449386119842529, "lr": 0.0005430679375370263, "grad_norm": 0.5416464805603027, "tokens": 24592384, "tok_per_s": 11715.626049715813, "t": 2099.75}
++{"step": 1520, "loss": 4.506931304931641, "lr": 0.0005412269818779387, "grad_norm": 0.5411835312843323, "tokens": 24920064, "tok_per_s": 11709.046886550848, "t": 2128.92}
++{"step": 1540, "loss": 4.46286678314209, "lr": 0.0005393603604061018, "grad_norm": 0.5418587327003479, "tokens": 25247744, "tok_per_s": 11712.523059721985, "t": 2156.26}
++{"step": 1560, "loss": 4.550535202026367, "lr": 0.0005374682999313519, "grad_norm": 0.9948550462722778, "tokens": 25575424, "tok_per_s": 11706.191924192692, "t": 2185.42}
++{"step": 1580, "loss": 4.395641326904297, "lr": 0.000535551030354573, "grad_norm": 0.5468602776527405, "tokens": 25903104, "tok_per_s": 11708.36429591705, "t": 2213.0}
++{"step": 1600, "loss": 4.452836036682129, "lr": 0.000533608784639762, "grad_norm": 0.5829842686653137, "tokens": 26230784, "tok_per_s": 11711.802983617088, "t": 2240.33}
++{"step": 1620, "loss": 4.391358375549316, "lr": 0.0005316417987857215, "grad_norm": 0.5233222246170044, "tokens": 26558464, "tok_per_s": 11704.320435493784, "t": 2269.76}
++{"step": 1640, "loss": 4.432625770568848, "lr": 0.0005296503117973845, "grad_norm": 0.5276442766189575, "tokens": 26886144, "tok_per_s": 11713.19016145784, "t": 2296.02}
++{"step": 1660, "loss": 4.441397666931152, "lr": 0.0005276345656567725, "grad_norm": 0.5338090062141418, "tokens": 27213824, "tok_per_s": 11709.577315533745, "t": 2324.71}
++{"step": 1680, "loss": 4.608345985412598, "lr": 0.000525594805293594, "grad_norm": 0.5458671450614929, "tokens": 27541504, "tok_per_s": 11706.816270307252, "t": 2353.25}
++{"step": 1700, "loss": 4.404858589172363, "lr": 0.0005235312785554818, "grad_norm": 0.5006349682807922, "tokens": 27869184, "tok_per_s": 11703.181470092059, "t": 2381.98}
++{"step": 1720, "loss": 4.4875993728637695, "lr": 0.000521444236177879, "grad_norm": 0.5281708240509033, "tokens": 28196864, "tok_per_s": 11706.901990885435, "t": 2409.21}
++{"step": 1740, "loss": 4.452470779418945, "lr": 0.0005193339317535718, "grad_norm": 0.5706871747970581, "tokens": 28524544, "tok_per_s": 11709.705702930381, "t": 2436.62}
++{"step": 1760, "loss": 4.406674385070801, "lr": 0.0005172006217018763, "grad_norm": 0.5338311195373535, "tokens": 28852224, "tok_per_s": 11705.233072294539, "t": 2465.54}
++{"step": 1780, "loss": 4.415951728820801, "lr": 0.0005150445652374809, "grad_norm": 0.54909747838974, "tokens": 29179904, "tok_per_s": 11705.049532957939, "t": 2493.58}
++{"step": 1800, "loss": 4.3329620361328125, "lr": 0.0005128660243389501, "grad_norm": 0.523220419883728, "tokens": 29507584, "tok_per_s": 11699.735345727024, "t": 2522.72}
++{"step": 1820, "loss": 4.487624645233154, "lr": 0.0005106652637168917, "grad_norm": 0.579258143901825, "tokens": 29835264, "tok_per_s": 11702.42285860611, "t": 2550.14}
++{"step": 1840, "loss": 4.522067070007324, "lr": 0.0005084425507817925, "grad_norm": 0.5385577082633972, "tokens": 30162944, "tok_per_s": 11697.895599891974, "t": 2579.14}
++{"step": 1860, "loss": 4.351419448852539, "lr": 0.0005061981556115255, "grad_norm": 0.5320832133293152, "tokens": 30490624, "tok_per_s": 11699.751976479256, "t": 2606.73}
++{"step": 1880, "loss": 4.346895694732666, "lr": 0.0005039323509185333, "grad_norm": 0.5495853424072266, "tokens": 30818304, "tok_per_s": 11702.60960328932, "t": 2634.1}
++{"step": 1900, "loss": 4.382790565490723, "lr": 0.0005016454120166914, "grad_norm": 0.6258963346481323, "tokens": 31145984, "tok_per_s": 11697.128749691117, "t": 2663.35}
++{"step": 1920, "loss": 4.427793502807617, "lr": 0.0004993376167878549, "grad_norm": 0.5516924858093262, "tokens": 31473664, "tok_per_s": 11691.72184032782, "t": 2692.6}
++{"step": 1940, "loss": 4.197183609008789, "lr": 0.0004970092456480945, "grad_norm": 0.588075578212738, "tokens": 31801344, "tok_per_s": 11690.4435519863, "t": 2720.93}
++{"step": 1960, "loss": 4.288908958435059, "lr": 0.0004946605815136225, "grad_norm": 0.5377606749534607, "tokens": 32129024, "tok_per_s": 11687.307367410403, "t": 2749.7}
++{"step": 1980, "loss": 4.317612171173096, "lr": 0.0004922919097664167, "grad_norm": 0.5370826125144958, "tokens": 32456704, "tok_per_s": 11689.41867836808, "t": 2777.23}
++{"step": 1999, "val_loss": 4.377403140068054, "val_ppl": 79.63097418334951, "tokens": 32768000, "t": 2810.66}
++{"step": 1999, "checkpoint_sha256": "0aa6a9f604a8834952c90f7fdbc128abdeda90a2fab99192c8d6e82ae49cb5fd", "t": 2811.09}
++{"step": 2000, "loss": 4.231821537017822, "lr": 0.0004899035182195446, "grad_norm": 0.5784196257591248, "tokens": 32784384, "tok_per_s": 11658.385223092659, "t": 2812.73}
++{"step": 2020, "loss": 4.3884429931640625, "lr": 0.00048749569708219075, "grad_norm": 0.525882363319397, "tokens": 33112064, "tok_per_s": 11656.431473337303, "t": 2841.31}
++{"step": 2040, "loss": 4.128255844116211, "lr": 0.0004850687389243948, "grad_norm": 0.5654372572898865, "tokens": 33439744, "tok_per_s": 11655.083826786376, "t": 2869.75}
++{"step": 2060, "loss": 4.392373561859131, "lr": 0.00048262293864150137, "grad_norm": 0.5647616982460022, "tokens": 33767424, "tok_per_s": 11651.209102049805, "t": 2898.83}
++{"step": 2080, "loss": 4.272017478942871, "lr": 0.0004801585934183281, "grad_norm": 0.5246736407279968, "tokens": 34095104, "tok_per_s": 11654.059240044917, "t": 2926.24}
++{"step": 2100, "loss": 4.351164817810059, "lr": 0.0004776760026930552, "grad_norm": 0.5639843344688416, "tokens": 34422784, "tok_per_s": 11651.157758820587, "t": 2955.09}
++{"step": 2120, "loss": 4.381256103515625, "lr": 0.0004751754681208411, "grad_norm": 0.5990827679634094, "tokens": 34750464, "tok_per_s": 11654.334181549966, "t": 2982.41}
++{"step": 2140, "loss": 4.293796539306641, "lr": 0.0004726572935371689, "grad_norm": 0.5889109373092651, "tokens": 35078144, "tok_per_s": 11657.838494768248, "t": 3009.62}
++{"step": 2160, "loss": 4.318106651306152, "lr": 0.00047012178492092805, "grad_norm": 0.5541678667068481, "tokens": 35405824, "tok_per_s": 11652.700123326464, "t": 3039.06}
++{"step": 2180, "loss": 4.370950698852539, "lr": 0.00046756925035723536, "grad_norm": 0.523361086845398, "tokens": 35733504, "tok_per_s": 11651.574937822988, "t": 3067.48}
++{"step": 2200, "loss": 4.3684163093566895, "lr": 0.00046499999999999997, "grad_norm": 0.5235266089439392, "tokens": 36061184, "tok_per_s": 11650.641940686393, "t": 3095.85}
++{"step": 2220, "loss": 4.416513442993164, "lr": 0.0004624143460342372, "grad_norm": 0.5547477006912231, "tokens": 36388864, "tok_per_s": 11650.49520771614, "t": 3124.02}
++{"step": 2240, "loss": 4.400991439819336, "lr": 0.00045981260263813534, "grad_norm": 0.6226858496665955, "tokens": 36716544, "tok_per_s": 11652.8716753309, "t": 3151.5}
++{"step": 2260, "loss": 4.15883731842041, "lr": 0.0004571950859448811, "grad_norm": 0.5616633296012878, "tokens": 37044224, "tok_per_s": 11655.973818399472, "t": 3178.77}
++{"step": 2280, "loss": 4.316667556762695, "lr": 0.0004545621140042457, "grad_norm": 0.568236231803894, "tokens": 37371904, "tok_per_s": 11653.730269233454, "t": 3207.5}
++{"step": 2300, "loss": 4.269564628601074, "lr": 0.00045191400674394013, "grad_norm": 0.5399973392486572, "tokens": 37699584, "tok_per_s": 11650.245063967917, "t": 3236.59}
++{"step": 2320, "loss": 4.225350379943848, "lr": 0.000449251085930741, "grad_norm": 0.5322178602218628, "tokens": 38027264, "tok_per_s": 11646.621966280994, "t": 3265.73}
++{"step": 2340, "loss": 4.254165172576904, "lr": 0.00044657367513139303, "grad_norm": 0.5262772440910339, "tokens": 38354944, "tok_per_s": 11649.606915520488, "t": 3293.02}
++{"step": 2360, "loss": 4.305544853210449, "lr": 0.00044388209967329364, "grad_norm": 0.5276159644126892, "tokens": 38682624, "tok_per_s": 11646.145968824367, "t": 3322.14}
++{"step": 2380, "loss": 4.318025588989258, "lr": 0.00044117668660496233, "grad_norm": 0.5362470149993896, "tokens": 39010304, "tok_per_s": 11649.147628536659, "t": 3349.41}
++{"step": 2400, "loss": 4.189087390899658, "lr": 0.00043845776465630177, "grad_norm": 0.5912144184112549, "tokens": 39337984, "tok_per_s": 11652.166209961133, "t": 3376.67}
++{"step": 2420, "loss": 4.167083263397217, "lr": 0.00043572566419865466, "grad_norm": 0.5396695733070374, "tokens": 39665664, "tok_per_s": 11653.654786173189, "t": 3404.35}
++{"step": 2440, "loss": 4.282366752624512, "lr": 0.0004329807172046608, "grad_norm": 0.5638473629951477, "tokens": 39993344, "tok_per_s": 11650.457568695849, "t": 3433.41}
++{"step": 2460, "loss": 4.179342269897461, "lr": 0.00043022325720791954, "grad_norm": 0.5641481876373291, "tokens": 40321024, "tok_per_s": 11647.07079450147, "t": 3462.54}
++{"step": 2480, "loss": 4.261195182800293, "lr": 0.000427453619262463, "grad_norm": 0.55649334192276, "tokens": 40648704, "tok_per_s": 11647.617384047178, "t": 3490.52}
++{"step": 2499, "val_loss": 4.253906965255737, "val_ppl": 70.37984751220374, "tokens": 40960000, "t": 3524.46}
++{"step": 2500, "loss": 4.270415782928467, "lr": 0.00042467213990204363, "grad_norm": 0.5214126706123352, "tokens": 40976384, "tok_per_s": 11623.909543881557, "t": 3525.82}
++{"step": 2520, "loss": 4.200478553771973, "lr": 0.0004218791570992436, "grad_norm": 0.5286369919776917, "tokens": 41304064, "tok_per_s": 11627.020290283132, "t": 3553.06}
++{"step": 2540, "loss": 4.271385192871094, "lr": 0.00041907501022440735, "grad_norm": 1.1362669467926025, "tokens": 41631744, "tok_per_s": 11624.355791304675, "t": 3582.07}
++{"step": 2560, "loss": 4.153648376464844, "lr": 0.00041626004000440593, "grad_norm": 0.5819516181945801, "tokens": 41959424, "tok_per_s": 11623.037056115738, "t": 3610.69}
++{"step": 2580, "loss": 4.239153861999512, "lr": 0.00041343458848123576, "grad_norm": 0.5144444108009338, "tokens": 42287104, "tok_per_s": 11619.319612031708, "t": 3640.02}
++{"step": 2600, "loss": 4.200573444366455, "lr": 0.0004105989989704575, "grad_norm": 0.562723696231842, "tokens": 42614784, "tok_per_s": 11622.04027606792, "t": 3667.36}
++{"step": 2620, "loss": 4.275785446166992, "lr": 0.0004077536160194806, "grad_norm": 0.5386180877685547, "tokens": 42942464, "tok_per_s": 11620.13482113355, "t": 3696.16}
++{"step": 2640, "loss": 4.140957355499268, "lr": 0.000404898785365698, "grad_norm": 0.51993727684021, "tokens": 43270144, "tok_per_s": 11622.032355083207, "t": 3723.76}
++{"step": 2660, "loss": 4.120570659637451, "lr": 0.000402034853894476, "grad_norm": 0.5389391183853149, "tokens": 43597824, "tok_per_s": 11624.461428055718, "t": 3751.17}
++{"step": 2680, "loss": 4.026679992675781, "lr": 0.000399162169597005, "grad_norm": 0.542602002620697, "tokens": 43925504, "tok_per_s": 11621.979828131509, "t": 3780.16}
++{"step": 2700, "loss": 4.190106391906738, "lr": 0.00039628108152801574, "grad_norm": 0.53343266248703, "tokens": 44253184, "tok_per_s": 11618.268985670818, "t": 3809.6}
++{"step": 2720, "loss": 4.038628578186035, "lr": 0.000393391939763366, "grad_norm": 0.5840466022491455, "tokens": 44580864, "tok_per_s": 11619.40309369032, "t": 3837.4}
++{"step": 2740, "loss": 4.085831642150879, "lr": 0.00039049509535750383, "grad_norm": 0.5408041477203369, "tokens": 44908544, "tok_per_s": 11618.701568373139, "t": 3865.84}
++{"step": 2760, "loss": 4.053369045257568, "lr": 0.0003875909003008116, "grad_norm": 0.5366416573524475, "tokens": 45236224, "tok_per_s": 11620.9278906342, "t": 3893.29}
++{"step": 2780, "loss": 4.154467582702637, "lr": 0.0003846797074768358, "grad_norm": 0.5264991521835327, "tokens": 45563904, "tok_per_s": 11618.84091486484, "t": 3922.2}
++{"step": 2800, "loss": 4.168743133544922, "lr": 0.00038176187061940935, "grad_norm": 0.536626398563385, "tokens": 45891584, "tok_per_s": 11616.338803380988, "t": 3951.25}
++{"step": 2820, "loss": 4.153240203857422, "lr": 0.0003788377442696697, "grad_norm": 0.533578634262085, "tokens": 46219264, "tok_per_s": 11615.253346568814, "t": 3979.83}
++{"step": 2840, "loss": 4.130046844482422, "lr": 0.00037590768373297907, "grad_norm": 0.5270524024963379, "tokens": 46546944, "tok_per_s": 11618.32541559027, "t": 4006.98}
++{"step": 2860, "loss": 4.124772071838379, "lr": 0.0003729720450357526, "grad_norm": 0.5412513613700867, "tokens": 46874624, "tok_per_s": 11617.065799558944, "t": 4035.62}
++{"step": 2880, "loss": 4.045408248901367, "lr": 0.0003700311848821973, "grad_norm": 0.5562382936477661, "tokens": 47202304, "tok_per_s": 11620.032790486088, "t": 4062.79}
++{"step": 2900, "loss": 4.053845405578613, "lr": 0.00036708546061097017, "grad_norm": 0.5519654154777527, "tokens": 47529984, "tok_per_s": 11622.144412185111, "t": 4090.25}
++{"step": 2920, "loss": 4.138660430908203, "lr": 0.00036413523015175846, "grad_norm": 0.5388080477714539, "tokens": 47857664, "tok_per_s": 11620.688842362795, "t": 4118.96}
++{"step": 2940, "loss": 4.136221885681152, "lr": 0.0003611808519817883, "grad_norm": 0.5571575164794922, "tokens": 48185344, "tok_per_s": 11617.611410784506, "t": 4148.25}
++{"step": 2960, "loss": 4.039947509765625, "lr": 0.0003582226850822664, "grad_norm": 0.535615861415863, "tokens": 48513024, "tok_per_s": 11618.557089478945, "t": 4176.12}
++{"step": 2980, "loss": 4.180006504058838, "lr": 0.0003552610888947618, "grad_norm": 0.5578852891921997, "tokens": 48840704, "tok_per_s": 11620.276163903865, "t": 4203.7}
++{"step": 2999, "val_loss": 4.158720552921295, "val_ppl": 63.989598897214684, "tokens": 49152000, "t": 4237.64}
++{"step": 2999, "checkpoint_sha256": "01088cbe14a4bbf4ef3d3f45ae99d689c24d2752c909afe60703401d5c293072", "t": 4238.01}
++{"step": 3000, "loss": 4.180983543395996, "lr": 0.0003522964232775297, "grad_norm": 0.5751718878746033, "tokens": 49168384, "tok_per_s": 11599.737328140918, "t": 4239.39}
++{"step": 3020, "loss": 4.092770576477051, "lr": 0.0003493290484617865, "grad_norm": 0.529112696647644, "tokens": 49496064, "tok_per_s": 11597.092272818163, "t": 4268.61}
++{"step": 3040, "loss": 4.0629377365112305, "lr": 0.0003463593250079383, "grad_norm": 0.5474892854690552, "tokens": 49823744, "tok_per_s": 11595.560939643, "t": 4297.44}
++{"step": 3060, "loss": 4.105502128601074, "lr": 0.00034338761376177, "grad_norm": 0.8342363834381104, "tokens": 50151424, "tok_per_s": 11597.205212677118, "t": 4325.08}
++{"step": 3080, "loss": 4.145478248596191, "lr": 0.00034041427581059977, "grad_norm": 0.5806731581687927, "tokens": 50479104, "tok_per_s": 11594.81427453882, "t": 4354.24}
++{"step": 3100, "loss": 4.065998554229736, "lr": 0.0003374396724394037, "grad_norm": 0.5266734957695007, "tokens": 50806784, "tok_per_s": 11597.401930260172, "t": 4381.52}
++{"step": 3120, "loss": 4.046976089477539, "lr": 0.0003344641650869168, "grad_norm": 0.5373362302780151, "tokens": 51134464, "tok_per_s": 11599.940600673119, "t": 4408.81}
++{"step": 3140, "loss": 4.0326666831970215, "lr": 0.0003314881153017151, "grad_norm": 0.5418212413787842, "tokens": 51462144, "tok_per_s": 11602.451890150127, "t": 4436.1}
++{"step": 3160, "loss": 4.104043006896973, "lr": 0.00032851188469828487, "grad_norm": 0.5828465223312378, "tokens": 51789824, "tok_per_s": 11599.187619918594, "t": 4465.6}
++{"step": 3180, "loss": 3.948253870010376, "lr": 0.0003255358349130832, "grad_norm": 0.5222381353378296, "tokens": 52117504, "tok_per_s": 11598.717562240729, "t": 4494.03}
++{"step": 3200, "loss": 4.206734657287598, "lr": 0.00032256032756059627, "grad_norm": 0.5611771941184998, "tokens": 52445184, "tok_per_s": 11598.044955429035, "t": 4522.54}
++{"step": 3220, "loss": 4.076472282409668, "lr": 0.00031958572418940007, "grad_norm": 0.5682035088539124, "tokens": 52772864, "tok_per_s": 11597.313869426645, "t": 4551.08}
++{"step": 3240, "loss": 3.881274700164795, "lr": 0.00031661238623822995, "grad_norm": 0.5410988330841064, "tokens": 53100544, "tok_per_s": 11599.778026789865, "t": 4578.36}
++{"step": 3260, "loss": 4.112787246704102, "lr": 0.00031364067499206177, "grad_norm": 0.5453398823738098, "tokens": 53428224, "tok_per_s": 11598.3489765245, "t": 4607.18}
++{"step": 3280, "loss": 4.042259216308594, "lr": 0.00031067095153821354, "grad_norm": 0.5254319906234741, "tokens": 53755904, "tok_per_s": 11596.062737123084, "t": 4636.35}
++{"step": 3300, "loss": 4.027780532836914, "lr": 0.0003077035767224703, "grad_norm": 0.5368713736534119, "tokens": 54083584, "tok_per_s": 11593.837629205618, "t": 4665.5}
++{"step": 3320, "loss": 4.048270225524902, "lr": 0.0003047389111052382, "grad_norm": 0.5798897743225098, "tokens": 54411264, "tok_per_s": 11594.94706767128, "t": 4693.31}
++{"step": 3340, "loss": 3.9507741928100586, "lr": 0.0003017773149177336, "grad_norm": 0.5504516363143921, "tokens": 54738944, "tok_per_s": 11594.143632157271, "t": 4721.9}
++{"step": 3360, "loss": 3.9647669792175293, "lr": 0.0002988191480182117, "grad_norm": 0.5416991114616394, "tokens": 55066624, "tok_per_s": 11596.184635204852, "t": 4749.33}
++{"step": 3380, "loss": 4.000699996948242, "lr": 0.0002958647698482415, "grad_norm": 0.5480679273605347, "tokens": 55394304, "tok_per_s": 11595.102761281029, "t": 4778.03}
++{"step": 3400, "loss": 3.971863031387329, "lr": 0.0002929145393890298, "grad_norm": 0.5375354290008545, "tokens": 55721984, "tok_per_s": 11594.402055720479, "t": 4806.58}
++{"step": 3420, "loss": 4.104773998260498, "lr": 0.0002899688151178026, "grad_norm": 0.5447196364402771, "tokens": 56049664, "tok_per_s": 11590.86988710522, "t": 4836.32}
++{"step": 3440, "loss": 4.046260833740234, "lr": 0.00028702795496424736, "grad_norm": 0.5488489866256714, "tokens": 56377344, "tok_per_s": 11592.886162092072, "t": 4863.74}
++{"step": 3460, "loss": 3.960139274597168, "lr": 0.00028409231626702093, "grad_norm": 0.5675356984138489, "tokens": 56705024, "tok_per_s": 11591.251066099452, "t": 4892.7}
++{"step": 3480, "loss": 3.994968891143799, "lr": 0.00028116225573033036, "grad_norm": 0.5636969208717346, "tokens": 57032704, "tok_per_s": 11592.733713437376, "t": 4920.34}
++{"step": 3499, "val_loss": 4.076284086704254, "val_ppl": 58.926098247434794, "tokens": 57344000, "t": 4953.42}
++{"step": 3500, "loss": 4.110962390899658, "lr": 0.00027823812938059065, "grad_norm": 0.5642924308776855, "tokens": 57360384, "tok_per_s": 11577.222286224182, "t": 4955.23}
++{"step": 3520, "loss": 3.9599759578704834, "lr": 0.0002753202925231642, "grad_norm": 0.5764612555503845, "tokens": 57688064, "tok_per_s": 11575.88157647093, "t": 4984.14}
++{"step": 3540, "loss": 3.9944944381713867, "lr": 0.0002724090996991884, "grad_norm": 0.5443633198738098, "tokens": 58015744, "tok_per_s": 11574.98417124486, "t": 5012.81}
++{"step": 3560, "loss": 4.018226623535156, "lr": 0.00026950490464249606, "grad_norm": 0.5742700099945068, "tokens": 58343424, "tok_per_s": 11573.307568254504, "t": 5041.85}
++{"step": 3580, "loss": 4.130144119262695, "lr": 0.0002666080602366339, "grad_norm": 0.5585811138153076, "tokens": 58671104, "tok_per_s": 11574.399101452585, "t": 5069.68}
++{"step": 3600, "loss": 4.045418739318848, "lr": 0.00026371891847198426, "grad_norm": 0.5662236213684082, "tokens": 58998784, "tok_per_s": 11573.813791413262, "t": 5098.25}
++{"step": 3620, "loss": 4.177013397216797, "lr": 0.00026083783040299483, "grad_norm": 0.5743661522865295, "tokens": 59326464, "tok_per_s": 11576.377095529744, "t": 5125.43}
++{"step": 3640, "loss": 3.959064483642578, "lr": 0.0002579651461055239, "grad_norm": 0.5499097108840942, "tokens": 59654144, "tok_per_s": 11578.98264556307, "t": 5152.58}
++{"step": 3660, "loss": 3.7604146003723145, "lr": 0.000255101214634302, "grad_norm": 0.5641483664512634, "tokens": 59981824, "tok_per_s": 11576.51833666578, "t": 5181.98}
++{"step": 3680, "loss": 4.021690845489502, "lr": 0.0002522463839805194, "grad_norm": 0.5463368892669678, "tokens": 60309504, "tok_per_s": 11575.705473660024, "t": 5210.65}
++{"step": 3700, "loss": 3.943697452545166, "lr": 0.00024940100102954256, "grad_norm": 0.5520758628845215, "tokens": 60637184, "tok_per_s": 11575.606176705642, "t": 5239.02}
++{"step": 3720, "loss": 3.904292583465576, "lr": 0.00024656541151876424, "grad_norm": 0.5493237972259521, "tokens": 60964864, "tok_per_s": 11577.859974252468, "t": 5266.28}
++{"step": 3740, "loss": 4.000242233276367, "lr": 0.00024373995999559398, "grad_norm": 0.5580956339836121, "tokens": 61292544, "tok_per_s": 11577.373754446042, "t": 5294.81}
++{"step": 3760, "loss": 3.917473554611206, "lr": 0.0002409249897755926, "grad_norm": 0.5592621564865112, "tokens": 61620224, "tok_per_s": 11575.636270353167, "t": 5323.91}
