Synchronize experiment results (part 7)

.gitattributes

@@ -96,3 +96,4 @@ results/cross_domain_multi/data.jsonl filter=lfs diff=lfs merge=lfs -text
 results/cross_domain_sweep/data.jsonl filter=lfs diff=lfs merge=lfs -text
 results/cross_domain_sweep_cap16/data.jsonl filter=lfs diff=lfs merge=lfs -text
 results/llm/gpt2_pipeline/semantics/llm_cache.sqlite3 filter=lfs diff=lfs merge=lfs -text
+results/llm/qwen2.5_0.5B_pipeline/semantics/llm_cache.sqlite3 filter=lfs diff=lfs merge=lfs -text

results/llm/qwen2.5_0.5B_pipeline/PIPELINE_MANIFEST.json

@@ -0,0 +1,30 @@
+{
+  "pipeline": "final learning-theoretic FCM pipeline (post-scan validation)",
+  "steps": [
+    "cross_fitted_operator_without_semantic_labels",
+    "K_1_to_Kmax_bootstrap_adaptive_block_matching",
+    "frequency_quantile_closed_block_detection",
+    "group_OMP_plus_training_only_swap_refinement",
+    "bootstrap_refit_heldout_module_width_certificate",
+    "frozen_object_parent_gated_semantics",
+    "untouched_partition_causal_validation"
+  ],
+  "operator_manifest": "/home/elicer/fcm_auto_discovery/results/llm/qwen2.5_0.5B_pipeline/operator/OPERATOR_FIT_MANIFEST.json",
+  "structural_scan": "/home/elicer/fcm_auto_discovery/results/llm/qwen2.5_0.5B_pipeline/structure/STABLE_BLOCK_SCAN.json",
+  "layers_requested": [
+    12,
+    13,
+    14,
+    15,
+    16,
+    17,
+    18,
+    19,
+    20,
+    21,
+    22,
+    23
+  ],
+  "semantic_summary": "/home/elicer/fcm_auto_discovery/results/llm/qwen2.5_0.5B_pipeline/semantics/SEMANTIC_BLOCK_SUMMARY.json",
+  "causal_summary": "/home/elicer/fcm_auto_discovery/results/llm/qwen2.5_0.5B_pipeline/causal/STABLE_BLOCK_CAUSAL_SUMMARY.json"
+}

results/llm/qwen2.5_0.5B_pipeline/causal/STABLE_BLOCK_CAUSAL_SUMMARY.json

@@ -0,0 +1,373 @@
+{
+  "aggregate": {
+    "median_dose_0.5_causal_correlation": 0.9781211685430513,
+    "median_dose_1.0_causal_correlation": 0.9831946272476291,
+    "median_fisher_relative_error": 0.9006691709312065,
+    "median_hook_relative_write_error": 0.0046528352531989125,
+    "median_jacobian_relative_error": 0.8394557391608667,
+    "n_layers": 7
+  },
+  "causal_partition": {
+    "causal_tokens": 50000,
+    "skip_tokens": 550000
+  },
+  "doses": [
+    0.0,
+    0.25,
+    0.5,
+    0.75,
+    1.0
+  ],
+  "downstream_metric": {
+    "finite_difference_step": 0.05,
+    "requested_method": "auto"
+  },
+  "experiment": "final stable-block causal validation summary",
+  "layers": [
+    {
+      "causal_fisher_relative_error": 0.9679715038130725,
+      "causal_float32_relative_write_error": 0.002562167689109037,
+      "causal_hook_relative_write_error": 0.002588326590496016,
+      "causal_jacobian_estimator": {
+        "automatic_fallback_reason": "NotImplementedError: Trying to use forward AD with _scaled_dot_product_flash_attention that does not support it because it has not been implemented yet.\nPlease file an issue to PyTorch at https://github.com/pytorch/pytorch/issues/new?template=feature-request.yml so that we can prioritize its implementation or submit a PR adding the implementation to derivatives.yaml",
+        "finite_difference_step": 0.05,
+        "methods_used": [
+          "finite_difference"
+        ],
+        "requested_method": "auto",
+        "resolved_method": "finite_difference"
+      },
+      "causal_jacobian_relative_error": 0.9535912092096361,
+      "dose_0.5": {
+        "full_module_mean_token_js": 0.00019690021812512548,
+        "full_to_module_mean_token_kl": 0.0007878188343872269,
+        "loss_delta_ratio": -6.562674001274612,
+        "loss_effect_regression_intercept": 0.0005178195322465665,
+        "loss_effect_regression_slope": 0.503826382177257,
+        "mean_abs_paired_loss_delta_difference": 0.001720172842753302,
+        "module_to_full_mean_token_kl": 0.0007879291727092386,
+        "per_sequence_loss_delta_correlation": 0.5773640337283498,
+        "rmse_paired_loss_delta_difference": 0.0021138484644388795
+      },
+      "dose_1.0": {
+        "full_module_mean_token_js": 0.00019957467561981343,
+        "full_to_module_mean_token_kl": 0.0007985719364405293,
+        "loss_delta_ratio": 0.8602160493827161,
+        "loss_effect_regression_intercept": 7.587002746573718e-05,
+        "loss_effect_regression_slope": 0.7649458947872652,
+        "mean_abs_paired_loss_delta_difference": 0.00154543414558332,
+        "module_to_full_mean_token_kl": 0.0007985879615496498,
+        "per_sequence_loss_delta_correlation": 0.7892962216483344,
+        "rmse_paired_loss_delta_difference": 0.0018786033212196981
+      },
+      "layer": 12,
+      "module_width": 6,
+      "result_file": "/home/elicer/fcm_auto_discovery/results/llm/qwen2.5_0.5B_pipeline/causal/layers/L12_top1_causal.json",
+      "selection_status": "high_confidence",
+      "support": [
+        2646,
+        4299,
+        2404,
+        3036,
+        3209,
+        1892
+      ],
+      "top_k": 1
+    },
+    {
+      "causal_fisher_relative_error": 0.9006691709312065,
+      "causal_float32_relative_write_error": 0.0009738639471939951,
+      "causal_hook_relative_write_error": 0.0010123318944013425,
+      "causal_jacobian_estimator": {
+        "automatic_fallback_reason": "NotImplementedError: Trying to use forward AD with _scaled_dot_product_flash_attention that does not support it because it has not been implemented yet.\nPlease file an issue to PyTorch at https://github.com/pytorch/pytorch/issues/new?template=feature-request.yml so that we can prioritize its implementation or submit a PR adding the implementation to derivatives.yaml",
+        "finite_difference_step": 0.05,
+        "methods_used": [
+          "finite_difference"
+        ],
+        "requested_method": "auto",
+        "resolved_method": "finite_difference"
+      },
+      "causal_jacobian_relative_error": 0.8394557391608667,
+      "dose_0.5": {
+        "full_module_mean_token_js": 0.00019116187033113861,
+        "full_to_module_mean_token_kl": 0.0007648701454301547,
+        "loss_delta_ratio": 1.0692942402923036,
+        "loss_effect_regression_intercept": 0.0001708932619555281,
+        "loss_effect_regression_slope": 1.0147559853889532,
+        "mean_abs_paired_loss_delta_difference": 0.0013858436309185224,
+        "module_to_full_mean_token_kl": 0.0007649346305172622,
+        "per_sequence_loss_delta_correlation": 0.9809041755109373,
+        "rmse_paired_loss_delta_difference": 0.0017923780366190512
+      },
+      "dose_1.0": {
+        "full_module_mean_token_js": 0.00019395076453724165,
+        "full_to_module_mean_token_kl": 0.0007760935611830852,
+        "loss_delta_ratio": 0.9947423471721226,
+        "loss_effect_regression_intercept": -3.922078784846645e-05,
+        "loss_effect_regression_slope": 0.9898607150256323,
+        "mean_abs_paired_loss_delta_difference": 0.0013727404407619201,
+        "module_to_full_mean_token_kl": 0.0007760429217182484,
+        "per_sequence_loss_delta_correlation": 0.9967067402738138,
+        "rmse_paired_loss_delta_difference": 0.0016174744091602407
+      },
+      "layer": 13,
+      "module_width": 1,
+      "result_file": "/home/elicer/fcm_auto_discovery/results/llm/qwen2.5_0.5B_pipeline/causal/layers/L13_top1_causal.json",
+      "selection_status": "high_confidence",
+      "support": [
+        1757
+      ],
+      "top_k": 1
+    },
+    {
+      "causal_fisher_relative_error": 0.8947071548882838,
+      "causal_float32_relative_write_error": 0.0013657923311508307,
+      "causal_hook_relative_write_error": 0.0014223118589259147,
+      "causal_jacobian_estimator": {
+        "automatic_fallback_reason": "NotImplementedError: Trying to use forward AD with _scaled_dot_product_flash_attention that does not support it because it has not been implemented yet.\nPlease file an issue to PyTorch at https://github.com/pytorch/pytorch/issues/new?template=feature-request.yml so that we can prioritize its implementation or submit a PR adding the implementation to derivatives.yaml",
+        "finite_difference_step": 0.05,
+        "methods_used": [
+          "finite_difference"
+        ],
+        "requested_method": "auto",
+        "resolved_method": "finite_difference"
+      },
+      "causal_jacobian_relative_error": 0.8248047951777114,
+      "dose_0.5": {
+        "full_module_mean_token_js": 0.00019227674596381287,
+        "full_to_module_mean_token_kl": 0.0007693542662009443,
+        "loss_delta_ratio": 1.0554075148727513,
+        "loss_effect_regression_intercept": 0.00020909157187050915,
+        "loss_effect_regression_slope": 0.9920524390883841,
+        "mean_abs_paired_loss_delta_difference": 0.0014742698866067473,
+        "module_to_full_mean_token_kl": 0.0007693830699919118,
+        "per_sequence_loss_delta_correlation": 0.981127825137671,
+        "rmse_paired_loss_delta_difference": 0.0017454401631977368
+      },
+      "dose_1.0": {
+        "full_module_mean_token_js": 0.00019466372842290774,
+        "full_to_module_mean_token_kl": 0.0007789658290122818,
+        "loss_delta_ratio": 1.0229331660444336,
+        "loss_effect_regression_intercept": -0.0003516838696427201,
+        "loss_effect_regression_slope": 0.9791478367910335,
+        "mean_abs_paired_loss_delta_difference": 0.0015591862275428378,
+        "module_to_full_mean_token_kl": 0.0007788754361153571,
+        "per_sequence_loss_delta_correlation": 0.9954689399359067,
+        "rmse_paired_loss_delta_difference": 0.00195340680001208
+      },
+      "layer": 13,
+      "module_width": 3,
+      "result_file": "/home/elicer/fcm_auto_discovery/results/llm/qwen2.5_0.5B_pipeline/causal/layers/L13_top2_causal.json",
+      "selection_status": "high_confidence",
+      "support": [
+        1757,
+        770,
+        2219
+      ],
+      "top_k": 2
+    },
+    {
+      "causal_fisher_relative_error": 0.8725793978503331,
+      "causal_float32_relative_write_error": 0.008551832680115962,
+      "causal_hook_relative_write_error": 0.008623535789185368,
+      "causal_jacobian_estimator": {
+        "automatic_fallback_reason": "NotImplementedError: Trying to use forward AD with _scaled_dot_product_flash_attention that does not support it because it has not been implemented yet.\nPlease file an issue to PyTorch at https://github.com/pytorch/pytorch/issues/new?template=feature-request.yml so that we can prioritize its implementation or submit a PR adding the implementation to derivatives.yaml",
+        "finite_difference_step": 0.05,
+        "methods_used": [
+          "finite_difference"
+        ],
+        "requested_method": "auto",
+        "resolved_method": "finite_difference"
+      },
+      "causal_jacobian_relative_error": 0.8109036972242986,
+      "dose_0.5": {
+        "full_module_mean_token_js": 0.00020105595752649663,
+        "full_to_module_mean_token_kl": 0.0008045018791009855,
+        "loss_delta_ratio": 1.0865862926891183,
+        "loss_effect_regression_intercept": 0.00028017489403450453,
+        "loss_effect_regression_slope": 0.999431562201443,
+        "mean_abs_paired_loss_delta_difference": 0.0015808208701536828,
+        "module_to_full_mean_token_kl": 0.0008045077382721315,
+        "per_sequence_loss_delta_correlation": 0.9781211685430513,
+        "rmse_paired_loss_delta_difference": 0.0019520773998573538
+      },
+      "dose_1.0": {
+        "full_module_mean_token_js": 0.00021777415216645975,
+        "full_to_module_mean_token_kl": 0.000871439322771639,
+        "loss_delta_ratio": 0.937620133153445,
+        "loss_effect_regression_intercept": 0.00024042628919594483,
+        "loss_effect_regression_slope": 0.9642325480780283,
+        "mean_abs_paired_loss_delta_difference": 0.0014842416822295828,
+        "module_to_full_mean_token_kl": 0.000871411783947158,
+        "per_sequence_loss_delta_correlation": 0.996687262190585,
+        "rmse_paired_loss_delta_difference": 0.0019058052115527118
+      },
+      "layer": 13,
+      "module_width": 3,
+      "result_file": "/home/elicer/fcm_auto_discovery/results/llm/qwen2.5_0.5B_pipeline/causal/layers/L13_top3_causal.json",
+      "selection_status": "high_confidence",
+      "support": [
+        1757,
+        770,
+        2219
+      ],
+      "top_k": 3
+    },
+    {
+      "causal_fisher_relative_error": 0.9852635410541086,
+      "causal_float32_relative_write_error": 0.004581114360546577,
+      "causal_hook_relative_write_error": 0.0046528352531989125,
+      "causal_jacobian_estimator": {
+        "automatic_fallback_reason": "NotImplementedError: Trying to use forward AD with _scaled_dot_product_flash_attention that does not support it because it has not been implemented yet.\nPlease file an issue to PyTorch at https://github.com/pytorch/pytorch/issues/new?template=feature-request.yml so that we can prioritize its implementation or submit a PR adding the implementation to derivatives.yaml",
+        "finite_difference_step": 0.05,
+        "methods_used": [
+          "finite_difference"
+        ],
+        "requested_method": "auto",
+        "resolved_method": "finite_difference"
+      },
+      "causal_jacobian_relative_error": 0.9628643734392903,
+      "dose_0.5": {
+        "full_module_mean_token_js": 0.00018461756306201142,
+        "full_to_module_mean_token_kl": 0.0007387188499668593,
+        "loss_delta_ratio": 0.5459097238630023,
+        "loss_effect_regression_intercept": -3.412514598113641e-05,
+        "loss_effect_regression_slope": 0.6157042168078029,
+        "mean_abs_paired_loss_delta_difference": 0.0013844327828318802,
+        "module_to_full_mean_token_kl": 0.0007387067757777332,
+        "per_sequence_loss_delta_correlation": 0.6286865720151205,
+        "rmse_paired_loss_delta_difference": 0.0017375831901857953
+      },
+      "dose_1.0": {
+        "full_module_mean_token_js": 0.0001860414469643429,
+        "full_to_module_mean_token_kl": 0.0007443596409710566,
+        "loss_delta_ratio": 0.9209963032373545,
+        "loss_effect_regression_intercept": 9.251508299088741e-05,
+        "loss_effect_regression_slope": 0.8330415590969005,
+        "mean_abs_paired_loss_delta_difference": 0.0014264534429176566,
+        "module_to_full_mean_token_kl": 0.0007444552233509873,
+        "per_sequence_loss_delta_correlation": 0.8649072082560829,
+        "rmse_paired_loss_delta_difference": 0.0018348949597987534
+      },
+      "layer": 14,
+      "module_width": 3,
+      "result_file": "/home/elicer/fcm_auto_discovery/results/llm/qwen2.5_0.5B_pipeline/causal/layers/L14_top1_causal.json",
+      "selection_status": "high_confidence",
+      "support": [
+        4278,
+        2634,
+        1698
+      ],
+      "top_k": 1
+    },
+    {
+      "causal_fisher_relative_error": 0.984156152225748,
+      "causal_float32_relative_write_error": 0.006824905209587115,
+      "causal_hook_relative_write_error": 0.006870400327692362,
+      "causal_jacobian_estimator": {
+        "automatic_fallback_reason": "NotImplementedError: Trying to use forward AD with _scaled_dot_product_flash_attention that does not support it because it has not been implemented yet.\nPlease file an issue to PyTorch at https://github.com/pytorch/pytorch/issues/new?template=feature-request.yml so that we can prioritize its implementation or submit a PR adding the implementation to derivatives.yaml",
+        "finite_difference_step": 0.05,
+        "methods_used": [
+          "finite_difference"
+        ],
+        "requested_method": "auto",
+        "resolved_method": "finite_difference"
+      },
+      "causal_jacobian_relative_error": 0.9624209406016443,
+      "dose_0.5": {
+        "full_module_mean_token_js": 0.00018628599369999253,
+        "full_to_module_mean_token_kl": 0.0007453639553523786,
+        "loss_delta_ratio": 2.1640016456390563,
+        "loss_effect_regression_intercept": 0.0002850437448007641,
+        "loss_effect_regression_slope": 0.5736378011203526,
+        "mean_abs_paired_loss_delta_difference": 0.0015016609860449722,
+        "module_to_full_mean_token_kl": 0.0007454147016312735,
+        "per_sequence_loss_delta_correlation": 0.617494013879319,
+        "rmse_paired_loss_delta_difference": 0.0018412136387724127
+      },
+      "dose_1.0": {
+        "full_module_mean_token_js": 0.00018972112675984726,
+        "full_to_module_mean_token_kl": 0.0007591388547149746,
+        "loss_delta_ratio": 0.9510791221862683,
+        "loss_effect_regression_intercept": -1.7724302407498567e-05,
+        "loss_effect_regression_slope": 0.9753096840792708,
+        "mean_abs_paired_loss_delta_difference": 0.0011706425971591595,
+        "module_to_full_mean_token_kl": 0.0007591345221447499,
+        "per_sequence_loss_delta_correlation": 0.9234363555698875,
+        "rmse_paired_loss_delta_difference": 0.0015158693535845323
+      },
+      "layer": 14,
+      "module_width": 5,
+      "result_file": "/home/elicer/fcm_auto_discovery/results/llm/qwen2.5_0.5B_pipeline/causal/layers/L14_top2_causal.json",
+      "selection_status": "high_confidence",
+      "support": [
+        4278,
+        2854,
+        481,
+        2634,
+        1698
+      ],
+      "top_k": 2
+    },
+    {
+      "causal_fisher_relative_error": 0.09610758359671492,
+      "causal_float32_relative_write_error": 0.007166055671173002,
+      "causal_hook_relative_write_error": 0.007328782530333054,
+      "causal_jacobian_estimator": {
+        "automatic_fallback_reason": "NotImplementedError: Trying to use forward AD with _scaled_dot_product_flash_attention that does not support it because it has not been implemented yet.\nPlease file an issue to PyTorch at https://github.com/pytorch/pytorch/issues/new?template=feature-request.yml so that we can prioritize its implementation or submit a PR adding the implementation to derivatives.yaml",
+        "finite_difference_step": 0.05,
+        "methods_used": [
+          "finite_difference"
+        ],
+        "requested_method": "auto",
+        "resolved_method": "finite_difference"
+      },
+      "causal_jacobian_relative_error": 0.06978660354592606,
+      "dose_0.5": {
+        "full_module_mean_token_js": 0.00016123155259752895,
+        "full_to_module_mean_token_kl": 0.0006463739267291175,
+        "loss_delta_ratio": 1.0174279961194443,
+        "loss_effect_regression_intercept": 0.0008141808297775142,
+        "loss_effect_regression_slope": 0.9729558108348993,
+        "mean_abs_paired_loss_delta_difference": 0.001418819132539415,
+        "module_to_full_mean_token_kl": 0.0006478559153788176,
+        "per_sequence_loss_delta_correlation": 0.9813341465216705,
+        "rmse_paired_loss_delta_difference": 0.0017707805199238368
+      },
+      "dose_1.0": {
+        "full_module_mean_token_js": 0.00018306936529829867,
+        "full_to_module_mean_token_kl": 0.0007409298883693427,
+        "loss_delta_ratio": 1.0196517638306166,
+        "loss_effect_regression_intercept": 0.001064145162260015,
+        "loss_effect_regression_slope": 0.974514365882781,
+        "mean_abs_paired_loss_delta_difference": 0.0015785472909199824,
+        "module_to_full_mean_token_kl": 0.0007424846209931312,
+        "per_sequence_loss_delta_correlation": 0.9831946272476291,
+        "rmse_paired_loss_delta_difference": 0.0018929377225538658
+      },
+      "layer": 21,
+      "module_width": 1,
+      "result_file": "/home/elicer/fcm_auto_discovery/results/llm/qwen2.5_0.5B_pipeline/causal/layers/L21_top1_causal.json",
+      "selection_status": "high_confidence",
+      "support": [
+        567
+      ],
+      "top_k": 1
+    }
+  ],
+  "layers_requested": [
+    12,
+    13,
+    14,
+    15,
+    16,
+    17,
+    18,
+    19,
+    20,
+    21,
+    22,
+    23
+  ],
+  "scan_json": "results/llm/qwen2.5_0.5B_pipeline/structure/STABLE_BLOCK_SCAN.json"
+}

