Add files using upload-large-folder tool

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/evaluations/S007/reward.json

@@ -0,0 +1,7 @@
+{
+  "faced": 0.4566559312936125,
+  "tuev": 0.6132832954348078,
+  "ad65": 0.726984126984127,
+  "openbmi": 0.78,
+  "longitudinal": 0.7630311481021734
+}

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/evaluations/S032/reward.json

@@ -0,0 +1,7 @@
+{
+  "faced": 0.48081052066559316,
+  "tuev": 0.6265397511057696,
+  "ad65": 0.6126984126984126,
+  "openbmi": 0.7790909090909092,
+  "longitudinal": 0.7649947533498158
+}

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/evaluations/S039/reward.json

@@ -0,0 +1,7 @@
+{
+  "faced": 0.48081052066559316,
+  "tuev": 0.6301147498183727,
+  "ad65": 0.6126984126984126,
+  "openbmi": 0.7790909090909092,
+  "longitudinal": 0.7649947533498158
+}

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/evaluations/S050/reward.json

@@ -0,0 +1,7 @@
+{
+  "faced": 0.48054213633923776,
+  "tuev": 0.6301147498183727,
+  "ad65": 0.6126984126984126,
+  "openbmi": 0.78,
+  "longitudinal": 0.7649978993280152
+}

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/evaluations/S053/report.json

@@ -0,0 +1,435 @@
+{
+  "schema": "bci-shared-v2",
+  "aggregate": null,
+  "scope": "full",
+  "tracks": {
+    "faced": {
+      "class_counts": [
+        207,
+        207,
+        207,
+        207,
+        276,
+        207,
+        207,
+        207,
+        207
+      ],
+      "predicted_class_counts": [
+        204,
+        174,
+        214,
+        248,
+        273,
+        190,
+        222,
+        198,
+        209
+      ],
+      "confusion_matrix": [
+        [
+          89,
+          14,
+          12,
+          17,
+          31,
+          9,
+          13,
+          11,
+          11
+        ],
+        [
+          10,
+          94,
+          11,
+          15,
+          32,
+          8,
+          12,
+          10,
+          15
+        ],
+        [
+          14,
+          5,
+          114,
+          16,
+          14,
+          13,
+          13,
+          8,
+          10
+        ],
+        [
+          15,
+          11,
+          9,
+          98,
+          23,
+          8,
+          16,
+          12,
+          15
+        ],
+        [
+          19,
+          23,
+          19,
+          42,
+          107,
+          14,
+          14,
+          21,
+          17
+        ],
+        [
+          10,
+          9,
+          14,
+          15,
+          15,
+          101,
+          13,
+          16,
+          14
+        ],
+        [
+          15,
+          7,
+          10,
+          13,
+          20,
+          17,
+          109,
+          5,
+          11
+        ],
+        [
+          13,
+          5,
+          10,
+          15,
+          15,
+          16,
+          12,
+          108,
+          13
+        ],
+        [
+          19,
+          6,
+          15,
+          17,
+          16,
+          4,
+          20,
+          7,
+          103
+        ]
+      ],
+      "per_class_recall": [
+        0.42995169082125606,
+        0.45410628019323673,
+        0.5507246376811594,
+        0.47342995169082125,
+        0.38768115942028986,
+        0.48792270531400966,
+        0.5265700483091788,
+        0.5217391304347826,
+        0.4975845410628019
+      ],
+      "balanced_accuracy": 0.48107890499194844,
+      "accuracy": 0.4777432712215321,
+      "weighted_f1": 0.478296255960509,
+      "macro_f1": 0.4815739085627028,
+      "kappa": 0.41174755866491763,
+      "samples": 1932,
+      "reward": 0.48107890499194844,
+      "subjects": 23,
+      "status": "ok",
+      "seconds": 10.132745108101517,
+      "sample_count": 1932,
+      "subject_count": 23,
+      "data_identity": "2eed85a6c50199b05cd777532a71116daafee58535a6d512cdac5c71cb20b3d4",
+      "package_sha256": "33fbf74b0c2ab74803edb96f25671b37e0640d10825ebbb3a2354c3caec93cc2"
+    },
+    "tuev": {
+      "class_counts": [
+        567,
+        4677,
+        1998,
+        329,
+        2204,
+        19646
+      ],
+      "predicted_class_counts": [
+        1111,
+        4047,
+        1615,
+        701,
+        4252,
+        17695
+      ],
+      "confusion_matrix": [
+        [
+          24,
+          263,
+          51,
+          108,
+          53,
+          68
+        ],
+        [
+          426,
+          3247,
+          229,
+          21,
+          222,
+          532
+        ],
+        [
+          255,
+          233,
+          1166,
+          84,
+          121,
+          139
+        ],
+        [
+          0,
+          0,
+          0,
+          311,
+          8,
+          10
+        ],
+        [
+          28,
+          6,
+          8,
+          69,
+          1488,
+          605
+        ],
+        [
+          378,
+          298,
+          161,
+          108,
+          2360,
+          16341
+        ]
+      ],
+      "per_class_recall": [
+        0.042328042328042326,
+        0.6942484498610221,
+        0.5835835835835835,
+        0.9452887537993921,
+        0.6751361161524501,
+        0.8317723709661
+      ],
+      "balanced_accuracy": 0.6287262194484317,
+      "accuracy": 0.7673770436083069,
+      "weighted_f1": 0.788441552016608,
+      "macro_f1": 0.559752801221674,
+      "kappa": 0.5853199311193578,
+      "samples": 29421,
+      "reward": 0.6287262194484317,
+      "subjects": 1,
+      "status": "ok",
+      "seconds": 28.602270467905328,
+      "sample_count": 29421,
+      "subject_count": 1,
+      "data_identity": "585d232b461e24820e1f9d45a7b6d7667e93c8f0db76d54875aa7c17ef23546a",
+      "package_sha256": "33fbf74b0c2ab74803edb96f25671b37e0640d10825ebbb3a2354c3caec93cc2"
+    },
+    "ad65": {
+      "class_counts": [
+        559,
+        357,
+        460
+      ],
+      "predicted_class_counts": [
+        446,
+        490,
+        440
+      ],
+      "confusion_matrix": [
+        [
+          275,
+          189,
+          95
+        ],
+        [
+          74,
+          204,
+          79
+        ],
+        [
+          97,
+          97,
+          266
+        ]
+      ],
+      "per_class_recall": [
+        0.4919499105545617,
+        0.5714285714285714,
+        0.5782608695652174
+      ],
+      "balanced_accuracy": 0.5472131171827835,
+      "accuracy": 0.5414244186046512,
+      "weighted_f1": 0.5449116651670216,
+      "macro_f1": 0.5400249702556351,
+      "kappa": 0.31457009059532715,
+      "samples": 1376,
+      "subject_metrics": {
+        "class_counts": [
+          7,
+          5,
+          6
+        ],
+        "predicted_class_counts": [
+          6,
+          6,
+          6
+        ],
+        "confusion_matrix": [
+          [
+            4,
+            2,
+            1
+          ],
+          [
+            1,
+            3,
+            1
+          ],
+          [
+            1,
+            1,
+            4
+          ]
+        ],
+        "per_class_recall": [
+          0.5714285714285714,
+          0.6,
+          0.6666666666666666
+        ],
+        "balanced_accuracy": 0.6126984126984126,
+        "accuracy": 0.6111111111111112,
+        "weighted_f1": 0.6130536130536131,
+        "macro_f1": 0.6091686091686092,
+        "kappa": 0.41666666666666663
+      },
+      "reward": 0.6126984126984126,
+      "subjects": 18,
+      "status": "ok",
+      "seconds": 9.776732075959444,
+      "sample_count": 1376,
+      "subject_count": 18,
+      "data_identity": "86749905e7423d4dbe2df8119caca8c3f21e0562b85ace01cc834572bff2ed4e",
+      "package_sha256": "33fbf74b0c2ab74803edb96f25671b37e0640d10825ebbb3a2354c3caec93cc2"
+    },
+    "openbmi": {
+      "class_counts": [
+        2200,
+        2200
+      ],
+      "predicted_class_counts": [
+        1958,
+        2442
+      ],
+      "confusion_matrix": [
+        [
+          1596,
+          604
+        ],
+        [
+          362,
+          1838
+        ]
+      ],
+      "per_class_recall": [
+        0.7254545454545455,
+        0.8354545454545454
+      ],
+      "balanced_accuracy": 0.7804545454545455,
+      "accuracy": 0.7804545454545454,
+      "weighted_f1": 0.7797884053808223,
+      "macro_f1": 0.7797884053808224,
+      "kappa": 0.5609090909090909,
+      "samples": 4400,
+      "subject_scores": {
+        "17": 0.725,
+        "01": 0.83,
+        "54": 0.6875,
+        "42": 0.7975,
+        "41": 0.7375,
+        "25": 0.88,
+        "30": 0.765,
+        "53": 0.725,
+        "14": 0.7375,
+        "43": 0.915,
+        "12": 0.7849999999999999
+      },
+      "reward": 0.7804545454545454,
+      "subjects": 11,
+      "status": "ok",
+      "seconds": 13.00368627300486,
+      "sample_count": 4400,
+      "subject_count": 11,
+      "data_identity": "7554d17ea4063ebae857d9bf791958338e34b2fe03f829f900aac0d16bb5e4cf",
+      "package_sha256": "33fbf74b0c2ab74803edb96f25671b37e0640d10825ebbb3a2354c3caec93cc2"
+    },
+    "longitudinal": {
+      "class_counts": [
+        4991,
+        5037
+      ],
+      "predicted_class_counts": [
+        6119,
+        3909
+      ],
+      "confusion_matrix": [
+        [
+          4357,
+          634
+        ],
+        [
+          1762,
+          3275
+        ]
+      ],
+      "per_class_recall": [
+        0.872971348427169,
+        0.650188604327973
+      ],
+      "balanced_accuracy": 0.761579976377571,
+      "accuracy": 0.7610690067810132,
+      "weighted_f1": 0.7581349676879711,
+      "macro_f1": 0.758254618218361,
+      "kappa": 0.5226206111188796,
+      "samples": 10028,
+      "subject_scores": {
+        "sub-G2022": 0.8416859958109955,
+        "sub-G2025": 0.542940781644146,
+        "sub-G2005": 0.8614230241653094,
+        "sub-G2026": 0.5215820730478056,
+        "sub-CTR004": 0.8877609732040112,
+        "sub-CTR001": 0.8132398668390053,
+        "sub-G2013": 0.8491150130745333,
+        "sub-G2012": 0.7866749526491074,
+        "sub-G2009": 0.7805584135172235
+      },
+      "reward": 0.7649978993280152,
+      "subjects": 9,
+      "status": "ok",
+      "seconds": 17.499888956081122,
+      "sample_count": 10028,
+      "subject_count": 9,
+      "data_identity": "96ec1f12076f070de8a25e225023645b13e751ccc507615cb4045d357519d742",
+      "package_sha256": "33fbf74b0c2ab74803edb96f25671b37e0640d10825ebbb3a2354c3caec93cc2"
+    }
+  },
+  "package_sha256": "33fbf74b0c2ab74803edb96f25671b37e0640d10825ebbb3a2354c3caec93cc2"
+}

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/evaluations/S055/report.json

@@ -0,0 +1,435 @@
+{
+  "schema": "bci-shared-v2",
+  "aggregate": null,
+  "scope": "full",
+  "tracks": {
+    "faced": {
+      "class_counts": [
+        207,
+        207,
+        207,
+        207,
+        276,
+        207,
+        207,
+        207,
+        207
+      ],
+      "predicted_class_counts": [
+        204,
+        174,
+        214,
+        248,
+        273,
+        190,
+        222,
+        198,
+        209
+      ],
+      "confusion_matrix": [
+        [
+          89,
+          14,
+          12,
+          17,
+          31,
+          9,
+          13,
+          11,
+          11
+        ],
+        [
+          10,
+          94,
+          11,
+          15,
+          32,
+          8,
+          12,
+          10,
+          15
+        ],
+        [
+          14,
+          5,
+          114,
+          16,
+          14,
+          13,
+          13,
+          8,
+          10
+        ],
+        [
+          15,
+          11,
+          9,
+          98,
+          23,
+          8,
+          16,
+          12,
+          15
+        ],
+        [
+          19,
+          23,
+          19,
+          42,
+          107,
+          14,
+          14,
+          21,
+          17
+        ],
+        [
+          10,
+          9,
+          14,
+          15,
+          15,
+          101,
+          13,
+          16,
+          14
+        ],
+        [
+          15,
+          7,
+          10,
+          13,
+          20,
+          17,
+          109,
+          5,
+          11
+        ],
+        [
+          13,
+          5,
+          10,
+          15,
+          15,
+          16,
+          12,
+          108,
+          13
+        ],
+        [
+          19,
+          6,
+          15,
+          17,
+          16,
+          4,
+          20,
+          7,
+          103
+        ]
+      ],
+      "per_class_recall": [
+        0.42995169082125606,
+        0.45410628019323673,
+        0.5507246376811594,
+        0.47342995169082125,
+        0.38768115942028986,
+        0.48792270531400966,
+        0.5265700483091788,
+        0.5217391304347826,
+        0.4975845410628019
+      ],
+      "balanced_accuracy": 0.48107890499194844,
+      "accuracy": 0.4777432712215321,
+      "weighted_f1": 0.478296255960509,
+      "macro_f1": 0.4815739085627028,
+      "kappa": 0.41174755866491763,
+      "samples": 1932,
+      "reward": 0.48107890499194844,
+      "subjects": 23,
+      "status": "ok",
+      "seconds": 9.76316216820851,
+      "sample_count": 1932,
+      "subject_count": 23,
+      "data_identity": "2eed85a6c50199b05cd777532a71116daafee58535a6d512cdac5c71cb20b3d4",
+      "package_sha256": "a87957715d812e9ccc3fa7167cc98d310ab0cdf1976bb2a25b8487d1b4e6caad"
+    },
+    "tuev": {
+      "class_counts": [
+        567,
+        4677,
+        1998,
+        329,
+        2204,
+        19646
+      ],
+      "predicted_class_counts": [
+        1111,
+        4047,
+        1615,
+        701,
+        4252,
+        17695
+      ],
+      "confusion_matrix": [
+        [
+          24,
+          263,
+          51,
+          108,
+          53,
+          68
+        ],
+        [
+          426,
+          3247,
+          229,
+          21,
+          222,
+          532
+        ],
+        [
+          255,
+          233,
+          1166,
+          84,
+          121,
+          139
+        ],
+        [
+          0,
+          0,
+          0,
+          311,
+          8,
+          10
+        ],
+        [
+          28,
+          6,
+          8,
+          69,
+          1488,
+          605
+        ],
+        [
+          378,
+          298,
+          161,
+          108,
+          2360,
+          16341
+        ]
+      ],
+      "per_class_recall": [
+        0.042328042328042326,
+        0.6942484498610221,
+        0.5835835835835835,
+        0.9452887537993921,
+        0.6751361161524501,
+        0.8317723709661
+      ],
+      "balanced_accuracy": 0.6287262194484317,
+      "accuracy": 0.7673770436083069,
+      "weighted_f1": 0.788441552016608,
+      "macro_f1": 0.559752801221674,
+      "kappa": 0.5853199311193578,
+      "samples": 29421,
+      "reward": 0.6287262194484317,
+      "subjects": 1,
+      "status": "ok",
+      "seconds": 27.513580065919086,
+      "sample_count": 29421,
+      "subject_count": 1,
+      "data_identity": "585d232b461e24820e1f9d45a7b6d7667e93c8f0db76d54875aa7c17ef23546a",
+      "package_sha256": "a87957715d812e9ccc3fa7167cc98d310ab0cdf1976bb2a25b8487d1b4e6caad"
+    },
+    "ad65": {
+      "class_counts": [
+        559,
+        357,
+        460
+      ],
+      "predicted_class_counts": [
+        446,
+        490,
+        440
+      ],
+      "confusion_matrix": [
+        [
+          275,
+          189,
+          95
+        ],
+        [
+          74,
+          204,
+          79
+        ],
+        [
+          97,
+          97,
+          266
+        ]
+      ],
+      "per_class_recall": [
+        0.4919499105545617,
+        0.5714285714285714,
+        0.5782608695652174
+      ],
+      "balanced_accuracy": 0.5472131171827835,
+      "accuracy": 0.5414244186046512,
+      "weighted_f1": 0.5449116651670216,
+      "macro_f1": 0.5400249702556351,
+      "kappa": 0.31457009059532715,
+      "samples": 1376,
+      "subject_metrics": {
+        "class_counts": [
+          7,
+          5,
+          6
+        ],
+        "predicted_class_counts": [
+          6,
+          6,
+          6
+        ],
+        "confusion_matrix": [
+          [
+            4,
+            2,
+            1
+          ],
+          [
+            1,
+            3,
+            1
+          ],
+          [
+            1,
+            1,
+            4
+          ]
+        ],
+        "per_class_recall": [
+          0.5714285714285714,
+          0.6,
+          0.6666666666666666
+        ],
+        "balanced_accuracy": 0.6126984126984126,
+        "accuracy": 0.6111111111111112,
+        "weighted_f1": 0.6130536130536131,
+        "macro_f1": 0.6091686091686092,
+        "kappa": 0.41666666666666663
+      },
+      "reward": 0.6126984126984126,
+      "subjects": 18,
+      "status": "ok",
+      "seconds": 7.737647537142038,
+      "sample_count": 1376,
+      "subject_count": 18,
+      "data_identity": "86749905e7423d4dbe2df8119caca8c3f21e0562b85ace01cc834572bff2ed4e",
+      "package_sha256": "a87957715d812e9ccc3fa7167cc98d310ab0cdf1976bb2a25b8487d1b4e6caad"
+    },
+    "openbmi": {
+      "class_counts": [
+        2200,
+        2200
+      ],
+      "predicted_class_counts": [
+        1958,
+        2442
+      ],
+      "confusion_matrix": [
+        [
+          1596,
+          604
+        ],
+        [
+          362,
+          1838
+        ]
+      ],
+      "per_class_recall": [
+        0.7254545454545455,
+        0.8354545454545454
+      ],
+      "balanced_accuracy": 0.7804545454545455,
+      "accuracy": 0.7804545454545454,
+      "weighted_f1": 0.7797884053808223,
+      "macro_f1": 0.7797884053808224,
+      "kappa": 0.5609090909090909,
+      "samples": 4400,
+      "subject_scores": {
+        "17": 0.725,
+        "01": 0.83,
+        "54": 0.6875,
+        "42": 0.7975,
+        "41": 0.7375,
+        "25": 0.88,
+        "30": 0.765,
+        "53": 0.725,
+        "14": 0.7375,
+        "43": 0.915,
+        "12": 0.7849999999999999
+      },
+      "reward": 0.7804545454545454,
+      "subjects": 11,
+      "status": "ok",
+      "seconds": 11.886369728948921,
+      "sample_count": 4400,
+      "subject_count": 11,
+      "data_identity": "7554d17ea4063ebae857d9bf791958338e34b2fe03f829f900aac0d16bb5e4cf",
+      "package_sha256": "a87957715d812e9ccc3fa7167cc98d310ab0cdf1976bb2a25b8487d1b4e6caad"
+    },
+    "longitudinal": {
+      "class_counts": [
+        4991,
+        5037
+      ],
+      "predicted_class_counts": [
+        6119,
+        3909
+      ],
+      "confusion_matrix": [
+        [
+          4357,
+          634
+        ],
+        [
+          1762,
+          3275
+        ]
+      ],
+      "per_class_recall": [
+        0.872971348427169,
+        0.650188604327973
+      ],
+      "balanced_accuracy": 0.761579976377571,
+      "accuracy": 0.7610690067810132,
+      "weighted_f1": 0.7581349676879711,
+      "macro_f1": 0.758254618218361,
+      "kappa": 0.5226206111188796,
+      "samples": 10028,
+      "subject_scores": {
+        "sub-G2022": 0.8416859958109955,
+        "sub-G2025": 0.542940781644146,
+        "sub-G2005": 0.8614230241653094,
+        "sub-G2026": 0.5215820730478056,
+        "sub-CTR004": 0.8877609732040112,
+        "sub-CTR001": 0.8132398668390053,
+        "sub-G2013": 0.8491150130745333,
+        "sub-G2012": 0.7866749526491074,
+        "sub-G2009": 0.7805584135172235
+      },
+      "reward": 0.7649978993280152,
+      "subjects": 9,
+      "status": "ok",
+      "seconds": 14.212455272907391,
+      "sample_count": 10028,
+      "subject_count": 9,
+      "data_identity": "96ec1f12076f070de8a25e225023645b13e751ccc507615cb4045d357519d742",
+      "package_sha256": "a87957715d812e9ccc3fa7167cc98d310ab0cdf1976bb2a25b8487d1b4e6caad"
+    }
+  },
+  "package_sha256": "a87957715d812e9ccc3fa7167cc98d310ab0cdf1976bb2a25b8487d1b4e6caad"
+}

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/evaluations/S059/reward.json

@@ -0,0 +1,7 @@
+{
+  "faced": 0.48550724637681164,
+  "tuev": 0.6281560533875664,
+  "ad65": 0.6126984126984126,
+  "openbmi": 0.7804545454545454,
+  "longitudinal": 0.7649978993280152
+}

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/evaluations/S060/report.json

@@ -0,0 +1,435 @@
+{
+  "schema": "bci-shared-v2",
+  "aggregate": null,
+  "scope": "full",
+  "tracks": {
+    "faced": {
+      "class_counts": [
+        207,
+        207,
+        207,
+        207,
+        276,
+        207,
+        207,
+        207,
+        207
+      ],
+      "predicted_class_counts": [
+        200,
+        178,
+        217,
+        244,
+        273,
+        194,
+        223,
+        197,
+        206
+      ],
+      "confusion_matrix": [
+        [
+          89,
+          15,
+          12,
+          16,
+          31,
+          9,
+          13,
+          11,
+          11
+        ],
+        [
+          10,
+          96,
+          11,
+          14,
+          31,
+          9,
+          12,
+          9,
+          15
+        ],
+        [
+          13,
+          5,
+          114,
+          16,
+          14,
+          14,
+          13,
+          8,
+          10
+        ],
+        [
+          15,
+          10,
+          9,
+          98,
+          24,
+          8,
+          16,
+          12,
+          15
+        ],
+        [
+          19,
+          24,
+          20,
+          41,
+          106,
+          14,
+          14,
+          21,
+          17
+        ],
+        [
+          10,
+          9,
+          15,
+          15,
+          13,
+          104,
+          11,
+          16,
+          14
+        ],
+        [
+          14,
+          8,
+          11,
+          13,
+          19,
+          16,
+          111,
+          5,
+          10
+        ],
+        [
+          12,
+          5,
+          10,
+          14,
+          16,
+          16,
+          13,
+          108,
+          13
+        ],
+        [
+          18,
+          6,
+          15,
+          17,
+          19,
+          4,
+          20,
+          7,
+          101
+        ]
+      ],
+      "per_class_recall": [
+        0.42995169082125606,
+        0.463768115942029,
+        0.5507246376811594,
+        0.47342995169082125,
+        0.38405797101449274,
+        0.5024154589371981,
+        0.5362318840579711,
+        0.5217391304347826,
+        0.48792270531400966
+      ],
+      "balanced_accuracy": 0.4833601717659689,
+      "accuracy": 0.4798136645962733,
+      "weighted_f1": 0.48021308679093977,
+      "macro_f1": 0.4836966585812809,
+      "kappa": 0.41407958023611724,
+      "samples": 1932,
+      "reward": 0.4833601717659689,
+      "subjects": 23,
+      "status": "ok",
+      "seconds": 10.570235804188997,
+      "sample_count": 1932,
+      "subject_count": 23,
+      "data_identity": "2eed85a6c50199b05cd777532a71116daafee58535a6d512cdac5c71cb20b3d4",
+      "package_sha256": "f15a31b65d7653d6d56327f83b4ea6bfe630cc5fabac06c04c8f711b5fd4a3d5"
+    },
+    "tuev": {
+      "class_counts": [
+        567,
+        4677,
+        1998,
+        329,
+        2204,
+        19646
+      ],
+      "predicted_class_counts": [
+        1120,
+        4038,
+        1615,
+        702,
+        4251,
+        17695
+      ],
+      "confusion_matrix": [
+        [
+          24,
+          263,
+          51,
+          108,
+          53,
+          68
+        ],
+        [
+          442,
+          3231,
+          229,
+          21,
+          222,
+          532
+        ],
+        [
+          248,
+          240,
+          1166,
+          85,
+          120,
+          139
+        ],
+        [
+          0,
+          0,
+          0,
+          311,
+          8,
+          10
+        ],
+        [
+          28,
+          6,
+          8,
+          69,
+          1488,
+          605
+        ],
+        [
+          378,
+          298,
+          161,
+          108,
+          2360,
+          16341
+        ]
+      ],
+      "per_class_recall": [
+        0.042328042328042326,
+        0.6908274534958306,
+        0.5835835835835835,
+        0.9452887537993921,
+        0.6751361161524501,
+        0.8317723709661
+      ],
+      "balanced_accuracy": 0.6281560533875664,
+      "accuracy": 0.7668332143706876,
+      "weighted_f1": 0.7879759099072944,
+      "macro_f1": 0.559157797000665,
+      "kappa": 0.5843837519086473,
+      "samples": 29421,
+      "reward": 0.6281560533875664,
+      "subjects": 1,
+      "status": "ok",
+      "seconds": 38.78919044393115,
+      "sample_count": 29421,
+      "subject_count": 1,
+      "data_identity": "585d232b461e24820e1f9d45a7b6d7667e93c8f0db76d54875aa7c17ef23546a",
+      "package_sha256": "f15a31b65d7653d6d56327f83b4ea6bfe630cc5fabac06c04c8f711b5fd4a3d5"
+    },
+    "ad65": {
+      "class_counts": [
+        559,
+        357,
+        460
+      ],
+      "predicted_class_counts": [
+        446,
+        490,
+        440
+      ],
+      "confusion_matrix": [
+        [
+          275,
+          189,
+          95
+        ],
+        [
+          74,
+          204,
+          79
+        ],
+        [
+          97,
+          97,
+          266
+        ]
+      ],
+      "per_class_recall": [
+        0.4919499105545617,
+        0.5714285714285714,
+        0.5782608695652174
+      ],
+      "balanced_accuracy": 0.5472131171827835,
+      "accuracy": 0.5414244186046512,
+      "weighted_f1": 0.5449116651670216,
+      "macro_f1": 0.5400249702556351,
+      "kappa": 0.31457009059532715,
+      "samples": 1376,
+      "subject_metrics": {
+        "class_counts": [
+          7,
+          5,
+          6
+        ],
+        "predicted_class_counts": [
+          6,
+          6,
+          6
+        ],
+        "confusion_matrix": [
+          [
+            4,
+            2,
+            1
+          ],
+          [
+            1,
+            3,
+            1
+          ],
+          [
+            1,
+            1,
+            4
+          ]
+        ],
+        "per_class_recall": [
+          0.5714285714285714,
+          0.6,
+          0.6666666666666666
+        ],
+        "balanced_accuracy": 0.6126984126984126,
+        "accuracy": 0.6111111111111112,
+        "weighted_f1": 0.6130536130536131,
+        "macro_f1": 0.6091686091686092,
+        "kappa": 0.41666666666666663
+      },
+      "reward": 0.6126984126984126,
+      "subjects": 18,
+      "status": "ok",
+      "seconds": 8.523222964955494,
+      "sample_count": 1376,
+      "subject_count": 18,
+      "data_identity": "86749905e7423d4dbe2df8119caca8c3f21e0562b85ace01cc834572bff2ed4e",
+      "package_sha256": "f15a31b65d7653d6d56327f83b4ea6bfe630cc5fabac06c04c8f711b5fd4a3d5"
+    },
+    "openbmi": {
+      "class_counts": [
+        2200,
+        2200
+      ],
+      "predicted_class_counts": [
+        1958,
+        2442
+      ],
+      "confusion_matrix": [
+        [
+          1596,
+          604
+        ],
+        [
+          362,
+          1838
+        ]
+      ],
+      "per_class_recall": [
+        0.7254545454545455,
+        0.8354545454545454
+      ],
+      "balanced_accuracy": 0.7804545454545455,
+      "accuracy": 0.7804545454545454,
+      "weighted_f1": 0.7797884053808223,
+      "macro_f1": 0.7797884053808224,
+      "kappa": 0.5609090909090909,
+      "samples": 4400,
+      "subject_scores": {
+        "17": 0.725,
+        "01": 0.83,
+        "54": 0.6875,
+        "42": 0.7975,
+        "41": 0.7375,
+        "25": 0.88,
+        "30": 0.765,
+        "53": 0.725,
+        "14": 0.7375,
+        "43": 0.915,
+        "12": 0.7849999999999999
+      },
+      "reward": 0.7804545454545454,
+      "subjects": 11,
+      "status": "ok",
+      "seconds": 14.376801694044843,
+      "sample_count": 4400,
+      "subject_count": 11,
+      "data_identity": "7554d17ea4063ebae857d9bf791958338e34b2fe03f829f900aac0d16bb5e4cf",
+      "package_sha256": "f15a31b65d7653d6d56327f83b4ea6bfe630cc5fabac06c04c8f711b5fd4a3d5"
+    },
+    "longitudinal": {
+      "class_counts": [
+        4991,
+        5037
+      ],
+      "predicted_class_counts": [
+        6119,
+        3909
+      ],
+      "confusion_matrix": [
+        [
+          4357,
+          634
+        ],
+        [
+          1762,
+          3275
+        ]
+      ],
+      "per_class_recall": [
+        0.872971348427169,
+        0.650188604327973
+      ],
+      "balanced_accuracy": 0.761579976377571,
+      "accuracy": 0.7610690067810132,
+      "weighted_f1": 0.7581349676879711,
+      "macro_f1": 0.758254618218361,
+      "kappa": 0.5226206111188796,
+      "samples": 10028,
+      "subject_scores": {
+        "sub-G2022": 0.8416859958109955,
+        "sub-G2025": 0.542940781644146,
+        "sub-G2005": 0.8614230241653094,
+        "sub-G2026": 0.5215820730478056,
+        "sub-CTR004": 0.8877609732040112,
+        "sub-CTR001": 0.8132398668390053,
+        "sub-G2013": 0.8491150130745333,
+        "sub-G2012": 0.7866749526491074,
+        "sub-G2009": 0.7805584135172235
+      },
+      "reward": 0.7649978993280152,
+      "subjects": 9,
+      "status": "ok",
+      "seconds": 18.027567972196266,
+      "sample_count": 10028,
+      "subject_count": 9,
+      "data_identity": "96ec1f12076f070de8a25e225023645b13e751ccc507615cb4045d357519d742",
+      "package_sha256": "f15a31b65d7653d6d56327f83b4ea6bfe630cc5fabac06c04c8f711b5fd4a3d5"
+    }
+  },
+  "package_sha256": "f15a31b65d7653d6d56327f83b4ea6bfe630cc5fabac06c04c8f711b5fd4a3d5"
+}

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/evaluations/S064/reward.json

@@ -0,0 +1,7 @@
+{
+  "faced": 0.4966451959205583,
+  "tuev": 0.6281560533875664,
+  "ad65": 0.6126984126984126,
+  "openbmi": 0.7804545454545454,
+  "longitudinal": 0.7660747910775001
+}

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/evaluations/S065/report.json

@@ -0,0 +1,435 @@
+{
+  "schema": "bci-shared-v2",
+  "aggregate": null,
+  "scope": "full",
+  "tracks": {
+    "faced": {
+      "class_counts": [
+        207,
+        207,
+        207,
+        207,
+        276,
+        207,
+        207,
+        207,
+        207
+      ],
+      "predicted_class_counts": [
+        201,
+        177,
+        224,
+        243,
+        270,
+        197,
+        220,
+        196,
+        204
+      ],
+      "confusion_matrix": [
+        [
+          93,
+          16,
+          14,
+          15,
+          29,
+          6,
+          13,
+          10,
+          11
+        ],
+        [
+          9,
+          98,
+          15,
+          13,
+          29,
+          8,
+          13,
+          7,
+          15
+        ],
+        [
+          14,
+          4,
+          116,
+          16,
+          16,
+          13,
+          12,
+          7,
+          9
+        ],
+        [
+          15,
+          9,
+          10,
+          100,
+          21,
+          8,
+          18,
+          12,
+          14
+        ],
+        [
+          17,
+          24,
+          21,
+          39,
+          111,
+          16,
+          16,
+          19,
+          13
+        ],
+        [
+          10,
+          8,
+          12,
+          15,
+          10,
+          111,
+          11,
+          15,
+          15
+        ],
+        [
+          14,
+          7,
+          12,
+          12,
+          21,
+          17,
+          108,
+          5,
+          11
+        ],
+        [
+          11,
+          6,
+          10,
+          14,
+          16,
+          14,
+          10,
+          114,
+          12
+        ],
+        [
+          18,
+          5,
+          14,
+          19,
+          17,
+          4,
+          19,
+          7,
+          104
+        ]
+      ],
+      "per_class_recall": [
+        0.4492753623188406,
+        0.47342995169082125,
+        0.5603864734299517,
+        0.4830917874396135,
+        0.40217391304347827,
+        0.5362318840579711,
+        0.5217391304347826,
+        0.5507246376811594,
+        0.5024154589371981
+      ],
+      "balanced_accuracy": 0.4977187332259796,
+      "accuracy": 0.4943064182194617,
+      "weighted_f1": 0.4948232538359813,
+      "macro_f1": 0.4980910259560767,
+      "kappa": 0.43043930876535497,
+      "samples": 1932,
+      "reward": 0.4977187332259796,
+      "subjects": 23,
+      "status": "ok",
+      "seconds": 10.571723641012795,
+      "sample_count": 1932,
+      "subject_count": 23,
+      "data_identity": "2eed85a6c50199b05cd777532a71116daafee58535a6d512cdac5c71cb20b3d4",
+      "package_sha256": "1758b08a248e9636dafd575a0f918a795e008f0ba3a6813f7f541e62483ce41a"
+    },
+    "tuev": {
+      "class_counts": [
+        567,
+        4677,
+        1998,
+        329,
+        2204,
+        19646
+      ],
+      "predicted_class_counts": [
+        1120,
+        4038,
+        1615,
+        702,
+        4251,
+        17695
+      ],
+      "confusion_matrix": [
+        [
+          24,
+          263,
+          51,
+          108,
+          53,
+          68
+        ],
+        [
+          442,
+          3231,
+          229,
+          21,
+          222,
+          532
+        ],
+        [
+          248,
+          240,
+          1166,
+          85,
+          120,
+          139
+        ],
+        [
+          0,
+          0,
+          0,
+          311,
+          8,
+          10
+        ],
+        [
+          28,
+          6,
+          8,
+          69,
+          1488,
+          605
+        ],
+        [
+          378,
+          298,
+          161,
+          108,
+          2360,
+          16341
+        ]
+      ],
+      "per_class_recall": [
+        0.042328042328042326,
+        0.6908274534958306,
+        0.5835835835835835,
+        0.9452887537993921,
+        0.6751361161524501,
+        0.8317723709661
+      ],
+      "balanced_accuracy": 0.6281560533875664,
+      "accuracy": 0.7668332143706876,
+      "weighted_f1": 0.7879759099072944,
+      "macro_f1": 0.559157797000665,
+      "kappa": 0.5843837519086473,
+      "samples": 29421,
+      "reward": 0.6281560533875664,
+      "subjects": 1,
+      "status": "ok",
+      "seconds": 36.78829408099409,
+      "sample_count": 29421,
+      "subject_count": 1,
+      "data_identity": "585d232b461e24820e1f9d45a7b6d7667e93c8f0db76d54875aa7c17ef23546a",
+      "package_sha256": "1758b08a248e9636dafd575a0f918a795e008f0ba3a6813f7f541e62483ce41a"
+    },
+    "ad65": {
+      "class_counts": [
+        559,
+        357,
+        460
+      ],
+      "predicted_class_counts": [
+        446,
+        490,
+        440
+      ],
+      "confusion_matrix": [
+        [
+          275,
+          189,
+          95
+        ],
+        [
+          74,
+          204,
+          79
+        ],
+        [
+          97,
+          97,
+          266
+        ]
+      ],
+      "per_class_recall": [
+        0.4919499105545617,
+        0.5714285714285714,
+        0.5782608695652174
+      ],
+      "balanced_accuracy": 0.5472131171827835,
+      "accuracy": 0.5414244186046512,
+      "weighted_f1": 0.5449116651670216,
+      "macro_f1": 0.5400249702556351,
+      "kappa": 0.31457009059532715,
+      "samples": 1376,
+      "subject_metrics": {
+        "class_counts": [
+          7,
+          5,
+          6
+        ],
+        "predicted_class_counts": [
+          6,
+          6,
+          6
+        ],
+        "confusion_matrix": [
+          [
+            4,
+            2,
+            1
+          ],
+          [
+            1,
+            3,
+            1
+          ],
+          [
+            1,
+            1,
+            4
+          ]
+        ],
+        "per_class_recall": [
+          0.5714285714285714,
+          0.6,
+          0.6666666666666666
+        ],
+        "balanced_accuracy": 0.6126984126984126,
+        "accuracy": 0.6111111111111112,
+        "weighted_f1": 0.6130536130536131,
+        "macro_f1": 0.6091686091686092,
+        "kappa": 0.41666666666666663
+      },
+      "reward": 0.6126984126984126,
+      "subjects": 18,
+      "status": "ok",
+      "seconds": 8.60704336600611,
+      "sample_count": 1376,
+      "subject_count": 18,
+      "data_identity": "86749905e7423d4dbe2df8119caca8c3f21e0562b85ace01cc834572bff2ed4e",
+      "package_sha256": "1758b08a248e9636dafd575a0f918a795e008f0ba3a6813f7f541e62483ce41a"
+    },
+    "openbmi": {
+      "class_counts": [
+        2200,
+        2200
+      ],
+      "predicted_class_counts": [
+        1958,
+        2442
+      ],
+      "confusion_matrix": [
+        [
+          1596,
+          604
+        ],
+        [
+          362,
+          1838
+        ]
+      ],
+      "per_class_recall": [
+        0.7254545454545455,
+        0.8354545454545454
+      ],
+      "balanced_accuracy": 0.7804545454545455,
+      "accuracy": 0.7804545454545454,
+      "weighted_f1": 0.7797884053808223,
+      "macro_f1": 0.7797884053808224,
+      "kappa": 0.5609090909090909,
+      "samples": 4400,
+      "subject_scores": {
+        "17": 0.725,
+        "01": 0.83,
+        "54": 0.6875,
+        "42": 0.7975,
+        "41": 0.7375,
+        "25": 0.88,
+        "30": 0.765,
+        "53": 0.725,
+        "14": 0.7375,
+        "43": 0.915,
+        "12": 0.7849999999999999
+      },
+      "reward": 0.7804545454545454,
+      "subjects": 11,
+      "status": "ok",
+      "seconds": 13.94210716796806,
+      "sample_count": 4400,
+      "subject_count": 11,
+      "data_identity": "7554d17ea4063ebae857d9bf791958338e34b2fe03f829f900aac0d16bb5e4cf",
+      "package_sha256": "1758b08a248e9636dafd575a0f918a795e008f0ba3a6813f7f541e62483ce41a"
+    },
+    "longitudinal": {
+      "class_counts": [
+        4991,
+        5037
+      ],
+      "predicted_class_counts": [
+        6121,
+        3907
+      ],
+      "confusion_matrix": [
+        [
+          4363,
+          628
+        ],
+        [
+          1758,
+          3279
+        ]
+      ],
+      "per_class_recall": [
+        0.8741735123221799,
+        0.6509827278141751
+      ],
+      "balanced_accuracy": 0.7625781200681775,
+      "accuracy": 0.7620662145991225,
+      "weighted_f1": 0.7591337024741095,
+      "macro_f1": 0.7592530790738887,
+      "kappa": 0.5246138815291013,
+      "samples": 10028,
+      "subject_scores": {
+        "sub-G2022": 0.840118582288983,
+        "sub-G2025": 0.5418442904160758,
+        "sub-G2005": 0.8635202439042169,
+        "sub-G2026": 0.5222995778307584,
+        "sub-CTR004": 0.8910899227354923,
+        "sub-CTR001": 0.8148803434285684,
+        "sub-G2013": 0.8522285329172483,
+        "sub-G2012": 0.7848808804108592,
+        "sub-G2009": 0.7838107457652987
+      },
+      "reward": 0.7660747910775001,
+      "subjects": 9,
+      "status": "ok",
+      "seconds": 17.662740516010672,
+      "sample_count": 10028,
+      "subject_count": 9,
+      "data_identity": "96ec1f12076f070de8a25e225023645b13e751ccc507615cb4045d357519d742",
+      "package_sha256": "1758b08a248e9636dafd575a0f918a795e008f0ba3a6813f7f541e62483ce41a"
+    }
+  },
+  "package_sha256": "1758b08a248e9636dafd575a0f918a795e008f0ba3a6813f7f541e62483ce41a"
+}

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/evaluations/S066/report.json