++{"step": 3780, "loss": 3.902782917022705, "lr": 0.00023812084290075633, "grad_norm": 0.5623103380203247, "tokens": 61947904, "tok_per_s": 11573.932404902851, "t": 5353.01}
++{"step": 3800, "loss": 4.045053482055664, "lr": 0.00023532786009795634, "grad_norm": 0.5493343472480774, "tokens": 62275584, "tok_per_s": 11576.10886221121, "t": 5380.31}
++{"step": 3820, "loss": 4.035120010375977, "lr": 0.0002325463807375371, "grad_norm": 0.5667306184768677, "tokens": 62603264, "tok_per_s": 11574.20379848079, "t": 5409.5}
++{"step": 3840, "loss": 4.002408027648926, "lr": 0.00022977674279208034, "grad_norm": 0.5492165088653564, "tokens": 62930944, "tok_per_s": 11576.377089696263, "t": 5436.79}
++{"step": 3860, "loss": 4.010754108428955, "lr": 0.00022701928279533916, "grad_norm": 0.5650079846382141, "tokens": 63258624, "tok_per_s": 11578.523310325681, "t": 5464.09}
++{"step": 3880, "loss": 4.072173118591309, "lr": 0.00022427433580134536, "grad_norm": 0.598127543926239, "tokens": 63586304, "tok_per_s": 11580.383044802815, "t": 5491.51}
++{"step": 3900, "loss": 3.9902126789093018, "lr": 0.0002215422353436983, "grad_norm": 0.5532431602478027, "tokens": 63913984, "tok_per_s": 11577.041542373525, "t": 5521.4}
++{"step": 3920, "loss": 3.8755486011505127, "lr": 0.00021882331339503767, "grad_norm": 0.5454591512680054, "tokens": 64241664, "tok_per_s": 11577.402404052884, "t": 5549.53}
++{"step": 3940, "loss": 3.9723997116088867, "lr": 0.00021611790032670633, "grad_norm": 0.5648732781410217, "tokens": 64569344, "tok_per_s": 11577.031145315133, "t": 5578.01}
++{"step": 3960, "loss": 4.01093053817749, "lr": 0.0002134263248686069, "grad_norm": 0.5775900483131409, "tokens": 64897024, "tok_per_s": 11578.616200441384, "t": 5605.55}
++{"step": 3980, "loss": 3.90216064453125, "lr": 0.00021074891406925903, "grad_norm": 0.5466880202293396, "tokens": 65224704, "tok_per_s": 11577.200550459853, "t": 5634.54}
++{"step": 3999, "val_loss": 4.0112474918365475, "val_ppl": 55.21570876292144, "tokens": 65536000, "t": 5668.99}
++{"step": 3999, "checkpoint_sha256": "1defac6e92e64f97536e0c6d03a5ad0f431d8bc6cbc78392fcd088a15efaf557", "t": 5669.37}
++{"step": 4000, "loss": 3.962883234024048, "lr": 0.00020808599325605992, "grad_norm": 0.5598580241203308, "tokens": 65552384, "tok_per_s": 11561.020262288332, "t": 5670.76}
++{"step": 4020, "loss": 3.9114115238189697, "lr": 0.0002054378859957543, "grad_norm": 0.5523264408111572, "tokens": 65880064, "tok_per_s": 11559.453836876024, "t": 5699.88}
++{"step": 4040, "loss": 3.91835355758667, "lr": 0.00020280491405511884, "grad_norm": 0.5684453845024109, "tokens": 66207744, "tok_per_s": 11561.263558356273, "t": 5727.33}
++{"step": 4060, "loss": 4.060446739196777, "lr": 0.00020018739736186444, "grad_norm": 0.5896223783493042, "tokens": 66535424, "tok_per_s": 11560.698828062677, "t": 5755.95}
++{"step": 4080, "loss": 4.083343029022217, "lr": 0.00019758565396576288, "grad_norm": 0.568533182144165, "tokens": 66863104, "tok_per_s": 11562.029005289642, "t": 5783.63}
++{"step": 4100, "loss": 3.9439985752105713, "lr": 0.00019500000000000005, "grad_norm": 0.5773836374282837, "tokens": 67190784, "tok_per_s": 11563.784358023158, "t": 5811.09}
++{"step": 4120, "loss": 3.813251256942749, "lr": 0.0001924307496427646, "grad_norm": 0.5718711614608765, "tokens": 67518464, "tok_per_s": 11560.73883504852, "t": 5840.97}
++{"step": 4140, "loss": 3.793492317199707, "lr": 0.000189878215079072, "grad_norm": 0.5597081184387207, "tokens": 67846144, "tok_per_s": 11560.29938081825, "t": 5869.53}
++{"step": 4160, "loss": 3.9493250846862793, "lr": 0.00018734270646283112, "grad_norm": 0.5616011023521423, "tokens": 68173824, "tok_per_s": 11560.322394282883, "t": 5897.87}
++{"step": 4180, "loss": 3.8800625801086426, "lr": 0.0001848245318791589, "grad_norm": 0.550301194190979, "tokens": 68501504, "tok_per_s": 11561.836976458999, "t": 5925.44}
++{"step": 4200, "loss": 3.836061954498291, "lr": 0.00018232399730694481, "grad_norm": 0.5758439898490906, "tokens": 68829184, "tok_per_s": 11561.382130904814, "t": 5954.01}
++{"step": 4220, "loss": 3.8811402320861816, "lr": 0.00017984140658167184, "grad_norm": 0.5658969879150391, "tokens": 69156864, "tok_per_s": 11560.042820097335, "t": 5983.05}
++{"step": 4240, "loss": 3.8178927898406982, "lr": 0.00017737706135849857, "grad_norm": 0.5584120750427246, "tokens": 69484544, "tok_per_s": 11558.705750259509, "t": 6012.09}
++{"step": 4260, "loss": 3.942742347717285, "lr": 0.00017493126107560525, "grad_norm": 0.5712693929672241, "tokens": 69812224, "tok_per_s": 11561.05464668217, "t": 6039.21}
++{"step": 4280, "loss": 3.9611594676971436, "lr": 0.00017250430291780933, "grad_norm": 0.5624386668205261, "tokens": 70139904, "tok_per_s": 11559.63639961476, "t": 6068.3}
++{"step": 4300, "loss": 4.0740766525268555, "lr": 0.0001700964817804554, "grad_norm": 0.5950373411178589, "tokens": 70467584, "tok_per_s": 11561.908158682603, "t": 6095.45}
++{"step": 4320, "loss": 3.9431636333465576, "lr": 0.00016770809023358322, "grad_norm": 0.5978004932403564, "tokens": 70795264, "tok_per_s": 11564.137920502184, "t": 6122.61}
++{"step": 4340, "loss": 3.7616939544677734, "lr": 0.00016533941848637757, "grad_norm": 0.5960965752601624, "tokens": 71122944, "tok_per_s": 11564.849615749808, "t": 6150.57}
++{"step": 4360, "loss": 3.8475029468536377, "lr": 0.00016299075435190547, "grad_norm": 0.5656423568725586, "tokens": 71450624, "tok_per_s": 11564.091210004042, "t": 6179.31}
++{"step": 4380, "loss": 3.911325454711914, "lr": 0.000160662383212145, "grad_norm": 0.5660343170166016, "tokens": 71778304, "tok_per_s": 11561.790236658328, "t": 6208.88}
++{"step": 4400, "loss": 3.8917365074157715, "lr": 0.00015835458798330867, "grad_norm": 0.5705775022506714, "tokens": 72105984, "tok_per_s": 11563.96163458225, "t": 6236.05}
++{"step": 4420, "loss": 3.8357012271881104, "lr": 0.00015606764908146668, "grad_norm": 0.6175817847251892, "tokens": 72433664, "tok_per_s": 11563.629199508314, "t": 6264.56}
++{"step": 4440, "loss": 3.9402458667755127, "lr": 0.00015380184438847444, "grad_norm": 0.5682572722434998, "tokens": 72761344, "tok_per_s": 11565.777961491118, "t": 6291.73}
++{"step": 4460, "loss": 3.9123647212982178, "lr": 0.0001515574492182075, "grad_norm": 0.5742660760879517, "tokens": 73089024, "tok_per_s": 11564.787390080484, "t": 6320.61}
++{"step": 4480, "loss": 3.8569695949554443, "lr": 0.00014933473628310837, "grad_norm": 0.5776249766349792, "tokens": 73416704, "tok_per_s": 11563.360530311495, "t": 6349.72}
++{"step": 4499, "val_loss": 3.957085204124451, "val_ppl": 52.304646175065805, "tokens": 73728000, "t": 6384.15}
++{"step": 4500, "loss": 3.900219440460205, "lr": 0.00014713397566104992, "grad_norm": 0.5595879554748535, "tokens": 73744384, "tok_per_s": 11549.85744394998, "t": 6385.52}
++{"step": 4520, "loss": 3.9554238319396973, "lr": 0.0001449554347625191, "grad_norm": 0.5791251063346863, "tokens": 74072064, "tok_per_s": 11551.858150088292, "t": 6412.78}
++{"step": 4540, "loss": 3.8150556087493896, "lr": 0.00014279937829812366, "grad_norm": 0.5590927600860596, "tokens": 74399744, "tok_per_s": 11550.828260871835, "t": 6441.72}
++{"step": 4560, "loss": 3.792018413543701, "lr": 0.00014066606824642808, "grad_norm": 0.5811460018157959, "tokens": 74727424, "tok_per_s": 11552.671204872138, "t": 6469.05}
++{"step": 4580, "loss": 3.759868621826172, "lr": 0.0001385557638221209, "grad_norm": 0.572047770023346, "tokens": 75055104, "tok_per_s": 11554.365421175793, "t": 6496.46}
++{"step": 4600, "loss": 3.800534248352051, "lr": 0.0001364687214445182, "grad_norm": 0.5741000175476074, "tokens": 75382784, "tok_per_s": 11554.324567762942, "t": 6524.85}
++{"step": 4620, "loss": 3.8696978092193604, "lr": 0.00013440519470640598, "grad_norm": 0.6237900853157043, "tokens": 75710464, "tok_per_s": 11554.072040147503, "t": 6553.35}
++{"step": 4640, "loss": 3.943920612335205, "lr": 0.00013236543434322733, "grad_norm": 0.5705210566520691, "tokens": 76038144, "tok_per_s": 11552.62606413038, "t": 6582.54}
++{"step": 4660, "loss": 3.7643725872039795, "lr": 0.00013034968820261554, "grad_norm": 0.5851072669029236, "tokens": 76365824, "tok_per_s": 11554.693671882093, "t": 6609.72}
++{"step": 4680, "loss": 3.8417506217956543, "lr": 0.00012835820121427852, "grad_norm": 0.5822275876998901, "tokens": 76693504, "tok_per_s": 11554.239013189506, "t": 6638.34}
++{"step": 4700, "loss": 3.8552558422088623, "lr": 0.00012639121536023803, "grad_norm": 0.5749747157096863, "tokens": 77021184, "tok_per_s": 11555.92077454996, "t": 6665.73}
++{"step": 4720, "loss": 3.877544403076172, "lr": 0.00012444896964542695, "grad_norm": 0.5795513987541199, "tokens": 77348864, "tok_per_s": 11555.125192265208, "t": 6694.54}
++{"step": 4740, "loss": 3.855560302734375, "lr": 0.00012253170006864812, "grad_norm": 0.5813860297203064, "tokens": 77676544, "tok_per_s": 11553.848127747267, "t": 6723.64}
++{"step": 4760, "loss": 3.9683942794799805, "lr": 0.00012063963959389822, "grad_norm": 0.5823618173599243, "tokens": 78004224, "tok_per_s": 11552.457505885966, "t": 6752.82}
++{"step": 4780, "loss": 3.827885150909424, "lr": 0.00011877301812206132, "grad_norm": 0.5629019141197205, "tokens": 78331904, "tok_per_s": 11554.108498220257, "t": 6780.21}
++{"step": 4800, "loss": 3.8732705116271973, "lr": 0.00011693206246297376, "grad_norm": 0.6126952171325684, "tokens": 78659584, "tok_per_s": 11553.165489637673, "t": 6809.13}
++{"step": 4820, "loss": 3.753878355026245, "lr": 0.00011511699630786462, "grad_norm": 0.6059414744377136, "tokens": 78987264, "tok_per_s": 11554.472144458165, "t": 6836.72}
++{"step": 4840, "loss": 4.004238605499268, "lr": 0.00011332804020217591, "grad_norm": 0.5791417360305786, "tokens": 79314944, "tok_per_s": 11555.287329584366, "t": 6864.6}
++{"step": 4860, "loss": 3.7956504821777344, "lr": 0.00011156541151876422, "grad_norm": 0.594039261341095, "tokens": 79642624, "tok_per_s": 11554.656880093267, "t": 6893.33}
++{"step": 4880, "loss": 3.903764486312866, "lr": 0.00010982932443148792, "grad_norm": 0.5707221031188965, "tokens": 79970304, "tok_per_s": 11551.781497839473, "t": 6923.42}
++{"step": 4900, "loss": 3.8113834857940674, "lr": 0.00010811998988918382, "grad_norm": 0.6023657917976379, "tokens": 80297984, "tok_per_s": 11553.437096988495, "t": 6950.78}
++{"step": 4920, "loss": 3.8437676429748535, "lr": 0.00010643761559003459, "grad_norm": 0.5618969798088074, "tokens": 80625664, "tok_per_s": 11552.96747676362, "t": 6979.43}
++{"step": 4940, "loss": 3.911839008331299, "lr": 0.00010478240595633218, "grad_norm": 0.5965155959129333, "tokens": 80953344, "tok_per_s": 11554.25529240919, "t": 7007.01}
++{"step": 4960, "loss": 3.9762771129608154, "lr": 0.00010315456210963839, "grad_norm": 0.5922918915748596, "tokens": 81281024, "tok_per_s": 11554.737882554518, "t": 7035.08}
++{"step": 4980, "loss": 3.800461530685425, "lr": 0.0001015542818463474, "grad_norm": 0.5719321966171265, "tokens": 81608704, "tok_per_s": 11554.03089926349, "t": 7063.87}
++{"step": 4999, "val_loss": 3.9170217752456664, "val_ppl": 50.250564225568326, "tokens": 81920000, "t": 7098.86}
++{"step": 4999, "checkpoint_sha256": "f09f039d5f72ac5fb5bdab0899b9f829650b3c0101cf509c62d1f4aaf0e6bb5f", "t": 7099.15}
++{"step": 5000, "loss": 3.9287047386169434, "lr": 9.998175961365173e-05, "grad_norm": 0.5943207144737244, "tokens": 81936384, "tok_per_s": 11540.023799157347, "t": 7100.83}
++{"step": 5020, "loss": 3.853404998779297, "lr": 9.843718648591489e-05, "grad_norm": 0.5844020247459412, "tokens": 82264064, "tok_per_s": 11541.101026861128, "t": 7128.56}
++{"step": 5040, "loss": 3.861715793609619, "lr": 9.692075014145502e-05, "grad_norm": 0.586184024810791, "tokens": 82591744, "tok_per_s": 11539.954421434737, "t": 7157.67}
++{"step": 5060, "loss": 3.7985501289367676, "lr": 9.543263483973994e-05, "grad_norm": 0.5837913155555725, "tokens": 82919424, "tok_per_s": 11541.87261737775, "t": 7184.87}
++{"step": 5080, "loss": 3.9008865356445312, "lr": 9.39730213989981e-05, "grad_norm": 0.5807478427886963, "tokens": 83247104, "tok_per_s": 11543.789493249493, "t": 7212.06}
++{"step": 5100, "loss": 3.8417551517486572, "lr": 9.254208717424797e-05, "grad_norm": 0.5791102647781372, "tokens": 83574784, "tok_per_s": 11545.686514623943, "t": 7239.26}
++{"step": 5120, "loss": 3.8045313358306885, "lr": 9.114000603574761e-05, "grad_norm": 0.5828171372413635, "tokens": 83902464, "tok_per_s": 11543.956937651712, "t": 7268.73}
++{"step": 5140, "loss": 3.9080166816711426, "lr": 8.976694834786837e-05, "grad_norm": 0.5843819379806519, "tokens": 84230144, "tok_per_s": 11543.657775211985, "t": 7297.3}
++{"step": 5160, "loss": 3.749263286590576, "lr": 8.842308094839385e-05, "grad_norm": 0.5649449229240417, "tokens": 84557824, "tok_per_s": 11544.058148752227, "t": 7325.43}
++{"step": 5180, "loss": 3.858259677886963, "lr": 8.710856712824807e-05, "grad_norm": 0.5771579742431641, "tokens": 84885504, "tok_per_s": 11544.122094995391, "t": 7353.78}
++{"step": 5200, "loss": 3.790924549102783, "lr": 8.582356661165409e-05, "grad_norm": 0.5768647789955139, "tokens": 85213184, "tok_per_s": 11545.853513065293, "t": 7381.06}
++{"step": 5220, "loss": 3.7866573333740234, "lr": 8.456823553672612e-05, "grad_norm": 0.56633460521698, "tokens": 85540864, "tok_per_s": 11545.18819581383, "t": 7409.86}
++{"step": 5240, "loss": 3.8559517860412598, "lr": 8.334272643649774e-05, "grad_norm": 0.6036170721054077, "tokens": 85868544, "tok_per_s": 11543.96703405992, "t": 7439.03}
++{"step": 5260, "loss": 3.857818603515625, "lr": 8.214718822038769e-05, "grad_norm": 0.590832531452179, "tokens": 86196224, "tok_per_s": 11543.041741204966, "t": 7468.02}
++{"step": 5280, "loss": 3.8089985847473145, "lr": 8.098176615610612e-05, "grad_norm": 0.5837736129760742, "tokens": 86523904, "tok_per_s": 11544.735299253369, "t": 7495.31}
++{"step": 5300, "loss": 3.729093313217163, "lr": 7.984660185200352e-05, "grad_norm": 0.5829435586929321, "tokens": 86851584, "tok_per_s": 11543.877992195296, "t": 7524.25}
++{"step": 5320, "loss": 3.79866886138916, "lr": 7.874183323986404e-05, "grad_norm": 0.5758870244026184, "tokens": 87179264, "tok_per_s": 11545.581373394647, "t": 7551.52}
++{"step": 5340, "loss": 3.85446834564209, "lr": 7.76675945581456e-05, "grad_norm": 0.5752629637718201, "tokens": 87506944, "tok_per_s": 11547.36954656337, "t": 7578.73}
++{"step": 5360, "loss": 3.789114475250244, "lr": 7.662401633566895e-05, "grad_norm": 0.5718693733215332, "tokens": 87834624, "tok_per_s": 11546.979012343047, "t": 7607.38}
++{"step": 5380, "loss": 3.6952404975891113, "lr": 7.561122537575734e-05, "grad_norm": 0.5910956859588623, "tokens": 88162304, "tok_per_s": 11546.87632803104, "t": 7635.81}
++{"step": 5400, "loss": 3.8815152645111084, "lr": 7.462934474082865e-05, "grad_norm": 0.5759338140487671, "tokens": 88489984, "tok_per_s": 11545.790709781168, "t": 7664.91}
++{"step": 5420, "loss": 4.046452045440674, "lr": 7.36784937374426e-05, "grad_norm": 0.6042420864105225, "tokens": 88817664, "tok_per_s": 11545.50851820901, "t": 7693.48}
++{"step": 5440, "loss": 3.7572991847991943, "lr": 7.27587879018039e-05, "grad_norm": 0.5804364681243896, "tokens": 89145344, "tok_per_s": 11546.780616024904, "t": 7721.01}
++{"step": 5460, "loss": 3.8507866859436035, "lr": 7.187033898572343e-05, "grad_norm": 0.5794432759284973, "tokens": 89473024, "tok_per_s": 11547.378430238676, "t": 7748.98}
++{"step": 5480, "loss": 3.731966733932495, "lr": 7.101325494303994e-05, "grad_norm": 0.5920389890670776, "tokens": 89800704, "tok_per_s": 11546.92697569402, "t": 7777.66}
++{"step": 5499, "val_loss": 3.887117636203766, "val_ppl": 48.770110536442886, "tokens": 90112000, "t": 7813.41}
++{"step": 5500, "loss": 3.9196672439575195, "lr": 7.018763991650238e-05, "grad_norm": 0.6001965403556824, "tokens": 90128384, "tok_per_s": 11534.021598855534, "t": 7814.78}
++{"step": 5520, "loss": 3.8390917778015137, "lr": 6.939359422511588e-05, "grad_norm": 0.5888873934745789, "tokens": 90456064, "tok_per_s": 11535.491951955515, "t": 7842.19}
++{"step": 5540, "loss": 3.8811357021331787, "lr": 6.863121435195203e-05, "grad_norm": 0.5845345258712769, "tokens": 90783744, "tok_per_s": 11534.633071129776, "t": 7871.18}
++{"step": 5560, "loss": 3.7738800048828125, "lr": 6.790059293242561e-05, "grad_norm": 0.5928100943565369, "tokens": 91111424, "tok_per_s": 11535.80731025614, "t": 7898.78}
++{"step": 5580, "loss": 3.835395574569702, "lr": 6.720181874303845e-05, "grad_norm": 0.5779862403869629, "tokens": 91439104, "tok_per_s": 11537.383975466677, "t": 7926.1}
++{"step": 5600, "loss": 3.815725803375244, "lr": 6.653497669059239e-05, "grad_norm": 0.579407811164856, "tokens": 91766784, "tok_per_s": 11535.913175032281, "t": 7955.52}
++{"step": 5620, "loss": 3.761955499649048, "lr": 6.590014780187247e-05, "grad_norm": 0.5867651700973511, "tokens": 92094464, "tok_per_s": 11535.090516233922, "t": 7984.5}
++{"step": 5640, "loss": 3.8077659606933594, "lr": 6.529740921380147e-05, "grad_norm": 0.5797607898712158, "tokens": 92422144, "tok_per_s": 11535.485589128413, "t": 8012.63}
++{"step": 5660, "loss": 3.789388656616211, "lr": 6.47268341640671e-05, "grad_norm": 0.5859403610229492, "tokens": 92749824, "tok_per_s": 11535.285421023631, "t": 8041.17}
++{"step": 5680, "loss": 3.8692054748535156, "lr": 6.418849198222306e-05, "grad_norm": 0.5814026594161987, "tokens": 93077504, "tok_per_s": 11536.586077977248, "t": 8068.67}
++{"step": 5700, "loss": 3.8076140880584717, "lr": 6.368244808126497e-05, "grad_norm": 0.5866917371749878, "tokens": 93405184, "tok_per_s": 11536.014408981831, "t": 8097.48}
++{"step": 5720, "loss": 3.7942938804626465, "lr": 6.320876394968206e-05, "grad_norm": 0.572772741317749, "tokens": 93732864, "tok_per_s": 11535.139908600137, "t": 8126.5}
++{"step": 5740, "loss": 3.8309621810913086, "lr": 6.276749714398588e-05, "grad_norm": 0.5950015783309937, "tokens": 94060544, "tok_per_s": 11534.120242321984, "t": 8155.62}
++{"step": 5760, "loss": 3.825157642364502, "lr": 6.23587012817167e-05, "grad_norm": 0.609603762626648, "tokens": 94388224, "tok_per_s": 11535.867126376139, "t": 8182.79}
++{"step": 5780, "loss": 3.729236125946045, "lr": 6.198242603492857e-05, "grad_norm": 0.5921071171760559, "tokens": 94715904, "tok_per_s": 11534.939763906073, "t": 8211.86}
++{"step": 5800, "loss": 3.8878514766693115, "lr": 6.163871712415363e-05, "grad_norm": 0.6161739826202393, "tokens": 95043584, "tok_per_s": 11536.696837804526, "t": 8239.01}
++{"step": 5820, "loss": 3.8525753021240234, "lr": 6.132761631284691e-05, "grad_norm": 0.5878046751022339, "tokens": 95371264, "tok_per_s": 11538.254949744933, "t": 8266.3}
++{"step": 5840, "loss": 3.8129560947418213, "lr": 6.10491614023115e-05, "grad_norm": 0.5896515846252441, "tokens": 95698944, "tok_per_s": 11538.439787215593, "t": 8294.57}
++{"step": 5860, "loss": 3.7627148628234863, "lr": 6.08033862271055e-05, "grad_norm": 0.6002752780914307, "tokens": 96026624, "tok_per_s": 11537.950833802404, "t": 8323.32}
++{"step": 5880, "loss": 3.917717933654785, "lr": 6.059032065093087e-05, "grad_norm": 0.5864716172218323, "tokens": 96354304, "tok_per_s": 11536.854347275696, "t": 8352.51}
++{"step": 5900, "loss": 3.671173572540283, "lr": 6.040999056300463e-05, "grad_norm": 0.596362292766571, "tokens": 96681984, "tok_per_s": 11537.971918323201, "t": 8380.1}
++{"step": 5920, "loss": 3.816864013671875, "lr": 6.026241787491321e-05, "grad_norm": 0.5997350811958313, "tokens": 97009664, "tok_per_s": 11537.623532511629, "t": 8408.76}
++{"step": 5940, "loss": 3.8800463676452637, "lr": 6.014762051794997e-05, "grad_norm": 0.5878621339797974, "tokens": 97337344, "tok_per_s": 11539.057999128861, "t": 8436.11}
++{"step": 5960, "loss": 3.8143668174743652, "lr": 6.0065612440936474e-05, "grad_norm": 0.599541425704956, "tokens": 97665024, "tok_per_s": 11540.68232762661, "t": 8463.32}
++{"step": 5980, "loss": 3.7237539291381836, "lr": 6.001640360852739e-05, "grad_norm": 0.5826242566108704, "tokens": 97992704, "tok_per_s": 11539.89272352153, "t": 8492.29}
++{"step": 5999, "loss": 3.8935680389404297, "lr": 6.000004100943553e-05, "grad_norm": 0.5838854908943176, "tokens": 98304000, "tok_per_s": 11538.776028145761, "t": 8520.09}
++{"step": 5999, "val_loss": 3.8689021825790406, "val_ppl": 47.88978297376415, "tokens": 98304000, "t": 8526.68}
++{"step": 5999, "checkpoint_sha256": "1552f1ec984ec895a40814cd466bde59abaf35e4bfe1ea35b9d43ff4cfe9b9d0", "t": 8527.03}

