Add DENIM evidence v3 protocols

Frozen data ablation, leave-one-family-out and stratified 16/64-trajectory adaptation evidence; no test leakage.

ARTICLE_EVIDENCE_V3_SHA256SUMS

@@ -0,0 +1,18 @@
+957eeda4af6adf7fa0c7388d46d74cba04cb1fb1654d9cb1151e0b8b5925ddc9  CONDITIONAL_V2_SHA256SUMS
+f99d140652e1c98ce39951be5e73690e58a66750336de8d16bf252c8777ace84  README.md
+54b5758b3c3ea75e3f11b701209b7a3f8bd73691f7b48bd701e4112c94dc68db  article_evidence_v3/README.md
+32d7ed305c6a32587be7b97448970388856e52482712c632409255e74efccb10  article_evidence_v3/evidence_metrics.json
+b86cdeb5e109beef399332d84fe174ff59e83f03f251451dd4d7beb359573260  article_evidence_v3/protocol.json
+53e80543f77326c9051dca766e045ab3390d96e5cb684cbf221df164c9ceb9d6  article_evidence_v3/run_t2_denim_evidence_v3.py
+84173c7a9e5fce83601b40d178f37cdb7030b46117839fb3b3fec4e62a269d7e  conditional_v2/README.md
+28725b2db1a303c86d0d9c9268ceea2bf6faaddfe2df167457fd3bd8e8e794de  conditional_v2/SHA256SUMS
+f2ba3f8d4af49eaec284e6ab776cc65eb80ac81139b800889765881dc14bfc0b  conditional_v2/artifacts/t2_denim_conditional_v2/model_metrics.json
+85c96f7d4954839fe99b8f741e6eaab19b0b6b47b87e19d3c3172853d0f3dba5  conditional_v2/artifacts/t2_denim_conditional_v2/physical_runtime_validation.json
+4c9fa10a3235892714d50d92e174663dbbac9e4162c964d739ffee28721b1873  conditional_v2/artifacts/t2_denim_conditional_v2/refined_metrics.json
+f7ef24867baa630310c9c391a7d4df3fc048fb501ea240a728a4cb6ade0ff7ae  conditional_v2/docs/T2_CONDITIONAL_DENIM_V2.md
+08f70f7cb3bebd63859cd0b2626775dda85cc738b8dad2fc77c5d8a80033a7b3  conditional_v2/src/infer_t2_denim_conditional_v2.py
+2d6b4b705f81250a447c7e59ded54815204cc923ccfbb4d87488844c47b10691  conditional_v2/src/refine_t2_denim_conditional_v2.py
+85b288ee63a862dc9698e501c9f691ec4b8ab772dde7f20109ab9222a1c6d874  conditional_v2/src/t2_denim_conditional.py
+40bd1d10cf1ea285bb7b88990ace757ba128e88e4dcf84d208dca8c2df1b5c4c  conditional_v2/src/train_t2_denim_conditional_v2.py
+47f5c2635a6702926b26fee2b28aacc94e333686f9a4bf672f6b58a41191a9d1  conditional_v2/src/validate_t2_denim_conditional_v2.py
+a940f94318c4143acbc9a360efdd84370e616bf1b0fd842d8c3c5d505f8eefc8  conditional_v2/tests/test_t2_denim_conditional.py

README.md

@@ -188,3 +188,8 @@ The releases are synthetic, small-strain material-point benchmarks. They do
 not constitute experimental calibration, fatigue-life prediction, damage,
 finite-strain plasticity or a production-certified learned material. Data are
 CC BY 4.0; code files are covered by the included code license.
+
+
+## Evidence protocol v3
+
+See `article_evidence_v3/` for frozen ablation and leave-one-family-out split definitions.

article_evidence_v3/README.md

@@ -0,0 +1,5 @@
+# DENIM evidence protocol v3
+
+This directory adds reproducible split and adaptation protocols to the existing 2,660-trajectory dataset. It does not duplicate or reassign the published trajectories. Training, validation, and test roles remain frozen; test trajectories are evaluation-only.
+
+The evidence matrix covers 80/330 incomplete-family trajectories, transfer from 1,154 declared-physics trajectories, leave-one-family-out zero-shot evaluation, and 16/64-trajectory adaptation.