@@ -0,0 +1,435 @@
+{
+  "schema": "bci-shared-v2",
+  "aggregate": null,
+  "scope": "full",
+  "tracks": {
+    "faced": {
+      "class_counts": [
+        207,
+        207,
+        207,
+        207,
+        276,
+        207,
+        207,
+        207,
+        207
+      ],
+      "predicted_class_counts": [
+        201,
+        177,
+        224,
+        243,
+        270,
+        197,
+        220,
+        196,
+        204
+      ],
+      "confusion_matrix": [
+        [
+          93,
+          16,
+          14,
+          15,
+          29,
+          6,
+          13,
+          10,
+          11
+        ],
+        [
+          9,
+          98,
+          15,
+          13,
+          29,
+          8,
+          13,
+          7,
+          15
+        ],
+        [
+          14,
+          4,
+          116,
+          16,
+          16,
+          13,
+          12,
+          7,
+          9
+        ],
+        [
+          15,
+          9,
+          10,
+          100,
+          21,
+          8,
+          18,
+          12,
+          14
+        ],
+        [
+          17,
+          24,
+          21,
+          39,
+          111,
+          16,
+          16,
+          19,
+          13
+        ],
+        [
+          10,
+          8,
+          12,
+          15,
+          10,
+          111,
+          11,
+          15,
+          15
+        ],
+        [
+          14,
+          7,
+          12,
+          12,
+          21,
+          17,
+          108,
+          5,
+          11
+        ],
+        [
+          11,
+          6,
+          10,
+          14,
+          16,
+          14,
+          10,
+          114,
+          12
+        ],
+        [
+          18,
+          5,
+          14,
+          19,
+          17,
+          4,
+          19,
+          7,
+          104
+        ]
+      ],
+      "per_class_recall": [
+        0.4492753623188406,
+        0.47342995169082125,
+        0.5603864734299517,
+        0.4830917874396135,
+        0.40217391304347827,
+        0.5362318840579711,
+        0.5217391304347826,
+        0.5507246376811594,
+        0.5024154589371981
+      ],
+      "balanced_accuracy": 0.4977187332259796,
+      "accuracy": 0.4943064182194617,
+      "weighted_f1": 0.4948232538359813,
+      "macro_f1": 0.4980910259560767,
+      "kappa": 0.43043930876535497,
+      "samples": 1932,
+      "reward": 0.4977187332259796,
+      "subjects": 23,
+      "status": "ok",
+      "seconds": 10.77297183900373,
+      "sample_count": 1932,
+      "subject_count": 23,
+      "data_identity": "2eed85a6c50199b05cd777532a71116daafee58535a6d512cdac5c71cb20b3d4",
+      "package_sha256": "1d4fb576f7947f44be8388c97a78ac46ad381a36c2e12ca56a186866c4dcc312"
+    },
+    "tuev": {
+      "class_counts": [
+        567,
+        4677,
+        1998,
+        329,
+        2204,
+        19646
+      ],
+      "predicted_class_counts": [
+        1120,
+        4038,
+        1615,
+        702,
+        4251,
+        17695
+      ],
+      "confusion_matrix": [
+        [
+          24,
+          263,
+          51,
+          108,
+          53,
+          68
+        ],
+        [
+          442,
+          3231,
+          229,
+          21,
+          222,
+          532
+        ],
+        [
+          248,
+          240,
+          1166,
+          85,
+          120,
+          139
+        ],
+        [
+          0,
+          0,
+          0,
+          311,
+          8,
+          10
+        ],
+        [
+          28,
+          6,
+          8,
+          69,
+          1488,
+          605
+        ],
+        [
+          378,
+          298,
+          161,
+          108,
+          2360,
+          16341
+        ]
+      ],
+      "per_class_recall": [
+        0.042328042328042326,
+        0.6908274534958306,
+        0.5835835835835835,
+        0.9452887537993921,
+        0.6751361161524501,
+        0.8317723709661
+      ],
+      "balanced_accuracy": 0.6281560533875664,
+      "accuracy": 0.7668332143706876,
+      "weighted_f1": 0.7879759099072944,
+      "macro_f1": 0.559157797000665,
+      "kappa": 0.5843837519086473,
+      "samples": 29421,
+      "reward": 0.6281560533875664,
+      "subjects": 1,
+      "status": "ok",
+      "seconds": 36.478118630999234,
+      "sample_count": 29421,
+      "subject_count": 1,
+      "data_identity": "585d232b461e24820e1f9d45a7b6d7667e93c8f0db76d54875aa7c17ef23546a",
+      "package_sha256": "1d4fb576f7947f44be8388c97a78ac46ad381a36c2e12ca56a186866c4dcc312"
+    },
+    "ad65": {
+      "class_counts": [
+        559,
+        357,
+        460
+      ],
+      "predicted_class_counts": [
+        446,
+        490,
+        440
+      ],
+      "confusion_matrix": [
+        [
+          275,
+          189,
+          95
+        ],
+        [
+          74,
+          204,
+          79
+        ],
+        [
+          97,
+          97,
+          266
+        ]
+      ],
+      "per_class_recall": [
+        0.4919499105545617,
+        0.5714285714285714,
+        0.5782608695652174
+      ],
+      "balanced_accuracy": 0.5472131171827835,
+      "accuracy": 0.5414244186046512,
+      "weighted_f1": 0.5449116651670216,
+      "macro_f1": 0.5400249702556351,
+      "kappa": 0.31457009059532715,
+      "samples": 1376,
+      "subject_metrics": {
+        "class_counts": [
+          7,
+          5,
+          6
+        ],
+        "predicted_class_counts": [
+          6,
+          6,
+          6
+        ],
+        "confusion_matrix": [
+          [
+            4,
+            2,
+            1
+          ],
+          [
+            1,
+            3,
+            1
+          ],
+          [
+            1,
+            1,
+            4
+          ]
+        ],
+        "per_class_recall": [
+          0.5714285714285714,
+          0.6,
+          0.6666666666666666
+        ],
+        "balanced_accuracy": 0.6126984126984126,
+        "accuracy": 0.6111111111111112,
+        "weighted_f1": 0.6130536130536131,
+        "macro_f1": 0.6091686091686092,
+        "kappa": 0.41666666666666663
+      },
+      "reward": 0.6126984126984126,
+      "subjects": 18,
+      "status": "ok",
+      "seconds": 8.99592928198399,
+      "sample_count": 1376,
+      "subject_count": 18,
+      "data_identity": "86749905e7423d4dbe2df8119caca8c3f21e0562b85ace01cc834572bff2ed4e",
+      "package_sha256": "1d4fb576f7947f44be8388c97a78ac46ad381a36c2e12ca56a186866c4dcc312"
+    },
+    "openbmi": {
+      "class_counts": [
+        2200,
+        2200
+      ],
+      "predicted_class_counts": [
+        1958,
+        2442
+      ],
+      "confusion_matrix": [
+        [
+          1596,
+          604
+        ],
+        [
+          362,
+          1838
+        ]
+      ],
+      "per_class_recall": [
+        0.7254545454545455,
+        0.8354545454545454
+      ],
+      "balanced_accuracy": 0.7804545454545455,
+      "accuracy": 0.7804545454545454,
+      "weighted_f1": 0.7797884053808223,
+      "macro_f1": 0.7797884053808224,
+      "kappa": 0.5609090909090909,
+      "samples": 4400,
+      "subject_scores": {
+        "17": 0.725,
+        "01": 0.83,
+        "54": 0.6875,
+        "42": 0.7975,
+        "41": 0.7375,
+        "25": 0.88,
+        "30": 0.765,
+        "53": 0.725,
+        "14": 0.7375,
+        "43": 0.915,
+        "12": 0.7849999999999999
+      },
+      "reward": 0.7804545454545454,
+      "subjects": 11,
+      "status": "ok",
+      "seconds": 13.920724708994385,
+      "sample_count": 4400,
+      "subject_count": 11,
+      "data_identity": "7554d17ea4063ebae857d9bf791958338e34b2fe03f829f900aac0d16bb5e4cf",
+      "package_sha256": "1d4fb576f7947f44be8388c97a78ac46ad381a36c2e12ca56a186866c4dcc312"
+    },
+    "longitudinal": {
+      "class_counts": [
+        4991,
+        5037
+      ],
+      "predicted_class_counts": [
+        6120,
+        3908
+      ],
+      "confusion_matrix": [
+        [
+          4361,
+          630
+        ],
+        [
+          1759,
+          3278
+        ]
+      ],
+      "per_class_recall": [
+        0.8737727910238429,
+        0.6507841969426246
+      ],
+      "balanced_accuracy": 0.7622784939832337,
+      "accuracy": 0.7617670522536897,
+      "weighted_f1": 0.7588362215640966,
+      "macro_f1": 0.7589556353920164,
+      "kappa": 0.5240157271618906,
+      "samples": 10028,
+      "subject_scores": {
+        "sub-G2022": 0.8392966753029181,
+        "sub-G2025": 0.5418442904160758,
+        "sub-G2005": 0.8635202439042169,
+        "sub-G2026": 0.5214880771050998,
+        "sub-CTR004": 0.8878431694887391,
+        "sub-CTR001": 0.8149225161681095,
+        "sub-G2013": 0.852293325580725,
+        "sub-G2012": 0.7840475470775259,
+        "sub-G2009": 0.784627739229351
+      },
+      "reward": 0.7655426204747512,
+      "subjects": 9,
+      "status": "ok",
+      "seconds": 17.544876814004965,
+      "sample_count": 10028,
+      "subject_count": 9,
+      "data_identity": "96ec1f12076f070de8a25e225023645b13e751ccc507615cb4045d357519d742",
+      "package_sha256": "1d4fb576f7947f44be8388c97a78ac46ad381a36c2e12ca56a186866c4dcc312"
+    }
+  },
+  "package_sha256": "1d4fb576f7947f44be8388c97a78ac46ad381a36c2e12ca56a186866c4dcc312"
+}

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/evaluations/S099/report.json

@@ -0,0 +1,435 @@
+{
+  "schema": "bci-shared-v2",
+  "aggregate": null,
+  "scope": "full",
+  "tracks": {
+    "faced": {
+      "class_counts": [
+        207,
+        207,
+        207,
+        207,
+        276,
+        207,
+        207,
+        207,
+        207
+      ],
+      "predicted_class_counts": [
+        182,
+        219,
+        210,
+        225,
+        275,
+        201,
+        239,
+        183,
+        198
+      ],
+      "confusion_matrix": [
+        [
+          91,
+          15,
+          11,
+          17,
+          28,
+          8,
+          21,
+          6,
+          10
+        ],
+        [
+          7,
+          112,
+          13,
+          14,
+          23,
+          8,
+          11,
+          4,
+          15
+        ],
+        [
+          12,
+          9,
+          118,
+          13,
+          18,
+          11,
+          10,
+          9,
+          7
+        ],
+        [
+          9,
+          18,
+          10,
+          89,
+          27,
+          10,
+          19,
+          12,
+          13
+        ],
+        [
+          22,
+          26,
+          22,
+          36,
+          112,
+          13,
+          13,
+          16,
+          16
+        ],
+        [
+          12,
+          13,
+          11,
+          12,
+          11,
+          114,
+          13,
+          13,
+          8
+        ],
+        [
+          12,
+          9,
+          7,
+          13,
+          22,
+          15,
+          118,
+          4,
+          7
+        ],
+        [
+          8,
+          7,
+          8,
+          17,
+          14,
+          14,
+          11,
+          115,
+          13
+        ],
+        [
+          9,
+          10,
+          10,
+          14,
+          20,
+          8,
+          23,
+          4,
+          109
+        ]
+      ],
+      "per_class_recall": [
+        0.4396135265700483,
+        0.5410628019323671,
+        0.5700483091787439,
+        0.42995169082125606,
+        0.4057971014492754,
+        0.5507246376811594,
+        0.5700483091787439,
+        0.5555555555555556,
+        0.5265700483091788
+      ],
+      "balanced_accuracy": 0.5099302200751477,
+      "accuracy": 0.5062111801242236,
+      "weighted_f1": 0.5067539649258648,
+      "macro_f1": 0.5104658312076522,
+      "kappa": 0.44378969286829784,
+      "samples": 1932,
+      "reward": 0.5099302200751477,
+      "subjects": 23,
+      "status": "ok",
+      "seconds": 11.372615861007944,
+      "sample_count": 1932,
+      "subject_count": 23,
+      "data_identity": "2eed85a6c50199b05cd777532a71116daafee58535a6d512cdac5c71cb20b3d4",
+      "package_sha256": "83881ee4a0b49af9e3d25542527ce1ac2881636f77f8a406b8d0a5d0b834565a"
+    },
+    "tuev": {
+      "class_counts": [
+        567,
+        4677,
+        1998,
+        329,
+        2204,
+        19646
+      ],
+      "predicted_class_counts": [
+        1099,
+        3951,
+        1669,
+        644,
+        4179,
+        17879
+      ],
+      "confusion_matrix": [
+        [
+          24,
+          263,
+          71,
+          88,
+          53,
+          68
+        ],
+        [
+          436,
+          3172,
+          228,
+          25,
+          222,
+          594
+        ],
+        [
+          233,
+          234,
+          1180,
+          86,
+          123,
+          142
+        ],
+        [
+          0,
+          0,
+          0,
+          311,
+          8,
+          10
+        ],
+        [
+          28,
+          6,
+          7,
+          48,
+          1492,
+          623
+        ],
+        [
+          378,
+          276,
+          183,
+          86,
+          2281,
+          16442
+        ]
+      ],
+      "per_class_recall": [
+        0.042328042328042326,
+        0.6782125293991875,
+        0.5905905905905906,
+        0.9452887537993921,
+        0.676950998185118,
+        0.8369133665886186
+      ],
+      "balanced_accuracy": 0.6283807134818248,
+      "accuracy": 0.7688725740117603,
+      "weighted_f1": 0.7884847833818526,
+      "macro_f1": 0.5651240184334573,
+      "kappa": 0.5853495195863888,
+      "samples": 29421,
+      "reward": 0.6283807134818248,
+      "subjects": 1,
+      "status": "ok",
+      "seconds": 48.59776473099191,
+      "sample_count": 29421,
+      "subject_count": 1,
+      "data_identity": "585d232b461e24820e1f9d45a7b6d7667e93c8f0db76d54875aa7c17ef23546a",
+      "package_sha256": "83881ee4a0b49af9e3d25542527ce1ac2881636f77f8a406b8d0a5d0b834565a"
+    },
+    "ad65": {
+      "class_counts": [
+        559,
+        357,
+        460
+      ],
+      "predicted_class_counts": [
+        446,
+        490,
+        440
+      ],
+      "confusion_matrix": [
+        [
+          275,
+          189,
+          95
+        ],
+        [
+          74,
+          204,
+          79
+        ],
+        [
+          97,
+          97,
+          266
+        ]
+      ],
+      "per_class_recall": [
+        0.4919499105545617,
+        0.5714285714285714,
+        0.5782608695652174
+      ],
+      "balanced_accuracy": 0.5472131171827835,
+      "accuracy": 0.5414244186046512,
+      "weighted_f1": 0.5449116651670216,
+      "macro_f1": 0.5400249702556351,
+      "kappa": 0.31457009059532715,
+      "samples": 1376,
+      "subject_metrics": {
+        "class_counts": [
+          7,
+          5,
+          6
+        ],
+        "predicted_class_counts": [
+          6,
+          6,
+          6
+        ],
+        "confusion_matrix": [
+          [
+            4,
+            2,
+            1
+          ],
+          [
+            1,
+            3,
+            1
+          ],
+          [
+            1,
+            1,
+            4
+          ]
+        ],
+        "per_class_recall": [
+          0.5714285714285714,
+          0.6,
+          0.6666666666666666
+        ],
+        "balanced_accuracy": 0.6126984126984126,
+        "accuracy": 0.6111111111111112,
+        "weighted_f1": 0.6130536130536131,
+        "macro_f1": 0.6091686091686092,
+        "kappa": 0.41666666666666663
+      },
+      "reward": 0.6126984126984126,
+      "subjects": 18,
+      "status": "ok",
+      "seconds": 9.610758816997986,
+      "sample_count": 1376,
+      "subject_count": 18,
+      "data_identity": "86749905e7423d4dbe2df8119caca8c3f21e0562b85ace01cc834572bff2ed4e",
+      "package_sha256": "83881ee4a0b49af9e3d25542527ce1ac2881636f77f8a406b8d0a5d0b834565a"
+    },
+    "openbmi": {
+      "class_counts": [
+        2200,
+        2200
+      ],
+      "predicted_class_counts": [
+        1923,
+        2477
+      ],
+      "confusion_matrix": [
+        [
+          1584,
+          616
+        ],
+        [
+          339,
+          1861
+        ]
+      ],
+      "per_class_recall": [
+        0.72,
+        0.8459090909090909
+      ],
+      "balanced_accuracy": 0.7829545454545455,
+      "accuracy": 0.7829545454545455,
+      "weighted_f1": 0.7820909118582631,
+      "macro_f1": 0.7820909118582631,
+      "kappa": 0.5659090909090909,
+      "samples": 4400,
+      "subject_scores": {
+        "17": 0.685,
+        "01": 0.7775,
+        "54": 0.735,
+        "42": 0.7825,
+        "41": 0.7475,
+        "25": 0.8925000000000001,
+        "30": 0.7825,
+        "53": 0.745,
+        "14": 0.745,
+        "43": 0.9,
+        "12": 0.82
+      },
+      "reward": 0.7829545454545456,
+      "subjects": 11,
+      "status": "ok",
+      "seconds": 15.334442339997622,
+      "sample_count": 4400,
+      "subject_count": 11,
+      "data_identity": "7554d17ea4063ebae857d9bf791958338e34b2fe03f829f900aac0d16bb5e4cf",
+      "package_sha256": "83881ee4a0b49af9e3d25542527ce1ac2881636f77f8a406b8d0a5d0b834565a"
+    },
+    "longitudinal": {
+      "class_counts": [
+        4991,
+        5037
+      ],
+      "predicted_class_counts": [
+        6200,
+        3828
+      ],
+      "confusion_matrix": [
+        [
+          4518,
+          473
+        ],
+        [
+          1682,
+          3355
+        ]
+      ],
+      "per_class_recall": [
+        0.905229412943298,
+        0.666071074052015
+      ],
+      "balanced_accuracy": 0.7856502434976564,
+      "accuracy": 0.7851017151974472,
+      "weighted_f1": 0.7820559857019868,
+      "macro_f1": 0.7821718695372153,
+      "kappa": 0.5706692693911197,
+      "samples": 10028,
+      "subject_scores": {
+        "sub-G2022": 0.8352376889770471,
+        "sub-G2025": 0.5619448738031315,
+        "sub-G2005": 0.8879750029665812,
+        "sub-G2026": 0.5529888004922688,
+        "sub-CTR004": 0.9326401446654611,
+        "sub-CTR001": 0.8180924552462331,
+        "sub-G2013": 0.8822652454271831,
+        "sub-G2012": 0.8020374673299534,
+        "sub-G2009": 0.839794545605719
+      },
+      "reward": 0.7903306916126198,
+      "subjects": 9,
+      "status": "ok",
+      "seconds": 21.817366016999586,
+      "sample_count": 10028,
+      "subject_count": 9,
+      "data_identity": "96ec1f12076f070de8a25e225023645b13e751ccc507615cb4045d357519d742",
+      "package_sha256": "83881ee4a0b49af9e3d25542527ce1ac2881636f77f8a406b8d0a5d0b834565a"
+    }
+  },
+  "package_sha256": "83881ee4a0b49af9e3d25542527ce1ac2881636f77f8a406b8d0a5d0b834565a"
+}

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/evaluations/S123/reward.json

@@ -0,0 +1,7 @@
+{
+  "faced": 0.504428341384863,
+  "tuev": 0.6286386210442109,
+  "ad65": 0.6126984126984126,
+  "openbmi": 0.7856818181818181,
+  "longitudinal": 0.7919130634585465
+}

tasks/brain-computer-interface-decoding/brain-computer-interface-decoding-astra-xhigh-1h100-48h-20261031a/submissions/S114/private-head-refresh-audit.json

@@ -0,0 +1,26 @@
+{
+  "reference": "/workspace/research/sources/emotion_private_input_source",
+  "donor": "/workspace/research/runs/emotion_input_adapt_full1/epoch-001",
+  "track": "longitudinal",
+  "source_sha256": {
+    "reference": "90904cd9b33d59b1ed531dcda90f95b9dce3e59661187242b6b26a06daba6d04",
+    "donor": "b4a858b3dab746a6600aa1fd194e7cc42b277fd11eb07d8561aff980f3a26b9d"
+  },
+  "changed_private_tensors": [
+    "adapters.longitudinal.emotion_projection.weight",
+    "adapters.longitudinal.emotion_projection.bias",
+    "heads.longitudinal.correction_gain",
+    "heads.longitudinal.projection.weight",
+    "heads.longitudinal.projection.bias"
+  ],
+  "all_common_old_input_normalizer_and_prior_head_tensors_exactly_equal": true,
+  "executable_code_and_architecture_exactly_equal": true,
+  "one_common_checkpoint": true,
+  "prediction_sources": {
+    "faced": "/workspace/research/runs/emotion_geometry_log_full1/faced-e002.npy",
+    "tuev": "/workspace/research/runs/emotion_geometry_log_full1/tuev-e002.npy",
+    "ad65": "/workspace/research/runs/emotion_geometry_log_full1/ad65-e002.npy",
+    "openbmi": "/workspace/research/runs/emotion_geometry_log_full1/openbmi-e002.npy",
+    "longitudinal": "/workspace/research/runs/emotion_input_adapt_full1/longitudinal-e001.npy"
+  }
+}

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/evaluations/S002/report.json

@@ -0,0 +1,75 @@
+{
+  "reward": 0.6911869614182907,
+  "schema": "peer-protein-report/v1",
+  "tasks": {
+    "beta_lactamase": {
+      "invalid_reason": null,
+      "reward": 0.8095661912700146,
+      "spearman": 0.8095661912700146
+    },
+    "binary_localization": {
+      "accuracy": 0.9205260148656375,
+      "balanced_accuracy": 0.9161364250566587,
+      "invalid_reason": null,
+      "macro_f1": 0.9182473096982013,
+      "reward": 0.9205260148656375
+    },
+    "fluorescence": {
+      "invalid_reason": null,
+      "reward": 0.6558959121791416,
+      "spearman": 0.6558959121791416
+    },
+    "fold": {
+      "accuracy": 0.23676880222841226,
+      "balanced_accuracy": 0.12646084511893338,
+      "invalid_reason": null,
+      "macro_f1": 0.013960067963442246,
+      "reward": 0.23676880222841226
+    },
+    "human_ppi": {
+      "accuracy": 0.8565400843881856,
+      "balanced_accuracy": 0.8570360347528843,
+      "invalid_reason": null,
+      "macro_f1": 0.8546536796536797,
+      "reward": 0.8565400843881856
+    },
+    "ppi_affinity": {
+      "invalid_reason": null,
+      "reward": 0.4861481483877837,
+      "rmse": 2.2237373560522875,
+      "spearman": 0.7174256511356671
+    },
+    "secondary_structure": {
+      "accuracy": 0.737632802124834,
+      "invalid_reason": null,
+      "reward": 0.737632802124834,
+      "valid_residues": 84336
+    },
+    "solubility": {
+      "accuracy": 0.7208604302151076,
+      "balanced_accuracy": 0.7208953953953954,
+      "invalid_reason": null,
+      "macro_f1": 0.7195042095416277,
+      "reward": 0.7208604302151076
+    },
+    "stability": {
+      "invalid_reason": null,
+      "reward": 0.8075337855787583,
+      "spearman": 0.8075337855787583
+    },
+    "subcellular_localization": {
+      "accuracy": 0.7446808510638298,
+      "balanced_accuracy": 0.5933066609923021,
+      "invalid_reason": null,
+      "macro_f1": 0.6033775936507972,
+      "reward": 0.7446808510638298
+    },
+    "yeast_ppi": {
+      "accuracy": 0.6269035532994924,
+      "balanced_accuracy": 0.6141859562912194,
+      "invalid_reason": null,
+      "macro_f1": 0.6028307515069227,
+      "reward": 0.6269035532994924
+    }
+  }
+}

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/evaluations/S010/report.json

@@ -0,0 +1,75 @@
+{
+  "reward": 0.6987365785255946,
+  "schema": "peer-protein-report/v1",
+  "tasks": {
+    "beta_lactamase": {
+      "invalid_reason": null,
+      "reward": 0.8149698701773106,
+      "spearman": 0.8149698701773106
+    },
+    "binary_localization": {
+      "accuracy": 0.9199542595769011,
+      "balanced_accuracy": 0.9158153367814647,
+      "invalid_reason": null,
+      "macro_f1": 0.9177073951093605,
+      "reward": 0.9199542595769011
+    },
+    "fluorescence": {
+      "invalid_reason": null,
+      "reward": 0.677462264993902,
+      "spearman": 0.677462264993902
+    },
+    "fold": {
+      "accuracy": 0.2479108635097493,
+      "balanced_accuracy": 0.1430335799821094,
+      "invalid_reason": null,
+      "macro_f1": 0.01581663882894986,
+      "reward": 0.2479108635097493
+    },
+    "human_ppi": {
+      "accuracy": 0.8607594936708861,
+      "balanced_accuracy": 0.8609172482552343,
+      "invalid_reason": null,
+      "macro_f1": 0.8606006594777649,
+      "reward": 0.8607594936708861
+    },
+    "ppi_affinity": {
+      "invalid_reason": null,
+      "reward": 0.5293729841014578,
+      "rmse": 1.961110326805573,
+      "spearman": 0.7905729059489769
+    },
+    "secondary_structure": {
+      "accuracy": 0.7341111743502182,
+      "invalid_reason": null,
+      "reward": 0.7341111743502182,
+      "valid_residues": 84336
+    },
+    "solubility": {
+      "accuracy": 0.7338669334667334,
+      "balanced_accuracy": 0.7339174174174175,
+      "invalid_reason": null,
+      "macro_f1": 0.7311487511376277,
+      "reward": 0.7338669334667334
+    },
+    "stability": {
+      "invalid_reason": null,
+      "reward": 0.8159854178422349,
+      "spearman": 0.8159854178422349
+    },
+    "subcellular_localization": {
+      "accuracy": 0.7324197619906239,
+      "balanced_accuracy": 0.5845469578891337,
+      "invalid_reason": null,
+      "macro_f1": 0.594284103393939,
+      "reward": 0.7324197619906239
+    },
+    "yeast_ppi": {
+      "accuracy": 0.6192893401015228,
+      "balanced_accuracy": 0.6153885943359627,
+      "invalid_reason": null,
+      "macro_f1": 0.6153245333125049,
+      "reward": 0.6192893401015228
+    }
+  }
+}

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S001/README.md

@@ -0,0 +1,50 @@
+# Unified PEER protein model
+
+One shared 12-layer, 384-channel, 6-head rotary encoder serves all eleven tasks.
+Exact counts: **21,306,271 shared encoder parameters and
+26,245,478 total parameters**, plus 769 fit-normalization buffer
+values. All learned weights originate from random initialization and training
+in this session. No external weights, embeddings or data are used.
+
+The seed-42 encoder received four active hours of MLM on mounted UniRef50 (9,127,835,498 input tokens). Seed-1 joint supervision on all eleven fit splits then ran for 30 active minutes, learning the shared ordered-bin projection. A complete-state checkpoint average was followed by accepted fit-only beta raw-label residual, affinity ridge and GFP rank readouts. The package averages four complete checkpoints with the same
+.25 coefficient for every tensor, selected at fixed time cutoffs. Tune selects
+checkpoints and subsequent fit-only readouts. Full ancestry and selection
+records are in training_history.json.
+
+Shared pooling combines mean, attention, maximum and terminal summaries;
+pair tasks add a shared product/difference projection. All global and pair tasks also use one shared projection of 8 ordered residue-bin summaries. Classification tasks
+use MLP heads, and secondary structure uses a residue MLP. Stability retains
+its scalar MLP; affinity uses a fitted final ridge layer. Beta uses a raw-label ridge
+plus a 64-unit residual. GFP uses a rank linear layer plus the mutation
+correction described below. Rank-scale output tasks: fluorescence.
+Original-label-scale output tasks: beta_lactamase, ppi_affinity, stability.
+
+Mutation corrections are installed for fluorescence.
+Each adds 770 linear coefficients over existing penultimate features, log
+mutation count, and summed contextual differences at substituted residues.
+The same shared encoder processes input and reference. Fixed references
+(fluorescence: 237 residues) are modal fit-length consensuses computed without labels or
+tune data and stored in configuration. Other-length inputs retain their
+original predictions. No input-keyed prediction table, feature cache or
+inference-time fitting is used. The small correction accumulates in float64
+over float32 features/coefficients; original encoder/head precision is preserved.
+
+Selection reviews the equal-weight mean and every task. Against the previous publication, five task point rewards rise and six decline. GFP, fold and secondary structure have positive paired intervals; the overall mean interval spans zero. This is not a Pareto improvement.
+
+The mutation study compared nine fixed ridge strengths with an existing-feature
+and mutation-count control. Adoption required at least .01 task-reward gain
+and positive paired intervals versus both parent and control. Full independent replay gives
+**0.713445 mean tune reward**, with all eleven tasks valid. This is visible
+tune evidence; repeated selection and homology dependence limit inference.
+
+Global/pair inputs truncate at 1022 residues. Secondary structure covers the
+whole chain with overlapping 1022-residue windows. Equal length does not imply
+alignment for unrelated proteins. Most same-length GFP tune substitutions
+already occur in fit, and no such tune chains have four or more substitutions;
+distant-mutant generalization remains unmeasured. Rare folds, low-activity beta
+variants and the small PPI tune sets remain limitations. Any subgroup diagnostics
+are research evidence only and do not route predictions.
+In the lowest fit-defined fluorescence quartile (1361 tune rows), conditional Spearman changes from 0.526670 to 0.533851; its paired change interval is [-0.01966, 0.03171]. In GFP examples with at least one unseen substitution (55 tune rows), conditional Spearman changes from 0.770130 to 0.761255; its paired change interval is [-0.06943, 0.04213]. These conditional correlations do not decompose the global score and are not used for routing predictions.
+
+Run offline: `bash infer.sh INPUT_ROOT OUTPUT_ROOT`. The package is independent
+of the trainer and loads one model.pt for all eleven tasks.

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S001/config.json

@@ -0,0 +1,33 @@
+{
+  "name": "unified_t12_21M",
+  "layers": 12,
+  "dim": 384,
+  "heads": 6,
+  "ordered_pooling_bins": 8,
+  "structured_pooling": true,
+  "full_chain_ss": true,
+  "full_precision_head_tasks": [
+    "fluorescence",
+    "stability",
+    "beta_lactamase",
+    "ppi_affinity"
+  ],
+  "weight_variant": "budget_snapshot_average",
+  "full_precision_tasks": [
+    "beta_lactamase",
+    "fluorescence",
+    "ppi_affinity"
+  ],
+  "residual_readouts": {
+    "beta_lactamase": 64
+  },
+  "rank_target_tasks": [
+    "fluorescence"
+  ],
+  "mutation_readouts": {
+    "fluorescence": {
+      "reference": "SKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLKFICTTGKLPVPWPTLVTTLSYGVQCFSRYPDHMKQHDFFKSAMPEGYVQERTIFFKDDGNYKTRAEVKFEGDTLVNRIELKGIDFKEDGNILGHKLEYNYNSHNVYIMADKQKNGIKVNFKIRHNIEDGSVQLADHYQQNTPIGDGPVLLPDNHYLSTQSALSKDPNEKRDHMVLLEFVTAAGITHGMDELYK",
+      "feature_policy": "Sum shared contextual substitution deltas; same-length only; fit-reference encoded by the same backbone."
+    }
+  }
+}

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S001/mutation_readout_runtime.py

@@ -0,0 +1,71 @@
+"""Inference-only contextual substitution residuals over one shared encoder."""
+from types import MethodType
+import torch
+from torch import nn
+import torch.nn.functional as F
+from esm2 import encode
+
+
+class MutationResidual(nn.Module):
+    def __init__(self, original_head, dim):
+        super().__init__()
+        self.original_head = original_head
+        self.dim = dim
+        self.correction = nn.Linear(dim*2+1, 1)
+        nn.init.zeros_(self.correction.weight)
+        nn.init.zeros_(self.correction.bias)
+
+    def forward(self, representation):
+        d = self.dim
+        original = representation[:, :d]
+        delta = representation[:, d:2*d]
+        count = representation[:, 2*d:2*d+1]
+        eligible = representation[:, 2*d+1:2*d+2].bool()
+        features = self.original_head[:-1](original)
+        base = self.original_head[-1](features)
+        joined = torch.cat([features, count, delta], dim=-1)
+        correction = F.linear(joined.double(), self.correction.weight.double(),
+                              self.correction.bias.double()).to(base.dtype)
+        return torch.where(eligible, base+correction, base)
+
+
+def _represent(self, task, batch, device, max_len=1024):
+    if task not in self._mutation_references:
+        return self._mutation_original_represent(task, batch, device, max_len)
+    reference_ids = self._mutation_references[task]
+    reference_length = len(reference_ids)-2
+    lengths = [len(row['sequence']) for row in batch]
+    captured = []
+    hook = self.encoder.register_forward_hook(lambda module, inputs, output: captured.append((inputs[0], output[0])))
+    try:
+        original = self._mutation_original_represent(task, batch, device, max_len)
+    finally:
+        hook.remove()
+    assert len(captured) == 1
+    tokens, hidden = captured[0]
+    eligible = torch.tensor(lengths, device=tokens.device).eq(reference_length)
+    delta = hidden.new_zeros((len(batch), hidden.shape[-1]))
+    count = hidden.new_zeros((len(batch), 1))
+    ids = eligible.nonzero(as_tuple=True)[0]
+    if len(ids):
+        reference_tokens = torch.tensor(reference_ids, device=tokens.device, dtype=torch.long)
+        reference_hidden = self.encoder(reference_tokens[None])[0][0]
+        different = tokens[ids, 1:reference_length+1].ne(reference_tokens[1:-1])
+        changes = hidden[ids, 1:reference_length+1]-reference_hidden[None, 1:-1]
+        delta[ids] = (changes*different[:, :, None]).sum(1)
+        count[ids, 0] = different.sum(1).to(hidden.dtype)
+    return torch.cat([original, delta, count.log1p(), eligible[:, None].to(hidden.dtype)], dim=-1)
+
+
+def install_mutation_readouts(model, specification):
+    assert not hasattr(model, '_mutation_references')
+    assert set(specification) <= {'fluorescence', 'beta_lactamase'}
+    model._mutation_references = {}
+    for task, info in specification.items():
+        reference = info['reference']
+        assert 0 < len(reference) <= 1022
+        model._mutation_references[task] = encode(reference)
+        device = next(model.heads[task].parameters()).device
+        model.heads[task] = MutationResidual(model.heads[task], model.encoder.dim).to(device)
+    model._mutation_original_represent = model.represent
+    model.represent = MethodType(_represent, model)