modified paper/appendix/prereg.tex +16 −2

++\subsection*{\texorpdfstring{Pre-registration --- WP31 PRISME pilot (v1.0, FROZEN 2026-10-30T04:04-0400 before any \texttt{test\_*} split was read)}{Pre-registration --- WP31 PRISME pilot (v1.0, FROZEN 2026-10-30T04:04-0400 before any test\_* split was read)}}\label{pre-registration-wp31-prisme-pilot-v1.0-frozen-2026-10-30t0404-0400-before-any-test_-split-was-read}
++Status: FROZEN. No file named \texttt{test*.jsonl} / \texttt{test.bin} had been opened by any evaluation script at freeze time (\texttt{protocol/locked\_test\_manifest.json} records the split hashes and the unsealing timestamp).
++
++\subsubsection*{0. Primary comparisons as executed (fixed at freeze, before unsealing)}\label{primary-comparisons-as-executed-fixed-at-freeze-before-unsealing}
++
++\begin{itemize}
++\tightlist
++\item
++  P1 (H1/H2, synthetic): exact match on the pooled test regimes at forced budget 2 for the sparse editor vs the frozen backbone, vs the global-recurrence adapter and vs the dense editor, 5 adapter seeds each (v2 backbone), paired by item.
++\item
++  P2 (H3, synthetic): dev-iid loss of the locked controller vs the best fixed schedule and vs budget 0 (already reported on dev; the test regimes are evaluated at budget 0 and 2 with forced random actions only, since the controller was worse than budget 0 on dev for every seed --- the controller is therefore evaluated on test for completeness, not as the primary system).
++\item
++  P3 (H1/H4, language): FineWeb-Edu locked test cross-entropy at budgets 0/1/2/4/8 for the 103M backbone, the sparse editor, the global adapter and the parameter-matched plain decoder.
++\item
++  Decision thresholds as in Section 4 below; no other analysis is pre-registered.
++\end{itemize}
+ \subsubsection*{1. Primary benchmark suite}\label{primary-benchmark-suite}

