spb/wp31_uqo

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

modified FAILURES.md +1 −0

		@@ -13,3 +13,4 @@
13	13	| 2026-10-30 | Ranking-loss controller (D12: pairwise + listwise on top of Huber/BCE/NLL, editor-unseen labels): calibration sign accuracy fell to 45 %, captured fraction −8 %; on dev-iid the locked policy (λ=0.3, κ=0.5) edits 0.07 tokens and matches budget 0 (0.4333 vs 0.4334); λ=0 is worse (+1.3 %). | `experiments/runs/controller_v1rank_seed01_*`, `results/raw/routing_v1rank.json` | Within-state ranking is not learnable from these labels at this scale; H3 null stands. |
14	14	| 2026-10-30 | Stage A ran 6,000 steps instead of the intended 12,000 (stale config on the node, see rsync exclusion bug above); the enlarged baseline ran the 12,000-step config until step ~6,540 and was killed. | `experiments/runs/pretrain103m_seed01_20261029232706/manifest.json` (resolved_train.steps = 6000), M3U96b log | Backbone budget fixed at 98.3M tokens (D13); baseline relaunched with the same 6,000-step recipe. |
15	15	| 2026-10-30 | 103M editor pilot at lr 1e-3 collapsed to inert edits within 500 steps (val loss identical at every budget, 3.9889; trainable gradient norm 0.00 from step 160) — the same failure as the tiny w=128 lr 3e-3 run. | M3U96a `experiments/logs/editor_pilot_103m_seed1.log`, run `editor_pilot_103m_seed01_*` (killed at step ~650) | relaunched with lr 3e-4 (D14). |
	16	+| 2026-10-30 | 103M dense-slot probe collapsed to inert edits at lr 1e-3 (step 280, gradient norm 0.00) and again at lr 3e-4 (val loss identical at every budget, 3.9394, at step 500). Killed; no 103M dense result. | M2U64 `experiments/logs/dense_probe_103m*_seed1.log` | The dense ablation at 103M is reported as not obtained (collapse), with the tiny-scale 5-seed result as the only dense evidence. |

modified PROGRESS.md +1 −0

		@@ -19,3 +19,4 @@
19	19	| 2026-10-30 | 5/6 routing evaluation (tiny v1 editor, dev-iid, 640 items / 3,417 answer tokens) | done | `evaluation.routing` → `results/raw/routing_controller_p4_hent_seed01_20261030002015_dev_iid.json` | budget 0: 0.4334; uniform b1 0.4301 (−0.8 %), random b1 0.4305, fixed pair b1 0.4306; b8 schedules +0.6 to +2.8 %; entropy router ≥ budget 0; norm router (thr 0.2) 0.4309; learned controller (locked λ=0.3) 0.4404 (+1.6 %) and worse at every λ; greedy one-step oracle 0.4025 (b1) → 0.3666 (b8, 2.26 edits/token, −15 %); two-step oracle gain 0.034 vs one-step 0.0233 vs greedy-then-best 0.0337 (myopia regret ≈ 0) | H2 null (uniform = random = pair); H3 null (controller < no routing on dev) at this scale; oracle opportunity large | labels from a split unseen by the editor (D11); 5-seed v2 matrix |
20	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	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 |

modified configs/experiments/controller_103m.yaml +1 −1

		@@ -2,7 +2,7 @@ name: controller_103m
2	2	seed: 1
3	3	mode: prisme
4	4	model: from_checkpoint
5		−checkpoint: EDITOR103M_CKPT
	5	+checkpoint: experiments/runs/editor_pilot_103m_lr3e4_seed01_20261030021200/checkpoint
6	6	labels:
7	7	train: LABELS_DIR/labels_train.pt
8	8	calib: LABELS_DIR/labels_calib.pt

modified configs/experiments/labels_103m.yaml +1 −1

		@@ -2,7 +2,7 @@ name: labels_103m
2	2	seed: 1
3	3	mode: prisme
4	4	model: from_checkpoint
5		−checkpoint: EDITOR103M_CKPT
	5	+checkpoint: experiments/runs/editor_pilot_103m_lr3e4_seed01_20261030021200/checkpoint
6	6	use_policy: false
7	7	data:
8	8	kind: tokens

modified configs/experiments/locked_evaluation.yaml +17 −10

		@@ -1,17 +1,23 @@
1		−# Locked evaluation of the 103M pilot: held-out token cross-entropy at forced budgets and under the controller.
2		−# Splits: val (development) now; test (LOCKED) only after protocol/preregistration.md is frozen and the manifest unsealed.
3	1	out_dir: results/raw
4		−data: {kind: tokens, dir: data/processed/fineweb_edu_300m}
5		−splits: [val]
	2	+data:
	3	+ kind: tokens
	4	+ dir: data/processed/fineweb_edu_300m
	5	+splits:
	6	+- val
6	7	seq_len: 512
7	8	batch_size: 16
8		−n_batches: 40 # 40 x 16 x 512 = 327,680 held-out tokens
9		−budgets: [0, 1, 2, 4, 8]
	9	+n_batches: 40
	10	+budgets:
	11	+- 0
	12	+- 1
	13	+- 2
	14	+- 4
	15	+- 8
10	16	systems:
11	17	backbone_103m:
12	18	mode: backbone_only
13	19	model: from_checkpoint
14		− checkpoint: PRETRAIN103M_CKPT
	20	+ checkpoint: experiments/runs/pretrain103m_seed01_20261029232706/checkpoint
15	21	controller: false
16	22	enlarged_102m:
17	23	mode: backbone_only
		@@ -21,11 +27,12 @@ systems:
21	27	prisme_103m_editor:
22	28	mode: prisme
23	29	model: from_checkpoint
24		− checkpoint: EDITOR103M_CKPT
	30	+ checkpoint: experiments/runs/editor_pilot_103m_lr3e4_seed01_20261030021200/checkpoint
25	31	controller: false
26	32	gr_103m:
27	33	mode: global_recurrence
28		− baseline_kwargs: {ffn_w: 2560}
	34	+ baseline_kwargs:
	35	+ ffn_w: 2560
29	36	model: from_checkpoint
30		− checkpoint: GR103M_CKPT
	37	+ checkpoint: experiments/runs/gr_pilot_103m_seed01_20261030015117/checkpoint
31	38	controller: false

added experiments/runs/controller_103m_seed01_20261030034451/config.yaml +65 −0

		@@ -0,0 +1,65 @@
	1	+experiment:
	2	+ name: controller_103m
	3	+ seed: 1
	4	+ mode: prisme
	5	+ model: from_checkpoint
	6	+ checkpoint: experiments/runs/editor_pilot_103m_lr3e4_seed01_20261030021200/checkpoint
	7	+ labels:
	8	+ train: experiments/runs/labels_103m_seed01_20261030034314/labels_train.pt
	9	+ calib: experiments/runs/labels_103m_seed01_20261030034314/labels_calib.pt
	10	+ train:
	11	+ steps: 1500
	12	+ lr: 0.001
	13	+ weight_decay: 0.01
	14	+ alpha: 1.0
	15	+ beta: 0.5
	16	+ gamma: 0.1
	17	+ huber_delta: 0.1
	18	+ log_every: 100
	19	+ eval_every: 250
	20	+ seq_len: 512
	21	+ lambda_grid:
	22	+ - 0.0
	23	+ - 0.01
	24	+ - 0.03
	25	+ - 0.1
	26	+ - 0.3
	27	+ kappa_grid:
	28	+ - 0.0
	29	+ - 0.5
	30	+ sweep_batches: 6
	31	+ max_budget: 8
	32	+ selection_rule: max_loss_reduction_per_ops
	33	+ model_overrides: {}
	34	+model:
	35	+ name: prisme_103m
	36	+ vocab_size: 32768
	37	+ d_model: 768
	38	+ n_layers: 12
	39	+ n_heads: 12
	40	+ n_kv_heads: 3
	41	+ ffn_width: 2048
	42	+ max_seq_len: 4096
	43	+ rope_theta: 10000.0
	44	+ norm_eps: 1.0e-05
	45	+ tie_embeddings: true
	46	+ init_std: 0.02
	47	+ use_workspace: true
	48	+ w: 256
	49	+ ffn_w: 768
	50	+ n_slots: 16
	51	+ n_editor_blocks: 2
	52	+ n_editor_heads: 4
	53	+ r_max: 8
	54	+ eta_max: 0.25
	55	+ clip_c: 1.0
	56	+ clip_eps: 1.0e-06
	57	+ readout_zero_init: false
	58	+ readout_init_std: 0.2
	59	+ readout_scale: 8.851566314697266
	60	+ controller_hidden: 0
	61	+ ctrl_input: h
	62	+ ctrl_use_entropy: true
	63	+ lambda_cost: 0.0
	64	+ kappa_risk: 0.0
	65	+ dense_writes: false

added experiments/runs/controller_103m_seed01_20261030034451/manifest.json +151 −0

		@@ -0,0 +1,151 @@
	1	+{
	2	+ "run_id": "controller_103m_seed01_20261030034451",
	3	+ "name": "controller_103m",
	4	+ "seed": 1,
	5	+ "code_commit": "unknown",
	6	+ "config_sha256": "cb64baa3261e25b692057f7be7ff4ff4d1fd1f0686f233f60c9906c26164e32e",
	7	+ "started_at": "2026-10-30T03:44:51-0400",
	8	+ "host": "MacStudnpierre2",
	9	+ "python": "3.12.13",
	10	+ "torch": "2.15.0",
	11	+ "command": "/Users/simon-pierreboucher/wp31/prisme-slm/training/train_controller.py --config configs/experiments/controller_103m.yaml --device mps",
	12	+ "state": "completed",
	13	+ "wall_seconds": 315.08,
	14	+ "device": "mps",
	15	+ "checkpoint_parent": "experiments/runs/editor_pilot_103m_lr3e4_seed01_20261030021200/checkpoint",
	16	+ "labels_train": "experiments/runs/labels_103m_seed01_20261030034314/labels_train.pt",
	17	+ "labels_calib": "experiments/runs/labels_103m_seed01_20261030034314/labels_calib.pt",
	18	+ "finished_at": "2026-10-30T03:50:06-0400",
	19	+ "eval_train": {
	20	+ "huber": 0.0005079921786091291,
	21	+ "sign_acc": 0.5165206909179687,
	22	+ "harm_brier": 0.23502638936042786,
	23	+ "harm_ece": 0.017360106110572815,
	24	+ "n_labels": 655360,
	25	+ "oracle_opportunity_per_state": 0.0076657113619148735,
	26	+ "regret_per_state": 0.0072254255879670385,
	27	+ "captured_fraction": 0.057435735988610515,
	28	+ "top1_match": 0.0565185546875,
	29	+ "n_audit_states": 40960
	30	+ },
	31	+ "eval_calib": {
	32	+ "huber": 0.0005022834433475509,
	33	+ "sign_acc": 0.5141215006510417,
	34	+ "harm_brier": 0.2452416201432546,
	35	+ "harm_ece": 0.03279714658856392,
	36	+ "n_labels": 491520,
	37	+ "oracle_opportunity_per_state": 0.008453587186522782,
	38	+ "regret_per_state": 0.007839534722734243,
	39	+ "captured_fraction": 0.07263809436631807,
	40	+ "top1_match": 0.055999755859375,
	41	+ "n_audit_states": 32768
	42	+ },
	43	+ "sweep": [
	44	+ {
	45	+ "lambda": 0.0,
	46	+ "kappa": 0.0,
	47	+ "loss": 3.8988457918167114,
	48	+ "edits_per_token": 4.8095703125,
	49	+ "extra_ops_ratio": 0.4544869214296341,
	50	+ "loss_reduction": 0.0015107393264770508,
	51	+ "reduction_per_ops": 0.0033240545662455347
	52	+ },
	53	+ {
	54	+ "lambda": 0.0,
	55	+ "kappa": 0.5,
	56	+ "loss": 3.8992578188578286,
	57	+ "edits_per_token": 2.7021891276041665,
	58	+ "extra_ops_ratio": 0.23505109051863352,
	59	+ "loss_reduction": 0.0010987122853598486,
	60	+ "reduction_per_ops": 0.00467435519203749
	61	+ },
	62	+ {
	63	+ "lambda": 0.01,
	64	+ "kappa": 0.0,
	65	+ "loss": 3.898860216140747,
	66	+ "edits_per_token": 4.6912841796875,
	67	+ "extra_ops_ratio": 0.43767033020655316,
	68	+ "loss_reduction": 0.0014963150024414062,
	69	+ "reduction_per_ops": 0.003418817541813352
	70	+ },
	71	+ {
	72	+ "lambda": 0.01,
	73	+ "kappa": 0.5,
	74	+ "loss": 3.8993543783823648,
	75	+ "edits_per_token": 2.5965576171875,
	76	+ "extra_ops_ratio": 0.2230608140428861,
	77	+ "loss_reduction": 0.0010021527608237157,
	78	+ "reduction_per_ops": 0.004492733361185708
	79	+ },
	80	+ {
	81	+ "lambda": 0.03,
	82	+ "kappa": 0.0,
	83	+ "loss": 3.8987175623575845,
	84	+ "edits_per_token": 4.466267903645833,
	85	+ "extra_ops_ratio": 0.4043350915114085,
	86	+ "loss_reduction": 0.0016389687856039892,
	87	+ "reduction_per_ops": 0.004053491324429715
	88	+ },
	89	+ {
	90	+ "lambda": 0.03,
	91	+ "kappa": 0.5,
	92	+ "loss": 3.8993316888809204,
	93	+ "edits_per_token": 2.40234375,
	94	+ "extra_ops_ratio": 0.20202602942784628,
	95	+ "loss_reduction": 0.0010248422622680664,
	96	+ "reduction_per_ops": 0.005072822869263435
	97	+ },
	98	+ {
	99	+ "lambda": 0.1,
	100	+ "kappa": 0.0,
	101	+ "loss": 3.8987818559010825,
	102	+ "edits_per_token": 3.6865641276041665,
	103	+ "extra_ops_ratio": 0.3067183345556259,
	104	+ "loss_reduction": 0.0015746752421059718,
	105	+ "reduction_per_ops": 0.00513394559339716
	106	+ },
	107	+ {
	108	+ "lambda": 0.1,
	109	+ "kappa": 0.5,
	110	+ "loss": 3.8996830383936563,
	111	+ "edits_per_token": 1.7566324869791667,
	112	+ "extra_ops_ratio": 0.14280162130792937,
	113	+ "loss_reduction": 0.0006734927495322118,
	114	+ "reduction_per_ops": 0.004716282233798523
	115	+ },
	116	+ {
	117	+ "lambda": 0.3,
	118	+ "kappa": 0.0,
	119	+ "loss": 3.8995965321858725,
	120	+ "edits_per_token": 1.7585856119791667,
	121	+ "extra_ops_ratio": 0.1424483023583889,
	122	+ "loss_reduction": 0.0007599989573159327,
	123	+ "reduction_per_ops": 0.0053352615983013555
	124	+ },
	125	+ {
	126	+ "lambda": 0.3,
	127	+ "kappa": 0.5,
	128	+ "loss": 3.900011897087097,
	129	+ "edits_per_token": 0.5972493489583334,
	130	+ "extra_ops_ratio": 0.048378171709676586,
	131	+ "loss_reduction": 0.0003446340560913086,
	132	+ "reduction_per_ops": 0.00712375114461746
	133	+ }
	134	+ ],
	135	+ "baseline_budget0": {
	136	+ "loss": 3.9003565311431885,
	137	+ "edits_per_token": 0.0,
	138	+ "extra_ops_ratio": 0.0
	139	+ },
	140	+ "selected": {
	141	+ "lambda": 0.3,
	142	+ "kappa": 0.5,
	143	+ "loss": 3.900011897087097,
	144	+ "edits_per_token": 0.5972493489583334,
	145	+ "extra_ops_ratio": 0.048378171709676586,
	146	+ "loss_reduction": 0.0003446340560913086,
	147	+ "reduction_per_ops": 0.00712375114461746
	148	+ },
	149	+ "checkpoint_sha256": "64d08e9120608983b261bd6d108eb2421da45c84ebc733fa5fb2b88780dae427",
	150	+ "selection_rule": "max_loss_reduction_per_ops"
	151	+}
		\ No newline at end of file

added experiments/runs/controller_103m_seed01_20261030034451/metrics.jsonl +23 −0