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S001/training_history.json

@@ -0,0 +1,2090 @@
+{
+  "config": {
+    "name": "unified_t12_21M",
+    "layers": 12,
+    "dim": 384,
+    "heads": 6,
+    "ordered_pooling_bins": 8,
+    "structured_pooling": true,
+    "full_chain_ss": true,
+    "full_precision_head_tasks": [
+      "fluorescence",
+      "stability",
+      "beta_lactamase",
+      "ppi_affinity"
+    ],
+    "weight_variant": "budget_snapshot_average",
+    "full_precision_tasks": [
+      "beta_lactamase",
+      "fluorescence",
+      "ppi_affinity"
+    ],
+    "residual_readouts": {
+      "beta_lactamase": 64
+    },
+    "rank_target_tasks": [
+      "fluorescence"
+    ],
+    "mutation_readouts": {
+      "fluorescence": {
+        "reference": "SKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLKFICTTGKLPVPWPTLVTTLSYGVQCFSRYPDHMKQHDFFKSAMPEGYVQERTIFFKDDGNYKTRAEVKFEGDTLVNRIELKGIDFKEDGNILGHKLEYNYNSHNVYIMADKQKNGIKVNFKIRHNIEDGSVQLADHYQQNTPIGDGPVLLPDNHYLSTQSALSKDPNEKRDHMVLLEFVTAAGITHGMDELYK",
+        "feature_policy": "Sum shared contextual substitution deltas; same-length only; fit-reference encoded by the same backbone."
+      }
+    }
+  },
+  "notes": {
+    "ancestor": "/workspace/research/pretrain21m_extended/pretrain.pt",
+    "seed": 1,
+    "data": "fit training; tune selection only",
+    "settings": {
+      "out": "/workspace/research/21m_4h_joint_ordered_pool",
+      "resume": "/workspace/research/pretrain21m_extended/pretrain.pt",
+      "hours": 0.5,
+      "lr": 0.0001,
+      "batch_size": 32,
+      "seed": 1,
+      "eval_seconds": 300.0,
+      "mean_pooling": false,
+      "normalized_regression": true,
+      "random_batches": true,
+      "task_weights": "{\"human_ppi\":0.25}",
+      "continue_state": false,
+      "model": "unified_t12_21M",
+      "head_lr_mult": 2.0,
+      "fold_contrastive": 0.0,
+      "ema_decay": 0.999,
+      "correlation_loss": 0.1,
+      "tune_limit": 0
+    },
+    "ancestor_provenance": {
+      "step": 211159,
+      "tokens_seen": 6876918950,
+      "total_pretrain_tokens": 9127835498,
+      "provenance": {
+        "random_initialization_seed": 42,
+        "stage_seed": 43,
+        "data": "/workspace/research/corpus/manifest.json",
+        "objective": "15% masked residues; 80% mask / 10% random / 10% unchanged",
+        "sampling": "Length buckets sampled by token mass; shuffled sequence cycles without replacement within each bucket",
+        "ancestry": {
+          "checkpoint": "/workspace/research/pretrain21m/pretrain.pt",
+          "prior_tokens": 2250916548,
+          "prior_provenance": {
+            "random_initialization_seed": 42,
+            "data": "/workspace/research/corpus/manifest.json",
+            "objective": "15% masked residues; 80% mask / 10% random / 10% unchanged",
+            "sampling": "Length buckets sampled by token mass; shuffled sequence cycles without replacement within each bucket",
+            "model": {
+              "name": "unified_t12_21M",
+              "layers": 12,
+              "dim": 384,
+              "heads": 6
+            },
+            "parameters": {
+              "backbone": 21306271,
+              "total": 25040355
+            }
+          }
+        },
+        "continued_ancestor_corpus": true,
+        "stage_active_hour_budget": 3.0,
+        "stage_peak_learning_rate": 0.0002,
+        "fit_text_probability": 0.0,
+        "fit_mixture_sampling": "Independent resumable RNG; mixture decisions do not consume the UniRef corpus sampler RNG",
+        "fit_text": null,
+        "model": {
+          "name": "unified_t12_21M",
+          "layers": 12,
+          "dim": 384,
+          "heads": 6
+        },
+        "parameters": {
+          "backbone": 21306271,
+          "total": 25040355
+        }
+      },
+      "complete": true
+    },
+    "compute_by_task": {
+      "fluorescence": {
+        "updates": 5697,
+        "examples": 182304,
+        "tokens": 43566251,
+        "seconds": 162.3441452680854
+      },
+      "stability": {
+        "updates": 8790,
+        "examples": 281280,
+        "tokens": 13289595,
+        "seconds": 238.59854101459496
+      },
+      "beta_lactamase": {
+        "updates": 2418,
+        "examples": 77376,
+        "tokens": 22284288,
+        "seconds": 69.57553572778124
+      },
+      "solubility": {
+        "updates": 9491,
+        "examples": 303712,
+        "tokens": 90589330,
+        "seconds": 295.11683348717634
+      },
+      "subcellular_localization": {
+        "updates": 3449,
+        "examples": 110368,
+        "tokens": 52209242,
+        "seconds": 147.72849281225353
+      },
+      "binary_localization": {
+        "updates": 2782,
+        "examples": 89024,
+        "tokens": 40134720,
+        "seconds": 118.94762054237071
+      },
+      "fold": {
+        "updates": 4281,
+        "examples": 136992,
+        "tokens": 23170407,
+        "seconds": 122.23370583902579
+      },
+      "secondary_structure": {
+        "updates": 3578,
+        "examples": 114496,
+        "tokens": 29435955,
+        "seconds": 110.94229607644957
+      },
+      "yeast_ppi": {
+        "updates": 2693,
+        "examples": 86176,
+        "tokens": 88063633,
+        "seconds": 269.12609137850814
+      },
+      "human_ppi": {
+        "updates": 1805,
+        "examples": 57760,
+        "tokens": 58388137,
+        "seconds": 171.54262299148832
+      },
+      "ppi_affinity": {
+        "updates": 1843,
+        "examples": 58976,
+        "tokens": 21239463,
+        "seconds": 93.84522322157864
+      }
+    },
+    "exposure_scope": "Top-level compute_by_task and auxiliary_mlm describe the last listed averaging parent, not the full selection trajectory. Exact cumulative exposure for each parent is recorded with its averaging entry; do not add repeated parents as independent training.",
+    "selection_training_extent": {
+      "active_seconds": 1800.0011083593126,
+      "steps": 46827,
+      "note": "Model selection inspected the full trajectory; averaged ancestors may come from earlier steps."
+    },
+    "weight_averaging": {
+      "policy": "One fixed equal average of COMPLETE best-mean checkpoints available by 600, 900, 1200, 1800 active seconds. Repeated parents are allowed when no later checkpoint improves the mean. Every tensor uses the same coefficients; no per-task mixtures. Policy chosen before evaluating the averaged predictor.",
+      "parents": [
+        {
+          "source": "/workspace/research/21m_4h_joint_ordered_pool/snapshots/step_15165_ema/package",
+          "cutoff_active_seconds": 600,
+          "step": 15165,
+          "active_seconds": 600.1121590716066,
+          "variant": "ema",
+          "weight": 0.25,
+          "sha256": "2b7c440f0bf409f0e814e34f7e1ba58984b3cd3021394ea756d3464c3f51436a",
+          "compute_by_task": {
+            "fluorescence": {
+              "updates": 1885,
+              "examples": 60320,
+              "tokens": 14415340,
+              "seconds": 55.9825758022489
+            },
+            "stability": {
+              "updates": 2868,
+              "examples": 91776,
+              "tokens": 4336945,
+              "seconds": 81.44490410888102
+            },
+            "beta_lactamase": {
+              "updates": 780,
+              "examples": 24960,
+              "tokens": 7188480,
+              "seconds": 23.228994634351693
+            },
+            "solubility": {
+              "updates": 3124,
+              "examples": 99968,
+              "tokens": 29801574,
+              "seconds": 100.85432327119634
+            },
+            "subcellular_localization": {
+              "updates": 1072,
+              "examples": 34304,
+              "tokens": 16146588,
+              "seconds": 47.07126678409986
+            },
+            "binary_localization": {
+              "updates": 890,
+              "examples": 28480,
+              "tokens": 13019550,
+              "seconds": 39.50789823872037
+            },
+            "fold": {
+              "updates": 1345,
+              "examples": 43040,
+              "tokens": 7339851,
+              "seconds": 40.318360643228516
+            },
+            "secondary_structure": {
+              "updates": 1172,
+              "examples": 37504,
+              "tokens": 9795282,
+              "seconds": 38.16913926193956
+            },
+            "yeast_ppi": {
+              "updates": 899,
+              "examples": 28768,
+              "tokens": 29646612,
+              "seconds": 91.2968919732375
+            },
+            "human_ppi": {
+              "updates": 523,
+              "examples": 16736,
+              "tokens": 16742620,
+              "seconds": 50.190283646807075
+            },
+            "ppi_affinity": {
+              "updates": 607,
+              "examples": 19424,
+              "tokens": 7093904,
+              "seconds": 32.04752070689574
+            }
+          },
+          "auxiliary_mlm": null
+        },
+        {
+          "source": "/workspace/research/21m_4h_joint_ordered_pool/snapshots/step_23765_ema/package",
+          "cutoff_active_seconds": 900,
+          "step": 23765,
+          "active_seconds": 900.1199180769036,
+          "variant": "ema",
+          "weight": 0.25,
+          "sha256": "738632dda1d70b36d3c02e8d523fa424a880a02109d60f11d639703f655a9fba",
+          "compute_by_task": {
+            "fluorescence": {
+              "updates": 2951,
+              "examples": 94432,
+              "tokens": 22567131,
+              "seconds": 81.38885841530282
+            },
+            "stability": {
+              "updates": 4459,
+              "examples": 142688,
+              "tokens": 6743261,
+              "seconds": 117.53654595965054
+            },
+            "beta_lactamase": {
+              "updates": 1238,
+              "examples": 39616,
+              "tokens": 11409408,
+              "seconds": 34.78296104166657
+            },
+            "solubility": {
+              "updates": 4854,
+              "examples": 155328,
+              "tokens": 46558435,
+              "seconds": 149.51968884689268
+            },
+            "subcellular_localization": {
+              "updates": 1719,
+              "examples": 55008,
+              "tokens": 25972822,
+              "seconds": 73.25578501832206
+            },
+            "binary_localization": {
+              "updates": 1407,
+              "examples": 45024,
+              "tokens": 20511693,
+              "seconds": 60.39200586779043
+            },
+            "fold": {
+              "updates": 2153,
+              "examples": 68896,
+              "tokens": 11792745,
+              "seconds": 60.37679525522981
+            },
+            "secondary_structure": {
+              "updates": 1798,
+              "examples": 57536,
+              "tokens": 14759896,
+              "seconds": 55.28974596003536
+            },
+            "yeast_ppi": {
+              "updates": 1375,
+              "examples": 44000,
+              "tokens": 45031183,
+              "seconds": 137.4598815512145
+            },
+            "human_ppi": {
+              "updates": 880,
+              "examples": 28160,
+              "tokens": 28327073,
+              "seconds": 83.37703397334553
+            },
+            "ppi_affinity": {
+              "updates": 931,
+              "examples": 29792,
+              "tokens": 10786887,
+              "seconds": 46.74061618745327
+            }
+          },
+          "auxiliary_mlm": null
+        },
+        {
+          "source": "/workspace/research/21m_4h_joint_ordered_pool/snapshots/step_31515_ema/package",
+          "cutoff_active_seconds": 1200,
+          "step": 31515,
+          "active_seconds": 1200.1399940178962,
+          "variant": "ema",
+          "weight": 0.25,
+          "sha256": "382b743232427cf84929bf1463165110ed02a32fb92392d4b3c3be374241dd3c",
+          "compute_by_task": {
+            "fluorescence": {
+              "updates": 3878,
+              "examples": 124096,
+              "tokens": 29656084,
+              "seconds": 108.29484692763072
+            },
+            "stability": {
+              "updates": 5948,
+              "examples": 190336,
+              "tokens": 8997202,
+              "seconds": 158.43155370745808
+            },
+            "beta_lactamase": {
+              "updates": 1671,
+              "examples": 53472,
+              "tokens": 15399936,
+              "seconds": 47.3976139058359
+            },
+            "solubility": {
+              "updates": 6371,
+              "examples": 203872,
+              "tokens": 61069885,
+              "seconds": 196.84498821105808
+            },
+            "subcellular_localization": {
+              "updates": 2302,
+              "examples": 73664,
+              "tokens": 34855339,
+              "seconds": 98.2246110033011
+            },
+            "binary_localization": {
+              "updates": 1858,
+              "examples": 59456,
+              "tokens": 27212128,
+              "seconds": 80.11429120693356
+            },
+            "fold": {
+              "updates": 2835,
+              "examples": 90720,
+              "tokens": 15470378,
+              "seconds": 79.91399675200228
+            },
+            "secondary_structure": {
+              "updates": 2426,
+              "examples": 77632,
+              "tokens": 20110862,
+              "seconds": 74.93963593046647
+            },
+            "yeast_ppi": {
+              "updates": 1801,
+              "examples": 57632,
+              "tokens": 59063301,
+              "seconds": 180.3279145406559
+            },
+            "human_ppi": {
+              "updates": 1190,
+              "examples": 38080,
+              "tokens": 38506419,
+              "seconds": 113.30054696486332
+            },
+            "ppi_affinity": {
+              "updates": 1235,
+              "examples": 39520,
+              "tokens": 14342617,
+              "seconds": 62.34999486769084
+            }
+          },
+          "auxiliary_mlm": null
+        },
+        {
+          "source": "/workspace/research/21m_4h_joint_ordered_pool/snapshots/step_46827_ema/package",
+          "cutoff_active_seconds": 1800,
+          "step": 46827,
+          "active_seconds": 1800.0011083593126,
+          "variant": "ema",
+          "weight": 0.25,
+          "sha256": "81d9d22214c61a22dc104f7865d8f0fe73291fb3a4f8814d77983d276843f647",
+          "compute_by_task": {
+            "fluorescence": {
+              "updates": 5697,
+              "examples": 182304,
+              "tokens": 43566251,
+              "seconds": 162.3441452680854
+            },
+            "stability": {
+              "updates": 8790,
+              "examples": 281280,
+              "tokens": 13289595,
+              "seconds": 238.59854101459496
+            },
+            "beta_lactamase": {
+              "updates": 2418,
+              "examples": 77376,
+              "tokens": 22284288,
+              "seconds": 69.57553572778124
+            },
+            "solubility": {
+              "updates": 9491,
+              "examples": 303712,
+              "tokens": 90589330,
+              "seconds": 295.11683348717634
+            },
+            "subcellular_localization": {
+              "updates": 3449,
+              "examples": 110368,
+              "tokens": 52209242,
+              "seconds": 147.72849281225353
+            },
+            "binary_localization": {
+              "updates": 2782,
+              "examples": 89024,
+              "tokens": 40134720,
+              "seconds": 118.94762054237071
+            },
+            "fold": {
+              "updates": 4281,
+              "examples": 136992,
+              "tokens": 23170407,
+              "seconds": 122.23370583902579
+            },
+            "secondary_structure": {
+              "updates": 3578,
+              "examples": 114496,
+              "tokens": 29435955,
+              "seconds": 110.94229607644957
+            },
+            "yeast_ppi": {
+              "updates": 2693,
+              "examples": 86176,
+              "tokens": 88063633,
+              "seconds": 269.12609137850814
+            },
+            "human_ppi": {
+              "updates": 1805,
+              "examples": 57760,
+              "tokens": 58388137,
+              "seconds": 171.54262299148832
+            },
+            "ppi_affinity": {
+              "updates": 1843,
+              "examples": 58976,
+              "tokens": 21239463,
+              "seconds": 93.84522322157864
+            }
+          },
+          "auxiliary_mlm": null
+        }
+      ]
+    },
+    "initial_export_config": {
+      "name": "unified_t12_21M",
+      "layers": 12,
+      "dim": 384,
+      "heads": 6,
+      "ordered_pooling_bins": 8,
+      "structured_pooling": true,
+      "full_chain_ss": true,
+      "full_precision_head_tasks": [
+        "fluorescence",
+        "stability",
+        "beta_lactamase",
+        "ppi_affinity"
+      ],
+      "weight_variant": "budget_snapshot_average"
+    }
+  },
+  "history": [
+    {
+      "step": 46827,
+      "active_seconds": 1800.0011083593126,
+      "variant": "budget_snapshot_average",
+      "reward": 0.6915543813634248,
+      "tasks": {
+        "beta_lactamase": {
+          "invalid_reason": null,
+          "reward": 0.6550991593684602,
+          "spearman": 0.6550991593684602
+        },
+        "binary_localization": {
+          "accuracy": 0.9519953730480046,
+          "balanced_accuracy": 0.9461171698013804,
+          "invalid_reason": null,
+          "macro_f1": 0.9498199375419822,
+          "reward": 0.9519953730480046
+        },
+        "fluorescence": {
+          "invalid_reason": null,
+          "reward": 0.756894724301284,
+          "spearman": 0.756894724301284
+        },
+        "fold": {
+          "accuracy": 0.30027173913043476,
+          "balanced_accuracy": 0.19779765505718316,
+          "invalid_reason": null,
+          "macro_f1": 0.04516428860793635,
+          "reward": 0.30027173913043476
+        },
+        "human_ppi": {
+          "accuracy": 0.780952380952381,
+          "balanced_accuracy": 0.7916666666666666,
+          "invalid_reason": null,
+          "macro_f1": 0.7783703310934138,
+          "reward": 0.780952380952381
+        },
+        "ppi_affinity": {
+          "invalid_reason": null,
+          "reward": 0.4754805237056252,
+          "rmse": 2.292145955107753,
+          "spearman": 0.7230777506843703
+        },
+        "secondary_structure": {
+          "accuracy": 0.7562211476825675,
+          "invalid_reason": null,
+          "reward": 0.7562211476825675,
+          "valid_residues": 556931
+        },
+        "solubility": {
+          "accuracy": 0.7523768366464996,
+          "balanced_accuracy": 0.7437736834447468,
+          "invalid_reason": null,
+          "macro_f1": 0.7445518036454949,
+          "reward": 0.7523768366464996
+        },
+        "stability": {
+          "invalid_reason": null,
+          "reward": 0.7739680057567921,
+          "spearman": 0.7739680057567921
+        },
+        "subcellular_localization": {
+          "accuracy": 0.7933119886161508,
+          "balanced_accuracy": 0.7042125985273946,
+          "invalid_reason": null,
+          "macro_f1": 0.7256092338006721,
+          "reward": 0.7933119886161508
+        },
+        "yeast_ppi": {
+          "accuracy": 0.6105263157894737,
+          "balanced_accuracy": 0.5840201465201464,
+          "invalid_reason": null,
+          "macro_f1": 0.5839744348443603,
+          "reward": 0.6105263157894737
+        }
+      },
+      "evaluation": "Independent frozen-package replay of complete averaged parameters."
+    }
+  ],
+  "parameters": {
+    "backbone": 21306271,
+    "total": 26245478
+  },
+  "regression_refinement": {
+    "beta_lactamase": {
+      "baseline": 0.6550991593684602,
+      "selected": {
+        "method": "ridge",
+        "relative_ridge": 0.001,
+        "reward": 0.7106118351392832
+      },
+      "trials": [
+        {
+          "method": "unchanged",
+          "reward": 0.6550991593684602
+        },
+        {
+          "method": "float32_original",
+          "reward": 0.6546228505780767
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 1e-06,
+          "reward": 0.7099944515539697
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 1e-05,
+          "reward": 0.7099543822156638
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 0.0001,
+          "reward": 0.7102049969375706
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 0.001,
+          "reward": 0.7106118351392832
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 0.01,
+          "reward": 0.7058432425020701
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 0.1,
+          "reward": 0.6893007079280463
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 1.0,
+          "reward": 0.6661544135568709
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 10.0,
+          "reward": 0.6563087925238985
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 100.0,
+          "reward": 0.653597263275794
+        }
+      ],
+      "feature_cache": {
+        "source_manifest": "/run-state/workspace/research/21m_4h_ordered_pool_budget_average/research_feature_cache/features/complete.json",
+        "source_model_sha256": "b55822b8d72982566add4343d3a1a55ab5d28f1f4fd480b7e2c8216bf8c64e93",
+        "consumer_model_sha256": "b55822b8d72982566add4343d3a1a55ab5d28f1f4fd480b7e2c8216bf8c64e93",
+        "device": "cuda",
+        "dtype": "float32",
+        "task": "beta_lactamase",
+        "feature_state_tensor_checks": "exact model hash",
+        "scope": "Research cache only; model selection and full independent replay remain required."
+      }
+    },
+    "ppi_affinity": {
+      "baseline": 0.4754805237056252,
+      "selected": {
+        "method": "ridge",
+        "relative_ridge": 0.0001,
+        "reward": 0.5193548557644798
+      },
+      "trials": [
+        {
+          "method": "unchanged",
+          "reward": 0.4754805237056252
+        },
+        {
+          "method": "float32_original",
+          "reward": 0.475524426029664
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 1e-06,
+          "reward": 0.5171982358723451
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 1e-05,
+          "reward": 0.5175600325241468
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 0.0001,
+          "reward": 0.5193548557644798
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 0.001,
+          "reward": 0.5187251417309807
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 0.01,
+          "reward": 0.5078914933023689
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 0.1,
+          "reward": 0.49430133294593714
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 1.0,
+          "reward": 0.4893950229304202
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 10.0,
+          "reward": 0.4991254559366401
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 100.0,
+          "reward": 0.4771850245399195
+        }
+      ],
+      "feature_cache": {
+        "source_manifest": "/run-state/workspace/research/21m_4h_ordered_pool_budget_average/research_feature_cache/features/complete.json",
+        "source_model_sha256": "b55822b8d72982566add4343d3a1a55ab5d28f1f4fd480b7e2c8216bf8c64e93",
+        "consumer_model_sha256": "b55822b8d72982566add4343d3a1a55ab5d28f1f4fd480b7e2c8216bf8c64e93",
+        "device": "cuda",
+        "dtype": "float32",
+        "task": "ppi_affinity",
+        "feature_state_tensor_checks": "exact model hash",
+        "scope": "Research cache only; model selection and full independent replay remain required."
+      }
+    },
+    "fluorescence": {
+      "baseline": 0.756894724301284,
+      "selected": {
+        "method": "rank_target_ridge",
+        "relative_ridge": 1e-06,
+        "reward": 0.7723428943870447
+      },
+      "target": "Fit-only average ranks, including ties, mapped to (rank-.5)/N. No tune labels enter fitted targets or coefficients.",
+      "output_scale": "Empirical fit-label quantile, not original physical units; only Spearman tasks are eligible.",
+      "paired_row_gain_95pct": [
+        0.01134136929120931,
+        0.01985160814480393
+      ],
+      "trials": [
+        {
+          "relative_ridge": 1e-06,
+          "reward": 0.7723428943870447
+        },
+        {
+          "relative_ridge": 1e-05,
+          "reward": 0.7723160406411347
+        },
+        {
+          "relative_ridge": 0.0001,
+          "reward": 0.7719732611277846
+        },
+        {
+          "relative_ridge": 0.001,
+          "reward": 0.7693700818934532
+        },
+        {
+          "relative_ridge": 0.01,
+          "reward": 0.7640541339423836
+        },
+        {
+          "relative_ridge": 0.1,
+          "reward": 0.7571043279918875
+        },
+        {
+          "relative_ridge": 1.0,
+          "reward": 0.752284575611104
+        },
+        {
+          "relative_ridge": 10.0,
+          "reward": 0.7541631333072059
+        },
+        {
+          "relative_ridge": 100.0,
+          "reward": 0.753894123263629
+        }
+      ]
+    },
+    "stability": {
+      "baseline": 0.7739680057567921,
+      "selected": {
+        "method": "unchanged",
+        "reward": 0.7739680057567921
+      },
+      "trials": [
+        {
+          "method": "unchanged",
+          "reward": 0.7739680057567921
+        },
+        {
+          "method": "float32_original",
+          "reward": 0.7741236302189023
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 1e-06,
+          "reward": 0.7672796329423932
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 1e-05,
+          "reward": 0.7672825460984762
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 0.0001,
+          "reward": 0.7672897366277059
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 0.001,
+          "reward": 0.7673673944190934
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 0.01,
+          "reward": 0.7678065625403476
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 0.1,
+          "reward": 0.7693224211883709
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 1.0,
+          "reward": 0.7714237530456988
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 10.0,
+          "reward": 0.7720087602863004
+        },
+        {
+          "method": "ridge",
+          "relative_ridge": 100.0,
+          "reward": 0.7707731884373721
+        }
+      ],
+      "feature_cache": {
+        "source_manifest": "/run-state/workspace/research/21m_4h_ordered_pool_budget_average/research_feature_cache/features/complete.json",
+        "source_model_sha256": "b55822b8d72982566add4343d3a1a55ab5d28f1f4fd480b7e2c8216bf8c64e93",
+        "consumer_model_sha256": "134bc25520d2c25ad08434f9dd55bddfe46c4ec7fef449151b1f70dec44c3b5c",
+        "device": "cuda",
+        "dtype": "float32",
+        "task": "stability",
+        "feature_state_tensor_checks": 223,
+        "scope": "Research cache only; model selection and full independent replay remain required."
+      }
+    }
+  },
+  "normalization_buffer_values": 769,
+  "residual_readout_refinement": {
+    "tasks": {
+      "beta_lactamase": {
+        "baseline": 0.7106118351392832,
+        "selection_reference": 0.7106118351392832,
+        "best": 0.8107092667901805,
+        "gain": 0.10009743165089735,
+        "selected": {
+          "width": 64,
+          "decay": 0.01,
+          "step": 16000,
+          "reward": 0.8107092667901805
+        },
+        "history": [
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 400,
+            "reward": 0.7225562531213887
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 800,
+            "reward": 0.7370098480595366
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 1200,
+            "reward": 0.7537553323385473
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 1600,
+            "reward": 0.7549423918199027
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 2000,
+            "reward": 0.7681788432687019
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 2400,
+            "reward": 0.7721023099792399
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 2800,
+            "reward": 0.785276545139168
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 3200,
+            "reward": 0.7873928377525242
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 3600,
+            "reward": 0.7919636867116655
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 4000,
+            "reward": 0.7983462330372756
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 4400,
+            "reward": 0.7858921364857074
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 4800,
+            "reward": 0.7986073451577161
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 5200,
+            "reward": 0.7971884467892298
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 5600,
+            "reward": 0.7940709327864479
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 6000,
+            "reward": 0.7905241349249257
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 6400,
+            "reward": 0.8005500210522755
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 6800,
+            "reward": 0.7897902558192981
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 7200,
+            "reward": 0.7991564103078563
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 7600,
+            "reward": 0.798500920312418
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 8000,
+            "reward": 0.8065441424238493
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 8400,
+            "reward": 0.8024169193613976
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 8800,
+            "reward": 0.7985610456560528
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 9200,
+            "reward": 0.7968774933619296
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 9600,
+            "reward": 0.8048230425879311
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 10000,
+            "reward": 0.8067912621406682
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 10400,
+            "reward": 0.8063756334344923
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 10800,
+            "reward": 0.7980782933405529
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 11200,
+            "reward": 0.8096389588632499
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 11600,
+            "reward": 0.7975013631447112
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 12000,
+            "reward": 0.8041350789328009
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 12400,
+            "reward": 0.8020163539953941
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 12800,
+            "reward": 0.8079087230162975
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 13200,
+            "reward": 0.8016118201006767
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 13600,
+            "reward": 0.8069077576609823
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 14000,
+            "reward": 0.8062183858182439
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 14400,
+            "reward": 0.8053305875398055
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 14800,
+            "reward": 0.8072232770299212
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 15200,
+            "reward": 0.8050798874731953
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 15600,
+            "reward": 0.8054001434731799
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 16000,
+            "reward": 0.8107092667901805
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 16400,
+            "reward": 0.8015877955666338
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 16800,
+            "reward": 0.8029335961965494
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 17200,
+            "reward": 0.8051754308687936
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 17600,
+            "reward": 0.8086111952710273
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 18000,
+            "reward": 0.8077652585696687
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 18400,
+            "reward": 0.7996888759131606
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 18800,
+            "reward": 0.808673112853436
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 19200,
+            "reward": 0.8066055520657937
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 19600,
+            "reward": 0.8085559345754937
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 20000,
+            "reward": 0.8019044244167186
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 20400,
+            "reward": 0.8031125640398452
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 20800,
+            "reward": 0.8092749210303559
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 21200,
+            "reward": 0.8006531003736074
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 21600,
+            "reward": 0.8096273093112184
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 22000,
+            "reward": 0.7994401387747314
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 22400,
+            "reward": 0.8038149936222233
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 22800,
+            "reward": 0.8047514810540238
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 23200,
+            "reward": 0.8011853952895045
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 23600,
+            "reward": 0.8003152633646962
+          },
+          {
+            "width": 64,
+            "decay": 0.01,
+            "step": 24000,
+            "reward": 0.8040724785927639
+          }
+        ],
+        "adopt_threshold": 0.02,
+        "initialization_seed": 200,
+        "feature_package": "/run-state/workspace/research/21m_4h_ordered_pool_budget_average_regression/package"
+      }
+    },
+    "feature_source": "Same frozen shared encoder and penultimate layers as source package; fit-only training, tune model selection"
+  },
+  "research_feature_cache": {
+    "manifest": "/run-state/workspace/research/21m_4h_ordered_pool_budget_average/research_feature_cache/features/complete.json",
+    "device": "cuda",
+    "model_sha256": "b55822b8d72982566add4343d3a1a55ab5d28f1f4fd480b7e2c8216bf8c64e93",
+    "tasks": [
+      "beta_lactamase",
+      "fold",
+      "ppi_affinity",
+      "yeast_ppi"
+    ]
+  },
+  "function_readout_study": {
+    "parent": "/run-state/workspace/research/21m_4h_ordered_pool_budget_average_refined/package",
+    "parent_sha256_model": "134bc25520d2c25ad08434f9dd55bddfe46c4ec7fef449151b1f70dec44c3b5c",
+    "policy": {
+      "recorded_utc": "2026-10-31T20:15:30.897237+00:00",
+      "source": "/run-state/workspace/research/21m_4h_ordered_pool_budget_average_refined/package",
+      "sha256_model": "134bc25520d2c25ad08434f9dd55bddfe46c4ec7fef449151b1f70dec44c3b5c",
+      "hypothesis": "The complete shared-model state may contain fluorescence/stability information that its native final layers do not fully use.",
+      "data": "Fit-only frozen shared-encoder penultimate features and labels. Tune selects ridge strength and residual checkpoints.",
+      "recipe": "Same existing nine ridge strengths, followed separately for GFP and stability by one64-wide residual,decay.01,6000steps,batch128,noise.02,dropout.1,seed200. 4 CPU threads; no GPU or shared-parameter updates.",
+      "adoption": "Ridge minimum gain.01 over the independently replayed parent; residual minimum.02 over the selected ridge/native score. Unaccepted heads are preserved exactly.",
+      "selection": "If promising, build one complete package preserving every unmodified tensor and all nine other task behaviors; independently replay all eleven tasks and compare to then-current publication. No automatic assembly or publication.",
+      "precomputed_features": {
+        "manifest": "/run-state/workspace/research/21m_4h_ordered_pool_budget_average/research_feature_cache/features/complete.json",
+        "device": "cuda",
+        "model_sha256": "b55822b8d72982566add4343d3a1a55ab5d28f1f4fd480b7e2c8216bf8c64e93",
+        "tasks": [
+          "fluorescence",
+          "stability"
+        ]
+      }
+    },
+    "changed_tasks": [
+      "fluorescence"
+    ],
+    "ridge_selection": {
+      "fluorescence": {
+        "baseline": 0.756894724301284,
+        "selected": {
+          "method": "unchanged",
+          "reward": 0.756894724301284
+        },
+        "trials": [
+          {
+            "method": "unchanged",
+            "reward": 0.756894724301284
+          },
+          {
+            "method": "float32_original",
+            "reward": 0.7568599737306553
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 1e-06,
+            "reward": 0.7514995788054283
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 1e-05,
+            "reward": 0.7517638687088956
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 0.0001,
+            "reward": 0.7532718154449811
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 0.001,
+            "reward": 0.7568819953432682
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 0.01,
+            "reward": 0.7571482881809775
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 0.1,
+            "reward": 0.7546210415421412
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 1.0,
+            "reward": 0.7548781526721023
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 10.0,
+            "reward": 0.7552520264655704
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 100.0,
+            "reward": 0.7511077620869435
+          }
+        ],
+        "feature_cache": {
+          "source_manifest": "/run-state/workspace/research/21m_4h_ordered_pool_budget_average/research_feature_cache/features/complete.json",
+          "source_model_sha256": "b55822b8d72982566add4343d3a1a55ab5d28f1f4fd480b7e2c8216bf8c64e93",
+          "consumer_model_sha256": "134bc25520d2c25ad08434f9dd55bddfe46c4ec7fef449151b1f70dec44c3b5c",
+          "device": "cuda",
+          "dtype": "float32",
+          "task": "fluorescence",
+          "feature_state_tensor_checks": 223,
+          "scope": "Research cache only; model selection and full independent replay remain required."
+        }
+      },
+      "stability": {
+        "baseline": 0.7739680057567921,
+        "selected": {
+          "method": "unchanged",
+          "reward": 0.7739680057567921
+        },
+        "trials": [
+          {
+            "method": "unchanged",
+            "reward": 0.7739680057567921
+          },
+          {
+            "method": "float32_original",
+            "reward": 0.7741236302189023
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 1e-06,
+            "reward": 0.7672796329423932
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 1e-05,
+            "reward": 0.7672825460984762
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 0.0001,
+            "reward": 0.7672897366277059
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 0.001,
+            "reward": 0.7673673944190934
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 0.01,
+            "reward": 0.7678065625403476
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 0.1,
+            "reward": 0.7693224211883709
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 1.0,
+            "reward": 0.7714237530456988
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 10.0,
+            "reward": 0.7720087602863004
+          },
+          {
+            "method": "ridge",
+            "relative_ridge": 100.0,
+            "reward": 0.7707731884373721
+          }
+        ],
+        "feature_cache": {
+          "source_manifest": "/run-state/workspace/research/21m_4h_ordered_pool_budget_average/research_feature_cache/features/complete.json",
+          "source_model_sha256": "b55822b8d72982566add4343d3a1a55ab5d28f1f4fd480b7e2c8216bf8c64e93",
+          "consumer_model_sha256": "134bc25520d2c25ad08434f9dd55bddfe46c4ec7fef449151b1f70dec44c3b5c",
+          "device": "cuda",
+          "dtype": "float32",
+          "task": "stability",
+          "feature_state_tensor_checks": 223,
+          "scope": "Research cache only; model selection and full independent replay remain required."
+        }
+      }
+    },
+    "residual_selection": {
+      "fluorescence": {
+        "baseline": 0.7568595603531254,
+        "selection_reference": 0.756894724301284,
+        "best": 0.7588742785752952,
+        "gain": 0.001979554274011197,
+        "selected": {
+          "width": 64,
+          "decay": 0.01,
+          "step": 4400,
+          "reward": 0.7588742785752952
+        },
+        "adopt_threshold": 0.02,
+        "initialization_seed": 200,
+        "feature_package": "/run-state/workspace/research/21m_4h_ordered_pool_average_function_readouts/regression/package"
+      },
+      "stability": {
+        "baseline": 0.7741236302189023,
+        "selection_reference": 0.7739680057567921,
+        "best": 0.7741236302189023,
+        "gain": 0.0001556244621101488,
+        "selected": null,
+        "adopt_threshold": 0.02,
+        "initialization_seed": 200,
+        "feature_package": "/run-state/workspace/research/21m_4h_ordered_pool_average_function_readouts/regression/package"
+      }
+    },
+    "rank_comparison": {
+      "results": {
+        "fluorescence": {
+          "baseline": 0.756894724301284,
+          "best": 0.7723428943870447,
+          "gain": 0.015448170085760715,
+          "selected": {
+            "relative_ridge": 1e-06,
+            "reward": 0.7723428943870447
+          },
+          "trials": [
+            {
+              "relative_ridge": 1e-06,
+              "reward": 0.7723428943870447
+            },
+            {
+              "relative_ridge": 1e-05,
+              "reward": 0.7723160406411347
+            },
+            {
+              "relative_ridge": 0.0001,
+              "reward": 0.7719732611277846
+            },
+            {
+              "relative_ridge": 0.001,
+              "reward": 0.7693700818934532
+            },
+            {
+              "relative_ridge": 0.01,
+              "reward": 0.7640541339423836
+            },
+            {
+              "relative_ridge": 0.1,
+              "reward": 0.7571043279918875
+            },
+            {
+              "relative_ridge": 1.0,
+              "reward": 0.752284575611104
+            },
+            {
+              "relative_ridge": 10.0,
+              "reward": 0.7541631333072059
+            },
+            {
+              "relative_ridge": 100.0,
+              "reward": 0.753894123263629
+            }
+          ],
+          "paired_row_gain_95pct": [
+            0.01134136929120931,
+            0.01985160814480393
+          ],
+          "adopt_threshold": 0.01,
+          "adopt": true,
+          "output_scale": "Empirical fit-label quantile, not original physical units; only Spearman tasks are eligible."
+        },
+        "stability": {
+          "baseline": 0.7739680057567921,
+          "best": 0.7745954734537845,
+          "gain": 0.0006274676969924009,
+          "selected": {
+            "relative_ridge": 10.0,
+            "reward": 0.7745954734537845
+          },
+          "trials": [
+            {
+              "relative_ridge": 1e-06,
+              "reward": 0.7678990098225509
+            },
+            {
+              "relative_ridge": 1e-05,
+              "reward": 0.7679080789892079
+            },
+            {
+              "relative_ridge": 0.0001,
+              "reward": 0.767915276710481
+            },
+            {
+              "relative_ridge": 0.001,
+              "reward": 0.7680092865587329
+            },
+            {
+              "relative_ridge": 0.01,
+              "reward": 0.7686162783179286
+            },
+            {
+              "relative_ridge": 0.1,
+              "reward": 0.77069032791628
+            },
+            {
+              "relative_ridge": 1.0,
+              "reward": 0.7735420800799776
+            },
+            {
+              "relative_ridge": 10.0,
+              "reward": 0.7745954734537845
+            },
+            {
+              "relative_ridge": 100.0,
+              "reward": 0.7724244282166632
+            }
+          ],
+          "paired_row_gain_95pct": [
+            -0.0005641879880377259,
+            0.001779585822624077
+          ],
+          "adopt_threshold": 0.01,
+          "adopt": false,
+          "output_scale": "Empirical fit-label quantile, not original physical units; only Spearman tasks are eligible."
+        }
+      },
+      "source": "/run-state/workspace/research/21m_4h_ordered_pool_budget_average_refined/package",
+      "policy": {
+        "recorded_utc": "2026-10-31T20:15:33.196032+00:00",
+        "source": "/run-state/workspace/research/21m_4h_ordered_pool_budget_average_refined/package",
+        "sha256_model": "134bc25520d2c25ad08434f9dd55bddfe46c4ec7fef449151b1f70dec44c3b5c",
+        "feature_study": "/run-state/workspace/research/21m_4h_ordered_pool_average_function_readouts",
+        "hypothesis": "Fit-label rank targets may better align a frozen linear readout with Spearman reward than squared error on the original labels.",
+        "targets": "Fit-only average ranks, including ties, mapped to (rank-.5)/N. No tune labels enter fitted targets or coefficients.",
+        "recipe": "Same nine ridge strengths as the raw-label control, same fit-centered/scaled penultimate features. One final Linear layer per task; no added parameters.",
+        "adoption": "At least .01 gain over the independently replayed parent and positive1,000-replicate paired row interval; model-selection uncertainty is not included.",
+        "publication": "Manual comparison against other completed readout studies, complete package assembly and independent eleven-task replay required. No automatic publication."
+      }
+    },
+    "rank_selected_tasks": [
+      "fluorescence"
+    ],
+    "head_selection_policy": "For each task, use the highest tune reward among heads that pass their recorded adoption rules. Rank-target heads require .01 gain and a positive paired interval; raw-label ridge requires .01 and residuals .02. All heads use this exact parent encoder and penultimate layers.",
+    "preservation": "All other state tensors and inference code are exactly unchanged."
+  },
+  "mutation_readouts": {
+    "parent": "/run-state/workspace/research/21m_4h_ordered_pool_gfp_rank_refined/package",
+    "parent_model_sha256": "e4810c40269f99e2a2c8c0ded049beef6012d1e123950c9dde7078c1d2de5df7",
+    "policy": {
+      "recorded_utc": "2026-10-31T20:16:59.733880+00:00",
+      "source": "/run-state/workspace/research/21m_4h_ordered_pool_gfp_rank_refined/package",
+      "model_sha256": "e4810c40269f99e2a2c8c0ded049beef6012d1e123950c9dde7078c1d2de5df7",
+      "hypothesis": "Test transfer of the established fixed GFP contextual-mutation recipe to the ordered-pooling joint encoder. Complete the same accepted-refinement comparison with the same shared encoder, fixed nine-ridge grid and quality gates.",
+      "reference": "Per-task modal fit length and per-position modal fit residue, deterministic lexical ties, no labels/tune used. Reference encoder is the same frozen jointly learned encoder.",
+      "features": "Current penultimate feature and log1p mutation count in both arms. Intervention also gets sum of input-minus-reference shared-encoder residue vectors at substitutions. Only equal-length sequences are eligible; all other predictions preserved exactly.",
+      "targets": "Fit-only empirical average-rank quantiles minus the frozen parent rank-head prediction. The shared encoder and all existing heads stay frozen.",
+      "tasks": [
+        "fluorescence"
+      ],
+      "ridge_strengths": [
+        1e-06,
+        1e-05,
+        0.0001,
+        0.001,
+        0.01,
+        0.1,
+        1.0,
+        10.0,
+        100.0
+      ],
+      "fit": "CPU8threads; newly initialized Linear seed802 then exact ridge minimizer on fit-centered/scaled eligible rows, unpenalized intercept; final coefficients float32. Tune chooses one of nine strengths or unchanged. Same grid for count control. No nonlinear expansion registered.",
+      "adoption": "Mutation-feature reward gain >= .01 versus the independently replayed parent AND positive lower endpoints for 1000 paired-row 95% intervals versus BOTH parent and count control. Same nine ridge strengths; no expansion.",
+      "publication": "Any accepted result must preserve unmodified parameters and all other task predictions, complete full independent11task replay, and receive root review. No automatic publication.",
+      "limitations": "One seed, repeated tune selection and shared ancestry limit confidence. Equal-length correspondence may be unsuitable for unrelated proteins; other lengths fall back. More distant GFP mutants are absent from visible same-length tune rows. All-task changes require review; beta recipes differ across lineages because fixed gates selected different targets.",
+      "device": "cpu",
+      "parent_completion_marker": "/run-state/workspace/research/21m_4h_ordered_pool_gfp_rank_refined/controlled_review_complete.json",
+      "lineage_summary": "Own seed42 four-hour UniRef50 MLM ancestor (9,127,835,498 input tokens), then seed1 all-eleven joint supervised training for30active minutes with one shared projection of eight ordered residue-bin summaries. A complete-state budget average is followed by accepted beta raw-label residual, affinity ridge and GFP rank ridge.",
+      "resources": "CPU only, eight threads, serialized with the existing readout pipeline lock. No GPU cache; prior mutation CUDA validation rejection remains in force.",
+      "scope": "GFP only, after independent parent replay/preservation succeeds, regardless of parent mean. No new beta mutation fit, nonlinear expansion, subgroup routing, or automatic assembly/publication.",
+      "comparison_references": [
+        "21m_4h_ordered_pool_gfp_rank_refined",
+        "21m_2h_gfp_mutation_refined",
+        "21m_4h_partner_attention_gfp_mutation_refined",
+        "21m_4h_mlm_regularized_gfp_mutation_refined"
+      ],
+      "numerical_implementation": "Use the already validated FP64 accumulation of the small correction over FP32 features/coefficients in fitting and runtime; cast correction to FP32 before adding unchanged parent prediction. No quality gate changes.",
+      "implementation_hashes": {
+        "extract_mutation_features.py": "ca936410e68cecb64359a6f680769a5b9a8e774ab1a0de9b7410f83e9b90c112",
+        "fit_mutation_readouts.py": "b98f3c957c27f341be77dfdcef29469ed8920a826a0a3a047af9bf2827d5af76",
+        "mutation_features.py": "6dad5f3faccc89de37b9e67fcbc63735f629ca9118e0fb6ef2107e1d106eceac",
+        "mutation_readout_runtime.py": "c17c51019f011b00539c6686dd46bc50cda2e426292972e339045bdca9d809b4"
+      },
+      "replay_validation": {
+        "registered_utc": "2026-10-31T20:16:59.733880+00:00",
+        "script": "audit_mutation_replay.py",
+        "sha256": "051b29020d31c0ec15f8d788ed6d26c04f2f55a7763507a48c4da3502452185d",
+        "policy": "Independent all-eleven replay; other ten files byte-identical and all ineligible GFP values exact. Reapply unchanged .01 gain and positive 1000-replicate paired intervals versus actual parent and recorded count control. Full replay/cached task reward difference must satisfy the same2e-5 implementation score tolerance used by the registered fitter. No change to extraction device, training grid or scientific gates."
+      }
+    },
+    "results": {
+      "recorded_utc": "2026-10-31T20:28:51.613149+00:00",
+      "model_sha256": "e4810c40269f99e2a2c8c0ded049beef6012d1e123950c9dde7078c1d2de5df7",
+      "tasks": {
+        "fluorescence": {
+          "baseline": 0.7723430213764729,
+          "official_baseline": 0.772343323215241,
+          "fit_baseline": 0.7863589989741424,
+          "variants": {
+            "count_control": {
+              "reward": 0.776790447777812,
+              "ridge": 1e-06,
+              "trainable_coefficients": 386,
+              "torch_reproduction_reward": 0.776790447777812,
+              "fit_reward": 0.7900097632599804,
+              "initialization_seed": 802,
+              "optimizer": "Exact ridge minimizer",
+              "history": [
+                {
+                  "ridge": 1e-06,
+                  "reward": 0.776790447777812
+                },
+                {
+                  "ridge": 1e-05,
+                  "reward": 0.7767805320171554
+                },
+                {
+                  "ridge": 0.0001,
+                  "reward": 0.7767515502440512
+                },
+                {
+                  "ridge": 0.001,
+                  "reward": 0.7767495968635165
+                },
+                {
+                  "ridge": 0.01,
+                  "reward": 0.776417447448294
+                },
+                {
+                  "ridge": 0.1,
+                  "reward": 0.7760944746335998
+                },
+                {
+                  "ridge": 1.0,
+                  "reward": 0.7748830010700987
+                },
+                {
+                  "ridge": 10.0,
+                  "reward": 0.772957989918246
+                },
+                {
+                  "ridge": 100.0,
+                  "reward": 0.7724844628929798
+                }
+              ],
+              "correction_accumulation": "Float64 over float32 features/coefficients, cast correction to float32 before adding base."
+            },
+            "mutation_delta": {
+              "reward": 0.798202901266472,
+              "ridge": 1e-05,
+              "trainable_coefficients": 770,
+              "torch_reproduction_reward": 0.798202901266472,
+              "fit_reward": 0.8184494140182572,
+              "initialization_seed": 802,
+              "optimizer": "Exact ridge minimizer",
+              "history": [
+                {
+                  "ridge": 1e-06,
+                  "reward": 0.7981796215218516
+                },
+                {
+                  "ridge": 1e-05,
+                  "reward": 0.798202901266472
+                },
+                {
+                  "ridge": 0.0001,
+                  "reward": 0.7981754469667142
+                },
+                {
+                  "ridge": 0.001,
+                  "reward": 0.7977612071069778
+                },
+                {
+                  "ridge": 0.01,
+                  "reward": 0.796146291196867
+                },
+                {
+                  "ridge": 0.1,
+                  "reward": 0.7909305870998008
+                },
+                {
+                  "ridge": 1.0,
+                  "reward": 0.7820077390436139
+                },
+                {
+                  "ridge": 10.0,
+                  "reward": 0.7751006571865364
+                },
+                {
+                  "ridge": 100.0,
+                  "reward": 0.772827596225293
+                }
+              ],
+              "correction_accumulation": "Float64 over float32 features/coefficients, cast correction to float32 before adding base."
+            }
+          },
+          "gain": 0.02585987988999905,
+          "gain_over_count_control": 0.021412453488659944,
+          "paired_row_gain_95pct": [
+            0.02038080894507352,
+            0.03128998874749907
+          ],
+          "paired_row_gain_over_control_95pct": [
+            0.016533974421942096,
+            0.026402356316020108
+          ],
+          "adopt": true,
+          "unchanged_different_length_tune_rows": 127,
+          "fit_eligible_rows": 20963,
+          "tune_eligible_rows": 5235,
+          "uncertainty": "Intervals exclude model-selection and homology dependence."
+        }
+      },
+      "next": "No automatic assembly/publication. Any accepted head requires independent complete eleven-task replay and preservation review."
+    },
+    "selected_tasks": [
+      "fluorescence"
+    ],
+    "references": {
+      "fluorescence": {
+        "sequence": "SKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLKFICTTGKLPVPWPTLVTTLSYGVQCFSRYPDHMKQHDFFKSAMPEGYVQERTIFFKDDGNYKTRAEVKFEGDTLVNRIELKGIDFKEDGNILGHKLEYNYNSHNVYIMADKQKNGIKVNFKIRHNIEDGSVQLADHYQQNTPIGDGPVLLPDNHYLSTQSALSKDPNEKRDHMVLLEFVTAAGITHGMDELYK",
+        "fit_rows": 21446,
+        "modal_length": 237,
+        "modal_length_rows": 20963,
+        "ties": "Lexicographically smallest residue"
+      }
+    },
+    "preservation": "Every original tensor is preserved exactly, including original heads. Selected head keys gain an original_head prefix. Only new correction coefficients are fitted.",
+    "subgroup_diagnostics": {
+      "recorded_utc": "2026-10-31T20:29:00.813932+00:00",
+      "scope": "Descriptive subgroups only; no fitting or alteration of predeclared acceptance rules.",
+      "limitations": "Conditional rank correlations do not decompose global reward. Global rank-error contributions do; intervals exclude selection/homology uncertainty.",
+      "tasks": {
+        "fluorescence": {
+          "fit_label_thresholds": [
+            3.256847381591797,
+            3.552252769470215,
+            3.677128791809082
+          ],
+          "groups": {
+            "fit_label_quartile_1": {
+              "rows": 1361,
+              "conditional_spearman": {
+                "base": 0.5266697511335511,
+                "count_control": 0.5237499741369044,
+                "mutation": 0.5338514525628313
+              },
+              "global_rho_change_contribution": 0.003322294104188563,
+              "global_rho_change_vs_count_control_contribution": 0.0030068212670968457,
+              "paired_conditional_gain_95pct": [
+                -0.019661179240733007,
+                0.031714895638367206
+              ]
+            },
+            "fit_label_quartile_2": {
+              "rows": 1297,
+              "conditional_spearman": {
+                "base": 0.4122631568417991,
+                "count_control": 0.41841400206050844,
+                "mutation": 0.4479224583250088
+              },
+              "global_rho_change_contribution": 0.008038911172027017,
+              "global_rho_change_vs_count_control_contribution": 0.0052589163937707024,
+              "paired_conditional_gain_95pct": [
+                0.017015931932606823,
+                0.05463600709331574
+              ]
+            },
+            "fit_label_quartile_3": {
+              "rows": 1384,
+              "conditional_spearman": {
+                "base": 0.18607304130725086,
+                "count_control": 0.19771895749919038,
+                "mutation": 0.23749204811606303
+              },
+              "global_rho_change_contribution": 0.008554293278452134,
+              "global_rho_change_vs_count_control_contribution": 0.007486135764047524,
+              "paired_conditional_gain_95pct": [
+                0.025552828827296335,
+                0.07658871688334705
+              ]
+            },
+            "fit_label_quartile_4": {
+              "rows": 1320,
+              "conditional_spearman": {
+                "base": 0.22482122707628116,
+                "count_control": 0.21610180054165207,
+                "mutation": 0.24994432075046086
+              },
+              "global_rho_change_contribution": 0.005944381335331244,
+              "global_rho_change_vs_count_control_contribution": 0.005660580063744807,
+              "paired_conditional_gain_95pct": [
+                -0.007185195872217796,
+                0.05597320652599361
+              ]
+            },
+            "mutations_0_or_1": {
+              "rows": 233,
+              "conditional_spearman": {
+                "base": 0.770087433343595,
+                "count_control": 0.7706699269903614,
+                "mutation": 0.7942187031253102
+              },
+              "global_rho_change_contribution": -0.0006073353012055145,
+              "global_rho_change_vs_count_control_contribution": 0.0007638252242460538,
+              "paired_conditional_gain_95pct": [
+                -0.008379313753226222,
+                0.05579036972752098
+              ]
+            },
+            "mutations_2": {
+              "rows": 2528,
+              "conditional_spearman": {
+                "base": 0.6889684903340906,
+                "count_control": 0.688807850337469,
+                "mutation": 0.719935872529225
+              },
+              "global_rho_change_contribution": 0.014026750885040486,
+              "global_rho_change_vs_count_control_contribution": 0.012652702663068617,
+              "paired_conditional_gain_95pct": [
+                0.021310743671952067,
+                0.04000200081647618
+              ]
+            },
+            "mutations_3": {
+              "rows": 2474,
+              "conditional_spearman": {
+                "base": 0.8094694522851252,
+                "count_control": 0.8100324150007207,
+                "mutation": 0.8290239354938609
+              },
+              "global_rho_change_contribution": 0.012521115893853784,
+              "global_rho_change_vs_count_control_contribution": 0.008108323315667024,
+              "paired_conditional_gain_95pct": [
+                0.013506687840631254,
+                0.026181799360849498
+              ]
+            },
+            "mutations_4plus": {
+              "rows": 0,
+              "conditional_spearman": {},
+              "global_rho_change_contribution": 0.0,
+              "global_rho_change_vs_count_control_contribution": 0.0
+            },
+            "different_length": {
+              "rows": 127,
+              "conditional_spearman": {
+                "base": 0.7519810177764953,
+                "count_control": 0.7519810177764953,
+                "mutation": 0.7519810177764953
+              },
+              "global_rho_change_contribution": -8.065158768979982e-05,
+              "global_rho_change_vs_count_control_contribution": -0.00011239771432181658,
+              "paired_conditional_gain_95pct": [
+                0.0,
+                0.0
+              ]
+            },
+            "at_least_one_unseen_substitution": {
+              "rows": 55,
+              "conditional_spearman": {
+                "base": 0.7701298701298701,
+                "count_control": 0.7515873015873016,
+                "mutation": 0.7612554112554113
+              },
+              "global_rho_change_contribution": -0.00033485385028152984,
+              "global_rho_change_vs_count_control_contribution": -3.1131152880709647e-06,
+              "paired_conditional_gain_95pct": [
+                -0.0694263914745223,
+                0.04213291963390857
+              ]
+            },
+            "only_fit_seen_substitutions": {
+              "rows": 5180,
+              "conditional_spearman": {
+                "base": 0.7642596271172903,
+                "count_control": 0.7692590587733675,
+                "mutation": 0.7921879713595679
+              },
+              "global_rho_change_contribution": 0.02627538532797029,
+              "global_rho_change_vs_count_control_contribution": 0.02152796431826977,
+              "paired_conditional_gain_95pct": [
+                0.022664277090117636,
+                0.033233160556254186
+              ]
+            }
+          },
+          "global_rho_change": 0.02585987988999905
+        }
+      }
+    },
+    "independent_replay_validation": {
+      "status": "passed",
+      "utc": "2026-10-31T20:31:46.502103+00:00",
+      "all_eleven_tasks_valid": true,
+      "parent_model_sha256": "e4810c40269f99e2a2c8c0ded049beef6012d1e123950c9dde7078c1d2de5df7",
+      "candidate_model_sha256": "55d3a4ac78013573513bad549fe800659672f3589dd40d1e915ab3f0c673bf78",
+      "identical_prediction_sha256": {
+        "beta_lactamase": "53408c9adf34de6cebeb286c754810d803f37c2bc3f645bba157a111b89677ee",
+        "binary_localization": "9b99415bdb4b7eaed6c20fe5b89697c21cb0bf70a2b295ba2aaceec7b0cd330f",
+        "fold": "bfaf0b6c78c2b5c813a1f242704f74cc40737d9207e084821fd0d793609c2509",
+        "human_ppi": "df27b7bdca1086e27bb6a2be010550bbb3081960f32c6bef3e65b841d5b78686",
+        "ppi_affinity": "0fd1b09dd9d43c766b62ef45268ef079ba7f469bb6fa37ca37beb0e72ef985b2",
+        "secondary_structure": "b73ed5ed2fd20d67d745401c7caed2ef58e6f9bb076c730b06277825ea4325fc",
+        "solubility": "51a047ec0bbd91091d70995310b628708c7500fdc5a9cca426aae4ceb8224c20",
+        "stability": "39376c54be40df028ee8c6442774a94d7d08f9e25ee5c920879bc4da706c401a",
+        "subcellular_localization": "76d60223f02d0c7cf2f15b10620fd58216c0cc533c4d920e5ae5f0c27c974d48",
+        "yeast_ppi": "a13ec6cee13d6ad15be8a306fba646088d02877a9cd80e40dd7f4595c43b7763"
+      },
+      "changed_tasks": {
+        "fluorescence": {
+          "official_reward": 0.7982027307273727,
+          "parent_reward": 0.772343323215241,
+          "cached_reward": 0.798202901266472,
+          "replay_minus_cached_reward": -1.7053909928677058e-07,
+          "max_abs_cached_replay_difference": 4.5299530029296875e-06,
+          "control_reward": 0.776790447777812,
+          "gain": 0.02585940751213167,
+          "paired_replay_gain_over_parent_95pct": [
+            0.02038053950513996,
+            0.03128861746995915
+          ],
+          "paired_replay_gain_over_count_control_95pct": [
+            0.016534458755953028,
+            0.02640360375129057
+          ],
+          "other_length_rows_exactly_preserved": 127
+        }
+      },
+      "limitations": "Count control uses the registered CPU frozen-feature computation; candidate and parent use independent full replay. All intervals exclude model-selection and homology uncertainty.",
+      "next": "All-task comparison against current publication and root review; no automatic publication."
+    }
+  },
+  "independent_replay": {
+    "reward": 0.7134447361359852,
+    "tasks": {
+      "beta_lactamase": {
+        "invalid_reason": null,
+        "reward": 0.8107105469607334,
+        "spearman": 0.8107105469607334
+      },
+      "binary_localization": {
+        "accuracy": 0.9519953730480046,
+        "balanced_accuracy": 0.9461171698013804,
+        "invalid_reason": null,
+        "macro_f1": 0.9498199375419822,
+        "reward": 0.9519953730480046
+      },
+      "fluorescence": {
+        "invalid_reason": null,
+        "reward": 0.7982027307273727,
+        "spearman": 0.7982027307273727
+      },
+      "fold": {
+        "accuracy": 0.30027173913043476,
+        "balanced_accuracy": 0.19779765505718316,
+        "invalid_reason": null,
+        "macro_f1": 0.04516428860793635,
+        "reward": 0.30027173913043476
+      },
+      "human_ppi": {
+        "accuracy": 0.780952380952381,
+        "balanced_accuracy": 0.7916666666666666,
+        "invalid_reason": null,
+        "macro_f1": 0.7783703310934138,
+        "reward": 0.780952380952381
+      },
+      "ppi_affinity": {
+        "invalid_reason": null,
+        "reward": 0.5193550321854273,
+        "rmse": 2.0200166397110486,
+        "spearman": 0.7476100989845718
+      },
+      "secondary_structure": {
+        "accuracy": 0.7562211476825675,
+        "invalid_reason": null,
+        "reward": 0.7562211476825675,
+        "valid_residues": 556931
+      },
+      "solubility": {
+        "accuracy": 0.7523768366464996,
+        "balanced_accuracy": 0.7437736834447468,
+        "invalid_reason": null,
+        "macro_f1": 0.7445518036454949,
+        "reward": 0.7523768366464996
+      },
+      "stability": {
+        "invalid_reason": null,
+        "reward": 0.7739680057567921,
+        "spearman": 0.7739680057567921
+      },
+      "subcellular_localization": {
+        "accuracy": 0.7933119886161508,
+        "balanced_accuracy": 0.7042125985273946,
+        "invalid_reason": null,
+        "macro_f1": 0.7256092338006721,
+        "reward": 0.7933119886161508
+      },
+      "yeast_ppi": {
+        "accuracy": 0.6105263157894737,
+        "balanced_accuracy": 0.5840201465201464,
+        "invalid_reason": null,
+        "macro_f1": 0.5839744348443603,
+        "reward": 0.6105263157894737
+      }
+    }
+  },
+  "publication_selection": {
+    "recorded_utc": "2026-10-31T20:32:53.831691+00:00",
+    "previous_publication": "/run-state/workspace/research/21m_2h_gfp_mutation_refined/publish_package",
+    "mean_reward": 0.7134447361359852,
+    "previous_mean": 0.7112407988729224,
+    "mean_delta": 0.002203937263062894,
+    "mean_intervals": {
+      "paired_row_or_ss_chain_95pct": [
+        -0.009605801691962384,
+        0.014426625158987648
+      ],
+      "paired_with_ppi_protein_components_95pct": [
+        -0.010322763408268208,
+        0.013717698113209295
+      ],
+      "point_delta": 0.002203937263062894
+    },
+    "task_change_counts": {
+      "gains": 5,
+      "declines": 6,
+      "ties": 0
+    },
+    "evidence": "Highest independently replayed complete tune candidate. Compared with the current publication, GFP, fold and secondary structure have positive paired intervals; human and subcellular point rewards also increase. Six tasks decline in point reward, with their intervals spanning zero. The mean gain interval spans zero, so aggregate superiority is uncertain. Chosen for higher equal-weight point reward and gains across function and both structure tasks, after explicit trade-off review, not a Pareto-improvement claim.",
+    "limitations": "Single-seed adaptive tune evidence; no hidden-test claim. Intervals exclude model-selection/homology uncertainty. Affinity declines .01638 versus previous publication. Stability is lower than the held partner-attention recipe.55novel-substitution GFP rows decline slightly with a wide interval;no aligned4+mutant tune rows. Rare-fold, long-context and small-PPI limitations remain."
+  }
+}

