PNS-Bind-25M: the 24 evaluated checkpoints (safetensors)

checkpoints/E3A_bind_s1/config.json

@@ -0,0 +1,37 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "E3A_bind_s1",
+ "batch": 128,
+ "cfg": {
+  "T_ev": 48,
+  "d": 192,
+  "enc_layers": 3,
+  "gate_bias": 0.0,
+  "heads": 6,
+  "mode": "bind",
+  "n_slots": 192,
+  "vocab": 8192
+ },
+ "group": "experiment2_stageA",
+ "kind": "bind",
+ "lr": 0.0003,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 4385515,
+ "role": "Stage A, binding-local writes, seed 1",
+ "run": "E3A_bind_s1",
+ "seed": 1,
+ "source_run_directory": "E3A_bind_s1",
+ "state_dict_tensor_total": 4385515,
+ "train_args": {
+  "batch": 128,
+  "eval_lifetimes": 400,
+  "lr": 0.0003,
+  "mode": "bind",
+  "run": "E3A_bind_s1",
+  "seed": 1,
+  "updates": 2600
+ },
+ "train_corpus": "e3_train",
+ "updates": 2600
+}

checkpoints/E3A_bind_s1/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f236a1059eb076147fb3e01a3c59856cf97d241087c3bf1c944522b8e800f45b
+size 17547772

checkpoints/E3A_bind_s2/config.json

@@ -0,0 +1,37 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "E3A_bind_s2",
+ "batch": 128,
+ "cfg": {
+  "T_ev": 48,
+  "d": 192,
+  "enc_layers": 3,
+  "gate_bias": 0.0,
+  "heads": 6,
+  "mode": "bind",
+  "n_slots": 192,
+  "vocab": 8192
+ },
+ "group": "experiment2_stageA",
+ "kind": "bind",
+ "lr": 0.0003,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 4385515,
+ "role": "Stage A, binding-local writes, seed 2",
+ "run": "E3A_bind_s2",
+ "seed": 2,
+ "source_run_directory": "E3A_bind_s2",
+ "state_dict_tensor_total": 4385515,
+ "train_args": {
+  "batch": 128,
+  "eval_lifetimes": 400,
+  "lr": 0.0003,
+  "mode": "bind",
+  "run": "E3A_bind_s2",
+  "seed": 2,
+  "updates": 2600
+ },
+ "train_corpus": "e3_train",
+ "updates": 2600
+}

checkpoints/E3A_bind_s2/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:801b7e0bdf4964293b33093c6447492438c6ed6f3952a6c89c58652d8617bfeb
+size 17547772

checkpoints/E3A_bind_s3/config.json

@@ -0,0 +1,37 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "E3A_bind_s3",
+ "batch": 128,
+ "cfg": {
+  "T_ev": 48,
+  "d": 192,
+  "enc_layers": 3,
+  "gate_bias": 0.0,
+  "heads": 6,
+  "mode": "bind",
+  "n_slots": 192,
+  "vocab": 8192
+ },
+ "group": "experiment2_stageA",
+ "kind": "bind",
+ "lr": 0.0003,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 4385515,
+ "role": "Stage A, binding-local writes, seed 3",
+ "run": "E3A_bind_s3",
+ "seed": 3,
+ "source_run_directory": "E3A_bind_s3",
+ "state_dict_tensor_total": 4385515,
+ "train_args": {
+  "batch": 128,
+  "eval_lifetimes": 400,
+  "lr": 0.0003,
+  "mode": "bind",
+  "run": "E3A_bind_s3",
+  "seed": 3,
+  "updates": 2600
+ },
+ "train_corpus": "e3_train",
+ "updates": 2600
+}

checkpoints/E3A_bind_s3/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8bb38f08665ccdc80f1332f6c353c2b0ac053ecac073d1c67a8455f16972fca
+size 17547772

checkpoints/E3A_unbound_s1/config.json

@@ -0,0 +1,37 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "E3A_unbound_s1",
+ "batch": 128,
+ "cfg": {
+  "T_ev": 48,
+  "d": 192,
+  "enc_layers": 3,
+  "gate_bias": 0.0,
+  "heads": 6,
+  "mode": "unbound",
+  "n_slots": 192,
+  "vocab": 8192
+ },
+ "group": "experiment2_stageA",
+ "kind": "unbound",
+ "lr": 0.0003,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 4385515,
+ "role": "Stage A, capacity-matched unstructured control, seed 1",
+ "run": "E3A_unbound_s1",
+ "seed": 1,
+ "source_run_directory": "E3A_unbound_s1",
+ "state_dict_tensor_total": 4385515,
+ "train_args": {
+  "batch": 128,
+  "eval_lifetimes": 400,
+  "lr": 0.0003,
+  "mode": "unbound",
+  "run": "E3A_unbound_s1",
+  "seed": 1,
+  "updates": 2600
+ },
+ "train_corpus": "e3_train",
+ "updates": 2600
+}

checkpoints/E3A_unbound_s1/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13484d5904ae8d3875b8e5b091e2de341b861fd31ed54f728175516984f6e890
+size 17547772

checkpoints/E3A_unbound_s2/config.json

@@ -0,0 +1,37 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "E3A_unbound_s2",
+ "batch": 128,
+ "cfg": {
+  "T_ev": 48,
+  "d": 192,
+  "enc_layers": 3,
+  "gate_bias": 0.0,
+  "heads": 6,
+  "mode": "unbound",
+  "n_slots": 192,
+  "vocab": 8192
+ },
+ "group": "experiment2_stageA",
+ "kind": "unbound",
+ "lr": 0.0003,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 4385515,
+ "role": "Stage A, capacity-matched unstructured control, seed 2",
+ "run": "E3A_unbound_s2",
+ "seed": 2,
+ "source_run_directory": "E3A_unbound_s2",
+ "state_dict_tensor_total": 4385515,
+ "train_args": {
+  "batch": 128,
+  "eval_lifetimes": 400,
+  "lr": 0.0003,
+  "mode": "unbound",
+  "run": "E3A_unbound_s2",
+  "seed": 2,
+  "updates": 2600
+ },
+ "train_corpus": "e3_train",
+ "updates": 2600
+}

checkpoints/E3A_unbound_s2/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae47667fce968fb2e006575414030e00912aed659741bc92dd4e1879a16f2754
+size 17547772

checkpoints/E3A_unbound_s3/config.json

@@ -0,0 +1,37 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "E3A_unbound_s3",
+ "batch": 128,
+ "cfg": {
+  "T_ev": 48,
+  "d": 192,
+  "enc_layers": 3,
+  "gate_bias": 0.0,
+  "heads": 6,
+  "mode": "unbound",
+  "n_slots": 192,
+  "vocab": 8192
+ },
+ "group": "experiment2_stageA",
+ "kind": "unbound",
+ "lr": 0.0003,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 4385515,
+ "role": "Stage A, capacity-matched unstructured control, seed 3",
+ "run": "E3A_unbound_s3",
+ "seed": 3,
+ "source_run_directory": "E3A_unbound_s3",
+ "state_dict_tensor_total": 4385515,
+ "train_args": {
+  "batch": 128,
+  "eval_lifetimes": 400,
+  "lr": 0.0003,
+  "mode": "unbound",
+  "run": "E3A_unbound_s3",
+  "seed": 3,
+  "updates": 2600
+ },
+ "train_corpus": "e3_train",
+ "updates": 2600
+}