modified paper/main.log +66 −65

				`@@ -1,4 +1,4 @@`
1	−This is pdfTeX, Version 3.141592653-2.6-1.40.29 (TeX Live 2026) (preloaded format=pdflatex 2026.3.22) 30 OCT 2026 04:04	1	+This is pdfTeX, Version 3.141592653-2.6-1.40.29 (TeX Live 2026) (preloaded format=pdflatex 2026.3.22) 30 OCT 2026 04:15
2	entering extended mode	2	entering extended mode
3	restricted \write18 enabled.	3	restricted \write18 enabled.
4	%&-line parsing enabled.	4	%&-line parsing enabled.
				`@@ -1159,7 +1159,7 @@ Package hyperref Info: Link coloring ON on input line 85.`
1159		1159
1160	\c@lstlisting=\count403	1160	\c@lstlisting=\count403
1161	(./sections/titlepage.tex	1161	(./sections/titlepage.tex
1162	−<uq_logo.jpg, id=199, 722.7pt x 361.35pt>	1162	+<uq_logo.jpg, id=203, 722.7pt x 361.35pt>
1163	File: uq_logo.jpg Graphic file (type jpg)	1163	File: uq_logo.jpg Graphic file (type jpg)
1164	<use uq_logo.jpg>	1164	<use uq_logo.jpg>
1165	Package pdftex.def Info: uq_logo.jpg used on input line 9.	1165	Package pdftex.def Info: uq_logo.jpg used on input line 9.
				`@@ -1168,10 +1168,9 @@ Package pdftex.def Info: uq_logo.jpg used on input line 9.`
1168		1168
1169	{/usr/local/texlive/2026/texmf-var/fonts/map/pdftex/updmap/pdftex.map}{/usr/loc	1169	{/usr/local/texlive/2026/texmf-var/fonts/map/pdftex/updmap/pdftex.map}{/usr/loc
1170	al/texlive/2026/texmf-dist/fonts/enc/dvips/base/8r.enc} <./uq_logo.jpg>] (./sec	1170	al/texlive/2026/texmf-dist/fonts/enc/dvips/base/8r.enc} <./uq_logo.jpg>] (./sec
1171	−tions/abstract.tex)	1171	+tions/abstract.tex
1172	−[2]) (./sections/introduction.tex
1173	LaTeX Font Info: Trying to load font information for OT1+ztmcm on input line	1172	LaTeX Font Info: Trying to load font information for OT1+ztmcm on input line
1174	− 8.	1173	+ 18.
1175		1174
1176	(/usr/local/texlive/2026/texmf-dist/tex/latex/psnfss/ot1ztmcm.fd	1175	(/usr/local/texlive/2026/texmf-dist/tex/latex/psnfss/ot1ztmcm.fd
1177	File: ot1ztmcm.fd 2000/01/03 Fontinst v1.801 font definitions for OT1/ztmcm.	1176	File: ot1ztmcm.fd 2000/01/03 Fontinst v1.801 font definitions for OT1/ztmcm.
				`@@ -1181,50 +1180,51 @@ Package microtype Info: Loading generic protrusion settings for font family`
1181	(microtype) For optimal results, create family-specific settings.	1180	(microtype) For optimal results, create family-specific settings.
1182	(microtype) See the microtype manual for details.	1181	(microtype) See the microtype manual for details.
1183	LaTeX Font Info: Trying to load font information for OML+ztmcm on input line	1182	LaTeX Font Info: Trying to load font information for OML+ztmcm on input line
1184	− 8.	1183	+ 18.
1185		1184
1186	(/usr/local/texlive/2026/texmf-dist/tex/latex/psnfss/omlztmcm.fd	1185	(/usr/local/texlive/2026/texmf-dist/tex/latex/psnfss/omlztmcm.fd
1187	File: omlztmcm.fd 2000/01/03 Fontinst v1.801 font definitions for OML/ztmcm.	1186	File: omlztmcm.fd 2000/01/03 Fontinst v1.801 font definitions for OML/ztmcm.
1188	)	1187	)
1189	LaTeX Font Info: Trying to load font information for OMS+ztmcm on input line	1188	LaTeX Font Info: Trying to load font information for OMS+ztmcm on input line
1190	− 8.	1189	+ 18.
1191		1190
1192	(/usr/local/texlive/2026/texmf-dist/tex/latex/psnfss/omsztmcm.fd	1191	(/usr/local/texlive/2026/texmf-dist/tex/latex/psnfss/omsztmcm.fd
1193	File: omsztmcm.fd 2000/01/03 Fontinst v1.801 font definitions for OMS/ztmcm.	1192	File: omsztmcm.fd 2000/01/03 Fontinst v1.801 font definitions for OMS/ztmcm.
1194	)	1193	)
1195	LaTeX Font Info: Trying to load font information for OMX+ztmcm on input line	1194	LaTeX Font Info: Trying to load font information for OMX+ztmcm on input line
1196	− 8.	1195	+ 18.
1197		1196
1198	(/usr/local/texlive/2026/texmf-dist/tex/latex/psnfss/omxztmcm.fd	1197	(/usr/local/texlive/2026/texmf-dist/tex/latex/psnfss/omxztmcm.fd
1199	File: omxztmcm.fd 2000/01/03 Fontinst v1.801 font definitions for OMX/ztmcm.	1198	File: omxztmcm.fd 2000/01/03 Fontinst v1.801 font definitions for OMX/ztmcm.
1200	)	1199	)
1201	−LaTeX Font Info: Trying to load font information for OT1+ptm on input line 8	1200	+LaTeX Font Info: Trying to load font information for OT1+ptm on input line 1
1202	−.	1201	+8.
1203		1202
1204	(/usr/local/texlive/2026/texmf-dist/tex/latex/psnfss/ot1ptm.fd	1203	(/usr/local/texlive/2026/texmf-dist/tex/latex/psnfss/ot1ptm.fd
1205	File: ot1ptm.fd 2001/06/04 font definitions for OT1/ptm.	1204	File: ot1ptm.fd 2001/06/04 font definitions for OT1/ptm.
1206	)	1205	)
1207	LaTeX Font Info: Font shape `OT1/ptm/bx/n' in size <12> not available	1206	LaTeX Font Info: Font shape `OT1/ptm/bx/n' in size <12> not available
1208	−(Font) Font shape `OT1/ptm/b/n' tried instead on input line 8.	1207	+(Font) Font shape `OT1/ptm/b/n' tried instead on input line 18.
1209	LaTeX Font Info: Font shape `OT1/ptm/bx/n' in size <9> not available	1208	LaTeX Font Info: Font shape `OT1/ptm/bx/n' in size <9> not available
1210	−(Font) Font shape `OT1/ptm/b/n' tried instead on input line 8.	1209	+(Font) Font shape `OT1/ptm/b/n' tried instead on input line 18.
1211	LaTeX Font Info: Font shape `OT1/ptm/bx/n' in size <7> not available	1210	LaTeX Font Info: Font shape `OT1/ptm/bx/n' in size <7> not available
1212	−(Font) Font shape `OT1/ptm/b/n' tried instead on input line 8.	1211	+(Font) Font shape `OT1/ptm/b/n' tried instead on input line 18.
1213		1212
1214	(/usr/local/texlive/2026/texmf-dist/tex/latex/microtype/mt-msa.cfg	1213	(/usr/local/texlive/2026/texmf-dist/tex/latex/microtype/mt-msa.cfg
1215	File: mt-msa.cfg 2006/02/04 v1.1 microtype config. file: AMS symbols (a) (RS)	1214	File: mt-msa.cfg 2006/02/04 v1.1 microtype config. file: AMS symbols (a) (RS)
1216	)	1215	)
1217	(/usr/local/texlive/2026/texmf-dist/tex/latex/microtype/mt-msb.cfg	1216	(/usr/local/texlive/2026/texmf-dist/tex/latex/microtype/mt-msb.cfg
1218	File: mt-msb.cfg 2005/06/01 v1.0 microtype config. file: AMS symbols (b) (RS)	1217	File: mt-msb.cfg 2005/06/01 v1.0 microtype config. file: AMS symbols (b) (RS)
1219	−) [3]	1218	+)) [2]
1220	−(./sections/findings.tex)) (./sections/related.tex	1219	+[3]) (./sections/introduction.tex [4] (./sections/findings.tex))
		1220	+(./sections/related.tex [5]
1221	Package microtype Info: Loading generic protrusion settings for font family	1221	Package microtype Info: Loading generic protrusion settings for font family
1222	(microtype) `cmtt' (encoding: T1).	1222	(microtype) `cmtt' (encoding: T1).
1223	(microtype) For optimal results, create family-specific settings.	1223	(microtype) For optimal results, create family-specific settings.
1224	(microtype) See the microtype manual for details.	1224	(microtype) See the microtype manual for details.
1225	− [4{/usr/local/texlive/2026/texmf-dist/fonts/enc/dvips/cm-super/cm-super-t1.enc	1225	+ [6{/usr/local/texlive/2026/texmf-dist/fonts/enc/dvips/cm-super/cm-super-t1.enc
1226	−}]) (./sections/problem.tex	1226	+}]) (./sections/problem.tex [7])
1227	−[5]) (./sections/architecture.tex (./figures/architecture.tex	1227	+(./sections/architecture.tex (./figures/architecture.tex
1228	LaTeX Font Info: Font shape `OT1/ptm/bx/n' in size <10> not available	1228	LaTeX Font Info: Font shape `OT1/ptm/bx/n' in size <10> not available
1229	(Font) Font shape `OT1/ptm/b/n' tried instead on input line 9.	1229	(Font) Font shape `OT1/ptm/b/n' tried instead on input line 9.
1230	LaTeX Font Info: Font shape `OT1/ptm/bx/n' in size <7.4> not available	1230	LaTeX Font Info: Font shape `OT1/ptm/bx/n' in size <7.4> not available
				@@ -1240,7 +1240,7 @@ LaTeX Font Info: Font shape `OT1/ptm/bx/n' in size <10.95> not available
1240	(Font) Font shape `OT1/ptm/b/n' tried instead on input line 13.	1240	(Font) Font shape `OT1/ptm/b/n' tried instead on input line 13.
1241	LaTeX Font Info: Font shape `OT1/ptm/bx/n' in size <8> not available	1241	LaTeX Font Info: Font shape `OT1/ptm/bx/n' in size <8> not available
1242	(Font) Font shape `OT1/ptm/b/n' tried instead on input line 13.	1242	(Font) Font shape `OT1/ptm/b/n' tried instead on input line 13.
1243	−[6]	1243	+
1244	Overfull \hbox (0.70027pt too wide) detected at line 28	1244	Overfull \hbox (0.70027pt too wide) detected at line 28
1245	[][][][] \OT1/ztmcm/m/n/12 = [][](\OML/ztmcm/m/it/12 H[]\OT1/ztmcm/m/n/12 )\OML	1245	[][][][] \OT1/ztmcm/m/n/12 = [][](\OML/ztmcm/m/it/12 H[]\OT1/ztmcm/m/n/12 )\OML
1246	/ztmcm/m/it/12 ; K[] \OT1/ztmcm/m/n/12 = []\OML/ztmcm/m/it/12 W[]; V[] \OT1/z	1246	/ztmcm/m/it/12 ; K[] \OT1/ztmcm/m/n/12 = []\OML/ztmcm/m/it/12 W[]; V[] \OT1/z
				`@@ -1249,12 +1249,12 @@ t/12 P []; W[]; W[]; P[] \OMS/ztmcm/m/n/12 2 \U/msb/m/n/12 R[]\OML/ztmcm/m/it/`
1249	12 :	1249	12 :
1250	[]	1250	[]
1251		1251
1252	−[7]	1252	+[8]
1253	Overfull \hbox (8.35876pt too wide) detected at line 60	1253	Overfull \hbox (8.35876pt too wide) detected at line 60
1254	[]	1254	[]
1255	[]	1255	[]
1256		1256
1257	−	1257	+[9]
1258	Overfull \hbox (22.86838pt too wide) detected at line 74	1258	Overfull \hbox (22.86838pt too wide) detected at line 74
1259	[][][]\OML/ztmcm/m/it/12 ^^N[] \OT1/ztmcm/m/n/12 = [] \OMS/ztmcm/m/n/12 ^^@ \OM	1259	[][][]\OML/ztmcm/m/it/12 ^^N[] \OT1/ztmcm/m/n/12 = [] \OMS/ztmcm/m/n/12 ^^@ \OM
1260	L/ztmcm/m/it/12 S[]; ^^N[] \OT1/ztmcm/m/n/12 = \OML/ztmcm/m/it/12 ^^N[] [] []	1260	L/ztmcm/m/it/12 S[]; ^^N[] \OT1/ztmcm/m/n/12 = \OML/ztmcm/m/it/12 ^^N[] [] []
				`@@ -1273,60 +1273,62 @@ mcm/m/n/12 (\OML/ztmcm/m/it/12 S; z[]\OT1/ztmcm/m/n/12 ) = [][] \OML/ztmcm/m/it`
1273	][](\OML/ztmcm/m/it/12 W[]z[]\OT1/ztmcm/m/n/12 ) \OMS/ztmcm/m/n/12 ^^A	1273	][](\OML/ztmcm/m/it/12 W[]z[]\OT1/ztmcm/m/n/12 ) \OMS/ztmcm/m/n/12 ^^A
1274	[]	1274	[]
1275		1275
1276	−[8] (./tables/configs.tex)) (./sections/policy.tex [9])	1276	+(./tables/configs.tex)) (./sections/policy.tex [10]) (./sections/learning.tex
1277	−(./sections/learning.tex [10]) (./sections/theory.tex [11]	1277	+[11] [12]) (./sections/theory.tex [13]
1278		1278
1279	LaTeX Font Warning: Font shape `T1/ptm/m/scit' undefined	1279	LaTeX Font Warning: Font shape `T1/ptm/m/scit' undefined
1280	(Font) using `T1/ptm/m/sc' instead on input line 37.	1280	(Font) using `T1/ptm/m/sc' instead on input line 37.
1281		1281
1282	−[12]) (./sections/implementation.tex (./tables/ops.tex	1282	+) (./sections/implementation.tex [14] (./tables/ops.tex
1283	Underfull \hbox (badness 10000) in paragraph at lines 5--5	1283	Underfull \hbox (badness 10000) in paragraph at lines 5--5
1284	[]\T1/ptm/b/n/10.95 (+20) Table 2. \|\T1/ptm/m/n/10.95 (+20) An-a-lytic op-er-a-	1284	[]\T1/ptm/b/n/10.95 (+20) Table 2. \|\T1/ptm/m/n/10.95 (+20) An-a-lytic op-er-a-
1285	tions per gen-er-ated to-ken, PRISME-459M, con-text $\OML/ztmcm/m/it/10.95 T \O	1285	tions per gen-er-ated to-ken, PRISME-459M, con-text $\OML/ztmcm/m/it/10.95 T \O
1286	T1/ztmcm/m/n/10.95 = 4096$	1286	T1/ztmcm/m/n/10.95 = 4096$
1287	[]	1287	[]
1288		1288
1289	−) [13]) (./sections/design.tex	1289	+)) (./sections/design.tex [15]
1290	Package microtype Info: Loading generic protrusion settings for font family	1290	Package microtype Info: Loading generic protrusion settings for font family
1291	(microtype) `cmss' (encoding: T1).	1291	(microtype) `cmss' (encoding: T1).
1292	(microtype) For optimal results, create family-specific settings.	1292	(microtype) For optimal results, create family-specific settings.
1293	(microtype) See the microtype manual for details.	1293	(microtype) See the microtype manual for details.
1294	− (./tables/families.tex) [14]	1294	+ (./tables/families.tex)
1295	Overfull \hbox (36.62411pt too wide) in paragraph at lines 19--24	1295	Overfull \hbox (36.62411pt too wide) in paragraph at lines 19--24
1296	\T1/ptm/m/n/12 (-20) 2.91M val-i-da-tion and 2.88M locked test to-kens (\T1/cmt	1296	\T1/ptm/m/n/12 (-20) 2.91M val-i-da-tion and 2.88M locked test to-kens (\T1/cmt
1297	t/m/n/12 data/processed/fineweb_edu_300m/manifest.json\T1/ptm/m/n/12 (-20) ).	1297	t/m/n/12 data/processed/fineweb_edu_300m/manifest.json\T1/ptm/m/n/12 (-20) ).
1298	[]	1298	[]
1299		1299
		1300	+[16]
1300		1301
1301	−LaTeX Warning: Reference `tab:controls' on page 15 undefined on input line 27.	1302	+LaTeX Warning: Reference `tab:controls' on page 17 undefined on input line 27.
1302		1303
1303	−[15]) (./sections/results.tex	1304	+) (./sections/results.tex
1304	Overfull \hbox (71.70984pt too wide) in paragraph at lines 38--1	1305	Overfull \hbox (71.70984pt too wide) in paragraph at lines 38--1
1305	\T1/ptm/m/n/12 (-20) The pri-mary out-come, met-rics, de-ci-sion thresh-olds an	1306	\T1/ptm/m/n/12 (-20) The pri-mary out-come, met-rics, de-ci-sion thresh-olds an
1306	d seed plan were writ-ten to \T1/cmtt/m/n/12 protocol/preregistration.md	1307	d seed plan were writ-ten to \T1/cmtt/m/n/12 protocol/preregistration.md
1307	[]	1308	[]
1308		1309
1309	−<../results/figures/learnability_bootstrap.pdf, id=482, 353.75287pt x 191.45125	1310	+[17]
		1311	+<../results/figures/learnability_bootstrap.pdf, id=494, 353.75287pt x 191.45125
1310	pt>	1312	pt>
1311	File: ../results/figures/learnability_bootstrap.pdf Graphic file (type pdf)	1313	File: ../results/figures/learnability_bootstrap.pdf Graphic file (type pdf)
1312	<use ../results/figures/learnability_bootstrap.pdf>	1314	<use ../results/figures/learnability_bootstrap.pdf>
1313	Package pdftex.def Info: ../results/figures/learnability_bootstrap.pdf used on	1315	Package pdftex.def Info: ../results/figures/learnability_bootstrap.pdf used on
1314	input line 26.	1316	input line 26.
1315	(pdftex.def) Requested size: 469.75502pt x 254.23257pt.	1317	(pdftex.def) Requested size: 469.75502pt x 254.23257pt.
1316	−(./tables/seed_matrix.tex	1318	+ (./tables/seed_matrix.tex
1317	Overfull \hbox (133.05286pt too wide) in paragraph at lines 7--21	1319	Overfull \hbox (133.05286pt too wide) in paragraph at lines 7--21
1318	[][]	1320	[][]
1319	[]	1321	[]
1320		1322
1321	)	1323	)
1322	−<../results/figures/frontier_v2.pdf, id=487, 305.58919pt x 208.87035pt>	1324	+<../results/figures/frontier_v2.pdf, id=499, 305.58919pt x 208.87035pt>
1323	File: ../results/figures/frontier_v2.pdf Graphic file (type pdf)	1325	File: ../results/figures/frontier_v2.pdf Graphic file (type pdf)