results/llm/qwen2.5_0.5B_pipeline/operator/OPERATOR_FIT_MANIFEST.json

@@ -1,175 +1,23 @@
 {
   "experiment": "final FCM cross-fitted operator fit",
   "layer_summaries": [
-    {
-      "layer": 0,
-      "n_modes": 12,
-      "noise_proxy": 0.0004578296502586454,
-      "top_eigenvalues": [
-        0.0025529107078909874,
-        0.001846993574872613,
-        0.0018409481272101402,
-        0.0016690139891579747,
-        0.001652997569181025,
-        0.0015749619342386723,
-        0.001513949828222394,
-        0.0014669835800305009,
-        0.001439486863091588,
-        0.001330690342001617,
-        0.0012903612805530429,
-        0.001259263837710023
-      ]
-    },
-    {
-      "layer": 1,
-      "n_modes": 12,
-      "noise_proxy": 0.0007243817672133446,
-      "top_eigenvalues": [
-        0.0016573303146287799,
-        0.0014361037174239755,
-        0.0012656296603381634,
-        0.0011564362794160843,
-        0.0009930180385708809,
-        0.0009625087841413915,
-        0.00091639842139557,
-        0.0009142282651737332,
-        0.000894141208846122,
-        0.0008776902686804533,
-        0.0008703446947038174,
-        0.0008498907554894686
-      ]
-    },
-    {
-      "layer": 2,
-      "n_modes": 12,
-      "noise_proxy": 0.0025943350046873093,
-      "top_eigenvalues": [
-        0.10447322577238083,
-        0.0014811063883826137,
-        0.0014659562148153782,
-        0.0010619326494634151,
-        0.0009653851739130914,
-        0.0009257346391677856,
-        0.0008696730947121978,
-        0.000813318882137537,
-        0.000811423291452229,
-        0.0007504104869440198,
-        0.0007007187814451754,
-        0.0006807986064814031
-      ]
-    },
-    {
-      "layer": 3,
-      "n_modes": 12,
-      "noise_proxy": 0.0006161753553897142,
-      "top_eigenvalues": [
-        0.108812615275383,
-        0.0014853139873594046,
-        0.001177590456791222,
-        0.0010743567254394293,
-        0.0010356560815125704,
-        0.0007879014010541141,
-        0.0007865044753998518,
-        0.000784696196205914,
-        0.0007659514085389674,
-        0.0006939597078599036,
-        0.0006880738073959947,
-        0.0006806555320508778
-      ]
-    },
-    {
-      "layer": 4,
-      "n_modes": 12,
-      "noise_proxy": 0.0035523874685168266,
-      "top_eigenvalues": [
-        0.003712416859343648,
-        0.0022392342798411846,
-        0.0017656482523307204,
-        0.0015337354270741343,
-        0.00143379473593086,
-        0.0014096475206315517,
-        0.0012295115739107132,
-        0.0011504573049023747,
-        0.0011152756633237004,
-        0.001109987497329712,
-        0.0010498241754248738,
-        0.0010346632916480303
-      ]
-    },
-    {
-      "layer": 5,
-      "n_modes": 12,
-      "noise_proxy": 0.009067830629646778,
-      "top_eigenvalues": [
-        0.009240304119884968,
-        0.006008951459079981,
-        0.00405877036973834,
-        0.0036186091601848602,
-        0.0024249586276710033,
-        0.0023161694407463074,
-        0.0022133903112262487,
-        0.002007170580327511,
-        0.0019215956563130021,
-        0.00186293909791857,
-        0.001741634914651513,
-        0.0016958648338913918
-      ]
-    },
-    {
-      "layer": 6,
-      "n_modes": 12,
-      "noise_proxy": 0.001615826040506363,
-      "top_eigenvalues": [
-        0.001999505329877138,
-        0.0014400719664990902,
-        0.001424510614015162,
-        0.0013461819617077708,
-        0.0013071974972262979,
-        0.0011832037707790732,
-        0.0010049182455986738,
-        0.0009268970461562276,
-        0.0009062953758984804,
-        0.0008914890349842608,
-        0.0008846995187923312,
-        0.0008547625620849431
-      ]
-    },
-    {
-      "layer": 7,
-      "n_modes": 12,
-      "noise_proxy": 0.0005503100692294538,
-      "top_eigenvalues": [
-        0.001771302311681211,
-        0.0014702205080538988,
-        0.0013954613823443651,
-        0.0010945197427645326,
-        0.0010202863486483693,
-        0.0009871127549558878,
-        0.0008797550108283758,
-        0.0008595152758061886,
-        0.0008358149789273739,
-        0.0008069726172834635,
-        0.0007955871406011283,
-        0.0007927389233373106
-      ]
-    },
     {
       "layer": 8,
       "n_modes": 12,
       "noise_proxy": 0.00035809201654046774,
       "top_eigenvalues": [
-        0.0011042995611205697,
-        0.0008600744768045843,
-        0.0007879775948822498,
-        0.0007819708553142846,
-        0.0007660360424779356,
+        0.0011042957194149494,
+        0.0008600710425525904,
+        0.0007879784679971635,
+        0.0007819749298505485,
+        0.0007660369738005102,
         0.000739099457859993,
-        0.0007161700050346553,
-        0.000702554767485708,
-        0.0006802285206504166,
-        0.0006751865148544312,
-        0.0006434704409912229,
-        0.0006374848890118301
+        0.000716174952685833,
+        0.0007025442901067436,
+        0.0006802307325415313,
+        0.0006751788896508515,
+        0.0006434576935134828,
+        0.0006374764489009976
       ]
     },
     {
@@ -177,312 +25,24 @@
       "n_modes": 12,
       "noise_proxy": 0.00030125415651127696,
       "top_eigenvalues": [
-        0.00128172489348799,
-        0.0008529043989256024,
-        0.0007900633499957621,
-        0.0007817356963641942,
-        0.0007544424734078348,
-        0.0007336981361731887,
-        0.0007272794027812779,
-        0.0007054239977151155,
-        0.0006889977375976741,
-        0.0006864757742732763,
-        0.0006852224469184875,
-        0.0006822206196375191
-      ]
-    },
-    {
-      "layer": 10,
-      "n_modes": 12,
-      "noise_proxy": 0.0003289701126050204,
-      "top_eigenvalues": [
-        0.0011117846006527543,
-        0.0010728632332757115,
-        0.0010562955867499113,
-        0.0009422693401575089,
-        0.0009188171825371683,
-        0.0009026590851135552,
-        0.0008145432802848518,
-        0.0007989321602508426,
-        0.0007767512579448521,
-        0.0007473158766515553,
-        0.000728330691345036,
-        0.0007233109208755195
-      ]
-    },
-    {
-      "layer": 11,
-      "n_modes": 12,
-      "noise_proxy": 0.0002047345624305308,
-      "top_eigenvalues": [
-        0.002310137962922454,
-        0.0022620246745646,
-        0.002253774320706725,
-        0.002220502123236656,
-        0.002215504879131913,
-        0.002188551938161254,
-        0.0021822352427989244,
-        0.002174628898501396,
-        0.002147188875824213,
-        0.0021393948700278997,
-        0.002116420539095998,
-        0.001978976419195533
-      ]
-    },
-    {
-      "layer": 12,
-      "n_modes": 12,
-      "noise_proxy": 0.00014893781917635351,
-      "top_eigenvalues": [
-        0.00137516378890723,
-        0.001246624975465238,
-        0.0009699046495370567,
-        0.0008294085273519158,
-        0.0008037644438445568,
-        0.0007852828712202609,
-        0.000739523908123374,
-        0.0007287413463927805,
-        0.0007156625506468117,
-        0.0007074682507663965,
-        0.0006836377433501184,
-        0.0006784487632103264
-      ]
-    },
-    {
-      "layer": 13,
-      "n_modes": 12,
-      "noise_proxy": 0.00074558868072927,
-      "top_eigenvalues": [
-        0.002215693471953273,
-        0.0015501020243391395,
-        0.0014483951963484287,
-        0.0008944656583480537,
-        0.00087796087609604,
-        0.0008171697263605893,
-        0.0007619653479196131,
-        0.0007518657366745174,
-        0.0007316024857573211,
-        0.0007199243409559131,
-        0.0007040360360406339,
-        0.0006888805073685944
-      ]
-    },
-    {
-      "layer": 14,
-      "n_modes": 12,
-      "noise_proxy": 0.0007072073058225214,
-      "top_eigenvalues": [
-        0.001213029376231134,
-        0.0011809697607532144,
-        0.0011740244226530194,
-        0.001086086151190102,
-        0.0010673822835087776,
-        0.0009612583089619875,
-        0.0009200417553074658,
-        0.0008845151751302183,
-        0.0008682254701852798,
-        0.0008360031642951071,
-        0.0008317838655784726,
-        0.0008003545226529241
-      ]
-    },
-    {
-      "layer": 15,
-      "n_modes": 12,
-      "noise_proxy": 0.0014647749485448003,
-      "top_eigenvalues": [
-        0.0019864016212522984,
-        0.0018195309676229954,
-        0.001772386603988707,
-        0.001669087097980082,
-        0.0015964909689500928,
-        0.0015281381784006953,
-        0.0013935146853327751,
-        0.0013220995897427201,
-        0.0012316794600337744,
-        0.0011373928282409906,
-        0.001118217594921589,
-        0.0010827797232195735
-      ]
-    },
-    {
-      "layer": 16,
-      "n_modes": 12,
-      "noise_proxy": 0.0012341620167717338,
-      "top_eigenvalues": [
-        0.0014217105926945806,
-        0.0014170955400913954,
-        0.0013846048386767507,
-        0.0013676694361492991,
-        0.0013143339892849326,
-        0.0012794567737728357,
-        0.0011783656664192677,
-        0.0011730565456673503,
-        0.0011676537105813622,
-        0.0011375972535461187,
-        0.001125900074839592,
-        0.0011192245874553919
-      ]
-    },
-    {
-      "layer": 17,
-      "n_modes": 12,
-      "noise_proxy": 0.01002126932144165,
-      "top_eigenvalues": [
-        0.01020407397300005,
-        0.002600831212475896,
-        0.002409868873655796,
-        0.0023856651969254017,
-        0.0021337782964110374,
-        0.0020099554676562548,
-        0.0019558160565793514,
-        0.0016858457820490003,
-        0.0016006388468667865,
-        0.0015264570247381926,
-        0.0014976696111261845,
-        0.001457395264878869
-      ]
-    },
-    {
-      "layer": 18,
-      "n_modes": 12,
-      "noise_proxy": 0.0015439848648384213,
-      "top_eigenvalues": [
-        0.0018585462821647525,
-        0.0017243503825739026,
-        0.0016948346747085452,
-        0.0016677732346579432,
-        0.0014411172596737742,
-        0.0012859748676419258,
-        0.0012627587420865893,
-        0.0012599454494193196,
-        0.0011534324148669839,
-        0.0011435139458626509,
-        0.0010838251328095794,
-        0.0010474930750206113
-      ]
-    },
-    {
-      "layer": 19,
-      "n_modes": 12,
-      "noise_proxy": 0.01198586355894804,
-      "top_eigenvalues": [
-        0.012130982242524624,
-        0.010460945777595043,
-        0.001632499392144382,
-        0.0015432165237143636,
-        0.0012960702879354358,
-        0.0012618962209671736,
-        0.0012591383419930935,
-        0.0012504609767347574,
-        0.0012364130234345794,
-        0.0011362977093085647,
-        0.0011290922993794084,
-        0.0011199087603017688
-      ]
-    },
-    {
-      "layer": 20,
-      "n_modes": 12,
-      "noise_proxy": 0.00534065393730998,
-      "top_eigenvalues": [
-        0.00558219850063324,
-        0.005346972960978746,
-        0.0044196792878210545,
-        0.0031733785290271044,
-        0.00193491589743644,
-        0.0016802209429442883,
-        0.0015803612768650055,
-        0.0015586006920784712,
-        0.0015285551780834794,
-        0.001424730638973415,
-        0.001381383161060512,
-        0.0013553269673138857
-      ]
-    },
-    {
-      "layer": 21,
-      "n_modes": 12,
-      "noise_proxy": 0.0058860876597464085,
-      "top_eigenvalues": [
-        0.05933131277561188,
-        0.006230871193110943,
-        0.004463900811970234,
-        0.004047534894198179,
-        0.0031463620252907276,
-        0.0020202239975333214,
-        0.0018641697242856026,
-        0.0018344890559092164,
-        0.0016679506516084075,
-        0.0016661640256643295,
-        0.0016021094052121043,
-        0.001541904523037374
-      ]
-    },
-    {
-      "layer": 22,
-      "n_modes": 12,
-      "noise_proxy": 0.007633066736161709,
-      "top_eigenvalues": [
-        0.007751170080155134,
-        0.007256270851939917,
-        0.00277764187194407,
-        0.002653731731697917,
-        0.0025857954751700163,
-        0.0022617534268647432,
-        0.0021908499766141176,
-        0.002110154600813985,
-        0.0020669030491262674,
-        0.002006187103688717,
-        0.001978958258405328,
-        0.0019665868021547794
-      ]
-    },
-    {
-      "layer": 23,
-      "n_modes": 12,
-      "noise_proxy": 0.008261716924607754,
-      "top_eigenvalues": [
-        0.008372127078473568,
-        0.007990064099431038,
-        0.005690667312592268,
-        0.004137176554650068,
-        0.003754871664568782,
-        0.003350531915202737,
-        0.0031643654219806194,
-        0.003044713521376252,
-        0.002895460231229663,
-        0.0028674171771854162,
-        0.002687951782718301,
-        0.00256848381832242
+        0.0012817481765523553,
+        0.0008529092301614583,
+        0.0007900725468061864,
+        0.0007817372097633779,
+        0.0007544419495388865,
+        0.0007336994167417288,
+        0.0007272816146723926,
+        0.000705425045453012,
+        0.0006890001823194325,
+        0.0006864730967208743,
+        0.0006852233200334013,
+        0.0006822223658673465
       ]
     }
   ],
   "layers": [
-    0,
-    1,
-    2,
-    3,
-    4,
-    5,
-    6,
-    7,
     8,
-    9,
-    10,
-    11,
-    12,
-    13,
-    14,
-    15,
-    16,
-    17,
-    18,
-    19,
-    20,
-    21,
-    22,
-    23
+    9
   ],
   "model": "Qwen/Qwen2.5-0.5B",
   "next_unused_skip_tokens": 200000,

results/llm/qwen2.5_0.5B_pipeline/operator/layers/L8_learning_geometry.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5e33cda6473ff0856ac3095de77d5dd2fb6d88de2f101fdc8b0cc842629a8957
-size 858395
+oid sha256:02c84df4f4e369d5ab0fdc15a625aaf42c8260f4b1fbd7171c47109262b817f0
+size 512283

results/llm/qwen2.5_0.5B_pipeline/operator/layers/L9_learning_geometry.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d416eae4e3f9ad800c421a4bbb21b740acd0feb6b6ef0936697b9236738d5f62
-size 858395
+oid sha256:f7623337b047506a1fcd69da2f5c9728aef4b38c044d64bd79075778e48009f8
+size 512283