article_evidence_v3/evidence_metrics.json

@@ -0,0 +1,704 @@
+{
+  "schema": "agentfem.denim-evidence.v3",
+  "frozen_protocol": {
+    "total_trajectories": 2660,
+    "train": 1484,
+    "validation": 407,
+    "test": 769,
+    "family_counts": {
+      "j2_linear_isotropic": {
+        "train": 577,
+        "validation": 171,
+        "test": 268
+      },
+      "chaboche_combined": {
+        "train": 577,
+        "validation": 170,
+        "test": 269
+      },
+      "hidden_three_memory_tabulated_hardening": {
+        "train": 330,
+        "validation": 66,
+        "test": 232
+      }
+    },
+    "selection_rule": "deterministic round-robin across path families; held-family training trajectories only"
+  },
+  "ablation": {
+    "hidden_80_scratch": {
+      "training": {
+        "steps": 1400,
+        "seconds": 18.30456762504764,
+        "best_validation_loss": 2.075452357530594e-05
+      },
+      "evaluation": {
+        "hidden_frozen_test": {
+          "trajectory_count": 32,
+          "rmse_mpa": 0.5097740292549133,
+          "mae_mpa": 0.28684791922569275,
+          "r2": 0.9999802112579346,
+          "state_point_count": 7712
+        },
+        "test_amplitude_ood": {
+          "trajectory_count": 60,
+          "rmse_mpa": 1.3261798620224,
+          "mae_mpa": 0.4734494984149933,
+          "r2": 0.9998748302459717,
+          "state_point_count": 14460
+        },
+        "test_discretization": {
+          "trajectory_count": 20,
+          "rmse_mpa": 0.4494419991970062,
+          "mae_mpa": 0.2459268420934677,
+          "r2": 0.9999846816062927,
+          "state_point_count": 8120
+        },
+        "test_long_horizon": {
+          "trajectory_count": 40,
+          "rmse_mpa": 9.0242338180542,
+          "mae_mpa": 3.596944570541382,
+          "r2": 0.9937672019004822,
+          "state_point_count": 9640
+        },
+        "test_path_ood": {
+          "trajectory_count": 80,
+          "rmse_mpa": 0.548997700214386,
+          "mae_mpa": 0.2917637825012207,
+          "r2": 0.999976396560669,
+          "state_point_count": 19280
+        }
+      }
+    },
+    "hidden_330_scratch": {
+      "training": {
+        "steps": 1400,
+        "seconds": 18.037176416022703,
+        "best_validation_loss": 2.017221959249582e-05,
+        "protocol": "hyperparameter_matched_to_hidden_80_scratch"
+      },
+      "evaluation": {
+        "hidden_frozen_test": {
+          "trajectory_count": 32,
+          "rmse_mpa": 0.5325977206230164,
+          "mae_mpa": 0.2931683659553528,
+          "r2": 0.9999784231185913,
+          "state_point_count": 7712
+        },
+        "test_amplitude_ood": {
+          "trajectory_count": 60,
+          "rmse_mpa": 1.385399341583252,
+          "mae_mpa": 0.4937373101711273,
+          "r2": 0.9998633861541748,
+          "state_point_count": 14460
+        },
+        "test_discretization": {
+          "trajectory_count": 20,
+          "rmse_mpa": 0.46980738639831543,
+          "mae_mpa": 0.25197768211364746,
+          "r2": 0.9999832510948181,
+          "state_point_count": 8120
+        },
+        "test_long_horizon": {
+          "trajectory_count": 40,
+          "rmse_mpa": 9.191946983337402,
+          "mae_mpa": 3.683844566345215,
+          "r2": 0.9935333728790283,
+          "state_point_count": 9640
+        },
+        "test_path_ood": {
+          "trajectory_count": 80,
+          "rmse_mpa": 0.5769456624984741,
+          "mae_mpa": 0.2999087870121002,
+          "r2": 0.9999739527702332,
+          "state_point_count": 19280
+        }
+      }
+    },
+    "known_1154_plus_hidden_80": {
+      "training": {
+        "steps": 1400,
+        "seconds": 39.2154094169382,
+        "best_validation_loss": 8.18048438304686e-05
+      },
+      "evaluation": {
+        "hidden_frozen_test": {
+          "trajectory_count": 32,
+          "rmse_mpa": 1.457561731338501,
+          "mae_mpa": 0.7716607451438904,
+          "r2": 0.9998383522033691,
+          "state_point_count": 7712
+        },
+        "test_amplitude_ood": {
+          "trajectory_count": 60,
+          "rmse_mpa": 3.7529003620147705,
+          "mae_mpa": 1.438835859298706,
+          "r2": 0.9989975690841675,
+          "state_point_count": 14460
+        },