1324	<use ../results/figures/frontier_v2.pdf>	1326	<use ../results/figures/frontier_v2.pdf>
1325	Package pdftex.def Info: ../results/figures/frontier_v2.pdf used on input line	1327	Package pdftex.def Info: ../results/figures/frontier_v2.pdf used on input line
1326	59.	1328	59.
1327	(pdftex.def) Requested size: 338.22418pt x 231.19547pt.	1329	(pdftex.def) Requested size: 338.22418pt x 231.19547pt.
1328	− [16] [17 <../results/figures/learnability_bootstrap.pdf>] [18]	1330	+ [18 <../results/figures/learnability_bootstrap.pdf>] [19]
1329	−<../results/figures/labels_analysis.pdf, id=528, 548.77519pt x 192.17598pt>	1331	+<../results/figures/labels_analysis.pdf, id=536, 548.77519pt x 192.17598pt>
1330	File: ../results/figures/labels_analysis.pdf Graphic file (type pdf)	1332	File: ../results/figures/labels_analysis.pdf Graphic file (type pdf)
1331	<use ../results/figures/labels_analysis.pdf>	1333	<use ../results/figures/labels_analysis.pdf>
1332	Package pdftex.def Info: ../results/figures/labels_analysis.pdf used on input	1334	Package pdftex.def Info: ../results/figures/labels_analysis.pdf used on input
				`@@ -1338,12 +1340,12 @@ Overfull \hbox (109.74734pt too wide) in paragraph at lines 8--15`
1338	[][]	1340	[][]
1339	[]	1341	[]
1340		1342
1341	−) [19 <../results/figures/frontier_v2.pdf>] (./tables/stats.tex)	1343	+) [20 <../results/figures/frontier_v2.pdf>] (./tables/stats.tex) [21 <../result
1342	−(./tables/latency.tex) [20 <../results/figures/labels_analysis.pdf>]	1344	+s/figures/labels_analysis.pdf>] (./tables/test.tex) (./tables/latency.tex)
1343	−(./sections/results_pilot.tex (./tables/pilot_ppl.tex)	1345	+(./sections/results_pilot.tex [22] (./tables/pilot_ppl.tex)
1344	−(./tables/pilot_routing.tex) [21] (./tables/pilot_latency.tex)))	1346	+(./tables/pilot_routing.tex) (./tables/pilot_latency.tex)))
1345	−(./sections/mechanistic.tex [22]) (./sections/limitations.tex [23])	1347	+(./sections/mechanistic.tex [23]) (./sections/limitations.tex [24])
1346	−(./sections/conclusion.tex) (./sections/reproducibility.tex)	1348	+(./sections/conclusion.tex [25]) (./sections/reproducibility.tex)
1347	Overfull \hbox (51.18217pt too wide) in paragraph at lines 2--101	1349	Overfull \hbox (51.18217pt too wide) in paragraph at lines 2--101
1348	\T1/ptm/m/n/12 (-20) num-ber in the manuscript is pro-duced by \T1/cmtt/m/n/12	1350	\T1/ptm/m/n/12 (-20) num-ber in the manuscript is pro-duced by \T1/cmtt/m/n/12
1349	scripts/make_tables.py \T1/ptm/m/n/12 (-20) and \T1/cmtt/m/n/12 scripts/reprodu	1351	scripts/make_tables.py \T1/ptm/m/n/12 (-20) and \T1/cmtt/m/n/12 scripts/reprodu
				`@@ -1357,18 +1359,18 @@ cript_claims.py \T1/ptm/m/n/12 (-20) checks the chain. \T1/cmtt/m/n/12 REPRODUC`
1357	E.md	1359	E.md
1358	[]	1360	[]
1359		1361
1360	−[24] [25] (./main.bbl [26] [27]	1362	+[26] (./main.bbl [27] [28]
1361	Underfull \hbox (badness 1297) in paragraph at lines 208--211	1363	Underfull \hbox (badness 1297) in paragraph at lines 208--211
1362	[]\T1/ptm/b/n/12 (+20) You, Yichen, Tianyu Fu, Aosong Feng, Xing-tai Lv, Xue-fe	1364	[]\T1/ptm/b/n/12 (+20) You, Yichen, Tianyu Fu, Aosong Feng, Xing-tai Lv, Xue-fe
1363	i Ning, Ning Ding, and	1365	i Ning, Ning Ding, and
1364	[]	1366	[]
1365		1367
1366	−[28]) (./appendix/appendix.tex (./appendix/notation.tex	1368	+[29]) (./appendix/appendix.tex (./appendix/notation.tex
1367	Overfull \hbox (12.19864pt too wide) in paragraph at lines 2--16	1369	Overfull \hbox (12.19864pt too wide) in paragraph at lines 2--16
1368	[]	1370	[]
1369	[]	1371	[]
1370		1372
1371	−) (./appendix/proofs.tex [29]) (./appendix/params.tex [30]	1373	+) (./appendix/proofs.tex [30]) (./appendix/params.tex [31]
1372	Overfull \hbox (92.32437pt too wide) detected at line 12	1374	Overfull \hbox (92.32437pt too wide) detected at line 12
1373	\OML/ztmcm/m/it/12 N[] \OT1/ztmcm/m/n/12 = \OML/ztmcm/m/it/12 d \OT1/ztmcm/m/n/	1375	\OML/ztmcm/m/it/12 N[] \OT1/ztmcm/m/n/12 = \OML/ztmcm/m/it/12 d \OT1/ztmcm/m/n/
1374	12 + 3\OML/ztmcm/m/it/12 dw \OT1/ztmcm/m/n/12 + \OML/ztmcm/m/it/12 Mw \OT1/ztmc	1376	12 + 3\OML/ztmcm/m/it/12 dw \OT1/ztmcm/m/n/12 + \OML/ztmcm/m/it/12 Mw \OT1/ztmc
				@@ -1477,20 +1479,20 @@ LaTeX Warning: `h' float specifier changed to `ht'.
1477	) (./appendix/prereg.tex	1479	) (./appendix/prereg.tex
1478	LaTeX Font Info: Font shape `T1/cmtt/bx/n' in size <12> not available	1480	LaTeX Font Info: Font shape `T1/cmtt/bx/n' in size <12> not available
1479	(Font) Font shape `T1/cmtt/m/n' tried instead on input line 1.	1481	(Font) Font shape `T1/cmtt/m/n' tried instead on input line 1.
1480	− [31]	1482	+ [32]
1481	LaTeX Font Info: Trying to load font information for TS1+ptm on input line 8	1483	LaTeX Font Info: Trying to load font information for TS1+ptm on input line 8
1482	.	1484	.
1483		1485
1484	(/usr/local/texlive/2026/texmf-dist/tex/latex/psnfss/ts1ptm.fd	1486	(/usr/local/texlive/2026/texmf-dist/tex/latex/psnfss/ts1ptm.fd
1485	File: ts1ptm.fd 2001/06/04 font definitions for TS1/ptm.	1487	File: ts1ptm.fd 2001/06/04 font definitions for TS1/ptm.
1486	)	1488	)
1487	−Overfull \hbox (80.09856pt too wide) in paragraph at lines 10--12	1489	+Overfull \hbox (80.09856pt too wide) in paragraph at lines 24--26
1488	[]\T1/ptm/b/n/12 (-20) Synthetic com-po-si-tional suite v1 \T1/ptm/m/n/12 (-20)	1490	[]\T1/ptm/b/n/12 (-20) Synthetic com-po-si-tional suite v1 \T1/ptm/m/n/12 (-20)
1489	(\T1/cmtt/m/n/12 data/splits/synthetic_v1\T1/ptm/m/n/12 (-20) , gen-er-a-tor \	1491	(\T1/cmtt/m/n/12 data/splits/synthetic_v1\T1/ptm/m/n/12 (-20) , gen-er-a-tor \
1490	T1/cmtt/m/n/12 data/generators/synthetic.py\T1/ptm/m/n/12 (-20) ,	1492	T1/cmtt/m/n/12 data/generators/synthetic.py\T1/ptm/m/n/12 (-20) ,
1491	[]	1493	[]
1492		1494
1493	−[32]) (./appendix/runs.tex	1495	+[33] [34]) (./appendix/runs.tex
1494	Overfull \hbox (121.91187pt too wide) in paragraph at lines 1--14	1496	Overfull \hbox (121.91187pt too wide) in paragraph at lines 1--14
1495	\T1/cmtt/m/n/12 training.pretrain\T1/ptm/m/n/12 (-20) , \T1/cmtt/m/n/12 trainin	1497	\T1/cmtt/m/n/12 training.pretrain\T1/ptm/m/n/12 (-20) , \T1/cmtt/m/n/12 trainin
1496	g.train_editor\T1/ptm/m/n/12 (-20) , \T1/cmtt/m/n/12 training.collect_intervent	1498	g.train_editor\T1/ptm/m/n/12 (-20) , \T1/cmtt/m/n/12 training.collect_intervent
				`@@ -1505,14 +1507,14 @@ tion.latency \T1/ptm/m/n/12 (-20) or \T1/cmtt/m/n/12 evaluation.run\T1/ptm/m/n/`
1505	12 (-20) ; launch scripts un-der \T1/cmtt/m/n/12 scripts/	1507	12 (-20) ; launch scripts un-der \T1/cmtt/m/n/12 scripts/
1506	[]	1508	[]
1507		1509
1508	−[33] (./appendix/runs_index.tex [34]	1510	+(./appendix/runs_index.tex [35] [36]
1509	Overfull \hbox (25.12868pt too wide) in paragraph at lines 52--54	1511	Overfull \hbox (25.12868pt too wide) in paragraph at lines 52--54
1510	[]\T1/ptm/b/n/10 (-20) run_id\T1/ptm/m/n/10 (-20) : p4_tiny_backbone_v2_seed01_	1512	[]\T1/ptm/b/n/10 (-20) run_id\T1/ptm/m/n/10 (-20) : p4_tiny_backbone_v2_seed01_
1511	20261029233624; \T1/ptm/b/n/10 (-20) state\T1/ptm/m/n/10 (-20) : stopped_overfi	1513	20261029233624; \T1/ptm/b/n/10 (-20) state\T1/ptm/m/n/10 (-20) : stopped_overfi
1512	tting_step9300_checkpoint_step5000_used;	1514	tting_step9300_checkpoint_step5000_used;
1513	[]	1515	[]
1514		1516
1515	−[35])) (./appendix/results.tex (./tables/em_families.tex))	1517	+)) (./appendix/results.tex (./tables/em_families.tex))
1516	Overfull \hbox (168.4564pt too wide) in paragraph at lines 2--16	1518	Overfull \hbox (168.4564pt too wide) in paragraph at lines 2--16
1517	\T1/ptm/m/n/12 (-20) records (\T1/cmtt/m/n/12 experiments/runs//em_.json\T1/p	1519	\T1/ptm/m/n/12 (-20) records (\T1/cmtt/m/n/12 experiments/runs//em_.json\T1/p
1518	tm/m/n/12 (-20) , \T1/cmtt/m/n/12 results/raw/items_*.json\T1/ptm/m/n/12 (-20)	1520	tm/m/n/12 (-20) , \T1/cmtt/m/n/12 results/raw/items_*.json\T1/ptm/m/n/12 (-20)
				`@@ -1536,7 +1538,7 @@ ileNotFoundError\T1/ptm/m/n/12 (-20) ).; \T1/ptm/b/n/12 (-20) ev-i-dence\T1/ptm`
1536	g	1538	g
1537	[]	1539	[]
1538		1540
1539	−	1541	+[37]
1540	Overfull \hbox (49.9369pt too wide) in paragraph at lines 10--12	1542	Overfull \hbox (49.9369pt too wide) in paragraph at lines 10--12
1541	\T1/ptm/m/n/12 (-20) be-fore ed-its can af-fect log-its.; \T1/ptm/b/n/12 (-20)	1543	\T1/ptm/m/n/12 (-20) be-fore ed-its can af-fect log-its.; \T1/ptm/b/n/12 (-20)
1542	ev-i-dence\T1/ptm/m/n/12 (-20) : \T1/cmtt/m/n/12 experiments/runs/p4_editor_v1_	1544	ev-i-dence\T1/ptm/m/n/12 (-20) : \T1/cmtt/m/n/12 experiments/runs/p4_editor_v1_
				`@@ -1550,7 +1552,6 @@ Overfull \hbox (47.44559pt too wide) in paragraph at lines 12--14`
1550	;	1552	;
1551	[]	1553	[]
1552		1554
1553	−[36]
1554	Package microtype Info: Loading generic protrusion settings for font family	1555	Package microtype Info: Loading generic protrusion settings for font family
1555	(microtype) `cmtt' (encoding: TS1).	1556	(microtype) `cmtt' (encoding: TS1).
1556	(microtype) For optimal results, create family-specific settings.	1557	(microtype) For optimal results, create family-specific settings.
				`@@ -1569,7 +1570,8 @@ Overfull \hbox (12.79033pt too wide) in paragraph at lines 18--20`
1569	run \T1/cmtt/m/n/12 p4_tiny_backbone_v2_seed01_* \T1/ptm/m/n/12 (-20) (state:	1570	run \T1/cmtt/m/n/12 p4_tiny_backbone_v2_seed01_* \T1/ptm/m/n/12 (-20) (state:
1570	[]	1571	[]
1571		1572
1572	−	1573	+[38{/usr/local/texlive/2026/texmf-dist/fonts/enc/dvips/cm-super/cm-super-ts1.en
		1574	+c}]
1573	Overfull \hbox (12.01894pt too wide) in paragraph at lines 22--24	1575	Overfull \hbox (12.01894pt too wide) in paragraph at lines 22--24
1574	\T1/ptm/m/n/12 (-20) 0.4334); $\OML/ztmcm/m/it/12 ^^U$\T1/ptm/m/n/12 (-20) =0 i	1576	\T1/ptm/m/n/12 (-20) 0.4334); $\OML/ztmcm/m/it/12 ^^U$\T1/ptm/m/n/12 (-20) =0 i
1575	s worse (+1.3 %).; \T1/ptm/b/n/12 (-20) ev-i-dence\T1/ptm/m/n/12 (-20) : \T1/cm	1577	s worse (+1.3 %).; \T1/ptm/b/n/12 (-20) ev-i-dence\T1/ptm/m/n/12 (-20) : \T1/cm
				`@@ -1590,15 +1592,14 @@ Overfull \hbox (107.55685pt too wide) in paragraph at lines 26--28`
1590	gs/editor_pilot_103m_seed1.log\T1/ptm/m/n/12 (-20) ,	1592	gs/editor_pilot_103m_seed1.log\T1/ptm/m/n/12 (-20) ,
1591	[]	1593	[]
1592		1594
1593	−[37{/usr/local/texlive/2026/texmf-dist/fonts/enc/dvips/cm-super/cm-super-ts1.en	1595	+
1594	−c}]
1595	Overfull \hbox (46.17226pt too wide) in paragraph at lines 28--29	1596	Overfull \hbox (46.17226pt too wide) in paragraph at lines 28--29
1596	\T1/ptm/m/n/12 (-20) Killed; no 103M dense re-sult.; \T1/ptm/b/n/12 (-20) ev-i-	1597	\T1/ptm/m/n/12 (-20) Killed; no 103M dense re-sult.; \T1/ptm/b/n/12 (-20) ev-i-
1597	dence\T1/ptm/m/n/12 (-20) : M2U64 \T1/cmtt/m/n/12 experiments/logs/dense_probe_	1598	dence\T1/ptm/m/n/12 (-20) : M2U64 \T1/cmtt/m/n/12 experiments/logs/dense_probe_
1598	103m*_seed1.log\T1/ptm/m/n/12 (-20) ;	1599	103m*_seed1.log\T1/ptm/m/n/12 (-20) ;
1599	[]	1600	[]
1600		1601
1601	−)) [38] [39] [40] [41] [42] [43] [44] (./main.aux)	1602	+)) [39] [40] [41] [42] [43] [44] [45] (./main.aux)
1602	***********	1603	***********
1603	LaTeX2e <2025-11-01>	1604	LaTeX2e <2025-11-01>
1604	L3 programming layer <2026-01-19>	1605	L3 programming layer <2026-01-19>
				`@@ -1608,16 +1609,16 @@ L3 programming layer <2026-01-19>`
1608	LaTeX Warning: There were undefined references.	1609	LaTeX Warning: There were undefined references.
1609		1610
1610	Package rerunfilecheck Info: File `main.out' has not changed.	1611	Package rerunfilecheck Info: File `main.out' has not changed.
1611	−(rerunfilecheck) Checksum: 6517FE2FB3D993CA26B075D85418F650;9020.	1612	+(rerunfilecheck) Checksum: 5C078824F258CF894B89F741D112AD48;9386.
1612	)	1613	)
1613	Here is how much of TeX's memory you used:	1614	Here is how much of TeX's memory you used:
1614	− 36143 strings out of 467525	1615	+ 36174 strings out of 467525
1615	− 702671 string characters out of 5418982	1616	+ 703202 string characters out of 5418982
1616	− 1149081 words of memory out of 5000000	1617	+ 1152637 words of memory out of 5000000
1617	− 63439 multiletter control sequences out of 15000+600000	1618	+ 63458 multiletter control sequences out of 15000+600000
1618	− 793579 words of font info for 669 fonts, out of 8000000 for 9000	1619	+ 793836 words of font info for 673 fonts, out of 8000000 for 9000
1619	1141 hyphenation exceptions out of 8191	1620	1141 hyphenation exceptions out of 8191
1620	− 84i,13n,104p,1001b,1501s stack positions out of 10000i,1000n,20000p,200000b,200000s	1621	+ 84i,13n,104p,1001b,1484s stack positions out of 10000i,1000n,20000p,200000b,200000s
1621	</usr/local/texlive/2026/texmf-dist/fonts/type1/public/amsfonts/cm/cmex10.pfb	1622	</usr/local/texlive/2026/texmf-dist/fonts/type1/public/amsfonts/cm/cmex10.pfb
1622	></usr/local/texlive/2026/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi10.pfb>	1623	></usr/local/texlive/2026/texmf-dist/fonts/type1/public/amsfonts/cm/cmmi10.pfb>
1623	</usr/local/texlive/2026/texmf-dist/fonts/type1/public/amsfonts/cm/cmr10.pfb></	1624	</usr/local/texlive/2026/texmf-dist/fonts/type1/public/amsfonts/cm/cmr10.pfb></
				`@@ -1633,10 +1634,10 @@ exlive/2026/texmf-dist/fonts/type1/urw/symbol/usyr.pfb></usr/local/texlive/2026`
1633	e1/urw/times/utmbi8a.pfb></usr/local/texlive/2026/texmf-dist/fonts/type1/urw/ti	1634	e1/urw/times/utmbi8a.pfb></usr/local/texlive/2026/texmf-dist/fonts/type1/urw/ti
1634	mes/utmr8a.pfb></usr/local/texlive/2026/texmf-dist/fonts/type1/urw/times/utmri8	1635	mes/utmr8a.pfb></usr/local/texlive/2026/texmf-dist/fonts/type1/urw/times/utmri8
1635	a.pfb>	1636	a.pfb>
1636	−Output written on main.pdf (44 pages, 442504 bytes).	1637	+Output written on main.pdf (45 pages, 447044 bytes).
1637	PDF statistics:	1638	PDF statistics:
1638	− 812 PDF objects out of 1000 (max. 8388607)	1639	+ 824 PDF objects out of 1000 (max. 8388607)
1639	− 699 compressed objects within 7 object streams	1640	+ 709 compressed objects within 8 object streams
1640	− 164 named destinations out of 1000 (max. 500000)	1641	+ 167 named destinations out of 1000 (max. 500000)
1641	− 143777 words of extra memory for PDF output out of 154059 (max. 10000000)	1642	+ 143785 words of extra memory for PDF output out of 154059 (max. 10000000)
1642		1643