		@@ -0,0 +1,23 @@
	1	+{"step": 0, "loss": 0.4732295870780945, "gain": 0.03798504173755646, "harm": 0.7073167562484741, "nll": 0.8158618211746216, "t": 1.59}
	2	+{"step": 100, "loss": 0.17227067053318024, "gain": 0.0041774241253733635, "harm": 0.6836851239204407, "nll": -1.7374930381774902, "t": 18.86}
	3	+{"step": 200, "loss": 0.12517236173152924, "gain": 0.001965296920388937, "harm": 0.670852780342102, "nll": -2.1221930980682373, "t": 34.25}
	4	+{"step": 249, "calib_huber": 0.000880393876771753, "calib_sign_acc": 0.5023295084635416, "calib_harm_brier": 0.24648866603771846, "calib_harm_ece": 0.03357856348156929, "calib_n_labels": 491520, "calib_oracle_opportunity_per_state": 0.008453587186522782, "calib_regret_per_state": 0.00804906408302486, "calib_captured_fraction": 0.047852242435357795, "calib_top1_match": 0.0699462890625, "calib_n_audit_states": 32768, "t": 46.9}
	5	+{"step": 300, "loss": 0.0760500431060791, "gain": 0.0009165772935375571, "harm": 0.6585628986358643, "nll": -2.5414798259735107, "t": 55.55}
	6	+{"step": 400, "loss": 0.06960448622703552, "gain": 0.0009221362415701151, "harm": 0.6586024761199951, "nll": -2.6061887741088867, "t": 72.36}
	7	+{"step": 499, "calib_huber": 0.0005022834433475509, "calib_sign_acc": 0.5141215006510417, "calib_harm_brier": 0.2452416201432546, "calib_harm_ece": 0.03279714658856392, "calib_n_labels": 491520, "calib_oracle_opportunity_per_state": 0.008453587186522782, "calib_regret_per_state": 0.007839534722734243, "calib_captured_fraction": 0.07263809436631807, "calib_top1_match": 0.055999755859375, "calib_n_audit_states": 32768, "t": 93.56}
	8	+{"step": 500, "loss": 0.034200459718704224, "gain": 0.00042655051220208406, "harm": 0.670921266078949, "nll": -3.016867160797119, "t": 93.76}
	9	+{"step": 600, "loss": 0.04568794369697571, "gain": 0.0006184718804433942, "harm": 0.662847638130188, "nll": -2.8635432720184326, "t": 110.57}
	10	+{"step": 700, "loss": -0.023632705211639404, "gain": 0.000218810499063693, "harm": 0.6303975582122803, "nll": -3.3905029296875, "t": 127.13}
	11	+{"step": 749, "calib_huber": 0.0002467253361828625, "calib_sign_acc": 0.4916748046875, "calib_harm_brier": 0.24779879599809645, "calib_harm_ece": 0.045154254883527756, "calib_n_labels": 491520, "calib_oracle_opportunity_per_state": 0.008453587186522782, "calib_regret_per_state": 0.008158762648236006, "calib_captured_fraction": 0.03487567251412538, "calib_top1_match": 0.0587158203125, "calib_n_audit_states": 32768, "t": 139.56}
	12	+{"step": 800, "loss": -0.007390350103378296, "gain": 0.00031520085758529603, "harm": 0.6132218837738037, "nll": -3.14316463470459, "t": 147.7}
	13	+{"step": 900, "loss": -0.010009020566940308, "gain": 0.00029699975857511163, "harm": 0.6271675825119019, "nll": -3.238898277282715, "t": 164.6}
	14	+{"step": 999, "calib_huber": 0.00017406620681867935, "calib_sign_acc": 0.50128173828125, "calib_harm_brier": 0.2535098448395729, "calib_harm_ece": 0.07380883395671844, "calib_n_labels": 491520, "calib_oracle_opportunity_per_state": 0.008453587186522782, "calib_regret_per_state": 0.008104458509478718, "calib_captured_fraction": 0.041299470785687964, "calib_top1_match": 0.065460205078125, "calib_n_audit_states": 32768, "t": 186.44}
	15	+{"step": 1000, "loss": -0.055137813091278076, "gain": 0.00015518153668381274, "harm": 0.6113837957382202, "nll": -3.609848976135254, "t": 186.6}
	16	+{"step": 1100, "loss": -0.03290015459060669, "gain": 0.00020753977878484875, "harm": 0.609285831451416, "nll": -3.3775060176849365, "t": 202.61}
	17	+{"step": 1200, "loss": -0.054915547370910645, "gain": 0.0002156073460355401, "harm": 0.5586880445480347, "nll": -3.3447518348693848, "t": 219.28}
	18	+{"step": 1249, "calib_huber": 0.0001583198141209626, "calib_sign_acc": 0.512738037109375, "calib_harm_brier": 0.2619521975517273, "calib_harm_ece": 0.10497063398361206, "calib_n_labels": 491520, "calib_oracle_opportunity_per_state": 0.008453587186522782, "calib_regret_per_state": 0.007955018023494631, "calib_captured_fraction": 0.05897723085212869, "calib_top1_match": 0.06988525390625, "calib_n_audit_states": 32768, "t": 231.67}
	19	+{"step": 1300, "loss": -0.046492695808410645, "gain": 0.0002507793833501637, "harm": 0.5760549306869507, "nll": -3.3477094173431396, "t": 239.51}
	20	+{"step": 1400, "loss": -0.10818794369697571, "gain": 0.00016302871517837048, "harm": 0.502214789390564, "nll": -3.594583511352539, "t": 255.57}
	21	+{"step": 1499, "loss": -0.09765100479125977, "gain": 0.0002172768145101145, "harm": 0.4987705647945404, "nll": -3.4725356101989746, "t": 271.95}
	22	+{"step": 1499, "calib_huber": 0.0001905371716323619, "calib_sign_acc": 0.5203938802083333, "calib_harm_brier": 0.2722386827071508, "calib_harm_ece": 0.1319294273853302, "calib_n_labels": 491520, "calib_oracle_opportunity_per_state": 0.008453587186522782, "calib_regret_per_state": 0.007843576604500413, "calib_captured_fraction": 0.07215996813694481, "calib_top1_match": 0.077239990234375, "calib_n_audit_states": 32768, "t": 276.65}
	23	+{"event": "restored_best_calibration_checkpoint", "step": 500, "score": -0.007839534722734243, "t": 276.65}

added experiments/runs/editor_pilot_103m_lr3e4_seed01_20261030021200/config.yaml +88 −0

		@@ -0,0 +1,88 @@
	1	+experiment:
	2	+ seed: 1
	3	+ model: configs/prisme_103m.yaml
	4	+ model_overrides:
	5	+ readout_zero_init: false
	6	+ readout_init_std: 0.2
	7	+ ctrl_input: h
	8	+ ctrl_use_entropy: true
	9	+ init:
	10	+ backbone_checkpoint: experiments/runs/pretrain103m_seed01_20261029232706/checkpoint
	11	+ freeze_backbone: true
	12	+ auto_readout_scale: true
	13	+ data:
	14	+ kind: tokens
	15	+ train_bin: data/processed/fineweb_edu_300m/train.bin
	16	+ val_bin: data/processed/fineweb_edu_300m/val.bin
	17	+ budgets:
	18	+ - 0
	19	+ - 1
	20	+ - 2
	21	+ - 4
	22	+ - 8
	23	+ budget_probs:
	24	+ - 0.2
	25	+ - 0.2
	26	+ - 0.2
	27	+ - 0.2
	28	+ - 0.2
	29	+ budget_per: position
	30	+ train:
	31	+ batch_size: 16
	32	+ seq_len: 512
	33	+ steps: 3000
	34	+ lr: 0.0003
	35	+ warmup: 100
	36	+ weight_decay: 0.1
	37	+ grad_clip: 1.0
	38	+ log_every: 20
	39	+ eval_every: 500
	40	+ eval_batches: 10
	41	+ ckpt_every: 1000
	42	+ final_exact_match: false
	43	+ name: editor_pilot_103m_lr3e4
	44	+ mode: prisme
	45	+model:
	46	+ name: prisme_103m
	47	+ vocab_size: 32768
	48	+ d_model: 768
	49	+ n_layers: 12
	50	+ n_heads: 12
	51	+ n_kv_heads: 3
	52	+ ffn_width: 2048
	53	+ max_seq_len: 4096
	54	+ rope_theta: 10000.0
	55	+ norm_eps: 1.0e-05
	56	+ tie_embeddings: true
	57	+ init_std: 0.02
	58	+ use_workspace: true
	59	+ w: 256
	60	+ ffn_w: 768
	61	+ n_slots: 16
	62	+ n_editor_blocks: 2
	63	+ n_editor_heads: 4
	64	+ r_max: 8
	65	+ eta_max: 0.25
	66	+ clip_c: 1.0
	67	+ clip_eps: 1.0e-06
	68	+ readout_zero_init: false
	69	+ readout_init_std: 0.2
	70	+ readout_scale: 1.0
	71	+ controller_hidden: 0
	72	+ ctrl_input: h
	73	+ ctrl_use_entropy: true
	74	+ lambda_cost: 0.0
	75	+ kappa_risk: 0.0
	76	+ dense_writes: false
	77	+resolved_train:
	78	+ batch_size: 16
	79	+ seq_len: 512
	80	+ steps: 3000
	81	+ lr: 0.0003
	82	+ warmup: 100
	83	+ weight_decay: 0.1
	84	+ grad_clip: 1.0
	85	+ log_every: 20
	86	+ eval_every: 500
	87	+ eval_batches: 10
	88	+ ckpt_every: 1000

added experiments/runs/editor_pilot_103m_lr3e4_seed01_20261030021200/manifest.json +28 −0

		@@ -0,0 +1,28 @@
	1	+{
	2	+ "run_id": "editor_pilot_103m_lr3e4_seed01_20261030021200",
	3	+ "name": "editor_pilot_103m_lr3e4",
	4	+ "seed": 1,
	5	+ "code_commit": "unknown",
	6	+ "config_sha256": "34555c00957d96128cbcc8a3db444e910de6724a1bcce91f80dbbc6e5faabb26",
	7	+ "started_at": "2026-10-30T02:12:01-0400",
	8	+ "host": "MacStudnpierre2",
	9	+ "python": "3.12.13",
	10	+ "torch": "2.15.0",
	11	+ "command": "/Users/simon-pierreboucher/wp31/prisme-slm/training/train_editor.py --config configs/experiments/editor_pilot_103m_lr3e4.yaml --device mps",
	12	+ "state": "completed",
	13	+ "wall_seconds": 5397.54,
	14	+ "device": "mps",
	15	+ "backend": "mps",
	16	+ "precision": "fp32",
	17	+ "n_params": 102265605,
	18	+ "n_trainable": 1907713,
	19	+ "checkpoint_parent": "experiments/runs/pretrain103m_seed01_20261029232706/checkpoint",
	20	+ "mode": "prisme",
	21	+ "freeze_backbone": true,
	22	+ "readout_scale": 8.851566314697266,
	23	+ "finished_at": "2026-10-30T03:41:58-0400",
	24	+ "train_tokens": 24576000,
	25	+ "peak_memory_bytes": 25939361792,
	26	+ "final_exact_match": {},
	27	+ "metrics_file": "experiments/runs/editor_pilot_103m_lr3e4_seed01_20261030021200/metrics.jsonl"
	28	+}
		\ No newline at end of file

added experiments/runs/editor_pilot_103m_lr3e4_seed01_20261030021200/metrics.jsonl +160 −0