+        "test_discretization": {
+          "trajectory_count": 20,
+          "rmse_mpa": 1.5815093517303467,
+          "mae_mpa": 0.7716865539550781,
+          "r2": 0.9998104572296143,
+          "state_point_count": 8120
+        },
+        "test_long_horizon": {
+          "trajectory_count": 40,
+          "rmse_mpa": 7.19581937789917,
+          "mae_mpa": 2.662142515182495,
+          "r2": 0.9960370063781738,
+          "state_point_count": 9640
+        },
+        "test_path_ood": {
+          "trajectory_count": 80,
+          "rmse_mpa": 1.4498724937438965,
+          "mae_mpa": 0.7752851247787476,
+          "r2": 0.9998355507850647,
+          "state_point_count": 19280
+        }
+      }
+    },
+    "known_1154_plus_hidden_330": {
+      "training": {
+        "steps": 1400,
+        "seconds": 37.73902383283712,
+        "best_validation_loss": 8.374246726816637e-05,
+        "protocol": "hyperparameter_matched_to_known_1154_plus_hidden_80"
+      },
+      "evaluation": {
+        "hidden_frozen_test": {
+          "trajectory_count": 32,
+          "rmse_mpa": 1.5190730094909668,
+          "mae_mpa": 0.8048877716064453,
+          "r2": 0.9998244047164917,
+          "state_point_count": 7712
+        },
+        "test_amplitude_ood": {
+          "trajectory_count": 60,
+          "rmse_mpa": 3.729703903198242,
+          "mae_mpa": 1.4087541103363037,
+          "r2": 0.9990099668502808,
+          "state_point_count": 14460
+        },
+        "test_discretization": {
+          "trajectory_count": 20,
+          "rmse_mpa": 1.656106948852539,
+          "mae_mpa": 0.8177827000617981,
+          "r2": 0.9997920989990234,
+          "state_point_count": 8120
+        },
+        "test_long_horizon": {
+          "trajectory_count": 40,
+          "rmse_mpa": 7.5080790519714355,
+          "mae_mpa": 2.7841875553131104,
+          "r2": 0.9956855773925781,
+          "state_point_count": 9640
+        },
+        "test_path_ood": {
+          "trajectory_count": 80,
+          "rmse_mpa": 1.5116416215896606,
+          "mae_mpa": 0.8086093664169312,
+          "r2": 0.9998212456703186,
+          "state_point_count": 19280
+        }
+      }
+    }
+  },
+  "leave_one_family_out": {
+    "j2_linear_isotropic": {
+      "held_out_family": "j2_linear_isotropic",
+      "base_training": {
+        "steps": 1400,
+        "seconds": 39.81002512504347,
+        "best_validation_loss": 0.0014799873461015522
+      },
+      "zero_shot": {
+        "trajectory_count": 268,
+        "rmse_mpa": 49.748207092285156,
+        "mae_mpa": 30.285388946533203,
+        "r2": 0.8084170818328857,
+        "state_point_count": 57388
+      },
+      "adaptation": {
+        "16": {
+          "training": {
+            "steps": 900,
+            "seconds": 20.13811704190448,
+            "best_validation_loss": 0.0008964448315964546
+          },
+          "selection": "path-family-balanced round-robin",
+          "training_trajectory_ids": [
+            337,
+            1,
+            1264,
+            1200,
+            1009,
+            1072,
+            1136,
+            252,
+            169,
+            85,
+            420,
+            338,
+            2,
+            1265,
+            1202,
+            1010
+          ],
+          "evaluation": {
+            "trajectory_count": 268,
+            "rmse_mpa": 3.864365577697754,
+            "mae_mpa": 1.7006479501724243,
+            "r2": 0.9988440275192261,
+            "state_point_count": 57388
+          }
+        },
+        "64": {
+          "training": {
+            "steps": 1200,
+            "seconds": 33.06111670914106,
+            "best_validation_loss": 0.00058859380078502
+          },
+          "selection": "path-family-balanced round-robin",
+          "training_trajectory_ids": [
+            337,
+            1,
+            1264,
+            1200,
+            1009,
+            1072,
+            1136,
+            252,
+            169,
+            85,
+            420,
+            338,
+            2,
+            1265,
+            1202,
+            1010,
+            1074,
+            1139,
+            253,
+            170,
+            86,
+            421,
+            339,
+            3,
+            1267,