results/llm/qwen2.5_0.5B_pipeline/pipeline_progress.log

@@ -10,3 +10,40 @@
 [2026-08-18 08:54:22] [2-5/7] DONE:  structural scan, matching, sparse fit, and certification (elapsed 02:14:07)
 [2026-08-18 08:54:22] [6/7] START: semantic validation
 [2026-08-18 08:54:22] FAILED during: semantic validation (EXPLAINER_MODEL and SCORER_MODEL are required unless SKIP_SEMANTICS=1)
+[2026-08-19 13:14:02] Post-scan pipeline started: model=allenai/OLMo-1B-hf output=results/llm/qwen2.5_0.5B_pipeline
+[2026-08-19 13:14:02] Using existing operator and structural scan artifacts
+[2026-08-19 13:14:02] [7/7] START: causal validation
+[2026-08-19 13:15:06] Post-scan pipeline started: model=Qwen/Qwen2.5-0.5B output=results/llm/qwen2.5_0.5B_pipeline
+[2026-08-19 13:15:06] Using existing operator and structural scan artifacts
+[2026-08-19 13:15:06] [6/7] START: semantic validation
+[2026-08-19 13:38:15] FAILED during: semantic validation (exit 1, post-scan elapsed 00:23:09)
+[2026-08-19 14:25:48] Post-scan pipeline started: model=Qwen/Qwen2.5-0.5B output=results/llm/qwen2.5_0.5B_pipeline
+[2026-08-19 14:25:48] Using existing operator and structural scan artifacts
+[2026-08-19 14:25:48] [6/7] START: semantic validation
+[2026-08-20 08:27:25] Pipeline started: model=Qwen/Qwen2.5-0.5B layers=[8 9] output=results/llm/qwen2.5_0.5B_pipeline
+[2026-08-20 08:27:25] Detailed stage output follows in this terminal; milestones are saved to results/llm/qwen2.5_0.5B_pipeline/pipeline_progress.log
+[2026-08-20 08:27:25] [1/7] START: cross-fitted operator
+[2026-08-20 08:29:34] [1/7] DONE:  cross-fitted operator (elapsed 00:02:09)
+[2026-08-20 08:29:34] [2-5/7] START: structural scan, matching, sparse fit, and certification
+[2026-08-20 08:34:32] Post-scan pipeline started: model=Qwen/Qwen2.5-0.5B output=results/llm/qwen2.5_0.5B_pipeline
+[2026-08-20 08:34:32] Using existing operator and structural scan artifacts
+[2026-08-20 08:34:32] [6/7] START: semantic validation
+[2026-08-20 08:34:51] FAILED during: semantic validation (exit 134, post-scan elapsed 00:00:19)
+[2026-08-20 08:41:11] Post-scan pipeline started: model=Qwen/Qwen2.5-0.5B output=results/llm/qwen2.5_0.5B_pipeline
+[2026-08-20 08:41:11] Using existing operator and structural scan artifacts
+[2026-08-20 08:41:11] Restricting semantic and causal validation to layers: [8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23]
+[2026-08-20 08:41:11] [6/7] START: semantic validation
+[2026-08-20 08:54:46] FAILED during: semantic validation (exit 1, post-scan elapsed 00:13:35)
+[2026-08-20 09:36:30] Post-scan pipeline started: model=Qwen/Qwen2.5-0.5B output=results/llm/qwen2.5_0.5B_pipeline
+[2026-08-20 09:36:30] Using existing operator and structural scan artifacts
+[2026-08-20 09:36:30] Restricting semantic and causal validation to layers: [12 13 14 15 16 17 18 19 20 21 22 23]
+[2026-08-20 09:36:30] [6/7] START: semantic validation
+[2026-08-20 09:38:09] FAILED during: semantic validation (exit 1, post-scan elapsed 00:01:39)
+[2026-08-20 09:40:58] Post-scan pipeline started: model=Qwen/Qwen2.5-0.5B output=results/llm/qwen2.5_0.5B_pipeline
+[2026-08-20 09:40:58] Using existing operator and structural scan artifacts
+[2026-08-20 09:40:58] Restricting semantic and causal validation to layers: [12 13 14 15 16 17 18 19 20 21 22 23]
+[2026-08-20 09:40:58] [6/7] START: semantic validation
+[2026-08-20 09:51:40] [6/7] DONE:  semantic validation (elapsed 00:10:42)
+[2026-08-20 09:51:40] [7/7] START: causal validation
+[2026-08-20 10:00:28] [7/7] DONE:  causal validation (elapsed 00:08:48)
+[2026-08-20 10:00:28] Post-scan pipeline complete in 00:19:30: results/llm/qwen2.5_0.5B_pipeline/PIPELINE_MANIFEST.json

results/llm/qwen2.5_0.5B_pipeline/semantics/L0B1.json

@@ -0,0 +1,66 @@
+{
+  "block_feature_id": "L0B1",
+  "certified_width": 1,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 0,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "heading or break marker",
+        "narrative or descriptive passage",
+        "'the' as passage opener"
+      ],
+      "exclusions": [
+        "'the' in dialogue",
+        "'the' in lists",
+        "'the' mid-paragraph"
+      ],
+      "feature_id": "L0B1__module_w1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"The computation fires at sentence-initial 'the' following a newline or section break.\",\"cues\":[\"preceding newline or break\",\"sentence-initial position\",\"definite article 'the'\"],\"exclusions\":[\"mid-sentence 'the'\",\"'the' after comma\",\"'the' after conjunction\"]},{\"text\":\"The computation activates when 'the' begins a new topical unit after structural formatting.\",\"cues\":[\"structural formatting cue\",\"new topical unit\",\"article 'the' at start\"],\"exclusions\":[\"continuation of prior sentence\",\"'the' within clause\",\"'the' after punctuation mid-line\"]},{\"text\":\"The computation triggers on 'the' that opens a descriptive or narrative passage after a heading or break.\",\"cues\":[\"heading or break marker\",\"narrative or descriptive passage\",\"'the' as passage opener\"],\"exclusions\":[\"'the' in dialogue\",\"'the' in lists\",\"'the' mid-paragraph\"]}]}",
+      "round_index": 0,
+      "text": "The computation triggers on 'the' that opens a descriptive or narrative passage after a heading or break."
+    },
+    "own_test_auc": 0.5,
+    "own_validation_auc": 0.515625,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L0B1__module_w1",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 4724121.09375,
+    "interpretation": {
+      "cues": [
+        "sequential or temporal markers",
+        "cohesive multi-sentence context",
+        "factual reporting style"
+      ],
+      "exclusions": [
+        "isolated factual fragments",
+        "heading or title context",
+        "list-like enumeration"
+      ],
+      "feature_id": "L0B1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Executes on 'the' within multi-clause narrative or expository sentences with three or more prepositional phrases.\",\"cues\":[\"multiple prepositional phrases\",\"multi-clause sentence structure\",\"narrative or expository context\"],\"exclusions\":[\"sentence-initial 'The'\",\"short simple sentences\",\"isolated single-phrase context\"]},{\"text\":\"Fires when 'the' appears mid-sentence after a verb or preposition, not at sentence start.\",\"cues\":[\"mid-sentence position\",\"preceded by verb or preposition\",\"lowercase 'the'\"],\"exclusions\":[\"sentence-initial capitalized 'The'\",\"clause-initial 'The'\",\"standalone title heading\"]},{\"text\":\"Activates on 'the' in passages with named entities and specific events or actions described.\",\"cues\":[\"named entities present\",\"specific events or actions\",\"descriptive detail density\"],\"exclusions\":[\"generic cultural descriptions\",\"abstract topic overviews\",\"definitional statements\"]},{\"text\":\"Triggers on 'the' within longer cohesive text spans containing sequential or temporal information.\",\"cues\":[\"sequential or temporal markers\",\"cohesive multi-sentence context\",\"factual reporting style\"],\"exclusions\":[\"isolated factual fragments\",\"heading or title context\",\"list-like enumeration\"]}]}",
+      "round_index": 0,
+      "text": "Triggers on 'the' within longer cohesive text spans containing sequential or temporal information."
+    },
+    "semantics_valid": false,
+    "test_auc": 0.5472412109375,
+    "validation_auc": 0.4912109375
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    2290
+  ],
+  "top_k": 1
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L0B3.json

@@ -0,0 +1,72 @@
+{
+  "block_feature_id": "L0B3",
+  "certified_width": 7,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 0,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "digit-dot-digit pattern",
+        "quantity or measurement",
+        "no hyphen present"
+      ],
+      "exclusions": [
+        "hyphenated adjectives",
+        "proper noun compounds",
+        "non-numeric hyphens"
+      ],
+      "feature_id": "L0B3__module_w7",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at decimal points in numeric measurements with unit conversions.\",\"cues\":[\"decimal point in measurement\",\"unit conversion context\",\"parenthetical alternate units\"],\"exclusions\":[\"hyphenated compound words\",\"dates without decimals\",\"proper names with hyphens\"]},{\"text\":\"Fires at decimal points separating digits in quantities, not at hyphens in compound words.\",\"cues\":[\"digit-dot-digit pattern\",\"quantity or measurement\",\"no hyphen present\"],\"exclusions\":[\"hyphenated adjectives\",\"proper noun compounds\",\"non-numeric hyphens\"]},{\"text\":\"Activates at decimal points within parenthetical unit-conversion expressions.\",\"cues\":[\"parenthetical expression\",\"decimal between digits\",\"unit follows\"],\"exclusions\":[\"hyphen between words\",\"no numeric context\",\"non-measurement hyphens\"]},{\"text\":\"Fires when decimal point separates fractional digits in a measured quantity with nearby unit.\",\"cues\":[\"decimal in number\",\"unit word nearby\",\"measurement context\"],\"exclusions\":[\"word-hyphen-word compounds\",\"dates or years\",\"proper names\"]}]}",
+      "round_index": 0,
+      "text": "Fires at decimal points separating digits in quantities, not at hyphens in compound words."
+    },
+    "own_test_auc": 0.639404296875,
+    "own_validation_auc": 0.6448974609375,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L0B3__module_w7",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.4307692307692308,
+    "interpretation": {
+      "cues": [
+        "hyphenated compound modifier",
+        "proper noun or technical term",
+        "two-token @>-<@ span"
+      ],
+      "exclusions": [
+        "simple hyphenated adjectives",
+        "U-boat style military prefixes",
+        "non-compound hyphen usage"
+      ],
+      "feature_id": "L0B3",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at decimal numbers in measurement or magnitude contexts, especially with @>.<@ decimal markers.\",\"cues\":[\"decimal number in measurement\",\"magnitude or size context\",\"@>.<@ decimal marker\"],\"exclusions\":[\"plain year numbers\",\"plain integer quantities\",\"time-of-day formats\"]},{\"text\":\"Computation fires at hyphenated proper-noun or technical compound modifiers spanning two marked tokens.\",\"cues\":[\"hyphenated compound modifier\",\"proper noun or technical term\",\"two-token @>-<@ span\"],\"exclusions\":[\"simple hyphenated adjectives\",\"U-boat style military prefixes\",\"non-compound hyphen usage\"]},{\"text\":\"Computation fires at split year digits in date ranges or multi-digit year fragments.\",\"cues\":[\"split year digit pairs\",\"date or historical context\",\"two adjacent marked digit groups\"],\"exclusions\":[\"single four-digit years\",\"non-year digit sequences\",\"page or section numbers\"]},{\"text\":\"Computation fires at decimal quantities in geographic, demographic, or distance contexts.\",\"cues\":[\"decimal quantity with units\",\"geographic or demographic context\",\"million or km or miles\"],\"exclusions\":[\"time durations\",\"plain counts without decimals\",\"non-geographic quantities\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires at hyphenated proper-noun or technical compound modifiers spanning two marked tokens."
+    },
+    "semantics_valid": false,
+    "test_auc": 0.59228515625,
+    "validation_auc": 0.7142333984375
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    2510,
+    2290,
+    4320,
+    3180,
+    2206,
+    902,
+    3557
+  ],
+  "top_k": 3
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L0B4.json

@@ -0,0 +1,73 @@
+{
+  "block_feature_id": "L0B4",
+  "certified_width": 9,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 0,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.5750732421875,
+    "own_interpretation": {
+      "cues": [
+        "verb of temporal event",
+        "digit in year position",
+        "preposition 'in' adjacent"
+      ],
+      "exclusions": [
+        "digits in measurements",
+        "digits in identifiers",
+        "digits in rankings"
+      ],
+      "feature_id": "L0B4__module_w9",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Executes when a year or date number follows a preposition like 'in'.\",\"cues\":[\"preposition 'in' before year\",\"4-digit year token\",\"temporal context\"],\"exclusions\":[\"ordinal rankings like 18th\",\"measurements with units\",\"non-date numbers\"]},{\"text\":\"Activates on numeric tokens within temporal phrases indicating event timing.\",\"cues\":[\"temporal preposition nearby\",\"numeric token marked\",\"event verb in context\"],\"exclusions\":[\"spatial preposition phrases\",\"measurement quantities\",\"ranking positions\"]},{\"text\":\"Fires on digits in date expressions following release, completion, or recognition verbs.\",\"cues\":[\"verb of temporal event\",\"digit in year position\",\"preposition 'in' adjacent\"],\"exclusions\":[\"digits in measurements\",\"digits in identifiers\",\"digits in rankings\"]},{\"text\":\"Triggers on numeric spans inside 'in + year' constructions across diverse topics.\",\"cues\":[\"'in' immediately before digits\",\"digits form valid year\",\"sentence-level temporal reference\"],\"exclusions\":[\"'in' before non-year nouns\",\"digits after 'the'\",\"measurement digits with units\"]}]}",
+      "round_index": 0,
+      "text": "Fires on digits in date expressions following release, completion, or recognition verbs."
+    },
+    "own_test_auc": 0.546875,
+    "own_validation_auc": 0.593505859375,
+    "retention_ratio_of_excess_over_chance": 0.39072426937738247
+  },
+  "module_feature_id": "L0B4__module_w9",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 1.5386119257086999,
+    "interpretation": {
+      "cues": [
+        "'in' before 'the'",
+        "temporal or spatial phrase",
+        "narrative context"
+      ],
+      "exclusions": [
+        "'in' before adjective like 'concise'",
+        "'in' before geographic direction"
+      ],
+      "feature_id": "L0B4",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires on definite article 'the' within narrative or descriptive passages containing dates, numbers, or proper nouns.\",\"cues\":[\"definite article 'the'\",\"nearby dates or numbers\",\"proper nouns nearby\"],\"exclusions\":[\"'the' in geographic comparisons\",\"'the' in herd hierarchy context\"]},{\"text\":\"Computation fires on preposition 'in' followed by a definite article or temporal/spatial reference within a sentence.\",\"cues\":[\"'in' before 'the'\",\"temporal or spatial phrase\",\"narrative context\"],\"exclusions\":[\"'in' before adjective like 'concise'\",\"'in' before geographic direction\"]},{\"text\":\"Computation fires on digits within dates, measurements, or identifiers embedded in encyclopedic prose.\",\"cues\":[\"digits in year-like patterns\",\"nearby 'the' or 'in'\",\"encyclopedic prose style\"],\"exclusions\":[\"digits in isolated numeric fragments\",\"digits without nearby articles\"]},{\"text\":\"Computation fires on 'the' in clauses with multiple prepositional phrases or appositives in expository text.\",\"cues\":[\"multiple 'the' in clause\",\"appositive or parenthetical\",\"expository register\"],\"exclusions\":[\"'the' in poetic or quoted speech\",\"'the' in simple comparison\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires on preposition 'in' followed by a definite article or temporal/spatial reference within a sentence."
+    },
+    "semantics_valid": true,
+    "test_auc": 0.692138671875,
+    "validation_auc": 0.6248779296875
+  },
+  "status": "SEMANTICS_NOT_RETAINED",
+  "support": [
+    760,
+    2510,
+    2290,
+    4320,
+    3180,
+    182,
+    2581,
+    2206,
+    902
+  ],
+  "top_k": 4
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L0B5.json

@@ -0,0 +1,74 @@
+{
+  "block_feature_id": "L0B5",
+  "certified_width": 9,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 0,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.856689453125,
+    "own_interpretation": {
+      "cues": [
+        "prose to header transition",
+        "labeled divider appears",
+        "document structure shift"
+      ],
+      "exclusions": [
+        "header to prose transition",
+        "prose continuing prose",
+        "inline quoted labels"
+      ],
+      "feature_id": "L0B5__module_w9",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates at section header transitions, especially when a new Wikipedia-style header follows prior content.\",\"cues\":[\"section header markers\",\"new subsection begins\",\"prior content ends\"],\"exclusions\":[\"mid-sentence text flow\",\"header-like text in body\",\"repeated same-level header\"]},{\"text\":\"Computation fires when a structural boundary is crossed from narrative prose into a labeled section divider.\",\"cues\":[\"prose to header transition\",\"labeled divider appears\",\"document structure shift\"],\"exclusions\":[\"header to prose transition\",\"prose continuing prose\",\"inline quoted labels\"]},{\"text\":\"Computation triggers at the first header after a completed topical paragraph or passage.\",\"cues\":[\"completed paragraph precedes\",\"new topic header follows\",\"topical shift occurs\"],\"exclusions\":[\"header mid-unfinished topic\",\"consecutive headers no prose\",\"header before any prose\"]},{\"text\":\"Computation engages at header tokens that are preceded by a period or closing punctuation ending prior content.\",\"cues\":[\"preceding sentence-final punctuation\",\"header immediately follows\",\"content boundary marked\"],\"exclusions\":[\"header preceded by comma\",\"header mid-sentence\",\"no preceding punctuation\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires when a structural boundary is crossed from narrative prose into a labeled section divider."
+    },
+    "own_test_auc": 0.7969970703125,
+    "own_validation_auc": 0.8408203125,
+    "retention_ratio_of_excess_over_chance": 0.9723793677204658
+  },
+  "module_feature_id": "L0B5__module_w9",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.9518530250237567,
+    "interpretation": {
+      "cues": [
+        "= = = section header",
+        "nested heading markers",
+        "Wikipedia article structure"
+      ],
+      "exclusions": [
+        "inline equals in numbers",
+        "hyphenated compound modifiers",
+        "mid-sentence punctuation"
+      ],
+      "feature_id": "L0B5",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates at Wikipedia section headers, especially nested subsections with multiple equals signs.\",\"cues\":[\"= = = section header\",\"nested heading markers\",\"Wikipedia article structure\"],\"exclusions\":[\"inline equals in numbers\",\"hyphenated compound modifiers\",\"mid-sentence punctuation\"]},{\"text\":\"Computation triggers at structural transition points where new topical sections begin after prior content.\",\"cues\":[\"section boundary transition\",\"new topic introduction\",\"structural document shift\"],\"exclusions\":[\"continuation within same section\",\"inline date or number formatting\",\"mid-paragraph content\"]},{\"text\":\"Computation fires when a heading marker sequence appears at line start, indicating document structure markup.\",\"cues\":[\"line-initial heading markers\",\"markup heading syntax\",\"structural formatting cue\"],\"exclusions\":[\"inline numeric separators\",\"hyphenated name connectors\",\"mid-sentence equals signs\"]}]}",
+      "round_index": 0,
+      "text": "Computation activates at Wikipedia section headers, especially nested subsections with multiple equals signs."
+    },
+    "semantics_valid": true,
+    "test_auc": 0.8668212890625,
+    "validation_auc": 0.8853759765625
+  },
+  "status": "SEMANTICS_RETAINED",
+  "support": [
+    191,
+    760,
+    2510,
+    2290,
+    4320,
+    3180,
+    182,
+    2581,
+    2206
+  ],
+  "top_k": 5
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L0B6.json