		@@ -0,0 +1,160 @@
	1	+{"step": 0, "loss": 3.8388280868530273, "lr": 2.9999999999999997e-06, "grad_norm": 0.16530217230319977, "tokens": 8192, "tok_per_s": 5555.570741903939, "t": 2.33}
	2	+{"step": 20, "loss": 3.683028221130371, "lr": 6.299999999999999e-05, "grad_norm": 0.14000000059604645, "tokens": 172032, "tok_per_s": 5153.812067721846, "t": 34.23}
	3	+{"step": 40, "loss": 4.018809795379639, "lr": 0.00012299999999999998, "grad_norm": 0.12102383375167847, "tokens": 335872, "tok_per_s": 5114.816955004463, "t": 66.52}
	4	+{"step": 60, "loss": 3.8221116065979004, "lr": 0.00018299999999999998, "grad_norm": 0.09996253252029419, "tokens": 499712, "tok_per_s": 4678.3844258352765, "t": 107.67}
	5	+{"step": 80, "loss": 3.7520482540130615, "lr": 0.000243, "grad_norm": 0.08119877427816391, "tokens": 663552, "tok_per_s": 4566.69780285139, "t": 146.16}
	6	+{"step": 100, "loss": 3.5873820781707764, "lr": 0.0003, "grad_norm": 0.05057114362716675, "tokens": 827392, "tok_per_s": 4404.052739043535, "t": 188.72}
	7	+{"step": 120, "loss": 3.880525827407837, "lr": 0.00029996831523249534, "grad_norm": 0.02119096927344799, "tokens": 991232, "tok_per_s": 4411.498032822185, "t": 225.55}
	8	+{"step": 140, "loss": 3.784116744995117, "lr": 0.00029987327580293684, "grad_norm": 0.012864955700933933, "tokens": 1155072, "tok_per_s": 4424.064223711705, "t": 261.94}
	9	+{"step": 160, "loss": 3.988798141479492, "lr": 0.00029971492632320935, "grad_norm": 0.015187021344900131, "tokens": 1318912, "tok_per_s": 4491.147841157171, "t": 294.52}
	10	+{"step": 180, "loss": 3.9943411350250244, "lr": 0.00029949334112318624, "grad_norm": 0.012424705550074577, "tokens": 1482752, "tok_per_s": 4375.225308050471, "t": 339.75}
	11	+{"step": 200, "loss": 3.7738120555877686, "lr": 0.00029920862421583855, "grad_norm": 0.015423587523400784, "tokens": 1646592, "tok_per_s": 4316.237828757673, "t": 382.34}
	12	+{"step": 220, "loss": 3.86301851272583, "lr": 0.00029886090924841085, "grad_norm": 0.01458093523979187, "tokens": 1810432, "tok_per_s": 4443.7918714043135, "t": 408.26}
	13	+{"step": 240, "loss": 3.718639373779297, "lr": 0.00029845035943968696, "grad_norm": 0.010044100694358349, "tokens": 1974272, "tok_per_s": 4483.892496917172, "t": 441.16}
	14	+{"step": 260, "loss": 3.63783860206604, "lr": 0.000297977167503374, "grad_norm": 0.01396073866635561, "tokens": 2138112, "tok_per_s": 4543.494007181648, "t": 471.44}
	15	+{"step": 280, "loss": 3.697767972946167, "lr": 0.0002974415555576423, "grad_norm": 0.014589798636734486, "tokens": 2301952, "tok_per_s": 4600.971590123155, "t": 501.17}
	16	+{"step": 300, "loss": 3.991833209991455, "lr": 0.00029684377502086165, "grad_norm": 0.006617589388042688, "tokens": 2465792, "tok_per_s": 4608.253979077773, "t": 535.93}
	17	+{"step": 320, "loss": 3.776200771331787, "lr": 0.0002961841064935852, "grad_norm": 0.009673132561147213, "tokens": 2629632, "tok_per_s": 4661.962361198249, "t": 564.91}
	18	+{"step": 340, "loss": 3.804020404815674, "lr": 0.0002954628596268339, "grad_norm": 0.008413425646722317, "tokens": 2793472, "tok_per_s": 4621.8243908042205, "t": 605.26}
	19	+{"step": 360, "loss": 3.6039321422576904, "lr": 0.00029468037297674526, "grad_norm": 0.011924357153475285, "tokens": 2957312, "tok_per_s": 4647.944307062156, "t": 637.12}
	20	+{"step": 380, "loss": 3.828267812728882, "lr": 0.000293837013845654, "grad_norm": 0.012362201698124409, "tokens": 3121152, "tok_per_s": 4671.112004620043, "t": 669.04}
	21	+{"step": 400, "loss": 3.7054331302642822, "lr": 0.0002929331781096783, "grad_norm": 0.013202954083681107, "tokens": 3284992, "tok_per_s": 4690.541827956888, "t": 701.2}
	22	+{"step": 420, "loss": 3.8094773292541504, "lr": 0.0002919692900328947, "grad_norm": 0.013429978862404823, "tokens": 3448832, "tok_per_s": 4650.34476065565, "t": 742.48}
	23	+{"step": 440, "loss": 3.832529306411743, "lr": 0.00029094580206818636, "grad_norm": 0.015276624821126461, "tokens": 3612672, "tok_per_s": 4700.278107237236, "t": 769.46}
	24	+{"step": 460, "loss": 3.6833889484405518, "lr": 0.00028986319464486, "grad_norm": 0.010125985369086266, "tokens": 3776512, "tok_per_s": 4592.850904398649, "t": 823.11}
	25	+{"step": 480, "loss": 3.6737911701202393, "lr": 0.0002887219759431305, "grad_norm": 0.013986553996801376, "tokens": 3940352, "tok_per_s": 4586.490897097214, "t": 859.97}
	26	+{"step": 499, "val_loss_b0": 3.9889174699783325, "val_loss_b1": 3.9885329008102417, "val_loss_b2": 3.988248109817505, "val_loss_b4": 3.988071346282959, "val_loss_b8": 3.989164853096008, "t": 918.99}
	27	+{"step": 500, "loss": 3.585740566253662, "lr": 0.00028752268165557917, "grad_norm": 0.015741195529699326, "tokens": 4104192, "tok_per_s": 4462.758202491632, "t": 920.51}
	28	+{"step": 520, "loss": 3.7321407794952393, "lr": 0.00028626587473569746, "grad_norm": 0.009654651395976543, "tokens": 4268032, "tok_per_s": 4392.5144225691865, "t": 972.52}
	29	+{"step": 540, "loss": 3.675868511199951, "lr": 0.00028495214513363427, "grad_norm": 0.012084590271115303, "tokens": 4431872, "tok_per_s": 4384.299847761102, "t": 1011.7}
	30	+{"step": 560, "loss": 3.8682422637939453, "lr": 0.0002835821095192712, "grad_norm": 0.012979291379451752, "tokens": 4595712, "tok_per_s": 4417.367114475134, "t": 1041.23}
	31	+{"step": 580, "loss": 3.8059096336364746, "lr": 0.00028215641099275517, "grad_norm": 0.014583375304937363, "tokens": 4759552, "tok_per_s": 4438.0271035006435, "t": 1073.3}
	32	+{"step": 600, "loss": 3.8229777812957764, "lr": 0.00028067571878262454, "grad_norm": 0.017459910362958908, "tokens": 4923392, "tok_per_s": 4466.848177326197, "t": 1103.06}
	33	+{"step": 620, "loss": 3.6982314586639404, "lr": 0.00027914072793167077, "grad_norm": 0.015852632001042366, "tokens": 5087232, "tok_per_s": 4492.692840303049, "t": 1133.19}
	34	+{"step": 640, "loss": 3.801331043243408, "lr": 0.00027755215897068257, "grad_norm": 0.012572171166539192, "tokens": 5251072, "tok_per_s": 4504.295626018583, "t": 1166.65}
	35	+{"step": 660, "loss": 3.7758383750915527, "lr": 0.0002759107575802257, "grad_norm": 0.014323781244456768, "tokens": 5414912, "tok_per_s": 4523.230886119979, "t": 1197.99}
	36	+{"step": 680, "loss": 3.9480979442596436, "lr": 0.00027421729424061787, "grad_norm": 0.013456886634230614, "tokens": 5578752, "tok_per_s": 4509.814087492459, "t": 1237.88}
	37	+{"step": 700, "loss": 3.779712438583374, "lr": 0.00027247256387026185, "grad_norm": 0.012655014172196388, "tokens": 5742592, "tok_per_s": 4532.351850816981, "t": 1267.88}
	38	+{"step": 720, "loss": 3.822575569152832, "lr": 0.000270677385452508, "grad_norm": 0.014358540996909142, "tokens": 5906432, "tok_per_s": 4547.740722835746, "t": 1299.61}
	39	+{"step": 740, "loss": 3.7792210578918457, "lr": 0.0002688326016512199, "grad_norm": 0.01548258401453495, "tokens": 6070272, "tok_per_s": 4553.915504804082, "t": 1333.83}
	40	+{"step": 760, "loss": 3.806464672088623, "lr": 0.0002669390784152248, "grad_norm": 0.013577770441770554, "tokens": 6234112, "tok_per_s": 4543.631112666273, "t": 1372.91}
	41	+{"step": 780, "loss": 3.8274717330932617, "lr": 0.00026499770457183376, "grad_norm": 0.013904247432947159, "tokens": 6397952, "tok_per_s": 4503.13781769418, "t": 1421.63}
	42	+{"step": 800, "loss": 3.7019476890563965, "lr": 0.0002630093914096226, "grad_norm": 0.03281262144446373, "tokens": 6561792, "tok_per_s": 4502.10486864367, "t": 1458.35}
	43	+{"step": 820, "loss": 3.932710647583008, "lr": 0.0002609750722506695, "grad_norm": 0.0163898766040802, "tokens": 6725632, "tok_per_s": 4502.591980769655, "t": 1494.58}
	44	+{"step": 840, "loss": 3.8267409801483154, "lr": 0.00025889570201244967, "grad_norm": 0.015150759369134903, "tokens": 6889472, "tok_per_s": 4514.623144027639, "t": 1526.89}
	45	+{"step": 860, "loss": 3.798384189605713, "lr": 0.00025677225675959357, "grad_norm": 0.016737058758735657, "tokens": 7053312, "tok_per_s": 4510.831425955238, "t": 1564.49}
	46	+{"step": 880, "loss": 3.679598808288574, "lr": 0.0002546057332457181, "grad_norm": 0.0183124803006649, "tokens": 7217152, "tok_per_s": 4463.914533024111, "t": 1617.63}
	47	+{"step": 900, "loss": 3.790926456451416, "lr": 0.00025239714844554667, "grad_norm": 0.014279105700552464, "tokens": 7380992, "tok_per_s": 4490.537276585615, "t": 1644.53}
	48	+{"step": 920, "loss": 3.7000484466552734, "lr": 0.00025014753907753737, "grad_norm": 0.01636924408376217, "tokens": 7544832, "tok_per_s": 4499.325213367083, "t": 1677.73}
	49	+{"step": 940, "loss": 3.832568645477295, "lr": 0.0002478579611172425, "grad_norm": 0.015563327819108963, "tokens": 7708672, "tok_per_s": 4515.3561128980455, "t": 1708.07}
	50	+{"step": 960, "loss": 3.7648119926452637, "lr": 0.0002455294893016306, "grad_norm": 0.014933220110833645, "tokens": 7872512, "tok_per_s": 4532.2022562195925, "t": 1737.87}
	51	+{"step": 980, "loss": 3.812920093536377, "lr": 0.00024316321662459957, "grad_norm": 0.01687745936214924, "tokens": 8036352, "tok_per_s": 4546.233485207428, "t": 1768.55}
	52	+{"step": 999, "val_loss_b0": 3.9889174699783325, "val_loss_b1": 3.988364601135254, "val_loss_b2": 3.9879326105117796, "val_loss_b4": 3.987476348876953, "val_loss_b8": 3.9878592252731324, "t": 1836.47}
	53	+{"step": 999, "checkpoint_sha256": "d9484ed8cfaee2f2838a7cf1e817d290aa9a8e8ba1bd415bb2d6ab7bf774f7ae", "t": 1836.88}
	54	+{"step": 1000, "loss": 3.744023323059082, "lr": 0.0002407602538239216, "grad_norm": 0.01775466836988926, "tokens": 8200192, "tok_per_s": 4462.841607778254, "t": 1838.29}
	55	+{"step": 1020, "loss": 3.6456146240234375, "lr": 0.0002383217288598584, "grad_norm": 0.015043514780700207, "tokens": 8364032, "tok_per_s": 4473.138341334909, "t": 1870.69}
	56	+{"step": 1040, "loss": 3.8070361614227295, "lr": 0.00023584878638569206, "grad_norm": 0.017470480874180794, "tokens": 8527872, "tok_per_s": 4484.145318430601, "t": 1902.64}
	57	+{"step": 1060, "loss": 3.9012932777404785, "lr": 0.00023334258721042025, "grad_norm": 0.017150765284895897, "tokens": 8691712, "tok_per_s": 4489.887152572901, "t": 1936.69}
	58	+{"step": 1080, "loss": 3.713714599609375, "lr": 0.00023080430775386808, "grad_norm": 0.0152369299903512, "tokens": 8855552, "tok_per_s": 4483.223476622849, "t": 1976.12}
	59	+{"step": 1100, "loss": 3.8825254440307617, "lr": 0.00022823513949447166, "grad_norm": 0.01471775397658348, "tokens": 9019392, "tok_per_s": 4461.2083640179635, "t": 2022.59}
	60	+{"step": 1120, "loss": 3.8597702980041504, "lr": 0.00022563628840999367, "grad_norm": 0.020681019872426987, "tokens": 9183232, "tok_per_s": 4460.321539731326, "t": 2059.73}
	61	+{"step": 1140, "loss": 3.8525748252868652, "lr": 0.00022300897441143247, "grad_norm": 0.015108371153473854, "tokens": 9347072, "tok_per_s": 4464.914655275153, "t": 2094.3}
	62	+{"step": 1160, "loss": 3.6822214126586914, "lr": 0.00022035443077039114, "grad_norm": 0.021639063954353333, "tokens": 9510912, "tok_per_s": 4465.419732137452, "t": 2130.76}
	63	+{"step": 1180, "loss": 3.7807776927948, "lr": 0.00021767390354017485, "grad_norm": 0.015806611627340317, "tokens": 9674752, "tok_per_s": 4457.327421713042, "t": 2171.38}
	64	+{"step": 1200, "loss": 3.7414073944091797, "lr": 0.00021496865097088845, "grad_norm": 0.016911961138248444, "tokens": 9838592, "tok_per_s": 4427.218627377195, "t": 2223.15}
	65	+{"step": 1220, "loss": 3.6769042015075684, "lr": 0.00021223994291880895, "grad_norm": 0.02072557620704174, "tokens": 10002432, "tok_per_s": 4443.716994735264, "t": 2251.77}
	66	+{"step": 1240, "loss": 3.855714797973633, "lr": 0.00020948906025030994, "grad_norm": 0.015918290242552757, "tokens": 10166272, "tok_per_s": 4451.796183563373, "t": 2284.49}
	67	+{"step": 1260, "loss": 3.812335968017578, "lr": 0.00020671729424061788, "grad_norm": 0.016744321212172508, "tokens": 10330112, "tok_per_s": 4464.24219340855, "t": 2314.82}
	68	+{"step": 1280, "loss": 3.8199563026428223, "lr": 0.00020392594596768237, "grad_norm": 0.01638096384704113, "tokens": 10493952, "tok_per_s": 4477.386252961841, "t": 2344.62}
	69	+{"step": 1300, "loss": 3.883105516433716, "lr": 0.00020111632570144484, "grad_norm": 0.015294700860977173, "tokens": 10657792, "tok_per_s": 4488.659379768258, "t": 2375.24}
	70	+{"step": 1320, "loss": 3.7806808948516846, "lr": 0.00019828975228879288, "grad_norm": 0.01800672523677349, "tokens": 10821632, "tok_per_s": 4497.1409745956535, "t": 2407.19}
	71	+{"step": 1340, "loss": 3.943575859069824, "lr": 0.00019544755253448823, "grad_norm": 0.016109060496091843, "tokens": 10985472, "tok_per_s": 4488.721913806638, "t": 2448.2}
	72	+{"step": 1360, "loss": 3.810106039047241, "lr": 0.00019259106057835926, "grad_norm": 0.016049453988671303, "tokens": 11149312, "tok_per_s": 4496.089783517685, "t": 2480.63}
	73	+{"step": 1380, "loss": 3.831757068634033, "lr": 0.00018972161726905055, "grad_norm": 0.01839154213666916, "tokens": 11313152, "tok_per_s": 4507.67956947905, "t": 2510.6}
	74	+{"step": 1400, "loss": 3.8676252365112305, "lr": 0.00018684056953462326, "grad_norm": 0.014575948938727379, "tokens": 11476992, "tok_per_s": 4514.820483486525, "t": 2542.92}
	75	+{"step": 1420, "loss": 3.9963245391845703, "lr": 0.00018394926975030145, "grad_norm": 0.018497414886951447, "tokens": 11640832, "tok_per_s": 4508.077077213495, "t": 2583.07}
	76	+{"step": 1440, "loss": 3.851541757583618, "lr": 0.00018104907510366235, "grad_norm": 0.017899619415402412, "tokens": 11804672, "tok_per_s": 4518.3079856051445, "t": 2613.48}
	77	+{"step": 1460, "loss": 3.8959813117980957, "lr": 0.00017814134695756664, "grad_norm": 0.017883889377117157, "tokens": 11968512, "tok_per_s": 4490.001312349696, "t": 2666.45}
	78	+{"step": 1480, "loss": 3.6291708946228027, "lr": 0.00017522745021112944, "grad_norm": 0.01759996823966503, "tokens": 12132352, "tok_per_s": 4495.778668172852, "t": 2699.46}
	79	+{"step": 1499, "val_loss_b0": 3.9889174699783325, "val_loss_b1": 3.9881173849105833, "val_loss_b2": 3.987528657913208, "val_loss_b4": 3.9869588375091554, "val_loss_b8": 3.9877641439437865, "t": 2762.58}
	80	+{"step": 1500, "loss": 3.90535569190979, "lr": 0.00017230875265903138, "grad_norm": 0.02068295143544674, "tokens": 12296192, "tok_per_s": 4450.055997063979, "t": 2764.01}
	81	+{"step": 1520, "loss": 3.7332606315612793, "lr": 0.00016938662434947037, "grad_norm": 0.020495988428592682, "tokens": 12460032, "tok_per_s": 4447.8016000180505, "t": 2802.25}
	82	+{"step": 1540, "loss": 3.713611602783203, "lr": 0.0001664624369410554, "grad_norm": 0.023707536980509758, "tokens": 12623872, "tok_per_s": 4421.60151321603, "t": 2855.9}
	83	+{"step": 1560, "loss": 3.7561423778533936, "lr": 0.0001635375630589446, "grad_norm": 0.024348311126232147, "tokens": 12787712, "tok_per_s": 4437.488887229454, "t": 2882.6}
	84	+{"step": 1580, "loss": 3.717986583709717, "lr": 0.00016061337565052966, "grad_norm": 0.020553236827254295, "tokens": 12951552, "tok_per_s": 4443.512346135477, "t": 2915.56}
	85	+{"step": 1600, "loss": 4.001625061035156, "lr": 0.00015769124734096862, "grad_norm": 0.023473521694540977, "tokens": 13115392, "tok_per_s": 4453.408975739985, "t": 2945.88}
	86	+{"step": 1620, "loss": 3.767557144165039, "lr": 0.0001547725497888706, "grad_norm": 0.018059588968753815, "tokens": 13279232, "tok_per_s": 4463.92539012233, "t": 2975.64}
	87	+{"step": 1640, "loss": 3.778282880783081, "lr": 0.0001518586530424334, "grad_norm": 0.01947772316634655, "tokens": 13443072, "tok_per_s": 4472.883608072359, "t": 3006.31}
	88	+{"step": 1660, "loss": 3.89432430267334, "lr": 0.00014895092489633762, "grad_norm": 0.021466944366693497, "tokens": 13606912, "tok_per_s": 4479.628182152929, "t": 3038.36}
	89	+{"step": 1680, "loss": 3.732147216796875, "lr": 0.0001460507302496985, "grad_norm": 0.019685208797454834, "tokens": 13770752, "tok_per_s": 4473.219439602959, "t": 3079.34}
	90	+{"step": 1700, "loss": 3.5457446575164795, "lr": 0.00014315943046537674, "grad_norm": 0.021686526015400887, "tokens": 13934592, "tok_per_s": 4479.236540919507, "t": 3111.78}
	91	+{"step": 1720, "loss": 3.7694344520568848, "lr": 0.00014027838273094942, "grad_norm": 0.016095783561468124, "tokens": 14098432, "tok_per_s": 4488.6670721878445, "t": 3141.75}
	92	+{"step": 1740, "loss": 3.877885580062866, "lr": 0.00013740893942164073, "grad_norm": 0.02273871935904026, "tokens": 14262272, "tok_per_s": 4494.629376677128, "t": 3174.03}
	93	+{"step": 1760, "loss": 3.595947265625, "lr": 0.00013455244746551177, "grad_norm": 0.02010590210556984, "tokens": 14426112, "tok_per_s": 4496.748110435041, "t": 3208.97}
	94	+{"step": 1780, "loss": 3.834798574447632, "lr": 0.00013171024771120715, "grad_norm": 0.018496211618185043, "tokens": 14589952, "tok_per_s": 4497.2367868885385, "t": 3245.06}
	95	+{"step": 1800, "loss": 3.7569198608398438, "lr": 0.0001288836742985552, "grad_norm": 0.017011066898703575, "tokens": 14753792, "tok_per_s": 4480.462218534476, "t": 3293.77}
	96	+{"step": 1820, "loss": 3.9124388694763184, "lr": 0.00012607405403231763, "grad_norm": 0.01866631768643856, "tokens": 14917632, "tok_per_s": 4479.462464200471, "t": 3331.08}
	97	+{"step": 1840, "loss": 3.813666582107544, "lr": 0.00012328270575938212, "grad_norm": 0.021668385714292526, "tokens": 15081472, "tok_per_s": 4485.040926548531, "t": 3363.47}
	98	+{"step": 1860, "loss": 3.759260654449463, "lr": 0.00012051093974969008, "grad_norm": 0.02275138534605503, "tokens": 15245312, "tok_per_s": 4485.084879003477, "t": 3399.97}
	99	+{"step": 1880, "loss": 3.715264320373535, "lr": 0.00011776005708119105, "grad_norm": 0.019421957433223724, "tokens": 15409152, "tok_per_s": 4490.093760354635, "t": 3432.66}
	100	+{"step": 1900, "loss": 3.7311267852783203, "lr": 0.00011503134902911152, "grad_norm": 0.02526695653796196, "tokens": 15572992, "tok_per_s": 4479.445462219193, "t": 3477.4}
	101	+{"step": 1920, "loss": 3.9095945358276367, "lr": 0.00011232609645982515, "grad_norm": 0.024148497730493546, "tokens": 15736832, "tok_per_s": 4471.055353422084, "t": 3520.57}
	102	+{"step": 1940, "loss": 3.6274752616882324, "lr": 0.00010964556922960883, "grad_norm": 0.020813388749957085, "tokens": 15900672, "tok_per_s": 4482.81961524903, "t": 3547.88}
	103	+{"step": 1960, "loss": 3.9998321533203125, "lr": 0.00010699102558856754, "grad_norm": 0.02876809798181057, "tokens": 16064512, "tok_per_s": 4485.545783655364, "t": 3582.25}
	104	+{"step": 1980, "loss": 3.557504653930664, "lr": 0.00010436371159000636, "grad_norm": 0.019393693655729294, "tokens": 16228352, "tok_per_s": 4493.9490049562355, "t": 3612.01}
	105	+{"step": 1999, "val_loss_b0": 3.9889174699783325, "val_loss_b1": 3.9878397941589356, "val_loss_b2": 3.9870748281478883, "val_loss_b4": 3.986360287666321, "val_loss_b8": 3.98707230091095, "t": 3666.57}
	106	+{"step": 1999, "checkpoint_sha256": "61126d1f41c3d4e1c92ee1136a36663df20fa31da9ede0d2bd2fff0de7f35f58", "t": 3667.02}
	107	+{"step": 2000, "loss": 3.812525749206543, "lr": 0.00010176486050552833, "grad_norm": 0.0217677503824234, "tokens": 16392192, "tok_per_s": 4467.965500930371, "t": 3669.68}
	108	+{"step": 2020, "loss": 3.7536354064941406, "lr": 9.919569224613192e-05, "grad_norm": 0.024728665128350258, "tokens": 16556032, "tok_per_s": 4476.2916397518375, "t": 3699.46}
	109	+{"step": 2040, "loss": 3.8425941467285156, "lr": 9.66574127895797e-05, "grad_norm": 0.020093586295843124, "tokens": 16719872, "tok_per_s": 4470.964526245769, "t": 3740.51}
	110	+{"step": 2060, "loss": 3.7483816146850586, "lr": 9.415121361430792e-05, "grad_norm": 0.021629303693771362, "tokens": 16883712, "tok_per_s": 4475.814541392472, "t": 3773.06}
	111	+{"step": 2080, "loss": 3.8599843978881836, "lr": 9.167827114014155e-05, "grad_norm": 0.020756013691425323, "tokens": 17047552, "tok_per_s": 4483.60948095026, "t": 3803.05}
	112	+{"step": 2100, "loss": 3.764911651611328, "lr": 8.92397461760784e-05, "grad_norm": 0.02337750419974327, "tokens": 17211392, "tok_per_s": 4488.301240996396, "t": 3835.58}
	113	+{"step": 2120, "loss": 4.088828086853027, "lr": 8.683678337540044e-05, "grad_norm": 0.023548651486635208, "tokens": 17375232, "tok_per_s": 4480.20498913092, "t": 3879.08}
	114	+{"step": 2140, "loss": 3.833620548248291, "lr": 8.447051069836941e-05, "grad_norm": 0.0264931321144104, "tokens": 17539072, "tok_per_s": 4487.815032221786, "t": 3909.01}
	115	+{"step": 2160, "loss": 3.9101006984710693, "lr": 8.214203888275743e-05, "grad_norm": 0.02647393010556698, "tokens": 17702912, "tok_per_s": 4471.573481150476, "t": 3959.84}
	116	+{"step": 2180, "loss": 3.7312822341918945, "lr": 7.985246092246267e-05, "grad_norm": 0.02162903919816017, "tokens": 17866752, "tok_per_s": 4473.032397135901, "t": 3995.18}
	117	+{"step": 2200, "loss": 3.8372957706451416, "lr": 7.760285155445326e-05, "grad_norm": 0.02322702668607235, "tokens": 18030592, "tok_per_s": 4476.022609599988, "t": 4029.12}
	118	+{"step": 2220, "loss": 4.003998279571533, "lr": 7.53942667542819e-05, "grad_norm": 0.021999096497893333, "tokens": 18194432, "tok_per_s": 4475.795164253909, "t": 4065.93}
	119	+{"step": 2240, "loss": 3.845353841781616, "lr": 7.322774324040643e-05, "grad_norm": 0.02525237947702408, "tokens": 18358272, "tok_per_s": 4450.846319413955, "t": 4125.52}
	120	+{"step": 2260, "loss": 3.7047626972198486, "lr": 7.11042979875503e-05, "grad_norm": 0.022539757192134857, "tokens": 18522112, "tok_per_s": 4461.855021141051, "t": 4152.07}
	121	+{"step": 2280, "loss": 3.5579464435577393, "lr": 6.902492774933048e-05, "grad_norm": 0.02275800332427025, "tokens": 18685952, "tok_per_s": 4473.976878177679, "t": 4177.44}
	122	+{"step": 2300, "loss": 3.917072296142578, "lr": 6.699060859037737e-05, "grad_norm": 0.027450494468212128, "tokens": 18849792, "tok_per_s": 4475.6279215284585, "t": 4212.51}
	123	+{"step": 2320, "loss": 3.77860164642334, "lr": 6.500229542816624e-05, "grad_norm": 0.02635987289249897, "tokens": 19013632, "tok_per_s": 4483.243541117967, "t": 4241.9}
	124	+{"step": 2340, "loss": 3.8698670864105225, "lr": 6.306092158477523e-05, "grad_norm": 0.02103089541196823, "tokens": 19177472, "tok_per_s": 4490.565259161523, "t": 4271.47}
	125	+{"step": 2360, "loss": 3.538996696472168, "lr": 6.116739834878009e-05, "grad_norm": 0.027196239680051804, "tokens": 19341312, "tok_per_s": 4495.306164625844, "t": 4303.41}
	126	+{"step": 2380, "loss": 3.8050360679626465, "lr": 5.9322614547491995e-05, "grad_norm": 0.02367325685918331, "tokens": 19505152, "tok_per_s": 4500.932601813845, "t": 4334.43}
	127	+{"step": 2400, "loss": 3.9583663940429688, "lr": 5.752743612973809e-05, "grad_norm": 0.024019066244363785, "tokens": 19668992, "tok_per_s": 4495.994951515232, "t": 4375.63}
	128	+{"step": 2420, "loss": 3.797323226928711, "lr": 5.578270575938211e-05, "grad_norm": 0.022699503228068352, "tokens": 19832832, "tok_per_s": 4499.898642157829, "t": 4408.25}
	129	+{"step": 2440, "loss": 3.8363139629364014, "lr": 5.408924241977427e-05, "grad_norm": 0.026374451816082, "tokens": 19996672, "tok_per_s": 4506.4186044389835, "t": 4438.23}
	130	+{"step": 2460, "loss": 3.8300018310546875, "lr": 5.244784102931745e-05, "grad_norm": 0.024398434907197952, "tokens": 20160512, "tok_per_s": 4510.375946585783, "t": 4470.66}
	131	+{"step": 2480, "loss": 3.7476959228515625, "lr": 5.085927206832921e-05, "grad_norm": 0.04752561077475548, "tokens": 20324352, "tok_per_s": 4502.40608964813, "t": 4514.96}
	132	+{"step": 2499, "val_loss_b0": 3.9889174699783325, "val_loss_b1": 3.9876109838485716, "val_loss_b2": 3.9867547750473022, "val_loss_b4": 3.9860803127288817, "val_loss_b8": 3.9870404481887816, "t": 4585.73}
	133	+{"step": 2500, "loss": 3.910581588745117, "lr": 4.9324281217375474e-05, "grad_norm": 0.029107971116900444, "tokens": 20488192, "tok_per_s": 4466.885087320102, "t": 4587.54}
	134	+{"step": 2520, "loss": 3.935685157775879, "lr": 4.78435890072448e-05, "grad_norm": 0.02710035629570484, "tokens": 20652032, "tok_per_s": 4466.65847609691, "t": 4624.45}
	135	+{"step": 2540, "loss": 3.789839267730713, "lr": 4.641789048072876e-05, "grad_norm": 0.021377671509981155, "tokens": 20815872, "tok_per_s": 4470.856416233407, "t": 4656.76}
	136	+{"step": 2560, "loss": 3.72589111328125, "lr": 4.504785486636568e-05, "grad_norm": 0.026580851525068283, "tokens": 20979712, "tok_per_s": 4471.037158516677, "t": 4693.21}
	137	+{"step": 2580, "loss": 3.7323179244995117, "lr": 4.373412526430252e-05, "grad_norm": 0.023492466658353806, "tokens": 21143552, "tok_per_s": 4471.1057294862485, "t": 4729.78}
	138	+{"step": 2600, "loss": 3.9327902793884277, "lr": 4.247731834442082e-05, "grad_norm": 0.02387898974120617, "tokens": 21307392, "tok_per_s": 4452.449317383295, "t": 4786.4}
	139	+{"step": 2620, "loss": 3.8893706798553467, "lr": 4.1278024056869486e-05, "grad_norm": 0.022287946194410324, "tokens": 21471232, "tok_per_s": 4461.553316090243, "t": 4813.35}
	140	+{"step": 2640, "loss": 3.728363037109375, "lr": 4.013680535514001e-05, "grad_norm": 0.02374396100640297, "tokens": 21635072, "tok_per_s": 4472.0444321199775, "t": 4838.7}
	141	+{"step": 2660, "loss": 3.68619441986084, "lr": 3.905419793181364e-05, "grad_norm": 0.0230452548712492, "tokens": 21798912, "tok_per_s": 4473.988572940419, "t": 4873.22}
	142	+{"step": 2680, "loss": 3.5850837230682373, "lr": 3.803070996710526e-05, "grad_norm": 0.024549061432480812, "tokens": 21962752, "tok_per_s": 4480.588428195816, "t": 4902.61}
	143	+{"step": 2700, "loss": 3.84997820854187, "lr": 3.7066821890321684e-05, "grad_norm": 0.024593722075223923, "tokens": 22126592, "tok_per_s": 4486.424080994401, "t": 4932.75}
	144	+{"step": 2720, "loss": 3.8746180534362793, "lr": 3.6162986154346006e-05, "grad_norm": 0.026537152007222176, "tokens": 22290432, "tok_per_s": 4489.576178719199, "t": 4965.78}
	145	+{"step": 2740, "loss": 3.584238290786743, "lr": 3.531962702325468e-05, "grad_norm": 0.0271969772875309, "tokens": 22454272, "tok_per_s": 4495.755865530468, "t": 4995.4}
	146	+{"step": 2760, "loss": 3.762524127960205, "lr": 3.4537140373166085e-05, "grad_norm": 0.023005293682217598, "tokens": 22618112, "tok_per_s": 4501.789992555519, "t": 5025.1}
	147	+{"step": 2780, "loss": 3.71954345703125, "lr": 3.3815893506414745e-05, "grad_norm": 0.023993074893951416, "tokens": 22781952, "tok_per_s": 4507.663435271643, "t": 5054.9}
	148	+{"step": 2800, "loss": 3.8339943885803223, "lr": 3.3156224979138305e-05, "grad_norm": 0.024403758347034454, "tokens": 22945792, "tok_per_s": 4514.798911326871, "t": 5083.2}
	149	+{"step": 2820, "loss": 3.7308526039123535, "lr": 3.255844444235771e-05, "grad_norm": 0.022225676104426384, "tokens": 23109632, "tok_per_s": 4521.21005386317, "t": 5112.24}
	150	+{"step": 2840, "loss": 3.7879652976989746, "lr": 3.202283249662597e-05, "grad_norm": 0.023512132465839386, "tokens": 23273472, "tok_per_s": 4527.090194869152, "t": 5141.79}
	151	+{"step": 2860, "loss": 3.8382761478424072, "lr": 3.154964056031306e-05, "grad_norm": 0.024715004488825798, "tokens": 23437312, "tok_per_s": 4533.258005850404, "t": 5170.93}
	152	+{"step": 2880, "loss": 3.820117950439453, "lr": 3.113909075158913e-05, "grad_norm": 0.02976880595088005, "tokens": 23601152, "tok_per_s": 4540.142502210738, "t": 5199.18}
	153	+{"step": 2900, "loss": 3.740880250930786, "lr": 3.0791375784161455e-05, "grad_norm": 0.02943800948560238, "tokens": 23764992, "tok_per_s": 4545.708717545187, "t": 5228.86}
	154	+{"step": 2920, "loss": 3.7544302940368652, "lr": 3.0506658876813734e-05, "grad_norm": 0.026308447122573853, "tokens": 23928832, "tok_per_s": 4551.5054035111825, "t": 5258.2}
	155	+{"step": 2940, "loss": 3.797105312347412, "lr": 3.0285073676790606e-05, "grad_norm": 0.024099133908748627, "tokens": 24092672, "tok_per_s": 4557.647528247589, "t": 5287.06}
	156	+{"step": 2960, "loss": 3.6772618293762207, "lr": 3.0126724197063143e-05, "grad_norm": 0.024978075176477432, "tokens": 24256512, "tok_per_s": 4563.641464905187, "t": 5316.02}
	157	+{"step": 2980, "loss": 3.8443057537078857, "lr": 3.0031684767504643e-05, "grad_norm": 0.02494724839925766, "tokens": 24420352, "tok_per_s": 4568.869186978655, "t": 5345.8}
	158	+{"step": 2999, "loss": 3.670595645904541, "lr": 3.0000079215009746e-05, "grad_norm": 0.025034256279468536, "tokens": 24576000, "tok_per_s": 4574.156629459247, "t": 5373.65}
	159	+{"step": 2999, "val_loss_b0": 3.9889174699783325, "val_loss_b1": 3.9875057220458983, "val_loss_b2": 3.9866103887557984, "val_loss_b4": 3.985898995399475, "val_loss_b8": 3.986349606513977, "t": 5397.13}
	160	+{"step": 2999, "checkpoint_sha256": "1203ea2e786b258268ea56204e9676a6b900d41628d9fce32af5a01ab0181ea3", "t": 5397.54}