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S002/PUBLISHED.json

@@ -0,0 +1,5 @@
+{
+  "source": "/run-state/workspace/research/21m_4p25h_uniref_function_rank_refined/publish_package",
+  "files": 7,
+  "bytes": 99304926
+}

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S002/README.md

@@ -0,0 +1,55 @@
+# Unified PEER sequence model
+
+One shared 12-layer rotary transformer (width384,6attention heads) serves all11tasks.
+Exact counts: **21,306,271 shared encoder parameters;24,769,764 total parameters**,
+plus769fit-derived normalization buffer values. Inference loads one model.pt.
+Shared pooling combines mean, task-query attention, maximum and terminal summaries;
+pair tasks also share product/difference processing. Every task was jointly trained
+through this same encoder.
+
+All learned weights originated from random initialization in this session: seed42,
+four hours of mounted UniRef50 MLM, followed by15active minutes of UniRef-only MLM
+(seed44;9,698,448,888 total input tokens). No external weights, embeddings or data.
+Then seed1 jointly trained all11fit tasks for1800retained active seconds with
+normalized regression, correlation coefficient0.1, square-root row sampling,
+human-PPI sampling multiplier0.25, head learning-rate multiplier2 and EMA0.999.
+Two interruptions discarded at least345.91additional active seconds before exact
+checkpoint recovery; retained and total consumed compute differ.
+
+The backbone and all native heads were averaged as complete checkpoints at
+steps17055,24750,32060,46433 (EMA), using0.25for every tensor. No task chooses its
+own encoder or checkpoint. Subsequent fit-only readouts use this frozen encoder:
+
+- Fluorescence: linear fit-label-rank ridge readout.
+- Stability: jointly trained scalar MLP.
+- Beta-lactamase: rank ridge base plus64-wide residual(selected11600of24000steps).
+- Affinity: linear ridge readout in original label units.
+- Yeast PPI: linear classifier on shared pooled sequence representations.
+- Solubility, subcellular/binary localization, fold and human PPI: joint MLP heads.
+- Secondary structure: joint per-residue MLP, using the same encoder.
+
+Data: mounted UniRef50 unlabeled sequences and all11labeled fit splits. Tune selects
+complete checkpoints and fixed ridge/readout grids; it does not fit these weights.
+GFP rank adoption requires≥0.01gain and a positive paired interval; beta rank
+requires≥0.02and a positive interval. Yeast required≥0.03and a positive paired-row
+interval versus its own native head. No training examples, feature caches or
+input lookup tables are packaged; no parameters are fitted at inference.
+
+Independent all11tune replay: **0.715300719 equal-weight mean**. This is not a hidden
+test score. Selection improves eight task point estimates versus the prior
+0.713444736package, while fluorescence(-0.040378),fold(-0.019022) and secondary
+structure(-0.001259) decline. Mean paired intervals spanzero; fluorescence/SS
+intervals are negative. Rare-fold accuracy remains1/92visible examples.
+Rank targets reduce conditional Spearman in the lowest fit-label quartile versus
+their raw-parent readouts: GFP0.5594→0.5216,beta0.5871→0.3237. Subgroup metrics do
+not add to overall Spearman. Tune has no aligned GFP examples with4+substitutions.
+One seed, repeated selection, small PPI splits and protein-family dependence limit
+confidence; the yeast gain over its native head has a grouped interval touching0.
+
+GFP/beta outputs use fit-label rank scales; stability/affinity use original units.
+Global and pair tasks truncate at1022residues. Secondary structure covers the full
+chain with overlapping1022-residue windows,stride894,taper64. Configured scalar and
+pooled readouts use FP32; other CUDA encoder paths use BF16. Empty/nonstandard
+chains and full1200-residue structure passed cold entry-point checks.
+
+Run offline: `bash infer.sh INPUT_ROOT OUTPUT_ROOT`.

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S002/config.json

@@ -0,0 +1,34 @@
+{
+  "name": "unified_t12_21M",
+  "layers": 12,
+  "dim": 384,
+  "heads": 6,
+  "structured_pooling": true,
+  "full_chain_ss": true,
+  "full_precision_head_tasks": [
+    "fluorescence",
+    "stability",
+    "beta_lactamase",
+    "ppi_affinity"
+  ],
+  "weight_variant": "budget_snapshot_average",
+  "full_precision_tasks": [
+    "beta_lactamase",
+    "fluorescence",
+    "ppi_affinity",
+    "yeast_ppi"
+  ],
+  "residual_readouts": {
+    "beta_lactamase": 64
+  },
+  "pooled_linear_classifiers": {
+    "yeast_ppi": 768
+  },
+  "fp32_head_tasks": [],
+  "rank_target_tasks": [
+    "fluorescence"
+  ],
+  "rank_residual_tasks": [
+    "beta_lactamase"
+  ]
+}

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S002/predict.py

@@ -0,0 +1,122 @@
+#!/usr/bin/env python3
+"""Self-contained offline inference, also used for in-training selection."""
+import json
+import sys
+from pathlib import Path
+
+import torch
+from esm2 import Model, encode, pad_batch
+
+TASKS = ('fluorescence', 'stability', 'beta_lactamase', 'solubility',
+         'subcellular_localization', 'binary_localization', 'fold',
+         'secondary_structure', 'yeast_ppi', 'human_ppi', 'ppi_affinity')
+
+
+def batches(rows, budget=16384, pair=False):
+    def length(i):
+        return min(1024, max(len(s) for s in rows[i]['features'].values()) + 2)
+    order = sorted(range(len(rows)), key=length)
+    result, width = [], 0
+    for i in order:
+        candidate = max(width, length(i))
+        if result and candidate * (len(result) + 1) * (2 if pair else 1) > budget:
+            yield result
+            result, width = [], 0
+        result.append(i)
+        width = max(width, candidate)
+    if result:
+        yield result
+
+
+@torch.inference_mode()
+def predict_records(model, rows, task, device, full_chain=True, budget=16384):
+    model.eval()
+    found = {}
+    use_amp = device.type == 'cuda' and task not in getattr(model, 'full_precision_tasks', ())
+    if task == 'secondary_structure':
+        windows = []
+        buffers = [torch.zeros(len(r['features']['sequence']), 3) for r in rows]
+        counts = [torch.zeros(len(r['features']['sequence']), 1) for r in rows]
+        for i, row in enumerate(rows):
+            sequence = row['features']['sequence']
+            last = max(0, len(sequence) - 1022)
+            starts = sorted(set(list(range(0, last, 894)) + [last])) if full_chain else [0]
+            for start in starts:
+                windows.append({'features': {'sequence': sequence[start:start+1022]},
+                                'row': i, 'start': start})
+        for indices in batches(windows, budget):
+            selected = [windows[i] for i in indices]
+            tokens = pad_batch([encode(w['features']['sequence']) for w in selected], device)
+            with torch.autocast(device.type, dtype=torch.bfloat16, enabled=use_amp):
+                hidden, _ = model.encoder(tokens)
+                probabilities = model.heads[task](hidden).float().softmax(-1).cpu()
+            for j, window in enumerate(selected):
+                i, start = window['row'], window['start']
+                length = len(window['features']['sequence'])
+                weight = torch.arange(1, length + 1).minimum(torch.arange(length, 0, -1)).clamp(max=64).float()[:, None]
+                buffers[i][start:start+length] += probabilities[j, 1:1+length] * weight
+                counts[i][start:start+length] += weight
+        for i, row in enumerate(rows):
+            score = buffers[i] / counts[i].clamp(min=1)
+            score[counts[i][:, 0] == 0] = 1 / 3
+            found[row['example_id']] = {'ss3_scores': score.tolist()}
+        return found
+    for indices in batches(rows, budget, task in Model.PAIR):
+        chunk = [rows[i] for i in indices]
+        with torch.autocast(device.type, dtype=torch.bfloat16, enabled=use_amp):
+            representation = model.represent(task, [r['features'] for r in chunk], device, 1024)
+            if task in getattr(model, 'full_precision_head_tasks', ()):
+                with torch.autocast(device.type, enabled=False):
+                    output = model.heads[task](representation.float())
+            elif task in getattr(model, 'fp32_head_tasks', ()):
+                hidden = model.heads[task][:-1](representation)
+                with torch.autocast(device.type, enabled=False):
+                    output = model.heads[task][-1](hidden.float())
+            else:
+                output = model.heads[task](representation).float()
+        if task in Model.CLASSES:
+            predictions = [{'scores': v} for v in output.softmax(-1).clamp(min=1e-9).cpu().tolist()]
+        else:
+            predictions = [{'value': v[0]} for v in output.cpu().tolist()]
+        found.update((row['example_id'], prediction) for row, prediction in zip(chunk, predictions))
+    return found
+
+
+def main():
+    if len(sys.argv) != 3:
+        raise SystemExit('usage: predict.py INPUT_ROOT OUTPUT_ROOT')
+    torch.set_num_threads(4)
+    torch.backends.cuda.matmul.allow_tf32 = False
+    here = Path(__file__).resolve().parent
+    source, out = Path(sys.argv[1]), Path(sys.argv[2]) / 'predictions'
+    out.mkdir(parents=True, exist_ok=False)
+    config = json.loads((here / 'config.json').read_text())
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    model_args = {k: config[k] for k in ['layers', 'dim', 'heads']}
+    if 'structured_pooling' in config:
+        model_args['structured_pooling'] = config['structured_pooling']
+    model = Model(**model_args).to(device)
+    if config.get('residual_readouts'):
+        from esm2 import install_residual_readouts
+        install_residual_readouts(model, config['residual_readouts'])
+    if config.get('pooled_linear_classifiers'):
+        from esm2 import install_pooled_classifiers
+        install_pooled_classifiers(model, config['pooled_linear_classifiers'])
+    model.load_state_dict(torch.load(here / 'model.pt', map_location=device, weights_only=True))
+    model.fp32_head_tasks = set(config.get('fp32_head_tasks', []))
+    model.full_precision_tasks = set(config.get('full_precision_tasks', []))
+    model.full_precision_head_tasks = set(config.get('full_precision_head_tasks', []))
+    model.eval()
+    for task in TASKS:
+        rows = [json.loads(line) for line in (source / f'{task}.jsonl').open()]
+        found = predict_records(model, rows, task, device,
+                                full_chain=config.get('full_chain_ss', True))
+        with (out / f'{task}.jsonl').open('w') as f:
+            for row in rows:
+                f.write(json.dumps({'example_id': row['example_id'],
+                                    'prediction': found[row['example_id']]}, separators=(',', ':')) + '\n')
+        print(f'{task}: {len(rows)} predictions', flush=True)
+
+
+if __name__ == '__main__':
+    main()

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S003/PUBLISHED.json

@@ -0,0 +1,5 @@
+{
+  "source": "/run-state/workspace/research/21m_4p25h_uniref_gfp_mutation_refined/publish_package",
+  "files": 8,
+  "bytes": 99314603
+}

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S003/config.json

@@ -0,0 +1,40 @@
+{
+  "name": "unified_t12_21M",
+  "layers": 12,
+  "dim": 384,
+  "heads": 6,
+  "structured_pooling": true,
+  "full_chain_ss": true,
+  "full_precision_head_tasks": [
+    "fluorescence",
+    "stability",
+    "beta_lactamase",
+    "ppi_affinity"
+  ],
+  "weight_variant": "budget_snapshot_average",
+  "full_precision_tasks": [
+    "beta_lactamase",
+    "fluorescence",
+    "ppi_affinity",
+    "yeast_ppi"
+  ],
+  "residual_readouts": {
+    "beta_lactamase": 64
+  },
+  "pooled_linear_classifiers": {
+    "yeast_ppi": 768
+  },
+  "fp32_head_tasks": [],
+  "rank_target_tasks": [
+    "fluorescence"
+  ],
+  "rank_residual_tasks": [
+    "beta_lactamase"
+  ],
+  "mutation_readouts": {
+    "fluorescence": {
+      "reference": "SKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLKFICTTGKLPVPWPTLVTTLSYGVQCFSRYPDHMKQHDFFKSAMPEGYVQERTIFFKDDGNYKTRAEVKFEGDTLVNRIELKGIDFKEDGNILGHKLEYNYNSHNVYIMADKQKNGIKVNFKIRHNIEDGSVQLADHYQQNTPIGDGPVLLPDNHYLSTQSALSKDPNEKRDHMVLLEFVTAAGITHGMDELYK",
+      "feature_policy": "Sum shared contextual substitution deltas; same-length only; fit-reference encoded by the same backbone."
+    }
+  }
+}

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S003/esm2.py

@@ -0,0 +1,281 @@
+#!/usr/bin/env python3
+"""An ESM-2 architecture, trained from scratch, as the reference baseline.
+
+Follows Lin et al. (Science 2023) / facebookresearch/esm in the ways that
+matter for a like-for-like architecture claim:
+
+  * rotary position embeddings, not learned positional embeddings -- this is
+    the defining architectural change from ESM-1b to ESM-2
+  * pre-LayerNorm blocks, GELU feed-forward at 4x expansion
+  * final LayerNorm before the LM head, and the LM head's projection weight
+    tied to the token embedding
+  * the ESM alphabet's special tokens (<cls>/<pad>/<eos>/<mask>) and the same
+    20 standard + rare amino-acid symbols
+
+The published size ladder is reproduced in CONFIGS so the baseline can be named
+by its ESM-2 counterpart rather than by an arbitrary width.
+"""
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# ESM-2's alphabet, in its canonical order.
+SPECIALS = ["<cls>", "<pad>", "<eos>", "<unk>"]
+RESIDUES = list("LAGVSERTIDPKQNFYMHWCXBUZO")
+EXTRA = ["<null_1>", "<mask>"]
+ALPHABET = SPECIALS + RESIDUES + EXTRA
+TOKEN = {symbol: index for index, symbol in enumerate(ALPHABET)}
+CLS, PAD, EOS, UNK = TOKEN["<cls>"], TOKEN["<pad>"], TOKEN["<eos>"], TOKEN["<unk>"]
+MASK = TOKEN["<mask>"]
+VOCAB_SIZE = len(ALPHABET)
+FIRST_RESIDUE, LAST_RESIDUE = TOKEN["L"], TOKEN["O"]
+
+# name -> (layers, embed_dim, attention_heads), from the ESM-2 release table.
+CONFIGS = {
+    "unified_t12_21M": (12, 384, 6),
+    "esm2_t6_8M":    (6, 320, 20),
+    "esm2_t12_35M":  (12, 480, 20),
+    "esm2_t30_150M": (30, 640, 20),
+    "esm2_t33_650M": (33, 1280, 20),
+}
+
+
+def encode(sequence, max_len=1024):
+    body = [TOKEN.get(c, UNK) for c in sequence[:max_len - 2]]
+    return [CLS] + body + [EOS]
+
+
+def pad_batch(sequences, device):
+    width = max(len(s) for s in sequences)
+    out = torch.full((len(sequences), width), PAD, dtype=torch.long)
+    for row, ids in enumerate(sequences):
+        out[row, :len(ids)] = torch.tensor(ids, dtype=torch.long)
+    return out.to(device, non_blocking=True)
+
+
+class Rotary(nn.Module):
+    """Rotary position embedding, as ESM-2 applies it to q and k."""
+
+    def __init__(self, dim, base=10000):
+        super().__init__()
+        inverse = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer("inverse", inverse, persistent=False)
+        self._cached_length = 0
+
+    def _tables(self, length, device, dtype):
+        if length > self._cached_length or self._cos.device != device:
+            position = torch.arange(length, device=device, dtype=torch.float)
+            angles = torch.outer(position, self.inverse.to(device))
+            table = torch.cat([angles, angles], dim=-1)
+            self._cos = table.cos()[None, None]
+            self._sin = table.sin()[None, None]
+            self._cached_length = length
+        return self._cos[:, :, :length].to(dtype), self._sin[:, :, :length].to(dtype)
+
+    @staticmethod
+    def _rotate_half(x):
+        half = x.shape[-1] // 2
+        return torch.cat([-x[..., half:], x[..., :half]], dim=-1)
+
+    def forward(self, q, k):
+        cos, sin = self._tables(q.shape[-2], q.device, q.dtype)
+        return (q * cos + self._rotate_half(q) * sin,
+                k * cos + self._rotate_half(k) * sin)
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads):
+        super().__init__()
+        self.heads = heads
+        self.head_dim = dim // heads
+        self.q = nn.Linear(dim, dim)
+        self.k = nn.Linear(dim, dim)
+        self.v = nn.Linear(dim, dim)
+        self.out = nn.Linear(dim, dim)
+        self.rotary = Rotary(self.head_dim)
+
+    def forward(self, x, pad_mask):
+        batch, length, _ = x.shape
+        shape = (batch, length, self.heads, self.head_dim)
+        q = self.q(x).view(shape).transpose(1, 2)
+        k = self.k(x).view(shape).transpose(1, 2)
+        v = self.v(x).view(shape).transpose(1, 2)
+        q, k = self.rotary(q, k)
+        attention_mask = None
+        if pad_mask is not None:
+            attention_mask = (~pad_mask)[:, None, None, :]
+        hidden = F.scaled_dot_product_attention(q, k, v, attn_mask=attention_mask)
+        hidden = hidden.transpose(1, 2).reshape(batch, length, -1)
+        return self.out(hidden)
+
+
+class Block(nn.Module):
+    def __init__(self, dim, heads, ffn):
+        super().__init__()
+        self.norm1 = nn.LayerNorm(dim)
+        self.attention = Attention(dim, heads)
+        self.norm2 = nn.LayerNorm(dim)
+        self.ffn = nn.Sequential(nn.Linear(dim, ffn), nn.GELU(), nn.Linear(ffn, dim))
+
+    def forward(self, x, pad_mask):
+        x = x + self.attention(self.norm1(x), pad_mask)
+        return x + self.ffn(self.norm2(x))
+
+
+class Encoder(nn.Module):
+    """The shared backbone. Every task reads from this one parameter set."""
+
+    def __init__(self, layers=12, dim=480, heads=20):
+        super().__init__()
+        self.dim = dim
+        self.embed = nn.Embedding(VOCAB_SIZE, dim, padding_idx=PAD)
+        self.blocks = nn.ModuleList([Block(dim, heads, dim * 4) for _ in range(layers)])
+        self.norm = nn.LayerNorm(dim)
+        self.lm_bias = nn.Parameter(torch.zeros(VOCAB_SIZE))
+        self.apply(self._init)
+
+    @staticmethod
+    def _init(module):
+        if isinstance(module, nn.Linear):
+            nn.init.normal_(module.weight, std=0.02)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, std=0.02)
+
+    def forward(self, tokens, unpadded=False):
+        pad_mask = tokens.eq(PAD)
+        hidden = self.embed(tokens)
+        for block in self.blocks:
+            hidden = block(hidden, None if unpadded else pad_mask)
+        return self.norm(hidden), pad_mask
+
+    def lm_logits(self, hidden):
+        # Weight tying, as in ESM-2's RobertaLMHead.
+        return F.linear(hidden, self.embed.weight) + self.lm_bias
+
+    def pool(self, tokens):
+        hidden, pad_mask = self.forward(tokens)
+        keep = (~pad_mask).to(hidden.dtype).unsqueeze(-1)
+        return (hidden * keep).sum(1) / keep.sum(1).clamp(min=1)
+
+
+class Model(nn.Module):
+    """ESM-2 backbone plus the eleven PEER task heads."""
+
+    CLASSES = {"solubility": 2, "subcellular_localization": 10,
+               "binary_localization": 2, "fold": 1195,
+               "yeast_ppi": 2, "human_ppi": 2}
+    REGRESSION = ("fluorescence", "stability", "beta_lactamase", "ppi_affinity")
+    PAIR = {"yeast_ppi", "human_ppi", "ppi_affinity"}
+
+    def __init__(self, layers=12, dim=480, heads=20, structured_pooling=True):
+        super().__init__()
+        self.structured_pooling = structured_pooling
+        self.encoder = Encoder(layers=layers, dim=dim, heads=heads)
+        self.task_names = list(self.CLASSES) + list(self.REGRESSION)
+        self.task_query = nn.ParameterDict({
+            task: nn.Parameter(torch.randn(dim) * 0.02) for task in self.task_names})
+        self.pool_projection = nn.Linear(dim * 4, dim, bias=False)
+        self.pool_gain = nn.ParameterDict({
+            task: nn.Parameter(torch.ones(dim)) for task in self.task_names})
+        self.pair_projection = nn.Linear(dim * 2, dim * 2, bias=False)
+        # Residual additions preserve the original mean-pool predictor when
+        # transferring this session's initial checkpoint.
+        nn.init.zeros_(self.pool_projection.weight)
+        nn.init.zeros_(self.pair_projection.weight)
+        if not structured_pooling:
+            # Consume identical initialization RNG for controlled comparisons,
+            # then remove parameters absent from this ablation's inference.
+            del self.task_query, self.pool_projection, self.pool_gain, self.pair_projection
+        self.heads = nn.ModuleDict()
+        for task, n in self.CLASSES.items():
+            width = dim * 2 if task in self.PAIR else dim
+            self.heads[task] = nn.Sequential(
+                nn.Linear(width, dim), nn.GELU(), nn.LayerNorm(dim), nn.Linear(dim, n))
+        for task in self.REGRESSION:
+            width = dim * 2 if task in self.PAIR else dim
+            self.heads[task] = nn.Sequential(
+                nn.Linear(width, dim), nn.GELU(), nn.LayerNorm(dim), nn.Linear(dim, 1))
+        self.heads["secondary_structure"] = nn.Sequential(
+            nn.Linear(dim, dim), nn.GELU(), nn.LayerNorm(dim), nn.Linear(dim, 3))
+
+    def pool_sequence(self, task, tokens):
+        if not self.structured_pooling:
+            return self.encoder.pool(tokens)
+        hidden, padding = self.encoder(tokens)
+        keep = (~padding).to(hidden.dtype).unsqueeze(-1)
+        mean = (hidden * keep).sum(1) / keep.sum(1).clamp(min=1)
+        residue = (tokens >= FIRST_RESIDUE) & (tokens <= LAST_RESIDUE)
+        safe = residue | ((~residue.any(1))[:, None] & tokens.eq(CLS))
+        score = torch.einsum('bld,d->bl', hidden, self.task_query[task])
+        score = score.masked_fill(~safe, -1e4)
+        attended = (hidden * score.float().softmax(-1).to(hidden.dtype).unsqueeze(-1)).sum(1)
+        maximum = hidden.masked_fill(~safe.unsqueeze(-1), -1e4).max(1).values
+        position = torch.arange(tokens.shape[1], device=tokens.device)[None, :]
+        end = (~padding).sum(1)[:, None] - 1
+        nmask = safe & (position <= 32)
+        cmask = safe & (position >= end - 32)
+        nterm = (hidden * nmask.unsqueeze(-1)).sum(1) / nmask.sum(1).clamp(min=1)[:, None]
+        cterm = (hidden * cmask.unsqueeze(-1)).sum(1) / cmask.sum(1).clamp(min=1)[:, None]
+        detail = torch.cat([attended - mean, maximum - mean,
+                            nterm - mean, cterm - mean], -1)
+        return mean + self.pool_gain[task] * self.pool_projection(detail)
+
+    def represent(self, task, batch, device, max_len=1024):
+        if task in self.PAIR:
+            a = self.pool_sequence(task, pad_batch(
+                [encode(x["sequence_a"], max_len) for x in batch], device))
+            b = self.pool_sequence(task, pad_batch(
+                [encode(x["sequence_b"], max_len) for x in batch], device))
+            base = torch.cat([a, b], dim=-1)
+            if not self.structured_pooling:
+                return base
+            interaction = torch.cat([a * b, torch.abs(a - b)], dim=-1)
+            return base + self.pair_projection(interaction)
+        return self.pool_sequence(task, pad_batch(
+            [encode(x["sequence"], max_len) for x in batch], device))
+
+
+def build(name="esm2_t12_35M", structured_pooling=True):
+    layers, dim, heads = CONFIGS[name]
+    model = Model(layers=layers, dim=dim, heads=heads,
+                  structured_pooling=structured_pooling)
+    return model, {"name": name, "layers": layers, "dim": dim, "heads": heads}
+
+
+if __name__ == "__main__":
+    for name in CONFIGS:
+        model, config = build(name)
+        backbone = sum(p.numel() for p in model.encoder.parameters())
+        total = sum(p.numel() for p in model.parameters())
+        print(f"{name:16s} backbone {backbone/1e6:7.2f}M   with heads {total/1e6:7.2f}M")
+
+
+class ResidualLinear(nn.Module):
+    """Small fit-trained nonlinear correction to one scalar Linear readout."""
+    def __init__(self, base, width):
+        super().__init__()
+        self.base = base
+        self.residual = nn.Sequential(nn.Linear(base.in_features, width), nn.GELU(),
+                                      nn.Dropout(0.1), nn.Linear(width, 1))
+        self.register_buffer("feature_center", torch.zeros(base.in_features))
+        self.register_buffer("feature_scale", torch.ones(base.in_features))
+        self.register_buffer("output_scale", torch.tensor(1.0))
+
+    def forward(self, hidden):
+        normalized = (hidden - self.feature_center) / self.feature_scale
+        return self.base(hidden) + self.output_scale * self.residual(normalized)
+
+
+def install_residual_readouts(model, specification):
+    for task, width in specification.items():
+        base = model.heads[task][-1]
+        model.heads[task][-1] = ResidualLinear(base, width).to(base.weight.device)
+
+
+def install_pooled_classifiers(model, specs):
+    for task, input_dim in specs.items():
+        model.heads[task] = torch.nn.Linear(input_dim, model.CLASSES[task]).to(next(model.parameters()).device)