+            1203,
+            1011,
+            1076,
+            1143,
+            257,
+            171,
+            87,
+            422,
+            340,
+            6,
+            1270,
+            1204,
+            1012,
+            1078,
+            1145,
+            259,
+            172,
+            88,
+            424,
+            341,
+            7,
+            1271,
+            1206,
+            1013,
+            1079,
+            1147,
+            261,
+            174,
+            91,
+            426,
+            343,
+            9,
+            1272,
+            1210,
+            1017,
+            1080,
+            1149,
+            262,
+            175
+          ],
+          "evaluation": {
+            "trajectory_count": 268,
+            "rmse_mpa": 3.2635509967803955,
+            "mae_mpa": 1.4564924240112305,
+            "r2": 0.999175488948822,
+            "state_point_count": 57388
+          }
+        }
+      }
+    },
+    "chaboche_combined": {
+      "held_out_family": "chaboche_combined",
+      "base_training": {
+        "steps": 1400,
+        "seconds": 41.68002520897426,
+        "best_validation_loss": 0.004121885751374066
+      },
+      "zero_shot": {
+        "trajectory_count": 269,
+        "rmse_mpa": 30.393577575683594,
+        "mae_mpa": 16.70126724243164,
+        "r2": 0.9328222870826721,
+        "state_point_count": 57629
+      },
+      "adaptation": {
+        "16": {
+          "training": {
+            "steps": 900,
+            "seconds": 21.83759491611272,
+            "best_validation_loss": 0.0010910335695371032
+          },
+          "selection": "path-family-balanced round-robin",
+          "training_trajectory_ids": [
+            841,
+            504,
+            1777,
+            1712,
+            1521,
+            1584,
+            1648,
+            760,
+            672,
+            588,
+            925,
+            842,
+            505,
+            1779,
+            1715,
+            1522
+          ],
+          "evaluation": {
+            "trajectory_count": 269,
+            "rmse_mpa": 20.846805572509766,
+            "mae_mpa": 10.1530122756958,
+            "r2": 0.9683961272239685,
+            "state_point_count": 57629
+          }
+        },
+        "64": {
+          "training": {
+            "steps": 1200,
+            "seconds": 34.28978379210457,
+            "best_validation_loss": 0.00030190416146069765
+          },
+          "selection": "path-family-balanced round-robin",
+          "training_trajectory_ids": [
+            841,
+            504,
+            1777,
+            1712,
+            1521,
+            1584,
+            1648,
+            760,
+            672,
+            588,
+            925,
+            842,
+            505,
+            1779,
+            1715,
+            1522,
+            1585,
+            1650,
+            761,
+            673,
+            589,
+            926,
+            843,
+            506,
+            1781,
+            1716,
+            1525,
+            1586,
+            1653,
+            764,
+            675,
+            590,
+            928,
+            846,
+            508,
+            1783,
+            1717,
+            1528,
+            1588,
+            1654,
+            765,
+            676,
+            591,
+            929,
+            848,
+            509,
+            1784,
+            1720,
+            1529,
+            1589,
+            1655,
+            767,
+            677,
+            592,
+            931,
+            849,
+            511,
+            1785,
+            1722,
+            1531,
+            1591,
+            1656,
+            769,
+            678
+          ],
+          "evaluation": {
+            "trajectory_count": 269,
+            "rmse_mpa": 8.491641998291016,
+            "mae_mpa": 4.113821983337402,
+            "r2": 0.9947562217712402,
+            "state_point_count": 57629
+          }
+        }
+      }
+    },
+    "hidden_three_memory_tabulated_hardening": {
+      "held_out_family": "hidden_three_memory_tabulated_hardening",
+      "base_training": {
+        "steps": 1400,
+        "seconds": 43.397430332843214,
+        "best_validation_loss": 0.0023462075041607022