added experiments/runs/labels_103m_seed01_20261030034314/config.yaml +55 −0

		@@ -0,0 +1,55 @@
	1	+experiment:
	2	+ name: labels_103m
	3	+ seed: 1
	4	+ mode: prisme
	5	+ model: from_checkpoint
	6	+ checkpoint: experiments/runs/editor_pilot_103m_lr3e4_seed01_20261030021200/checkpoint
	7	+ use_policy: false
	8	+ data:
	9	+ kind: tokens
	10	+ collect:
	11	+ batch_size: 8
	12	+ seq_len: 512
	13	+ max_prefix: 3
	14	+ audit_every: 4
	15	+ partitions:
	16	+ train:
	17	+ bin: data/processed/fineweb_edu_300m/train.bin
	18	+ n_batches: 120
	19	+ stream_seed: 777
	20	+ calib:
	21	+ bin: data/processed/fineweb_edu_300m/val.bin
	22	+ n_batches: 30
	23	+ stream_seed: 778
	24	+model:
	25	+ name: prisme_103m
	26	+ vocab_size: 32768
	27	+ d_model: 768
	28	+ n_layers: 12
	29	+ n_heads: 12
	30	+ n_kv_heads: 3
	31	+ ffn_width: 2048
	32	+ max_seq_len: 4096
	33	+ rope_theta: 10000.0
	34	+ norm_eps: 1.0e-05
	35	+ tie_embeddings: true
	36	+ init_std: 0.02
	37	+ use_workspace: true
	38	+ w: 256
	39	+ ffn_w: 768
	40	+ n_slots: 16
	41	+ n_editor_blocks: 2
	42	+ n_editor_heads: 4
	43	+ r_max: 8
	44	+ eta_max: 0.25
	45	+ clip_c: 1.0
	46	+ clip_eps: 1.0e-06
	47	+ readout_zero_init: false
	48	+ readout_init_std: 0.2
	49	+ readout_scale: 8.851566314697266
	50	+ controller_hidden: 0
	51	+ ctrl_input: h
	52	+ ctrl_use_entropy: true
	53	+ lambda_cost: 0.0
	54	+ kappa_risk: 0.0
	55	+ dense_writes: false