@@ -0,0 +1,75 @@
+{
+  "block_feature_id": "L0B6",
+  "certified_width": 10,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 0,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.8663330078125,
+    "own_interpretation": {
+      "cues": [
+        "section title word",
+        "equals signs delimiter",
+        "topic shift marker"
+      ],
+      "exclusions": [
+        "prose continuation",
+        "inline notation",
+        "no heading word"
+      ],
+      "feature_id": "L0B6__module_w10",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at Wikipedia section headers (lines of = signs) following sentence-final punctuation.\",\"cues\":[\"line of = signs\",\"preceding sentence period\",\"section heading structure\"],\"exclusions\":[\"inline = in prose\",\"no preceding punctuation\",\"mid-sentence position\"]},{\"text\":\"Computation fires at section headers introducing new topical subsections in articles.\",\"cues\":[\"section title word\",\"equals signs delimiter\",\"topic shift marker\"],\"exclusions\":[\"prose continuation\",\"inline notation\",\"no heading word\"]},{\"text\":\"Computation fires at structural boundaries marked by repeated equals signs after completed sentences.\",\"cues\":[\"equals sign line\",\"sentence boundary before\",\"new topic introduction\"],\"exclusions\":[\"mid-paragraph text\",\"inline hyphenation\",\"no equals signs\"]},{\"text\":\"Computation fires at Wikipedia-style heading delimiters separating article sections.\",\"cues\":[\"multiple equals signs\",\"heading keyword present\",\"post-sentence position\"],\"exclusions\":[\"prose without headings\",\"inline dash notation\",\"continuous narrative\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires at section headers introducing new topical subsections in articles."
+    },
+    "own_test_auc": 0.8616943359375,
+    "own_validation_auc": 0.8770751953125,
+    "retention_ratio_of_excess_over_chance": 0.9168958142377024
+  },
+  "module_feature_id": "L0B6__module_w10",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 1.043353522473701,
+    "interpretation": {
+      "cues": [
+        "topical section boundaries",
+        "organizational markup tokens",
+        "content category shifts"
+      ],
+      "exclusions": [
+        "content within sections",
+        "footnote markers",
+        "citation boundaries"
+      ],
+      "feature_id": "L0B6",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at section header markers like equals signs\",\"cues\":[\"equals sign section headers\",\"wiki section markup\",\"heading delimiter patterns\"],\"exclusions\":[\"narrative body text\",\"inline equals signs\",\"mathematical equations\"]},{\"text\":\"Activates at structural transitions between subsections in articles\",\"cues\":[\"subsection boundary transitions\",\"nested heading levels\",\"article structure shifts\"],\"exclusions\":[\"mid-paragraph transitions\",\"sentence-level transitions\",\"topic shifts within paragraphs\"]},{\"text\":\"Fires when encountering hierarchical heading depth changes\",\"cues\":[\"heading depth increases\",\"heading depth decreases\",\"level transitions in structure\"],\"exclusions\":[\"same-level heading continuations\",\"non-heading structural elements\",\"list item transitions\"]},{\"text\":\"Triggers on document organization markers indicating new topical sections\",\"cues\":[\"topical section boundaries\",\"organizational markup tokens\",\"content category shifts\"],\"exclusions\":[\"content within sections\",\"footnote markers\",\"citation boundaries\"]}]}",
+      "round_index": 0,
+      "text": "Triggers on document organization markers indicating new topical sections"
+    },
+    "semantics_valid": true,
+    "test_auc": 0.8995361328125,
+    "validation_auc": 0.8829345703125
+  },
+  "status": "SEMANTICS_RETAINED",
+  "support": [
+    191,
+    760,
+    2510,
+    2581,
+    2290,
+    4320,
+    3180,
+    182,
+    2144,
+    2206
+  ],
+  "top_k": 6
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L0B7.json

@@ -0,0 +1,77 @@
+{
+  "block_feature_id": "L0B7",
+  "certified_width": 12,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 0,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.84326171875,
+    "own_interpretation": {
+      "cues": [
+        "completed sentence before",
+        "new heading begins",
+        "section boundary transition"
+      ],
+      "exclusions": [
+        "heading at article start",
+        "heading inside paragraph",
+        "no preceding sentence"
+      ],
+      "feature_id": "L0B7__module_w12",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at section header tokens following a prior section's body text.\",\"cues\":[\"wiki section header markup\",\"preceding body sentence text\",\"transition between sections\"],\"exclusions\":[\"header at document start\",\"header with no preceding text\",\"inline non-header markup\"]},{\"text\":\"Computation fires at heading tokens that begin a new topical subsection after completed prior content.\",\"cues\":[\"new subsection heading\",\"prior content completed\",\"topical shift marker\"],\"exclusions\":[\"heading mid-sentence\",\"heading with no prior context\",\"inline emphasis markup\"]},{\"text\":\"Computation fires at section boundary tokens where a completed sentence precedes a new heading.\",\"cues\":[\"completed sentence before\",\"new heading begins\",\"section boundary transition\"],\"exclusions\":[\"heading at article start\",\"heading inside paragraph\",\"no preceding sentence\"]},{\"text\":\"Computation fires at heading tokens preceded by body text from a different section.\",\"cues\":[\"preceding different section text\",\"heading token follows\",\"section transition point\"],\"exclusions\":[\"heading at article beginning\",\"consecutive headings only\",\"heading after heading\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires at section boundary tokens where a completed sentence precedes a new heading."
+    },
+    "own_test_auc": 0.68603515625,
+    "own_validation_auc": 0.8763427734375,
+    "retention_ratio_of_excess_over_chance": 0.9118028534370947
+  },
+  "module_feature_id": "L0B7__module_w12",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.9796696315120712,
+    "interpretation": {
+      "cues": [
+        "preceding body text",
+        "following new topic sentence",
+        "heading delimiter pattern"
+      ],
+      "exclusions": [
+        "heading without preceding text",
+        "consecutive headings only",
+        "heading mid-paragraph"
+      ],
+      "feature_id": "L0B7",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at section header markers with equals signs\",\"cues\":[\"equals sign section headers\",\"wiki heading markup\",\"structural transition point\"],\"exclusions\":[\"inline equals signs\",\"mathematical expressions\",\"mid-sentence punctuation\"]},{\"text\":\"Activates on Wikipedia-style heading transitions between content blocks\",\"cues\":[\"preceding body text\",\"following new topic sentence\",\"heading delimiter pattern\"],\"exclusions\":[\"heading without preceding text\",\"consecutive headings only\",\"heading mid-paragraph\"]},{\"text\":\"Triggers at structural boundaries marked by repeated equals delimiters\",\"cues\":[\"2+ consecutive equals signs\",\"topic shift indicator\",\"document structure marker\"],\"exclusions\":[\"single equals sign\",\"comparison operators\",\"attribution dashes\"]},{\"text\":\"Fires when heading markup follows completed sentence or clause\",\"cues\":[\"sentence-final punctuation before\",\"heading markup after\",\"content boundary transition\"],\"exclusions\":[\"heading after fragment\",\"heading after comma\",\"inline section reference\"]}]}",
+      "round_index": 0,
+      "text": "Activates on Wikipedia-style heading transitions between content blocks"
+    },
+    "semantics_valid": true,
+    "test_auc": 0.87646484375,
+    "validation_auc": 0.88427734375
+  },
+  "status": "SEMANTICS_RETAINED",
+  "support": [
+    191,
+    760,
+    2510,
+    2581,
+    2290,
+    3557,
+    4320,
+    2206,
+    902,
+    3180,
+    182,
+    1615
+  ],
+  "top_k": 7
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L11B1.json

@@ -0,0 +1,65 @@
+{
+  "block_feature_id": "L11B1",
+  "certified_width": 1,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 11,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "coordinate named entities",
+        "list separator boundary",
+        "entity conjunction"
+      ],
+      "exclusions": [
+        "narrative action sentences",
+        "non-entity coordination"
+      ],
+      "feature_id": "L11B1__module_w1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at morpheme boundaries where suffixes or inflections attach to proper names or lexical stems.\",\"cues\":[\"proper name stem\",\"suffix boundary marker\",\"inflectional ending\"],\"exclusions\":[\"common noun phrases\",\"generic sentence endings\"]},{\"text\":\"Activates at multi-word proper noun boundaries, especially where titles or descriptors follow names.\",\"cues\":[\"multi-word proper noun\",\"title or descriptor\",\"name continuation token\"],\"exclusions\":[\"single common nouns\",\"generic location phrases\"]},{\"text\":\"Fires at conjunction/punctuation boundaries between named entities in coordinate lists.\",\"cues\":[\"coordinate named entities\",\"list separator boundary\",\"entity conjunction\"],\"exclusions\":[\"narrative action sentences\",\"non-entity coordination\"]},{\"text\":\"Activates at compound noun boundaries where second element is a known geographic or institutional term.\",\"cues\":[\"compound noun head\",\"geographic or institutional term\",\"proper noun modifier\"],\"exclusions\":[\"simple common nouns\",\"verb phrase boundaries\"]}]}",
+      "round_index": 0,
+      "text": "Fires at conjunction/punctuation boundaries between named entities in coordinate lists."
+    },
+    "own_test_auc": 0.4578857421875,
+    "own_validation_auc": 0.551513671875,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L11B1__module_w1",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 5200195.3125,
+    "interpretation": {
+      "cues": [
+        "@-@ connector in proper noun",
+        "hyphenated place name",
+        "compound geographic name"
+      ],
+      "exclusions": [
+        "standalone punctuation",
+        "sentence-final period",
+        "comma after common noun"
+      ],
+      "feature_id": "L11B1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at morpheme or word-internal boundaries where suffixes or inflectional endings attach to roots.\",\"cues\":[\"word-internal morpheme boundary\",\"suffix attaching to root\",\"inflectional ending\"],\"exclusions\":[\"word-final punctuation boundary\",\"sentence-level boundary\",\"clause boundary\"]},{\"text\":\"Activates at hyphen/at-symbol connectors within proper nouns or compound place names.\",\"cues\":[\"@-@ connector in proper noun\",\"hyphenated place name\",\"compound geographic name\"],\"exclusions\":[\"standalone punctuation\",\"sentence-final period\",\"comma after common noun\"]},{\"text\":\"Fires at boundaries between a proper name and following coordinating conjunction in lists.\",\"cues\":[\"proper name before conjunction\",\"list of proper nouns\",\"coordination in name list\"],\"exclusions\":[\"common noun before conjunction\",\"clause-level coordination\",\"sentence-initial conjunction\"]},{\"text\":\"Activates at word-internal boundaries in multi-syllabic words with embedded shorter words or morphemes.\",\"cues\":[\"embedded word within longer word\",\"syllable boundary in long word\",\"recognizable substring in token\"],\"exclusions\":[\"short common word boundary\",\"function word sequence\",\"single-syllable word\"]}]}",
+      "round_index": 0,
+      "text": "Activates at hyphen/at-symbol connectors within proper nouns or compound place names."
+    },
+    "semantics_valid": false,
+    "test_auc": 0.552001953125,
+    "validation_auc": 0.4993896484375
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    2506
+  ],
+  "top_k": 1
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L11B2.json

@@ -0,0 +1,69 @@
+{
+  "block_feature_id": "L11B2",
+  "certified_width": 4,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 11,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.5704345703125,
+    "own_interpretation": {
+      "cues": [
+        "punctuation within entity name",
+        "designation codes like Tu-2",
+        "coordinate proper noun lists"
+      ],
+      "exclusions": [
+        "punctuation in narrative prose",
+        "common noun punctuation",
+        "dialogue punctuation"
+      ],
+      "feature_id": "L11B2__module_w4",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates at morpheme or token boundaries where hyphenated compounds, numeric spans, or multi-word proper nouns are being assembled.\",\"cues\":[\"hyphenated compound tokens\",\"numeric span boundaries\",\"proper noun multi-token sequences\"],\"exclusions\":[\"single common nouns\",\"generic adjectives\",\"non-boundary mid-token positions\"]},{\"text\":\"Activates at punctuation-adjacent positions within multi-token entity names or technical designations.\",\"cues\":[\"punctuation within entity name\",\"designation codes like Tu-2\",\"coordinate proper noun lists\"],\"exclusions\":[\"punctuation in narrative prose\",\"common noun punctuation\",\"dialogue punctuation\"]},{\"text\":\"Fires at token transitions where the next token completes a known named entity, place, or title.\",\"cues\":[\"entity-initial token followed by completion\",\"place name continuation\",\"title or designation completion\"],\"exclusions\":[\"common noun continuations\",\"verb phrase continuations\",\"generic adjective continuations\"]},{\"text\":\"Activates at boundaries between tokens forming compound modifiers, coordinate structures, or prefixed words.\",\"cues\":[\"compound modifier boundaries\",\"coordinate conjunction boundaries\",\"prefix-stem boundaries\"],\"exclusions\":[\"simple subject-verb boundaries\",\"single modifier-noun pairs\",\"adverb-verb boundaries\"]}]}",
+      "round_index": 0,
+      "text": "Activates at punctuation-adjacent positions within multi-token entity names or technical designations."
+    },
+    "own_test_auc": 0.73583984375,
+    "own_validation_auc": 0.6107177734375,
+    "retention_ratio_of_excess_over_chance": 0.5464015151515151
+  },
+  "module_feature_id": "L11B2__module_w4",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 1.3101736972704714,
+    "interpretation": {
+      "cues": [
+        "comma in large numbers",
+        "decimal point in measurements",
+        "digit before unit token"
+      ],
+      "exclusions": [
+        "century ordinal numbers",
+        "percentage after digits",
+        "year digit sequences"
+      ],
+      "feature_id": "L11B2",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Executes at token boundaries within multi-word proper nouns and hyphenated compounds, especially geographic or institutional names.\",\"cues\":[\"hyphenated proper noun spans\",\"multi-token place names\",\"institutional name continuations\"],\"exclusions\":[\"single common nouns\",\"present-day temporal phrases\",\"adjectival modifiers\"]},{\"text\":\"Activates at numeric formatting boundaries: thousands separators, decimal points, and digit-to-unit transitions.\",\"cues\":[\"comma in large numbers\",\"decimal point in measurements\",\"digit before unit token\"],\"exclusions\":[\"century ordinal numbers\",\"percentage after digits\",\"year digit sequences\"]},{\"text\":\"Fires at morpheme boundaries in inflected proper nouns and compound word splits across tokens.\",\"cues\":[\"inflected proper noun suffixes\",\"compound word token splits\",\"possessive or plural on names\"],\"exclusions\":[\"common noun plurals\",\"verb inflection boundaries\",\"adjective comparative forms\"]},{\"text\":\"Triggers at conjunction/punctuation boundaries following named entities or large numeric values.\",\"cues\":[\"comma after proper noun\",\"conjunction after number\",\"closing quote after entity\"],\"exclusions\":[\"comma after common noun\",\"conjunction in narrative text\",\"punctuation in dialogue\"]}]}",
+      "round_index": 0,
+      "text": "Activates at numeric formatting boundaries: thousands separators, decimal points, and digit-to-unit transitions."
+    },
+    "semantics_valid": true,
+    "test_auc": 0.62890625,
+    "validation_auc": 0.598388671875
+  },
+  "status": "SEMANTICS_RETAINED",
+  "support": [
+    2506,
+    2308,
+    3599,
+    3706
+  ],
+  "top_k": 2
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L11B3.json

@@ -0,0 +1,70 @@
+{
+  "block_feature_id": "L11B3",
+  "certified_width": 5,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 11,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "numeric delimiter boundary",
+        "digit-grouping comma",
+        "hyphenated measurement"
+      ],
+      "exclusions": [
+        "plain digit sequences",
+        "year-only numbers",
+        "percentages without delimiters"
+      ],
+      "feature_id": "L11B3__module_w5",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at token boundaries within multi-word proper nouns or hyphenated place names.\",\"cues\":[\"proper noun continuation\",\"hyphenated place name\",\"multi-word entity span\"],\"exclusions\":[\"common noun phrases\",\"generic geographic terms\",\"single-word place names\"]},{\"text\":\"Activates at boundaries inside numeric expressions with delimiters like commas or hyphens.\",\"cues\":[\"numeric delimiter boundary\",\"digit-grouping comma\",\"hyphenated measurement\"],\"exclusions\":[\"plain digit sequences\",\"year-only numbers\",\"percentages without delimiters\"]},{\"text\":\"Fires at morpheme boundaries within hyphenated or compound technical terms.\",\"cues\":[\"compound term boundary\",\"hyphenated technical term\",\"morpheme split across tokens\"],\"exclusions\":[\"simple adjectives\",\"non-compound common nouns\",\"regular inflections\"]},{\"text\":\"Activates at boundaries within named entities split across multiple subword tokens.\",\"cues\":[\"subword-split entity name\",\"person name token boundary\",\"ship or vessel name\"],\"exclusions\":[\"common noun subword splits\",\"regular verb inflections\",\"pronoun boundaries\"]}]}",
+      "round_index": 0,
+      "text": "Activates at boundaries inside numeric expressions with delimiters like commas or hyphens."
+    },
+    "own_test_auc": 0.4764404296875,
+    "own_validation_auc": 0.62060546875,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L11B3__module_w5",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 1.631268436578171,
+    "interpretation": {
+      "cues": [
+        "numeric with comma/decimal",
+        "list or enumeration context",
+        "quantity expression"
+      ],
+      "exclusions": [
+        "standalone year numbers",
+        "isolated digit in tech specs",
+        "plain integer without comma"
+      ],
+      "feature_id": "L11B3",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at token boundaries following proper nouns, place names, or titles, especially mid-sentence.\",\"cues\":[\"proper noun preceding boundary\",\"place name context\",\"mid-sentence position\"],\"exclusions\":[\"common noun boundaries\",\"sentence-final position\",\"generic activity words\"]},{\"text\":\"Fires at boundaries after numeric expressions with commas or decimal formatting in lists.\",\"cues\":[\"numeric with comma/decimal\",\"list or enumeration context\",\"quantity expression\"],\"exclusions\":[\"standalone year numbers\",\"isolated digit in tech specs\",\"plain integer without comma\"]},{\"text\":\"Activates at boundaries within multi-word named entities, especially hyphenated or compound place names.\",\"cues\":[\"compound entity internal boundary\",\"hyphenated name components\",\"entity continuation tokens\"],\"exclusions\":[\"common adjective-noun pairs\",\"generic compound nouns\",\"non-entity hyphenated words\"]},{\"text\":\"Fires at boundaries after titles, ranks, or honorifics followed by personal names.\",\"cues\":[\"title or rank token\",\"personal name follows\",\"honorific abbreviation\"],\"exclusions\":[\"common noun before common noun\",\"role description without name\",\"generic occupational label\"]}]}",
+      "round_index": 0,
+      "text": "Fires at boundaries after numeric expressions with commas or decimal formatting in lists."
+    },
+    "semantics_valid": false,
+    "test_auc": 0.5675048828125,
+    "validation_auc": 0.5413818359375
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    2506,
+    2308,
+    2965,
+    3599,
+    3706
+  ],
+  "top_k": 3
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L11B4.json