+      },
+      "zero_shot": {
+        "trajectory_count": 232,
+        "rmse_mpa": 50.57877731323242,
+        "mae_mpa": 30.57018280029297,
+        "r2": 0.806849479675293,
+        "state_point_count": 59212
+      },
+      "adaptation": {
+        "16": {
+          "training": {
+            "steps": 900,
+            "seconds": 23.648075999924913,
+            "best_validation_loss": 0.00018563247112979298
+          },
+          "selection": "path-family-balanced round-robin",
+          "training_trajectory_ids": [
+            2096,
+            2080,
+            2032,
+            2048,
+            2064,
+            2097,
+            2081,
+            2033,
+            2049,
+            2065,
+            2098,
+            2082,
+            2034,
+            2050,
+            2066,
+            2099
+          ],
+          "evaluation": {
+            "trajectory_count": 232,
+            "rmse_mpa": 3.847371816635132,
+            "mae_mpa": 1.2943682670593262,
+            "r2": 0.9988824129104614,
+            "state_point_count": 59212
+          }
+        },
+        "64": {
+          "training": {
+            "steps": 1200,
+            "seconds": 34.54037875006907,
+            "best_validation_loss": 8.566799533582525e-05
+          },
+          "selection": "path-family-balanced round-robin",
+          "training_trajectory_ids": [
+            2096,
+            2080,
+            2032,
+            2048,
+            2064,
+            2097,
+            2081,
+            2033,
+            2049,
+            2065,
+            2098,
+            2082,
+            2034,
+            2050,
+            2066,
+            2099,
+            2083,
+            2035,
+            2051,
+            2067,
+            2100,
+            2084,
+            2036,
+            2052,
+            2068,
+            2101,
+            2085,
+            2037,
+            2053,
+            2069,
+            2102,
+            2086,
+            2038,
+            2054,
+            2070,
+            2103,
+            2087,
+            2039,
+            2055,
+            2071,
+            2104,
+            2088,
+            2040,
+            2056,
+            2072,
+            2105,
+            2089,
+            2041,
+            2057,
+            2073,
+            2106,
+            2090,
+            2042,
+            2058,
+            2074,
+            2107,
+            2091,
+            2043,
+            2059,
+            2075,
+            2108,
+            2092,
+            2044,
+            2060
+          ],
+          "evaluation": {
+            "trajectory_count": 232,
+            "rmse_mpa": 3.7123770713806152,
+            "mae_mpa": 1.1961064338684082,
+            "r2": 0.9989594221115112,
+            "state_point_count": 59212
+          }
+        }
+      }
+    }
+  },
+  "artifacts": {
+    "hidden_80_scratch": {
+      "path": "models/t2_denim_evidence_v3/hidden_80_scratch.safetensors",
+      "sha256": "b438ecbc97e5c63bf39a020a81858da6eabd13be00bc51336df544843558aba1"
+    },
+    "known_1154_plus_hidden_80": {
+      "path": "models/t2_denim_evidence_v3/known_1154_plus_hidden_80.safetensors",
+      "sha256": "cc21b64c3c9d2321cfabf261f1728e7811f114efbbdd4904ded124ee0b16abf2"
+    },
+    "lofo_j2_linear_isotropic_zero": {
+      "path": "models/t2_denim_evidence_v3/lofo_j2_linear_isotropic_zero.safetensors",
+      "sha256": "e5a2e66708b054002c0a13949d53022caa492c3d0d9b7d0400ce00a73e0d3d60"
+    },
+    "lofo_j2_linear_isotropic_16": {
+      "path": "models/t2_denim_evidence_v3/lofo_j2_linear_isotropic_16.safetensors",
+      "sha256": "2a597c038c7f1c794dc2ac0afceb06292786a9c14f694ea566f5eabaa9d4d491"
+    },
+    "lofo_j2_linear_isotropic_64": {
+      "path": "models/t2_denim_evidence_v3/lofo_j2_linear_isotropic_64.safetensors",
+      "sha256": "760b1429a631d6e53c53412086c73b9051fb19a03bddde350abfc90d0d4b1099"
+    },
+    "lofo_chaboche_combined_zero": {
+      "path": "models/t2_denim_evidence_v3/lofo_chaboche_combined_zero.safetensors",
+      "sha256": "d3fd0ed95602d76bb10ac7242485f39b0efe15c7c6792ce8f1f990694ec13427"
+    },
+    "lofo_chaboche_combined_16": {
+      "path": "models/t2_denim_evidence_v3/lofo_chaboche_combined_16.safetensors",