added experiments/runs/labels_103m_seed01_20261030034314/manifest.json +40 −0

		@@ -0,0 +1,40 @@
	1	+{
	2	+ "run_id": "labels_103m_seed01_20261030034314",
	3	+ "name": "labels_103m",
	4	+ "seed": 1,
	5	+ "code_commit": "unknown",
	6	+ "config_sha256": "810649993595f24664a37f2c3db5b3a114b40b2dbbda7765502e421c383bcf6c",
	7	+ "started_at": "2026-10-30T03:43:14-0400",
	8	+ "host": "MacStudnpierre2",
	9	+ "python": "3.12.13",
	10	+ "torch": "2.15.0",
	11	+ "command": "/Users/simon-pierreboucher/wp31/prisme-slm/training/collect_interventions.py --config configs/experiments/labels_103m.yaml --device mps",
	12	+ "state": "completed",
	13	+ "wall_seconds": 94.09,
	14	+ "device": "mps",
	15	+ "checkpoint_parent": "experiments/runs/editor_pilot_103m_lr3e4_seed01_20261030021200/checkpoint",
	16	+ "checkpoint_sha256": "1203ea2e786b258268ea56204e9676a6b900d41628d9fce32af5a01ab0181ea3",
	17	+ "finished_at": "2026-10-30T03:44:48-0400",
	18	+ "partitions": {
	19	+ "train": {
	20	+ "file": "experiments/runs/labels_103m_seed01_20261030034314/labels_train.pt",
	21	+ "n_states": 491520,
	22	+ "n_labels": 1966080,
	23	+ "mean_gain": 0.0008011759491637349,
	24	+ "frac_positive": 0.531177282333374,
	25	+ "frac_harm": 0.4687851071357727,
	26	+ "seconds": 74.4,
	27	+ "n_audit_batches": 30
	28	+ },
	29	+ "calib": {
	30	+ "file": "experiments/runs/labels_103m_seed01_20261030034314/labels_calib.pt",
	31	+ "n_states": 122880,
	32	+ "n_labels": 491520,
	33	+ "mean_gain": 0.0007764770998619497,
	34	+ "frac_positive": 0.5318562984466553,
	35	+ "frac_harm": 0.4680948853492737,
	36	+ "seconds": 19.7,
	37	+ "n_audit_batches": 8
	38	+ }
	39	+ }
	40	+}
		\ No newline at end of file

added experiments/runs/labels_103m_seed01_20261030034314/metrics.jsonl +0 −0

modified paper/appendix/failures.tex +6 −0

		@@ -20,4 +20,10 @@
20	20	\textbf{date}: 2026-10-30; \textbf{what failed}: Killed to free nodes for the 5-seed matrix: the degraded joint-finetune reference (M4M64a) and the v1 dense-writes run on the slow m4mn node (superseded by the v2 matrix \texttt{p4v2\_dense}).; \textbf{evidence}: manifests remain in \texttt{experiments/runs/} on the nodes (state: running $\rightarrow$ killed)
21	21	\item
22	22	\textbf{date}: 2026-10-30; \textbf{what failed}: Ranking-loss controller (D12: pairwise + listwise on top of Huber/BCE/NLL, editor-unseen labels): calibration sign accuracy fell to 45 \%, captured fraction --8 \%; on dev-iid the locked policy ($\lambda$=0.3, $\kappa$=0.5) edits 0.07 tokens and matches budget 0 (0.4333 vs 0.4334); $\lambda$=0 is worse (+1.3 \%).; \textbf{evidence}: \texttt{experiments/runs/controller\_v1rank\_seed01\_*}, \texttt{results/raw/routing\_v1rank.json}; \textbf{resolution}: Within-state ranking is not learnable from these labels at this scale; H3 null stands.
	23	+\item
	24	+ \textbf{date}: 2026-10-30; \textbf{what failed}: Stage A ran 6,000 steps instead of the intended 12,000 (stale config on the node, see rsync exclusion bug above); the enlarged baseline ran the 12,000-step config until step \textasciitilde6,540 and was killed.; \textbf{evidence}: \texttt{experiments/runs/pretrain103m\_seed01\_20261029232706/manifest.json} (resolved\_train.steps = 6000), M3U96b log; \textbf{resolution}: Backbone budget fixed at 98.3M tokens (D13); baseline relaunched with the same 6,000-step recipe.
	25	+\item
	26	+ \textbf{date}: 2026-10-30; \textbf{what failed}: 103M editor pilot at lr 1e-3 collapsed to inert edits within 500 steps (val loss identical at every budget, 3.9889; trainable gradient norm 0.00 from step 160) --- the same failure as the tiny w=128 lr 3e-3 run.; \textbf{evidence}: M3U96a \texttt{experiments/logs/editor\_pilot\_103m\_seed1.log}, run \texttt{editor\_pilot\_103m\_seed01\_*} (killed at step \textasciitilde650); \textbf{resolution}: relaunched with lr 3e-4 (D14).
	27	+\item
	28	+ \textbf{date}: 2026-10-30; \textbf{what failed}: 103M dense-slot probe collapsed to inert edits at lr 1e-3 (step 280, gradient norm 0.00) and again at lr 3e-4 (val loss identical at every budget, 3.9394, at step 500). Killed; no 103M dense result.; \textbf{evidence}: M2U64 \texttt{experiments/logs/dense\_probe\_103m*\_seed1.log}; \textbf{resolution}: The dense ablation at 103M is reported as not obtained (collapse), with the tiny-scale 5-seed result as the only dense evidence.
23	29	\end{itemize}

modified paper/main.log +47 −25

		@@ -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 02:28
	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
2	2	entering extended mode
3	3	restricted \write18 enabled.
4	4	%&-line parsing enabled.
		@@ -1340,9 +1340,10 @@ Overfull \hbox (109.74734pt too wide) in paragraph at lines 8--15
1340	1340
1341	1341	) [19 <../results/figures/frontier_v2.pdf>] (./tables/stats.tex)
1342	1342	(./tables/latency.tex) [20 <../results/figures/labels_analysis.pdf>]
1343		−(./sections/results_pilot.tex)) (./sections/mechanistic.tex [21])
1344		−(./sections/limitations.tex [22]) (./sections/conclusion.tex)
1345		−(./sections/reproducibility.tex)
	1343	+(./sections/results_pilot.tex (./tables/pilot_ppl.tex)
	1344	+(./tables/pilot_routing.tex) [21] (./tables/pilot_latency.tex)))
	1345	+(./sections/mechanistic.tex [22]) (./sections/limitations.tex [23])
	1346	+(./sections/conclusion.tex) (./sections/reproducibility.tex)
1346	1347	Overfull \hbox (51.18217pt too wide) in paragraph at lines 2--101
1347	1348	\T1/ptm/m/n/12 (-20) num-ber in the manuscript is pro-duced by \T1/cmtt/m/n/12
1348	1349	scripts/make_tables.py \T1/ptm/m/n/12 (-20) and \T1/cmtt/m/n/12 scripts/reprodu
		@@ -1356,18 +1357,18 @@ cript_claims.py \T1/ptm/m/n/12 (-20) checks the chain. \T1/cmtt/m/n/12 REPRODUC
1356	1357	E.md
1357	1358	[]
1358	1359
1359		−[23] [24] (./main.bbl [25] [26]
	1360	+[24] [25] (./main.bbl [26] [27]
1360	1361	Underfull \hbox (badness 1297) in paragraph at lines 208--211
1361	1362	[]\T1/ptm/b/n/12 (+20) You, Yichen, Tianyu Fu, Aosong Feng, Xing-tai Lv, Xue-fe
1362	1363	i Ning, Ning Ding, and
1363	1364	[]
1364	1365
1365		−[27]) (./appendix/appendix.tex (./appendix/notation.tex
	1366	+[28]) (./appendix/appendix.tex (./appendix/notation.tex
1366	1367	Overfull \hbox (12.19864pt too wide) in paragraph at lines 2--16
1367	1368	[]
1368	1369	[]
1369	1370
1370		−) (./appendix/proofs.tex [28]) (./appendix/params.tex [29]
	1371	+) (./appendix/proofs.tex [29]) (./appendix/params.tex [30]
1371	1372	Overfull \hbox (92.32437pt too wide) detected at line 12
1372	1373	\OML/ztmcm/m/it/12 N[] \OT1/ztmcm/m/n/12 = \OML/ztmcm/m/it/12 d \OT1/ztmcm/m/n/
1373	1374	12 + 3\OML/ztmcm/m/it/12 dw \OT1/ztmcm/m/n/12 + \OML/ztmcm/m/it/12 Mw \OT1/ztmc
		@@ -1476,7 +1477,7 @@ LaTeX Warning: `h' float specifier changed to `ht'.
1476	1477	) (./appendix/prereg.tex
1477	1478	LaTeX Font Info: Font shape `T1/cmtt/bx/n' in size <12> not available
1478	1479	(Font) Font shape `T1/cmtt/m/n' tried instead on input line 1.
1479		− [30]
	1480	+ [31]
1480	1481	LaTeX Font Info: Trying to load font information for TS1+ptm on input line 8
1481	1482	.
1482	1483
		@@ -1489,7 +1490,7 @@ Overfull \hbox (80.09856pt too wide) in paragraph at lines 10--12
1489	1490	T1/cmtt/m/n/12 data/generators/synthetic.py\T1/ptm/m/n/12 (-20) ,
1490	1491	[]
1491	1492
1492		−[31]) (./appendix/runs.tex
	1493	+[32]) (./appendix/runs.tex
1493	1494	Overfull \hbox (121.91187pt too wide) in paragraph at lines 1--14
1494	1495	\T1/cmtt/m/n/12 training.pretrain\T1/ptm/m/n/12 (-20) , \T1/cmtt/m/n/12 trainin
1495	1496	g.train_editor\T1/ptm/m/n/12 (-20) , \T1/cmtt/m/n/12 training.collect_intervent
		@@ -1504,14 +1505,14 @@ tion.latency \T1/ptm/m/n/12 (-20) or \T1/cmtt/m/n/12 evaluation.run\T1/ptm/m/n/
1504	1505	12 (-20) ; launch scripts un-der \T1/cmtt/m/n/12 scripts/
1505	1506	[]
1506	1507
1507		−[32] (./appendix/runs_index.tex [33]
	1508	+[33] (./appendix/runs_index.tex [34]
1508	1509	Overfull \hbox (25.12868pt too wide) in paragraph at lines 52--54
1509	1510	[]\T1/ptm/b/n/10 (-20) run_id\T1/ptm/m/n/10 (-20) : p4_tiny_backbone_v2_seed01_
1510	1511	20261029233624; \T1/ptm/b/n/10 (-20) state\T1/ptm/m/n/10 (-20) : stopped_overfi
1511	1512	tting_step9300_checkpoint_step5000_used;
1512	1513	[]
1513	1514
1514		−[34])) (./appendix/results.tex (./tables/em_families.tex))
	1515	+[35])) (./appendix/results.tex (./tables/em_families.tex))
1515	1516	Overfull \hbox (168.4564pt too wide) in paragraph at lines 2--16
1516	1517	\T1/ptm/m/n/12 (-20) records (\T1/cmtt/m/n/12 experiments/runs/*/em_*.json\T1/p
1517	1518	tm/m/n/12 (-20) , \T1/cmtt/m/n/12 results/raw/items_*.json\T1/ptm/m/n/12 (-20)
		@@ -1549,7 +1550,7 @@ Overfull \hbox (47.44559pt too wide) in paragraph at lines 12--14
1549	1550	;
1550	1551	[]
1551	1552
1552		−[35]
	1553	+[36]
1553	1554	Package microtype Info: Loading generic protrusion settings for font family
1554	1555	(microtype) `cmtt' (encoding: TS1).
1555	1556	(microtype) For optimal results, create family-specific settings.
		@@ -1569,14 +1570,35 @@ run \T1/cmtt/m/n/12 p4_tiny_backbone_v2_seed01_* \T1/ptm/m/n/12 (-20) (state:
1569	1570	[]
1570	1571
1571	1572
1572		−Overfull \hbox (12.01894pt too wide) in paragraph at lines 22--23
	1573	+Overfull \hbox (12.01894pt too wide) in paragraph at lines 22--24
1573	1574	\T1/ptm/m/n/12 (-20) 0.4334); $\OML/ztmcm/m/it/12 ^^U$\T1/ptm/m/n/12 (-20) =0 i
1574	1575	s worse (+1.3 %).; \T1/ptm/b/n/12 (-20) ev-i-dence\T1/ptm/m/n/12 (-20) : \T1/cm
1575	1576	tt/m/n/12 experiments/runs/controller_v1rank_seed01_*\T1/ptm/m/n/12 (-20) ,
1576	1577	[]
1577	1578
1578		−)) [36{/usr/local/texlive/2026/texmf-dist/fonts/enc/dvips/cm-super/cm-super-ts1
1579		−.enc}] [37] [38] [39] [40] (./main.aux)
	1579	+
	1580	+Overfull \hbox (146.55865pt too wide) in paragraph at lines 24--26
	1581	+\T1/ptm/m/n/12 (-20) un-til step ~6,540 and was killed.; \T1/ptm/b/n/12 (-20) e
	1582	+v-i-dence\T1/ptm/m/n/12 (-20) : \T1/cmtt/m/n/12 experiments/runs/pretrain103m_s
	1583	+eed01_20261029232706/manifest.json
	1584	+ []
	1585	+
	1586	+
	1587	+Overfull \hbox (107.55685pt too wide) in paragraph at lines 26--28
	1588	+\T1/ptm/m/n/12 (-20) same fail-ure as the tiny w=128 lr 3e-3 run.; \T1/ptm/b/n/
	1589	+12 (-20) ev-i-dence\T1/ptm/m/n/12 (-20) : M3U96a \T1/cmtt/m/n/12 experiments/lo
	1590	+gs/editor_pilot_103m_seed1.log\T1/ptm/m/n/12 (-20) ,
	1591	+ []
	1592	+
	1593	+[37{/usr/local/texlive/2026/texmf-dist/fonts/enc/dvips/cm-super/cm-super-ts1.en
	1594	+c}]
	1595	+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	+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	+ []
	1600	+
	1601	+)) [38] [39] [40] [41] [42] [43] [44] (./main.aux)
1580	1602	***********
1581	1603	LaTeX2e <2025-11-01>
1582	1604	L3 programming layer <2026-01-19>
		@@ -1586,14 +1608,14 @@ L3 programming layer <2026-01-19>
1586	1608	LaTeX Warning: There were undefined references.
1587	1609
1588	1610	Package rerunfilecheck Info: File `main.out' has not changed.
1589		−(rerunfilecheck) Checksum: 445BE7B843FC16B6CF1F9C8E67681D49;9020.
	1611	+(rerunfilecheck) Checksum: 6517FE2FB3D993CA26B075D85418F650;9020.
1590	1612	)
1591	1613	Here is how much of TeX's memory you used:
1592		− 36075 strings out of 467525
1593		− 701562 string characters out of 5418982
1594		− 1130351 words of memory out of 5000000
1595		− 63396 multiletter control sequences out of 15000+600000
1596		− 793198 words of font info for 663 fonts, out of 8000000 for 9000
	1614	+ 36143 strings out of 467525
	1615	+ 702671 string characters out of 5418982
	1616	+ 1149081 words of memory out of 5000000
	1617	+ 63439 multiletter control sequences out of 15000+600000
	1618	+ 793579 words of font info for 669 fonts, out of 8000000 for 9000
1597	1619	1141 hyphenation exceptions out of 8191
1598	1620	84i,13n,104p,1001b,1501s stack positions out of 10000i,1000n,20000p,200000b,200000s
1599	1621	</usr/local/texlive/2026/texmf-dist/fonts/type1/public/amsfonts/cm/cmex10.pfb
		@@ -1611,10 +1633,10 @@ exlive/2026/texmf-dist/fonts/type1/urw/symbol/usyr.pfb></usr/local/texlive/2026
1611	1633	e1/urw/times/utmbi8a.pfb></usr/local/texlive/2026/texmf-dist/fonts/type1/urw/ti
1612	1634	mes/utmr8a.pfb></usr/local/texlive/2026/texmf-dist/fonts/type1/urw/times/utmri8
1613	1635	a.pfb>
1614		−Output written on main.pdf (40 pages, 434428 bytes).
	1636	+Output written on main.pdf (44 pages, 442504 bytes).
1615	1637	PDF statistics:
1616		− 785 PDF objects out of 1000 (max. 8388607)
1617		− 676 compressed objects within 7 object streams
1618		− 157 named destinations out of 1000 (max. 500000)
	1638	+ 812 PDF objects out of 1000 (max. 8388607)
	1639	+ 699 compressed objects within 7 object streams
	1640	+ 164 named destinations out of 1000 (max. 500000)
1619	1641	143777 words of extra memory for PDF output out of 154059 (max. 10000000)
1620	1642

modified paper/main.pdf +0 −0

Binary file not shown.