checkpoints/E3A_unbound_s3/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54bfb2ddee00c7da6d8005e7b825f69713af5cde083fb0acbd6e9e8a2c9d1414
+size 17547772

checkpoints/E3B_bind_s1/config.json

@@ -0,0 +1,37 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "E3B_bind_s1",
+ "batch": 96,
+ "cfg": {
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 8,
+  "full_heads": true,
+  "gate_bias": 0.0,
+  "heads": 8,
+  "mode": "bind",
+  "n_slots": 192,
+  "vocab": 8192
+ },
+ "group": "experiment2_stageB",
+ "kind": "bind",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24265440,
+ "role": "Stage B, binding-local writes, seed 1",
+ "run": "E3B_bind_s1",
+ "seed": 1,
+ "source_run_directory": "E3B_bind_s1",
+ "state_dict_tensor_total": 27411168,
+ "train_args": {
+  "batch": 96,
+  "lr": 0.0004,
+  "mode": "bind",
+  "run": "E3B_bind_s1",
+  "seed": 1,
+  "updates": 6000
+ },
+ "train_corpus": "e3_train",
+ "updates": 6000
+}

checkpoints/E3B_bind_s1/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db549fda0dacc874a96dc3d8c83c8991617f83767989db4d3672c7dad2630a32
+size 109659632

checkpoints/E3B_bind_s2/config.json

@@ -0,0 +1,37 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "E3B_bind_s2",
+ "batch": 96,
+ "cfg": {
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 8,
+  "full_heads": true,
+  "gate_bias": 0.0,
+  "heads": 8,
+  "mode": "bind",
+  "n_slots": 192,
+  "vocab": 8192
+ },
+ "group": "experiment2_stageB",
+ "kind": "bind",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24265440,
+ "role": "Stage B, binding-local writes, seed 2",
+ "run": "E3B_bind_s2",
+ "seed": 2,
+ "source_run_directory": "E3B_bind_s2",
+ "state_dict_tensor_total": 27411168,
+ "train_args": {
+  "batch": 96,
+  "lr": 0.0004,
+  "mode": "bind",
+  "run": "E3B_bind_s2",
+  "seed": 2,
+  "updates": 6000
+ },
+ "train_corpus": "e3_train",
+ "updates": 6000
+}

checkpoints/E3B_bind_s2/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b692713a9730855faf59525abf56c742c2c2766619c9de4cf6904783a7f790d9
+size 109659632

checkpoints/E3B_bind_s3/config.json

@@ -0,0 +1,37 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "E3B_bind_s3",
+ "batch": 96,
+ "cfg": {
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 8,
+  "full_heads": true,
+  "gate_bias": 0.0,
+  "heads": 8,
+  "mode": "bind",
+  "n_slots": 192,
+  "vocab": 8192
+ },
+ "group": "experiment2_stageB",
+ "kind": "bind",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24265440,
+ "role": "Stage B, binding-local writes, seed 3",
+ "run": "E3B_bind_s3",
+ "seed": 3,
+ "source_run_directory": "E3B_bind_s3",
+ "state_dict_tensor_total": 27411168,
+ "train_args": {
+  "batch": 96,
+  "lr": 0.0004,
+  "mode": "bind",
+  "run": "E3B_bind_s3",
+  "seed": 3,
+  "updates": 6000
+ },
+ "train_corpus": "e3_train",
+ "updates": 6000
+}

checkpoints/E3B_bind_s3/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:466d889a73a74b341b4fe52f071615be3dbe4b8de00f1d2c4fd98bb0c9da2de2
+size 109659632

checkpoints/E3B_unbound_s1/config.json

@@ -0,0 +1,37 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "E3B_unbound_s1",
+ "batch": 96,
+ "cfg": {
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 8,
+  "full_heads": true,
+  "gate_bias": 0.0,
+  "heads": 8,
+  "mode": "unbound",
+  "n_slots": 192,
+  "vocab": 8192
+ },
+ "group": "experiment2_stageB",
+ "kind": "unbound",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24265440,
+ "role": "Stage B, capacity/parameter-matched unstructured control, seed 1",
+ "run": "E3B_unbound_s1",
+ "seed": 1,
+ "source_run_directory": "E3B_unbound_s1",
+ "state_dict_tensor_total": 27411168,
+ "train_args": {
+  "batch": 96,
+  "lr": 0.0004,
+  "mode": "unbound",
+  "run": "E3B_unbound_s1",
+  "seed": 1,
+  "updates": 6000
+ },
+ "train_corpus": "e3_train",
+ "updates": 6000
+}

checkpoints/E3B_unbound_s1/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:018324ec3f42ec10a0ee2868145675c1eb56d2947916d3c79446693a6c87370c
+size 109659632

checkpoints/E3B_unbound_s2/config.json

@@ -0,0 +1,37 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "E3B_unbound_s2",
+ "batch": 96,
+ "cfg": {
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 8,
+  "full_heads": true,
+  "gate_bias": 0.0,
+  "heads": 8,
+  "mode": "unbound",
+  "n_slots": 192,
+  "vocab": 8192
+ },
+ "group": "experiment2_stageB",
+ "kind": "unbound",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24265440,
+ "role": "Stage B, capacity/parameter-matched unstructured control, seed 2",
+ "run": "E3B_unbound_s2",
+ "seed": 2,
+ "source_run_directory": "E3B_unbound_s2",
+ "state_dict_tensor_total": 27411168,
+ "train_args": {
+  "batch": 96,
+  "lr": 0.0004,
+  "mode": "unbound",
+  "run": "E3B_unbound_s2",
+  "seed": 2,
+  "updates": 6000
+ },
+ "train_corpus": "e3_train",
+ "updates": 6000
+}

checkpoints/E3B_unbound_s2/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46f09e502c1450bb194426253dff48009bbf5d97436d4437535723f12042ea28
+size 109659632

checkpoints/E3B_unbound_s3/config.json

@@ -0,0 +1,37 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "E3B_unbound_s3",
+ "batch": 96,
+ "cfg": {
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 8,
+  "full_heads": true,
+  "gate_bias": 0.0,
+  "heads": 8,
+  "mode": "unbound",
+  "n_slots": 192,
+  "vocab": 8192
+ },
+ "group": "experiment2_stageB",
+ "kind": "unbound",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24265440,
+ "role": "Stage B, capacity/parameter-matched unstructured control, seed 3",
+ "run": "E3B_unbound_s3",
+ "seed": 3,
+ "source_run_directory": "E3B_unbound_s3",
+ "state_dict_tensor_total": 27411168,
+ "train_args": {
+  "batch": 96,
+  "lr": 0.0004,
+  "mode": "unbound",
+  "run": "E3B_unbound_s3",
+  "seed": 3,
+  "updates": 6000
+ },
+ "train_corpus": "e3_train",
+ "updates": 6000
+}

checkpoints/E3B_unbound_s3/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:feb06a34e4385e05d54d0e98fb1567afccca659a6b5d1578ec87b288352edd6e
+size 109659632