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S003/predict.py

@@ -0,0 +1,125 @@
+#!/usr/bin/env python3
+"""Self-contained offline inference, also used for in-training selection."""
+import json
+import sys
+from pathlib import Path
+
+import torch
+from esm2 import Model, encode, pad_batch
+
+TASKS = ('fluorescence', 'stability', 'beta_lactamase', 'solubility',
+         'subcellular_localization', 'binary_localization', 'fold',
+         'secondary_structure', 'yeast_ppi', 'human_ppi', 'ppi_affinity')
+
+
+def batches(rows, budget=16384, pair=False):
+    def length(i):
+        return min(1024, max(len(s) for s in rows[i]['features'].values()) + 2)
+    order = sorted(range(len(rows)), key=length)
+    result, width = [], 0
+    for i in order:
+        candidate = max(width, length(i))
+        if result and candidate * (len(result) + 1) * (2 if pair else 1) > budget:
+            yield result
+            result, width = [], 0
+        result.append(i)
+        width = max(width, candidate)
+    if result:
+        yield result
+
+
+@torch.inference_mode()
+def predict_records(model, rows, task, device, full_chain=True, budget=16384):
+    model.eval()
+    found = {}
+    use_amp = device.type == 'cuda' and task not in getattr(model, 'full_precision_tasks', ())
+    if task == 'secondary_structure':
+        windows = []
+        buffers = [torch.zeros(len(r['features']['sequence']), 3) for r in rows]
+        counts = [torch.zeros(len(r['features']['sequence']), 1) for r in rows]
+        for i, row in enumerate(rows):
+            sequence = row['features']['sequence']
+            last = max(0, len(sequence) - 1022)
+            starts = sorted(set(list(range(0, last, 894)) + [last])) if full_chain else [0]
+            for start in starts:
+                windows.append({'features': {'sequence': sequence[start:start+1022]},
+                                'row': i, 'start': start})
+        for indices in batches(windows, budget):
+            selected = [windows[i] for i in indices]
+            tokens = pad_batch([encode(w['features']['sequence']) for w in selected], device)
+            with torch.autocast(device.type, dtype=torch.bfloat16, enabled=use_amp):
+                hidden, _ = model.encoder(tokens)
+                probabilities = model.heads[task](hidden).float().softmax(-1).cpu()
+            for j, window in enumerate(selected):
+                i, start = window['row'], window['start']
+                length = len(window['features']['sequence'])
+                weight = torch.arange(1, length + 1).minimum(torch.arange(length, 0, -1)).clamp(max=64).float()[:, None]
+                buffers[i][start:start+length] += probabilities[j, 1:1+length] * weight
+                counts[i][start:start+length] += weight
+        for i, row in enumerate(rows):
+            score = buffers[i] / counts[i].clamp(min=1)
+            score[counts[i][:, 0] == 0] = 1 / 3
+            found[row['example_id']] = {'ss3_scores': score.tolist()}
+        return found
+    for indices in batches(rows, budget, task in Model.PAIR):
+        chunk = [rows[i] for i in indices]
+        with torch.autocast(device.type, dtype=torch.bfloat16, enabled=use_amp):
+            representation = model.represent(task, [r['features'] for r in chunk], device, 1024)
+            if task in getattr(model, 'full_precision_head_tasks', ()):
+                with torch.autocast(device.type, enabled=False):
+                    output = model.heads[task](representation.float())
+            elif task in getattr(model, 'fp32_head_tasks', ()):
+                hidden = model.heads[task][:-1](representation)
+                with torch.autocast(device.type, enabled=False):
+                    output = model.heads[task][-1](hidden.float())
+            else:
+                output = model.heads[task](representation).float()
+        if task in Model.CLASSES:
+            predictions = [{'scores': v} for v in output.softmax(-1).clamp(min=1e-9).cpu().tolist()]
+        else:
+            predictions = [{'value': v[0]} for v in output.cpu().tolist()]
+        found.update((row['example_id'], prediction) for row, prediction in zip(chunk, predictions))
+    return found
+
+
+def main():
+    if len(sys.argv) != 3:
+        raise SystemExit('usage: predict.py INPUT_ROOT OUTPUT_ROOT')
+    torch.set_num_threads(4)
+    torch.backends.cuda.matmul.allow_tf32 = False
+    here = Path(__file__).resolve().parent
+    source, out = Path(sys.argv[1]), Path(sys.argv[2]) / 'predictions'
+    out.mkdir(parents=True, exist_ok=False)
+    config = json.loads((here / 'config.json').read_text())
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    model_args = {k: config[k] for k in ['layers', 'dim', 'heads']}
+    if 'structured_pooling' in config:
+        model_args['structured_pooling'] = config['structured_pooling']
+    model = Model(**model_args).to(device)
+    if config.get('residual_readouts'):
+        from esm2 import install_residual_readouts
+        install_residual_readouts(model, config['residual_readouts'])
+    if config.get('pooled_linear_classifiers'):
+        from esm2 import install_pooled_classifiers
+        install_pooled_classifiers(model, config['pooled_linear_classifiers'])
+    if config.get('mutation_readouts'):
+        from mutation_readout_runtime import install_mutation_readouts
+        install_mutation_readouts(model, config['mutation_readouts'])
+    model.load_state_dict(torch.load(here / 'model.pt', map_location=device, weights_only=True))
+    model.fp32_head_tasks = set(config.get('fp32_head_tasks', []))
+    model.full_precision_tasks = set(config.get('full_precision_tasks', []))
+    model.full_precision_head_tasks = set(config.get('full_precision_head_tasks', []))
+    model.eval()
+    for task in TASKS:
+        rows = [json.loads(line) for line in (source / f'{task}.jsonl').open()]
+        found = predict_records(model, rows, task, device,
+                                full_chain=config.get('full_chain_ss', True))
+        with (out / f'{task}.jsonl').open('w') as f:
+            for row in rows:
+                f.write(json.dumps({'example_id': row['example_id'],
+                                    'prediction': found[row['example_id']]}, separators=(',', ':')) + '\n')
+        print(f'{task}: {len(rows)} predictions', flush=True)
+
+
+if __name__ == '__main__':
+    main()

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S004/PUBLISHED.json

@@ -0,0 +1,5 @@
+{
+  "source": "/run-state/workspace/research/21m_4h_joint_gfp_function_rank_refined/publish_package",
+  "files": 7,
+  "bytes": 99262666
+}

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S004/README.md

@@ -0,0 +1,91 @@
+# Unified PEER protein sequence model
+
+One shared rotary transformer serves all eleven tasks: 12 layers,
+width 384, 6 attention heads. Exact counts: **21,306,271 shared
+encoder parameters; 24,769,764 total parameters**, plus 769 normalization
+buffer values fitted on training data. There is one model.pt and one complete parameter set.
+
+
+
+Shared pooling combines mean, task-query attention, maximum and terminal
+summaries through a learned shared projection. Pair tasks also use a shared
+product/difference projection.
+
+
+
+
+
+
+
+
+Classification MLPs serve
+solubility, subcellular_localization, binary_localization, fold, human_ppi. Scalar MLPs serve
+fluorescence, stability, beta-lactamase and affinity. Secondary structure uses
+a per-residue MLP.
+Additional scalar residual readouts and hidden widths: {"beta_lactamase": 64}.
+Additional classification residual readouts and hidden widths: {}.
+Direct linear classifiers over shared pooled states are fitted by ridge regression
+on fit examples and replace the entire MLP for:
+yeast_ppi.
+
+
+All learned weights originate from random initialization and training in this
+session. The seed-42 encoder and initial heads used no external checkpoints or
+embeddings. MLM consumed 9,127,835,498 input tokens from the full mounted decontaminated UniRef50 corpus.
+
+
+Joint supervised training used all eleven
+fit pools, normalized regression losses, sqrt(row-count) task sampling and seed 1.
+Task sampling multipliers: {"human_ppi":0.25}.
+Correlation-loss coefficient: 0.1.
+Encoder peak learning rate: 0.0001. Nonencoder
+parameters, including shared pooling and task heads, use peak rate
+0.0002.
+
+
+Joint training transforms only the fit labels of fluorescence to empirical average-rank quantiles and recomputes their fit normalization. Their native outputs use a rank scale; original tune labels are retained for Spearman scoring. This changes the targets for their supervised and correlation losses while all eleven tasks still update the same encoder. Later readout refinements, if present, are described below.
+
+
+
+Before readout refinements, the joint parameter set was initialized by averaging complete joint snapshots with the same coefficients for every tensor: step 15493 (ema) at 600.1s, weight 0.25, step 23025 (ema) at 900.1s, weight 0.25, step 23025 (ema) at 900.1s, weight 0.25, step 46713 (ema) at 1800.0s, weight 0.25. The coefficients were fixed before evaluating the average. The readout changes described here follow that initialization; inference uses one model.pt, with no backbone ensemble.
+
+
+EMA decay was 0.999.
+
+
+Fit-only regression refinement replaced final linear layers with ridge readouts
+of this same frozen shared encoder for:
+beta_lactamase, ppi_affinity.
+Rank-target ridge readouts replace the final linear layer for fluorescence. They fit empirical fit-label ranks on this same shared representation, using the same nine ridge strengths. Their values estimate fit-label quantiles rather than the original physical scale; these tasks are evaluated by Spearman correlation. Adoption requires at least 0.01 tune gain and a positive paired row interval before independent full replay.
+Rank-target ridge bases and residual readouts replace the existing scalar readouts for beta_lactamase. They use fit-label average-rank quantiles on the same frozen shared representation, with unchanged residual capacity. Their outputs use a rank scale rather than physical units. Adoption requires at least 0.02 tune gain and a positive paired row interval before independent full replay.
+Residual readouts were initialized and trained in this session on
+fit hidden states. Classifier final layers refined by fit-only ridge regression:
+none.
+
+No training sequences, feature caches or input lookup tables
+are packaged. No parameters are fitted at inference.
+
+Model selection uses all eleven tune rewards, their equal-weight mean and
+per-task changes. Tune selects joint snapshots, ridge regularization, residual
+width and checkpoint. Scalar ridge adoption requires 0.01 tune gain; residual
+adoption requires 0.02. Automatic classifier checks require a positive paired bootstrap
+interval and at least 0.02 gain for fold, 0.03 for PPI, or 0.01 otherwise.
+yeast_ppi: pooled-head gain is measured against native; its paired interval compares with its own unchanged float32 head.
+
+
+
+Independent frozen-package replay gives **0.721618 mean
+tune reward**. Detailed ancestry, settings and every task score are in
+training_history.json. This is not a held-out test result.
+
+Tasks using full float32 inference: beta_lactamase, fluorescence, ppi_affinity, yeast_ppi.
+Other
+regression heads use float32 with a bfloat16 backbone on CUDA. Global and pair
+inputs truncate at1022 residues; secondary structure covers every residue with
+overlapping windows, without cross-window long-range context. Small PPI tune
+sets and repeated model selection limit confidence in small differences.
+
+Run: bash infer.sh INPUT_ROOT OUTPUT_ROOT. The package loads independently of
+the trainer and runs offline with PyTorch.
+
+Publication selection: mean tune reward 0.7216184066 versus the previous0.7180249283. Seven task point scores improve and four decline. Both aggregate paired95% intervals span zero; GFP falls0.02129 with a negative interval. This is a measured selection, not a demonstrated hidden-test gain. Beta's original readout gate is narrowly positive and sensitive to bootstrap Monte Carlo sampling. Its lowest fit-label quartile correlation declines0.20343 versus the common parent, while the third quartile improves. Rare-fold support1-5 accuracy is1/92. No aligned4+GFPmutants appear in tune. The prior live arm used15additional MLMminutes and incurred discarded interrupted training compute.

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S004/esm2.py

@@ -0,0 +1,284 @@
+#!/usr/bin/env python3
+"""An ESM-2 architecture, trained from scratch, as the reference baseline.
+
+Follows Lin et al. (Science 2023) / facebookresearch/esm in the ways that
+matter for a like-for-like architecture claim:
+
+  * rotary position embeddings, not learned positional embeddings -- this is
+    the defining architectural change from ESM-1b to ESM-2
+  * pre-LayerNorm blocks, GELU feed-forward at 4x expansion
+  * final LayerNorm before the LM head, and the LM head's projection weight
+    tied to the token embedding
+  * the ESM alphabet's special tokens (<cls>/<pad>/<eos>/<mask>) and the same
+    20 standard + rare amino-acid symbols
+
+The published size ladder is reproduced in CONFIGS so the baseline can be named
+by its ESM-2 counterpart rather than by an arbitrary width.
+"""
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# ESM-2's alphabet, in its canonical order.
+SPECIALS = ["<cls>", "<pad>", "<eos>", "<unk>"]
+RESIDUES = list("LAGVSERTIDPKQNFYMHWCXBUZO")
+EXTRA = ["<null_1>", "<mask>"]
+ALPHABET = SPECIALS + RESIDUES + EXTRA
+TOKEN = {symbol: index for index, symbol in enumerate(ALPHABET)}
+CLS, PAD, EOS, UNK = TOKEN["<cls>"], TOKEN["<pad>"], TOKEN["<eos>"], TOKEN["<unk>"]
+MASK = TOKEN["<mask>"]
+VOCAB_SIZE = len(ALPHABET)
+FIRST_RESIDUE, LAST_RESIDUE = TOKEN["L"], TOKEN["O"]
+
+# name -> (layers, embed_dim, attention_heads), from the ESM-2 release table.
+CONFIGS = {
+    "unified_t12_21M": (12, 384, 6),
+    "esm2_t6_8M":    (6, 320, 20),
+    "esm2_t12_35M":  (12, 480, 20),
+    "esm2_t30_150M": (30, 640, 20),
+    "esm2_t33_650M": (33, 1280, 20),
+}
+
+
+def encode(sequence, max_len=1024):
+    body = [TOKEN.get(c, UNK) for c in sequence[:max_len - 2]]
+    return [CLS] + body + [EOS]
+
+
+def pad_batch(sequences, device):
+    width = max(len(s) for s in sequences)
+    out = torch.full((len(sequences), width), PAD, dtype=torch.long)
+    for row, ids in enumerate(sequences):
+        out[row, :len(ids)] = torch.tensor(ids, dtype=torch.long)
+    return out.to(device, non_blocking=True)
+
+
+class Rotary(nn.Module):
+    """Rotary position embedding, as ESM-2 applies it to q and k."""
+
+    def __init__(self, dim, base=10000):
+        super().__init__()
+        inverse = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer("inverse", inverse, persistent=False)
+        self._cached_length = 0
+
+    def _tables(self, length, device, dtype):
+        if length > self._cached_length or self._cos.device != device:
+            # Evaluation can first encounter a longer sequence. Cache ordinary
+            # tensors so subsequent training can save them for backward.
+            with torch.inference_mode(False), torch.no_grad():
+                position = torch.arange(length, device=device, dtype=torch.float)
+                angles = torch.outer(position, self.inverse.to(device))
+                table = torch.cat([angles, angles], dim=-1)
+                self._cos = table.cos()[None, None]
+                self._sin = table.sin()[None, None]
+            self._cached_length = length
+        return self._cos[:, :, :length].to(dtype), self._sin[:, :, :length].to(dtype)
+
+    @staticmethod
+    def _rotate_half(x):
+        half = x.shape[-1] // 2
+        return torch.cat([-x[..., half:], x[..., :half]], dim=-1)
+
+    def forward(self, q, k):
+        cos, sin = self._tables(q.shape[-2], q.device, q.dtype)
+        return (q * cos + self._rotate_half(q) * sin,
+                k * cos + self._rotate_half(k) * sin)
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads):
+        super().__init__()
+        self.heads = heads
+        self.head_dim = dim // heads
+        self.q = nn.Linear(dim, dim)
+        self.k = nn.Linear(dim, dim)
+        self.v = nn.Linear(dim, dim)
+        self.out = nn.Linear(dim, dim)
+        self.rotary = Rotary(self.head_dim)
+
+    def forward(self, x, pad_mask):
+        batch, length, _ = x.shape
+        shape = (batch, length, self.heads, self.head_dim)
+        q = self.q(x).view(shape).transpose(1, 2)
+        k = self.k(x).view(shape).transpose(1, 2)
+        v = self.v(x).view(shape).transpose(1, 2)
+        q, k = self.rotary(q, k)
+        attention_mask = None
+        if pad_mask is not None:
+            attention_mask = (~pad_mask)[:, None, None, :]
+        hidden = F.scaled_dot_product_attention(q, k, v, attn_mask=attention_mask)
+        hidden = hidden.transpose(1, 2).reshape(batch, length, -1)
+        return self.out(hidden)
+
+
+class Block(nn.Module):
+    def __init__(self, dim, heads, ffn):
+        super().__init__()
+        self.norm1 = nn.LayerNorm(dim)
+        self.attention = Attention(dim, heads)
+        self.norm2 = nn.LayerNorm(dim)
+        self.ffn = nn.Sequential(nn.Linear(dim, ffn), nn.GELU(), nn.Linear(ffn, dim))
+
+    def forward(self, x, pad_mask):
+        x = x + self.attention(self.norm1(x), pad_mask)
+        return x + self.ffn(self.norm2(x))
+
+
+class Encoder(nn.Module):
+    """The shared backbone. Every task reads from this one parameter set."""
+
+    def __init__(self, layers=12, dim=480, heads=20):
+        super().__init__()
+        self.dim = dim
+        self.embed = nn.Embedding(VOCAB_SIZE, dim, padding_idx=PAD)
+        self.blocks = nn.ModuleList([Block(dim, heads, dim * 4) for _ in range(layers)])
+        self.norm = nn.LayerNorm(dim)
+        self.lm_bias = nn.Parameter(torch.zeros(VOCAB_SIZE))
+        self.apply(self._init)
+
+    @staticmethod
+    def _init(module):
+        if isinstance(module, nn.Linear):
+            nn.init.normal_(module.weight, std=0.02)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, std=0.02)
+
+    def forward(self, tokens, unpadded=False):
+        pad_mask = tokens.eq(PAD)
+        hidden = self.embed(tokens)
+        for block in self.blocks:
+            hidden = block(hidden, None if unpadded else pad_mask)
+        return self.norm(hidden), pad_mask
+
+    def lm_logits(self, hidden):
+        # Weight tying, as in ESM-2's RobertaLMHead.
+        return F.linear(hidden, self.embed.weight) + self.lm_bias
+
+    def pool(self, tokens):
+        hidden, pad_mask = self.forward(tokens)
+        keep = (~pad_mask).to(hidden.dtype).unsqueeze(-1)
+        return (hidden * keep).sum(1) / keep.sum(1).clamp(min=1)
+
+
+class Model(nn.Module):
+    """ESM-2 backbone plus the eleven PEER task heads."""
+
+    CLASSES = {"solubility": 2, "subcellular_localization": 10,
+               "binary_localization": 2, "fold": 1195,
+               "yeast_ppi": 2, "human_ppi": 2}
+    REGRESSION = ("fluorescence", "stability", "beta_lactamase", "ppi_affinity")
+    PAIR = {"yeast_ppi", "human_ppi", "ppi_affinity"}
+
+    def __init__(self, layers=12, dim=480, heads=20, structured_pooling=True):
+        super().__init__()
+        self.structured_pooling = structured_pooling
+        self.encoder = Encoder(layers=layers, dim=dim, heads=heads)
+        self.task_names = list(self.CLASSES) + list(self.REGRESSION)
+        self.task_query = nn.ParameterDict({
+            task: nn.Parameter(torch.randn(dim) * 0.02) for task in self.task_names})
+        self.pool_projection = nn.Linear(dim * 4, dim, bias=False)
+        self.pool_gain = nn.ParameterDict({
+            task: nn.Parameter(torch.ones(dim)) for task in self.task_names})
+        self.pair_projection = nn.Linear(dim * 2, dim * 2, bias=False)
+        # Residual additions preserve the original mean-pool predictor when
+        # transferring this session's initial checkpoint.
+        nn.init.zeros_(self.pool_projection.weight)
+        nn.init.zeros_(self.pair_projection.weight)
+        if not structured_pooling:
+            # Consume identical initialization RNG for controlled comparisons,
+            # then remove parameters absent from this ablation's inference.
+            del self.task_query, self.pool_projection, self.pool_gain, self.pair_projection
+        self.heads = nn.ModuleDict()
+        for task, n in self.CLASSES.items():
+            width = dim * 2 if task in self.PAIR else dim
+            self.heads[task] = nn.Sequential(
+                nn.Linear(width, dim), nn.GELU(), nn.LayerNorm(dim), nn.Linear(dim, n))
+        for task in self.REGRESSION:
+            width = dim * 2 if task in self.PAIR else dim
+            self.heads[task] = nn.Sequential(
+                nn.Linear(width, dim), nn.GELU(), nn.LayerNorm(dim), nn.Linear(dim, 1))
+        self.heads["secondary_structure"] = nn.Sequential(
+            nn.Linear(dim, dim), nn.GELU(), nn.LayerNorm(dim), nn.Linear(dim, 3))
+
+    def pool_sequence(self, task, tokens):
+        if not self.structured_pooling:
+            return self.encoder.pool(tokens)
+        hidden, padding = self.encoder(tokens)
+        keep = (~padding).to(hidden.dtype).unsqueeze(-1)
+        mean = (hidden * keep).sum(1) / keep.sum(1).clamp(min=1)
+        residue = (tokens >= FIRST_RESIDUE) & (tokens <= LAST_RESIDUE)
+        safe = residue | ((~residue.any(1))[:, None] & tokens.eq(CLS))
+        score = torch.einsum('bld,d->bl', hidden, self.task_query[task])
+        score = score.masked_fill(~safe, -1e4)
+        attended = (hidden * score.float().softmax(-1).to(hidden.dtype).unsqueeze(-1)).sum(1)
+        maximum = hidden.masked_fill(~safe.unsqueeze(-1), -1e4).max(1).values
+        position = torch.arange(tokens.shape[1], device=tokens.device)[None, :]
+        end = (~padding).sum(1)[:, None] - 1
+        nmask = safe & (position <= 32)
+        cmask = safe & (position >= end - 32)
+        nterm = (hidden * nmask.unsqueeze(-1)).sum(1) / nmask.sum(1).clamp(min=1)[:, None]
+        cterm = (hidden * cmask.unsqueeze(-1)).sum(1) / cmask.sum(1).clamp(min=1)[:, None]
+        detail = torch.cat([attended - mean, maximum - mean,
+                            nterm - mean, cterm - mean], -1)
+        return mean + self.pool_gain[task] * self.pool_projection(detail)
+
+    def represent(self, task, batch, device, max_len=1024):
+        if task in self.PAIR:
+            a = self.pool_sequence(task, pad_batch(
+                [encode(x["sequence_a"], max_len) for x in batch], device))
+            b = self.pool_sequence(task, pad_batch(
+                [encode(x["sequence_b"], max_len) for x in batch], device))
+            base = torch.cat([a, b], dim=-1)
+            if not self.structured_pooling:
+                return base
+            interaction = torch.cat([a * b, torch.abs(a - b)], dim=-1)
+            return base + self.pair_projection(interaction)
+        return self.pool_sequence(task, pad_batch(
+            [encode(x["sequence"], max_len) for x in batch], device))
+
+
+def build(name="esm2_t12_35M", structured_pooling=True):
+    layers, dim, heads = CONFIGS[name]
+    model = Model(layers=layers, dim=dim, heads=heads,
+                  structured_pooling=structured_pooling)
+    return model, {"name": name, "layers": layers, "dim": dim, "heads": heads}
+
+
+if __name__ == "__main__":
+    for name in CONFIGS:
+        model, config = build(name)
+        backbone = sum(p.numel() for p in model.encoder.parameters())
+        total = sum(p.numel() for p in model.parameters())
+        print(f"{name:16s} backbone {backbone/1e6:7.2f}M   with heads {total/1e6:7.2f}M")
+
+
+class ResidualLinear(nn.Module):
+    """Small fit-trained nonlinear correction to one scalar Linear readout."""
+    def __init__(self, base, width):
+        super().__init__()
+        self.base = base
+        self.residual = nn.Sequential(nn.Linear(base.in_features, width), nn.GELU(),
+                                      nn.Dropout(0.1), nn.Linear(width, 1))
+        self.register_buffer("feature_center", torch.zeros(base.in_features))
+        self.register_buffer("feature_scale", torch.ones(base.in_features))
+        self.register_buffer("output_scale", torch.tensor(1.0))
+
+    def forward(self, hidden):
+        normalized = (hidden - self.feature_center) / self.feature_scale
+        return self.base(hidden) + self.output_scale * self.residual(normalized)
+
+
+def install_residual_readouts(model, specification):
+    for task, width in specification.items():
+        base = model.heads[task][-1]
+        model.heads[task][-1] = ResidualLinear(base, width).to(base.weight.device)
+
+
+def install_pooled_classifiers(model, specs):
+    for task, input_dim in specs.items():
+        model.heads[task] = torch.nn.Linear(input_dim, model.CLASSES[task]).to(next(model.parameters()).device)

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S004/infer.sh

@@ -0,0 +1,5 @@
+#!/usr/bin/env bash
+set -euo pipefail
+export OMP_NUM_THREADS=4
+export MKL_NUM_THREADS=4
+exec python3 "$(dirname "$0")/predict.py" "$1" "$2"

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S004/predict.py

@@ -0,0 +1,122 @@
+#!/usr/bin/env python3
+"""Self-contained offline inference, also used for in-training selection."""
+import json
+import sys
+from pathlib import Path
+
+import torch
+from esm2 import Model, encode, pad_batch
+
+TASKS = ('fluorescence', 'stability', 'beta_lactamase', 'solubility',
+         'subcellular_localization', 'binary_localization', 'fold',
+         'secondary_structure', 'yeast_ppi', 'human_ppi', 'ppi_affinity')
+
+
+def batches(rows, budget=16384, pair=False):
+    def length(i):
+        return min(1024, max(len(s) for s in rows[i]['features'].values()) + 2)
+    order = sorted(range(len(rows)), key=length)
+    result, width = [], 0
+    for i in order:
+        candidate = max(width, length(i))
+        if result and candidate * (len(result) + 1) * (2 if pair else 1) > budget:
+            yield result
+            result, width = [], 0
+        result.append(i)
+        width = max(width, candidate)
+    if result:
+        yield result
+
+
+@torch.inference_mode()
+def predict_records(model, rows, task, device, full_chain=True, budget=16384):
+    model.eval()
+    found = {}
+    use_amp = device.type == 'cuda' and task not in getattr(model, 'full_precision_tasks', ())
+    if task == 'secondary_structure':
+        windows = []
+        buffers = [torch.zeros(len(r['features']['sequence']), 3) for r in rows]
+        counts = [torch.zeros(len(r['features']['sequence']), 1) for r in rows]
+        for i, row in enumerate(rows):
+            sequence = row['features']['sequence']
+            last = max(0, len(sequence) - 1022)
+            starts = sorted(set(list(range(0, last, 894)) + [last])) if full_chain else [0]
+            for start in starts:
+                windows.append({'features': {'sequence': sequence[start:start+1022]},
+                                'row': i, 'start': start})
+        for indices in batches(windows, budget):
+            selected = [windows[i] for i in indices]
+            tokens = pad_batch([encode(w['features']['sequence']) for w in selected], device)
+            with torch.autocast(device.type, dtype=torch.bfloat16, enabled=use_amp):
+                hidden, _ = model.encoder(tokens)
+                probabilities = model.heads[task](hidden).float().softmax(-1).cpu()
+            for j, window in enumerate(selected):
+                i, start = window['row'], window['start']
+                length = len(window['features']['sequence'])
+                weight = torch.arange(1, length + 1).minimum(torch.arange(length, 0, -1)).clamp(max=64).float()[:, None]
+                buffers[i][start:start+length] += probabilities[j, 1:1+length] * weight
+                counts[i][start:start+length] += weight
+        for i, row in enumerate(rows):
+            score = buffers[i] / counts[i].clamp(min=1)
+            score[counts[i][:, 0] == 0] = 1 / 3
+            found[row['example_id']] = {'ss3_scores': score.tolist()}
+        return found
+    for indices in batches(rows, budget, task in Model.PAIR):
+        chunk = [rows[i] for i in indices]
+        with torch.autocast(device.type, dtype=torch.bfloat16, enabled=use_amp):
+            representation = model.represent(task, [r['features'] for r in chunk], device, 1024)
+            if task in getattr(model, 'full_precision_head_tasks', ()):
+                with torch.autocast(device.type, enabled=False):
+                    output = model.heads[task](representation.float())
+            elif task in getattr(model, 'fp32_head_tasks', ()):
+                hidden = model.heads[task][:-1](representation)
+                with torch.autocast(device.type, enabled=False):
+                    output = model.heads[task][-1](hidden.float())
+            else:
+                output = model.heads[task](representation).float()
+        if task in Model.CLASSES:
+            predictions = [{'scores': v} for v in output.softmax(-1).clamp(min=1e-9).cpu().tolist()]
+        else:
+            predictions = [{'value': v[0]} for v in output.cpu().tolist()]
+        found.update((row['example_id'], prediction) for row, prediction in zip(chunk, predictions))
+    return found
+
+
+def main():
+    if len(sys.argv) != 3:
+        raise SystemExit('usage: predict.py INPUT_ROOT OUTPUT_ROOT')
+    torch.set_num_threads(4)
+    torch.backends.cuda.matmul.allow_tf32 = False
+    here = Path(__file__).resolve().parent
+    source, out = Path(sys.argv[1]), Path(sys.argv[2]) / 'predictions'
+    out.mkdir(parents=True, exist_ok=False)
+    config = json.loads((here / 'config.json').read_text())
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    model_args = {k: config[k] for k in ['layers', 'dim', 'heads']}
+    if 'structured_pooling' in config:
+        model_args['structured_pooling'] = config['structured_pooling']
+    model = Model(**model_args).to(device)
+    if config.get('residual_readouts'):
+        from esm2 import install_residual_readouts
+        install_residual_readouts(model, config['residual_readouts'])
+    if config.get('pooled_linear_classifiers'):
+        from esm2 import install_pooled_classifiers
+        install_pooled_classifiers(model, config['pooled_linear_classifiers'])
+    model.load_state_dict(torch.load(here / 'model.pt', map_location=device, weights_only=True))
+    model.fp32_head_tasks = set(config.get('fp32_head_tasks', []))
+    model.full_precision_tasks = set(config.get('full_precision_tasks', []))
+    model.full_precision_head_tasks = set(config.get('full_precision_head_tasks', []))
+    model.eval()
+    for task in TASKS:
+        rows = [json.loads(line) for line in (source / f'{task}.jsonl').open()]
+        found = predict_records(model, rows, task, device,
+                                full_chain=config.get('full_chain_ss', True))
+        with (out / f'{task}.jsonl').open('w') as f:
+            for row in rows:
+                f.write(json.dumps({'example_id': row['example_id'],
+                                    'prediction': found[row['example_id']]}, separators=(',', ':')) + '\n')
+        print(f'{task}: {len(rows)} predictions', flush=True)
+
+
+if __name__ == '__main__':
+    main()

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S005/PUBLISHED.json

@@ -0,0 +1,5 @@
+{
+  "source": "/run-state/workspace/research/21m_4h_joint_gfp_mutation_refined/publish_package",
+  "files": 8,
+  "bytes": 99302981
+}

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S005/config.json

@@ -0,0 +1,43 @@
+{
+  "name": "unified_t12_21M",
+  "layers": 12,
+  "dim": 384,
+  "heads": 6,
+  "structured_pooling": true,
+  "full_chain_ss": true,
+  "full_precision_head_tasks": [
+    "fluorescence",
+    "stability",
+    "beta_lactamase",
+    "ppi_affinity"
+  ],
+  "rank_supervised_tasks": [
+    "fluorescence"
+  ],
+  "weight_variant": "budget_snapshot_average",
+  "full_precision_tasks": [
+    "beta_lactamase",
+    "fluorescence",
+    "ppi_affinity",
+    "yeast_ppi"
+  ],
+  "residual_readouts": {
+    "beta_lactamase": 64
+  },
+  "pooled_linear_classifiers": {
+    "yeast_ppi": 768
+  },
+  "fp32_head_tasks": [],
+  "rank_target_tasks": [
+    "fluorescence"
+  ],
+  "rank_residual_tasks": [
+    "beta_lactamase"
+  ],
+  "mutation_readouts": {
+    "fluorescence": {
+      "reference": "SKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLKFICTTGKLPVPWPTLVTTLSYGVQCFSRYPDHMKQHDFFKSAMPEGYVQERTIFFKDDGNYKTRAEVKFEGDTLVNRIELKGIDFKEDGNILGHKLEYNYNSHNVYIMADKQKNGIKVNFKIRHNIEDGSVQLADHYQQNTPIGDGPVLLPDNHYLSTQSALSKDPNEKRDHMVLLEFVTAAGITHGMDELYK",
+      "feature_policy": "Sum shared contextual substitution deltas; same-length only; fit-reference encoded by the same backbone."
+    }
+  }
+}

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S005/esm2.py

@@ -0,0 +1,284 @@
+#!/usr/bin/env python3
+"""An ESM-2 architecture, trained from scratch, as the reference baseline.
+
+Follows Lin et al. (Science 2023) / facebookresearch/esm in the ways that
+matter for a like-for-like architecture claim:
+
+  * rotary position embeddings, not learned positional embeddings -- this is
+    the defining architectural change from ESM-1b to ESM-2
+  * pre-LayerNorm blocks, GELU feed-forward at 4x expansion
+  * final LayerNorm before the LM head, and the LM head's projection weight
+    tied to the token embedding
+  * the ESM alphabet's special tokens (<cls>/<pad>/<eos>/<mask>) and the same
+    20 standard + rare amino-acid symbols
+
+The published size ladder is reproduced in CONFIGS so the baseline can be named
+by its ESM-2 counterpart rather than by an arbitrary width.
+"""
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# ESM-2's alphabet, in its canonical order.
+SPECIALS = ["<cls>", "<pad>", "<eos>", "<unk>"]
+RESIDUES = list("LAGVSERTIDPKQNFYMHWCXBUZO")
+EXTRA = ["<null_1>", "<mask>"]
+ALPHABET = SPECIALS + RESIDUES + EXTRA
+TOKEN = {symbol: index for index, symbol in enumerate(ALPHABET)}
+CLS, PAD, EOS, UNK = TOKEN["<cls>"], TOKEN["<pad>"], TOKEN["<eos>"], TOKEN["<unk>"]
+MASK = TOKEN["<mask>"]
+VOCAB_SIZE = len(ALPHABET)
+FIRST_RESIDUE, LAST_RESIDUE = TOKEN["L"], TOKEN["O"]
+
+# name -> (layers, embed_dim, attention_heads), from the ESM-2 release table.
+CONFIGS = {
+    "unified_t12_21M": (12, 384, 6),
+    "esm2_t6_8M":    (6, 320, 20),
+    "esm2_t12_35M":  (12, 480, 20),
+    "esm2_t30_150M": (30, 640, 20),
+    "esm2_t33_650M": (33, 1280, 20),
+}
+
+
+def encode(sequence, max_len=1024):
+    body = [TOKEN.get(c, UNK) for c in sequence[:max_len - 2]]
+    return [CLS] + body + [EOS]
+
+
+def pad_batch(sequences, device):
+    width = max(len(s) for s in sequences)
+    out = torch.full((len(sequences), width), PAD, dtype=torch.long)
+    for row, ids in enumerate(sequences):
+        out[row, :len(ids)] = torch.tensor(ids, dtype=torch.long)
+    return out.to(device, non_blocking=True)
+
+
+class Rotary(nn.Module):
+    """Rotary position embedding, as ESM-2 applies it to q and k."""
+
+    def __init__(self, dim, base=10000):
+        super().__init__()
+        inverse = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer("inverse", inverse, persistent=False)
+        self._cached_length = 0
+
+    def _tables(self, length, device, dtype):
+        if length > self._cached_length or self._cos.device != device:
+            # Evaluation can first encounter a longer sequence. Cache ordinary
+            # tensors so subsequent training can save them for backward.
+            with torch.inference_mode(False), torch.no_grad():
+                position = torch.arange(length, device=device, dtype=torch.float)
+                angles = torch.outer(position, self.inverse.to(device))
+                table = torch.cat([angles, angles], dim=-1)
+                self._cos = table.cos()[None, None]
+                self._sin = table.sin()[None, None]
+            self._cached_length = length
+        return self._cos[:, :, :length].to(dtype), self._sin[:, :, :length].to(dtype)
+
+    @staticmethod
+    def _rotate_half(x):
+        half = x.shape[-1] // 2
+        return torch.cat([-x[..., half:], x[..., :half]], dim=-1)
+
+    def forward(self, q, k):
+        cos, sin = self._tables(q.shape[-2], q.device, q.dtype)
+        return (q * cos + self._rotate_half(q) * sin,
+                k * cos + self._rotate_half(k) * sin)
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads):
+        super().__init__()
+        self.heads = heads
+        self.head_dim = dim // heads
+        self.q = nn.Linear(dim, dim)
+        self.k = nn.Linear(dim, dim)
+        self.v = nn.Linear(dim, dim)
+        self.out = nn.Linear(dim, dim)
+        self.rotary = Rotary(self.head_dim)
+
+    def forward(self, x, pad_mask):
+        batch, length, _ = x.shape
+        shape = (batch, length, self.heads, self.head_dim)
+        q = self.q(x).view(shape).transpose(1, 2)
+        k = self.k(x).view(shape).transpose(1, 2)
+        v = self.v(x).view(shape).transpose(1, 2)
+        q, k = self.rotary(q, k)
+        attention_mask = None
+        if pad_mask is not None:
+            attention_mask = (~pad_mask)[:, None, None, :]
+        hidden = F.scaled_dot_product_attention(q, k, v, attn_mask=attention_mask)
+        hidden = hidden.transpose(1, 2).reshape(batch, length, -1)
+        return self.out(hidden)
+
+
+class Block(nn.Module):
+    def __init__(self, dim, heads, ffn):
+        super().__init__()
+        self.norm1 = nn.LayerNorm(dim)
+        self.attention = Attention(dim, heads)
+        self.norm2 = nn.LayerNorm(dim)
+        self.ffn = nn.Sequential(nn.Linear(dim, ffn), nn.GELU(), nn.Linear(ffn, dim))
+
+    def forward(self, x, pad_mask):
+        x = x + self.attention(self.norm1(x), pad_mask)
+        return x + self.ffn(self.norm2(x))
+
+
+class Encoder(nn.Module):
+    """The shared backbone. Every task reads from this one parameter set."""
+
+    def __init__(self, layers=12, dim=480, heads=20):
+        super().__init__()
+        self.dim = dim
+        self.embed = nn.Embedding(VOCAB_SIZE, dim, padding_idx=PAD)
+        self.blocks = nn.ModuleList([Block(dim, heads, dim * 4) for _ in range(layers)])
+        self.norm = nn.LayerNorm(dim)
+        self.lm_bias = nn.Parameter(torch.zeros(VOCAB_SIZE))
+        self.apply(self._init)
+
+    @staticmethod
+    def _init(module):
+        if isinstance(module, nn.Linear):
+            nn.init.normal_(module.weight, std=0.02)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, std=0.02)
+
+    def forward(self, tokens, unpadded=False):
+        pad_mask = tokens.eq(PAD)
+        hidden = self.embed(tokens)
+        for block in self.blocks:
+            hidden = block(hidden, None if unpadded else pad_mask)
+        return self.norm(hidden), pad_mask
+
+    def lm_logits(self, hidden):
+        # Weight tying, as in ESM-2's RobertaLMHead.
+        return F.linear(hidden, self.embed.weight) + self.lm_bias
+
+    def pool(self, tokens):
+        hidden, pad_mask = self.forward(tokens)
+        keep = (~pad_mask).to(hidden.dtype).unsqueeze(-1)
+        return (hidden * keep).sum(1) / keep.sum(1).clamp(min=1)
+
+
+class Model(nn.Module):
+    """ESM-2 backbone plus the eleven PEER task heads."""
+
+    CLASSES = {"solubility": 2, "subcellular_localization": 10,
+               "binary_localization": 2, "fold": 1195,
+               "yeast_ppi": 2, "human_ppi": 2}
+    REGRESSION = ("fluorescence", "stability", "beta_lactamase", "ppi_affinity")
+    PAIR = {"yeast_ppi", "human_ppi", "ppi_affinity"}
+
+    def __init__(self, layers=12, dim=480, heads=20, structured_pooling=True):
+        super().__init__()
+        self.structured_pooling = structured_pooling
+        self.encoder = Encoder(layers=layers, dim=dim, heads=heads)
+        self.task_names = list(self.CLASSES) + list(self.REGRESSION)
+        self.task_query = nn.ParameterDict({
+            task: nn.Parameter(torch.randn(dim) * 0.02) for task in self.task_names})
+        self.pool_projection = nn.Linear(dim * 4, dim, bias=False)
+        self.pool_gain = nn.ParameterDict({
+            task: nn.Parameter(torch.ones(dim)) for task in self.task_names})
+        self.pair_projection = nn.Linear(dim * 2, dim * 2, bias=False)
+        # Residual additions preserve the original mean-pool predictor when
+        # transferring this session's initial checkpoint.
+        nn.init.zeros_(self.pool_projection.weight)
+        nn.init.zeros_(self.pair_projection.weight)
+        if not structured_pooling:
+            # Consume identical initialization RNG for controlled comparisons,
+            # then remove parameters absent from this ablation's inference.
+            del self.task_query, self.pool_projection, self.pool_gain, self.pair_projection
+        self.heads = nn.ModuleDict()
+        for task, n in self.CLASSES.items():
+            width = dim * 2 if task in self.PAIR else dim
+            self.heads[task] = nn.Sequential(
+                nn.Linear(width, dim), nn.GELU(), nn.LayerNorm(dim), nn.Linear(dim, n))
+        for task in self.REGRESSION:
+            width = dim * 2 if task in self.PAIR else dim
+            self.heads[task] = nn.Sequential(
+                nn.Linear(width, dim), nn.GELU(), nn.LayerNorm(dim), nn.Linear(dim, 1))
+        self.heads["secondary_structure"] = nn.Sequential(
+            nn.Linear(dim, dim), nn.GELU(), nn.LayerNorm(dim), nn.Linear(dim, 3))
+
+    def pool_sequence(self, task, tokens):
+        if not self.structured_pooling:
+            return self.encoder.pool(tokens)
+        hidden, padding = self.encoder(tokens)
+        keep = (~padding).to(hidden.dtype).unsqueeze(-1)
+        mean = (hidden * keep).sum(1) / keep.sum(1).clamp(min=1)
+        residue = (tokens >= FIRST_RESIDUE) & (tokens <= LAST_RESIDUE)
+        safe = residue | ((~residue.any(1))[:, None] & tokens.eq(CLS))
+        score = torch.einsum('bld,d->bl', hidden, self.task_query[task])
+        score = score.masked_fill(~safe, -1e4)
+        attended = (hidden * score.float().softmax(-1).to(hidden.dtype).unsqueeze(-1)).sum(1)
+        maximum = hidden.masked_fill(~safe.unsqueeze(-1), -1e4).max(1).values
+        position = torch.arange(tokens.shape[1], device=tokens.device)[None, :]
+        end = (~padding).sum(1)[:, None] - 1
+        nmask = safe & (position <= 32)
+        cmask = safe & (position >= end - 32)
+        nterm = (hidden * nmask.unsqueeze(-1)).sum(1) / nmask.sum(1).clamp(min=1)[:, None]
+        cterm = (hidden * cmask.unsqueeze(-1)).sum(1) / cmask.sum(1).clamp(min=1)[:, None]
+        detail = torch.cat([attended - mean, maximum - mean,
+                            nterm - mean, cterm - mean], -1)
+        return mean + self.pool_gain[task] * self.pool_projection(detail)
+
+    def represent(self, task, batch, device, max_len=1024):
+        if task in self.PAIR:
+            a = self.pool_sequence(task, pad_batch(
+                [encode(x["sequence_a"], max_len) for x in batch], device))
+            b = self.pool_sequence(task, pad_batch(
+                [encode(x["sequence_b"], max_len) for x in batch], device))
+            base = torch.cat([a, b], dim=-1)
+            if not self.structured_pooling:
+                return base
+            interaction = torch.cat([a * b, torch.abs(a - b)], dim=-1)
+            return base + self.pair_projection(interaction)
+        return self.pool_sequence(task, pad_batch(
+            [encode(x["sequence"], max_len) for x in batch], device))
+
+
+def build(name="esm2_t12_35M", structured_pooling=True):
+    layers, dim, heads = CONFIGS[name]
+    model = Model(layers=layers, dim=dim, heads=heads,
+                  structured_pooling=structured_pooling)
+    return model, {"name": name, "layers": layers, "dim": dim, "heads": heads}
+
+
+if __name__ == "__main__":
+    for name in CONFIGS:
+        model, config = build(name)
+        backbone = sum(p.numel() for p in model.encoder.parameters())
+        total = sum(p.numel() for p in model.parameters())
+        print(f"{name:16s} backbone {backbone/1e6:7.2f}M   with heads {total/1e6:7.2f}M")
+
+
+class ResidualLinear(nn.Module):
+    """Small fit-trained nonlinear correction to one scalar Linear readout."""
+    def __init__(self, base, width):
+        super().__init__()
+        self.base = base
+        self.residual = nn.Sequential(nn.Linear(base.in_features, width), nn.GELU(),
+                                      nn.Dropout(0.1), nn.Linear(width, 1))
+        self.register_buffer("feature_center", torch.zeros(base.in_features))
+        self.register_buffer("feature_scale", torch.ones(base.in_features))
+        self.register_buffer("output_scale", torch.tensor(1.0))
+
+    def forward(self, hidden):
+        normalized = (hidden - self.feature_center) / self.feature_scale
+        return self.base(hidden) + self.output_scale * self.residual(normalized)
+
+
+def install_residual_readouts(model, specification):
+    for task, width in specification.items():
+        base = model.heads[task][-1]
+        model.heads[task][-1] = ResidualLinear(base, width).to(base.weight.device)
+
+
+def install_pooled_classifiers(model, specs):
+    for task, input_dim in specs.items():
+        model.heads[task] = torch.nn.Linear(input_dim, model.CLASSES[task]).to(next(model.parameters()).device)