modified paper/sections/results_pilot.tex +39 −1

		@@ -1 +1,39 @@
1		−% filled after the 103M pilot completes
	1	+\paragraph{Stage A.} The 103M backbone (Table~\ref{tab:configs}) was pretrained on \pilotbbtokens{}M tokens of
	2	+FineWeb-Edu (6,000 steps of $32\times512$, cosine schedule, fp32) on one Apple M3 Ultra in \pilotbbhours{} hours
	3	+(13.8k tokens/s, peak \pilotbbpeakgb{} GB, $6NP\approx\pilotbbflops\times10^{16}$ FLOPs), reaching a validation
	4	+perplexity of \pilotbbppl{}. This is one token per parameter---far from compute-optimal and one-half of the
	5	+budget announced in the proposal's progression (a stale configuration on the node, Appendix~\ref{app:failures});
	6	+the parameter-matched plain decoder (feed-forward width 2144, 102.16M) was trained with the identical recipe
	7	+(validation perplexity \pilotenppl{}, \pilotenhours{} h). Only one backbone seed was affordable.
	8	+
	9	+\paragraph{Stage B.} On the frozen backbone, the sparse editor (\pilotedparams{} trainable parameters with the
	10	+residual-plus-entropy controller features, $s=\rms(H)=8.85$) and the global-recurrence adapter
	11	+(\pilotgrparams{}) were trained for 3,000 steps of $16\times512$ tokens with the forced-budget recipe. At the
	12	+learning rate that worked at the tiny scale ($10^{-3}$) the sparse editor collapsed to inert edits within 500
	13	+steps (every budget identical to budget 0, gradient norm zero); at $3\times10^{-4}$ it trained (\pilotedminutes{}
	14	+min, peak \pilotedpeakgb{} GB) but the effect on held-out web text is negligible: $-\pilotedgainbone$,
	15	+$-\pilotedgainbtwo$, $-\pilotedgainbfour$ and $-\pilotedgainbeight\,\%$ of loss at budgets 1, 2, 4 and 8. The
	16	+global adapter, which did not collapse at $10^{-3}$, gives $-\pilotgrgainbone$, $-\pilotgrgainbtwo$,
	17	+$-\pilotgrgainbfour$ and $-\pilotgrgainbeight\,\%$ in \pilotgrminutes{} min. The dense sixteen-slot probe
	18	+collapsed at both learning rates and yields no 103M result. On general web text, therefore, neither adapter
	19	+moves a frozen 100M backbone by more than a tenth of a percent of loss with 24.6M adapter tokens---less than the
	20	+run-to-run noise of pretraining itself---so H1 is not supported outside the synthetic suite, and the
	21	+parameter-matched plain decoder is the relevant comparison (Table~\ref{tab:pilotppl}).
	22	+\input{tables/pilot_ppl.tex}
	23	+
	24	+\paragraph{Stages C--D and routing.} Counterfactual labels were collected on a validation-disjoint stream of
	25	+web text and a controller fitted with the tiny-scale recipe (Table~\ref{tab:pilotrouting}). The locked
	26	+controller changes the validation loss by \pilotroutingctrl\,\% (\pilotroutingctrledits{} edits per token);
	27	+the one-step oracle reaches \pilotroutingoracle\,\% with \pilotroutingoracleedits{} edits per token, and the
	28	+best fixed schedule (round-robin, budget 1) \pilotroutinguniformone\,\%. The picture of
	29	+Section~\ref{sec:results-routing} is unchanged at 103M: there is oracle opportunity (now on natural text), and
	30	+the controller does not capture it.
	31	+\input{tables/pilot_routing.tex}
	32	+
	33	+\paragraph{Latency at 103M.} Table~\ref{tab:pilotlatency} repeats the latency protocol with the 103M model on
	34	+the M3 Ultra (prompt 256, 32 generated tokens, fp32). Budget 0 costs \pilotlatbzero{} ms per token at batch 1;
	35	+eight single-slot edits \pilotlatseight{} ms; eight dense writes \pilotlatdeight{} ms; the controller with its
	36	+locked penalty \pilotlatctrl{} ms. The sparse-versus-dense gap is now visible but remains far below the
	37	+analytic ratio (3.1$\times$ more editor operations for dense writes), confirming that at this size the
	38	+per-action dispatch, the slot re-keying and the bank read---not the lane arithmetic---set the cost.
	39	+\input{tables/pilot_latency.tex}

modified paper/tables/numbers.tex +31 −0

		@@ -38,6 +38,37 @@
38	38	\newcommand{\paramsMediumWs}{10,505,221}
39	39	\newcommand{\paramsSmallTotal}{102,134,021}
40	40	\newcommand{\paramsSmallWs}{2,631,173}
	41	+\newcommand{\pilotbbflops}{5.9}
	42	+\newcommand{\pilotbbhours}{2.3}
	43	+\newcommand{\pilotbbpeakgb}{28}
	44	+\newcommand{\pilotbbppl}{48.2}
	45	+\newcommand{\pilotbbtokens}{98.3}
	46	+\newcommand{\pilotedgainbeight}{0.06}
	47	+\newcommand{\pilotedgainbfour}{0.08}
	48	+\newcommand{\pilotedgainbone}{0.04}
	49	+\newcommand{\pilotedgainbtwo}{0.06}
	50	+\newcommand{\pilotedminutes}{90}
	51	+\newcommand{\pilotedparams}{1,907,713}
	52	+\newcommand{\pilotedpeakgb}{25.9}
	53	+\newcommand{\pilotenhours}{n/a}
	54	+\newcommand{\pilotenppl}{n/a}
	55	+\newcommand{\pilotgrgainbeight}{0.07}
	56	+\newcommand{\pilotgrgainbfour}{0.08}
	57	+\newcommand{\pilotgrgainbone}{0.04}
	58	+\newcommand{\pilotgrgainbtwo}{0.06}
	59	+\newcommand{\pilotgrminutes}{21}
	60	+\newcommand{\pilotgrparams}{2,623,488}
	61	+\newcommand{\pilotgrpeakgb}{9.7}
	62	+\newcommand{\pilotlatbzero}{7.6}
	63	+\newcommand{\pilotlatctrl}{11.0}
	64	+\newcommand{\pilotlatdeight}{26.9}
	65	+\newcommand{\pilotlatseight}{22.9}
	66	+\newcommand{\pilotpplgrm}{48.16}
	67	+\newcommand{\pilotroutingctrl}{-0.01}
	68	+\newcommand{\pilotroutingctrledits}{0.61}
	69	+\newcommand{\pilotroutingoracle}{-3.6}
	70	+\newcommand{\pilotroutingoracleedits}{4.28}
	71	+\newcommand{\pilotroutinguniformone}{-0.02}
41	72	\newcommand{\routingbzero}{0.4334}
42	73	\newcommand{\routingctrllambda}{0.3}
43	74	\newcommand{\routingctrlledits}{0.22}

added paper/tables/pilot_latency.tex +33 −0

		@@ -0,0 +1,33 @@
	1	+\begin{table}[t]\centering\small
	2	+\caption{103M pilot: measured decode latency per generated token on mps (fp32, prompt 256 tokens, 32 generated, three
	3	+repetitions, median and p95), batch 1 and 8.}
	4	+\label{tab:pilotlatency}
	5	+\begin{tabular}{llrrrr}\toprule
	6	+Policy & budget & batch & median ms & p95 ms & cache bytes\\\midrule
	7	+budget0 & 0 & 1 & 7.55 & 8.27 & 5,898,240\\
	8	+forced single & 1 & 1 & 10.20 & 11.00 & 5,898,240\\
	9	+forced pair & 1 & 1 & 10.23 & 10.95 & 5,898,240\\
	10	+forced single & 2 & 1 & 12.38 & 12.70 & 5,898,240\\
	11	+forced pair & 2 & 1 & 11.51 & 12.13 & 5,898,240\\
	12	+forced single & 4 & 1 & 15.87 & 54.16 & 5,898,240\\
	13	+forced pair & 4 & 1 & 15.46 & 27.41 & 5,898,240\\
	14	+forced single & 8 & 1 & 22.88 & 80.61 & 5,898,240\\
	15	+forced pair & 8 & 1 & 23.29 & 24.80 & 5,898,240\\
	16	+controller & 8 & 1 & 11.01 & 27.54 & 5,898,240\\
	17	+dense forced & 1 & 1 & 10.65 & 17.74 & 5,898,240\\
	18	+dense forced & 4 & 1 & 17.85 & 65.12 & 5,898,240\\
	19	+dense forced & 8 & 1 & 26.91 & 27.86 & 5,898,240\\
	20	+budget0 & 0 & 8 & 10.27 & 18.75 & 47,185,920\\
	21	+forced single & 1 & 8 & 13.08 & 23.51 & 47,185,920\\
	22	+forced pair & 1 & 8 & 13.04 & 57.55 & 47,185,920\\
	23	+forced single & 2 & 8 & 15.26 & 16.93 & 47,185,920\\
	24	+forced pair & 2 & 8 & 15.20 & 16.37 & 47,185,920\\
	25	+forced single & 4 & 8 & 18.81 & 20.15 & 47,185,920\\
	26	+forced pair & 4 & 8 & 20.04 & 21.58 & 47,185,920\\
	27	+forced single & 8 & 8 & 28.97 & 29.76 & 47,185,920\\
	28	+forced pair & 8 & 8 & 29.06 & 33.15 & 47,185,920\\
	29	+controller & 8 & 8 & 32.67 & 77.22 & 47,185,920\\
	30	+dense forced & 1 & 8 & 13.03 & 26.24 & 47,185,920\\
	31	+dense forced & 4 & 8 & 21.36 & 88.63 & 47,185,920\\
	32	+dense forced & 8 & 8 & 32.10 & 34.51 & 47,185,920\\
	33	+\bottomrule\end{tabular}\end{table}

added paper/tables/pilot_ppl.tex +12 −0

		@@ -0,0 +1,12 @@
	1	+\begin{table}[t]\centering\small
	2	+\caption{103M pilot: held-out cross-entropy (nats/token) and perplexity on 327,680 FineWeb-Edu validation tokens
	3	+(40 $\times$ 16 $\times$ 512) under forced random budgets and, for PRISME, under the locked controller.}
	4	+\label{tab:pilotppl}
	5	+\begin{tabular}{llrr}\toprule
	6	+System & policy & loss & ppl\\\midrule
	7	+gr 103m & forced b0 & 3.8746 & 48.16\\
	8	+gr 103m & forced b1 & 3.8731 & 48.09\\
	9	+gr 103m & forced b2 & 3.8722 & 48.05\\
	10	+gr 103m & forced b4 & 3.8714 & 48.01\\
	11	+gr 103m & forced b8 & 3.8718 & 48.03\\
	12	+\bottomrule\end{tabular}\end{table}

added paper/tables/pilot_routing.tex +34 −0

		@@ -0,0 +1,34 @@
	1	+\begin{table}[t]\centering\small
	2	+\caption{103M pilot: routing comparison on validation text (20 $\times$ 8 $\times$ 512 tokens). Oracles are diagnostic.}
	3	+\label{tab:pilotrouting}
	4	+\begin{tabular}{llrrr}\toprule
	5	+Policy & setting & loss & edits/token & extra ops\\\midrule
	6	+budget0 & & 3.9889 & 0.00 & 0.000\\
	7	+random forced & budget=1 & 3.9875 & 1.00 & 0.091\\
	8	+uniform & budget=1 & 3.9881 & 1.00 & 0.081\\
	9	+fixed pair & budget=1 & 3.9875 & 1.00 & 0.102\\
	10	+random forced & budget=2 & 3.9866 & 2.00 & 0.183\\
	11	+uniform & budget=2 & 3.9875 & 2.00 & 0.162\\
	12	+fixed pair & budget=2 & 3.9866 & 2.00 & 0.203\\
	13	+random forced & budget=4 & 3.9859 & 4.00 & 0.365\\
	14	+uniform & budget=4 & 3.9865 & 4.00 & 0.324\\
	15	+fixed pair & budget=4 & 3.9858 & 4.00 & 0.406\\
	16	+random forced & budget=8 & 3.9864 & 8.00 & 0.731\\
	17	+uniform & budget=8 & 3.9858 & 8.00 & 0.648\\
	18	+fixed pair & budget=8 & 3.9887 & 8.00 & 0.813\\
	19	+entropy & threshold=1.0, budget=8 & 3.9861 & 7.41 & 0.652\\
	20	+entropy & threshold=2.0, budget=8 & 3.9865 & 6.78 & 0.597\\
	21	+entropy & threshold=4.0, budget=8 & 3.9870 & 4.45 & 0.391\\
	22	+norm & threshold=0.05, budget=8 & 3.9872 & 1.65 & 0.145\\
	23	+norm & threshold=0.1, budget=8 & 3.9876 & 1.00 & 0.088\\
	24	+norm & threshold=0.2, budget=8 & 3.9876 & 1.00 & 0.088\\
	25	+controller locked & lam=0.3, kap=0.5, budget=8 & 3.9886 & 0.61 & 0.050\\
	26	+controller & lam=0.0, kap=0.5, budget=8 & 3.9879 & 3.04 & 0.266\\
	27	+controller & lam=0.03, kap=0.5, budget=8 & 3.9879 & 2.69 & 0.228\\
	28	+controller & lam=0.1, kap=0.5, budget=8 & 3.9881 & 1.98 & 0.161\\
	29	+controller & lam=0.3, kap=0.5, budget=8 & 3.9886 & 0.61 & 0.050\\
	30	+oracle1 greedy (diag.) & budget=1 & 3.8905 & 0.57 & 0.057\\
	31	+oracle1 greedy (diag.) & budget=2 & 3.8828 & 1.12 & 0.113\\
	32	+oracle1 greedy (diag.) & budget=4 & 3.8687 & 2.21 & 0.222\\
	33	+oracle1 greedy (diag.) & budget=8 & 3.8445 & 4.28 & 0.429\\
	34	+\bottomrule\end{tabular}\end{table}

modified results/claim_registry.csv +6 −0

		@@ -14,3 +14,9 @@ Fixed schedules: random = uniform = fixed pair (H2 null),loss at budget 1,see nu
14	14	Two-step oracle vs one-step: myopia regret small,gain nats,see numbers.tex oracletwostep;oracleonestep,synthetic dev-iid v1 editor (1 batch),controller_v1ctrl_seed01,results/raw/routing_v1ctrl.json,evaluation/routing.py,complete
15	15	Dispatch overhead dominates at tiny scale; dense not slower than sparse,median ms/token batch 1,see paper/tables/latency.tex,laptop MPS tiny model,p4_editor_v1_Ustd02_seed01,results/raw/latency_p4_editor_v1_Ustd02_seed01_20261029235050_mps.json,evaluation/latency.py,complete
16	16	Miniature end-to-end pipeline runs on CPU,wall seconds,56,mini configs,mini_backbone_seed01;mini_editor_seed01;labels_mini_seed01;controller_mini_seed01,results/raw/routing_mini.json,scripts/reproduce_pilot.sh,complete
	17	+103M backbone pretrained on 98.3M tokens (ppl 48.2),val loss/ppl,3.8746 / 48.16,FineWeb-Edu val,pretrain103m_seed01_20261029232706,experiments/runs/pretrain103m_seed01_20261029232706/manifest.json,training/pretrain.py,complete
	18	+103M sparse editor (lr 3e-4) changes held-out loss by less than 0.1% at any budget,val loss relative change,see numbers.tex piloted*,FineWeb-Edu val (training-time eval),editor_pilot_103m_lr3e4_seed01_20261030021200,experiments/runs/editor_pilot_103m_lr3e4_seed01_20261030021200/metrics.jsonl,training/train_editor.py,complete
	19	+103M sparse editor at lr 1e-3 collapsed to inert edits,val loss by budget,identical 3.9889,FineWeb-Edu val,editor_pilot_103m_seed01_20261030015139,experiments/runs/editor_pilot_103m_seed01_20261030015139/metrics.jsonl,,stopped-run-used
	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
	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

modified results/figures/frontier_v2.pdf +0 −0

Binary file not shown.

modified results/figures/labels_analysis.pdf +0 −0

Binary file not shown.

added results/figures/latency_103m_mps.pdf +0 −0

Binary file not shown.

added results/figures/latency_103m_mps.png +0 −0

Binary file not shown.

modified results/figures/latency_p4_editor_v1_Ustd02_seed01_20261029235050_mps.pdf +0 −0

Binary file not shown.

modified results/figures/learnability_bootstrap.pdf +0 −0

Binary file not shown.

modified results/figures/learnability_v2.pdf +0 −0

Binary file not shown.