checkpoints/G_CONVEX_FREE/config.json

@@ -0,0 +1,46 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "G_CONVEX_FREE",
+ "batch": 160,
+ "cfg": {
+  "K": 4,
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 7,
+  "heads": 8,
+  "n_workspace": 4,
+  "state_tau": 16.0,
+  "update_rule": "convex",
+  "vocab": 8192,
+  "write_mode": "independent",
+  "write_temp": 1.0
+ },
+ "group": "experiment1",
+ "kind": "pnsr",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24614453,
+ "role": "E2-GATED arm: convex gated update, no clamp",
+ "run": "G_CONVEX_FREE",
+ "seed": 1,
+ "source_run_directory": "G_CONVEX_FREE",
+ "state_dict_tensor_total": 27760181,
+ "train_args": {
+  "batch": 160,
+  "gate_bias": 0.0,
+  "hold_frac": null,
+  "lr": 0.0004,
+  "model": "pnsr",
+  "resume_from": null,
+  "run": "G_CONVEX_FREE",
+  "seed": 1,
+  "state_tau": 16.0,
+  "update_rule": "convex",
+  "updates": 6000,
+  "w_enum": 2.0,
+  "w_other": 1.0
+ },
+ "train_corpus": "train",
+ "updates": 6000
+}

checkpoints/G_CONVEX_FREE/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b8a4854e61a5127c53f60e92b4fee19ddb8cd6e9021109dcabb08a2ef4b3700
+size 111057564

checkpoints/G_CONVEX_RETAIN/config.json

@@ -0,0 +1,45 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "G_CONVEX_RETAIN",
+ "batch": 160,
+ "cfg": {
+  "K": 4,
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 7,
+  "heads": 8,
+  "n_workspace": 4,
+  "state_tau": 16.0,
+  "update_rule": "convex",
+  "vocab": 8192,
+  "write_mode": "independent",
+  "write_temp": 1.0
+ },
+ "group": "experiment1",
+ "kind": "pnsr",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24614453,
+ "role": "E2-GATED arm: convex gated update, retention-biased",
+ "run": "G_CONVEX_RETAIN",
+ "seed": 1,
+ "source_run_directory": "G_CONVEX_RETAIN",
+ "state_dict_tensor_total": 27760181,
+ "train_args": {
+  "batch": 160,
+  "gate_bias": -2.0,
+  "hold_frac": null,
+  "lr": 0.0004,
+  "model": "pnsr",
+  "resume_from": null,
+  "run": "G_CONVEX_RETAIN",
+  "seed": 1,
+  "update_rule": "convex",
+  "updates": 6000,
+  "w_enum": 2.0,
+  "w_other": 1.0
+ },
+ "train_corpus": "train",
+ "updates": 6000
+}

checkpoints/G_CONVEX_RETAIN/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd49eec82cf5f3c367558737350819cbd37e18b3638bfb429a9be41c5c7fdfcf
+size 111057564

checkpoints/G_TAU512/config.json

@@ -0,0 +1,46 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "G_TAU512",
+ "batch": 160,
+ "cfg": {
+  "K": 4,
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 7,
+  "heads": 8,
+  "n_workspace": 4,
+  "state_tau": 512.0,
+  "update_rule": "additive_clamp",
+  "vocab": 8192,
+  "write_mode": "independent",
+  "write_temp": 1.0
+ },
+ "group": "experiment1",
+ "kind": "pnsr",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24614453,
+ "role": "isolating control: same additive rule, clamp raised 16 -> 512",
+ "run": "G_TAU512",
+ "seed": 1,
+ "source_run_directory": "G_TAU512",
+ "state_dict_tensor_total": 27760181,
+ "train_args": {
+  "batch": 160,
+  "gate_bias": 0.0,
+  "hold_frac": null,
+  "lr": 0.0004,
+  "model": "pnsr",
+  "resume_from": null,
+  "run": "G_TAU512",
+  "seed": 1,
+  "state_tau": 512.0,
+  "update_rule": "additive_clamp",
+  "updates": 6000,
+  "w_enum": 2.0,
+  "w_other": 1.0
+ },
+ "train_corpus": "train",
+ "updates": 6000
+}

checkpoints/G_TAU512/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edd0f07d1140a5a40c3d238d71f086f1b2a40719c00b6e7082daf4f89a7c73c6
+size 111057564

checkpoints/PNSR_K1_fin/config.json

@@ -0,0 +1,36 @@
+{
+ "admission_note": "amendment A6: held at stage 1 for a protocol-matched comparison with PNSR_K4_fin",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "PNSR_K1_v2",
+ "batch": 160,
+ "cfg": {
+  "K": 1,
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 7,
+  "heads": 8,
+  "n_workspace": 4,
+  "state_tau": 16.0,
+  "vocab": 8192
+ },
+ "group": "experiment1",
+ "kind": "pnsr_k1",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24614453,
+ "role": "deliberation-depth control K=1 (E5)",
+ "run": "PNSR_K1_fin",
+ "seed": 1,
+ "source_run_directory": "PNSR_K1_fin",
+ "state_dict_tensor_total": 27760181,
+ "train_args": {
+  "batch": 160,
+  "lr": 0.0004,
+  "model": "pnsr_k1",
+  "run": "PNSR_K1_v2",
+  "seed": 1,
+  "updates": 6000
+ },
+ "train_corpus": "train",
+ "updates": 6000
+}

checkpoints/PNSR_K1_fin/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1d90f6f3ea4bd0d67ac38abb896105191d1f925ecf3f14cb3f129a05c08610b
+size 111057564

checkpoints/PNSR_K4_fin/config.json

@@ -0,0 +1,36 @@
+{
+ "admission_note": "amendment A6: the 8x-enum stage-2 continuation failed its own preregistered admission rule, so the stage-1 checkpoint is the admitted one",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "PNSR_K4_v2",
+ "batch": 160,
+ "cfg": {
+  "K": 4,
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 7,
+  "heads": 8,
+  "n_workspace": 4,
+  "state_tau": 16.0,
+  "vocab": 8192
+ },
+ "group": "experiment1",
+ "kind": "pnsr",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24614453,
+ "role": "PNSR typed learned state, seed 1 (additive write, clamp tau=16)",
+ "run": "PNSR_K4_fin",
+ "seed": 1,
+ "source_run_directory": "PNSR_K4_fin",
+ "state_dict_tensor_total": 27760181,
+ "train_args": {
+  "batch": 160,
+  "lr": 0.0004,
+  "model": "pnsr",
+  "run": "PNSR_K4_v2",
+  "seed": 1,
+  "updates": 6000
+ },
+ "train_corpus": "train",
+ "updates": 6000
+}

checkpoints/PNSR_K4_fin/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe9c4cc706e503d36b581d9b74c2e00e0342b07855c570ee05bf4c281b2944bb
+size 111057564

