1claim,metric,value,benchmark,run_ids,artifact,analysis_script,status2Backbone parameter counts equal the proposal for 103M/459M/1.18B,unique parameters,99502848;446744064;1149339648,meta instantiation,,results/tables/param_counts.json,scripts/count_params.py,complete3Implemented workspace+controller counts differ from the proposal by a closed-form 7w^2,unique parameters,2631173;10505221;21655813,meta instantiation,,results/tables/param_counts.json,scripts/count_params.py,complete4Analytic ops table reproduces the proposal within 1.3%,ops/token at 459M T=4096,1.296;1.666;1.787;2.219;5.633 e9,analytic,,results/tables/ops_459m_T4096.json,scripts/estimate_flops.py,complete515 mandated invariants + cache equivalence pass,tests,23 passed,pytest,,tests/,python -m pytest tests,complete6Zero-U bootstrap does not learn (scale mismatch),val loss by budget after 700 steps,0.4144/0.4143/0.4143/0.4146/0.4164,synthetic dev-iid v1 backbone,p4_editor_v1_seed01_20261029233624,experiments/runs/p4_editor_v1_seed01_20261029233624/metrics.jsonl,,stopped-run-used7Overfitting test passes through the workspace path,train loss 64 items budget 8,0.347->0.0024 in 200 steps,synthetic train subset,p4_overfit64_b8_seed01,experiments/runs/p4_overfit64_b8_seed01_20261029235111/metrics.jsonl,,complete8Sparse editor lowers held-out loss vs frozen backbone (H1 weak),dev-iid answer loss relative change by budget,see numbers.tex vtwoprismegainb*,synthetic dev-iid v2 backbone 5 seeds,p4v2_prisme,results/analyzed/seed_matrix_v2.json,scripts/make_results_tables.py,complete9Global-recurrence adapter matches or exceeds the sparse editor at every budget,dev-iid answer loss by budget,see numbers.tex vtwogrgainb*,synthetic dev-iid v2 backbone 5 seeds,p4v2_gr,results/analyzed/seed_matrix_v2.json,scripts/make_results_tables.py,complete10Variance of intervention gains is mostly between states,within-state share,see numbers.tex labelscalibwithin,editor-unseen label stream (calib audit),labels_v1ctrl_seed01_20261030004018,results/analyzed/labels_labels_v1ctrl_seed01_20261030004018.json,evaluation/label_analysis.py,complete11One-step oracle lowers dev-iid loss by ~15% at budget 8 (diagnostic),loss relative change,see numbers.tex routingoraclebeight,synthetic dev-iid v1 editor,controller_v1ctrl_seed01,results/raw/routing_v1ctrl.json,evaluation/routing.py,complete12Learned controller is worse than no routing on dev-iid (H3 null),loss relative change at locked lambda,see numbers.tex routingctrllocked;vtwoctrlrel,synthetic dev-iid v1 and v2 editors,controller_v1ctrl_seed01;controller_v2s2;controller_v2s3;controller_v2s4;controller_v2s5,results/raw/routing_v1ctrl.json;results/analyzed/routing_seeds_v2.json,evaluation/routing.py,complete13Fixed schedules: random = uniform = fixed pair (H2 null),loss at budget 1,see numbers.tex routing*bone,synthetic dev-iid v1 editor,controller_v1ctrl_seed01,results/raw/routing_v1ctrl.json,evaluation/routing.py,complete14Two-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,complete15Dispatch 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,complete16Miniature 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,complete17103M 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,complete18103M 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,complete19103M 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-used20103M 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,complete21103M 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,complete22103M 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,complete23Pre-registered primary outcome: pooled test EM gain of sparse editor at b2 is +0.9 pp (< 3 pp threshold),exact match pp,see numbers.tex testprismebtwovsbzero*,synthetic locked test regimes (1000 items each),p4v2_prisme_seed01_20261030001749;p4v2_prisme_seed02_20261030001749;p4v2_prisme_seed03_20261030001749;p4v2_prisme_seed04_20261030001749;p4v2_prisme_seed05_20261030001839,results/analyzed/test_statistics.json;paper/tables/test.tex,scripts/make_results_tables.py,complete24Sparse editor vs global recurrence on locked test: -0.3 pp [-0.6 +0.1]; vs dense +0.4 pp [+0.1 +0.7],exact match pp,see numbers.tex testprismevs*,synthetic locked test regimes,p4v2_prisme;p4v2_gr;p4v2_dense,results/analyzed/test_statistics.json,scripts/make_results_tables.py,complete25Parameter-matched plain decoder (102.2M) beats backbone+adapters on validation text,val loss,3.8689 vs 3.8746,FineWeb-Edu val,pretrain_enlarged102m_seed01_20261030015017,experiments/runs/pretrain_enlarged102m_seed01_20261030015017/manifest.json,training/pretrain.py,complete26Locked FineWeb test: backbone ppl 45.72; sparse editor and global adapter change test loss by -0.08% at budget 4; enlarged decoder lower,test ppl / relative loss,see numbers.tex pilotppl*test;pilottest*,FineWeb-Edu locked test (327680 tokens),pretrain103m_seed01_20261029232706;editor_pilot_103m_lr3e4_seed01_20261030021200;gr_pilot_103m_seed01_20261030015117;pretrain_enlarged102m_seed01_20261030015017,results/raw/eval_backbone_103m_test.json;results/raw/eval_prisme_103m_editor_test.json;results/raw/eval_gr_103m_test.json;results/raw/eval_enlarged_102m_test.json,evaluation/run.py,complete27Ablations: M in {4 8 32} / L_w in {1 4} / offset pairs all within 0.003 nats of the reference at b2,dev-iid loss relative change,see numbers.tex abl*gainbtwo,synthetic dev-iid v2 backbone 3 seeds,p4v2_abl_M4;p4v2_abl_M8;p4v2_abl_M32;p4v2_abl_L1;p4v2_abl_L4;p4v2_abl_pairoffset,paper/tables/ablations.tex,scripts/make_results_tables.py,complete28