modified results/figures/pretrain_curves.pdf +0 −0

Binary file not shown.

added results/figures/routing_103m_val.pdf +0 −0

Binary file not shown.

added results/figures/routing_103m_val.png +0 −0

Binary file not shown.

modified results/figures/routing_controller_p4_hent_seed01_20261030002015_dev_iid.pdf +0 −0

Binary file not shown.

modified results/figures/routing_mini.pdf +0 −0

Binary file not shown.

modified results/figures/routing_v1ctrl.pdf +0 −0

Binary file not shown.

modified results/figures/routing_v1rank.pdf +0 −0

Binary file not shown.

modified results/figures/routing_v2s1.pdf +0 −0

Binary file not shown.

modified results/figures/routing_v2s2.pdf +0 −0

Binary file not shown.

modified results/figures/routing_v2s3.pdf +0 −0

Binary file not shown.

modified results/figures/routing_v2s4.pdf +0 −0

Binary file not shown.

modified results/figures/routing_v2s5.pdf +0 −0

Binary file not shown.

added results/raw/eval_gr_103m_val.json +40 −0

		@@ -0,0 +1,40 @@
	1	+{
	2	+ "system": "gr_103m",
	3	+ "split": "val",
	4	+ "checkpoint": "experiments/runs/gr_pilot_103m_seed01_20261030015117/checkpoint",
	5	+ "checkpoint_sha256": "1e3e5cdd103672376ad23bbfaafc50bcc68b937d808fae74308285a54564938a",
	6	+ "config_sha256": "a3f0c8dc75a01e38b6eb8a9d199aa2e29f574b887bdd0e62092e2f550e7a6373",
	7	+ "code_commit": "unknown",
	8	+ "device": "mps",
	9	+ "precision": "fp32",
	10	+ "evaluated_at": "2026-10-30T03:43:14-0400",
	11	+ "data_sha256": "d8753bdfa9609fce56a7ec2da7c108666185d021e3f921721e4d2241e1d4fe54",
	12	+ "perplexity": {
	13	+ "forced_b0": {
	14	+ "loss": 3.8746298491954803,
	15	+ "ppl": 48.16486672694994,
	16	+ "n_tokens": 327680
	17	+ },
	18	+ "forced_b1": {
	19	+ "loss": 3.873138165473938,
	20	+ "ppl": 48.09307353898038,
	21	+ "n_tokens": 327680
	22	+ },
	23	+ "forced_b2": {
	24	+ "loss": 3.8722472012043,
	25	+ "ppl": 48.050243411729866,
	26	+ "n_tokens": 327680
	27	+ },
	28	+ "forced_b4": {
	29	+ "loss": 3.8714304685592653,
	30	+ "ppl": 48.01101523098215,
	31	+ "n_tokens": 327680
	32	+ },
	33	+ "forced_b8": {
	34	+ "loss": 3.871782386302948,
	35	+ "ppl": 48.02791413247139,
	36	+ "n_tokens": 327680
	37	+ }
	38	+ },
	39	+ "seconds": 45.8
	40	+}
		\ No newline at end of file

added results/raw/latency_103m_mps.json +302 −0

		@@ -0,0 +1,302 @@
	1	+{
	2	+ "config": "configs/experiments/latency_103m.yaml",
	3	+ "checkpoint_sha256": "64d08e9120608983b261bd6d108eb2421da45c84ebc733fa5fb2b88780dae427",
	4	+ "device": "mps",
	5	+ "torch": "2.15.0",
	6	+ "precision": "fp32",
	7	+ "prompt_len": 256,
	8	+ "n_new": 32,
	9	+ "code_commit": "unknown",
	10	+ "peak_memory_bytes": 3273342976,
	11	+ "analytic_ops_per_token": {
	12	+ "budget0": 209622528,
	13	+ "4_k1": 339414528,
	14	+ "4_k2": 355540480,
	15	+ "8_k2": 435656192,
	16	+ "8_dense": 887182848
	17	+ },
	18	+ "records": [
	19	+ {
	20	+ "policy": "budget0",
	21	+ "batch": 1,
	22	+ "budget": 0,
	23	+ "median_ms": 7.554021023679525,
	24	+ "p95_ms": 8.273145940620452,
	25	+ "mean_ms": 7.569623681774829,
	26	+ "n": 96,
	27	+ "cache_bytes": 5898240
	28	+ },
	29	+ {
	30	+ "policy": "forced_single",
	31	+ "batch": 1,
	32	+ "budget": 1,
	33	+ "forced_k": 1,
	34	+ "median_ms": 10.197187482845038,
	35	+ "p95_ms": 10.996218799846247,
	36	+ "mean_ms": 10.16590669072078,
	37	+ "n": 96,
	38	+ "cache_bytes": 5898240
	39	+ },
	40	+ {
	41	+ "policy": "forced_pair",
	42	+ "batch": 1,
	43	+ "budget": 1,
	44	+ "forced_k": 2,
	45	+ "median_ms": 10.232354514300823,
	46	+ "p95_ms": 10.949291929136962,
	47	+ "mean_ms": 10.210288580613755,
	48	+ "n": 96,
	49	+ "cache_bytes": 5898240
	50	+ },
	51	+ {
	52	+ "policy": "forced_single",
	53	+ "batch": 1,
	54	+ "budget": 2,
	55	+ "forced_k": 1,
	56	+ "median_ms": 12.383270543068647,
	57	+ "p95_ms": 12.69960420904681,
	58	+ "mean_ms": 11.684006077606076,
	59	+ "n": 96,
	60	+ "cache_bytes": 5898240
	61	+ },
	62	+ {
	63	+ "policy": "forced_pair",
	64	+ "batch": 1,
	65	+ "budget": 2,
	66	+ "forced_k": 2,
	67	+ "median_ms": 11.509437521453947,
	68	+ "p95_ms": 12.131708004744723,
	69	+ "mean_ms": 11.05400784460168,
	70	+ "n": 96,
	71	+ "cache_bytes": 5898240
	72	+ },
	73	+ {
	74	+ "policy": "forced_single",
	75	+ "batch": 1,
	76	+ "budget": 4,
	77	+ "forced_k": 1,
	78	+ "median_ms": 15.866020519752055,
	79	+ "p95_ms": 54.16153179248795,
	80	+ "mean_ms": 19.282480483525433,
	81	+ "n": 96,
	82	+ "cache_bytes": 5898240
	83	+ },
	84	+ {
	85	+ "policy": "forced_pair",
	86	+ "batch": 1,
	87	+ "budget": 4,
	88	+ "forced_k": 2,
	89	+ "median_ms": 15.460958005860448,
	90	+ "p95_ms": 27.407583547756076,
	91	+ "mean_ms": 16.31527776650425,
	92	+ "n": 96,
	93	+ "cache_bytes": 5898240
	94	+ },
	95	+ {
	96	+ "policy": "forced_single",
	97	+ "batch": 1,
	98	+ "budget": 8,
	99	+ "forced_k": 1,
	100	+ "median_ms": 22.882541525177658,
	101	+ "p95_ms": 80.60674997977912,
	102	+ "mean_ms": 28.589312497691328,
	103	+ "n": 96,
	104	+ "cache_bytes": 5898240
	105	+ },
	106	+ {
	107	+ "policy": "forced_pair",
	108	+ "batch": 1,
	109	+ "budget": 8,
	110	+ "forced_k": 2,
	111	+ "median_ms": 23.291583522222936,
	112	+ "p95_ms": 24.80255250702612,
	113	+ "mean_ms": 23.393341953730367,
	114	+ "n": 96,
	115	+ "cache_bytes": 5898240
	116	+ },
	117	+ {
	118	+ "policy": "controller",
	119	+ "batch": 1,
	120	+ "budget": 8,
	121	+ "median_ms": 11.006999993696809,
	122	+ "p95_ms": 27.535593806533143,
	123	+ "mean_ms": 12.026253437700992,
	124	+ "n": 96,
	125	+ "cache_bytes": 5898240
	126	+ },
	127	+ {
	128	+ "policy": "dense_forced",
	129	+ "batch": 1,
	130	+ "budget": 1,
	131	+ "forced_k": 1,
	132	+ "median_ms": 10.64931252039969,
	133	+ "p95_ms": 17.744635028066114,
	134	+ "mean_ms": 11.291276908499034,
	135	+ "n": 96,
	136	+ "cache_bytes": 5898240
	137	+ },
	138	+ {
	139	+ "policy": "dense_forced",
	140	+ "batch": 1,
	141	+ "budget": 4,
	142	+ "forced_k": 1,
	143	+ "median_ms": 17.851791460998356,
	144	+ "p95_ms": 65.12488526641391,
	145	+ "mean_ms": 22.8273315927557,
	146	+ "n": 96,
	147	+ "cache_bytes": 5898240
	148	+ },
	149	+ {
	150	+ "policy": "dense_forced",
	151	+ "batch": 1,
	152	+ "budget": 8,
	153	+ "forced_k": 1,
	154	+ "median_ms": 26.906583516392857,
	155	+ "p95_ms": 27.862822695169598,
	156	+ "mean_ms": 27.034480945682542,
	157	+ "n": 96,
	158	+ "cache_bytes": 5898240
	159	+ },
	160	+ {
	161	+ "policy": "budget0",
	162	+ "batch": 8,
	163	+ "budget": 0,
	164	+ "median_ms": 10.26502053719014,
	165	+ "p95_ms": 18.752072588540614,
	166	+ "mean_ms": 10.654087217214206,
	167	+ "n": 96,
	168	+ "cache_bytes": 47185920
	169	+ },
	170	+ {
	171	+ "policy": "forced_single",
	172	+ "batch": 8,
	173	+ "budget": 1,
	174	+ "forced_k": 1,
	175	+ "median_ms": 13.082749966997653,
	176	+ "p95_ms": 23.506614292273298,
	177	+ "mean_ms": 13.859784298498804,
	178	+ "n": 96,
	179	+ "cache_bytes": 47185920
	180	+ },
	181	+ {
	182	+ "policy": "forced_pair",
	183	+ "batch": 8,
	184	+ "budget": 1,
	185	+ "forced_k": 2,
	186	+ "median_ms": 13.039500452578068,
	187	+ "p95_ms": 57.54836479900405,
	188	+ "mean_ms": 17.062923668466585,
	189	+ "n": 96,
	190	+ "cache_bytes": 47185920
	191	+ },
	192	+ {
	193	+ "policy": "forced_single",
	194	+ "batch": 8,
	195	+ "budget": 2,
	196	+ "forced_k": 1,
	197	+ "median_ms": 15.261041990015656,
	198	+ "p95_ms": 16.932114987866953,
	199	+ "mean_ms": 16.069337232086884,
	200	+ "n": 96,
	201	+ "cache_bytes": 47185920
	202	+ },
	203	+ {
	204	+ "policy": "forced_pair",
	205	+ "batch": 8,
	206	+ "budget": 2,
	207	+ "forced_k": 2,
	208	+ "median_ms": 15.204499941319227,
	209	+ "p95_ms": 16.368166747270152,
	210	+ "mean_ms": 14.876095493188283,
	211	+ "n": 96,
	212	+ "cache_bytes": 47185920
	213	+ },
	214	+ {
	215	+ "policy": "forced_single",
	216	+ "batch": 8,
	217	+ "budget": 4,
	218	+ "forced_k": 1,
	219	+ "median_ms": 18.81216699257493,
	220	+ "p95_ms": 20.145801798207685,
	221	+ "mean_ms": 18.852500479018392,
	222	+ "n": 96,
	223	+ "cache_bytes": 47185920
	224	+ },
	225	+ {
	226	+ "policy": "forced_pair",
	227	+ "batch": 8,
	228	+ "budget": 4,
	229	+ "forced_k": 2,
	230	+ "median_ms": 20.03912499640137,
	231	+ "p95_ms": 21.57740603433922,
	232	+ "mean_ms": 19.88851258768894,
	233	+ "n": 96,
	234	+ "cache_bytes": 47185920
	235	+ },
	236	+ {
	237	+ "policy": "forced_single",
	238	+ "batch": 8,
	239	+ "budget": 8,
	240	+ "forced_k": 1,
	241	+ "median_ms": 28.970604529604316,
	242	+ "p95_ms": 29.755865019978955,
	243	+ "mean_ms": 28.583446644915966,
	244	+ "n": 96,
	245	+ "cache_bytes": 47185920
	246	+ },
	247	+ {
	248	+ "policy": "forced_pair",
	249	+ "batch": 8,
	250	+ "budget": 8,
	251	+ "forced_k": 2,
	252	+ "median_ms": 29.06233351677656,
	253	+ "p95_ms": 33.15305270371027,
	254	+ "mean_ms": 28.535953611329507,
	255	+ "n": 96,
	256	+ "cache_bytes": 47185920
	257	+ },
	258	+ {
	259	+ "policy": "controller",
	260	+ "batch": 8,
	261	+ "budget": 8,
	262	+ "median_ms": 32.673625042662024,
	263	+ "p95_ms": 77.21810429939069,
	264	+ "mean_ms": 30.23307767701529,
	265	+ "n": 96,
	266	+ "cache_bytes": 47185920
	267	+ },
	268	+ {
	269	+ "policy": "dense_forced",
	270	+ "batch": 8,
	271	+ "budget": 1,
	272	+ "forced_k": 1,
	273	+ "median_ms": 13.034250470809639,
	274	+ "p95_ms": 26.244479202432558,
	275	+ "mean_ms": 14.254568138615772,
	276	+ "n": 96,
	277	+ "cache_bytes": 47185920
	278	+ },
	279	+ {
	280	+ "policy": "dense_forced",
	281	+ "batch": 8,
	282	+ "budget": 4,
	283	+ "forced_k": 1,
	284	+ "median_ms": 21.35618799366057,
	285	+ "p95_ms": 88.62939599202946,
	286	+ "mean_ms": 26.533539952652063,
	287	+ "n": 96,
	288	+ "cache_bytes": 47185920
	289	+ },
	290	+ {
	291	+ "policy": "dense_forced",
	292	+ "batch": 8,
	293	+ "budget": 8,
	294	+ "forced_k": 1,
	295	+ "median_ms": 32.09874941967428,
	296	+ "p95_ms": 34.511145524447784,
	297	+ "mean_ms": 32.247447521513095,
	298	+ "n": 96,
	299	+ "cache_bytes": 47185920
	300	+ }
	301	+ ]
	302	+}
		\ No newline at end of file