checkpoints/PNSR_K4_s2/config.json

@@ -0,0 +1,46 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "PNSR_K4_s2",
+ "batch": 160,
+ "cfg": {
+  "K": 4,
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 7,
+  "heads": 8,
+  "n_workspace": 4,
+  "state_tau": 16.0,
+  "update_rule": "additive_clamp",
+  "vocab": 8192,
+  "write_mode": "independent",
+  "write_temp": 1.0
+ },
+ "group": "experiment1",
+ "kind": "pnsr",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24614453,
+ "role": "PNSR typed learned state, seed 2",
+ "run": "PNSR_K4_s2",
+ "seed": 2,
+ "source_run_directory": "PNSR_K4_s2",
+ "state_dict_tensor_total": 27760181,
+ "train_args": {
+  "batch": 160,
+  "gate_bias": 0.0,
+  "hold_frac": null,
+  "lr": 0.0004,
+  "model": "pnsr",
+  "resume_from": null,
+  "run": "PNSR_K4_s2",
+  "seed": 2,
+  "state_tau": 16.0,
+  "update_rule": "additive_clamp",
+  "updates": 6000,
+  "w_enum": 2.0,
+  "w_other": 1.0
+ },
+ "train_corpus": "train",
+ "updates": 6000
+}

checkpoints/PNSR_K4_s2/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96b014fc56653b97560432374a325f1c5a154a9586209362107a4673fb293af3
+size 111057564

checkpoints/PNSR_K4_s3/config.json

@@ -0,0 +1,46 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "PNSR_K4_s3",
+ "batch": 160,
+ "cfg": {
+  "K": 4,
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 7,
+  "heads": 8,
+  "n_workspace": 4,
+  "state_tau": 16.0,
+  "update_rule": "additive_clamp",
+  "vocab": 8192,
+  "write_mode": "independent",
+  "write_temp": 1.0
+ },
+ "group": "experiment1",
+ "kind": "pnsr",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24614453,
+ "role": "PNSR typed learned state, seed 3",
+ "run": "PNSR_K4_s3",
+ "seed": 3,
+ "source_run_directory": "PNSR_K4_s3",
+ "state_dict_tensor_total": 27760181,
+ "train_args": {
+  "batch": 160,
+  "gate_bias": 0.0,
+  "hold_frac": null,
+  "lr": 0.0004,
+  "model": "pnsr",
+  "resume_from": null,
+  "run": "PNSR_K4_s3",
+  "seed": 3,
+  "state_tau": 16.0,
+  "update_rule": "additive_clamp",
+  "updates": 6000,
+  "w_enum": 2.0,
+  "w_other": 1.0
+ },
+ "train_corpus": "train",
+ "updates": 6000
+}

checkpoints/PNSR_K4_s3/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:21455bb390a5e7d3fb98f6f484c6698f40b20af097a1b2f2a727a0e2676a4ea4
+size 111057564

checkpoints/PNSR_ON_E3/config.json

@@ -0,0 +1,47 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "PNSR_ON_E3",
+ "batch": 96,
+ "cfg": {
+  "K": 4,
+  "T_ev": 48,
+  "d": 384,
+  "enc_layers": 7,
+  "heads": 8,
+  "n_workspace": 4,
+  "state_tau": 16.0,
+  "update_rule": "additive_clamp",
+  "vocab": 8192,
+  "write_mode": "independent",
+  "write_temp": 1.0
+ },
+ "group": "corrected_corpus",
+ "kind": "pnsr",
+ "lr": 0.0004,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 24614453,
+ "role": "E1 architecture retrained from scratch on the corrected corpus",
+ "run": "PNSR_ON_E3",
+ "seed": 1,
+ "source_run_directory": "PNSR_ON_E3",
+ "state_dict_tensor_total": 27760181,
+ "train_args": {
+  "batch": 96,
+  "gate_bias": 0.0,
+  "hold_frac": null,
+  "lr": 0.0004,
+  "model": "pnsr",
+  "resume_from": null,
+  "run": "PNSR_ON_E3",
+  "seed": 1,
+  "shards_split": "e3_train",
+  "state_tau": 16.0,
+  "update_rule": "additive_clamp",
+  "updates": 6000,
+  "w_enum": 2.0,
+  "w_other": 1.0
+ },
+ "train_corpus": "e3_train",
+ "updates": 6000
+}

checkpoints/PNSR_ON_E3/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10acf934a6ed2661b6c4b8d5e9c05853bde89f1dddb0d36ae18544017cff2e81
+size 111057564

checkpoints/RMT_s1/config.json

@@ -0,0 +1,41 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "RMT_s1",
+ "batch": 160,
+ "cfg": {
+  "T_ev": 48,
+  "d": 384,
+  "heads": 8,
+  "layers": 10,
+  "n_mem": 32,
+  "vocab": 8192
+ },
+ "group": "experiment1",
+ "kind": "rmt",
+ "lr": 0.0003,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 23945525,
+ "role": "memory-token recurrent baseline (RMT-style), seed 1",
+ "run": "RMT_s1",
+ "seed": 1,
+ "source_run_directory": "RMT_s1",
+ "state_dict_tensor_total": 27091253,
+ "train_args": {
+  "batch": 160,
+  "gate_bias": 0.0,
+  "hold_frac": null,
+  "lr": 0.0003,
+  "model": "rmt",
+  "resume_from": null,
+  "run": "RMT_s1",
+  "seed": 1,
+  "state_tau": 16.0,
+  "update_rule": "additive_clamp",
+  "updates": 6000,
+  "w_enum": 2.0,
+  "w_other": 1.0
+ },
+ "train_corpus": "train",
+ "updates": 6000
+}

checkpoints/RMT_s1/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5dc513bddfab863c314ee2e41882037c2a77efe130d6147fb5bde40a0ef8078b
+size 108380916

checkpoints/TX768_fin/config.json

@@ -0,0 +1,37 @@
+{
+ "admission_note": "amendment A7: v2/v3 invalid (v3 was frozen mid-run by the A3 skip-guard, AUDIT 2); v4 retrained with the guard removed, lr 1.5e-4, exposure-matched; best.pt (update 18250) is the admitted checkpoint",
+ "admitted_checkpoint": "best.pt",
+ "admitted_from_run": "TX768_v4",
+ "batch": 384,
+ "cfg": {
+  "d": 384,
+  "heads": 8,
+  "layers": 10,
+  "vocab": 8192,
+  "window": 768
+ },
+ "group": "experiment1",
+ "kind": "tx768",
+ "lr": 0.00015,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 23892917,
+ "role": "transcript-window baseline, 768 tokens (~40 events)",
+ "run": "TX768_fin",
+ "seed": 1,
+ "source_run_directory": "TX768_fin",
+ "state_dict_tensor_total": 27038645,
+ "train_args": {
+  "batch": 384,
+  "hold_frac": 0.0,
+  "lr": 0.00015,
+  "model": "tx768",
+  "resume_from": null,
+  "run": "TX768_v4",
+  "seed": 1,
+  "updates": 18850,
+  "w_enum": 2.0,
+  "w_other": 1.0
+ },
+ "train_corpus": "train",
+ "updates": 18850
+}

checkpoints/TX768_fin/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c2224e807fad0cf576651874ec06770966490fedfa3d7efc6030462c3ee1f10
+size 108169660