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S005/predict.py

@@ -0,0 +1,125 @@
+#!/usr/bin/env python3
+"""Self-contained offline inference, also used for in-training selection."""
+import json
+import sys
+from pathlib import Path
+
+import torch
+from esm2 import Model, encode, pad_batch
+
+TASKS = ('fluorescence', 'stability', 'beta_lactamase', 'solubility',
+         'subcellular_localization', 'binary_localization', 'fold',
+         'secondary_structure', 'yeast_ppi', 'human_ppi', 'ppi_affinity')
+
+
+def batches(rows, budget=16384, pair=False):
+    def length(i):
+        return min(1024, max(len(s) for s in rows[i]['features'].values()) + 2)
+    order = sorted(range(len(rows)), key=length)
+    result, width = [], 0
+    for i in order:
+        candidate = max(width, length(i))
+        if result and candidate * (len(result) + 1) * (2 if pair else 1) > budget:
+            yield result
+            result, width = [], 0
+        result.append(i)
+        width = max(width, candidate)
+    if result:
+        yield result
+
+
+@torch.inference_mode()
+def predict_records(model, rows, task, device, full_chain=True, budget=16384):
+    model.eval()
+    found = {}
+    use_amp = device.type == 'cuda' and task not in getattr(model, 'full_precision_tasks', ())
+    if task == 'secondary_structure':
+        windows = []
+        buffers = [torch.zeros(len(r['features']['sequence']), 3) for r in rows]
+        counts = [torch.zeros(len(r['features']['sequence']), 1) for r in rows]
+        for i, row in enumerate(rows):
+            sequence = row['features']['sequence']
+            last = max(0, len(sequence) - 1022)
+            starts = sorted(set(list(range(0, last, 894)) + [last])) if full_chain else [0]
+            for start in starts:
+                windows.append({'features': {'sequence': sequence[start:start+1022]},
+                                'row': i, 'start': start})
+        for indices in batches(windows, budget):
+            selected = [windows[i] for i in indices]
+            tokens = pad_batch([encode(w['features']['sequence']) for w in selected], device)
+            with torch.autocast(device.type, dtype=torch.bfloat16, enabled=use_amp):
+                hidden, _ = model.encoder(tokens)
+                probabilities = model.heads[task](hidden).float().softmax(-1).cpu()
+            for j, window in enumerate(selected):
+                i, start = window['row'], window['start']
+                length = len(window['features']['sequence'])
+                weight = torch.arange(1, length + 1).minimum(torch.arange(length, 0, -1)).clamp(max=64).float()[:, None]
+                buffers[i][start:start+length] += probabilities[j, 1:1+length] * weight
+                counts[i][start:start+length] += weight
+        for i, row in enumerate(rows):
+            score = buffers[i] / counts[i].clamp(min=1)
+            score[counts[i][:, 0] == 0] = 1 / 3
+            found[row['example_id']] = {'ss3_scores': score.tolist()}
+        return found
+    for indices in batches(rows, budget, task in Model.PAIR):
+        chunk = [rows[i] for i in indices]
+        with torch.autocast(device.type, dtype=torch.bfloat16, enabled=use_amp):
+            representation = model.represent(task, [r['features'] for r in chunk], device, 1024)
+            if task in getattr(model, 'full_precision_head_tasks', ()):
+                with torch.autocast(device.type, enabled=False):
+                    output = model.heads[task](representation.float())
+            elif task in getattr(model, 'fp32_head_tasks', ()):
+                hidden = model.heads[task][:-1](representation)
+                with torch.autocast(device.type, enabled=False):
+                    output = model.heads[task][-1](hidden.float())
+            else:
+                output = model.heads[task](representation).float()
+        if task in Model.CLASSES:
+            predictions = [{'scores': v} for v in output.softmax(-1).clamp(min=1e-9).cpu().tolist()]
+        else:
+            predictions = [{'value': v[0]} for v in output.cpu().tolist()]
+        found.update((row['example_id'], prediction) for row, prediction in zip(chunk, predictions))
+    return found
+
+
+def main():
+    if len(sys.argv) != 3:
+        raise SystemExit('usage: predict.py INPUT_ROOT OUTPUT_ROOT')
+    torch.set_num_threads(4)
+    torch.backends.cuda.matmul.allow_tf32 = False
+    here = Path(__file__).resolve().parent
+    source, out = Path(sys.argv[1]), Path(sys.argv[2]) / 'predictions'
+    out.mkdir(parents=True, exist_ok=False)
+    config = json.loads((here / 'config.json').read_text())
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    model_args = {k: config[k] for k in ['layers', 'dim', 'heads']}
+    if 'structured_pooling' in config:
+        model_args['structured_pooling'] = config['structured_pooling']
+    model = Model(**model_args).to(device)
+    if config.get('residual_readouts'):
+        from esm2 import install_residual_readouts
+        install_residual_readouts(model, config['residual_readouts'])
+    if config.get('pooled_linear_classifiers'):
+        from esm2 import install_pooled_classifiers
+        install_pooled_classifiers(model, config['pooled_linear_classifiers'])
+    if config.get('mutation_readouts'):
+        from mutation_readout_runtime import install_mutation_readouts
+        install_mutation_readouts(model, config['mutation_readouts'])
+    model.load_state_dict(torch.load(here / 'model.pt', map_location=device, weights_only=True))
+    model.fp32_head_tasks = set(config.get('fp32_head_tasks', []))
+    model.full_precision_tasks = set(config.get('full_precision_tasks', []))
+    model.full_precision_head_tasks = set(config.get('full_precision_head_tasks', []))
+    model.eval()
+    for task in TASKS:
+        rows = [json.loads(line) for line in (source / f'{task}.jsonl').open()]
+        found = predict_records(model, rows, task, device,
+                                full_chain=config.get('full_chain_ss', True))
+        with (out / f'{task}.jsonl').open('w') as f:
+            for row in rows:
+                f.write(json.dumps({'example_id': row['example_id'],
+                                    'prediction': found[row['example_id']]}, separators=(',', ':')) + '\n')
+        print(f'{task}: {len(rows)} predictions', flush=True)
+
+
+if __name__ == '__main__':
+    main()

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S006/README.md

@@ -0,0 +1,13 @@
+# Unified protein sequence model
+
+One shared 12-layer,384-channel,six-head rotary encoder serves all eleven PEER tasks. Exact counts: **21,306,271 encoder parameters;25,065,060 total parameters**, plus769 normalization buffer values. Every learned value originates from random initialization and training in this session using mounted data only.
+
+The encoder was pretrained for four hours on permitted UniRef50 (9,127,835,498 input tokens), then jointly trained on all eleven fit tasks for1800active seconds, seed1. Tune selected the final online state at update55298. All eleven tasks updated the encoder for300seconds; human PPI, solubility and secondary structure subsequently updated their heads while the other eight continued shared learning. Inference loads one complete state. No checkpoint averaging is used.
+
+Global representations combine mean, task-query, maximum and terminal summaries. Pair tasks add shared product/difference processing. Beta uses a fit-label rank ridge (relative strength0.0001) plus64-unit residual, selected at step14000of24000 under fixed magnitude, paired-interval and replay-agreement gates. Fluorescence, stability, affinity and all classification tasks retain their jointly trained MLP heads; secondary structure uses a residue MLP. Beta outputs rank units; the other three scalar tasks use original label units. All scalar heads useFP32; beta's full path usesFP32.
+
+Tune selects complete states and fixed fit-only readout recipes. Independent eleven-task mean is **0.724671**, versus the previous package0.723342: six task scores improve andfive decline. GFP andaffinity improve with positive paired intervals; human PPI's component interval reacheszero. Secondary structure andsolubility decline with negative intervals. The aggregate interval spanszero. Rare-fold accuracy remains1/92; yeast drops7/95correct predictions. Beta lowest-quartile ordering declines versus its raw-readout parent, despite its overall gain. Repeated selection, homology, one seed andPPI dependence limit generalization claims. Full results andconditional beta diagnostics are in training_history.json.
+
+Global/pair chains truncate at1022residues. Secondary structure covers full chains with overlapping1022-residue windows,stride894,taper64. Beta batches use8192tokens; others16384. The beta-only readout change preserves ten parent prediction files exactly andmatches its cached fitting reward exactly. The first49-fixture cold replay passed, including alteredIDs, empty/nonstandard chains and1200-residue secondary structure.
+
+Run offline: `bash infer.sh INPUT_ROOT OUTPUT_ROOT`. The package contains inference code, one parameter set, configuration and provenance. It uses no external weights, embeddings, lookup table, feature cache or inference-time fitting.

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S007/predict.py

@@ -0,0 +1,122 @@
+#!/usr/bin/env python3
+"""Self-contained offline inference, also used for in-training selection."""
+import json
+import sys
+from pathlib import Path
+
+import torch
+from esm2 import Model, encode, pad_batch
+
+TASKS = ('fluorescence', 'stability', 'beta_lactamase', 'solubility',
+         'subcellular_localization', 'binary_localization', 'fold',
+         'secondary_structure', 'yeast_ppi', 'human_ppi', 'ppi_affinity')
+
+
+def batches(rows, budget=16384, pair=False):
+    def length(i):
+        return min(1024, max(len(s) for s in rows[i]['features'].values()) + 2)
+    order = sorted(range(len(rows)), key=length)
+    result, width = [], 0
+    for i in order:
+        candidate = max(width, length(i))
+        if result and candidate * (len(result) + 1) * (2 if pair else 1) > budget:
+            yield result
+            result, width = [], 0
+        result.append(i)
+        width = max(width, candidate)
+    if result:
+        yield result
+
+
+@torch.inference_mode()
+def predict_records(model, rows, task, device, full_chain=True, budget=16384):
+    model.eval()
+    found = {}
+    use_amp = device.type == 'cuda' and task not in getattr(model, 'full_precision_tasks', ())
+    if task == 'secondary_structure':
+        windows = []
+        buffers = [torch.zeros(len(r['features']['sequence']), 3) for r in rows]
+        counts = [torch.zeros(len(r['features']['sequence']), 1) for r in rows]
+        for i, row in enumerate(rows):
+            sequence = row['features']['sequence']
+            last = max(0, len(sequence) - 1022)
+            starts = sorted(set(list(range(0, last, 894)) + [last])) if full_chain else [0]
+            for start in starts:
+                windows.append({'features': {'sequence': sequence[start:start+1022]},
+                                'row': i, 'start': start})
+        for indices in batches(windows, budget):
+            selected = [windows[i] for i in indices]
+            tokens = pad_batch([encode(w['features']['sequence']) for w in selected], device)
+            with torch.autocast(device.type, dtype=torch.bfloat16, enabled=use_amp):
+                hidden, _ = model.encoder(tokens)
+                probabilities = model.heads[task](hidden).float().softmax(-1).cpu()
+            for j, window in enumerate(selected):
+                i, start = window['row'], window['start']
+                length = len(window['features']['sequence'])
+                weight = torch.arange(1, length + 1).minimum(torch.arange(length, 0, -1)).clamp(max=64).float()[:, None]
+                buffers[i][start:start+length] += probabilities[j, 1:1+length] * weight
+                counts[i][start:start+length] += weight
+        for i, row in enumerate(rows):
+            score = buffers[i] / counts[i].clamp(min=1)
+            score[counts[i][:, 0] == 0] = 1 / 3
+            found[row['example_id']] = {'ss3_scores': score.tolist()}
+        return found
+    for indices in batches(rows, budget, task in Model.PAIR):
+        chunk = [rows[i] for i in indices]
+        with torch.autocast(device.type, dtype=torch.bfloat16, enabled=use_amp):
+            representation = model.represent(task, [r['features'] for r in chunk], device, 1024)
+            if task in getattr(model, 'full_precision_head_tasks', ()):
+                with torch.autocast(device.type, enabled=False):
+                    output = model.heads[task](representation.float())
+            elif task in getattr(model, 'fp32_head_tasks', ()):
+                hidden = model.heads[task][:-1](representation)
+                with torch.autocast(device.type, enabled=False):
+                    output = model.heads[task][-1](hidden.float())
+            else:
+                output = model.heads[task](representation).float()
+        if task in Model.CLASSES:
+            predictions = [{'scores': v} for v in output.softmax(-1).clamp(min=1e-9).cpu().tolist()]
+        else:
+            predictions = [{'value': v[0]} for v in output.cpu().tolist()]
+        found.update((row['example_id'], prediction) for row, prediction in zip(chunk, predictions))
+    return found
+
+
+def main():
+    if len(sys.argv) != 3:
+        raise SystemExit('usage: predict.py INPUT_ROOT OUTPUT_ROOT')
+    torch.set_num_threads(4)
+    torch.backends.cuda.matmul.allow_tf32 = False
+    here = Path(__file__).resolve().parent
+    source, out = Path(sys.argv[1]), Path(sys.argv[2]) / 'predictions'
+    out.mkdir(parents=True, exist_ok=False)
+    config = json.loads((here / 'config.json').read_text())
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    model_args = {k: config[k] for k in ['layers', 'dim', 'heads']}
+    if 'structured_pooling' in config:
+        model_args['structured_pooling'] = config['structured_pooling']
+    model = Model(**model_args).to(device)
+    if config.get('residual_readouts'):
+        from esm2 import install_residual_readouts
+        install_residual_readouts(model, config['residual_readouts'])
+    if config.get('mutation_readouts'):
+        from mutation_readout_runtime import install_mutation_readouts
+        install_mutation_readouts(model, config['mutation_readouts'])
+    model.load_state_dict(torch.load(here / 'model.pt', map_location=device, weights_only=True))
+    model.fp32_head_tasks = set(config.get('fp32_head_tasks', []))
+    model.full_precision_tasks = set(config.get('full_precision_tasks', []))
+    model.full_precision_head_tasks = set(config.get('full_precision_head_tasks', []))
+    model.eval()
+    for task in TASKS:
+        rows = [json.loads(line) for line in (source / f'{task}.jsonl').open()]
+        found = predict_records(model, rows, task, device,
+                                full_chain=config.get('full_chain_ss', True), budget=config.get('inference_token_budgets', {}).get(task, 16384))
+        with (out / f'{task}.jsonl').open('w') as f:
+            for row in rows:
+                f.write(json.dumps({'example_id': row['example_id'],
+                                    'prediction': found[row['example_id']]}, separators=(',', ':')) + '\n')
+        print(f'{task}: {len(rows)} predictions', flush=True)
+
+
+if __name__ == '__main__':
+    main()

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S009/PUBLISHED.json

@@ -0,0 +1,5 @@
+{
+  "source": "/run-state/workspace/research/21m_4h_two_task_fixed_readouts/candidate/package",
+  "files": 8,
+  "bytes": 100501085
+}

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S010/mutation_readout_runtime.py

@@ -0,0 +1,71 @@
+"""Inference-only contextual substitution residuals over one shared encoder."""
+from types import MethodType
+import torch
+from torch import nn
+import torch.nn.functional as F
+from esm2 import encode
+
+
+class MutationResidual(nn.Module):
+    def __init__(self, original_head, dim):
+        super().__init__()
+        self.original_head = original_head
+        self.dim = dim
+        self.correction = nn.Linear(dim*2+1, 1)
+        nn.init.zeros_(self.correction.weight)
+        nn.init.zeros_(self.correction.bias)
+
+    def forward(self, representation):
+        d = self.dim
+        original = representation[:, :d]
+        delta = representation[:, d:2*d]
+        count = representation[:, 2*d:2*d+1]
+        eligible = representation[:, 2*d+1:2*d+2].bool()
+        features = self.original_head[:-1](original)
+        base = self.original_head[-1](features)
+        joined = torch.cat([features, count, delta], dim=-1)
+        correction = F.linear(joined.double(), self.correction.weight.double(),
+                              self.correction.bias.double()).to(base.dtype)
+        return torch.where(eligible, base+correction, base)
+
+
+def _represent(self, task, batch, device, max_len=1024):
+    if task not in self._mutation_references:
+        return self._mutation_original_represent(task, batch, device, max_len)
+    reference_ids = self._mutation_references[task]
+    reference_length = len(reference_ids)-2
+    lengths = [len(row['sequence']) for row in batch]
+    captured = []
+    hook = self.encoder.register_forward_hook(lambda module, inputs, output: captured.append((inputs[0], output[0])))
+    try:
+        original = self._mutation_original_represent(task, batch, device, max_len)
+    finally:
+        hook.remove()
+    assert len(captured) == 1
+    tokens, hidden = captured[0]
+    eligible = torch.tensor(lengths, device=tokens.device).eq(reference_length)
+    delta = hidden.new_zeros((len(batch), hidden.shape[-1]))
+    count = hidden.new_zeros((len(batch), 1))
+    ids = eligible.nonzero(as_tuple=True)[0]
+    if len(ids):
+        reference_tokens = torch.tensor(reference_ids, device=tokens.device, dtype=torch.long)
+        reference_hidden = self.encoder(reference_tokens[None])[0][0]
+        different = tokens[ids, 1:reference_length+1].ne(reference_tokens[1:-1])
+        changes = hidden[ids, 1:reference_length+1]-reference_hidden[None, 1:-1]
+        delta[ids] = (changes*different[:, :, None]).sum(1)
+        count[ids, 0] = different.sum(1).to(hidden.dtype)
+    return torch.cat([original, delta, count.log1p(), eligible[:, None].to(hidden.dtype)], dim=-1)
+
+
+def install_mutation_readouts(model, specification):
+    assert not hasattr(model, '_mutation_references')
+    assert set(specification) <= {'fluorescence', 'beta_lactamase'}
+    model._mutation_references = {}
+    for task, info in specification.items():
+        reference = info['reference']
+        assert 0 < len(reference) <= 1022
+        model._mutation_references[task] = encode(reference)
+        device = next(model.heads[task].parameters()).device
+        model.heads[task] = MutationResidual(model.heads[task], model.encoder.dim).to(device)
+    model._mutation_original_represent = model.represent
+    model.represent = MethodType(_represent, model)

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S010/predict.py

@@ -0,0 +1,121 @@
+#!/usr/bin/env python3
+"""Self-contained offline inference, also used for in-training selection."""
+import json
+import sys
+from pathlib import Path
+
+import torch
+from esm2 import Model, encode, pad_batch
+
+TASKS = ('fluorescence', 'stability', 'beta_lactamase', 'solubility',
+         'subcellular_localization', 'binary_localization', 'fold',
+         'secondary_structure', 'yeast_ppi', 'human_ppi', 'ppi_affinity')
+
+
+def batches(rows, budget=16384, pair=False):
+    def length(i):
+        return min(1024, max(len(s) for s in rows[i]['features'].values()) + 2)
+    order = sorted(range(len(rows)), key=length)
+    result, width = [], 0
+    for i in order:
+        candidate = max(width, length(i))
+        if result and candidate * (len(result) + 1) * (2 if pair else 1) > budget:
+            yield result
+            result, width = [], 0
+        result.append(i)
+        width = max(width, candidate)
+    if result:
+        yield result
+
+
+@torch.inference_mode()
+def predict_records(model, rows, task, device, full_chain=True, budget=16384):
+    model.eval()
+    found = {}
+    use_amp = device.type == 'cuda' and task not in getattr(model, 'full_precision_tasks', ())
+    if task == 'secondary_structure':
+        windows = []
+        buffers = [torch.zeros(len(r['features']['sequence']), 3) for r in rows]
+        counts = [torch.zeros(len(r['features']['sequence']), 1) for r in rows]
+        for i, row in enumerate(rows):
+            sequence = row['features']['sequence']
+            last = max(0, len(sequence) - 1022)
+            starts = sorted(set(list(range(0, last, 894)) + [last])) if full_chain else [0]
+            for start in starts:
+                windows.append({'features': {'sequence': sequence[start:start+1022]},
+                                'row': i, 'start': start})
+        for indices in batches(windows, budget):
+            selected = [windows[i] for i in indices]
+            tokens = pad_batch([encode(w['features']['sequence']) for w in selected], device)
+            with torch.autocast(device.type, dtype=torch.bfloat16, enabled=use_amp):
+                hidden, _ = model.encoder(tokens)
+                probabilities = model.heads[task](hidden).float().softmax(-1).cpu()
+            for j, window in enumerate(selected):
+                i, start = window['row'], window['start']
+                length = len(window['features']['sequence'])
+                weight = torch.arange(1, length + 1).minimum(torch.arange(length, 0, -1)).clamp(max=64).float()[:, None]
+                buffers[i][start:start+length] += probabilities[j, 1:1+length] * weight
+                counts[i][start:start+length] += weight
+        for i, row in enumerate(rows):
+            score = buffers[i] / counts[i].clamp(min=1)
+            score[counts[i][:, 0] == 0] = 1 / 3
+            found[row['example_id']] = {'ss3_scores': score.tolist()}
+        return found
+    for indices in batches(rows, budget, task in Model.PAIR):
+        chunk = [rows[i] for i in indices]
+        with torch.autocast(device.type, dtype=torch.bfloat16, enabled=use_amp):
+            representation = model.represent(task, [r['features'] for r in chunk], device, 1024)
+            if task in getattr(model, 'full_precision_head_tasks', ()):
+                with torch.autocast(device.type, enabled=False):
+                    output = model.heads[task](representation.float())
+            elif task in getattr(model, 'fp32_head_tasks', ()):
+                hidden = model.heads[task][:-1](representation)
+                with torch.autocast(device.type, enabled=False):
+                    output = model.heads[task][-1](hidden.float())
+            else:
+                output = model.heads[task](representation).float()
+        if task in Model.CLASSES:
+            predictions = [{'scores': v} for v in output.softmax(-1).clamp(min=1e-9).cpu().tolist()]
+        else:
+            predictions = [{'value': v[0]} for v in output.cpu().tolist()]
+        found.update((row['example_id'], prediction) for row, prediction in zip(chunk, predictions))
+    return found
+
+
+def main():
+    if len(sys.argv) != 3:
+        raise SystemExit('usage: predict.py INPUT_ROOT OUTPUT_ROOT')
+    torch.set_num_threads(4)
+    torch.backends.cuda.matmul.allow_tf32 = False
+    here = Path(__file__).resolve().parent
+    source, out = Path(sys.argv[1]), Path(sys.argv[2]) / 'predictions'
+    out.mkdir(parents=True, exist_ok=False)
+    config = json.loads((here / 'config.json').read_text())
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    model_args = {k: config[k] for k in ['layers', 'dim', 'heads']}
+    if 'structured_pooling' in config:
+        model_args['structured_pooling'] = config['structured_pooling']
+    model = Model(**model_args).to(device)
+    from esm2 import install_residual_readouts
+    install_residual_readouts(model, config['residual_readouts'])
+    from mutation_readout_runtime import install_mutation_readouts
+    install_mutation_readouts(model, config['mutation_readouts'])
+    model.load_state_dict(torch.load(here / 'model.pt', map_location=device, weights_only=True))
+    model.fp32_head_tasks = set(config.get('fp32_head_tasks', []))
+    model.full_precision_tasks = set(config.get('full_precision_tasks', []))
+    model.full_precision_head_tasks = set(config.get('full_precision_head_tasks', []))
+    model.eval()
+    for task in TASKS:
+        rows = [json.loads(line) for line in (source / f'{task}.jsonl').open()]
+        found = predict_records(model, rows, task, device,
+                                full_chain=config.get('full_chain_ss', True),
+                                budget=config.get('inference_token_budgets', {}).get(task, 16384))
+        with (out / f'{task}.jsonl').open('w') as f:
+            for row in rows:
+                f.write(json.dumps({'example_id': row['example_id'],
+                                    'prediction': found[row['example_id']]}, separators=(',', ':')) + '\n')
+        print(f'{task}: {len(rows)} predictions', flush=True)
+
+
+if __name__ == '__main__':
+    main()

tasks/protein-sequence-modeling/protein-sequence-modeling-astra-xhigh-1h100-48h-20261031a/submissions/S010/training_history.json