@@ -0,0 +1,71 @@
+{
+  "block_feature_id": "L11B4",
+  "certified_width": 6,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 11,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "place name subword boundary",
+        "entity name internal token",
+        "geographic name fragment"
+      ],
+      "exclusions": [
+        "common noun subword boundary",
+        "generic location reference",
+        "single-token place name"
+      ],
+      "feature_id": "L11B4__module_w6",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at token boundaries inside multi-token numeric expressions with punctuation separators\",\"cues\":[\"numeric token with @ separator\",\"digit token after comma/period\",\"number span with internal punctuation\"],\"exclusions\":[\"single standalone number token\",\"numeric token in non-numeric context\",\"isolated digit without separator\"]},{\"text\":\"Computation activates at morpheme boundaries within hyphenated or compound proper names\",\"cues\":[\"proper name internal morpheme\",\"hyphenated compound token boundary\",\"multi-token proper name segment\"],\"exclusions\":[\"simple single-token proper name\",\"common noun compound\",\"non-hyphenated regular word\"]},{\"text\":\"Computation triggers at subword boundaries in multi-token place or entity names\",\"cues\":[\"place name subword boundary\",\"entity name internal token\",\"geographic name fragment\"],\"exclusions\":[\"common noun subword boundary\",\"generic location reference\",\"single-token place name\"]},{\"text\":\"Computation fires at digit tokens that are part of larger multi-digit year or quantity spans\",\"cues\":[\"digit within year span\",\"digit in multi-digit number\",\"numeric digit in quantity expression\"],\"exclusions\":[\"standalone single digit\",\"digit in date expression\",\"ordinal number token\"]}]}",
+      "round_index": 0,
+      "text": "Computation triggers at subword boundaries in multi-token place or entity names"
+    },
+    "own_test_auc": 0.5836181640625,
+    "own_validation_auc": 0.6455078125,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L11B4__module_w6",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.10121457489878542,
+    "interpretation": {
+      "cues": [
+        "proper noun lists",
+        "coordinate conjunctions",
+        "parallel entity names"
+      ],
+      "exclusions": [
+        "conjunctions in prose",
+        "simple sentence conjunctions",
+        "narrative conjunctions"
+      ],
+      "feature_id": "L11B4",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates at morpheme boundaries in hyphenated or compound proper nouns, especially multi-word entities.\",\"cues\":[\"hyphenated proper nouns\",\"compound entity names\",\"multi-word place names\"],\"exclusions\":[\"simple hyphenated common words\",\"dates with hyphens\",\"single proper nouns\"]},{\"text\":\"Activates at number-punctuation boundaries in large numeric quantities with thousands separators.\",\"cues\":[\"large numbers with commas\",\"numeric quantities\",\"measurement values\"],\"exclusions\":[\"years and dates\",\"small integers\",\"decimal numbers\"]},{\"text\":\"Activates at syllable boundaries within multi-syllable proper nouns, especially foreign or unusual names.\",\"cues\":[\"foreign proper names\",\"uncommon syllable patterns\",\"biographical name segments\"],\"exclusions\":[\"common English words\",\"regular English names\",\"generic nouns\"]},{\"text\":\"Activates at conjunction boundaries linking parallel proper noun lists or coordinate phrases.\",\"cues\":[\"proper noun lists\",\"coordinate conjunctions\",\"parallel entity names\"],\"exclusions\":[\"conjunctions in prose\",\"simple sentence conjunctions\",\"narrative conjunctions\"]}]}",
+      "round_index": 0,
+      "text": "Activates at conjunction boundaries linking parallel proper noun lists or coordinate phrases."
+    },
+    "semantics_valid": false,
+    "test_auc": 0.51220703125,
+    "validation_auc": 0.62060546875
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    2506,
+    2308,
+    2965,
+    4052,
+    3599,
+    1911
+  ],
+  "top_k": 4
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L11B5.json

@@ -0,0 +1,72 @@
+{
+  "block_feature_id": "L11B5",
+  "certified_width": 7,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 11,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "digit sequences in identifiers",
+        "scientific notation tokens",
+        "measurement unit tokens"
+      ],
+      "exclusions": [
+        "plain year dates",
+        "unstructured numbers",
+        "counts of common objects"
+      ],
+      "feature_id": "L11B5__module_w7",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires on multi-token proper nouns, compound names, and hyphenated entities being completed.\",\"cues\":[\"proper noun continuation\",\"hyphenated entity tokens\",\"compound name completion\"],\"exclusions\":[\"common noun phrases\",\"generic descriptive text\",\"single common words\"]},{\"text\":\"Fires on numeric or alphanumeric identifiers like issue numbers, coordinates, and measurements.\",\"cues\":[\"digit sequences in identifiers\",\"scientific notation tokens\",\"measurement unit tokens\"],\"exclusions\":[\"plain year dates\",\"unstructured numbers\",\"counts of common objects\"]},{\"text\":\"Activates on taxonomic or scientific name fragments and author citations in lists.\",\"cues\":[\"species name fragments\",\"authority name citations\",\"repeated name patterns\"],\"exclusions\":[\"narrative prose\",\"single author mentions\",\"non-scientific names\"]},{\"text\":\"Fires on structural or formatting tokens like section headers, punctuation patterns, and list separators.\",\"cues\":[\"heading marker sequences\",\"punctuation-heavy spans\",\"list item formatting\"],\"exclusions\":[\"normal sentence punctuation\",\"flowing narrative text\",\"dialogue passages\"]}]}",
+      "round_index": 0,
+      "text": "Fires on numeric or alphanumeric identifiers like issue numbers, coordinates, and measurements."
+    },
+    "own_test_auc": 0.494140625,
+    "own_validation_auc": 0.603271484375,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L11B5__module_w7",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.4680483592400691,
+    "interpretation": {
+      "cues": [
+        "hyphenated place name",
+        "punctuated proper noun",
+        "multi-token entity"
+      ],
+      "exclusions": [
+        "single common nouns",
+        "generic noun phrases",
+        "plain geographic names"
+      ],
+      "feature_id": "L11B5",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires on numeric tokens embedded in proper-noun or technical designations.\",\"cues\":[\"digit within named entity\",\"alphanumeric model/designation\",\"number in title/code\"],\"exclusions\":[\"standalone sports scores\",\"plain counts of items\",\"dates alone\"]},{\"text\":\"Activates on multi-word proper-noun spans containing internal punctuation or hyphenation.\",\"cues\":[\"hyphenated place name\",\"punctuated proper noun\",\"multi-token entity\"],\"exclusions\":[\"single common nouns\",\"generic noun phrases\",\"plain geographic names\"]},{\"text\":\"Fires on numeric sub-tokens within larger alphanumeric identifiers or version strings.\",\"cues\":[\"digit in identifier\",\"alphanumeric code token\",\"version/designation number\"],\"exclusions\":[\"isolated measurements\",\"plain quantities\",\"standalone years\"]},{\"text\":\"Activates on digit tokens that are part of chart-ranking or monetary-amount constructions.\",\"cues\":[\"digit in chart position\",\"number in monetary figure\",\"digit in relief-supply count\"],\"exclusions\":[\"rugby match scores\",\"plain age ranges\",\"generic numeric quantities\"]}]}",
+      "round_index": 0,
+      "text": "Activates on multi-word proper-noun spans containing internal punctuation or hyphenation."
+    },
+    "semantics_valid": false,
+    "test_auc": 0.5330810546875,
+    "validation_auc": 0.5706787109375
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    2308,
+    2506,
+    3599,
+    2965,
+    4052,
+    3900,
+    2594
+  ],
+  "top_k": 5
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L11B6.json

@@ -0,0 +1,72 @@
+{
+  "block_feature_id": "L11B6",
+  "certified_width": 7,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 11,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "mid-word token boundary",
+        "proper noun completion",
+        "compound term formation"
+      ],
+      "exclusions": [
+        "single common nouns",
+        "sentence-initial tokens",
+        "generic adjectives"
+      ],
+      "feature_id": "L11B6__module_w7",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at token boundaries within multi-token proper nouns or compound terms being completed.\",\"cues\":[\"mid-word token boundary\",\"proper noun completion\",\"compound term formation\"],\"exclusions\":[\"single common nouns\",\"sentence-initial tokens\",\"generic adjectives\"]},{\"text\":\"Activates at subword splits inside numbers, dates, or hyphenated spans.\",\"cues\":[\"numeric token splits\",\"date/year fragments\",\"hyphenated spans\"],\"exclusions\":[\"standalone numbers\",\"month names alone\",\"non-numeric common words\"]},{\"text\":\"Fires at morphological suffix boundaries in plural or inflected forms.\",\"cues\":[\"plural suffix split\",\"inflection boundary\",\"derivational suffix\"],\"exclusions\":[\"uninflected base words\",\"irregular plurals\",\"compound suffixes\"]},{\"text\":\"Activates at wiki markup section headers and list structure boundaries.\",\"cues\":[\"section header markup\",\"list item markers\",\"heading equals signs\"],\"exclusions\":[\"inline punctuation\",\"prose commas\",\"quotation marks\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires at token boundaries within multi-token proper nouns or compound terms being completed."
+    },
+    "own_test_auc": 0.6099853515625,
+    "own_validation_auc": 0.555908203125,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L11B6__module_w7",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.15681983953318746,
+    "interpretation": {
+      "cues": [
+        "list or header formatting",
+        "enumerated sequences",
+        "section markers like ==="
+      ],
+      "exclusions": [
+        "flowing narrative sentences",
+        "isolated clauses without structure",
+        "unstructured dialogue"
+      ],
+      "feature_id": "L11B6",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at token boundaries within multi-token proper nouns, compound names, or hyphenated entities.\",\"cues\":[\"multi-token proper noun spans\",\"hyphenated or compound entities\",\"taxonomic or place names\"],\"exclusions\":[\"single common nouns\",\"generic verbs or adjectives\",\"isolated punctuation\"]},{\"text\":\"Activates at subword boundaries where tokens reconstruct technical, geographic, or biographical entity names.\",\"cues\":[\"subword-boundary entity reconstruction\",\"technical or scientific terminology\",\"biographical name fragments\"],\"exclusions\":[\"common single-token words\",\"generic narrative prose\",\"standalone numbers\"]},{\"text\":\"Fires when tokens form part of structured lists, section headers, or enumerated/sequential content.\",\"cues\":[\"list or header formatting\",\"enumerated sequences\",\"section markers like ===\"],\"exclusions\":[\"flowing narrative sentences\",\"isolated clauses without structure\",\"unstructured dialogue\"]},{\"text\":\"Activates at boundaries of multi-token place names, dates, or numerical expressions with separators.\",\"cues\":[\"multi-token place names\",\"dates with numerical subwords\",\"numbers with comma or hyphen separators\"],\"exclusions\":[\"single-token dates\",\"plain numbers without separators\",\"common nouns without entity structure\"]}]}",
+      "round_index": 0,
+      "text": "Fires when tokens form part of structured lists, section headers, or enumerated/sequential content."
+    },
+    "semantics_valid": false,
+    "test_auc": 0.5262451171875,
+    "validation_auc": 0.6673583984375
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    3706,
+    2308,
+    2506,
+    3599,
+    2965,
+    4052,
+    3900
+  ],
+  "top_k": 6
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L12B1.json

@@ -0,0 +1,71 @@
+{
+  "block_feature_id": "L12B1",
+  "certified_width": 6,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 12,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "release date construction",
+        "media publication context",
+        "month-day-year sequence"
+      ],
+      "exclusions": [
+        "non-publication dates",
+        "general temporal phrases",
+        "century references"
+      ],
+      "feature_id": "L12B1__module_w6",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at sentence-initial temporal/date expressions with day-month-year format.\",\"cues\":[\"sentence-initial date pattern\",\"day number before month\",\"year digits split across tokens\"],\"exclusions\":[\"mid-sentence dates\",\"dates without day numbers\",\"non-temporal sentence starts\"]},{\"text\":\"Computation activates at clause boundaries introducing new narrative episodes or events.\",\"cues\":[\"clause-initial position\",\"new event introduced\",\"narrative continuation marker\"],\"exclusions\":[\"mid-clause descriptive content\",\"list items\",\"general factual statements\"]},{\"text\":\"Computation fires at numeric year tokens within biographical or historical event contexts.\",\"cues\":[\"year number in biography\",\"event date in history\",\"split digit tokens in year\"],\"exclusions\":[\"quantities not years\",\"century references\",\"non-date numbers\"]},{\"text\":\"Computation activates at publication/release date constructions in media or album articles.\",\"cues\":[\"release date construction\",\"media publication context\",\"month-day-year sequence\"],\"exclusions\":[\"non-publication dates\",\"general temporal phrases\",\"century references\"]}]}",
+      "round_index": 0,
+      "text": "Computation activates at publication/release date constructions in media or album articles."
+    },
+    "own_test_auc": 0.63134765625,
+    "own_validation_auc": 0.5675048828125,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L12B1__module_w6",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.3474218089602705,
+    "interpretation": {
+      "cues": [
+        "release or publication verb",
+        "date or year nearby",
+        "album or title context"
+      ],
+      "exclusions": [
+        "non-publication events",
+        "dates without release context",
+        "fictional plot events"
+      ],
+      "feature_id": "L12B1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at sentence-initial temporal or date expressions marking narrative transitions.\",\"cues\":[\"sentence-initial date or time\",\"narrative transition marker\",\"temporal preposition at boundary\"],\"exclusions\":[\"mid-sentence date mentions\",\"non-temporal sentence starts\",\"dates embedded in clauses\"]},{\"text\":\"Computation fires at clause boundaries introducing background or historical context.\",\"cues\":[\"clause boundary punctuation\",\"historical or background context\",\"definite article after boundary\"],\"exclusions\":[\"dialogue or fictional narrative\",\"conditional or hypothetical clauses\",\"action-sequence boundaries\"]},{\"text\":\"Computation fires at boundaries introducing geographic or political entities in expository text.\",\"cues\":[\"geographic or political entity\",\"expository listing structure\",\"definite article before entity\"],\"exclusions\":[\"fictional character actions\",\"conditional reasoning\",\"narrative action sequences\"]},{\"text\":\"Computation fires at boundaries marking publication or release events with dates.\",\"cues\":[\"release or publication verb\",\"date or year nearby\",\"album or title context\"],\"exclusions\":[\"non-publication events\",\"dates without release context\",\"fictional plot events\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires at boundaries marking publication or release events with dates."
+    },
+    "semantics_valid": false,
+    "test_auc": 0.5501708984375,
+    "validation_auc": 0.6444091796875
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    2646,
+    4299,
+    2404,
+    3036,
+    3209,
+    1892
+  ],
+  "top_k": 1
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L13B1.json

@@ -0,0 +1,65 @@
+{
+  "block_feature_id": "L13B1",
+  "certified_width": 1,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 13,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.5941162109375,
+    "own_interpretation": {
+      "cues": [
+        "descriptive context precedes",
+        "new entity follows",
+        "narrative transition point"
+      ],
+      "exclusions": [
+        "same-entity continuation",
+        "generic transition word",
+        "within same topic"
+      ],
+      "feature_id": "L13B1__module_w1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at sentence-final positions before a new topic or section begins.\",\"cues\":[\"sentence-final punctuation\",\"preceding section header\",\"topic shift boundary\"],\"exclusions\":[\"mid-sentence clause boundary\",\"within-list enumeration\",\"parenthetical aside end\"]},{\"text\":\"Computation fires at end of a paragraph or list item before structural transition.\",\"cues\":[\"paragraph end\",\"list item boundary\",\"section transition nearby\"],\"exclusions\":[\"mid-paragraph sentence\",\"within running prose\",\"inside parenthetical\"]},{\"text\":\"Computation fires at end of a clause introducing a subject before elaboration follows.\",\"cues\":[\"subject-introducing clause\",\"elaboration follows immediately\",\"entity name at boundary\"],\"exclusions\":[\"action description clause\",\"generic pronoun boundary\",\"mid-sentence conjunction\"]},{\"text\":\"Computation fires at boundary between descriptive context and new entity introduction.\",\"cues\":[\"descriptive context precedes\",\"new entity follows\",\"narrative transition point\"],\"exclusions\":[\"same-entity continuation\",\"generic transition word\",\"within same topic\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires at boundary between descriptive context and new entity introduction."
+    },
+    "own_test_auc": 0.6282958984375,
+    "own_validation_auc": 0.8265380859375,
+    "retention_ratio_of_excess_over_chance": 0.225900966891298
+  },
+  "module_feature_id": "L13B1__module_w1",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 1.4056836902800658,
+    "interpretation": {
+      "cues": [
+        "descriptive clause end",
+        "sequential marker follows",
+        "contrast follows"
+      ],
+      "exclusions": [
+        "mid-description",
+        "no transition follows"
+      ],
+      "feature_id": "L13B1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Executes at sentence-final positions before new topic or section transitions\",\"cues\":[\"sentence-final punctuation\",\"section header follows\",\"topic shift\"],\"exclusions\":[\"mid-sentence continuation\",\"same topic maintained\"]},{\"text\":\"Activates at clause boundaries where narrative shifts to new event or entity\",\"cues\":[\"clause boundary\",\"new entity introduced\",\"event transition\"],\"exclusions\":[\"within ongoing description\",\"elaboration of same subject\"]},{\"text\":\"Fires at end of descriptive clause before contrasting or sequential content\",\"cues\":[\"descriptive clause end\",\"sequential marker follows\",\"contrast follows\"],\"exclusions\":[\"mid-description\",\"no transition follows\"]},{\"text\":\"Triggers at period before coordinate clause introducing new subject\",\"cues\":[\"period before clause\",\"new subject follows\",\"coordinate structure\"],\"exclusions\":[\"subordinate clause continuation\",\"same subject maintained\"]}]}",
+      "round_index": 0,
+      "text": "Fires at end of descriptive clause before contrasting or sequential content"
+    },
+    "semantics_valid": true,
+    "test_auc": 0.9166259765625,
+    "validation_auc": 0.79638671875
+  },
+  "status": "SEMANTICS_NOT_RETAINED",
+  "support": [
+    1757
+  ],
+  "top_k": 1
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L13B2.json