checkpoints/TXE_ON_E3/config.json

@@ -0,0 +1,41 @@
+{
+ "admission_note": "the run's own final checkpoint",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "TXE_ON_E3",
+ "batch": 384,
+ "cfg": {
+  "d": 384,
+  "heads": 8,
+  "layers": 10,
+  "vocab": 8192,
+  "window": 64
+ },
+ "group": "corrected_corpus",
+ "kind": "txe",
+ "lr": 0.0003,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 23622581,
+ "role": "memory-free control on the corrected corpus (Experiment 2 comparator)",
+ "run": "TXE_ON_E3",
+ "seed": 1,
+ "source_run_directory": "TXE_ON_E3",
+ "state_dict_tensor_total": 26768309,
+ "train_args": {
+  "batch": 384,
+  "gate_bias": 0.0,
+  "hold_frac": null,
+  "lr": 0.0003,
+  "model": "txe",
+  "resume_from": null,
+  "run": "TXE_ON_E3",
+  "seed": 1,
+  "shards_split": "e3_train",
+  "state_tau": 16.0,
+  "update_rule": "additive_clamp",
+  "updates": 20000,
+  "w_enum": 2.0,
+  "w_other": 1.0
+ },
+ "train_corpus": "e3_train",
+ "updates": 20000
+}

checkpoints/TXE_ON_E3/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2888f76103ec4def71b4b37be97b144e18f63482bae9b1c73078f782ced7f8c4
+size 107088308

checkpoints/TXE_fin/config.json

@@ -0,0 +1,33 @@
+{
+ "admission_note": "TXE_v2c diverged and was killed before completion (amendment A5 auto-fail)",
+ "admitted_checkpoint": "final.pt",
+ "admitted_from_run": "TXE_v2",
+ "batch": 384,
+ "cfg": {
+  "d": 384,
+  "heads": 8,
+  "layers": 10,
+  "vocab": 8192,
+  "window": 64
+ },
+ "group": "experiment1",
+ "kind": "txe",
+ "lr": 0.0003,
+ "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+ "params": 23622581,
+ "role": "memory-free control: current event only",
+ "run": "TXE_fin",
+ "seed": 1,
+ "source_run_directory": "TXE_fin",
+ "state_dict_tensor_total": 26768309,
+ "train_args": {
+  "batch": 384,
+  "lr": 0.0003,
+  "model": "txe",
+  "run": "TXE_v2",
+  "seed": 1,
+  "updates": 20000
+ },
+ "train_corpus": "train",
+ "updates": 20000
+}

checkpoints/TXE_fin/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8dc5eb958b9732f8819826b6f30b2e05010c9cbe455f63a31b2b0c36f8ed60ee
+size 107088308