modified paper/main.pdf +0 −0

Binary file not shown.

modified paper/sections/abstract.tex +6 −3

				`@@ -15,6 +15,9 @@ not matter for fixed schedules. An oracle that reads the target finds large oppo`
15	of loss), but \labelscalibwithin{}\,\% of the variance of intervention gains is state-level rather than action-level,	15	of loss), but \labelscalibwithin{}\,\% of the variance of intervention gains is state-level rather than action-level,
16	the opportunity is concentrated in high-uncertainty states where a wrong action costs more than a right one gains,	16	the opportunity is concentrated in high-uncertainty states where a wrong action costs more than a right one gains,
17	and the learned controller ranks actions within a state near chance; it is worse than no routing on held-out data.	17	and the learned controller ranks actions within a state near chance; it is worse than no routing on held-out data.
18	−Measured on Apple Silicon, per-action dispatch overhead erases the arithmetic savings of sparsity at small scale.	18	+On the locked test regimes the pre-registered primary effect is $\testprismebtwovsbzero$ points against a three-point
19	−We report the 103M web-text pilot, the complete negative result on learned spatial routing, the mechanistic	19	+threshold. At 103M parameters on 98M tokens of web text, neither the sparse editor nor the global adapter moves the frozen
20	−explanation, and the conditions under which the hypothesis could still be rescued.	20	+backbone by more than 0.1\,\% of loss and a parameter-matched plain decoder does better than both. Measured on Apple
		21	+Silicon, per-action dispatch overhead erases most of the arithmetic savings of sparsity. We report the complete negative
		22	+result on learned spatial routing, its mechanistic explanation, and the conditions under which the hypothesis could
		23	+still be rescued.