@@ -0,0 +1,2068 @@
+{
+  "config": {
+    "name": "unified_t12_21M",
+    "layers": 12,
+    "dim": 384,
+    "heads": 6,
+    "structured_pooling": true,
+    "full_chain_ss": true,
+    "full_precision_head_tasks": [
+      "fluorescence",
+      "stability",
+      "beta_lactamase",
+      "ppi_affinity"
+    ],
+    "weight_variant": "ema",
+    "full_precision_tasks": [
+      "beta_lactamase",
+      "fluorescence"
+    ],
+    "inference_token_budgets": {
+      "beta_lactamase": 8192
+    },
+    "residual_readouts": {
+      "beta_lactamase": 64
+    },
+    "rank_residual_tasks": [
+      "beta_lactamase"
+    ],
+    "mutation_readouts": {
+      "fluorescence": {
+        "reference": "SKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLKFICTTGKLPVPWPTLVTTLSYGVQCFSRYPDHMKQHDFFKSAMPEGYVQERTIFFKDDGNYKTRAEVKFEGDTLVNRIELKGIDFKEDGNILGHKLEYNYNSHNVYIMADKQKNGIKVNFKIRHNIEDGSVQLADHYQQNTPIGDGPVLLPDNHYLSTQSALSKDPNEKRDHMVLLEFVTAAGITHGMDELYK",
+        "feature_policy": "Fit-only consensus and shared contextual substitution delta; other lengths unchanged."
+      }
+    }
+  },
+  "notes": {
+    "ancestor": "/workspace/research/pretrain21m_extended/pretrain.pt",
+    "seed": 1,
+    "data": "fit training; tune selection only",
+    "settings": {
+      "out": "/workspace/research/21m_4h_joint_fold_mask",
+      "resume": "/workspace/research/pretrain21m_extended/pretrain.pt",
+      "hours": 0.5,
+      "lr": 0.0001,
+      "batch_size": 32,
+      "seed": 1,
+      "eval_seconds": 300.0,
+      "mean_pooling": false,
+      "normalized_regression": true,
+      "random_batches": true,
+      "task_weights": "{\"human_ppi\":0.25}",
+      "continue_state": false,
+      "model": "unified_t12_21M",
+      "head_lr_mult": 2.0,
+      "fold_contrastive": 0.0,
+      "ema_decay": 0.999,
+      "correlation_loss": 0.1,
+      "stop_shared_tasks": "[\"human_ppi\", \"solubility\"]",
+      "stop_shared_after_seconds": 300.0,
+      "tune_limit": 0
+    },
+    "ancestor_provenance": {
+      "step": 211159,
+      "tokens_seen": 6876918950,
+      "total_pretrain_tokens": 9127835498,
+      "provenance": {
+        "random_initialization_seed": 42,
+        "stage_seed": 43,
+        "data": "/workspace/research/corpus/manifest.json",
+        "objective": "15% masked residues; 80% mask / 10% random / 10% unchanged",
+        "sampling": "Length buckets sampled by token mass; shuffled sequence cycles without replacement within each bucket",
+        "ancestry": {
+          "checkpoint": "/workspace/research/pretrain21m/pretrain.pt",
+          "prior_tokens": 2250916548,
+          "prior_provenance": {
+            "random_initialization_seed": 42,
+            "data": "/workspace/research/corpus/manifest.json",
+            "objective": "15% masked residues; 80% mask / 10% random / 10% unchanged",
+            "sampling": "Length buckets sampled by token mass; shuffled sequence cycles without replacement within each bucket",
+            "model": {
+              "name": "unified_t12_21M",
+              "layers": 12,
+              "dim": 384,
+              "heads": 6
+            },
+            "parameters": {
+              "backbone": 21306271,
+              "total": 25040355
+            }
+          }
+        },
+        "continued_ancestor_corpus": true,
+        "stage_active_hour_budget": 3.0,
+        "stage_peak_learning_rate": 0.0002,
+        "fit_text_probability": 0.0,
+        "fit_mixture_sampling": "Independent resumable RNG; mixture decisions do not consume the UniRef corpus sampler RNG",
+        "fit_text": null,
+        "model": {
+          "name": "unified_t12_21M",
+          "layers": 12,
+          "dim": 384,
+          "heads": 6
+        },
+        "parameters": {
+          "backbone": 21306271,
+          "total": 25040355
+        }
+      },
+      "complete": true,
+      "source_counts": {
+        "uniref": {
+          "updates": 211159,
+          "input_tokens": 6876918950
+        },
+        "fit_text": {
+          "updates": 0,
+          "input_tokens": 0
+        }
+      }
+    },
+    "shared_gradient_gate": {
+      "tasks": [
+        "human_ppi",
+        "solubility"
+      ],
+      "after_seconds": 300.0,
+      "updates": {
+        "fluorescence": 0,
+        "stability": 0,
+        "beta_lactamase": 0,
+        "solubility": 9194,
+        "subcellular_localization": 0,
+        "binary_localization": 0,
+        "fold": 0,
+        "secondary_structure": 0,
+        "yeast_ppi": 0,
+        "human_ppi": 1795,
+        "ppi_affinity": 0
+      },
+      "scope": "All representation parameters receive no gradient from these tasks after the cutoff. Their MLP heads continue learning; other tasks continue updating the same shared representation. No added parameters."
+    },
+    "fold_mask_augmentation": {
+      "seed": 1611,
+      "row_probability": 0.5,
+      "residue_probability": 0.1,
+      "updates": 4820,
+      "rows": 154240,
+      "eligible_residues": 25746939,
+      "selected_rows": 76975,
+      "masked_rows": 76872,
+      "masked_residues": 1283235,
+      "scope": "Fit fold inputs only. No label changes, auxiliary MLM loss or inference masking. Special/padding/UNK tokens preserved; masked positions retain their original pooling positions. No per-row lookup assets exported."
+    },
+    "compute_by_task": {
+      "fluorescence": {
+        "updates": 6472,
+        "examples": 207104,
+        "tokens": 49492867,
+        "seconds": 187.698781524261
+      },
+      "stability": {
+        "updates": 10023,
+        "examples": 320736,
+        "tokens": 15153763,
+        "seconds": 276.75090847232786
+      },
+      "beta_lactamase": {
+        "updates": 2744,
+        "examples": 87808,
+        "tokens": 25288704,
+        "seconds": 79.99926061385486
+      },
+      "solubility": {
+        "updates": 10816,
+        "examples": 346112,
+        "tokens": 103011696,
+        "seconds": 167.9670603771956
+      },
+      "subcellular_localization": {
+        "updates": 3924,
+        "examples": 125568,
+        "tokens": 59655796,
+        "seconds": 172.02873353951145
+      },
+      "binary_localization": {
+        "updates": 3143,
+        "examples": 100576,
+        "tokens": 45368420,
+        "seconds": 137.227227671523
+      },
+      "fold": {
+        "updates": 4820,
+        "examples": 154240,
+        "tokens": 26055419,
+        "seconds": 141.78438254390494
+      },
+      "secondary_structure": {
+        "updates": 4097,
+        "examples": 131104,
+        "tokens": 33763745,
+        "seconds": 133.5113292203023
+      },
+      "yeast_ppi": {
+        "updates": 3083,
+        "examples": 98656,
+        "tokens": 100729762,
+        "seconds": 309.4622458037484
+      },
+      "human_ppi": {
+        "updates": 2051,
+        "examples": 65632,
+        "tokens": 66106750,
+        "seconds": 84.95186259552429
+      },
+      "ppi_affinity": {
+        "updates": 2121,
+        "examples": 67872,
+        "tokens": 24460152,
+        "seconds": 108.62219824487693
+      }
+    },
+    "continuations": [
+      {
+        "checkpoint": "/workspace/research/21m_4h_joint_fold_mask/ck/finetune.pt",
+        "step": 36201,
+        "active_seconds": 1200.0549627174623,
+        "restored": "model, optimizer, EMA, Python/Torch/CUDA RNG, task exposure, full selection history and best reward"
+      }
+    ]
+  },
+  "history": [
+    {
+      "step": 7672,
+      "active_seconds": 300.00514184945496,
+      "variant": "online",
+      "reward": 0.630689327475037,
+      "tasks": {
+        "fluorescence": {
+          "spearman": 0.662602076321563,
+          "reward": 0.662602076321563
+        },
+        "stability": {
+          "spearman": 0.6842833989995151,
+          "reward": 0.6842833989995151
+        },
+        "beta_lactamase": {
+          "spearman": 0.48047899888349094,
+          "reward": 0.48047899888349094
+        },
+        "solubility": {
+          "accuracy": 0.7496398732353788,
+          "balanced_accuracy": 0.7383881625124324,
+          "macro_f1": 0.7402483293276466,
+          "reward": 0.7496398732353788
+        },
+        "subcellular_localization": {
+          "accuracy": 0.6531483457844184,
+          "balanced_accuracy": 0.5511694467213236,
+          "macro_f1": 0.5544142762748534,
+          "reward": 0.6531483457844184
+        },
+        "binary_localization": {
+          "accuracy": 0.924812030075188,
+          "balanced_accuracy": 0.911793372319688,
+          "macro_f1": 0.9202155603262228,
+          "reward": 0.924812030075188
+        },
+        "fold": {
+          "accuracy": 0.20923913043478262,
+          "balanced_accuracy": 0.09853702961688442,
+          "macro_f1": 0.019837649237673546,
+          "reward": 0.20923913043478262
+        },
+        "secondary_structure": {
+          "accuracy": 0.7474283169728386,
+          "valid_residues": 556931,
+          "reward": 0.7474283169728386
+        },
+        "yeast_ppi": {
+          "accuracy": 0.5789473684210527,
+          "balanced_accuracy": 0.5377747252747254,
+          "macro_f1": 0.5287698412698413,
+          "reward": 0.5789473684210527
+        },
+        "human_ppi": {
+          "accuracy": 0.8031746031746032,
+          "balanced_accuracy": 0.7988945578231292,
+          "macro_f1": 0.8004210268149117,
+          "reward": 0.8031746031746032
+        },
+        "ppi_affinity": {
+          "rmse": 2.504541197254769,
+          "spearman": 0.6340510687158502,
+          "reward": 0.4438284599225761
+        }
+      }
+    },
+    {
+      "step": 7672,
+      "active_seconds": 300.00514184945496,
+      "variant": "ema",
+      "reward": 0.6521213572878972,
+      "tasks": {
+        "fluorescence": {
+          "spearman": 0.6509414450729477,
+          "reward": 0.6509414450729477
+        },
+        "stability": {
+          "spearman": 0.7019963547461268,
+          "reward": 0.7019963547461268
+        },
+        "beta_lactamase": {
+          "spearman": 0.5268092997333048,
+          "reward": 0.5268092997333048
+        },
+        "solubility": {
+          "accuracy": 0.7565543071161048,
+          "balanced_accuracy": 0.7407946415732911,
+          "macro_f1": 0.7444303719707693,
+          "reward": 0.7565543071161048
+        },
+        "subcellular_localization": {
+          "accuracy": 0.7488438278192814,
+          "balanced_accuracy": 0.5844029444923964,
+          "macro_f1": 0.5898563439824718,
+          "reward": 0.7488438278192814
+        },
+        "binary_localization": {
+          "accuracy": 0.9329091960670908,
+          "balanced_accuracy": 0.9237658711342922,
+          "macro_f1": 0.9294472398577154,
+          "reward": 0.9329091960670908
+        },
+        "fold": {
+          "accuracy": 0.22418478260869565,
+          "balanced_accuracy": 0.10198668038958422,
+          "macro_f1": 0.0203231348785858,
+          "reward": 0.22418478260869565
+        },
+        "secondary_structure": {
+          "accuracy": 0.7556842768673319,
+          "valid_residues": 556931,
+          "reward": 0.7556842768673319
+        },
+        "yeast_ppi": {
+          "accuracy": 0.5894736842105263,
+          "balanced_accuracy": 0.5583791208791209,
+          "macro_f1": 0.5562342795544376,
+          "reward": 0.5894736842105263
+        },
+        "human_ppi": {
+          "accuracy": 0.8317460317460318,
+          "balanced_accuracy": 0.8307823129251701,
+          "macro_f1": 0.8309212991563788,
+          "reward": 0.8317460317460318
+        },
+        "ppi_affinity": {
+          "rmse": 2.433376875995891,
+          "spearman": 0.6286745086813996,
+          "reward": 0.4541917241794273
+        }
+      }
+    },
+    {
+      "step": 17079,
+      "active_seconds": 600.007775742939,
+      "variant": "online",
+      "reward": 0.670020300759891,
+      "tasks": {
+        "fluorescence": {
+          "spearman": 0.707789297048708,
+          "reward": 0.707789297048708
+        },
+        "stability": {
+          "spearman": 0.7545616535010345,
+          "reward": 0.7545616535010345
+        },
+        "beta_lactamase": {
+          "spearman": 0.5180760726197687,
+          "reward": 0.5180760726197687
+        },
+        "solubility": {
+          "accuracy": 0.7483434168827428,
+          "balanced_accuracy": 0.7409493986575773,
+          "macro_f1": 0.741101993719804,
+          "reward": 0.7483434168827428
+        },
+        "subcellular_localization": {
+          "accuracy": 0.7623621487015297,
+          "balanced_accuracy": 0.6629158106231113,
+          "macro_f1": 0.6701746279318399,
+          "reward": 0.7623621487015297
+        },
+        "binary_localization": {
+          "accuracy": 0.951417004048583,
+          "balanced_accuracy": 0.9478689215531321,
+          "macro_f1": 0.9494746949055842,
+          "reward": 0.951417004048583
+        },
+        "fold": {
+          "accuracy": 0.266304347826087,
+          "balanced_accuracy": 0.16802592950505843,
+          "macro_f1": 0.03682564229435009,
+          "reward": 0.266304347826087
+        },
+        "secondary_structure": {
+          "accuracy": 0.7546679929829727,
+          "valid_residues": 556931,
+          "reward": 0.7546679929829727
+        },
+        "yeast_ppi": {
+          "accuracy": 0.6210526315789474,
+          "balanced_accuracy": 0.5618131868131868,
+          "macro_f1": 0.5343137254901961,
+          "reward": 0.6210526315789474
+        },
+        "human_ppi": {
+          "accuracy": 0.8031746031746032,
+          "balanced_accuracy": 0.8056972789115646,
+          "macro_f1": 0.8031250000000001,
+          "reward": 0.8031746031746032
+        },
+        "ppi_affinity": {
+          "rmse": 2.2471268230866137,
+          "spearman": 0.690737340413114,
+          "reward": 0.4824741399938253
+        }
+      }
+    },
+    {
+      "step": 17079,
+      "active_seconds": 600.007775742939,
+      "variant": "ema",
+      "reward": 0.6760988617070517,
+      "tasks": {
+        "fluorescence": {
+          "spearman": 0.7115087259027314,
+          "reward": 0.7115087259027314
+        },
+        "stability": {
+          "spearman": 0.7728268245558844,
+          "reward": 0.7728268245558844
+        },
+        "beta_lactamase": {
+          "spearman": 0.6068508114000871,
+          "reward": 0.6068508114000871
+        },
+        "solubility": {
+          "accuracy": 0.7522327859406511,
+          "balanced_accuracy": 0.7388007627343918,
+          "macro_f1": 0.7415399321436905,
+          "reward": 0.7522327859406511
+        },
+        "subcellular_localization": {
+          "accuracy": 0.7911775168979011,
+          "balanced_accuracy": 0.6478454007877607,
+          "macro_f1": 0.6664727010569773,
+          "reward": 0.7911775168979011
+        },
+        "binary_localization": {
+          "accuracy": 0.9537304800462695,
+          "balanced_accuracy": 0.9502660555292135,
+          "macro_f1": 0.9518806618148421,
+          "reward": 0.9537304800462695
+        },
+        "fold": {
+          "accuracy": 0.27717391304347827,
+          "balanced_accuracy": 0.16411162441107996,
+          "macro_f1": 0.03553839215700225,
+          "reward": 0.27717391304347827
+        },
+        "secondary_structure": {
+          "accuracy": 0.757379280377641,
+          "valid_residues": 556931,
+          "reward": 0.757379280377641
+        },
+        "yeast_ppi": {
+          "accuracy": 0.5578947368421052,
+          "balanced_accuracy": 0.5393772893772895,
+          "macro_f1": 0.5394736842105263,
+          "reward": 0.5578947368421052
+        },
+        "human_ppi": {
+          "accuracy": 0.8,
+          "balanced_accuracy": 0.7967687074829932,
+          "macro_f1": 0.7979144815225914,
+          "reward": 0.8
+        },
+        "ppi_affinity": {
+          "rmse": 2.419014493778693,
+          "spearman": 0.6994692296686369,
+          "reward": 0.4563124037708191
+        }
+      }
+    },
+    {
+      "step": 26646,
+      "active_seconds": 900.0244509173208,
+      "variant": "online",
+      "reward": 0.6921122266932366,
+      "tasks": {
+        "fluorescence": {
+          "spearman": 0.7536902950805767,
+          "reward": 0.7536902950805767
+        },
+        "stability": {
+          "spearman": 0.7780988747095005,
+          "reward": 0.7780988747095005
+        },
+        "beta_lactamase": {
+          "spearman": 0.6138726322274654,
+          "reward": 0.6138726322274654
+        },
+        "solubility": {
+          "accuracy": 0.747479112647652,
+          "balanced_accuracy": 0.7408445205453149,
+          "macro_f1": 0.7406185849766191,
+          "reward": 0.747479112647652
+        },
+        "subcellular_localization": {
+          "accuracy": 0.7908217716115261,
+          "balanced_accuracy": 0.6857218159177666,
+          "macro_f1": 0.6985840134990137,
+          "reward": 0.7908217716115261
+        },
+        "binary_localization": {
+          "accuracy": 0.9508386350491613,
+          "balanced_accuracy": 0.949396765186239,
+          "macro_f1": 0.949094938381547,
+          "reward": 0.9508386350491613
+        },
+        "fold": {
+          "accuracy": 0.27445652173913043,
+          "balanced_accuracy": 0.18473331971517096,
+          "macro_f1": 0.04106685362943855,
+          "reward": 0.27445652173913043
+        },
+        "secondary_structure": {
+          "accuracy": 0.7544166153437319,
+          "valid_residues": 556931,
+          "reward": 0.7544166153437319
+        },
+        "yeast_ppi": {
+          "accuracy": 0.6105263157894737,
+          "balanced_accuracy": 0.5684523809523809,
+          "macro_f1": 0.5606799150106236,
+          "reward": 0.6105263157894737
+        },
+        "human_ppi": {
+          "accuracy": 0.8317460317460318,
+          "balanced_accuracy": 0.8341836734693877,
+          "macro_f1": 0.8316849650666909,
+          "reward": 0.8317460317460318
+        },
+        "ppi_affinity": {
+          "rmse": 2.0925011048393993,
+          "spearman": 0.7079759001607104,
+          "reward": 0.5072876876813527
+        }
+      }
+    },
+    {
+      "step": 26646,
+      "active_seconds": 900.0244509173208,
+      "variant": "ema",
+      "reward": 0.6965855514841848,
+      "tasks": {
+        "fluorescence": {
+          "spearman": 0.7545942584485884,
+          "reward": 0.7545942584485884
+        },
+        "stability": {
+          "spearman": 0.7811837790587203,
+          "reward": 0.7811837790587203
+        },
+        "beta_lactamase": {
+          "spearman": 0.6218453637415401,
+          "reward": 0.6218453637415401
+        },
+        "solubility": {
+          "accuracy": 0.7500720253529243,
+          "balanced_accuracy": 0.7378772977288214,
+          "macro_f1": 0.7401121078424804,
+          "reward": 0.7500720253529243
+        },
+        "subcellular_localization": {
+          "accuracy": 0.8064745642120242,
+          "balanced_accuracy": 0.7015209248431455,
+          "macro_f1": 0.7185374784865061,
+          "reward": 0.8064745642120242
+        },
+        "binary_localization": {
+          "accuracy": 0.9560439560439561,
+          "balanced_accuracy": 0.9540066382171646,
+          "macro_f1": 0.9544133727804083,
+          "reward": 0.9560439560439561
+        },
+        "fold": {
+          "accuracy": 0.28532608695652173,
+          "balanced_accuracy": 0.185908082641295,
+          "macro_f1": 0.042895310434463485,
+          "reward": 0.28532608695652173
+        },
+        "secondary_structure": {
+          "accuracy": 0.7564689342126762,
+          "valid_residues": 556931,
+          "reward": 0.7564689342126762
+        },
+        "yeast_ppi": {
+          "accuracy": 0.6210526315789474,
+          "balanced_accuracy": 0.612408424908425,
+          "macro_f1": 0.6113636363636363,
+          "reward": 0.6210526315789474
+        },
+        "human_ppi": {
+          "accuracy": 0.8444444444444444,
+          "balanced_accuracy": 0.8435374149659864,
+          "macro_f1": 0.8436819558238219,
+          "reward": 0.8444444444444444
+        },
+        "ppi_affinity": {
+          "rmse": 2.2314407535138563,
+          "spearman": 0.709239707361978,
+          "reward": 0.48493502227569046
+        }
+      }
+    },
+    {
+      "step": 36201,
+      "active_seconds": 1200.0549627174623,
+      "variant": "online",
+      "reward": 0.7068145259166966,
+      "tasks": {
+        "fluorescence": {
+          "spearman": 0.7697553344883555,
+          "reward": 0.7697553344883555
+        },
+        "stability": {
+          "spearman": 0.7815438487087465,
+          "reward": 0.7815438487087465
+        },
+        "beta_lactamase": {
+          "spearman": 0.7030052474026528,
+          "reward": 0.7030052474026528
+        },
+        "solubility": {
+          "accuracy": 0.7503601267646212,
+          "balanced_accuracy": 0.7371938406083102,
+          "macro_f1": 0.7397922830353822,
+          "reward": 0.7503601267646212
+        },
+        "subcellular_localization": {
+          "accuracy": 0.8054073283528993,
+          "balanced_accuracy": 0.6970237166426081,
+          "macro_f1": 0.7104151826115701,
+          "reward": 0.8054073283528993
+        },
+        "binary_localization": {
+          "accuracy": 0.9606709080393291,
+          "balanced_accuracy": 0.9576813655761025,
+          "macro_f1": 0.9591177910560693,
+          "reward": 0.9606709080393291
+        },
+        "fold": {
+          "accuracy": 0.28804347826086957,
+          "balanced_accuracy": 0.1992417040511415,
+          "macro_f1": 0.047253443158646755,
+          "reward": 0.28804347826086957
+        },
+        "secondary_structure": {
+          "accuracy": 0.7517394434858178,
+          "valid_residues": 556931,
+          "reward": 0.7517394434858178
+        },
+        "yeast_ppi": {
+          "accuracy": 0.6,
+          "balanced_accuracy": 0.5828754578754579,
+          "macro_f1": 0.5833333333333333,
+          "reward": 0.6
+        },
+        "human_ppi": {
+          "accuracy": 0.8380952380952381,
+          "balanced_accuracy": 0.83843537414966,
+          "macro_f1": 0.8376764441390738,
+          "reward": 0.8380952380952381
+        },
+        "ppi_affinity": {
+          "rmse": 1.978832859750178,
+          "spearman": 0.7206986085239818,
+          "reward": 0.5263388314851328
+        }
+      }
+    },
+    {
+      "step": 36201,
+      "active_seconds": 1200.0549627174623,
+      "variant": "ema",
+      "reward": 0.7059204921603738,
+      "tasks": {
+        "fluorescence": {
+          "spearman": 0.7735445686605613,
+          "reward": 0.7735445686605613
+        },
+        "stability": {
+          "spearman": 0.7864386648331025,
+          "reward": 0.7864386648331025
+        },
+        "beta_lactamase": {
+          "spearman": 0.6914435427441995,
+          "reward": 0.6914435427441995
+        },
+        "solubility": {
+          "accuracy": 0.7503601267646212,
+          "balanced_accuracy": 0.736997908837965,
+          "macro_f1": 0.7396658834237162,
+          "reward": 0.7503601267646212
+        },
+        "subcellular_localization": {
+          "accuracy": 0.8071860547847741,
+          "balanced_accuracy": 0.7138959826937248,
+          "macro_f1": 0.7212493372383939,
+          "reward": 0.8071860547847741
+        },
+        "binary_localization": {
+          "accuracy": 0.9618276460381724,
+          "balanced_accuracy": 0.959327748801433,
+          "macro_f1": 0.9603571517902056,
+          "reward": 0.9618276460381724
+        },
+        "fold": {
+          "accuracy": 0.2907608695652174,
+          "balanced_accuracy": 0.19440267335004188,
+          "macro_f1": 0.044564923681055114,
+          "reward": 0.2907608695652174
+        },
+        "secondary_structure": {
+          "accuracy": 0.753764829036272,
+          "valid_residues": 556931,
+          "reward": 0.753764829036272
+        },
+        "yeast_ppi": {
+          "accuracy": 0.5894736842105263,
+          "balanced_accuracy": 0.5739468864468864,
+          "macro_f1": 0.5741868750718309,
+          "reward": 0.5894736842105263
+        },
+        "human_ppi": {
+          "accuracy": 0.8412698412698413,
+          "balanced_accuracy": 0.8414115646258504,
+          "macro_f1": 0.8408061777310585,
+          "reward": 0.8412698412698413
+        },
+        "ppi_affinity": {
+          "rmse": 2.021794847226883,
+          "spearman": 0.7103500215206399,
+          "reward": 0.5190555858568232
+        }
+      }
+    },
+    {
+      "step": 44746,
+      "active_seconds": 1500.0611601371202,
+      "variant": "online",
+      "reward": 0.7125092710156866,
+      "tasks": {
+        "fluorescence": {
+          "spearman": 0.7849124490375919,
+          "reward": 0.7849124490375919
+        },
+        "stability": {
+          "spearman": 0.7810242901567164,
+          "reward": 0.7810242901567164
+        },
+        "beta_lactamase": {
+          "spearman": 0.730575625515941,
+          "reward": 0.730575625515941
+        },
+        "solubility": {
+          "accuracy": 0.7545375972342264,
+          "balanced_accuracy": 0.741464274239623,
+          "macro_f1": 0.7441311054547298,
+          "reward": 0.7545375972342264
+        },
+        "subcellular_localization": {
+          "accuracy": 0.7986481679117752,
+          "balanced_accuracy": 0.687525081142312,
+          "macro_f1": 0.7004379594385564,
+          "reward": 0.7986481679117752
+        },
+        "binary_localization": {
+          "accuracy": 0.9624060150375939,
+          "balanced_accuracy": 0.9604868025920656,
+          "macro_f1": 0.9610025757697207,
+          "reward": 0.9624060150375939
+        },
+        "fold": {
+          "accuracy": 0.2907608695652174,
+          "balanced_accuracy": 0.20076805126714387,
+          "macro_f1": 0.04643683671356658,
+          "reward": 0.2907608695652174
+        },
+        "secondary_structure": {
+          "accuracy": 0.7507860040112689,
+          "valid_residues": 556931,
+          "reward": 0.7507860040112689
+        },
+        "yeast_ppi": {
+          "accuracy": 0.6,
+          "balanced_accuracy": 0.5789835164835164,
+          "macro_f1": 0.5794501397949674,
+          "reward": 0.6
+        },
+        "human_ppi": {
+          "accuracy": 0.8412698412698413,
+          "balanced_accuracy": 0.8414115646258504,
+          "macro_f1": 0.8408061777310585,
+          "reward": 0.8412698412698413
+        },
+        "ppi_affinity": {
+          "rmse": 1.8845585216485148,
+          "spearman": 0.7387534069747805,
+          "reward": 0.5426811214323801
+        }
+      }
+    },
+    {
+      "step": 44746,
+      "active_seconds": 1500.0611601371202,
+      "variant": "ema",
+      "reward": 0.7143941863830476,
+      "tasks": {
+        "fluorescence": {
+          "spearman": 0.7867063065181349,
+          "reward": 0.7867063065181349
+        },
+        "stability": {
+          "spearman": 0.7831602755914945,
+          "reward": 0.7831602755914945
+        },
+        "beta_lactamase": {
+          "spearman": 0.7276808611893643,
+          "reward": 0.7276808611893643
+        },
+        "solubility": {
+          "accuracy": 0.7529530394698934,
+          "balanced_accuracy": 0.7379493214283733,
+          "macro_f1": 0.7412394545644425,
+          "reward": 0.7529530394698934
+        },
+        "subcellular_localization": {
+          "accuracy": 0.8036286019210246,
+          "balanced_accuracy": 0.6999584425146658,
+          "macro_f1": 0.712152588252362,
+          "reward": 0.8036286019210246
+        },
+        "binary_localization": {
+          "accuracy": 0.9635627530364372,
+          "balanced_accuracy": 0.9619092776987513,
+          "macro_f1": 0.9622196440370219,
+          "reward": 0.9635627530364372
+        },
+        "fold": {
+          "accuracy": 0.2921195652173913,
+          "balanced_accuracy": 0.19564788407801117,
+          "macro_f1": 0.046417944884858654,
+          "reward": 0.2921195652173913
+        },
+        "secondary_structure": {
+          "accuracy": 0.7515957991205374,
+          "valid_residues": 556931,
+          "reward": 0.7515957991205374
+        },
+        "yeast_ppi": {
+          "accuracy": 0.6,
+          "balanced_accuracy": 0.5789835164835164,
+          "macro_f1": 0.5794501397949674,
+          "reward": 0.6
+        },
+        "human_ppi": {
+          "accuracy": 0.8476190476190476,
+          "balanced_accuracy": 0.8473639455782314,
+          "macro_f1": 0.8470626314937693,
+          "reward": 0.8476190476190476
+        },
+        "ppi_affinity": {
+          "rmse": 1.8471262780671165,
+          "spearman": 0.7455006455345559,
+          "reward": 0.549309800530199
+        }
+      }
+    },
+    {
+      "step": 53294,
+      "active_seconds": 1800.0039906070306,
+      "variant": "online",
+      "reward": 0.7159026524944158,
+      "tasks": {
+        "fluorescence": {
+          "spearman": 0.7916969155160615,
+          "reward": 0.7916969155160615
+        },
+        "stability": {
+          "spearman": 0.7800266830014776,
+          "reward": 0.7800266830014776
+        },
+        "beta_lactamase": {
+          "spearman": 0.7389957766382556,
+          "reward": 0.7389957766382556
+        },
+        "solubility": {
+          "accuracy": 0.7533851915874388,
+          "balanced_accuracy": 0.7399365867166623,
+          "macro_f1": 0.7427105941881498,
+          "reward": 0.7533851915874388
+        },
+        "subcellular_localization": {
+          "accuracy": 0.80042689434365,
+          "balanced_accuracy": 0.6903173652477832,
+          "macro_f1": 0.7048552924485659,
+          "reward": 0.80042689434365
+        },
+        "binary_localization": {
+          "accuracy": 0.9624060150375939,
+          "balanced_accuracy": 0.9609346188293557,
+          "macro_f1": 0.9610378354391542,
+          "reward": 0.9624060150375939
+        },
+        "fold": {
+          "accuracy": 0.29483695652173914,
+          "balanced_accuracy": 0.19567773928028018,
+          "macro_f1": 0.047001763948018305,
+          "reward": 0.29483695652173914
+        },
+        "secondary_structure": {
+          "accuracy": 0.7503353198152015,
+          "valid_residues": 556931,
+          "reward": 0.7503353198152015
+        },
+        "yeast_ppi": {
+          "accuracy": 0.5894736842105263,
+          "balanced_accuracy": 0.570054945054945,
+          "macro_f1": 0.5704347826086956,
+          "reward": 0.5894736842105263
+        },
+        "human_ppi": {
+          "accuracy": 0.8507936507936508,
+          "balanced_accuracy": 0.8503401360544218,
+          "macro_f1": 0.8501897293195042,
+          "reward": 0.8507936507936508
+        },
+        "ppi_affinity": {
+          "rmse": 1.7736807556553096,
+          "spearman": 0.7516747955847892,
+          "reward": 0.5625520899729786
+        }
+      }
+    },
+    {
+      "step": 53294,
+      "active_seconds": 1800.0039906070306,
+      "variant": "ema",
+      "reward": 0.7163740028281073,
+      "tasks": {
+        "fluorescence": {
+          "spearman": 0.7918192514022675,
+          "reward": 0.7918192514022675
+        },
+        "stability": {
+          "spearman": 0.780208804465382,
+          "reward": 0.780208804465382
+        },
+        "beta_lactamase": {
+          "spearman": 0.742143502666086,
+          "reward": 0.742143502666086
+        },
+        "solubility": {
+          "accuracy": 0.7538173437049841,
+          "balanced_accuracy": 0.7404053807847767,
+          "macro_f1": 0.74317716963567,
+          "reward": 0.7538173437049841
+        },
+        "subcellular_localization": {
+          "accuracy": 0.8018498754891498,
+          "balanced_accuracy": 0.6917805302447911,
+          "macro_f1": 0.7068148859422272,
+          "reward": 0.8018498754891498
+        },
+        "binary_localization": {
+          "accuracy": 0.9624060150375939,
+          "balanced_accuracy": 0.9609346188293557,
+          "macro_f1": 0.9610378354391542,
+          "reward": 0.9624060150375939
+        },
+        "fold": {
+          "accuracy": 0.29483695652173914,
+          "balanced_accuracy": 0.19703615098351948,
+          "macro_f1": 0.04734856523385019,
+          "reward": 0.29483695652173914
+        },
+        "secondary_structure": {
+          "accuracy": 0.7503389109243336,
+          "valid_residues": 556931,
+          "reward": 0.7503389109243336
+        },
+        "yeast_ppi": {
+          "accuracy": 0.5894736842105263,
+          "balanced_accuracy": 0.570054945054945,
+          "macro_f1": 0.5704347826086956,
+          "reward": 0.5894736842105263
+        },
+        "human_ppi": {
+          "accuracy": 0.8507936507936508,
+          "balanced_accuracy": 0.8503401360544218,
+          "macro_f1": 0.8501897293195042,
+          "reward": 0.8507936507936508
+        },
+        "ppi_affinity": {
+          "rmse": 1.7743717036904072,
+          "spearman": 0.7498163821577267,
+          "reward": 0.5624260358934674
+        }
+      }
+    }
+  ],
+  "parameters": {
+    "backbone": 21306271,
+    "total": 25065830
+  },
+  "execution_recovery": {
+    "preflight": {
+      "utc": "2026-11-01T15:29:37.519257+00:00",
+      "status": "passed",
+      "old_pid_namespace": 4026551139,
+      "new_pid_namespace": "pid:[4026550898]",
+      "original_process_handles_absent": true,
+      "source_checkpoint_sha256": "091f83530d8976b879ef7dad2bf752655f7dac346e0dcf7f3fa34bd8170018d3",
+      "source_step": 36201,
+      "source_active_seconds": 1200.0549627174623,
+      "original_horizon_seconds": 1800,
+      "remaining_nominal_seconds": 599.9450372825377,
+      "retained_history_records": 8,
+      "source_best_reward": 0.7068145259166966,
+      "model_and_ema_tensors": 284,
+      "optimizer_entries": 283,
+      "mask_state_keys": [
+        "config",
+        "rng_state",
+        "statistics",
+        "updates",
+        "rows"
+      ],
+      "last_logged_step": 39000,
+      "last_logged_active_seconds_rounded": 1286,
+      "discarded_suffix_logged_updates_lower_bound": 2799,
+      "discarded_suffix_active_seconds_conservative_lower_bound": 84.94503728253767,
+      "discarded_suffix_exact_duration_unknown": true,
+      "original_recipe_code_hashes": {
+        "esm2.py": "8e652c3a3db1ccb7114ea353170912360e304841ef42afa915b9b432af448209",
+        "predict.py": "03cbc374ba04fb23d31702216481b4c15cd67cbe05516a2ec3bc03f859b80110",
+        "train.py": "758c20d4dc130d5576de0c1eda96b8a49e32655fa26ba7d5fca995d60bfece37",
+        "finetune.py": "5dc6b7e2a36c07fb63c275428b1012179eddc9a0af9fa915f13648b10e0d61ee",
+        "task_gradient_gate.py": "76268983c198ec6e6b104f3d4732f5335924e046f6941a4ae05c219aef35cff6",
+        "fold_mask.py": "80f7aa3434693801356e145cb2fface432c731aa63e588627276db60459f1875"
+      },
+      "resume_command": [
+        "python3",
+        "-u",
+        "/workspace/research/unified_fold_mask/finetune.py",
+        "--resume",
+        "/workspace/research/21m_4h_joint_fold_mask/ck/finetune.pt",
+        "--model",
+        "unified_t12_21M",
+        "--normalized-regression",
+        "--head-lr-mult",
+        "2",
+        "--random-batches",
+        "--ema-decay",
+        ".999",
+        "--correlation-loss",
+        ".1",
+        "--task-weights",
+        "{\"human_ppi\":0.25}",
+        "--stop-shared-tasks",
+        "[\"human_ppi\", \"solubility\"]",
+        "--out",
+        "/workspace/research/21m_4h_joint_fold_mask",
+        "--hours",
+        ".5",
+        "--eval-seconds",
+        "300",
+        "--stop-shared-after-seconds",
+        "300",
+        "--continue-state"
+      ],
+      "scope": "Own full supervised checkpoint, original1800s horizon, optimizer/model/EMA/Python-Torch-CUDA-mask RNG/task-gate-mask counts/fullhistory preserved. No prior smoke/replay repeated. Active-time LR after recovery need not be bit-identical to discarded suffix."
+    },
+    "actual_continuation_audit": {
+      "status": "passed",
+      "source_checkpoint_sha256": "091f83530d8976b879ef7dad2bf752655f7dac346e0dcf7f3fa34bd8170018d3",
+      "source_step": 36201,
+      "new_step": 44746,
+      "source_active_seconds": 1200.0549627174623,
+      "new_active_seconds": 1500.0611601371202,
+      "preserved_history_records": 8,
+      "current_history_records": 10,
+      "preserved_best_reward": 0.7068145259166966,
+      "current_best_reward": 0.7143941863830476,
+      "changed_model_tensors": 279,
+      "total_model_tensors": 284,
+      "repeated_sampling_log_steps": [
+        36500,
+        37000,
+        37500,
+        38000,
+        38500,
+        39000
+      ],
+      "checks": [
+        "Complete selection history and previous best retained",
+        "Optimizer step counters and all task exposure retained and advanced",
+        "Original recipe and ancestor provenance retained",
+        "Model and EMA shapes/finite values valid",
+        "Task sampling matches the discarded suffix at every shared logged update"
+      ],
+      "limitations": "Active-time learning rates depend on runtime; resumed losses and later parameters need not be bit-identical to discarded unsaved work.",
+      "fold_mask_continuation": {
+        "config": {
+          "seed": 1611,
+          "row_probability": 0.5,
+          "residue_probability": 0.1
+        },
+        "before_updates": 3238,
+        "after_updates": 4074,
+        "before_statistics": [
+          17412525,
+          51786,
+          51721,
+          867860
+        ],
+        "after_statistics": [
+          21800573,
+          65120,
+          65033,
+          1087733
+        ],
+        "rng_advanced": true
+      },
+      "shared_gate_counters_preserved": true
+    },
+    "actual_continuation_audit_exit": {
+      "utc": "2026-11-01T15:35:22.002794+00:00",
+      "actual_exit_code": 0
+    }
+  },
+  "fixed_readout_precision_control": {
+    "policy": {
+      "utc": "2026-11-01T15:38:09.324940+00:00",
+      "status": "bound_after_completed_native_review",
+      "source_run": "/workspace/research/21m_4h_joint_fold_mask",
+      "source": "/workspace/research/21m_4h_joint_fold_mask/package",
+      "source_selection": "Only this run\u2019s single complete online/EMA state selected by its original equal-weight11-task tune mean. Bind exact model/source-review hashes after full native replay/source/root review, before full extraction/fitting. No alternate body, average or per-task state selection.",
+      "source_code_sha256": {
+        "esm2.py": "8e652c3a3db1ccb7114ea353170912360e304841ef42afa915b9b432af448209",
+        "predict.py": "03cbc374ba04fb23d31702216481b4c15cd67cbe05516a2ec3bc03f859b80110",
+        "train.py": "758c20d4dc130d5576de0c1eda96b8a49e32655fa26ba7d5fca995d60bfece37",
+        "finetune.py": "5dc6b7e2a36c07fb63c275428b1012179eddc9a0af9fa915f13648b10e0d61ee",
+        "task_gradient_gate.py": "76268983c198ec6e6b104f3d4732f5335924e046f6941a4ae05c219aef35cff6",
+        "fold_mask.py": "80f7aa3434693801356e145cb2fface432c731aa63e588627276db60459f1875"
+      },
+      "parameters_before_readouts": {
+        "backbone": 21306271,
+        "total": 25040355
+      },
+      "hypothesis": "At1200and1500seconds fold masking improves fold and secondary structure in both native variants versus the gated control, while beta remains weak. The already published fixed beta/GFP readouts may transfer to this single shared body and retain its nine other task predictions. Test once with the identical fixed hyperparameters and original gates.",
+      "scope": "One whole-state fold-mask native body selected by its original11taskmean; fixed beta/GFP only. No alternative state, average, per-task body, ridge, width, step, seed or precision search. Preserve nine other full prediction files if both pass.",
+      "selection_change_from_older_studies": "Fix the previously published hyperparameters in advance, instead of repeating their nine-ridge/checkpoint searches. This is a fixed hyperparameter transfer, not the complete older selection protocol.",
+      "precision_control": "Copy the exact native weights; enable full FP32 for beta and GFP, beta token budget8192 and all others16384. Independently replay all eleven once and require nine unchanged full prediction files before fitting.",
+      "beta": {
+        "relative_ridge": 0.0001,
+        "width": 64,
+        "steps": 14000,
+        "seed": 200,
+        "batch": 128,
+        "lr": 0.001,
+        "decay": 0.01,
+        "dropout": 0.1,
+        "feature_noise": 0.02,
+        "clip": 1,
+        "target": "Empirical fit-label average-rank quantiles",
+        "selection": "Use the final14000-step state only; no intermediate quality evaluation or selection. Report ridge-only control.",
+        "adoption": "Gain>=.02 and positive1000 paired-row95percent interval versus FP32 native control; independently replayed cached reward agreement within2e-5. Also report gain/interval versus original native."
+      },
+      "gfp": {
+        "relative_ridge": 0.1,
+        "seed": 802,
+        "target": "Original fit labels minus unchanged FP32 native prediction",
+        "features": "Native penultimate384+log1p substitution count1+sum shared-encoder contextual substitution deltas384, plus intercept; fit-only237-residue consensus.",
+        "control": "Same fixed.1 ridge using penultimate+logcount only; unchanged parent also retained.",
+        "adoption": "Gain>=.01 versus FP32 native AND original native, positive1000 paired intervals versus parent and count control; same2e-5 cached/replay tolerance. No exception to a failed floor."
+      },
+      "resources": "OneH100 remains serialized for training/replay. All beta and GFP feature extraction and fitting CPU-only under the shared readout lock, at most8threads plus4trainerthreads. Beta uses existing generic FP32/TF32off CPU extractor,8192token batches; this differs from the publication\u2019s CUDA beta cache, so the original2e-5 numerical replay gate remains mandatory, with no rescue/tolerance change. Precision-control and candidate fullreplays use existing paused trainer boundaries during already registered later trainers, initially foldweight41. Preserve their training budgets and prior45queue order. No additional GPU training stage.",
+      "assembly": "Only accepted fixed readouts may be assembled; preserve nine other tasks exactly. If only one passes, assemble that one and preserve the other ten tasks. No automatic publication.",
+      "publication": "One independently loaded cold check, exact protected weights/predictions, complete offline eleven-task replay, source/provenance/count audit, all-task paired row/SSchain/PPIcomponent comparison to native and live, GFP strata review if changed. Publish only with stronger complete evidence and atomic publisher.",
+      "failure": "Preserve first numeric/extraction/fit/cold/replay failure and stop dependent work. No grid expansion, precision repair or tolerance relaxation. Bookkeeping fixes may reuse completed outputs without new fitting or inference.",
+      "limitations": "Adaptive research from the visible1500s masked trajectory and earlier fixed transfers; one own4hancestor/seed, repeated tune selection, CPU/GPU numerics and unmodeled homology dependence. Source run includes an infrastructure restart and at least84.945discarded active seconds; record recovery in final provenance. No hidden-test improvement or full independent replication is claimed.",
+      "preparation": {
+        "scope": "Reuse unchanged fixed published beta/GFP mathematics, validated generic feature extraction/runtime/export gates, and the existing base architecture already used by published and fold-variant fixed studies. Helper/source binding preparation follows; no extraction, fitting, new fixture or replay has begun.",
+        "base_model_sha256": "8e652c3a3db1ccb7114ea353170912360e304841ef42afa915b9b432af448209",
+        "base_predict_sha256": "03cbc374ba04fb23d31702216481b4c15cd67cbe05516a2ec3bc03f859b80110",
+        "closed_prior_transfer": "/workspace/research/21m_4h_content_gates_fixed_readouts/root_final_review.json"
+      },
+      "worker_status": "Not launched; prepare isolated path-specific helpers and execution hash manifest before a single waiting worker. Bind only after completed native source/root review.",
+      "registration_context": {
+        "utc": "2026-11-01T15:38:09.324940+00:00",
+        "available_source_point": {
+          "step": 44746,
+          "active_seconds": 1500.0611601371202,
+          "variant": "ema",
+          "reward": 0.7143941863830476
+        },
+        "source_1500_review_sha256": "59ed83867a197d29888201eefd8ccbe4b34ddc90449b509bceebb4d79f195ba8",
+        "final_native_evaluation_available": false,
+        "complete_previous_transfer": "Content-gates fixed transfer closed HOLD at0.7265155418; small aggregate near-tie with function/human regressions. This motivates preserving all-task review, not assuming readout transfer makes a better publication."
+      },
+      "bound_utc": "2026-11-01T15:42:39.100936+00:00",
+      "model_sha256": "5a59a5ace990511aa072a7818f5d160907e4fcaa8833b2ec86a4c9a1b3ca0367",
+      "source_review_sha256": "c83522d7c082dcf558f2ebee97666fad41def6f3243a1be2e55c84be1b18565d",
+      "central_policy_sha256": "51282e50721a05daed94610032ab928c16900d218986395f373f1812bd1be751"
+    },
+    "source_config": {
+      "name": "unified_t12_21M",
+      "layers": 12,
+      "dim": 384,
+      "heads": 6,
+      "structured_pooling": true,
+      "full_chain_ss": true,
+      "full_precision_head_tasks": [
+        "fluorescence",
+        "stability",
+        "beta_lactamase",
+        "ppi_affinity"
+      ],
+      "weight_variant": "ema"
+    },
+    "changed_tasks": [
+      "beta_lactamase",
+      "fluorescence"
+    ]
+  },
+  "normalization_buffer_values": 769,
+  "fixed_native_readout_transfer": {
+    "central_policy": {
+      "utc": "2026-11-01T15:38:09.324940+00:00",
+      "status": "bound_after_completed_native_review",
+      "source_run": "/workspace/research/21m_4h_joint_fold_mask",
+      "source": "/workspace/research/21m_4h_joint_fold_mask/package",
+      "source_selection": "Only this run\u2019s single complete online/EMA state selected by its original equal-weight11-task tune mean. Bind exact model/source-review hashes after full native replay/source/root review, before full extraction/fitting. No alternate body, average or per-task state selection.",
+      "source_code_sha256": {
+        "esm2.py": "8e652c3a3db1ccb7114ea353170912360e304841ef42afa915b9b432af448209",
+        "predict.py": "03cbc374ba04fb23d31702216481b4c15cd67cbe05516a2ec3bc03f859b80110",
+        "train.py": "758c20d4dc130d5576de0c1eda96b8a49e32655fa26ba7d5fca995d60bfece37",
+        "finetune.py": "5dc6b7e2a36c07fb63c275428b1012179eddc9a0af9fa915f13648b10e0d61ee",
+        "task_gradient_gate.py": "76268983c198ec6e6b104f3d4732f5335924e046f6941a4ae05c219aef35cff6",
+        "fold_mask.py": "80f7aa3434693801356e145cb2fface432c731aa63e588627276db60459f1875"
+      },
+      "parameters_before_readouts": {
+        "backbone": 21306271,
+        "total": 25040355
+      },
+      "hypothesis": "At1200and1500seconds fold masking improves fold and secondary structure in both native variants versus the gated control, while beta remains weak. The already published fixed beta/GFP readouts may transfer to this single shared body and retain its nine other task predictions. Test once with the identical fixed hyperparameters and original gates.",
+      "scope": "One whole-state fold-mask native body selected by its original11taskmean; fixed beta/GFP only. No alternative state, average, per-task body, ridge, width, step, seed or precision search. Preserve nine other full prediction files if both pass.",
+      "selection_change_from_older_studies": "Fix the previously published hyperparameters in advance, instead of repeating their nine-ridge/checkpoint searches. This is a fixed hyperparameter transfer, not the complete older selection protocol.",
+      "precision_control": "Copy the exact native weights; enable full FP32 for beta and GFP, beta token budget8192 and all others16384. Independently replay all eleven once and require nine unchanged full prediction files before fitting.",
+      "beta": {
+        "relative_ridge": 0.0001,
+        "width": 64,
+        "steps": 14000,
+        "seed": 200,
+        "batch": 128,
+        "lr": 0.001,
+        "decay": 0.01,
+        "dropout": 0.1,
+        "feature_noise": 0.02,
+        "clip": 1,
+        "target": "Empirical fit-label average-rank quantiles",
+        "selection": "Use the final14000-step state only; no intermediate quality evaluation or selection. Report ridge-only control.",