+      "sha256": "4afaa868bdeba09f87880f51487f7b396584b3dcf8666954ed4543e3fdd6fb15"
+    },
+    "lofo_chaboche_combined_64": {
+      "path": "models/t2_denim_evidence_v3/lofo_chaboche_combined_64.safetensors",
+      "sha256": "e1f3ef0b9371e11598135712e73a39da21a640fe7483efb8aa0f88553e3f63ab"
+    },
+    "lofo_hidden_three_memory_tabulated_hardening_zero": {
+      "path": "models/t2_denim_evidence_v3/lofo_hidden_three_memory_tabulated_hardening_zero.safetensors",
+      "sha256": "b0e6868b3d58ab1a8395ccc4754d563965f882bc106bc5fad44165310c68bf26"
+    },
+    "lofo_hidden_three_memory_tabulated_hardening_16": {
+      "path": "models/t2_denim_evidence_v3/lofo_hidden_three_memory_tabulated_hardening_16.safetensors",
+      "sha256": "af624369788d1372ac43041c8a5092f3db417eadb8bbb87283cec0fca2f9718f"
+    },
+    "lofo_hidden_three_memory_tabulated_hardening_64": {
+      "path": "models/t2_denim_evidence_v3/lofo_hidden_three_memory_tabulated_hardening_64.safetensors",
+      "sha256": "76106c05dbe5aeaa8579e2e765a3f12f744d8e4e050e6048783b093fab3e35e9"
+    }
+  },
+  "elapsed_seconds": 1603.7899112082087,
+  "lofo_refinement": {
+    "learning_rate": 0.0012,
+    "steps_16": 900,
+    "steps_64": 1200,
+    "reason": "balanced path-family coverage for fair few-shot boundary measurement"
+  },
+  "production_checkpoint_reference": {
+    "training": {
+      "reused": "t2_denim_conditional_v2/denim_conditional_v2_refined.pt"
+    },
+    "evaluation": {
+      "hidden_frozen_test": {
+        "trajectory_count": 32,
+        "rmse_mpa": 0.7980589270591736,
+        "mae_mpa": 0.39469894766807556,
+        "r2": 0.9999515414237976,
+        "state_point_count": 7712
+      },
+      "test_amplitude_ood": {
+        "trajectory_count": 60,
+        "rmse_mpa": 2.67646861076355,
+        "mae_mpa": 0.9395037889480591,
+        "r2": 0.9994901418685913,
+        "state_point_count": 14460
+      },
+      "test_discretization": {
+        "trajectory_count": 20,
+        "rmse_mpa": 0.9083492159843445,
+        "mae_mpa": 0.4616833031177521,
+        "r2": 0.9999374747276306,
+        "state_point_count": 8120
+      },
+      "test_long_horizon": {
+        "trajectory_count": 40,
+        "rmse_mpa": 4.262782096862793,
+        "mae_mpa": 1.5307230949401855,
+        "r2": 0.9986092448234558,
+        "state_point_count": 9640
+      },
+      "test_path_ood": {
+        "trajectory_count": 80,
+        "rmse_mpa": 0.782439112663269,
+        "mae_mpa": 0.39284950494766235,
+        "r2": 0.9999520778656006,
+        "state_point_count": 19280
+      }
+    }
+  },
+  "ablation_protocol": {
+    "steps": 1400,
+    "scratch_seed": 20260930,
+    "transfer_seed": 20260931,
+    "same_architecture_optimizer_and_validation_policy": true,
+    "production_checkpoint_reported_separately": true
+  }
+}

article_evidence_v3/protocol.json

@@ -0,0 +1,41 @@
+{
+  "schema": "agentfem.denim-dataset-protocol.v3",
+  "data_are_additive": true,
+  "trajectory_total": 2660,
+  "frozen_protocol": {
+    "total_trajectories": 2660,
+    "train": 1484,
+    "validation": 407,
+    "test": 769,
+    "family_counts": {
+      "j2_linear_isotropic": {
+        "train": 577,
+        "validation": 171,
+        "test": 268
+      },
+      "chaboche_combined": {
+        "train": 577,
+        "validation": 170,
+        "test": 269
+      },
+      "hidden_three_memory_tabulated_hardening": {
+        "train": 330,
+        "validation": 66,
+        "test": 232
+      }
+    },
+    "selection_rule": "deterministic round-robin across path families; held-family training trajectories only"
+  },
+  "ablation_training_counts": [
+    80,
+    330,
+    1234,
+    1484
+  ],
+  "leave_one_family_out_adaptation_counts": [
+    0,
+    16,
+    64
+  ],
+  "test_trajectories_never_used_for_optimization": true
+}

article_evidence_v3/run_t2_denim_evidence_v3.py

@@ -0,0 +1,256 @@
+"""Run the frozen ablation and leave-one-family-out DENIM evidence matrix.
+
+This script never optimizes on validation or test trajectories.  It answers
+two release questions: how much the 80/330 incomplete-family trajectories add,
+and how a shared conditional closure behaves when one material family is
+entirely unseen before 16/64-trajectory adaptation.