checkpoints/index.json

@@ -0,0 +1,946 @@
+{
+ "E3A_bind_s1": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "E3A_bind_s1",
+  "batch": 128,
+  "cfg": {
+   "T_ev": 48,
+   "d": 192,
+   "enc_layers": 3,
+   "gate_bias": 0.0,
+   "heads": 6,
+   "mode": "bind",
+   "n_slots": 192,
+   "vocab": 8192
+  },
+  "group": "experiment2_stageA",
+  "kind": "bind",
+  "lr": 0.0003,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 4385515,
+  "role": "Stage A, binding-local writes, seed 1",
+  "run": "E3A_bind_s1",
+  "seed": 1,
+  "source_run_directory": "E3A_bind_s1",
+  "state_dict_tensor_total": 4385515,
+  "train_args": {
+   "batch": 128,
+   "eval_lifetimes": 400,
+   "lr": 0.0003,
+   "mode": "bind",
+   "run": "E3A_bind_s1",
+   "seed": 1,
+   "updates": 2600
+  },
+  "train_corpus": "e3_train",
+  "updates": 2600
+ },
+ "E3A_bind_s2": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "E3A_bind_s2",
+  "batch": 128,
+  "cfg": {
+   "T_ev": 48,
+   "d": 192,
+   "enc_layers": 3,
+   "gate_bias": 0.0,
+   "heads": 6,
+   "mode": "bind",
+   "n_slots": 192,
+   "vocab": 8192
+  },
+  "group": "experiment2_stageA",
+  "kind": "bind",
+  "lr": 0.0003,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 4385515,
+  "role": "Stage A, binding-local writes, seed 2",
+  "run": "E3A_bind_s2",
+  "seed": 2,
+  "source_run_directory": "E3A_bind_s2",
+  "state_dict_tensor_total": 4385515,
+  "train_args": {
+   "batch": 128,
+   "eval_lifetimes": 400,
+   "lr": 0.0003,
+   "mode": "bind",
+   "run": "E3A_bind_s2",
+   "seed": 2,
+   "updates": 2600
+  },
+  "train_corpus": "e3_train",
+  "updates": 2600
+ },
+ "E3A_bind_s3": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "E3A_bind_s3",
+  "batch": 128,
+  "cfg": {
+   "T_ev": 48,
+   "d": 192,
+   "enc_layers": 3,
+   "gate_bias": 0.0,
+   "heads": 6,
+   "mode": "bind",
+   "n_slots": 192,
+   "vocab": 8192
+  },
+  "group": "experiment2_stageA",
+  "kind": "bind",
+  "lr": 0.0003,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 4385515,
+  "role": "Stage A, binding-local writes, seed 3",
+  "run": "E3A_bind_s3",
+  "seed": 3,
+  "source_run_directory": "E3A_bind_s3",
+  "state_dict_tensor_total": 4385515,
+  "train_args": {
+   "batch": 128,
+   "eval_lifetimes": 400,
+   "lr": 0.0003,
+   "mode": "bind",
+   "run": "E3A_bind_s3",
+   "seed": 3,
+   "updates": 2600
+  },
+  "train_corpus": "e3_train",
+  "updates": 2600
+ },
+ "E3A_unbound_s1": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "E3A_unbound_s1",
+  "batch": 128,
+  "cfg": {
+   "T_ev": 48,
+   "d": 192,
+   "enc_layers": 3,
+   "gate_bias": 0.0,
+   "heads": 6,
+   "mode": "unbound",
+   "n_slots": 192,
+   "vocab": 8192
+  },
+  "group": "experiment2_stageA",
+  "kind": "unbound",
+  "lr": 0.0003,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 4385515,
+  "role": "Stage A, capacity-matched unstructured control, seed 1",
+  "run": "E3A_unbound_s1",
+  "seed": 1,
+  "source_run_directory": "E3A_unbound_s1",
+  "state_dict_tensor_total": 4385515,
+  "train_args": {
+   "batch": 128,
+   "eval_lifetimes": 400,
+   "lr": 0.0003,
+   "mode": "unbound",
+   "run": "E3A_unbound_s1",
+   "seed": 1,
+   "updates": 2600
+  },
+  "train_corpus": "e3_train",
+  "updates": 2600
+ },
+ "E3A_unbound_s2": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "E3A_unbound_s2",
+  "batch": 128,
+  "cfg": {
+   "T_ev": 48,
+   "d": 192,
+   "enc_layers": 3,
+   "gate_bias": 0.0,
+   "heads": 6,
+   "mode": "unbound",
+   "n_slots": 192,
+   "vocab": 8192
+  },
+  "group": "experiment2_stageA",
+  "kind": "unbound",
+  "lr": 0.0003,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 4385515,
+  "role": "Stage A, capacity-matched unstructured control, seed 2",
+  "run": "E3A_unbound_s2",
+  "seed": 2,
+  "source_run_directory": "E3A_unbound_s2",
+  "state_dict_tensor_total": 4385515,
+  "train_args": {
+   "batch": 128,
+   "eval_lifetimes": 400,
+   "lr": 0.0003,
+   "mode": "unbound",
+   "run": "E3A_unbound_s2",
+   "seed": 2,
+   "updates": 2600
+  },
+  "train_corpus": "e3_train",
+  "updates": 2600
+ },
+ "E3A_unbound_s3": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "E3A_unbound_s3",
+  "batch": 128,
+  "cfg": {
+   "T_ev": 48,
+   "d": 192,
+   "enc_layers": 3,
+   "gate_bias": 0.0,
+   "heads": 6,
+   "mode": "unbound",
+   "n_slots": 192,
+   "vocab": 8192
+  },
+  "group": "experiment2_stageA",
+  "kind": "unbound",
+  "lr": 0.0003,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 4385515,
+  "role": "Stage A, capacity-matched unstructured control, seed 3",
+  "run": "E3A_unbound_s3",
+  "seed": 3,
+  "source_run_directory": "E3A_unbound_s3",
+  "state_dict_tensor_total": 4385515,
+  "train_args": {
+   "batch": 128,
+   "eval_lifetimes": 400,
+   "lr": 0.0003,
+   "mode": "unbound",
+   "run": "E3A_unbound_s3",
+   "seed": 3,
+   "updates": 2600
+  },
+  "train_corpus": "e3_train",
+  "updates": 2600
+ },
+ "E3B_bind_s1": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "E3B_bind_s1",
+  "batch": 96,
+  "cfg": {
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 8,
+   "full_heads": true,
+   "gate_bias": 0.0,
+   "heads": 8,
+   "mode": "bind",
+   "n_slots": 192,
+   "vocab": 8192
+  },
+  "group": "experiment2_stageB",
+  "kind": "bind",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24265440,
+  "role": "Stage B, binding-local writes, seed 1",
+  "run": "E3B_bind_s1",
+  "seed": 1,
+  "source_run_directory": "E3B_bind_s1",
+  "state_dict_tensor_total": 27411168,
+  "train_args": {
+   "batch": 96,
+   "lr": 0.0004,
+   "mode": "bind",
+   "run": "E3B_bind_s1",
+   "seed": 1,
+   "updates": 6000
+  },
+  "train_corpus": "e3_train",
+  "updates": 6000
+ },
+ "E3B_bind_s2": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "E3B_bind_s2",
+  "batch": 96,
+  "cfg": {
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 8,
+   "full_heads": true,
+   "gate_bias": 0.0,
+   "heads": 8,
+   "mode": "bind",
+   "n_slots": 192,
+   "vocab": 8192
+  },
+  "group": "experiment2_stageB",
+  "kind": "bind",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24265440,
+  "role": "Stage B, binding-local writes, seed 2",
+  "run": "E3B_bind_s2",
+  "seed": 2,
+  "source_run_directory": "E3B_bind_s2",
+  "state_dict_tensor_total": 27411168,
+  "train_args": {
+   "batch": 96,
+   "lr": 0.0004,
+   "mode": "bind",
+   "run": "E3B_bind_s2",
+   "seed": 2,
+   "updates": 6000
+  },
+  "train_corpus": "e3_train",
+  "updates": 6000
+ },
+ "E3B_bind_s3": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "E3B_bind_s3",
+  "batch": 96,
+  "cfg": {
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 8,
+   "full_heads": true,
+   "gate_bias": 0.0,
+   "heads": 8,
+   "mode": "bind",
+   "n_slots": 192,
+   "vocab": 8192
+  },
+  "group": "experiment2_stageB",
+  "kind": "bind",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24265440,
+  "role": "Stage B, binding-local writes, seed 3",
+  "run": "E3B_bind_s3",
+  "seed": 3,
+  "source_run_directory": "E3B_bind_s3",
+  "state_dict_tensor_total": 27411168,
+  "train_args": {
+   "batch": 96,
+   "lr": 0.0004,
+   "mode": "bind",
+   "run": "E3B_bind_s3",
+   "seed": 3,