modified paper/sections/findings.tex +16 −1

++On a frozen synthetic backbone with five adapter seeds, sparse one-or-two-slot edits lower the held-out
++answer-span loss by about 3\,\% and raise exact match by about two points at budget 2---but a parameter-matched
++global-recurrence adapter trained with the identical recipe does at least as well at every budget and dominates
++the quality-versus-cost frontier, and dense sixteen-slot writes do slightly better than sparse ones; on the locked
++test regimes the pre-registered primary effect is $\testprismebtwovsbzero$ points against a three-point threshold.
++Which slot is written does not matter for fixed schedules. An exhaustive one-step oracle that reads the target
++lowers the loss by 15\,\% on the synthetic suite and 3.6\,\% on web text, but the learned value controller---four
++feature sets, three loss variants, labels from an editor-unseen stream, calibration-based early stopping---never
++beats not routing on held-out data, because \labelscalibwithin{}\,\% of the variance of intervention gains is
++action-level, opportunity concentrates in high-uncertainty states where a wrong action costs more than a right one
++gains, and the controller's within-state ranking is at chance. On 98M tokens of web text at 103M parameters, neither
++adapter moves the frozen backbone by more than a tenth of a percent of loss, the parameter-matched plain decoder
++does better than both, and the zero-initialised readout prescribed by the proposal does not learn without a
++scale-aware bootstrap. Measured on Apple Silicon, eight sparse edits cost three times the backbone's decode
++latency at 103M against an analytic 1.6 times, and dense writes are only 17\,\% slower than sparse ones. The exact
++zero-budget bypass holds bitwise; the parameter and operation accounting of the proposal is reproduced.

modified paper/sections/results.tex +14 −0

				`@@ -139,6 +139,20 @@ items per thousand from wrong to right and \vtwoprismetransminus{} from right to`
139	No adapter changes the held-out regimes by more than a point, and the compose family stays at 2--3\,\%.	139	No adapter changes the held-out regimes by more than a point, and the compose family stays at 2--3\,\%.
140	\input{tables/stats.tex}	140	\input{tables/stats.tex}
141		141
		142	+\subsection{Pre-registered primary outcome on the locked test regimes}\label{sec:results-test}
		143	+The test streams (seed 9) were unsealed on 2026-10-30 after the pre-registration was frozen
		144	+(Appendix~\ref{app:prereg}); every system was evaluated once, at budget 0 and at forced random budget 2, on 1,000
		145	+stratified items per regime (Table~\ref{tab:test}). The pre-registered primary outcome---pooled exact match over
		146	+the iid, template and depth regimes at budget 2 versus the frozen backbone---is $\testprismebtwovsbzero$ points
		147	+for the sparse editor (95\,\% $t$-interval over \testprismebtwovsbzeron{} seeds
		148	+$[\testprismebtwovsbzerolo,\testprismebtwovsbzerohi]$), far below the pre-registered minimum effect of three
		149	+points; against the global-recurrence adapter the paired difference is $\testprismevsgr$ points
		150	+$[\testprismevsgrlo,\testprismevsgrhi]$ and against the dense editor $\testprismevsdense$ points
		151	+$[\testprismevsdenselo,\testprismevsdensehi]$. The compose regime stays at 4--5\,\% for every system. By the
		152	+decision rule fixed in advance, the architecture does not advance: it neither exceeds the strongest matched
		153	+baseline by three points at comparable cost nor reduces cost at non-inferior quality.
		154	+\input{tables/test.tex}
		155	+
142	\subsection{Latency and memory}\label{sec:results-latency}	156	\subsection{Latency and memory}\label{sec:results-latency}
143	Table~\ref{tab:latency} gives measured batch-1 and batch-8 decode latencies for the tiny model on the	157	Table~\ref{tab:latency} gives measured batch-1 and batch-8 decode latencies for the tiny model on the
144	development laptop (Apple M5 Max, fp32, PyTorch 2.15, MPS). Budget 0 costs \latencybzero{} ms per token; each	158	development laptop (Apple M5 Max, fp32, PyTorch 2.15, MPS). Budget 0 costs \latencybzero{} ms per token; each

modified paper/tables/numbers.tex +14 −2

				`@@ -50,8 +50,8 @@`
50	\newcommand{\pilotedminutes}{90}	50	\newcommand{\pilotedminutes}{90}
51	\newcommand{\pilotedparams}{1,907,713}	51	\newcommand{\pilotedparams}{1,907,713}
52	\newcommand{\pilotedpeakgb}{25.9}	52	\newcommand{\pilotedpeakgb}{25.9}
53	−\newcommand{\pilotenhours}{n/a}	53	+\newcommand{\pilotenhours}{2.4}
54	−\newcommand{\pilotenppl}{n/a}	54	+\newcommand{\pilotenppl}{47.9}
55	\newcommand{\pilotgrgainbeight}{0.07}	55	\newcommand{\pilotgrgainbeight}{0.07}
56	\newcommand{\pilotgrgainbfour}{0.08}	56	\newcommand{\pilotgrgainbfour}{0.08}
57	\newcommand{\pilotgrgainbone}{0.04}	57	\newcommand{\pilotgrgainbone}{0.04}
				`@@ -88,6 +88,18 @@`
88	\newcommand{\routinguniformbeight}{+0.6}	88	\newcommand{\routinguniformbeight}{+0.6}
89	\newcommand{\routinguniformbone}{-0.8}	89	\newcommand{\routinguniformbone}{-0.8}
90	\newcommand{\routinguniformbtwo}{-0.7}	90	\newcommand{\routinguniformbtwo}{-0.7}
		91	+\newcommand{\testprismebtwovsbzero}{+0.9}
		92	+\newcommand{\testprismebtwovsbzerohi}{+1.0}
		93	+\newcommand{\testprismebtwovsbzerolo}{+0.9}
		94	+\newcommand{\testprismebtwovsbzeron}{5}
		95	+\newcommand{\testprismevsdense}{+0.4}
		96	+\newcommand{\testprismevsdensehi}{+0.7}
		97	+\newcommand{\testprismevsdenselo}{+0.1}
		98	+\newcommand{\testprismevsdensen}{5}
		99	+\newcommand{\testprismevsgr}{-0.3}
		100	+\newcommand{\testprismevsgrhi}{+0.1}
		101	+\newcommand{\testprismevsgrlo}{-0.6}
		102	+\newcommand{\testprismevsgrn}{5}
91	\newcommand{\vtwobestfixedrel}{-4.2}	103	\newcommand{\vtwobestfixedrel}{-4.2}
92	\newcommand{\vtwobzeroloss}{0.4621}	104	\newcommand{\vtwobzeroloss}{0.4621}
93	\newcommand{\vtwoctrlbetter}{3}	105	\newcommand{\vtwoctrlbetter}{3}

added paper/tables/test.tex +14 −0

++\begin{table}[t]\centering\small
++\caption{LOCKED test regimes (synthetic suite, 1,000 stratified items per regime, unsealed 2026-10-30 after the
++pre-registration was frozen): exact match (\%) at budget 0 (frozen backbone, identical for all systems) and at forced
++random budget 2, mean $\pm$ SD over adapter seeds. ``pooled'' = the pre-registered primary outcome (iid+template+depth).}
++\label{tab:test}
++\begin{tabular}{llrrrrr}\toprule
++System & budget & iid & template & depth & compose & pooled\\\midrule
++PRISME sparse editor & 0 & 36.8 $\pm$ 0.0 & 25.4 $\pm$ 0.0 & 31.2 $\pm$ 0.0 & 4.6 $\pm$ 0.0 & 31.1 $\pm$ 0.0\\
++PRISME sparse editor & 2 & 38.3 $\pm$ 0.3 & 25.9 $\pm$ 0.5 & 32.1 $\pm$ 0.2 & 4.3 $\pm$ 0.3 & 32.1 $\pm$ 0.1\\
++Global recurrence & 0 & 36.8 $\pm$ 0.0 & 25.4 $\pm$ 0.0 & 31.2 $\pm$ 0.0 & 4.6 $\pm$ 0.0 & 31.1 $\pm$ 0.0\\
++Global recurrence & 2 & 38.3 $\pm$ 0.4 & 25.6 $\pm$ 0.4 & 33.2 $\pm$ 0.4 & 4.3 $\pm$ 0.5 & 32.3 $\pm$ 0.3\\
++Dense 16-slot & 0 & 36.8 $\pm$ 0.0 & 25.4 $\pm$ 0.0 & 31.2 $\pm$ 0.0 & 4.6 $\pm$ 0.0 & 31.1 $\pm$ 0.0\\
++Dense 16-slot & 2 & 37.7 $\pm$ 0.1 & 25.4 $\pm$ 0.4 & 32.0 $\pm$ 0.2 & 4.7 $\pm$ 0.2 & 31.7 $\pm$ 0.2\\
++\bottomrule\end{tabular}\end{table}

modified protocol/locked_test_manifest.json +5 −4

				`@@ -1,8 +1,8 @@`
1	{	1	{
2	− "status": "sealed",	2	+ "status": "unsealed",
3	− "unsealed": false,	3	+ "unsealed": true,
4	"sealed_at": "2026-10-29T23:30:00-0400",	4	"sealed_at": "2026-10-29T23:30:00-0400",
5	− "unsealed_at": null,	5	+ "unsealed_at": "2026-10-30T04:04:55-0400",
6	"locked_splits": {	6	"locked_splits": {
7	"test_iid": "f342d78bb14c8d19c513e9040a1b4f07dee5f866ad6c1a0fc5729a9d011ef45b",	7	"test_iid": "f342d78bb14c8d19c513e9040a1b4f07dee5f866ad6c1a0fc5729a9d011ef45b",
8	"test_template": "f9a7402e39b4304695625cd82a8b7ddd7d9b4f607860a5b26d9d7b2ceda50705",	8	"test_template": "f9a7402e39b4304695625cd82a8b7ddd7d9b4f607860a5b26d9d7b2ceda50705",
				`@@ -11,5 +11,6 @@`
11	},	11	},
12	"fineweb_test_sha256": "1ce9ca3ab6d659fe08042844fc0589951e09386ee79d2ef76504ffaaebb0947b",	12	"fineweb_test_sha256": "1ce9ca3ab6d659fe08042844fc0589951e09386ee79d2ef76504ffaaebb0947b",
13	"fineweb_test_tokens": 2883436,	13	"fineweb_test_tokens": 2883436,
14	− "rule": "evaluation.run refuses any split whose name starts with 'test' while unsealed is false; unsealing requires protocol/preregistration.md to be frozen (status FROZEN) and is recorded here with a timestamp."	14	+ "rule": "evaluation.run refuses any split whose name starts with 'test' while unsealed is false; unsealing requires protocol/preregistration.md to be frozen (status FROZEN) and is recorded here with a timestamp.",
		15	+ "preregistration_version": "1.0 frozen"
15	}	16	}
	\ No newline at end of file

modified protocol/preregistration.md +14 −3

				`@@ -1,7 +1,18 @@`
1	−# Pre-registration — WP31 PRISME pilot (DRAFT v0.1, 2026-10-29; to be FROZEN before any `test_*` split is read)	1	+# Pre-registration — WP31 PRISME pilot (v1.0, FROZEN 2026-10-30T04:04-0400 before any `test_*` split was read)
2		2
3	−Status: DRAFT. No file named `test*.jsonl` / `test.bin` has been opened by any evaluation script so far	3	+Status: FROZEN. No file named `test*.jsonl` / `test.bin` had been opened by any evaluation script at freeze time
4	−(`protocol/locked_test_manifest.json` records the split hashes and the unsealing timestamp once frozen).	4	+(`protocol/locked_test_manifest.json` records the split hashes and the unsealing timestamp).
		5	+
		6	+## 0. Primary comparisons as executed (fixed at freeze, before unsealing)
		7	+* P1 (H1/H2, synthetic): exact match on the pooled test regimes at forced budget 2 for the sparse editor vs the frozen
		8	+ backbone, vs the global-recurrence adapter and vs the dense editor, 5 adapter seeds each (v2 backbone), paired by item.
		9	+* P2 (H3, synthetic): dev-iid loss of the locked controller vs the best fixed schedule and vs budget 0 (already reported
		10	+ on dev; the test regimes are evaluated at budget 0 and 2 with forced random actions only, since the controller was
		11	+ worse than budget 0 on dev for every seed — the controller is therefore evaluated on test for completeness, not as the
		12	+ primary system).
		13	+* P3 (H1/H4, language): FineWeb-Edu locked test cross-entropy at budgets 0/1/2/4/8 for the 103M backbone, the sparse
		14	+ editor, the global adapter and the parameter-matched plain decoder.
		15	+* Decision thresholds as in Section 4 below; no other analysis is pre-registered.
5		16
6	## 1. Primary benchmark suite	17	## 1. Primary benchmark suite
7	* Synthetic compositional suite v1 (`data/splits/synthetic_v1`, generator `data/generators/synthetic.py`,	18	* Synthetic compositional suite v1 (`data/splits/synthetic_v1`, generator `data/generators/synthetic.py`,

added results/analyzed/test_statistics.json +69 −0

++{
++ "pooled_per_seed": {
++  "p4v2_prisme|budget0": [
++   0.3113333333333333,
++   0.3113333333333333,
++   0.3113333333333333,
++   0.3113333333333333,
++   0.3113333333333333
++  ],
++  "p4v2_prisme|forced_b2": [
++   0.3216666666666667,
++   0.32033333333333336,
++   0.3213333333333333,
++   0.32066666666666666,
++   0.32
++  ],
++  "p4v2_gr|budget0": [
++   0.3113333333333333,
++   0.3113333333333333,
++   0.3113333333333333,
++   0.3113333333333333,
++   0.3113333333333333
++  ],
++  "p4v2_gr|forced_b2": [
++   0.322,
++   0.3236666666666667,
++   0.32033333333333336,
++   0.3273333333333333,
++   0.3233333333333333
++  ],
++  "p4v2_dense|budget0": [
++   0.3113333333333333,
++   0.3113333333333333,
++   0.3113333333333333,
++   0.3113333333333333,
++   0.3113333333333333
++  ],
++  "p4v2_dense|forced_b2": [
++   0.3153333333333333,
++   0.31833333333333336,
++   0.3156666666666667,
++   0.31866666666666665,
++   0.316
++  ]
++ },
++ "stats": {
++  "prismebtwovsbzero": {
++   "mean": 0.009466666666666668,
++   "sd": 0.000691214711777591,
++   "ci_low": 0.008608410244023642,
++   "ci_high": 0.010324923089309695,
++   "n": 5
++  },
++  "prismevsgr": {
++   "mean": -0.0025333333333333336,
++   "sd": 0.002987008908962647,
++   "ci_low": -0.0062421947579061575,
++   "ci_high": 0.0011755280912394898,
++   "n": 5
++  },
++  "prismevsdense": {
++   "mean": 0.004,
++   "sd": 0.002013840995599095,
++   "ci_low": 0.0014994861376577788,
++   "ci_high": 0.006500513862342221,
++   "n": 5
++  }
++ }
++}
+ \ No newline at end of file