@@ -0,0 +1,68 @@
+{
+  "block_feature_id": "L13B2",
+  "certified_width": 3,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 13,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.66748046875,
+    "own_interpretation": {
+      "cues": [
+        "text after section header",
+        "text after topic shift",
+        "text after formatting break"
+      ],
+      "exclusions": [
+        "text within same section",
+        "text mid-paragraph",
+        "text after inline clause"
+      ],
+      "feature_id": "L13B2__module_w3",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at section headers or structural transitions marked by formatting cues.\",\"cues\":[\"equals signs formatting\",\"section header markers\",\"structural transition\"],\"exclusions\":[\"mid-sentence dates\",\"inline parentheticals\",\"chart positions\"]},{\"text\":\"Computation activates at year tokens following sentence-final punctuation or clause boundaries.\",\"cues\":[\"year after period\",\"year after comma clause\",\"year starting new sentence\"],\"exclusions\":[\"year mid-sentence\",\"year after preposition\",\"year in parenthetical\"]},{\"text\":\"Computation triggers on dates that begin a new clause or sentence, not embedded in narrative.\",\"cues\":[\"date starts new clause\",\"date after sentence boundary\",\"date after structural break\"],\"exclusions\":[\"date within flowing narrative\",\"date after preposition in\",\"date mid-sentence\"]},{\"text\":\"Computation fires at text following section breaks, headers, or major topic shifts.\",\"cues\":[\"text after section header\",\"text after topic shift\",\"text after formatting break\"],\"exclusions\":[\"text within same section\",\"text mid-paragraph\",\"text after inline clause\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires at text following section breaks, headers, or major topic shifts."
+    },
+    "own_test_auc": 0.723388671875,
+    "own_validation_auc": 0.5345458984375,
+    "retention_ratio_of_excess_over_chance": 0.6964467005076143
+  },
+  "module_feature_id": "L13B2__module_w3",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 1.1595055915244261,
+    "interpretation": {
+      "cues": [
+        "sentence-final punctuation",
+        "new entity introduction",
+        "subtopic onset"
+      ],
+      "exclusions": [
+        "mid-sentence temporal clauses",
+        "inline appositives",
+        "continuation of same subject"
+      ],
+      "feature_id": "L13B2",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation executes at section header boundaries or major topic transitions in Wikipedia-style articles.\",\"cues\":[\"section header markers\",\"topic shift boundaries\",\"structural article transitions\"],\"exclusions\":[\"mid-sentence date references\",\"inline parenthetical clauses\",\"continuous narrative flow\"]},{\"text\":\"Computation fires when a sentence ends and the next span begins a new subtopic or named entity.\",\"cues\":[\"sentence-final punctuation\",\"new entity introduction\",\"subtopic onset\"],\"exclusions\":[\"mid-sentence temporal clauses\",\"inline appositives\",\"continuation of same subject\"]},{\"text\":\"Computation activates at paragraph or section breaks marked by headers like 'Development' or 'Casualties'.\",\"cues\":[\"wiki section headers\",\"equals-sign headers\",\"paragraph breaks\"],\"exclusions\":[\"in-paragraph dates\",\"inline temporal phrases\",\"same-paragraph continuations\"]},{\"text\":\"Computation triggers at boundaries between descriptive narrative and new event/entity exposition.\",\"cues\":[\"narrative-to-event shift\",\"new entity naming\",\"expository onset\"],\"exclusions\":[\"same-entity continuation\",\"mid-clause dates\",\"parenthetical asides\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires when a sentence ends and the next span begins a new subtopic or named entity."
+    },
+    "semantics_valid": true,
+    "test_auc": 0.740478515625,
+    "validation_auc": 0.7073974609375
+  },
+  "status": "SEMANTICS_RETAINED",
+  "support": [
+    1757,
+    770,
+    2219
+  ],
+  "top_k": 2
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L13B3.json

@@ -0,0 +1,68 @@
+{
+  "block_feature_id": "L13B3",
+  "certified_width": 3,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 13,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.628173828125,
+    "own_interpretation": {
+      "cues": [
+        "section header markers",
+        "topic shift",
+        "Wikipedia article structure"
+      ],
+      "exclusions": [
+        "mid-sentence dates",
+        "chart statistics",
+        "intra-section content"
+      ],
+      "feature_id": "L13B3__module_w3",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at section header boundaries or major topic transitions in Wikipedia-style articles.\",\"cues\":[\"section header markers\",\"topic shift\",\"Wikipedia article structure\"],\"exclusions\":[\"mid-sentence dates\",\"chart statistics\",\"intra-section content\"]},{\"text\":\"Activates at narrative transitions introducing new entities, events, or temporal shifts following prior context.\",\"cues\":[\"new entity introduction\",\"temporal transition\",\"narrative continuation\"],\"exclusions\":[\"same-topic elaboration\",\"statistical data\",\"chart entries\"]},{\"text\":\"Triggers at paragraph boundaries where discourse shifts from descriptive content to new narrative episodes.\",\"cues\":[\"paragraph boundary\",\"discourse shift\",\"new episode start\"],\"exclusions\":[\"within-paragraph continuation\",\"data listing\",\"chart formatting\"]},{\"text\":\"Fires when text transitions from specific factual content to broader historical or contextual narrative.\",\"cues\":[\"factual-to-contextual shift\",\"broadening scope\",\"historical narrative onset\"],\"exclusions\":[\"narrowing to specifics\",\"statistical detail\",\"same-scope continuation\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires at section header boundaries or major topic transitions in Wikipedia-style articles."
+    },
+    "own_test_auc": 0.5919189453125,
+    "own_validation_auc": 0.6043701171875,
+    "retention_ratio_of_excess_over_chance": 0.4697986577181208
+  },
+  "module_feature_id": "L13B3__module_w3",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.9410526315789474,
+    "interpretation": {
+      "cues": [
+        "paragraph ending with period",
+        "new thematic content",
+        "shift in subject matter"
+      ],
+      "exclusions": [
+        "continuation of same topic",
+        "mid-paragraph date",
+        "same narrative flow"
+      ],
+      "feature_id": "L13B3",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates at section header boundaries, especially Wikipedia-style headers with equals signs.\",\"cues\":[\"= = section header\",\"header followed by prose\",\"structural transition marker\"],\"exclusions\":[\"inline dates in prose\",\"mid-sentence year references\",\"narrative continuation\"]},{\"text\":\"Activates at topic transitions where new subtopic begins after prior paragraph closure.\",\"cues\":[\"paragraph ending with period\",\"new thematic content\",\"shift in subject matter\"],\"exclusions\":[\"continuation of same topic\",\"mid-paragraph date\",\"same narrative flow\"]},{\"text\":\"Fires at temporal or narrative shifts introducing new events or phases.\",\"cues\":[\"temporal phrase like following\",\"event transition\",\"new chronological phase\"],\"exclusions\":[\"ongoing same event\",\"continuous action description\",\"same time period\"]},{\"text\":\"Triggers on structural boundaries marked by punctuation closure and new content onset.\",\"cues\":[\"sentence-final punctuation\",\"new sentence new topic\",\"discourse boundary\"],\"exclusions\":[\"mid-sentence continuation\",\"same sentence topic\",\"inline numerical data\"]}]}",
+      "round_index": 0,
+      "text": "Activates at topic transitions where new subtopic begins after prior paragraph closure."
+    },
+    "semantics_valid": true,
+    "test_auc": 0.7728271484375,
+    "validation_auc": 0.7899169921875
+  },
+  "status": "SEMANTICS_NOT_RETAINED",
+  "support": [
+    1757,
+    770,
+    2219
+  ],
+  "top_k": 3
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L14B1.json

@@ -0,0 +1,68 @@
+{
+  "block_feature_id": "L14B1",
+  "certified_width": 3,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 14,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.906982421875,
+    "own_interpretation": {
+      "cues": [
+        "digit within compound number",
+        "compound number in parens",
+        "conversion unit nearby"
+      ],
+      "exclusions": [
+        "single standalone digit",
+        "digit in year or date",
+        "digit in non-conversion context"
+      ],
+      "feature_id": "L14B1__module_w3",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Executes when a unit token (m, km, ft, mi, \u00b0C, mm, in, kg, lb, mph) is followed by a parenthetical unit conversion.\",\"cues\":[\"unit token before parenthesis\",\"parenthetical conversion follows\",\"numeric value inside parens\"],\"exclusions\":[\"no unit token before parenthesis\",\"parenthetical contains non-numeric content\",\"no conversion relationship\"]},{\"text\":\"Activates on numeric tokens inside parenthetical unit-conversion expressions, especially multi-token numbers with @,@ or @.@ separators.\",\"cues\":[\"numeric token inside parens\",\"separator tokens @,@ or @.@\",\"adjacent unit token\"],\"exclusions\":[\"numeric token outside parens\",\"date or year number\",\"identifier-like number\"]},{\"text\":\"Fires on the first numeric token immediately after an opening parenthesis preceded by a unit and conversion cue.\",\"cues\":[\"opening parenthesis after unit\",\"first number in conversion\",\"unit conversion context\"],\"exclusions\":[\"number after closing parenthesis\",\"number in non-conversion clause\",\"standalone numeric token\"]},{\"text\":\"Triggers on numeric sub-tokens within multi-digit conversion values, especially digits of compound numbers like 35,056 or 98.5.\",\"cues\":[\"digit within compound number\",\"compound number in parens\",\"conversion unit nearby\"],\"exclusions\":[\"single standalone digit\",\"digit in year or date\",\"digit in non-conversion context\"]}]}",
+      "round_index": 0,
+      "text": "Triggers on numeric sub-tokens within multi-digit conversion values, especially digits of compound numbers like 35,056 or 98.5."
+    },
+    "own_test_auc": 0.781494140625,
+    "own_validation_auc": 0.8641357421875,
+    "retention_ratio_of_excess_over_chance": 1.0537294563843236
+  },
+  "module_feature_id": "L14B1__module_w3",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 1.1823617339312407,
+    "interpretation": {
+      "cues": [
+        "digit token in prose",
+        "year-like number",
+        "quantity with digits"
+      ],
+      "exclusions": [
+        "parenthesized conversion values",
+        "header markup lines",
+        "unit symbols adjacent"
+      ],
+      "feature_id": "L14B1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Executes when a numeric measurement in parentheses follows a unit, indicating unit conversion.\",\"cues\":[\"parenthesized number after unit\",\"unit conversion context\",\"measurement value in parens\"],\"exclusions\":[\"section headers with equals signs\",\"dates in parentheses\",\"proper names in parentheses\"]},{\"text\":\"Executes at section header boundaries marked by equals signs.\",\"cues\":[\"equals signs delimiting header\",\"header title text\",\"blank lines around header\"],\"exclusions\":[\"parenthesized unit conversions\",\"measurement values\",\"narrative prose text\"]},{\"text\":\"Executes on digit sequences embedded in prose, especially years or quantities.\",\"cues\":[\"digit token in prose\",\"year-like number\",\"quantity with digits\"],\"exclusions\":[\"parenthesized conversion values\",\"header markup lines\",\"unit symbols adjacent\"]}]}",
+      "round_index": 0,
+      "text": "Executes on digit sequences embedded in prose, especially years or quantities."
+    },
+    "semantics_valid": true,
+    "test_auc": 0.88623046875,
+    "validation_auc": 0.82666015625
+  },
+  "status": "SEMANTICS_RETAINED",
+  "support": [
+    4278,
+    2634,
+    1698
+  ],
+  "top_k": 1
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L14B2.json

@@ -0,0 +1,70 @@
+{
+  "block_feature_id": "L14B2",
+  "certified_width": 5,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 14,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.897216796875,
+    "own_interpretation": {
+      "cues": [
+        "parenthetical conversion context",
+        "measurement with unit precedes",
+        "numeric conversion value"
+      ],
+      "exclusions": [
+        "year tokens in dates",
+        "percentages",
+        "casualty counts"
+      ],
+      "feature_id": "L14B2__module_w5",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Executes at numeric tokens inside parenthetical unit conversions following a measurement with unit.\",\"cues\":[\"parenthetical conversion context\",\"measurement with unit precedes\",\"numeric conversion value\"],\"exclusions\":[\"year tokens in dates\",\"percentages\",\"casualty counts\"]},{\"text\":\"Executes at digit tokens within four-digit year references in narrative text.\",\"cues\":[\"four-digit year pattern\",\"narrative date context\",\"century prefix present\"],\"exclusions\":[\"parenthetical conversions\",\"measurement values\",\"non-year numbers\"]},{\"text\":\"Executes at numeric tokens in speed/range parentheticals with knots or km/h units.\",\"cues\":[\"knots or km/h context\",\"parenthetical speed conversion\",\"range distance value\"],\"exclusions\":[\"plain year references\",\"non-speed measurements\",\"percentages\"]},{\"text\":\"Executes at decade year tokens ending in 0s within historical narrative sentences.\",\"cues\":[\"decade suffix 0s\",\"historical narrative context\",\"century prefix\"],\"exclusions\":[\"exact four-digit years\",\"parenthetical conversions\",\"non-decade numbers\"]}]}",
+      "round_index": 0,
+      "text": "Executes at numeric tokens inside parenthetical unit conversions following a measurement with unit."
+    },
+    "own_test_auc": 0.8038330078125,
+    "own_validation_auc": 0.8480224609375,
+    "retention_ratio_of_excess_over_chance": 1.093781512605042
+  },
+  "module_feature_id": "L14B2__module_w5",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 1.2262984336356142,
+    "interpretation": {
+      "cues": [
+        "parentheses with operators",
+        "unit conversion context",
+        "numeric digit target"
+      ],
+      "exclusions": [
+        "year in date context",
+        "digit in non-conversion text",
+        "no parentheses present"
+      ],
+      "feature_id": "L14B2",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Executes when a unit conversion value follows a measurement with parenthetical format containing comparison operators.\",\"cues\":[\"parenthetical unit conversion\",\"comparison operators present\",\"measurement value preceding\"],\"exclusions\":[\"standalone year digits\",\"narrative date references\",\"no unit conversion context\"]},{\"text\":\"Activates on numeric digits inside parenthetical conversion expressions with angle bracket markup.\",\"cues\":[\"parenthetical expression context\",\"angle bracket markup\",\"numeric conversion target\"],\"exclusions\":[\"year in date phrase\",\"digit in narrative sentence\",\"no parenthetical conversion\"]},{\"text\":\"Fires when digits appear within unit conversion parentheses marked by comparison operator sequences.\",\"cues\":[\"parentheses with operators\",\"unit conversion context\",\"numeric digit target\"],\"exclusions\":[\"year in date context\",\"digit in non-conversion text\",\"no parentheses present\"]},{\"text\":\"Triggers on digits embedded in measurement conversion parentheticals containing operator chains.\",\"cues\":[\"measurement conversion parenthetical\",\"operator chain markup\",\"digit as conversion result\"],\"exclusions\":[\"standalone four-digit year\",\"date narrative context\",\"no conversion operators\"]}]}",
+      "round_index": 0,
+      "text": "Fires when digits appear within unit conversion parentheses marked by comparison operator sequences."
+    },
+    "semantics_valid": true,
+    "test_auc": 0.8631591796875,
+    "validation_auc": 0.796142578125
+  },
+  "status": "SEMANTICS_RETAINED",
+  "support": [
+    4278,
+    2854,
+    481,
+    2634,
+    1698
+  ],
+  "top_k": 2
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L1B1.json

@@ -0,0 +1,67 @@
+{
+  "block_feature_id": "L1B1",
+  "certified_width": 2,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 1,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.718994140625,
+    "own_interpretation": {
+      "cues": [
+        "narrative prose sentence",
+        "biographical or historical context",
+        "year as temporal anchor"
+      ],
+      "exclusions": [
+        "chart or ranking context",
+        "list of competition years",
+        "inline parenthetical date"
+      ],
+      "feature_id": "L1B1__module_w2",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation executes when a single digit completes a four-digit year token at a sentence-internal position.\",\"cues\":[\"single digit after 3 year digits\",\"sentence-internal year token\",\"year followed by comma or period\"],\"exclusions\":[\"digit inside larger number sequence\",\"digit in date day/month field\",\"digit in non-year numeric context\"]},{\"text\":\"Computation fires when the marked digit is the final digit of a standalone year appearing after a preposition or temporal phrase.\",\"cues\":[\"year after preposition like In/On\",\"temporal phrase before year\",\"year as standalone token\"],\"exclusions\":[\"year in list of multiple years\",\"year as part of range\",\"digit in non-temporal sentence\"]},{\"text\":\"Computation activates for year-final digits in narrative or biographical sentences, not in tabular or list contexts.\",\"cues\":[\"narrative prose sentence\",\"biographical or historical context\",\"year as temporal anchor\"],\"exclusions\":[\"chart or ranking context\",\"list of competition years\",\"inline parenthetical date\"]}]}",
+      "round_index": 0,
+      "text": "Computation activates for year-final digits in narrative or biographical sentences, not in tabular or list contexts."
+    },
+    "own_test_auc": 0.6695556640625,
+    "own_validation_auc": 0.7188720703125,
+    "retention_ratio_of_excess_over_chance": 1.0965770171149145
+  },
+  "module_feature_id": "L1B1__module_w2",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.6935142009325985,
+    "interpretation": {
+      "cues": [
+        "year in date phrase",
+        "year after month name",
+        "year in historical narrative"
+      ],
+      "exclusions": [
+        "digit in technical specification",
+        "digit in UTC timestamp",
+        "digit in non-date number"
+      ],
+      "feature_id": "L1B1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation executes when a single digit follows a year prefix like 201 or 200, forming a complete year token.\",\"cues\":[\"digit preceded by 201 or 200\",\"digit preceded by 192 or 193\",\"digit preceded by 194\"],\"exclusions\":[\"digit preceded by non-year context\",\"digit as part of time format\",\"digit preceded by 1 only\"]},{\"text\":\"Computation fires when the digit completes a four-digit year in a date or temporal expression.\",\"cues\":[\"year in date phrase\",\"year after month name\",\"year in historical narrative\"],\"exclusions\":[\"digit in technical specification\",\"digit in UTC timestamp\",\"digit in non-date number\"]},{\"text\":\"Computation activates when the marked digit is the final digit of a standalone year reference.\",\"cues\":[\"year as standalone token\",\"year in sentence-initial position\",\"year in parenthetical\"],\"exclusions\":[\"digit within multi-digit technical number\",\"digit in coordinate\",\"digit in version number\"]},{\"text\":\"Computation triggers on digits that complete year references across centuries, especially in biographical or historical contexts.\",\"cues\":[\"year in biography\",\"year in military history\",\"year in publication context\"],\"exclusions\":[\"digit in product model number\",\"digit in population statistic\",\"digit in measurement value\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires when the digit completes a four-digit year in a date or temporal expression."
+    },
+    "semantics_valid": true,
+    "test_auc": 0.69970703125,
+    "validation_auc": 0.7879638671875
+  },
+  "status": "SEMANTICS_RETAINED",
+  "support": [
+    4058,
+    2700
+  ],
+  "top_k": 1
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L21B1.json