+"""
+
+from __future__ import annotations
+
+import copy
+import hashlib
+import json
+import random
+import time
+from collections import defaultdict
+from pathlib import Path
+
+import numpy as np
+import torch
+from safetensors.torch import save_file
+
+from src.t2_denim_conditional import ConditionalDENIM
+from src.train_t2_denim_conditional_v2 import (
+    ARTIFACT_DIR as V2_ARTIFACT_DIR,
+    MODEL_DIR as V2_MODEL_DIR,
+    Trajectory,
+    _train_stage,
+    evaluate,
+    load_all_trajectories,
+    transitions,
+)
+
+
+ROOT = Path(__file__).resolve().parents[1]
+OUTPUT = ROOT / "artifacts" / "t2_denim_evidence_v3"
+MODELS = ROOT / "models" / "t2_denim_evidence_v3"
+FAMILIES = (
+    "j2_linear_isotropic",
+    "chaboche_combined",
+    "hidden_three_memory_tabulated_hardening",
+)
+
+
+def _load(path: Path) -> ConditionalDENIM:
+    checkpoint = torch.load(path, map_location="cpu", weights_only=False)
+    model = ConditionalDENIM()
+    model.load_state_dict(checkpoint["state_dict"])
+    model.eval()
+    return model
+
+
+def _save(model: ConditionalDENIM, name: str, metadata: dict[str, object]) -> dict[str, object]:
+    MODELS.mkdir(parents=True, exist_ok=True)
+    path = MODELS / f"{name}.safetensors"
+    save_file({key: value.detach().cpu().contiguous() for key, value in model.state_dict().items()}, path)
+    digest = hashlib.sha256(path.read_bytes()).hexdigest()
+    (MODELS / f"{name}.json").write_text(
+        json.dumps({**metadata, "sha256": digest}, indent=2) + "\n", encoding="utf-8"
+    )
+    return {"path": str(path.relative_to(ROOT)), "sha256": digest}
+
+
+def _train(
+    initial: ConditionalDENIM | None,
+    train_groups: list[list[Trajectory]],
+    validation_groups: list[list[Trajectory]],
+    *,
+    steps: int,
+    seed: int,
+) -> tuple[ConditionalDENIM, dict[str, object]]:
+    torch.manual_seed(seed)
+    model = ConditionalDENIM() if initial is None else copy.deepcopy(initial)
+    history = _train_stage(
+        model,
+        [transitions(group) for group in train_groups],
+        [transitions(group) for group in validation_groups],
+        steps=steps,
+        seed=seed,
+    )
+    return model, {key: value for key, value in history.items() if key != "history"}
+
+
+def _ids(trajectories: list[Trajectory], selected: list[Trajectory]) -> list[int]:
+    lookup = {id(value): index for index, value in enumerate(trajectories)}
+    return [lookup[id(value)] for value in selected]
+
+
+def _hidden_protocols(trajectories: list[Trajectory]) -> dict[str, list[Trajectory]]:
+    groups: dict[str, list[Trajectory]] = defaultdict(list)
+    for item in trajectories:
+        if item.material == FAMILIES[2] and item.role == "test":
+            groups[item.evaluation].append(item)
+    return dict(groups)
+
+
+def stratified_few_shot(values: list[Trajectory], count: int) -> list[Trajectory]:
+    """Select a deterministic path-family-balanced adaptation cohort."""
+
+    groups: dict[str, list[Trajectory]] = defaultdict(list)
+    for item in values:
+        groups[item.family].append(item)
+    selected: list[Trajectory] = []
+    depth = 0
+    names = sorted(groups)
+    while len(selected) < count:
+        changed = False
+        for name in names:
+            if depth < len(groups[name]) and len(selected) < count:
+                selected.append(groups[name][depth])
+                changed = True
+        if not changed:
+            break
+        depth += 1
+    if len(selected) != count:
+        raise RuntimeError(f"Cannot select {count} stratified trajectories from {len(values)}.")
+    return selected
+
+
+def _evaluate_hidden(model: ConditionalDENIM, groups: dict[str, list[Trajectory]]) -> dict[str, object]:
+    return {name: evaluate(model, group) for name, group in sorted(groups.items())}
+
+
+def run(*, steps: int = 1400, adaptation_steps: int = 900) -> dict[str, object]:
+    random.seed(20260925)
+    np.random.seed(20260925)
+    torch.manual_seed(20260925)
+    trajectories = load_all_trajectories()
+    train = {family: [x for x in trajectories if x.material == family and x.role == "train"] for family in FAMILIES}
+    validation = {family: [x for x in trajectories if x.material == family and x.role == "validation"] for family in FAMILIES}
+    test = {family: [x for x in trajectories if x.material == family and x.role == "test"] for family in FAMILIES}
+    hidden_groups = _hidden_protocols(trajectories)
+    closure80 = [x for x in train[FAMILIES[2]] if x.source == "denim_closure_v1"]
+    if len(closure80) != 80 or len(train[FAMILIES[2]]) != 330:
+        raise RuntimeError("Frozen hidden-family training cohorts changed.")