+   "updates": 6000
+  },
+  "train_corpus": "e3_train",
+  "updates": 6000
+ },
+ "E3B_unbound_s1": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "E3B_unbound_s1",
+  "batch": 96,
+  "cfg": {
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 8,
+   "full_heads": true,
+   "gate_bias": 0.0,
+   "heads": 8,
+   "mode": "unbound",
+   "n_slots": 192,
+   "vocab": 8192
+  },
+  "group": "experiment2_stageB",
+  "kind": "unbound",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24265440,
+  "role": "Stage B, capacity/parameter-matched unstructured control, seed 1",
+  "run": "E3B_unbound_s1",
+  "seed": 1,
+  "source_run_directory": "E3B_unbound_s1",
+  "state_dict_tensor_total": 27411168,
+  "train_args": {
+   "batch": 96,
+   "lr": 0.0004,
+   "mode": "unbound",
+   "run": "E3B_unbound_s1",
+   "seed": 1,
+   "updates": 6000
+  },
+  "train_corpus": "e3_train",
+  "updates": 6000
+ },
+ "E3B_unbound_s2": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "E3B_unbound_s2",
+  "batch": 96,
+  "cfg": {
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 8,
+   "full_heads": true,
+   "gate_bias": 0.0,
+   "heads": 8,
+   "mode": "unbound",
+   "n_slots": 192,
+   "vocab": 8192
+  },
+  "group": "experiment2_stageB",
+  "kind": "unbound",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24265440,
+  "role": "Stage B, capacity/parameter-matched unstructured control, seed 2",
+  "run": "E3B_unbound_s2",
+  "seed": 2,
+  "source_run_directory": "E3B_unbound_s2",
+  "state_dict_tensor_total": 27411168,
+  "train_args": {
+   "batch": 96,
+   "lr": 0.0004,
+   "mode": "unbound",
+   "run": "E3B_unbound_s2",
+   "seed": 2,
+   "updates": 6000
+  },
+  "train_corpus": "e3_train",
+  "updates": 6000
+ },
+ "E3B_unbound_s3": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "E3B_unbound_s3",
+  "batch": 96,
+  "cfg": {
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 8,
+   "full_heads": true,
+   "gate_bias": 0.0,
+   "heads": 8,
+   "mode": "unbound",
+   "n_slots": 192,
+   "vocab": 8192
+  },
+  "group": "experiment2_stageB",
+  "kind": "unbound",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24265440,
+  "role": "Stage B, capacity/parameter-matched unstructured control, seed 3",
+  "run": "E3B_unbound_s3",
+  "seed": 3,
+  "source_run_directory": "E3B_unbound_s3",
+  "state_dict_tensor_total": 27411168,
+  "train_args": {
+   "batch": 96,
+   "lr": 0.0004,
+   "mode": "unbound",
+   "run": "E3B_unbound_s3",
+   "seed": 3,
+   "updates": 6000
+  },
+  "train_corpus": "e3_train",
+  "updates": 6000
+ },
+ "G_CONVEX_FREE": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "G_CONVEX_FREE",
+  "batch": 160,
+  "cfg": {
+   "K": 4,
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 7,
+   "heads": 8,
+   "n_workspace": 4,
+   "state_tau": 16.0,
+   "update_rule": "convex",
+   "vocab": 8192,
+   "write_mode": "independent",
+   "write_temp": 1.0
+  },
+  "group": "experiment1",
+  "kind": "pnsr",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24614453,
+  "role": "E2-GATED arm: convex gated update, no clamp",
+  "run": "G_CONVEX_FREE",
+  "seed": 1,
+  "source_run_directory": "G_CONVEX_FREE",
+  "state_dict_tensor_total": 27760181,
+  "train_args": {
+   "batch": 160,
+   "gate_bias": 0.0,
+   "hold_frac": null,
+   "lr": 0.0004,
+   "model": "pnsr",
+   "resume_from": null,
+   "run": "G_CONVEX_FREE",
+   "seed": 1,
+   "state_tau": 16.0,
+   "update_rule": "convex",
+   "updates": 6000,
+   "w_enum": 2.0,
+   "w_other": 1.0
+  },
+  "train_corpus": "train",
+  "updates": 6000
+ },
+ "G_CONVEX_RETAIN": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "G_CONVEX_RETAIN",
+  "batch": 160,
+  "cfg": {
+   "K": 4,
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 7,
+   "heads": 8,
+   "n_workspace": 4,
+   "state_tau": 16.0,
+   "update_rule": "convex",
+   "vocab": 8192,
+   "write_mode": "independent",
+   "write_temp": 1.0
+  },
+  "group": "experiment1",
+  "kind": "pnsr",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24614453,
+  "role": "E2-GATED arm: convex gated update, retention-biased",
+  "run": "G_CONVEX_RETAIN",
+  "seed": 1,
+  "source_run_directory": "G_CONVEX_RETAIN",
+  "state_dict_tensor_total": 27760181,
+  "train_args": {
+   "batch": 160,
+   "gate_bias": -2.0,
+   "hold_frac": null,
+   "lr": 0.0004,
+   "model": "pnsr",
+   "resume_from": null,
+   "run": "G_CONVEX_RETAIN",
+   "seed": 1,
+   "update_rule": "convex",
+   "updates": 6000,
+   "w_enum": 2.0,
+   "w_other": 1.0
+  },
+  "train_corpus": "train",
+  "updates": 6000
+ },
+ "G_TAU512": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "G_TAU512",
+  "batch": 160,
+  "cfg": {
+   "K": 4,
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 7,
+   "heads": 8,
+   "n_workspace": 4,
+   "state_tau": 512.0,
+   "update_rule": "additive_clamp",
+   "vocab": 8192,
+   "write_mode": "independent",
+   "write_temp": 1.0
+  },
+  "group": "experiment1",
+  "kind": "pnsr",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24614453,
+  "role": "isolating control: same additive rule, clamp raised 16 -> 512",
+  "run": "G_TAU512",
+  "seed": 1,
+  "source_run_directory": "G_TAU512",
+  "state_dict_tensor_total": 27760181,
+  "train_args": {
+   "batch": 160,
+   "gate_bias": 0.0,
+   "hold_frac": null,
+   "lr": 0.0004,
+   "model": "pnsr",
+   "resume_from": null,
+   "run": "G_TAU512",
+   "seed": 1,
+   "state_tau": 512.0,
+   "update_rule": "additive_clamp",
+   "updates": 6000,
+   "w_enum": 2.0,
+   "w_other": 1.0
+  },
+  "train_corpus": "train",
+  "updates": 6000
+ },
+ "PNSR_K1_fin": {
+  "admission_note": "amendment A6: held at stage 1 for a protocol-matched comparison with PNSR_K4_fin",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "PNSR_K1_v2",
+  "batch": 160,
+  "cfg": {
+   "K": 1,
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 7,
+   "heads": 8,
+   "n_workspace": 4,
+   "state_tau": 16.0,
+   "vocab": 8192
+  },
+  "group": "experiment1",
+  "kind": "pnsr_k1",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24614453,
+  "role": "deliberation-depth control K=1 (E5)",
+  "run": "PNSR_K1_fin",
+  "seed": 1,
+  "source_run_directory": "PNSR_K1_fin",
+  "state_dict_tensor_total": 27760181,
+  "train_args": {
+   "batch": 160,
+   "lr": 0.0004,
+   "model": "pnsr_k1",
+   "run": "PNSR_K1_v2",
+   "seed": 1,
+   "updates": 6000
+  },
+  "train_corpus": "train",
+  "updates": 6000
+ },
+ "PNSR_K4_fin": {
+  "admission_note": "amendment A6: the 8x-enum stage-2 continuation failed its own preregistered admission rule, so the stage-1 checkpoint is the admitted one",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "PNSR_K4_v2",
+  "batch": 160,
+  "cfg": {
+   "K": 4,
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 7,
+   "heads": 8,
+   "n_workspace": 4,
+   "state_tau": 16.0,
+   "vocab": 8192
+  },
+  "group": "experiment1",
+  "kind": "pnsr",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24614453,
+  "role": "PNSR typed learned state, seed 1 (additive write, clamp tau=16)",
+  "run": "PNSR_K4_fin",
+  "seed": 1,
+  "source_run_directory": "PNSR_K4_fin",
+  "state_dict_tensor_total": 27760181,