@@ -0,0 +1,65 @@
+{
+  "block_feature_id": "L21B1",
+  "certified_width": 1,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 21,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "author names",
+        "comic book creators",
+        "artistic works"
+      ],
+      "exclusions": [
+        "generic creative discussion",
+        "unnamed references"
+      ],
+      "feature_id": "L21B1__module_w1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at sentence-initial tokens following structural breaks like section headers or list markers.\",\"cues\":[\"section header markers\",\"list numbering\",\"structural break\",\"article headings\"],\"exclusions\":[\"mid-sentence continuation\",\"paragraph-internal tokens\"]},{\"text\":\"Computation activates at tokens beginning named-entity lists or enumerations of proper nouns.\",\"cues\":[\"proper noun sequences\",\"enumeration of people\",\"list of artists\"],\"exclusions\":[\"single proper noun\",\"common noun lists\"]},{\"text\":\"Computation fires at tokens starting disaster or conflict narrative descriptions involving destruction.\",\"cues\":[\"earthquake aftermath\",\"building destruction\",\"conflict narrative\"],\"exclusions\":[\"general historical narrative\",\"technical descriptions\"]},{\"text\":\"Computation activates at tokens beginning biographical or creative-work discussion involving named creators.\",\"cues\":[\"author names\",\"comic book creators\",\"artistic works\"],\"exclusions\":[\"generic creative discussion\",\"unnamed references\"]}]}",
+      "round_index": 0,
+      "text": "Computation activates at tokens beginning biographical or creative-work discussion involving named creators."
+    },
+    "own_test_auc": 0.5262451171875,
+    "own_validation_auc": 0.48291015625,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L21B1__module_w1",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": -8007812.5,
+    "interpretation": {
+      "cues": [
+        "digits in year or date",
+        "measurement units nearby",
+        "parenthetical numeric conversions"
+      ],
+      "exclusions": [
+        "non-numeric prose tokens",
+        "place names without numbers",
+        "abstract concepts"
+      ],
+      "feature_id": "L21B1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates at numeric tokens within dates, measurements, and quantities in encyclopedic text.\",\"cues\":[\"digits in year or date\",\"measurement units nearby\",\"parenthetical numeric conversions\"],\"exclusions\":[\"non-numeric prose tokens\",\"place names without numbers\",\"abstract concepts\"]},{\"text\":\"Activates on tokens immediately following sentence boundaries in biographical or historical narrative.\",\"cues\":[\"period or newline preceding\",\"proper noun follows\",\"encyclopedic register\"],\"exclusions\":[\"mid-sentence punctuation\",\"dialogue or quotation\",\"casual register\"]},{\"text\":\"Fires on tokens within multi-token numeric sequences like years, dimensions, or large quantities.\",\"cues\":[\"adjacent digit tokens\",\"conversion parenthesis pattern\",\"quantity or measurement context\"],\"exclusions\":[\"single isolated digit\",\"non-numeric descriptive text\",\"list item markers\"]},{\"text\":\"Activates at content word boundaries in factual expository passages with named entities and dates.\",\"cues\":[\"named entity present\",\"factual expository style\",\"temporal or numeric data\"],\"exclusions\":[\"opinion or commentary text\",\"fictional narrative dialogue\",\"procedural instructions\"]}]}",
+      "round_index": 0,
+      "text": "Computation activates at numeric tokens within dates, measurements, and quantities in encyclopedic text."
+    },
+    "semantics_valid": false,
+    "test_auc": 0.419921875,
+    "validation_auc": 0.4937744140625
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    567
+  ],
+  "top_k": 1
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L2B1.json

@@ -0,0 +1,66 @@
+{
+  "block_feature_id": "L2B1",
+  "certified_width": 1,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 2,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "equals signs nearby",
+        "newline boundary",
+        "wiki formatting artifacts"
+      ],
+      "exclusions": [
+        "continuous prose",
+        "no formatting markers",
+        "embedded within paragraph"
+      ],
+      "feature_id": "L2B1__module_w1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Executes at sentence-initial or clause-initial positions following section headers, list markers, or structural breaks.\",\"cues\":[\"section header precedes\",\"list marker precedes\",\"structural break precedes\"],\"exclusions\":[\"mid-sentence tokens\",\"narrative continuation\",\"within descriptive passage\"]},{\"text\":\"Activates at document or paragraph boundaries marked by wiki formatting artifacts like equals signs or newlines.\",\"cues\":[\"equals signs nearby\",\"newline boundary\",\"wiki formatting artifacts\"],\"exclusions\":[\"continuous prose\",\"no formatting markers\",\"embedded within paragraph\"]},{\"text\":\"Fires on tokens beginning new informational units after completed prior units, especially in encyclopedic or technical text.\",\"cues\":[\"new informational unit\",\"encyclopedic text style\",\"prior unit completed\"],\"exclusions\":[\"narrative fiction text\",\"biographical anecdote\",\"opinion or dialogue\"]},{\"text\":\"Triggers on initial tokens of multi-token technical or proper-noun spans, particularly in spec-style or reference text.\",\"cues\":[\"technical specification context\",\"proper noun span start\",\"reference text style\"],\"exclusions\":[\"common noun mid-sentence\",\"informal language\",\"first-person narrative\"]}]}",
+      "round_index": 0,
+      "text": "Activates at document or paragraph boundaries marked by wiki formatting artifacts like equals signs or newlines."
+    },
+    "own_test_auc": 0.4376220703125,
+    "own_validation_auc": 0.4798583984375,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L2B1__module_w1",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": -9655761.71875,
+    "interpretation": {
+      "cues": [
+        "split digit sequences in dates",
+        "alphanumeric identifiers",
+        "numeric ranges or quantities"
+      ],
+      "exclusions": [
+        "single-token whole numbers",
+        "spelled-out numbers",
+        "non-numeric common words"
+      ],
+      "feature_id": "L2B1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Activates at token boundaries within multi-token named entities, proper nouns, or technical terms being decomposed into subword units.\",\"cues\":[\"subword tokenization of proper nouns\",\"multi-token entity names\",\"technical or specialized vocabulary\"],\"exclusions\":[\"common single-token words\",\"function words in isolation\",\"generic vocabulary\"]},{\"text\":\"Fires at punctuation-adjacent positions where structural formatting tokens (headings, list markers, section breaks) are tokenized into multiple pieces.\",\"cues\":[\"section heading markers\",\"list formatting tokens\",\"markup punctuation sequences\"],\"exclusions\":[\"mid-sentence punctuation\",\"standard comma or period usage\",\"quotation marks in dialogue\"]},{\"text\":\"Triggers on numeric or alphanumeric tokens split across boundaries, especially dates, measurements, or identifiers embedded in text.\",\"cues\":[\"split digit sequences in dates\",\"alphanumeric identifiers\",\"numeric ranges or quantities\"],\"exclusions\":[\"single-token whole numbers\",\"spelled-out numbers\",\"non-numeric common words\"]},{\"text\":\"Activates at word-internal morpheme boundaries where complex words are split into subword units during tokenization.\",\"cues\":[\"compound word splits\",\"affixed words at morpheme boundaries\",\"inflected or derived forms\"],\"exclusions\":[\"simple monomorphemic words\",\"already short common words\",\"uninflected base forms\"]}]}",
+      "round_index": 0,
+      "text": "Triggers on numeric or alphanumeric tokens split across boundaries, especially dates, measurements, or identifiers embedded in text."
+    },
+    "semantics_valid": false,
+    "test_auc": 0.4034423828125,
+    "validation_auc": 0.4913330078125
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    813
+  ],
+  "top_k": 1
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L2B2.json

@@ -0,0 +1,66 @@
+{
+  "block_feature_id": "L2B2",
+  "certified_width": 2,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 2,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "proper noun onset",
+        "capitalized entity",
+        "place or ship name"
+      ],
+      "exclusions": [
+        "lowercase common nouns",
+        "mid-phrase tokens"
+      ],
+      "feature_id": "L2B2__module_w2",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Executes at token immediately following a hyphen or @-@ joiner in compound words.\",\"cues\":[\"@-@ token boundary\",\"compound word joiner\",\"hyphenated construction\"],\"exclusions\":[\"plain hyphen in prose\",\"non-compound tokens\"]},{\"text\":\"Executes at section header markup tokens like =>> or =<<.\",\"cues\":[\"wiki header syntax\",\"=>> or =<< marker\",\"section divider\"],\"exclusions\":[\"prose punctuation\",\"parenthetical content\"]},{\"text\":\"Executes at word-initial tokens in clause-initial position after sentence boundary.\",\"cues\":[\"sentence-initial position\",\"clause start\",\"following period or comma\"],\"exclusions\":[\"mid-sentence tokens\",\"tokens after other words\"]},{\"text\":\"Executes at tokens starting multi-word proper nouns or capitalized entities.\",\"cues\":[\"proper noun onset\",\"capitalized entity\",\"place or ship name\"],\"exclusions\":[\"lowercase common nouns\",\"mid-phrase tokens\"]}]}",
+      "round_index": 0,
+      "text": "Executes at tokens starting multi-word proper nouns or capitalized entities."
+    },
+    "own_test_auc": 0.4339599609375,
+    "own_validation_auc": 0.646484375,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L2B2__module_w2",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.6666666666666666,
+    "interpretation": {
+      "cues": [
+        "proper noun clause subject",
+        "pronoun opening clause",
+        "subject at clause start"
+      ],
+      "exclusions": [
+        "proper nouns in apposition",
+        "nouns in prepositional phrases",
+        "tokens mid-clause"
+      ],
+      "feature_id": "L2B2",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at tokens immediately following section header markup or punctuation boundaries.\",\"cues\":[\"token after header markup\",\"token after punctuation mark\",\"boundary or transition point\"],\"exclusions\":[\"tokens within running prose\",\"tokens mid-sentence\",\"tokens inside header text\"]},{\"text\":\"Computation activates at tokens that are subjects of passive or transferred actions in narrative.\",\"cues\":[\"subject of passive verb\",\"entity receiving action\",\"narrative event participant\"],\"exclusions\":[\"abstract section labels\",\"geographic descriptors alone\",\"tokens in definitional context\"]},{\"text\":\"Computation triggers at tokens following hyphenated compound or special punctuation constructions.\",\"cues\":[\"after hyphenated compound\",\"after special punctuation\",\"compound token boundary\"],\"exclusions\":[\"simple unbroken words\",\"tokens in plain lists\",\"standard sentence tokens\"]},{\"text\":\"Computation fires at proper nouns or pronouns serving as clause-initial subjects.\",\"cues\":[\"proper noun clause subject\",\"pronoun opening clause\",\"subject at clause start\"],\"exclusions\":[\"proper nouns in apposition\",\"nouns in prepositional phrases\",\"tokens mid-clause\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires at proper nouns or pronouns serving as clause-initial subjects."
+    },
+    "semantics_valid": false,
+    "test_auc": 0.515625,
+    "validation_auc": 0.5234375
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    4693,
+    813
+  ],
+  "top_k": 2
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L2B3.json

@@ -0,0 +1,67 @@
+{
+  "block_feature_id": "L2B3",
+  "certified_width": 3,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 2,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "latter half of split date",
+        "second digit pair of year",
+        "continuation of split number"
+      ],
+      "exclusions": [
+        "complete single-token dates",
+        "unsplit numbers or years"
+      ],
+      "feature_id": "L2B3__module_w3",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates at token immediately following a word boundary or morpheme split, including hyphenated compounds and concatenated tokens.\",\"cues\":[\"token after word break\",\"hyphenated compound boundary\",\"concatenated morpheme split\"],\"exclusions\":[\"mid-word continuations\",\"tokens within unbroken words\"]},{\"text\":\"Activates at the first token of a multi-token word or at the boundary where a single conceptual word spans multiple tokens.\",\"cues\":[\"start of multi-token word\",\"first token after split\",\"word-internal boundary token\"],\"exclusions\":[\"single-token complete words\",\"tokens not at word boundaries\"]},{\"text\":\"Fires at tokens that are the second part of a split word, where the first part is a recognizable word stem or prefix.\",\"cues\":[\"second part of split word\",\"suffix or continuation token\",\"completes a word fragment\"],\"exclusions\":[\"standalone complete words\",\"tokens within normal unsplit text\"]},{\"text\":\"Activates at tokens representing the latter half of a date, number, or compound entity split across token boundaries.\",\"cues\":[\"latter half of split date\",\"second digit pair of year\",\"continuation of split number\"],\"exclusions\":[\"complete single-token dates\",\"unsplit numbers or years\"]}]}",
+      "round_index": 0,
+      "text": "Activates at tokens representing the latter half of a date, number, or compound entity split across token boundaries."
+    },
+    "own_test_auc": 0.5537109375,
+    "own_validation_auc": 0.5924072265625,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L2B3__module_w3",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": -3.2789968652037618,
+    "interpretation": {
+      "cues": [
+        "preceded by punctuation",
+        "sentence-initial position",
+        "after clause break"
+      ],
+      "exclusions": [
+        "mid-sentence common nouns",
+        "tokens after commas in lists",
+        "descriptive adjectives mid-sentence"
+      ],
+      "feature_id": "L2B3",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Executes at token following a clause boundary or sentence start\",\"cues\":[\"preceded by punctuation\",\"sentence-initial position\",\"after clause break\"],\"exclusions\":[\"mid-sentence common nouns\",\"tokens after commas in lists\",\"descriptive adjectives mid-sentence\"]},{\"text\":\"Activates at section headers, markup boundaries, or structural delimiters\",\"cues\":[\"wiki heading markers\",\"section title tokens\",\"markup delimiter symbols\"],\"exclusions\":[\"inline hyphenated words\",\"regular sentence content\",\"common nouns in prose\"]},{\"text\":\"Fires on proper nouns that are subjects of historical or geopolitical events\",\"cues\":[\"country or place names\",\"followed by action verbs\",\"historical event context\"],\"exclusions\":[\"common nouns in lists\",\"activity or hobby nouns\",\"abstract descriptive nouns\"]},{\"text\":\"Triggers at numeric tokens in date expressions or structural numbering\",\"cues\":[\"digits in date context\",\"section or issue numbers\",\"year or day values\"],\"exclusions\":[\"percentage values\",\"measurement quantities\",\"counts in descriptive text\"]}]}",
+      "round_index": 0,
+      "text": "Executes at token following a clause boundary or sentence start"
+    },
+    "semantics_valid": false,
+    "test_auc": 0.372314453125,
+    "validation_auc": 0.5389404296875
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    4693,
+    813,
+    3575
+  ],
+  "top_k": 3
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L2B4.json

@@ -0,0 +1,70 @@
+{
+  "block_feature_id": "L2B4",
+  "certified_width": 5,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 2,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "numeric date components",
+        "day-month date pattern",
+        "year in date context"
+      ],
+      "exclusions": [
+        "percentage numbers",
+        "edition numbers",
+        "age thresholds"
+      ],
+      "feature_id": "L2B4__module_w5",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates at hyphenated compound modifiers like adjective-noun or noun-noun pairs\",\"cues\":[\"hyphenated compound modifier\",\"adjective-noun hyphenation\",\"noun-noun hyphenation\"],\"exclusions\":[\"proper name hyphenation\",\"multi-hyphen proper names\",\"place name hyphens\"]},{\"text\":\"Activates at numeric date tokens forming day-month or year patterns in date expressions\",\"cues\":[\"numeric date components\",\"day-month date pattern\",\"year in date context\"],\"exclusions\":[\"percentage numbers\",\"edition numbers\",\"age thresholds\"]},{\"text\":\"Activates at section header delimiters like equals signs marking wiki-style headings\",\"cues\":[\"equals sign headers\",\"section heading delimiters\",\"wiki markup headers\"],\"exclusions\":[\"regular prose text\",\"parenthetical content\",\"inline punctuation\"]},{\"text\":\"Activates at sentence-initial common words in narrative or expository contexts\",\"cues\":[\"sentence-initial common word\",\"narrative continuation\",\"expository transition\"],\"exclusions\":[\"mid-sentence function words\",\"proper nouns\",\"quoted speech\"]}]}",
+      "round_index": 0,
+      "text": "Activates at numeric date tokens forming day-month or year patterns in date expressions"
+    },
+    "own_test_auc": 0.517333984375,
+    "own_validation_auc": 0.5653076171875,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L2B4__module_w5",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": -1.1217292377701935,
+    "interpretation": {
+      "cues": [
+        "sentence-initial position",
+        "function words",
+        "discourse markers"
+      ],
+      "exclusions": [
+        "mid-sentence function words",
+        "content words",
+        "clause-internal tokens"
+      ],
+      "feature_id": "L2B4",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at punctuation, symbols, and structural delimiters in text.\",\"cues\":[\"punctuation marks\",\"special symbols\",\"section headers\"],\"exclusions\":[\"common nouns\",\"regular verbs\",\"conjunctions\"]},{\"text\":\"Fires at hyphenated compound word boundaries and @-@ connector tokens.\",\"cues\":[\"@-@ tokens\",\"compound word splits\",\"hyphen boundaries\"],\"exclusions\":[\"standalone words\",\"regular hyphen words\",\"unconnected tokens\"]},{\"text\":\"Activates at wiki section header markers like = = and => sequences.\",\"cues\":[\"= = markers\",\"=> headers\",\"section delimiters\"],\"exclusions\":[\"body text\",\"regular equals signs\",\"mathematical expressions\"]},{\"text\":\"Fires at sentence-initial function words and discourse markers.\",\"cues\":[\"sentence-initial position\",\"function words\",\"discourse markers\"],\"exclusions\":[\"mid-sentence function words\",\"content words\",\"clause-internal tokens\"]}]}",
+      "round_index": 0,
+      "text": "Fires at sentence-initial function words and discourse markers."
+    },
+    "semantics_valid": false,
+    "test_auc": 0.379638671875,
+    "validation_auc": 0.6072998046875
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    4693,
+    285,
+    813,
+    3575,
+    3977
+  ],
+  "top_k": 4
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L2B6.json