+
+    started = time.perf_counter()
+    records: dict[str, object] = {}
+    artifacts: dict[str, object] = {}
+
+    pretrained = _load(V2_MODEL_DIR / "denim_conditional_pretrained.pt")
+    full = _load(V2_MODEL_DIR / "denim_conditional_v2_refined.pt")
+    scratch330 = _load(V2_MODEL_DIR / "denim_conditional_scratch.pt")
+
+    scratch80, training = _train(
+        None, [closure80], [validation[FAMILIES[2]]], steps=steps, seed=20260930
+    )
+    records["hidden_80_scratch"] = {
+        "training": training,
+        "evaluation": _evaluate_hidden(scratch80, hidden_groups),
+    }
+    artifacts["hidden_80_scratch"] = _save(
+        scratch80, "hidden_80_scratch", {"training_trajectory_ids": _ids(trajectories, closure80)}
+    )
+
+    records["hidden_330_scratch"] = {
+        "training": {"reused": "t2_denim_conditional_v2/denim_conditional_scratch.pt"},
+        "evaluation": _evaluate_hidden(scratch330, hidden_groups),
+    }
+
+    transfer80, training = _train(
+        pretrained,
+        [train[FAMILIES[0]] + train[FAMILIES[1]], closure80],
+        [validation[FAMILIES[0]] + validation[FAMILIES[1]], validation[FAMILIES[2]]],
+        steps=steps,
+        seed=20260931,
+    )
+    records["known_1154_plus_hidden_80"] = {
+        "training": training,
+        "evaluation": _evaluate_hidden(transfer80, hidden_groups),
+    }
+    artifacts["known_1154_plus_hidden_80"] = _save(
+        transfer80,
+        "known_1154_plus_hidden_80",
+        {"hidden_training_trajectory_ids": _ids(trajectories, closure80)},
+    )
+    records["known_1154_plus_hidden_330"] = {
+        "training": {"reused": "t2_denim_conditional_v2/denim_conditional_v2_refined.pt"},
+        "evaluation": _evaluate_hidden(full, hidden_groups),
+    }
+
+    lofo: dict[str, object] = {}
+    for target_index, target in enumerate(FAMILIES):
+        others = [family for family in FAMILIES if family != target]
+        base, base_training = _train(
+            None,
+            [train[family] for family in others],
+            [validation[family] for family in others],
+            steps=steps,
+            seed=20261010 + target_index,
+        )
+        family_record: dict[str, object] = {
+            "held_out_family": target,
+            "base_training": base_training,
+            "zero_shot": evaluate(base, test[target]),
+            "adaptation": {},
+        }
+        artifacts[f"lofo_{target}_zero"] = _save(
+            base,
+            f"lofo_{target}_zero",
+            {"held_out_family": target, "adaptation_trajectories": 0},
+        )
+        for count in (16, 64):
+            selected = stratified_few_shot(train[target], count)
+            adapted, adaptation = _train(
+                base,
+                [sum((train[family] for family in others), []), selected],
+                [sum((validation[family] for family in others), []), validation[target]],
+                steps=adaptation_steps,
+                seed=20261100 + 10 * target_index + count,
+            )
+            family_record["adaptation"][str(count)] = {
+                "training": adaptation,
+                "training_trajectory_ids": _ids(trajectories, selected),
+                "evaluation": evaluate(adapted, test[target]),
+            }
+            artifacts[f"lofo_{target}_{count}"] = _save(
+                adapted,
+                f"lofo_{target}_{count}",
+                {
+                    "held_out_family": target,
+                    "adaptation_trajectories": count,
+                    "training_trajectory_ids": _ids(trajectories, selected),
+                },
+            )
+        lofo[target] = family_record
+        print(json.dumps({"completed_lofo": target, "metrics": family_record}, indent=2), flush=True)
+
+    result = {
+        "schema": "agentfem.denim-evidence.v3",
+        "frozen_protocol": {
+            "total_trajectories": len(trajectories),
+            "train": sum(len(values) for values in train.values()),
+            "validation": sum(len(values) for values in validation.values()),
+            "test": sum(len(values) for values in test.values()),
+            "family_counts": {
+                family: {"train": len(train[family]), "validation": len(validation[family]), "test": len(test[family])}
+                for family in FAMILIES
+            },
+            "selection_rule": "deterministic round-robin across path families; held-family training trajectories only",
+        },
+        "ablation": records,
+        "leave_one_family_out": lofo,
+        "artifacts": artifacts,
+        "elapsed_seconds": time.perf_counter() - started,
+    }
+    OUTPUT.mkdir(parents=True, exist_ok=True)
+    (OUTPUT / "evidence_metrics.json").write_text(
+        json.dumps(result, indent=2) + "\n", encoding="utf-8"
+    )
+    print(json.dumps({"elapsed_seconds": result["elapsed_seconds"], "output": str(OUTPUT)}, indent=2))
+    return result
+
+
+if __name__ == "__main__":
+    run()