added results/raw/routing_103m_val.json +262 −0

		@@ -0,0 +1,262 @@
	1	+{
	2	+ "config": "configs/experiments/routing_103m.yaml",
	3	+ "checkpoint": "experiments/runs/controller_103m_seed01_20261030034451/checkpoint",
	4	+ "checkpoint_sha256": "64d08e9120608983b261bd6d108eb2421da45c84ebc733fa5fb2b88780dae427",
	5	+ "split": "val",
	6	+ "code_commit": "unknown",
	7	+ "device": "mps",
	8	+ "seconds": 214.7,
	9	+ "records": [
	10	+ {
	11	+ "policy": "budget0",
	12	+ "loss": 3.988917422294617,
	13	+ "edits_per_token": 0.0,
	14	+ "extra_ops_ratio": 0.0,
	15	+ "n_tokens": 81920
	16	+ },
	17	+ {
	18	+ "policy": "random_forced",
	19	+ "budget": 1,
	20	+ "loss": 3.9874816298484803,
	21	+ "edits_per_token": 1.0,
	22	+ "extra_ops_ratio": 0.09125293754041194,
	23	+ "n_tokens": 81920
	24	+ },
	25	+ {
	26	+ "policy": "uniform",
	27	+ "budget": 1,
	28	+ "loss": 3.988114631175995,
	29	+ "edits_per_token": 1.0,
	30	+ "extra_ops_ratio": 0.0810016319155693,
	31	+ "n_tokens": 81920
	32	+ },
	33	+ {
	34	+ "policy": "fixed_pair",
	35	+ "budget": 1,
	36	+ "loss": 3.9875080823898315,
	37	+ "edits_per_token": 1.0,
	38	+ "extra_ops_ratio": 0.10161040723323822,
	39	+ "n_tokens": 81920
	40	+ },
	41	+ {
	42	+ "policy": "random_forced",
	43	+ "budget": 2,
	44	+ "loss": 3.986599826812744,
	45	+ "edits_per_token": 2.0,
	46	+ "extra_ops_ratio": 0.18254537209868432,
	47	+ "n_tokens": 81920
	48	+ },
	49	+ {
	50	+ "policy": "uniform",
	51	+ "budget": 2,
	52	+ "loss": 3.9874621868133544,
	53	+ "edits_per_token": 2.0,
	54	+ "extra_ops_ratio": 0.1620032638311386,
	55	+ "n_tokens": 81920
	56	+ },
	57	+ {
	58	+ "policy": "fixed_pair",
	59	+ "budget": 2,
	60	+ "loss": 3.986551058292389,
	61	+ "edits_per_token": 2.0,
	62	+ "extra_ops_ratio": 0.20322081446647644,
	63	+ "n_tokens": 81920
	64	+ },
	65	+ {
	66	+ "policy": "random_forced",
	67	+ "budget": 4,
	68	+ "loss": 3.9859113812446596,
	69	+ "edits_per_token": 4.0,
	70	+ "extra_ops_ratio": 0.3652265787124634,
	71	+ "n_tokens": 81920
	72	+ },
	73	+ {
	74	+ "policy": "uniform",
	75	+ "budget": 4,
	76	+ "loss": 3.98651465177536,
	77	+ "edits_per_token": 4.0,
	78	+ "extra_ops_ratio": 0.3240065276622772,
	79	+ "n_tokens": 81920
	80	+ },
	81	+ {
	82	+ "policy": "fixed_pair",
	83	+ "budget": 4,
	84	+ "loss": 3.9858184456825256,
	85	+ "edits_per_token": 4.0,
	86	+ "extra_ops_ratio": 0.4064416289329529,
	87	+ "n_tokens": 81920
	88	+ },
	89	+ {
	90	+ "policy": "random_forced",
	91	+ "budget": 8,
	92	+ "loss": 3.9863768219947815,
	93	+ "edits_per_token": 8.0,
	94	+ "extra_ops_ratio": 0.7305014580488205,
	95	+ "n_tokens": 81920
	96	+ },
	97	+ {
	98	+ "policy": "uniform",
	99	+ "budget": 8,
	100	+ "loss": 3.9858006834983826,
	101	+ "edits_per_token": 8.0,
	102	+ "extra_ops_ratio": 0.6480130553245544,
	103	+ "n_tokens": 81920
	104	+ },
	105	+ {
	106	+ "policy": "fixed_pair",
	107	+ "budget": 8,
	108	+ "loss": 3.9886863827705383,
	109	+ "edits_per_token": 8.0,
	110	+ "extra_ops_ratio": 0.8128832578659058,
	111	+ "n_tokens": 81920
	112	+ },
	113	+ {
	114	+ "policy": "entropy",
	115	+ "threshold": 1.0,
	116	+ "budget": 8,
	117	+ "loss": 3.986129975318909,
	118	+ "edits_per_token": 7.41322021484375,
	119	+ "extra_ops_ratio": 0.6520211786031723,
	120	+ "n_tokens": 81920
	121	+ },
	122	+ {
	123	+ "policy": "entropy",
	124	+ "threshold": 2.0,
	125	+ "budget": 8,
	126	+ "loss": 3.986458885669708,
	127	+ "edits_per_token": 6.782666015625,
	128	+ "extra_ops_ratio": 0.5965691864490509,
	129	+ "n_tokens": 81920
	130	+ },
	131	+ {
	132	+ "policy": "entropy",
	133	+ "threshold": 4.0,
	134	+ "budget": 8,
	135	+ "loss": 3.9870473384857177,
	136	+ "edits_per_token": 4.45064697265625,
	137	+ "extra_ops_ratio": 0.3914341315627098,
	138	+ "n_tokens": 81920
	139	+ },
	140	+ {
	141	+ "policy": "norm",
	142	+ "threshold": 0.05,
	143	+ "budget": 8,
	144	+ "loss": 3.987158489227295,
	145	+ "edits_per_token": 1.64981689453125,
	146	+ "extra_ops_ratio": 0.14528941884636878,
	147	+ "n_tokens": 81920
	148	+ },
	149	+ {
	150	+ "policy": "norm",
	151	+ "threshold": 0.1,
	152	+ "budget": 8,
	153	+ "loss": 3.9875789999961855,
	154	+ "edits_per_token": 1.00203857421875,
	155	+ "extra_ops_ratio": 0.08833706304430962,
	156	+ "n_tokens": 81920
	157	+ },
	158	+ {
	159	+ "policy": "norm",
	160	+ "threshold": 0.2,
	161	+ "budget": 8,
	162	+ "loss": 3.987588274478912,
	163	+ "edits_per_token": 1.0,
	164	+ "extra_ops_ratio": 0.0881563350558281,
	165	+ "n_tokens": 81920
	166	+ },
	167	+ {
	168	+ "policy": "controller_locked",
	169	+ "lam": 0.3,
	170	+ "kap": 0.5,
	171	+ "budget": 8,
	172	+ "loss": 3.988595390319824,
	173	+ "edits_per_token": 0.61395263671875,
	174	+ "extra_ops_ratio": 0.04973116461187601,
	175	+ "n_tokens": 81920
	176	+ },
	177	+ {
	178	+ "policy": "controller",
	179	+ "lam": 0.0,
	180	+ "kap": 0.5,
	181	+ "budget": 8,
	182	+ "loss": 3.987903368473053,
	183	+ "edits_per_token": 3.03917236328125,
	184	+ "extra_ops_ratio": 0.26631725803017614,
	185	+ "n_tokens": 81920
	186	+ },
	187	+ {
	188	+ "policy": "controller",
	189	+ "lam": 0.03,
	190	+ "kap": 0.5,
	191	+ "budget": 8,
	192	+ "loss": 3.9879065990447997,
	193	+ "edits_per_token": 2.6948974609375,
	194	+ "extra_ops_ratio": 0.22800378128886223,
	195	+ "n_tokens": 81920
	196	+ },
	197	+ {
	198	+ "policy": "controller",
	199	+ "lam": 0.1,
	200	+ "kap": 0.5,
	201	+ "budget": 8,
	202	+ "loss": 3.9881483793258665,
	203	+ "edits_per_token": 1.97513427734375,
	204	+ "extra_ops_ratio": 0.16089022792875768,
	205	+ "n_tokens": 81920
	206	+ },
	207	+ {
	208	+ "policy": "controller",
	209	+ "lam": 0.3,
	210	+ "kap": 0.5,
	211	+ "budget": 8,
	212	+ "loss": 3.988595390319824,
	213	+ "edits_per_token": 0.61395263671875,
	214	+ "extra_ops_ratio": 0.04973116461187601,
	215	+ "n_tokens": 81920
	216	+ },
	217	+ {
	218	+ "policy": "oracle1_greedy",
	219	+ "budget": 1,
	220	+ "diagnostic": true,
	221	+ "loss": 3.8904590606689453,
	222	+ "edits_per_token": 0.5654296875,
	223	+ "extra_ops_ratio": 0.056877437978982925,
	224	+ "n_tokens": 8192
	225	+ },
	226	+ {
	227	+ "policy": "oracle1_greedy",
	228	+ "budget": 2,
	229	+ "diagnostic": true,
	230	+ "loss": 3.8827518224716187,
	231	+ "edits_per_token": 1.1217041015625,
	232	+ "extra_ops_ratio": 0.11277931556105614,
	233	+ "n_tokens": 8192
	234	+ },
	235	+ {
	236	+ "policy": "oracle1_greedy",
	237	+ "budget": 4,
	238	+ "diagnostic": true,
	239	+ "loss": 3.8686580657958984,
	240	+ "edits_per_token": 2.21142578125,
	241	+ "extra_ops_ratio": 0.22216549515724182,
	242	+ "n_tokens": 8192
	243	+ },
	244	+ {
	245	+ "policy": "oracle1_greedy",
	246	+ "budget": 8,
	247	+ "diagnostic": true,
	248	+ "loss": 3.844477653503418,
	249	+ "edits_per_token": 4.278076171875,
	250	+ "extra_ops_ratio": 0.42921778559684753,
	251	+ "n_tokens": 8192
	252	+ },
	253	+ {
	254	+ "policy": "oracle_two_step",
	255	+ "diagnostic": true,
	256	+ "one_step_gain": 0.00875546969473362,
	257	+ "two_step_gain": 0.016163207590579987,
	258	+ "greedy_then_best_gain": 0.016161995008587837,
	259	+ "n_tokens": 4096
	260	+ }
	261	+ ]
	262	+}
		\ No newline at end of file

modified scripts/make_results_tables.py +99 −1

		@@ -2,7 +2,7 @@
2	2	experiments/runs (local copies), results/raw/*.json and results/analyzed/*.json. Every number is traceable to a run id."""
3	3	from __future__ import annotations
4	4
5		−import glob, json, math
	5	+import glob, json, math, re
6	6	from collections import defaultdict
7	7	from pathlib import Path
8	8
		@@ -316,3 +316,101 @@ System & budget & """ + " & ".join(fams) + r""" & transitions ($+$/$-$)\\\midrul
316	316	_old_make_all2 = make_all
317	317	def make_all(macro): # noqa: F811
318	318	_old_make_all2(macro); table_em_families(macro)
	319	+
	320	+
	321	+def table_pilot(macro):
	322	+ """103M pilot: Stage A/B manifests, held-out perplexity by policy (evaluation.run), routing and latency at 103M."""
	323	+ import math
	324	+ rows = []
	325	+ def man(prefix):
	326	+ ms = sorted(glob.glob(f"{RUNS}/{prefix}_seed*/manifest.json"))
	327	+ return json.load(open(ms[-1])) if ms else None
	328	+ pre = man("pretrain103m"); en = man("pretrain_enlarged102m"); ed = man("editor_pilot_103m_lr3e4"); gr = man("gr_pilot_103m")
	329	+ if pre:
	330	+ macro("pilotbbtokens", f"{pre['train_tokens']/1e6:.1f}"); macro("pilotbbhours", f"{pre['wall_seconds']/3600:.1f}"); macro("pilotbbppl", f"{math.exp(pre['final_val_loss']):.1f}")
	331	+ macro("pilotbbpeakgb", f"{pre['peak_memory_bytes']/1e9:.0f}"); macro("pilotbbflops", f"{pre['estimated_train_flops']/1e16:.1f}")
	332	+ if en and en.get("state") == "completed":
	333	+ macro("pilotenppl", f"{math.exp(en['final_val_loss']):.1f}"); macro("pilotenhours", f"{en['wall_seconds']/3600:.1f}")
	334	+ else:
	335	+ macro("pilotenppl", "n/a"); macro("pilotenhours", "n/a")
	336	+ for m, tag in [(ed, "piloted"), (gr, "pilotgr")]:
	337	+ if m:
	338	+ fv = _final_val(f"{RUNS}/{m['run_id']}")
	339	+ b0 = fv.get("val_loss_b0");
	340	+ for b in (1, 2, 4, 8):
	341	+ macro(f"{tag}gainb{ {1:'one',2:'two',4:'four',8:'eight'}[b] }", f"{100*(1 - fv.get(f'val_loss_b{b}', b0)/b0):.2f}" if b0 else "n/a")
	342	+ macro(f"{tag}minutes", f"{m['wall_seconds']/60:.0f}"); macro(f"{tag}peakgb", f"{m['peak_memory_bytes']/1e9:.1f}"); macro(f"{tag}params", f"{m['n_trainable']:,}")
	343	+ else:
	344	+ for bn in ("one", "two", "four", "eight"):
	345	+ macro(f"{tag}gainb{bn}", "n/a")
	346	+ macro(f"{tag}minutes", "n/a"); macro(f"{tag}peakgb", "n/a"); macro(f"{tag}params", "n/a")
	347	+ # perplexity table from evaluation.run outputs
	348	+ evs = sorted(glob.glob("results/raw/eval_*_val.json"))
	349	+ if evs:
	350	+ head = r"""\begin{table}[t]\centering\small
	351	+\caption{103M pilot: held-out cross-entropy (nats/token) and perplexity on 327,680 FineWeb-Edu validation tokens
	352	+(40 $\times$ 16 $\times$ 512) under forced random budgets and, for PRISME, under the locked controller.}
	353	+\label{tab:pilotppl}
	354	+\begin{tabular}{llrr}\toprule
	355	+System & policy & loss & ppl\\\midrule
	356	+"""
	357	+ body = ""
	358	+ for f in evs:
	359	+ d = json.load(open(f))
	360	+ for pol, r in d.get("perplexity", {}).items():
	361	+ body += f"{d['system'].replace('_', ' ')} & {pol.replace('_', ' ')} & {r['loss']:.4f} & {r['ppl']:.2f}\\\\\n"
	362	+ if pol == "forced_b0":
	363	+ macro("pilotppl" + re.sub(r"[^a-z]", "", d["system"]), f"{r['ppl']:.2f}")
	364	+ (OUT / "pilot_ppl.tex").write_text(head + body + r"\bottomrule\end{tabular}\end{table}" + "\n")
	365	+ rt = Path("results/raw/routing_103m_val.json")
	366	+ if rt.exists():
	367	+ d = json.load(open(rt)); recs = d["records"]
	368	+ def pick(pol, **kw):
	369	+ for r in recs:
	370	+ if r["policy"] == pol and all(r.get(k) == v for k, v in kw.items()):
	371	+ return r
	372	+ b0 = pick("budget0"); ctl = pick("controller_locked"); ora = pick("oracle1_greedy", budget=8); uni1 = pick("uniform", budget=1)
	373	+ if b0 and ctl:
	374	+ macro("pilotroutingctrl", f"{100*(ctl['loss']/b0['loss']-1):+.2f}"); macro("pilotroutingctrledits", f"{ctl['edits_per_token']:.2f}")
	375	+ if b0 and ora:
	376	+ macro("pilotroutingoracle", f"{100*(ora['loss']/b0['loss']-1):+.1f}"); macro("pilotroutingoracleedits", f"{ora['edits_per_token']:.2f}")
	377	+ if b0 and uni1:
	378	+ macro("pilotroutinguniformone", f"{100*(uni1['loss']/b0['loss']-1):+.2f}")
	379	+ head = r"""\begin{table}[t]\centering\small
	380	+\caption{103M pilot: routing comparison on validation text (20 $\times$ 8 $\times$ 512 tokens). Oracles are diagnostic.}
	381	+\label{tab:pilotrouting}
	382	+\begin{tabular}{llrrr}\toprule
	383	+Policy & setting & loss & edits/token & extra ops\\\midrule
	384	+"""
	385	+ body = ""
	386	+ for r in recs:
	387	+ if r["policy"] == "oracle_two_step":
	388	+ continue
	389	+ setting = ", ".join(f"{k}={v}" for k, v in r.items() if k in ("budget", "threshold", "lam", "kap") and v is not None)
	390	+ body += f"{r['policy'].replace('_', ' ')}{' (diag.)' if r.get('diagnostic') else ''} & {setting} & {r['loss']:.4f} & {r['edits_per_token']:.2f} & {r['extra_ops_ratio']:.3f}\\\\\n"
	391	+ (OUT / "pilot_routing.tex").write_text(head + body + r"\bottomrule\end{tabular}\end{table}" + "\n")
	392	+ lt = Path("results/raw/latency_103m_mps.json")
	393	+ if lt.exists():
	394	+ d = json.load(open(lt))
	395	+ head = r"""\begin{table}[t]\centering\small
	396	+\caption{103M pilot: measured decode latency per generated token on %s (fp32, prompt %d tokens, %d generated, three
	397	+repetitions, median and p95), batch 1 and 8.}
	398	+\label{tab:pilotlatency}
	399	+\begin{tabular}{llrrrr}\toprule
	400	+Policy & budget & batch & median ms & p95 ms & cache bytes\\\midrule
	401	+""" % (d["device"], d["prompt_len"], d["n_new"])
	402	+ body = ""
	403	+ for r in d["records"]:
	404	+ body += f"{r['policy'].replace('_', ' ')} & {r['budget']} & {r['batch']} & {r['median_ms']:.2f} & {r['p95_ms']:.2f} & {r['cache_bytes']:,}\\\\\n"
	405	+ (OUT / "pilot_latency.tex").write_text(head + body + r"\bottomrule\end{tabular}\end{table}" + "\n")
	406	+ b0 = [r for r in d["records"] if r["policy"] == "budget0" and r["batch"] == 1]; s8 = [r for r in d["records"] if r["policy"] == "forced_single" and r["budget"] == 8 and r["batch"] == 1]
	407	+ d8 = [r for r in d["records"] if r["policy"] == "dense_forced" and r["budget"] == 8 and r["batch"] == 1]; c = [r for r in d["records"] if r["policy"] == "controller" and r["batch"] == 1]
	408	+ if b0: macro("pilotlatbzero", f"{b0[0]['median_ms']:.1f}")
	409	+ if s8: macro("pilotlatseight", f"{s8[0]['median_ms']:.1f}")
	410	+ if d8: macro("pilotlatdeight", f"{d8[0]['median_ms']:.1f}")
	411	+ if c: macro("pilotlatctrl", f"{c[0]['median_ms']:.1f}")
	412	+
	413	+
	414	+_old_make_all3 = make_all
	415	+def make_all(macro): # noqa: F811
	416	+ _old_make_all3(macro); table_pilot(macro)
319	417