modified results/claim_registry.csv +3 −0

				`@@ -20,3 +20,6 @@ Miniature end-to-end pipeline runs on CPU,wall seconds,56,mini configs,mini_back`
20	103M global-recurrence adapter changes held-out loss by less than 0.1%,val ppl by budget,48.16 -> 48.01,FineWeb-Edu val,gr_pilot_103m_seed01_20261030015117,results/raw/eval_gr_103m_val.json,evaluation/run.py,complete	20	103M global-recurrence adapter changes held-out loss by less than 0.1%,val ppl by budget,48.16 -> 48.01,FineWeb-Edu val,gr_pilot_103m_seed01_20261030015117,results/raw/eval_gr_103m_val.json,evaluation/run.py,complete
21	103M controller: oracle opportunity exists (-3.6% at b8) but the controller captures none (-0.01%),val loss relative change,see numbers.tex pilotrouting*,FineWeb-Edu val,controller_103m_seed01,results/raw/routing_103m_val.json,evaluation/routing.py,complete	21	103M controller: oracle opportunity exists (-3.6% at b8) but the controller captures none (-0.01%),val loss relative change,see numbers.tex pilotrouting*,FineWeb-Edu val,controller_103m_seed01,results/raw/routing_103m_val.json,evaluation/routing.py,complete
22	103M latency: 8 sparse edits 3.0x budget-0 median latency; dense 3.6x; controller 1.5x,median ms/token batch 1,7.55 / 22.9 / 26.9 / 11.0,M3 Ultra MPS fp32,controller_103m_seed01,results/raw/latency_103m_mps.json,evaluation/latency.py,complete	22	103M latency: 8 sparse edits 3.0x budget-0 median latency; dense 3.6x; controller 1.5x,median ms/token batch 1,7.55 / 22.9 / 26.9 / 11.0,M3 Ultra MPS fp32,controller_103m_seed01,results/raw/latency_103m_mps.json,evaluation/latency.py,complete
		23	+Pre-registered primary outcome: pooled test EM gain of sparse editor at b2 is +0.9 pp (< 3 pp threshold),exact match pp,see numbers.tex testprismebtwovsbzero*,synthetic locked test regimes (1000 items each),p4v2_prisme (5 seeds),results/analyzed/test_statistics.json;paper/tables/test.tex,scripts/make_results_tables.py,complete
		24	+Sparse editor vs global recurrence on locked test: -0.3 pp [-0.6 +0.1]; vs dense +0.4 pp [+0.1 +0.7],exact match pp,see numbers.tex testprismevs*,synthetic locked test regimes,p4v2_prisme;p4v2_gr;p4v2_dense,results/analyzed/test_statistics.json,scripts/make_results_tables.py,complete
		25	+Parameter-matched plain decoder (102.2M) beats backbone+adapters on validation text,val loss,3.8689 vs 3.8746,FineWeb-Edu val,pretrain_enlarged102m_seed01_20261030015017,experiments/runs/pretrain_enlarged102m_seed01_20261030015017/manifest.json,training/pretrain.py,complete

added results/raw/eval_p4v2_dense_s1_test_compose.json +31 −0

++{
++ "system": "p4v2_dense_s1",
++ "split": "test_compose",
++ "checkpoint": "experiments/runs/p4v2_dense_seed01_20261030013421/checkpoint",
++ "checkpoint_sha256": "a31c97935f20348d4763244744bc1606292a50f1bcd8fddfc0df68b7550c29aa",
++ "config_sha256": "ed25665359232b92666bd3a813caa56cf6b32ba4c23af58a3c562f5c4b72d9e2",
++ "code_commit": "unknown",
++ "device": "mps",
++ "precision": "fp32",
++ "evaluated_at": "2026-10-30T04:06:12-0400",
++ "data_sha256": "06620e1cf60b7250aac71f4fb5ee74e3705f4b9aab102e22b5c5bdbed76c4ae2",
++ "exact_match": {
++  "budget0": {
++   "accuracy": 0.046296296296296294,
++   "n": 540,
++   "by_family": {
++    "compose": 0.046296296296296294
++   },
++   "edits_per_token": 0.0
++  },
++  "forced_b2": {
++   "accuracy": 0.046296296296296294,
++   "n": 540,
++   "by_family": {
++    "compose": 0.046296296296296294
++   },
++   "edits_per_token": 2.0
++  }
++ },
++ "seconds": 1.8
++}
+ \ No newline at end of file

added results/raw/eval_p4v2_dense_s1_test_depth.json +39 −0

++{
++ "system": "p4v2_dense_s1",
++ "split": "test_depth",
++ "checkpoint": "experiments/runs/p4v2_dense_seed01_20261030013421/checkpoint",
++ "checkpoint_sha256": "a31c97935f20348d4763244744bc1606292a50f1bcd8fddfc0df68b7550c29aa",
++ "config_sha256": "ed25665359232b92666bd3a813caa56cf6b32ba4c23af58a3c562f5c4b72d9e2",
++ "code_commit": "unknown",
++ "device": "mps",
++ "precision": "fp32",
++ "evaluated_at": "2026-10-30T04:05:59-0400",
++ "data_sha256": "06620e1cf60b7250aac71f4fb5ee74e3705f4b9aab102e22b5c5bdbed76c4ae2",
++ "exact_match": {
++  "budget0": {
++   "accuracy": 0.312,
++   "n": 1000,
++   "by_family": {
++    "arith": 0.0,
++    "bind": 0.02,
++    "fsm": 0.425,
++    "excl": 0.81,
++    "sched": 0.305
++   },
++   "edits_per_token": 0.0
++  },
++  "forced_b2": {
++   "accuracy": 0.317,
++   "n": 1000,
++   "by_family": {
++    "arith": 0.0,
++    "bind": 0.025,
++    "fsm": 0.42,
++    "excl": 0.805,
++    "sched": 0.335
++   },
++   "edits_per_token": 2.0
++  }
++ },
++ "seconds": 12.7
++}
+ \ No newline at end of file

added results/raw/eval_p4v2_dense_s1_test_iid.json +39 −0

++{
++ "system": "p4v2_dense_s1",
++ "split": "test_iid",
++ "checkpoint": "experiments/runs/p4v2_dense_seed01_20261030013421/checkpoint",
++ "checkpoint_sha256": "a31c97935f20348d4763244744bc1606292a50f1bcd8fddfc0df68b7550c29aa",
++ "config_sha256": "ed25665359232b92666bd3a813caa56cf6b32ba4c23af58a3c562f5c4b72d9e2",
++ "code_commit": "unknown",
++ "device": "mps",
++ "precision": "fp32",
++ "evaluated_at": "2026-10-30T04:05:31-0400",
++ "data_sha256": "06620e1cf60b7250aac71f4fb5ee74e3705f4b9aab102e22b5c5bdbed76c4ae2",
++ "exact_match": {
++  "budget0": {
++   "accuracy": 0.368,
++   "n": 1000,
++   "by_family": {
++    "bind": 0.2119815668202765,
++    "fsm": 0.4358974358974359,
++    "arith": 0.06701030927835051,
++    "excl": 0.3348623853211009,
++    "sched": 0.7813953488372093
++   },
++   "edits_per_token": 0.0
++  },
++  "forced_b2": {
++   "accuracy": 0.378,
++   "n": 1000,
++   "by_family": {
++    "bind": 0.2119815668202765,
++    "fsm": 0.4807692307692308,
++    "arith": 0.07731958762886598,
++    "excl": 0.3348623853211009,
++    "sched": 0.786046511627907
++   },
++   "edits_per_token": 2.0
++  }
++ },
++ "seconds": 15.0
++}
+ \ No newline at end of file

added results/raw/eval_p4v2_dense_s1_test_template.json +39 −0

++{
++ "system": "p4v2_dense_s1",
++ "split": "test_template",
++ "checkpoint": "experiments/runs/p4v2_dense_seed01_20261030013421/checkpoint",
++ "checkpoint_sha256": "a31c97935f20348d4763244744bc1606292a50f1bcd8fddfc0df68b7550c29aa",
++ "config_sha256": "ed25665359232b92666bd3a813caa56cf6b32ba4c23af58a3c562f5c4b72d9e2",
++ "code_commit": "unknown",
++ "device": "mps",
++ "precision": "fp32",
++ "evaluated_at": "2026-10-30T04:05:45-0400",
++ "data_sha256": "06620e1cf60b7250aac71f4fb5ee74e3705f4b9aab102e22b5c5bdbed76c4ae2",
++ "exact_match": {
++  "budget0": {
++   "accuracy": 0.254,
++   "n": 1000,
++   "by_family": {
++    "fsm": 0.430939226519337,
++    "arith": 0.024875621890547265,
++    "bind": 0.10679611650485436,
++    "sched": 0.4854368932038835,
++    "excl": 0.23786407766990292
++   },
++   "edits_per_token": 0.0
++  },
++  "forced_b2": {
++   "accuracy": 0.251,
++   "n": 1000,
++   "by_family": {
++    "fsm": 0.4530386740331492,
++    "arith": 0.029850746268656716,
++    "bind": 0.0825242718446602,
++    "sched": 0.4854368932038835,
++    "excl": 0.22330097087378642
++   },
++   "edits_per_token": 2.0
++  }
++ },
++ "seconds": 13.6
++}
+ \ No newline at end of file

added results/raw/eval_p4v2_dense_s2_test_compose.json +31 −0

++{
++ "system": "p4v2_dense_s2",
++ "split": "test_compose",
++ "checkpoint": "experiments/runs/p4v2_dense_seed02_20261030004544/checkpoint",
++ "checkpoint_sha256": "a7371f8c402f8a7fd98f784bb1c6750d70ccb46f53060656ba762df1166582b0",
++ "config_sha256": "dbf7526c8c09e5cb0257a92cb81256910e8697c6a3507f64b37a3af184e37f46",
++ "code_commit": "unknown",
++ "device": "mps",
++ "precision": "fp32",
++ "evaluated_at": "2026-10-30T04:06:59-0400",
++ "data_sha256": "06620e1cf60b7250aac71f4fb5ee74e3705f4b9aab102e22b5c5bdbed76c4ae2",
++ "exact_match": {
++  "budget0": {
++   "accuracy": 0.046296296296296294,
++   "n": 540,
++   "by_family": {
++    "compose": 0.046296296296296294
++   },
++   "edits_per_token": 0.0
++  },
++  "forced_b2": {
++   "accuracy": 0.04814814814814815,
++   "n": 540,
++   "by_family": {
++    "compose": 0.04814814814814815
++   },
++   "edits_per_token": 2.0
++  }
++ },
++ "seconds": 1.5
++}
+ \ No newline at end of file

added results/raw/eval_p4v2_dense_s2_test_depth.json +39 −0

++{
++ "system": "p4v2_dense_s2",
++ "split": "test_depth",
++ "checkpoint": "experiments/runs/p4v2_dense_seed02_20261030004544/checkpoint",
++ "checkpoint_sha256": "a7371f8c402f8a7fd98f784bb1c6750d70ccb46f53060656ba762df1166582b0",
++ "config_sha256": "dbf7526c8c09e5cb0257a92cb81256910e8697c6a3507f64b37a3af184e37f46",
++ "code_commit": "unknown",
++ "device": "mps",
++ "precision": "fp32",
++ "evaluated_at": "2026-10-30T04:06:49-0400",
++ "data_sha256": "06620e1cf60b7250aac71f4fb5ee74e3705f4b9aab102e22b5c5bdbed76c4ae2",
++ "exact_match": {
++  "budget0": {
++   "accuracy": 0.312,
++   "n": 1000,
++   "by_family": {
++    "arith": 0.0,
++    "bind": 0.02,
++    "fsm": 0.425,
++    "excl": 0.81,
++    "sched": 0.305
++   },
++   "edits_per_token": 0.0
++  },
++  "forced_b2": {
++   "accuracy": 0.319,
++   "n": 1000,
++   "by_family": {
++    "arith": 0.0,
++    "bind": 0.02,
++    "fsm": 0.425,
++    "excl": 0.805,
++    "sched": 0.345
++   },
++   "edits_per_token": 2.0
++  }
++ },
++ "seconds": 10.3
++}
+ \ No newline at end of file

added results/raw/eval_p4v2_dense_s2_test_iid.json +39 −0

++{
++ "system": "p4v2_dense_s2",
++ "split": "test_iid",
++ "checkpoint": "experiments/runs/p4v2_dense_seed02_20261030004544/checkpoint",
++ "checkpoint_sha256": "a7371f8c402f8a7fd98f784bb1c6750d70ccb46f53060656ba762df1166582b0",
++ "config_sha256": "dbf7526c8c09e5cb0257a92cb81256910e8697c6a3507f64b37a3af184e37f46",
++ "code_commit": "unknown",
++ "device": "mps",
++ "precision": "fp32",
++ "evaluated_at": "2026-10-30T04:06:25-0400",
++ "data_sha256": "06620e1cf60b7250aac71f4fb5ee74e3705f4b9aab102e22b5c5bdbed76c4ae2",
++ "exact_match": {
++  "budget0": {
++   "accuracy": 0.368,
++   "n": 1000,
++   "by_family": {
++    "bind": 0.2119815668202765,
++    "fsm": 0.4358974358974359,
++    "arith": 0.06701030927835051,
++    "excl": 0.3348623853211009,
++    "sched": 0.7813953488372093
++   },
++   "edits_per_token": 0.0
++  },
++  "forced_b2": {
++   "accuracy": 0.377,
++   "n": 1000,
++   "by_family": {
++    "bind": 0.22119815668202766,
++    "fsm": 0.4807692307692308,
++    "arith": 0.07731958762886598,
++    "excl": 0.3256880733944954,
++    "sched": 0.7813953488372093
++   },
++   "edits_per_token": 2.0
++  }
++ },
++ "seconds": 12.2
++}
+ \ No newline at end of file

added results/raw/eval_p4v2_dense_s2_test_template.json +39 −0

++{
++ "system": "p4v2_dense_s2",
++ "split": "test_template",
++ "checkpoint": "experiments/runs/p4v2_dense_seed02_20261030004544/checkpoint",
++ "checkpoint_sha256": "a7371f8c402f8a7fd98f784bb1c6750d70ccb46f53060656ba762df1166582b0",
++ "config_sha256": "dbf7526c8c09e5cb0257a92cb81256910e8697c6a3507f64b37a3af184e37f46",
++ "code_commit": "unknown",
++ "device": "mps",
++ "precision": "fp32",
++ "evaluated_at": "2026-10-30T04:06:37-0400",
++ "data_sha256": "06620e1cf60b7250aac71f4fb5ee74e3705f4b9aab102e22b5c5bdbed76c4ae2",
++ "exact_match": {
++  "budget0": {
++   "accuracy": 0.254,
++   "n": 1000,
++   "by_family": {
++    "fsm": 0.430939226519337,
++    "arith": 0.024875621890547265,
++    "bind": 0.10679611650485436,
++    "sched": 0.4854368932038835,
++    "excl": 0.23786407766990292
++   },
++   "edits_per_token": 0.0
++  },
++  "forced_b2": {
++   "accuracy": 0.259,
++   "n": 1000,
++   "by_family": {
++    "fsm": 0.44751381215469616,
++    "arith": 0.03980099502487562,
++    "bind": 0.07766990291262135,
++    "sched": 0.5097087378640777,
++    "excl": 0.23786407766990292
++   },
++   "edits_per_token": 2.0
++  }
++ },
++ "seconds": 11.5
++}
+ \ No newline at end of file

added results/raw/eval_p4v2_dense_s3_test_compose.json +31 −0

Large diff (32 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_dense_s3_test_depth.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_dense_s3_test_iid.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_dense_s3_test_template.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_dense_s4_test_compose.json +31 −0

Large diff (32 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_dense_s4_test_depth.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_dense_s4_test_iid.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_dense_s4_test_template.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_dense_s5_test_compose.json +31 −0

Large diff (32 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_dense_s5_test_depth.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_dense_s5_test_iid.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_dense_s5_test_template.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s1_test_compose.json +31 −0

Large diff (32 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s1_test_depth.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s1_test_iid.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s1_test_template.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s2_test_compose.json +31 −0

Large diff (32 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s2_test_depth.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s2_test_iid.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s2_test_template.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s3_test_compose.json +31 −0

Large diff (32 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s3_test_depth.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s3_test_iid.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s3_test_template.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s4_test_compose.json +31 −0

Large diff (32 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s4_test_depth.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s4_test_iid.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s4_test_template.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s5_test_compose.json +31 −0

Large diff (32 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s5_test_depth.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s5_test_iid.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_gr_s5_test_template.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s1_test_compose.json +31 −0

Large diff (32 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s1_test_depth.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s1_test_iid.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s1_test_template.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s2_test_compose.json +31 −0

Large diff (32 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s2_test_depth.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s2_test_iid.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s2_test_template.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s3_test_compose.json +31 −0

Large diff (32 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s3_test_depth.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s3_test_iid.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s3_test_template.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s4_test_compose.json +31 −0

Large diff (32 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s4_test_depth.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s4_test_iid.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s4_test_template.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s5_test_compose.json +31 −0

Large diff (32 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s5_test_depth.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s5_test_iid.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/eval_p4v2_prisme_s5_test_template.json +39 −0

Large diff (40 lines) is not loaded by default to keep the page fast.Load diff

added results/raw/items_p4v2_dense_s1_test_compose_budget0.json +1 −0