+  "train_args": {
+   "batch": 160,
+   "lr": 0.0004,
+   "model": "pnsr",
+   "run": "PNSR_K4_v2",
+   "seed": 1,
+   "updates": 6000
+  },
+  "train_corpus": "train",
+  "updates": 6000
+ },
+ "PNSR_K4_s2": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "PNSR_K4_s2",
+  "batch": 160,
+  "cfg": {
+   "K": 4,
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 7,
+   "heads": 8,
+   "n_workspace": 4,
+   "state_tau": 16.0,
+   "update_rule": "additive_clamp",
+   "vocab": 8192,
+   "write_mode": "independent",
+   "write_temp": 1.0
+  },
+  "group": "experiment1",
+  "kind": "pnsr",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24614453,
+  "role": "PNSR typed learned state, seed 2",
+  "run": "PNSR_K4_s2",
+  "seed": 2,
+  "source_run_directory": "PNSR_K4_s2",
+  "state_dict_tensor_total": 27760181,
+  "train_args": {
+   "batch": 160,
+   "gate_bias": 0.0,
+   "hold_frac": null,
+   "lr": 0.0004,
+   "model": "pnsr",
+   "resume_from": null,
+   "run": "PNSR_K4_s2",
+   "seed": 2,
+   "state_tau": 16.0,
+   "update_rule": "additive_clamp",
+   "updates": 6000,
+   "w_enum": 2.0,
+   "w_other": 1.0
+  },
+  "train_corpus": "train",
+  "updates": 6000
+ },
+ "PNSR_K4_s3": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "PNSR_K4_s3",
+  "batch": 160,
+  "cfg": {
+   "K": 4,
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 7,
+   "heads": 8,
+   "n_workspace": 4,
+   "state_tau": 16.0,
+   "update_rule": "additive_clamp",
+   "vocab": 8192,
+   "write_mode": "independent",
+   "write_temp": 1.0
+  },
+  "group": "experiment1",
+  "kind": "pnsr",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24614453,
+  "role": "PNSR typed learned state, seed 3",
+  "run": "PNSR_K4_s3",
+  "seed": 3,
+  "source_run_directory": "PNSR_K4_s3",
+  "state_dict_tensor_total": 27760181,
+  "train_args": {
+   "batch": 160,
+   "gate_bias": 0.0,
+   "hold_frac": null,
+   "lr": 0.0004,
+   "model": "pnsr",
+   "resume_from": null,
+   "run": "PNSR_K4_s3",
+   "seed": 3,
+   "state_tau": 16.0,
+   "update_rule": "additive_clamp",
+   "updates": 6000,
+   "w_enum": 2.0,
+   "w_other": 1.0
+  },
+  "train_corpus": "train",
+  "updates": 6000
+ },
+ "PNSR_ON_E3": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "PNSR_ON_E3",
+  "batch": 96,
+  "cfg": {
+   "K": 4,
+   "T_ev": 48,
+   "d": 384,
+   "enc_layers": 7,
+   "heads": 8,
+   "n_workspace": 4,
+   "state_tau": 16.0,
+   "update_rule": "additive_clamp",
+   "vocab": 8192,
+   "write_mode": "independent",
+   "write_temp": 1.0
+  },
+  "group": "corrected_corpus",
+  "kind": "pnsr",
+  "lr": 0.0004,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 24614453,
+  "role": "E1 architecture retrained from scratch on the corrected corpus",
+  "run": "PNSR_ON_E3",
+  "seed": 1,
+  "source_run_directory": "PNSR_ON_E3",
+  "state_dict_tensor_total": 27760181,
+  "train_args": {
+   "batch": 96,
+   "gate_bias": 0.0,
+   "hold_frac": null,
+   "lr": 0.0004,
+   "model": "pnsr",
+   "resume_from": null,
+   "run": "PNSR_ON_E3",
+   "seed": 1,
+   "shards_split": "e3_train",
+   "state_tau": 16.0,
+   "update_rule": "additive_clamp",
+   "updates": 6000,
+   "w_enum": 2.0,
+   "w_other": 1.0
+  },
+  "train_corpus": "e3_train",
+  "updates": 6000
+ },
+ "RMT_s1": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "RMT_s1",
+  "batch": 160,
+  "cfg": {
+   "T_ev": 48,
+   "d": 384,
+   "heads": 8,
+   "layers": 10,
+   "n_mem": 32,
+   "vocab": 8192
+  },
+  "group": "experiment1",
+  "kind": "rmt",
+  "lr": 0.0003,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 23945525,
+  "role": "memory-token recurrent baseline (RMT-style), seed 1",
+  "run": "RMT_s1",
+  "seed": 1,
+  "source_run_directory": "RMT_s1",
+  "state_dict_tensor_total": 27091253,
+  "train_args": {
+   "batch": 160,
+   "gate_bias": 0.0,
+   "hold_frac": null,
+   "lr": 0.0003,
+   "model": "rmt",
+   "resume_from": null,
+   "run": "RMT_s1",
+   "seed": 1,
+   "state_tau": 16.0,
+   "update_rule": "additive_clamp",
+   "updates": 6000,
+   "w_enum": 2.0,
+   "w_other": 1.0
+  },
+  "train_corpus": "train",
+  "updates": 6000
+ },
+ "TX768_fin": {
+  "admission_note": "amendment A7: v2/v3 invalid (v3 was frozen mid-run by the A3 skip-guard, AUDIT 2); v4 retrained with the guard removed, lr 1.5e-4, exposure-matched; best.pt (update 18250) is the admitted checkpoint",
+  "admitted_checkpoint": "best.pt",
+  "admitted_from_run": "TX768_v4",
+  "batch": 384,
+  "cfg": {
+   "d": 384,
+   "heads": 8,
+   "layers": 10,
+   "vocab": 8192,
+   "window": 768
+  },
+  "group": "experiment1",
+  "kind": "tx768",
+  "lr": 0.00015,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 23892917,
+  "role": "transcript-window baseline, 768 tokens (~40 events)",
+  "run": "TX768_fin",
+  "seed": 1,
+  "source_run_directory": "TX768_fin",
+  "state_dict_tensor_total": 27038645,
+  "train_args": {
+   "batch": 384,
+   "hold_frac": 0.0,
+   "lr": 0.00015,
+   "model": "tx768",
+   "resume_from": null,
+   "run": "TX768_v4",
+   "seed": 1,
+   "updates": 18850,
+   "w_enum": 2.0,
+   "w_other": 1.0
+  },
+  "train_corpus": "train",
+  "updates": 18850
+ },
+ "TXE_ON_E3": {
+  "admission_note": "the run's own final checkpoint",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "TXE_ON_E3",
+  "batch": 384,
+  "cfg": {
+   "d": 384,
+   "heads": 8,
+   "layers": 10,
+   "vocab": 8192,
+   "window": 64
+  },
+  "group": "corrected_corpus",
+  "kind": "txe",
+  "lr": 0.0003,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 23622581,
+  "role": "memory-free control on the corrected corpus (Experiment 2 comparator)",
+  "run": "TXE_ON_E3",
+  "seed": 1,
+  "source_run_directory": "TXE_ON_E3",
+  "state_dict_tensor_total": 26768309,
+  "train_args": {
+   "batch": 384,
+   "gate_bias": 0.0,
+   "hold_frac": null,
+   "lr": 0.0003,
+   "model": "txe",
+   "resume_from": null,
+   "run": "TXE_ON_E3",
+   "seed": 1,
+   "shards_split": "e3_train",
+   "state_tau": 16.0,
+   "update_rule": "additive_clamp",
+   "updates": 20000,
+   "w_enum": 2.0,
+   "w_other": 1.0
+  },
+  "train_corpus": "e3_train",
+  "updates": 20000
+ },
+ "TXE_fin": {
+  "admission_note": "TXE_v2c diverged and was killed before completion (amendment A5 auto-fail)",
+  "admitted_checkpoint": "final.pt",
+  "admitted_from_run": "TXE_v2",
+  "batch": 384,
+  "cfg": {
+   "d": 384,
+   "heads": 8,
+   "layers": 10,
+   "vocab": 8192,
+   "window": 64
+  },
+  "group": "experiment1",
+  "kind": "txe",
+  "lr": 0.0003,
+  "param_count_note": "params = distinct trainable parameters (the paper's matched-budget figure). The event encoder and record encoder share one tied 8192x384 token embedding, so state_dict lists it twice; state_dict_tensor_total is that raw sum and is ~3.15M larger for the 25M models.",
+  "params": 23622581,
+  "role": "memory-free control: current event only",
+  "run": "TXE_fin",
+  "seed": 1,
+  "source_run_directory": "TXE_fin",
+  "state_dict_tensor_total": 26768309,
+  "train_args": {
+   "batch": 384,
+   "lr": 0.0003,
+   "model": "txe",
+   "run": "TXE_v2",
+   "seed": 1,
+   "updates": 20000
+  },
+  "train_corpus": "train",
+  "updates": 20000
+ }
+}