+        "adoption": "Gain>=.02 and positive1000 paired-row95percent interval versus FP32 native control; independently replayed cached reward agreement within2e-5. Also report gain/interval versus original native."
+      },
+      "gfp": {
+        "relative_ridge": 0.1,
+        "seed": 802,
+        "target": "Original fit labels minus unchanged FP32 native prediction",
+        "features": "Native penultimate384+log1p substitution count1+sum shared-encoder contextual substitution deltas384, plus intercept; fit-only237-residue consensus.",
+        "control": "Same fixed.1 ridge using penultimate+logcount only; unchanged parent also retained.",
+        "adoption": "Gain>=.01 versus FP32 native AND original native, positive1000 paired intervals versus parent and count control; same2e-5 cached/replay tolerance. No exception to a failed floor."
+      },
+      "resources": "OneH100 remains serialized for training/replay. All beta and GFP feature extraction and fitting CPU-only under the shared readout lock, at most8threads plus4trainerthreads. Beta uses existing generic FP32/TF32off CPU extractor,8192token batches; this differs from the publication\u2019s CUDA beta cache, so the original2e-5 numerical replay gate remains mandatory, with no rescue/tolerance change. Precision-control and candidate fullreplays use existing paused trainer boundaries during already registered later trainers, initially foldweight41. Preserve their training budgets and prior45queue order. No additional GPU training stage.",
+      "assembly": "Only accepted fixed readouts may be assembled; preserve nine other tasks exactly. If only one passes, assemble that one and preserve the other ten tasks. No automatic publication.",
+      "publication": "One independently loaded cold check, exact protected weights/predictions, complete offline eleven-task replay, source/provenance/count audit, all-task paired row/SSchain/PPIcomponent comparison to native and live, GFP strata review if changed. Publish only with stronger complete evidence and atomic publisher.",
+      "failure": "Preserve first numeric/extraction/fit/cold/replay failure and stop dependent work. No grid expansion, precision repair or tolerance relaxation. Bookkeeping fixes may reuse completed outputs without new fitting or inference.",
+      "limitations": "Adaptive research from the visible1500s masked trajectory and earlier fixed transfers; one own4hancestor/seed, repeated tune selection, CPU/GPU numerics and unmodeled homology dependence. Source run includes an infrastructure restart and at least84.945discarded active seconds; record recovery in final provenance. No hidden-test improvement or full independent replication is claimed.",
+      "preparation": {
+        "scope": "Reuse unchanged fixed published beta/GFP mathematics, validated generic feature extraction/runtime/export gates, and the existing base architecture already used by published and fold-variant fixed studies. Helper/source binding preparation follows; no extraction, fitting, new fixture or replay has begun.",
+        "base_model_sha256": "8e652c3a3db1ccb7114ea353170912360e304841ef42afa915b9b432af448209",
+        "base_predict_sha256": "03cbc374ba04fb23d31702216481b4c15cd67cbe05516a2ec3bc03f859b80110",
+        "closed_prior_transfer": "/workspace/research/21m_4h_content_gates_fixed_readouts/root_final_review.json"
+      },
+      "worker_status": "Not launched; prepare isolated path-specific helpers and execution hash manifest before a single waiting worker. Bind only after completed native source/root review.",
+      "registration_context": {
+        "utc": "2026-11-01T15:38:09.324940+00:00",
+        "available_source_point": {
+          "step": 44746,
+          "active_seconds": 1500.0611601371202,
+          "variant": "ema",
+          "reward": 0.7143941863830476
+        },
+        "source_1500_review_sha256": "59ed83867a197d29888201eefd8ccbe4b34ddc90449b509bceebb4d79f195ba8",
+        "final_native_evaluation_available": false,
+        "complete_previous_transfer": "Content-gates fixed transfer closed HOLD at0.7265155418; small aggregate near-tie with function/human regressions. This motivates preserving all-task review, not assuming readout transfer makes a better publication."
+      },
+      "bound_utc": "2026-11-01T15:42:39.100936+00:00",
+      "model_sha256": "5a59a5ace990511aa072a7818f5d160907e4fcaa8833b2ec86a4c9a1b3ca0367",
+      "source_review_sha256": "c83522d7c082dcf558f2ebee97666fad41def6f3243a1be2e55c84be1b18565d",
+      "central_policy_sha256": "51282e50721a05daed94610032ab928c16900d218986395f373f1812bd1be751"
+    },
+    "root_fit_review": {
+      "utc": "2026-11-01T16:00:39.677338+00:00",
+      "status": "cached_fixed_readout_review_complete",
+      "accepted_tasks": [
+        "beta_lactamase",
+        "fluorescence"
+      ],
+      "source_model_sha256": "5a59a5ace990511aa072a7818f5d160907e4fcaa8833b2ec86a4c9a1b3ca0367",
+      "beta": {
+        "baseline": 0.7403514772537801,
+        "rank_ridge_best": 0.7872673810383397,
+        "best": 0.8456129035533306,
+        "gain": 0.10526142629955049,
+        "best_step": 14000,
+        "paired_row_gain_95pct": [
+          0.0726677734073657,
+          0.14099973147734582
+        ],
+        "adopt_threshold": 0.02,
+        "adopt": true,
+        "ridge_trials": [
+          {
+            "relative_ridge": 0.0001,
+            "reward": 0.7872673810383397
+          }
+        ],
+        "residual_history": [
+          {
+            "step": 14000,
+            "reward": 0.8456129035533306
+          }
+        ],
+        "source_model_sha256": "5a59a5ace990511aa072a7818f5d160907e4fcaa8833b2ec86a4c9a1b3ca0367",
+        "unchanged_feature_tensors": 222,
+        "output_scale": "Fit-label rank quantiles, not original physical units.",
+        "uncertainty": "Paired row interval excludes model-selection and homology dependence."
+      },
+      "gfp": {
+        "baseline": 0.79199891808651,
+        "official_baseline": 0.7919983783106186,
+        "fit_baseline": 0.8015985276579471,
+        "variants": {
+          "count_control": {
+            "reward": 0.7941345130236414,
+            "ridge": 0.1,
+            "trainable_coefficients": 386,
+            "torch_reproduction_reward": 0.7941345130236414,
+            "fit_reward": 0.8031659726615605,
+            "initialization_seed": 802,
+            "optimizer": "Exact ridge minimizer",
+            "history": [
+              {
+                "ridge": 0.1,
+                "reward": 0.7941345130236414
+              }
+            ],
+            "correction_accumulation": "Float64 over float32 features/coefficients, cast correction to float32 before adding base."
+          },
+          "mutation_delta": {
+            "reward": 0.80349773201962,
+            "ridge": 0.1,
+            "trainable_coefficients": 770,
+            "torch_reproduction_reward": 0.80349773201962,
+            "fit_reward": 0.8118414408701651,
+            "initialization_seed": 802,
+            "optimizer": "Exact ridge minimizer",
+            "history": [
+              {
+                "ridge": 0.1,
+                "reward": 0.80349773201962
+              }
+            ],
+            "correction_accumulation": "Float64 over float32 features/coefficients, cast correction to float32 before adding base."
+          }
+        },
+        "gain": 0.01149881393311003,
+        "gain_over_count_control": 0.009363218995978695,
+        "paired_row_gain_95pct": [
+          0.008131903291103443,
+          0.014888659571684124
+        ],
+        "paired_row_gain_over_control_95pct": [
+          0.006516736807961018,
+          0.012521238826549685
+        ],
+        "adopt": true,
+        "unchanged_different_length_tune_rows": 127,
+        "fit_eligible_rows": 20963,
+        "tune_eligible_rows": 5235,
+        "uncertainty": "Intervals exclude model-selection and homology dependence."
+      },
+      "gfp_original_native_gain": 0.011679138103189435,
+      "gfp_cached_gain_over_official_fp32_native": 0.011499353709001459,
+      "hypothetical_complete_mean_from_cached_readouts": 0.7268420032776292,
+      "hypothetical_delta_from_live": 0.0003333666570310978,
+      "hypothetical_all_task_deltas_from_live": {
+        "beta_lactamase": 0.017120446233956033,
+        "binary_localization": 0.006940427993059495,
+        "fluorescence": -0.005010069113147386,
+        "fold": 0.021739130434782594,
+        "human_ppi": 0.0031746031746031633,
+        "ppi_affinity": -0.0036920705560915534,
+        "secondary_structure": 0.004138753274642637,
+        "solubility": -0.004033419763756929,
+        "stability": 0.0004140606979395889,
+        "subcellular_localization": 0.004980434009249413,
+        "yeast_ppi": -0.04210526315789476
+      },
+      "publication_reference": "/workspace/research/21m_4h_two_task_fixed_readouts/candidate/replay/report.json",
+      "publication_reference_sha256": "7e7b32ed3339d52c0a2ef2c2ba6deefa4f1ffb962a934797d787f67f91034f7f",
+      "subgroup_and_all_task_review": [
+        "Accept the two prospectively fixed readout recipes for one independent assembled-candidate evaluation. Beta uses only ridge0.0001 plus64residual units, exactly14000updates/seed200; cachedreward0.8456129035533306 versusFP32native0.7403514772537801, gain0.1052614263,paired95[0.0726677734,0.1409997315]. All extraction/fitting source checks are complete. This is cached evidence, not a full package score.",
+        "GFP raw-unit correction uses only ridge0.1/seed802 and770coefficients. Cachedreward0.80349773201962 versusofficialFP32native0.7919983783106186 andcountcontrol0.7941345130236414. Its improvement exceeds0.01 versus bothFP32 andoriginalnative, and the existing paired intervals are positive versusparent[0.0081319033,0.0148886596] andcountcontrol[0.0065167368,0.0125212388]. The tinyCPUcached-parent versusGPUofficial-parent difference remains subject to the original2e-5 fullreplay tolerance; no precision repair or waiver is authorized.",
+        "All four fit-label quartiles make positive global-rank contributions. Conditional correlations improve with positive paired intervals in quartiles1,2and3. Quartile4 has a negative conditional point change,0.2390557846to0.2303228627,withinterval[-0.0304665416,0.0140238721]includingzero. Global contribution andwithin-subgroup correlation answer different questions; positive contribution does not establish every subgroup improves.",
+        "Two-substitution andthree-substitution groups have positive conditional paired intervals andpositive global contributions. Zero-or-one substitutions have a positive conditional point change withintervalincludingzero butnegative global contribution. No alignedtune example has4+substitutions, so the study provides no distant-mutation evidence. The55examples containinganunseen substitution decline0.8150072150to0.8002886003,withinterval[-0.0558712717,0.0310380116]includingzero andnegative global contribution. The5180fit-seen-substitutionexamples improve withpositiveinterval. No subgroup-specific inference routing is introduced.",
+        "The127different-length GFP predictions are exactly preserved in cached features. Their global-rank contribution neednotbezero because ranks relative to changed examples canmove; their conditional correlation anddirectpredictions are unchanged. Fullindependent replay must confirm the same127outputs andnineother complete task files, plus original gain andpaired gates.",
+        "Combining accepted cached heads with the original native body projects a small aggregate improvement over the publication, withseven task gains(beta,binary,fold,human,secondary structure,stability,subcellular) andfour declines(fluorescence,affinity,solubility,yeast). GFP remains belowthepublished GFPhead even though it improves thisbody. The projected overall difference is not an independent package score or certain improvement. Review everytask androw/chain/PPIcomponent intervals after fullreplay before choosing publication.",
+        "Proceed once to accepted-only assembly, strict export/preservation audit,111-row standalone cold check, inference-file hash binding andoneall11GPU replay at the nextpausedtraining boundary. Keep all original thresholds unchanged. No alternative body, readoutsearch, average, extractionrepeat, bootstraprepeat orautomaticpublication. Preserve the masked-body runtime-recovery anddiscardedcompute provenance."
+      ],
+      "subprocess_exits": {
+        "beta_extract": {
+          "utc": "2026-11-01T15:48:56.020331+00:00",
+          "actual_exit_code": 0,
+          "command": [
+            "python3",
+            "-u",
+            "/workspace/research/cache_frozen_features.py",
+            "--package",
+            "/workspace/research/21m_4h_fold_mask_fixed_readouts/precision_control/package",
+            "--out",
+            "/workspace/research/21m_4h_fold_mask_fixed_readouts/beta_features",
+            "--tasks",
+            "beta_lactamase",
+            "--device",
+            "cpu",
+            "--threads",
+            "8"
+          ]
+        },
+        "beta_fit": {
+          "utc": "2026-11-01T15:49:18.095970+00:00",
+          "actual_exit_code": 0,
+          "command": [
+            "python3",
+            "-u",
+            "/workspace/research/21m_4h_fold_mask_fixed_readouts/fit_beta_fixed.py",
+            "--source",
+            "/workspace/research/21m_4h_fold_mask_fixed_readouts/precision_control/package",
+            "--features",
+            "/workspace/research/21m_4h_fold_mask_fixed_readouts/beta_features",
+            "--out",
+            "/workspace/research/21m_4h_fold_mask_fixed_readouts/beta_study",
+            "--threads",
+            "4"
+          ]
+        },
+        "gfp_extract": {
+          "utc": "2026-11-01T15:59:08.639046+00:00",
+          "actual_exit_code": 0,
+          "command": [
+            "python3",
+            "-u",
+            "/workspace/research/extract_mutation_features.py",
+            "--source",
+            "/workspace/research/21m_4h_fold_mask_fixed_readouts/precision_control/package",
+            "--out",
+            "/workspace/research/21m_4h_fold_mask_fixed_readouts/gfp_study/features",
+            "--threads",
+            "8",
+            "--device",
+            "cpu",
+            "--tasks",
+            "fluorescence"
+          ]
+        },
+        "gfp_fit": {
+          "utc": "2026-11-01T15:59:34.320574+00:00",
+          "actual_exit_code": 0,
+          "command": [
+            "python3",
+            "-u",
+            "/workspace/research/fit_raw_mutation_readouts.py",
+            "--study",
+            "/workspace/research/21m_4h_fold_mask_fixed_readouts/gfp_study",
+            "--threads",
+            "8"
+          ]
+        },
+        "gfp_strata": {
+          "utc": "2026-11-01T16:00:17.968657+00:00",
+          "actual_exit_code": 0,
+          "command": [
+            "python3",
+            "/workspace/research/analyze_mutation_readouts.py",
+            "--study",
+            "/workspace/research/21m_4h_fold_mask_fixed_readouts/gfp_study"
+          ]
+        }
+      },
+      "decision": "Proceed once to accepted-only assembly, export audit, cold check and full replay; no publication yet.",
+      "artifact_sha256": {
+        "hypothesis.json": "51282e50721a05daed94610032ab928c16900d218986395f373f1812bd1be751",
+        "bound_policy.json": "31bd81c53973ed869580d016ccb2a8cf66a6e92ae19830dacbbe1175843ab256",
+        "source_binding.json": "49c5d09ec11c01d4c88af45314c1e7e1a84d279abba974ff3b6d93142bc33551",
+        "beta_study/results.json": "ef0b3060216f2f01f9333c43bb19e7a3f62b0b3252f9aec505b5286fbe356d44",
+        "gfp_study/results.json": "f0fe85ae83dc57676c5a923f74aa32e65e4b1bfcd3a39ae4b210d9b591810545",
+        "gfp_study/strata.json": "bcb941b96da211a224391a9e968613cca881e9f096e75f5d32d1bdffa2276733",
+        "beta_cache_audit.json": "f846ca6f1e076257553984f900065bbbb2ec4332466ccd28a6328822fb81b4bc",
+        "precision_control/root_precision_audit.json": "d5bb0717147eb776f6e75d70e76a337f7045fb06b49e78b2cd46221e3cd71b51"
+      },
+      "limitations": "Cached projections are not independent package scores. Original numerical, gain, paired, protected-task and ineligible-GFP gates remain mandatory. One source lineage, repeated tune use, CPU/GPU numerics and homology dependence remain limitations."
+    },
+    "readouts": {
+      "beta_lactamase": {
+        "policy": {
+          "recorded_utc": "2026-11-01T15:48:57.261611+00:00",
+          "source": "/workspace/research/21m_4h_fold_mask_fixed_readouts/precision_control/package",
+          "features": "/workspace/research/21m_4h_fold_mask_fixed_readouts/beta_features",
+          "hypothesis": "Fit-label rank targets may improve beta Spearman reward with the same 64-unit residual capacity and training budget as the raw-label readout.",
+          "targets": "Average ranks of fit labels, including ties, mapped to (rank-.5)/N. Tune labels only evaluate the fixed final readout; no strength/checkpoint selection.",
+          "recipe": "One fixed ridge.0001, then one64-wide residual; seed200,14000steps,batch128,AdamW.001,decay.01,dropout.1,feature noise.02,normalized squared residual loss,clip1; evaluate final14000 only, no intermediate selection. Hyperparameters fixed before extraction.",
+          "adoption": "At least.02 gain over the independently replayed complete parent and positive1000-replicate paired row interval. No threshold override.",
+          "publication": "No assembly or publication in this study; any accepted result needs manual assessment and full independent package replay."
+        },
+        "result": {
+          "baseline": 0.7403514772537801,
+          "rank_ridge_best": 0.7872673810383397,
+          "best": 0.8456129035533306,
+          "gain": 0.10526142629955049,
+          "best_step": 14000,
+          "paired_row_gain_95pct": [
+            0.0726677734073657,
+            0.14099973147734582
+          ],
+          "adopt_threshold": 0.02,
+          "adopt": true,
+          "ridge_trials": [
+            {
+              "relative_ridge": 0.0001,
+              "reward": 0.7872673810383397
+            }
+          ],
+          "residual_history": [
+            {
+              "step": 14000,
+              "reward": 0.8456129035533306
+            }
+          ],
+          "source_model_sha256": "5a59a5ace990511aa072a7818f5d160907e4fcaa8833b2ec86a4c9a1b3ca0367",
+          "unchanged_feature_tensors": 222,
+          "output_scale": "Fit-label rank quantiles, not original physical units.",
+          "uncertainty": "Paired row interval excludes model-selection and homology dependence."
+        },
+        "export": {
+          "status": "passed",
+          "stored_fit_reward": 0.8456129035533306,
+          "reconstructed_fit_reward": 0.8456129035533306,
+          "exported_reward": 0.8456129035533306,
+          "score_difference": 0.0,
+          "all_saved_head_tensors_exact": true,
+          "normalization_exact": true,
+          "residual_max_absolute_error": 0.0,
+          "base_max_absolute_difference": 6.4373016357421875e-06,
+          "base_numpy_max_error_vs_fp64": 6.126665955008548e-06,
+          "base_torch_max_error_vs_fp64": 4.141718488170909e-06,
+          "float32_unit_roundoff": 5.960464477539063e-08,
+          "rounding_operations_bound": 770,
+          "maximum_base_rounding_bound": 0.005847996316499563,
+          "baseline": 0.7403514772537801,
+          "exported_gain": 0.10526142629955049,
+          "exported_paired_row_gain_95pct": [
+            0.0726677734073657,
+            0.14099973147734582
+          ],
+          "adoption_threshold": 0.02,
+          "scope": "Exact saved weights and fit-path reproduction; arithmetic rounding checks; unchanged gain and 1000-paired-interval gates reapplied to exported outputs. No fitting or checkpoint selection. Independent all-task replay remains required."
+        }
+      },
+      "fluorescence": {
+        "policy": {
+          "reference": "Per-task modal fit length and per-position modal fit residue, deterministic lexical ties, no labels/tune used. Reference encoder is the same frozen jointly learned encoder.",
+          "features": "Current penultimate feature and log1p mutation count in both arms. Intervention also gets sum of input-minus-reference shared-encoder residue vectors at substitutions. Only equal-length sequences are eligible; all other predictions preserved exactly.",
+          "targets": "Original fit fluorescence labels minus unchanged native FP32 GFPprediction; residualandfinaloutputremaininoriginalunits. No empirical-rank target, affine target fit or target sweep.",
+          "tasks": [
+            "fluorescence"
+          ],
+          "target_scale": "original_label_units",
+          "device": "cpu",
+          "numerical_implementation": "Use the already validated FP64 accumulation of the small correction over FP32 features/coefficients in fitting and runtime; cast correction to FP32 before adding unchanged parent prediction. No quality gate changes.",
+          "replay_validation": {
+            "registered_utc": "2026-11-01T15:38:09.324940+00:00",
+            "script": "audit_fold_mask_fixed_replay.py",
+            "sha256": "ab30f4dd5c60258588c83641faabc8f6a816411f6a3588fc1b14a1d8c6e13e55",
+            "policy": "Independent full11replay; all unchanged task files exact versus original native; all ineligible GFP values exact. Reapply original.01 gain and1000paired intervals versus actualFP32parent/countcontrol, extra.01original-nativefloor,2e-5 cached/replay tolerance. No exceptions."
+          },
+          "ridge_strengths": [
+            0.1
+          ],
+          "fit": "One fixed.1 ridge in each arm; seed802, exact minimizer. No ridge grid. Only unchanged fallback allowed.",
+          "adoption": "Gain>=.01 versus FP32 native AND original native, positive1000 paired intervals versus parent and count control; same2e-5 cached/replay tolerance. No exception to a failed floor.",
+          "publication": "One independently loaded cold check, exact protected weights/predictions, complete offline eleven-task replay, source/provenance/count audit, all-task paired row/SSchain/PPIcomponent comparison to native and live, GFP strata review if changed. Publish only with stronger complete evidence and atomic publisher.",
+          "limitations": "Adaptive research from the visible1500s masked trajectory and earlier fixed transfers; one own4hancestor/seed, repeated tune selection, CPU/GPU numerics and unmodeled homology dependence. Source run includes an infrastructure restart and at least84.945discarded active seconds; record recovery in final provenance. No hidden-test improvement or full independent replication is claimed.",
+          "resources": "OneH100 remains serialized for training/replay. All beta and GFP feature extraction and fitting CPU-only under the shared readout lock, at most8threads plus4trainerthreads. Beta uses existing generic FP32/TF32off CPU extractor,8192token batches; this differs from the publication\u2019s CUDA beta cache, so the original2e-5 numerical replay gate remains mandatory, with no rescue/tolerance change. Precision-control and candidate fullreplays use existing paused trainer boundaries during already registered later trainers, initially foldweight41. Preserve their training budgets and prior45queue order. No additional GPU training stage.",
+          "scope": "Fixed raw-unit GFP correction and count control only; original CPU extraction; no additional selections.",
+          "implementation_hashes": {
+            "extract_mutation_features.py": "5e97b7931d9b6ccd094ad4c287fe6f9acea0aca8dd6e20b07ec52b99c6c36910",
+            "fit_raw_mutation_readouts.py": "39eae0d420b8165ed9d05006cd2c1f165f1b777dbfbd16f337375e1fbd5eea68",
+            "mutation_features.py": "6dad5f3faccc89de37b9e67fcbc63735f629ca9118e0fb6ef2107e1d106eceac",
+            "mutation_readout_runtime.py": "c17c51019f011b00539c6686dd46bc50cda2e426292972e339045bdca9d809b4"
+          },
+          "recorded_utc": "2026-11-01T15:42:39.100936+00:00",
+          "source": "/run-state/workspace/research/21m_4h_fold_mask_fixed_readouts/precision_control/package",
+          "model_sha256": "5a59a5ace990511aa072a7818f5d160907e4fcaa8833b2ec86a4c9a1b3ca0367",
+          "parent_completion_marker": "/workspace/research/21m_4h_fold_mask_fixed_readouts/precision_control/root_precision_audit.json",
+          "central_policy": "/workspace/research/21m_4h_fold_mask_fixed_readouts/hypothesis.json",
+          "central_policy_sha256": "51282e50721a05daed94610032ab928c16900d218986395f373f1812bd1be751"
+        },
+        "result": {
+          "baseline": 0.79199891808651,
+          "official_baseline": 0.7919983783106186,
+          "fit_baseline": 0.8015985276579471,
+          "variants": {
+            "count_control": {
+              "reward": 0.7941345130236414,
+              "ridge": 0.1,
+              "trainable_coefficients": 386,
+              "torch_reproduction_reward": 0.7941345130236414,
+              "fit_reward": 0.8031659726615605,
+              "initialization_seed": 802,
+              "optimizer": "Exact ridge minimizer",
+              "history": [
+                {
+                  "ridge": 0.1,
+                  "reward": 0.7941345130236414
+                }
+              ],
+              "correction_accumulation": "Float64 over float32 features/coefficients, cast correction to float32 before adding base."
+            },
+            "mutation_delta": {
+              "reward": 0.80349773201962,
+              "ridge": 0.1,
+              "trainable_coefficients": 770,
+              "torch_reproduction_reward": 0.80349773201962,
+              "fit_reward": 0.8118414408701651,
+              "initialization_seed": 802,
+              "optimizer": "Exact ridge minimizer",
+              "history": [
+                {
+                  "ridge": 0.1,
+                  "reward": 0.80349773201962
+                }
+              ],
+              "correction_accumulation": "Float64 over float32 features/coefficients, cast correction to float32 before adding base."
+            }
+          },
+          "gain": 0.01149881393311003,
+          "gain_over_count_control": 0.009363218995978695,
+          "paired_row_gain_95pct": [
+            0.008131903291103443,
+            0.014888659571684124
+          ],
+          "paired_row_gain_over_control_95pct": [
+            0.006516736807961018,
+            0.012521238826549685
+          ],
+          "adopt": true,
+          "unchanged_different_length_tune_rows": 127,
+          "fit_eligible_rows": 20963,
+          "tune_eligible_rows": 5235,
+          "uncertainty": "Intervals exclude model-selection and homology dependence."
+        },
+        "reference": {
+          "sequence": "SKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLKFICTTGKLPVPWPTLVTTLSYGVQCFSRYPDHMKQHDFFKSAMPEGYVQERTIFFKDDGNYKTRAEVKFEGDTLVNRIELKGIDFKEDGNILGHKLEYNYNSHNVYIMADKQKNGIKVNFKIRHNIEDGSVQLADHYQQNTPIGDGPVLLPDNHYLSTQSALSKDPNEKRDHMVLLEFVTAAGITHGMDELYK",
+          "fit_rows": 21446,
+          "modal_length": 237,
+          "modal_length_rows": 20963,
+          "ties": "Lexicographically smallest residue"
+        }
+      }
+    }
+  },
+  "mutation_readouts": {
+    "parent": "/workspace/research/21m_4h_fold_mask_fixed_readouts/precision_control/package",
+    "parent_model_sha256": "5a59a5ace990511aa072a7818f5d160907e4fcaa8833b2ec86a4c9a1b3ca0367",
+    "policy": {
+      "reference": "Per-task modal fit length and per-position modal fit residue, deterministic lexical ties, no labels/tune used. Reference encoder is the same frozen jointly learned encoder.",
+      "features": "Current penultimate feature and log1p mutation count in both arms. Intervention also gets sum of input-minus-reference shared-encoder residue vectors at substitutions. Only equal-length sequences are eligible; all other predictions preserved exactly.",
+      "targets": "Original fit fluorescence labels minus unchanged native FP32 GFPprediction; residualandfinaloutputremaininoriginalunits. No empirical-rank target, affine target fit or target sweep.",
+      "tasks": [
+        "fluorescence"
+      ],
+      "target_scale": "original_label_units",
+      "device": "cpu",
+      "numerical_implementation": "Use the already validated FP64 accumulation of the small correction over FP32 features/coefficients in fitting and runtime; cast correction to FP32 before adding unchanged parent prediction. No quality gate changes.",
+      "replay_validation": {
+        "registered_utc": "2026-11-01T15:38:09.324940+00:00",
+        "script": "audit_fold_mask_fixed_replay.py",
+        "sha256": "ab30f4dd5c60258588c83641faabc8f6a816411f6a3588fc1b14a1d8c6e13e55",
+        "policy": "Independent full11replay; all unchanged task files exact versus original native; all ineligible GFP values exact. Reapply original.01 gain and1000paired intervals versus actualFP32parent/countcontrol, extra.01original-nativefloor,2e-5 cached/replay tolerance. No exceptions."
+      },
+      "ridge_strengths": [
+        0.1
+      ],
+      "fit": "One fixed.1 ridge in each arm; seed802, exact minimizer. No ridge grid. Only unchanged fallback allowed.",
+      "adoption": "Gain>=.01 versus FP32 native AND original native, positive1000 paired intervals versus parent and count control; same2e-5 cached/replay tolerance. No exception to a failed floor.",
+      "publication": "One independently loaded cold check, exact protected weights/predictions, complete offline eleven-task replay, source/provenance/count audit, all-task paired row/SSchain/PPIcomponent comparison to native and live, GFP strata review if changed. Publish only with stronger complete evidence and atomic publisher.",
+      "limitations": "Adaptive research from the visible1500s masked trajectory and earlier fixed transfers; one own4hancestor/seed, repeated tune selection, CPU/GPU numerics and unmodeled homology dependence. Source run includes an infrastructure restart and at least84.945discarded active seconds; record recovery in final provenance. No hidden-test improvement or full independent replication is claimed.",
+      "resources": "OneH100 remains serialized for training/replay. All beta and GFP feature extraction and fitting CPU-only under the shared readout lock, at most8threads plus4trainerthreads. Beta uses existing generic FP32/TF32off CPU extractor,8192token batches; this differs from the publication\u2019s CUDA beta cache, so the original2e-5 numerical replay gate remains mandatory, with no rescue/tolerance change. Precision-control and candidate fullreplays use existing paused trainer boundaries during already registered later trainers, initially foldweight41. Preserve their training budgets and prior45queue order. No additional GPU training stage.",
+      "scope": "Fixed raw-unit GFP correction and count control only; original CPU extraction; no additional selections.",
+      "implementation_hashes": {
+        "extract_mutation_features.py": "5e97b7931d9b6ccd094ad4c287fe6f9acea0aca8dd6e20b07ec52b99c6c36910",
+        "fit_raw_mutation_readouts.py": "39eae0d420b8165ed9d05006cd2c1f165f1b777dbfbd16f337375e1fbd5eea68",
+        "mutation_features.py": "6dad5f3faccc89de37b9e67fcbc63735f629ca9118e0fb6ef2107e1d106eceac",
+        "mutation_readout_runtime.py": "c17c51019f011b00539c6686dd46bc50cda2e426292972e339045bdca9d809b4"
+      },
+      "recorded_utc": "2026-11-01T15:42:39.100936+00:00",
+      "source": "/run-state/workspace/research/21m_4h_fold_mask_fixed_readouts/precision_control/package",
+      "model_sha256": "5a59a5ace990511aa072a7818f5d160907e4fcaa8833b2ec86a4c9a1b3ca0367",
+      "parent_completion_marker": "/workspace/research/21m_4h_fold_mask_fixed_readouts/precision_control/root_precision_audit.json",
+      "central_policy": "/workspace/research/21m_4h_fold_mask_fixed_readouts/hypothesis.json",
+      "central_policy_sha256": "51282e50721a05daed94610032ab928c16900d218986395f373f1812bd1be751",
+      "lineage_summary": "Own four-hour UniRef MLM,9,127,835,498tokens;1800active seconds of all11fit supervision. Selectedema step53294 at1800.0039906070306active seconds. Training-only fold masking regularizes the shared encoder; human/solubility stop shared gradients after300seconds, while all nine other tasks continue. Fixed beta/GFP hyperparameters transferred without a new grid; CPUFP32 feature extraction. Native training resumed its own complete1200-second state after a runtime restart; at least84.945additional active seconds were discarded, as documented in training history."
+    },
+    "results": {
+      "recorded_utc": "2026-11-01T15:59:33.993546+00:00",
+      "model_sha256": "5a59a5ace990511aa072a7818f5d160907e4fcaa8833b2ec86a4c9a1b3ca0367",
+      "tasks": {
+        "fluorescence": {
+          "baseline": 0.79199891808651,
+          "official_baseline": 0.7919983783106186,
+          "fit_baseline": 0.8015985276579471,
+          "variants": {
+            "count_control": {
+              "reward": 0.7941345130236414,
+              "ridge": 0.1,
+              "trainable_coefficients": 386,
+              "torch_reproduction_reward": 0.7941345130236414,
+              "fit_reward": 0.8031659726615605,
+              "initialization_seed": 802,
+              "optimizer": "Exact ridge minimizer",
+              "history": [
+                {
+                  "ridge": 0.1,
+                  "reward": 0.7941345130236414
+                }
+              ],
+              "correction_accumulation": "Float64 over float32 features/coefficients, cast correction to float32 before adding base."
+            },
+            "mutation_delta": {
+              "reward": 0.80349773201962,
+              "ridge": 0.1,
+              "trainable_coefficients": 770,
+              "torch_reproduction_reward": 0.80349773201962,
+              "fit_reward": 0.8118414408701651,
+              "initialization_seed": 802,
+              "optimizer": "Exact ridge minimizer",
+              "history": [
+                {
+                  "ridge": 0.1,
+                  "reward": 0.80349773201962
+                }
+              ],
+              "correction_accumulation": "Float64 over float32 features/coefficients, cast correction to float32 before adding base."
+            }
+          },
+          "gain": 0.01149881393311003,
+          "gain_over_count_control": 0.009363218995978695,
+          "paired_row_gain_95pct": [
+            0.008131903291103443,
+            0.014888659571684124
+          ],
+          "paired_row_gain_over_control_95pct": [
+            0.006516736807961018,
+            0.012521238826549685
+          ],
+          "adopt": true,
+          "unchanged_different_length_tune_rows": 127,
+          "fit_eligible_rows": 20963,
+          "tune_eligible_rows": 5235,
+          "uncertainty": "Intervals exclude model-selection and homology dependence."
+        }
+      },
+      "next": "No automatic assembly/publication. Any accepted head requires independent complete eleven-task replay and preservation review."
+    },
+    "selected_tasks": [
+      "fluorescence"
+    ],
+    "subgroup_diagnostics": {
+      "recorded_utc": "2026-11-01T15:59:34.514940+00:00",
+      "scope": "Descriptive subgroups only; no fitting or alteration of predeclared acceptance rules.",
+      "limitations": "Conditional rank correlations do not decompose global reward. Global rank-error contributions do; intervals exclude selection/homology uncertainty.",
+      "tasks": {
+        "fluorescence": {
+          "fit_label_thresholds": [
+            3.256847381591797,
+            3.552252769470215,
+            3.677128791809082
+          ],
+          "groups": {
+            "fit_label_quartile_1": {
+              "rows": 1361,
+              "conditional_spearman": {
+                "base": 0.5902464398299688,
+                "count_control": 0.6033304407785361,
+                "mutation": 0.6122380258910823
+              },
+              "global_rho_change_contribution": 0.0017785806549375153,
+              "global_rho_change_vs_count_control_contribution": 0.0012298227412430877,
+              "paired_conditional_gain_95pct": [
+                0.002724715662720134,
+                0.04179508673813816
+              ]
+            },
+            "fit_label_quartile_2": {
+              "rows": 1297,
+              "conditional_spearman": {
+                "base": 0.44151365270705784,
+                "count_control": 0.4444034547831131,
+                "mutation": 0.45497768987920806
+              },
+              "global_rho_change_contribution": 0.003723691370474727,
+              "global_rho_change_vs_count_control_contribution": 0.0028860231800414394,
+              "paired_conditional_gain_95pct": [
+                0.0021970537145898847,
+                0.025211052203393963
+              ]
+            },
+            "fit_label_quartile_3": {
+              "rows": 1384,
+              "conditional_spearman": {
+                "base": 0.2076473309514199,
+                "count_control": 0.21814163150270519,
+                "mutation": 0.24357852995482093
+              },
+              "global_rho_change_contribution": 0.004367255730781997,
+              "global_rho_change_vs_count_control_contribution": 0.0036651242172731587,
+              "paired_conditional_gain_95pct": [
+                0.016850443150831434,
+                0.05412158336502228
+              ]
+            },
+            "fit_label_quartile_4": {
+              "rows": 1320,
+              "conditional_spearman": {
+                "base": 0.23905578458561658,
+                "count_control": 0.22048107498094563,
+                "mutation": 0.23032286274322997
+              },
+              "global_rho_change_contribution": 0.0016292861769158396,
+              "global_rho_change_vs_count_control_contribution": 0.0015822488574209527,
+              "paired_conditional_gain_95pct": [
+                -0.030466541609212597,
+                0.014023872146553366
+              ]
+            },
+            "mutations_0_or_1": {
+              "rows": 233,
+              "conditional_spearman": {
+                "base": 0.7685420226748939,
+                "count_control": 0.7649645250495584,
+                "mutation": 0.7815950032707549
+              },
+              "global_rho_change_contribution": -0.0008731330512296009,
+              "global_rho_change_vs_count_control_contribution": 0.0018896590896428577,
+              "paired_conditional_gain_95pct": [
+                -0.013331535928997298,
+                0.04014347099125209
+              ]
+            },
+            "mutations_2": {
+              "rows": 2528,
+              "conditional_spearman": {
+                "base": 0.698027263188347,
+                "count_control": 0.7002796614488055,
+                "mutation": 0.7148476367428673
+              },
+              "global_rho_change_contribution": 0.0073290293094369,
+              "global_rho_change_vs_count_control_contribution": 0.006153299603157906,
+              "paired_conditional_gain_95pct": [
+                0.010889253474149072,
+                0.024111682991964964
+              ]
+            },
+            "mutations_3": {
+              "rows": 2474,
+              "conditional_spearman": {
+                "base": 0.840244719944968,
+                "count_control": 0.841712918607403,
+                "mutation": 0.8479607993579165
+              },
+              "global_rho_change_contribution": 0.004900652803623974,
+              "global_rho_change_vs_count_control_contribution": 0.001326595757979534,
+              "paired_conditional_gain_95pct": [
+                0.0034951710741375763,
+                0.011986767110728654
+              ]
+            },
+            "mutations_4plus": {
+              "rows": 0,
+              "conditional_spearman": {},
+              "global_rho_change_contribution": 0.0,
+              "global_rho_change_vs_count_control_contribution": 0.0
+            },
+            "different_length": {
+              "rows": 127,
+              "conditional_spearman": {
+                "base": 0.7918415776085357,
+                "count_control": 0.7918415776085357,
+                "mutation": 0.7918415776085357
+              },
+              "global_rho_change_contribution": 0.00014226487127880718,
+              "global_rho_change_vs_count_control_contribution": -6.335454801658159e-06,
+              "paired_conditional_gain_95pct": [
+                0.0,
+                0.0
+              ]
+            },
+            "at_least_one_unseen_substitution": {
+              "rows": 55,
+              "conditional_spearman": {
+                "base": 0.815007215007215,
+                "count_control": 0.8041125541125541,
+                "mutation": 0.8002886002886003
+              },
+              "global_rho_change_contribution": -0.00026462047301932724,
+              "global_rho_change_vs_count_control_contribution": -6.089922896184901e-05,
+              "paired_conditional_gain_95pct": [
+                -0.05587127167534611,
+                0.031038011641875505
+              ]
+            },
+            "only_fit_seen_substitutions": {
+              "rows": 5180,
+              "conditional_spearman": {
+                "base": 0.786671754218523,
+                "count_control": 0.7891191165784951,
+                "mutation": 0.7991834832073826
+              },
+              "global_rho_change_contribution": 0.011621169534850602,
+              "global_rho_change_vs_count_control_contribution": 0.009430453679742146,
+              "paired_conditional_gain_95pct": [
+                0.008814417425801028,
+                0.01620050339326211
+              ]
+            }
+          },
+          "global_rho_change": 0.01149881393311003
+        }
+      }
+    }
+  },
+  "independent_replay": {
+    "reward": 0.7268424717655811,
+    "tasks": {
+      "beta_lactamase": {
+        "invalid_reason": null,
+        "reward": 0.8456191337166881,
+        "spearman": 0.8456191337166881
+      },
+      "binary_localization": {
+        "accuracy": 0.9624060150375939,
+        "balanced_accuracy": 0.9609346188293557,
+        "invalid_reason": null,
+        "macro_f1": 0.9610378354391542,
+        "reward": 0.9624060150375939
+      },
+      "fluorescence": {
+        "invalid_reason": null,
+        "reward": 0.8034966552237341,
+        "spearman": 0.8034966552237341
+      },
+      "fold": {
+        "accuracy": 0.29483695652173914,
+        "balanced_accuracy": 0.19703615098351948,
+        "invalid_reason": null,
+        "macro_f1": 0.04734856523385019,
+        "reward": 0.29483695652173914
+      },
+      "human_ppi": {
+        "accuracy": 0.8507936507936508,
+        "balanced_accuracy": 0.8503401360544218,
+        "invalid_reason": null,
+        "macro_f1": 0.8501897293195042,
+        "reward": 0.8507936507936508
+      },
+      "ppi_affinity": {
+        "invalid_reason": null,
+        "reward": 0.5624258031382015,
+        "rmse": 1.7743729796494956,
+        "spearman": 0.7498163821577267
+      },
+      "secondary_structure": {
+        "accuracy": 0.7503389109243336,
+        "invalid_reason": null,
+        "reward": 0.7503389109243336,
+        "valid_residues": 556931
+      },
+      "solubility": {
+        "accuracy": 0.7538173437049841,
+        "balanced_accuracy": 0.7404053807847767,
+        "invalid_reason": null,
+        "macro_f1": 0.74317716963567,
+        "reward": 0.7538173437049841
+      },
+      "stability": {
+        "invalid_reason": null,
+        "reward": 0.7802091606607918,
+        "spearman": 0.7802091606607918
+      },
+      "subcellular_localization": {
+        "accuracy": 0.8018498754891498,
+        "balanced_accuracy": 0.6917805302447911,
+        "invalid_reason": null,
+        "macro_f1": 0.7068148859422272,
+        "reward": 0.8018498754891498
+      },
+      "yeast_ppi": {
+        "accuracy": 0.5894736842105263,
+        "balanced_accuracy": 0.570054945054945,
+        "invalid_reason": null,
+        "macro_f1": 0.5704347826086956,
+        "reward": 0.5894736842105263
+      }
+    }
+  },
+  "root_publication_decision": {
+    "utc": "2026-11-01T16:04:17.164625+00:00",
+    "decision": "Selected for publication following root evidence review",
+    "reward": 0.7268424717655811,
+    "interpretation": [
+      "Select this complete independently replayed candidate for atomic publication. Mean0.7268424717655811 exceeds the current0.7265086366205981 by0.00033383514498308564,withseven task gains andfour declines. The decision uses the higher measuredmean and broader point improvement,including structure, rather than claiming a Pareto or statistically certain aggregate gain. No acceptance gate changed and no additional search was performed on thisbody.",
+      "Beta improves0.0171266764 to0.8456191337,withpairedrowinterval[-0.0099489251,0.0449645934]includingzero. Binarylocalization improves0.0069404280(30corrected,18newerrors),subcellular0.0049804340(164corrected,150newerrors),stability0.0004140607 andhumanPPI0.0031746032(13corrected,12newerrors);theirrowintervalsincludezero,andhuman'scomponentintervalalsoincludeszero. These are favorable pointchanges,not independently established improvements.",
+      "Fold improves0.0217391304 to217/736correct versus201,with54corrected and38newerrors. Its pairedinterval[-0.0027173913,0.0462296196]includeszero. Allfourcoarsefit-supportgroups havepositiveaccuracychanges, butrare1\u20135exampleclasses remainonly3/92versus2/92. The alreadyclosed nativefoldaudit shows no singleton top1/top5/top10 success andmixed rare-ranking behavior; thispackage doesnot solve rarefoldgeneralization. Reuse that existing audit becausefoldpredictionsareexact; no extra fold bootstrap or inference.",
+      "Secondarystructure improves0.0041387533 to0.7503389109 acrossall556931tuner esidues,withpositivechaininterval[0.0026131548,0.0054343886]. This is the clearest favorable pairedsignal. Fullchainwindowing remains active; global andpairchains stilltruncateat1022residues.",
+      "Fourtasks decline: fluorescence-0.0050111459,solubility-0.0040334198,affinity-0.0036920706 andyeastPPI-0.0421052632. Allfourrowintervalsincludezero;yeast andaffinitycomponentintervalsalsoincludezero. Yeast hasfour fewer correctexamples(10corrected,14newerrors)outof95,an important pointregression despitewideuncertainty. GFP'srowinterval[-0.0103088511,0.0003718571]nearlyexcludeszero andmustnotbe describedas unchanged. No taskloss is repaired by mixing task-specific backbone states.",
+      "Aggregatepairedrow/chain95interval[-0.0087820636,0.0102981347]andcomponent-awareinterval[-0.0106261114,0.0071172432]bothincludezero. Onepretrainedancestor, adaptive repeatedtuneselection, omittedhomologyandcross-task dependence limit inference. HumanPPIhas284/315rowsinonecomponentof27,so evencomponentresamplingislimited. No hidden-test gain or independentreplication isclaimed.",
+      "Against its originalnative parent,thefixedreadouts improvebeta0.1034728573 andGFP0.0116780613,bothwithpositivepairedintervals;theotherninefullpredictionfilesareexact. Parentmeangain0.0104682653hasinterval[0.0075594704,0.0137127666]. The candidatepreserves282originaltensors andall21,306,271encoderparameters. Total25,065,830learnedparametersplus769normalizationvalues;one jointlylearnedencoder suppliesalltasks andthefit-onlyGFPconsensus.",
+      "Originalindependentreplaygates allPASS: betaactual-minus-cache+0.000006230163357434648,GFP-0.0000010767958859014826,bothwithin2e-5;requiredmagnitudegains andallpairedparent/countintervalsPASS. Ninefulltaskfilesand127other-lengthGFPoutputsexact. Standalone111-rowcoldcheckpassed. No numericalrepair,tolerancewaiver,alternatebody,readoutgrid or interpolation wasused.",
+      "GFPsubgroupssupportsomebenefitrelative tothisparent: positiveconditionalintervalsinfit-labelquartiles1\u20133 andtwo/three-substitutiongroups. Quartile4 and55unseen-substitutionexamples have negativepointchangeswithintervalsincludingzero. No4+alignedtuneexamplesareavailable. Allfitting,extraction,strata,cold,replay,sourcepreservation andcomparisonsarenowclosed;do notrepeat.",
+      "Publicationuses the suppliedatomicpublishercompatibilitywrapper afterfinalmetadatareview. Preserve thepriorlivepackageandselectionrecordunderresearch. Afterpublicationverifyallsource/livefilesandmarkeronce,updatethecurrentpublicationindex,andcontinueactivefoldweightthenregistered42\u201346. Retainruntime-recoveryprovenance: own1200secondfullstatecontinuation andatleast84.945discardedactivesecondsoutsideitsretained1800secondtrajectory."
+    ]
+  }
+}