@@ -0,0 +1,73 @@
+{
+  "block_feature_id": "L2B6",
+  "certified_width": 8,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 2,
+  "module": {
+    "frozen_parent_predicate_test_auc": null,
+    "own_interpretation": {
+      "cues": [
+        "4-digit year context",
+        "mid-digit of year",
+        "date or year phrase"
+      ],
+      "exclusions": [
+        "hyphenated compound words",
+        "non-year numeric quantities",
+        "measurement values"
+      ],
+      "feature_id": "L2B6__module_w8",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates on year digits in 4-digit year spans, especially mid-digit positions.\",\"cues\":[\"4-digit year context\",\"mid-digit of year\",\"date or year phrase\"],\"exclusions\":[\"hyphenated compound words\",\"non-year numeric quantities\",\"measurement values\"]},{\"text\":\"Activates on tokens immediately following prepositions or function words introducing noun phrases.\",\"cues\":[\"preposition followed by noun\",\"function word before content\",\"phrase-initial position\"],\"exclusions\":[\"mid-sentence content words\",\"proper nouns in isolation\",\"verb-internal positions\"]},{\"text\":\"Fires on individual digit tokens within multi-digit numbers, particularly in technical or military contexts.\",\"cues\":[\"digit within multi-digit number\",\"technical or military context\",\"specification or designation\"],\"exclusions\":[\"hyphenated technical terms\",\"measurement unit tokens\",\"non-numeric adjectives\"]},{\"text\":\"Activates on repeated or parallel structural tokens like hyphens in compound constructions.\",\"cues\":[\"repeated structural punctuation\",\"compound word boundaries\",\"parallel construction patterns\"],\"exclusions\":[\"single isolated hyphen\",\"non-repeated punctuation marks\",\"sentence-final punctuation\"]}]}",
+      "round_index": 0,
+      "text": "Computation activates on year digits in 4-digit year spans, especially mid-digit positions."
+    },
+    "own_test_auc": 0.381103515625,
+    "own_validation_auc": 0.5032958984375,
+    "retention_ratio_of_excess_over_chance": null
+  },
+  "module_feature_id": "L2B6__module_w8",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": -139.5,
+    "interpretation": {
+      "cues": [
+        "token after month name",
+        "day-of-month position",
+        "date construction pattern"
+      ],
+      "exclusions": [
+        "month name as standalone",
+        "proper noun fragments",
+        "non-date month mentions"
+      ],
+      "feature_id": "L2B6",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates on digits within date or year spans, especially multi-digit sequences forming dates.\",\"cues\":[\"digit in date context\",\"multi-digit year span\",\"date-adjacent tokens\"],\"exclusions\":[\"isolated percentage digits\",\"highway route numbers\",\"non-date numeric tokens\"]},{\"text\":\"Activates on tokens immediately following month names or within full date constructions.\",\"cues\":[\"token after month name\",\"day-of-month position\",\"date construction pattern\"],\"exclusions\":[\"month name as standalone\",\"proper noun fragments\",\"non-date month mentions\"]},{\"text\":\"Activates on hyphen-adjacent or symbol-adjacent tokens in technical or compound constructions.\",\"cues\":[\"hyphen-adjacent token\",\"compound construction\",\"technical notation context\"],\"exclusions\":[\"hyphen in proper names\",\"hyphen in place names\",\"non-technical hyphen usage\"]},{\"text\":\"Activates on single characters or symbols that are part of larger structural patterns like dates or notation.\",\"cues\":[\"single char in pattern\",\"structural component token\",\"pattern-internal position\"],\"exclusions\":[\"single char in name\",\"isolated punctuation\",\"character in word fragment\"]}]}",
+      "round_index": 0,
+      "text": "Activates on tokens immediately following month names or within full date constructions."
+    },
+    "semantics_valid": false,
+    "test_auc": 0.4659423828125,
+    "validation_auc": 0.500244140625
+  },
+  "status": "PARENT_SEMANTICS_UNRESOLVED",
+  "support": [
+    4693,
+    285,
+    276,
+    813,
+    3977,
+    3575,
+    3737,
+    1680
+  ],
+  "top_k": 6
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L2B8.json

@@ -0,0 +1,75 @@
+{
+  "block_feature_id": "L2B8",
+  "certified_width": 10,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 2,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.3646240234375,
+    "own_interpretation": {
+      "cues": [
+        "preposition introducing noun phrase",
+        "function word in biography",
+        "historical narrative context"
+      ],
+      "exclusions": [
+        "prepositions in technical specs",
+        "prepositions in earthquake context",
+        "prepositions in music review"
+      ],
+      "feature_id": "L2B8__module_w10",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates on digits within dates, years, and numerical quantities, especially multi-digit sequences.\",\"cues\":[\"digits in date/year context\",\"multi-digit number sequences\",\"numerical quantities\"],\"exclusions\":[\"digits in product specs\",\"digits in measurements\",\"isolated single digits\"]},{\"text\":\"Computation activates on tokens that are part of larger compound or hyphenated constructions.\",\"cues\":[\"hyphenated compound parts\",\"compound word components\",\"prefix/suffix tokens\"],\"exclusions\":[\"standalone complete words\",\"sentence-final punctuation\",\"possessive markers\"]},{\"text\":\"Computation activates on prepositions and function words introducing noun phrases in biographical or historical contexts.\",\"cues\":[\"preposition introducing noun phrase\",\"function word in biography\",\"historical narrative context\"],\"exclusions\":[\"prepositions in technical specs\",\"prepositions in earthquake context\",\"prepositions in music review\"]},{\"text\":\"Computation activates on section heading markers and structural formatting tokens in article text.\",\"cues\":[\"heading equality markers\",\"section structure tokens\",\"article formatting elements\"],\"exclusions\":[\"inline punctuation marks\",\"quotation marks\",\"sentence boundaries\"]}]}",
+      "round_index": 0,
+      "text": "Computation activates on prepositions and function words introducing noun phrases in biographical or historical contexts."
+    },
+    "own_test_auc": 0.2177734375,
+    "own_validation_auc": 0.5380859375,
+    "retention_ratio_of_excess_over_chance": -2.6279620853080567
+  },
+  "module_feature_id": "L2B8__module_w10",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.6374622356495468,
+    "interpretation": {
+      "cues": [
+        "after sentence period",
+        "clause boundary start",
+        "post-comma new clause"
+      ],
+      "exclusions": [
+        "mid-sentence numeric tokens",
+        "tokens within flowing prose",
+        "inline technical notation"
+      ],
+      "feature_id": "L2B8",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires on tokens immediately following section headers or structural markup like equals signs.\",\"cues\":[\"preceding = = heading\",\"section header markup\",\"wiki section break\"],\"exclusions\":[\"mid-sentence numeric tokens\",\"dates in prose\",\"inline @-@ hyphenation\"]},{\"text\":\"Fires on tokens at sentence-initial or clause-initial positions following a period or strong punctuation break.\",\"cues\":[\"after sentence period\",\"clause boundary start\",\"post-comma new clause\"],\"exclusions\":[\"mid-sentence numeric tokens\",\"tokens within flowing prose\",\"inline technical notation\"]},{\"text\":\"Fires on tokens that are the first content word after a quotation, parenthesis, or structural delimiter closes.\",\"cues\":[\"after closing quote\",\"post-parenthesis content\",\"after structural delimiter\"],\"exclusions\":[\"tokens inside quotation\",\"mid-clause common words\",\"inline hyphenated terms\"]},{\"text\":\"Fires on tokens at the start of a new syntactic unit like a list item, coordinate clause, or discourse-new sentence.\",\"cues\":[\"new list item\",\"coordinate clause onset\",\"discourse-new sentence start\"],\"exclusions\":[\"continuation of same clause\",\"mid-sentence function words\",\"embedded date fragments\"]}]}",
+      "round_index": 0,
+      "text": "Fires on tokens at sentence-initial or clause-initial positions following a period or strong punctuation break."
+    },
+    "semantics_valid": true,
+    "test_auc": 0.551513671875,
+    "validation_auc": 0.580810546875
+  },
+  "status": "SEMANTICS_NOT_RETAINED",
+  "support": [
+    4693,
+    1719,
+    285,
+    276,
+    813,
+    3977,
+    1550,
+    3575,
+    3737,
+    1680
+  ],
+  "top_k": 8
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L3B1.json

@@ -0,0 +1,66 @@
+{
+  "block_feature_id": "L3B1",
+  "certified_width": 1,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 3,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.4364013671875,
+    "own_interpretation": {
+      "cues": [
+        "narrative clause continuation",
+        "coherent event description",
+        "sequential clause tokens"
+      ],
+      "exclusions": [
+        "list-item token sequences",
+        "factual enumeration tokens",
+        "coordinate-list continuations"
+      ],
+      "feature_id": "L3B1__module_w1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates at clause-initial coordinators and discourse transitions in narrative text.\",\"cues\":[\"clause-initial coordinator\",\"narrative discourse context\",\"sentence/clause boundary\"],\"exclusions\":[\"mid-clause coordinators\",\"coordinate structures within lists\",\"coordinators in factual summaries\"]},{\"text\":\"Activates at punctuation or tokens marking structural breaks between sentences or sections.\",\"cues\":[\"sentence-final punctuation\",\"section break markers\",\"structural transition tokens\"],\"exclusions\":[\"punctuation within running lists\",\"mid-sentence punctuation only\",\"punctuation in dense factual passages\"]},{\"text\":\"Fires on tokens beginning a new descriptive or narrative clause after prior context.\",\"cues\":[\"new clause onset\",\"descriptive/narrative continuation\",\"post-comma clause start\"],\"exclusions\":[\"clause onsets in list enumerations\",\"factual summary clauses\",\"coordinate-list members\"]},{\"text\":\"Activates at multi-token spans where each token continues a coherent narrative clause.\",\"cues\":[\"narrative clause continuation\",\"coherent event description\",\"sequential clause tokens\"],\"exclusions\":[\"list-item token sequences\",\"factual enumeration tokens\",\"coordinate-list continuations\"]}]}",
+      "round_index": 0,
+      "text": "Activates at multi-token spans where each token continues a coherent narrative clause."
+    },
+    "own_test_auc": 0.371826171875,
+    "own_validation_auc": 0.53173828125,
+    "retention_ratio_of_excess_over_chance": -0.6376988984088128
+  },
+  "module_feature_id": "L3B1__module_w1",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.6653094462540716,
+    "interpretation": {
+      "cues": [
+        "post-punctuation positions",
+        "encyclopedic narrative context",
+        "structural section transitions"
+      ],
+      "exclusions": [
+        "mid-sentence common words",
+        "chart or listing contexts",
+        "dialogue-only passages"
+      ],
+      "feature_id": "L3B1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires at tokens within multi-token word completions or morphological continuations across diverse content domains.\",\"cues\":[\"mid-word token boundaries\",\"subword continuation tokens\",\"cross-domain content diversity\"],\"exclusions\":[\"complete standalone words\",\"single-domain narrow context\",\"sentence-initial tokens\"]},{\"text\":\"Activates at tokens that are part of incomplete words being assembled from subword units, regardless of topic.\",\"cues\":[\"subword tokenization splits\",\"partial word fragments\",\"topic-agnostic firing\"],\"exclusions\":[\"whole-word single tokens\",\"topic-specific vocabulary\",\"proper noun completions only\"]},{\"text\":\"Fires at continuation tokens following punctuation or structural markers in encyclopedic or narrative text.\",\"cues\":[\"post-punctuation positions\",\"encyclopedic narrative context\",\"structural section transitions\"],\"exclusions\":[\"mid-sentence common words\",\"chart or listing contexts\",\"dialogue-only passages\"]},{\"text\":\"Executes at tokens forming multi-token words in informational text spanning history, geography, science, and biography.\",\"cues\":[\"informational prose genre\",\"multi-token word spans\",\"broad topical coverage\"],\"exclusions\":[\"entertainment chart listings\",\"sports season narratives\",\"single-topic narrow scope\"]}]}",
+      "round_index": 0,
+      "text": "Fires at continuation tokens following punctuation or structural markers in encyclopedic or narrative text."
+    },
+    "semantics_valid": true,
+    "test_auc": 0.5997314453125,
+    "validation_auc": 0.64990234375
+  },
+  "status": "SEMANTICS_NOT_RETAINED",
+  "support": [
+    2247
+  ],
+  "top_k": 1
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L7B1.json

@@ -0,0 +1,67 @@
+{
+  "block_feature_id": "L7B1",
+  "certified_width": 2,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 7,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.7010498046875,
+    "own_interpretation": {
+      "cues": [
+        "marked spans are consecutive digits",
+        "marked spans form full year",
+        "marked spans adjacent no gap"
+      ],
+      "exclusions": [
+        "marked digits split by text",
+        "single marked digit pair",
+        "marked spans separated by words"
+      ],
+      "feature_id": "L7B1__module_w2",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation executes when marked digits form a complete 4-digit year token, not split across other tokens.\",\"cues\":[\"marked spans are consecutive digits\",\"marked spans form full year\",\"marked spans adjacent no gap\"],\"exclusions\":[\"marked digits split by text\",\"single marked digit pair\",\"marked spans separated by words\"]},{\"text\":\"Executes when marked digit spans are the sole content of a year token in a date expression.\",\"cues\":[\"marked spans = entire year token\",\"preceded by month or season\",\"year in date context\"],\"exclusions\":[\"year split across non-marked tokens\",\"marked spans within larger number\",\"decade like 1980s\"]},{\"text\":\"Triggers when marked spans complete a 4-digit year that would otherwise be missing digits.\",\"cues\":[\"context implies missing digits\",\"marked spans fill year gap\",\"surrounding text has date cue\"],\"exclusions\":[\"all digits already present\",\"marked spans redundant\",\"decade expression not year\"]},{\"text\":\"Executes for year tokens where marked spans provide the century or decade digits completing the number.\",\"cues\":[\"marked spans supply century digits\",\"marked spans supply decade digits\",\"year token needs marked digits\"],\"exclusions\":[\"marked spans supply ones digit only\",\"non-year numeric context\",\"century already complete\"]}]}",
+      "round_index": 0,
+      "text": "Computation executes when marked digits form a complete 4-digit year token, not split across other tokens."
+    },
+    "own_test_auc": 0.867919921875,
+    "own_validation_auc": 0.783447265625,
+    "retention_ratio_of_excess_over_chance": 0.4591580708112629
+  },
+  "module_feature_id": "L7B1__module_w2",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.9787175989085948,
+    "interpretation": {
+      "cues": [
+        "year range 1200-2000",
+        "post-medieval dates",
+        "modern era years"
+      ],
+      "exclusions": [
+        "years before 1200",
+        "ancient dynasty dates",
+        "early medieval years"
+      ],
+      "feature_id": "L7B1",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates on four-digit years in the 1800s\u20132000s range, especially mid-19th to late-20th century dates.\",\"cues\":[\"year token 18xx/19xx/20xx\",\"four-digit date pattern\",\"centuries 1800s-2000s\"],\"exclusions\":[\"years before 1700\",\"two-digit numbers\",\"non-year numeric quantities\"]},{\"text\":\"Computation triggers on years embedded in narrative or historical event contexts, not standalone measurements.\",\"cues\":[\"narrative date context\",\"event or biographical context\",\"temporal clause markers\"],\"exclusions\":[\"measurement dimensions\",\"century-only references\",\"isolated numeric values\"]},{\"text\":\"Computation fires on years 1000\u20132000 excluding pre-1200 dates, distinguishing recent historical from ancient dates.\",\"cues\":[\"year range 1200-2000\",\"post-medieval dates\",\"modern era years\"],\"exclusions\":[\"years before 1200\",\"ancient dynasty dates\",\"early medieval years\"]},{\"text\":\"Computation activates on year tokens following prepositions, months, or ordinal dates in sentence contexts.\",\"cues\":[\"year after month name\",\"year after preposition\",\"year in date phrase\"],\"exclusions\":[\"year after AD suffix\",\"year in list of numbers\",\"year as standalone label\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires on years 1000\u20132000 excluding pre-1200 dates, distinguishing recent historical from ancient dates."
+    },
+    "semantics_valid": true,
+    "test_auc": 0.9378662109375,
+    "validation_auc": 0.9473876953125
+  },
+  "status": "SEMANTICS_NOT_RETAINED",
+  "support": [
+    2523,
+    2413
+  ],
+  "top_k": 1
+}

results/llm/qwen2.5_0.5B_pipeline/semantics/L7B2.json

@@ -0,0 +1,67 @@
+{
+  "block_feature_id": "L7B2",
+  "certified_width": 2,
+  "gate": {
+    "parent_min_generalization_ratio": 0.5,
+    "parent_min_test_auc": 0.55,
+    "parent_min_validation_auc": 0.55,
+    "retention_min_auc": 0.55,
+    "retention_min_excess_ratio": 0.5
+  },
+  "layer": 7,
+  "module": {
+    "frozen_parent_predicate_test_auc": 0.857177734375,
+    "own_interpretation": {
+      "cues": [
+        "four-digit year token",
+        "preceded by month name or day",
+        "mid-sentence date context"
+      ],
+      "exclusions": [
+        "non-date numeric tokens",
+        "proper noun fragments",
+        "non-year numbers"
+      ],
+      "feature_id": "L7B2__module_w2",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation activates when parsing four-digit year tokens in date expressions, especially mid-sentence.\",\"cues\":[\"four-digit year token\",\"preceded by month name or day\",\"mid-sentence date context\"],\"exclusions\":[\"non-date numeric tokens\",\"proper noun fragments\",\"non-year numbers\"]},{\"text\":\"Activates on year tokens that complete a full date phrase with month and/or day already present.\",\"cues\":[\"month name before year\",\"day number before year\",\"date phrase completion\"],\"exclusions\":[\"standalone year references\",\"year ranges with dash\",\"years in non-date contexts\"]},{\"text\":\"Fires on year tokens forming the end of a date expression, where the year is the final numeric element.\",\"cues\":[\"year as final date element\",\"numeric year completion\",\"date phrase ending\"],\"exclusions\":[\"year followed by additional numbers\",\"year in compound expression\",\"non-numeric year context\"]},{\"text\":\"Activates on four-digit year tokens that are split or marked across multiple token positions in the text.\",\"cues\":[\"split digit groups\",\"marked year token spans\",\"multi-position year encoding\"],\"exclusions\":[\"single-token years\",\"unmarked numeric sequences\",\"non-year digit groups\"]}]}",
+      "round_index": 0,
+      "text": "Computation activates when parsing four-digit year tokens in date expressions, especially mid-sentence."
+    },
+    "own_test_auc": 0.7947998046875,
+    "own_validation_auc": 0.826416015625,
+    "retention_ratio_of_excess_over_chance": 1.3391304347826087
+  },
+  "module_feature_id": "L7B2__module_w2",
+  "parent": {
+    "generalization_ratio_of_excess_over_chance": 0.6462585034013606,
+    "interpretation": {
+      "cues": [
+        "4-digit year after month/day",
+        "year preceded by date preposition",
+        "year as sentence-final timestamp"
+      ],
+      "exclusions": [
+        "century reference with 'th'",
+        "partial year awaiting more digits",
+        "year mid-list without date context"
+      ],
+      "feature_id": "L7B2",
+      "kind": "when",
+      "metadata": {},
+      "raw_response": "{\"hypotheses\":[{\"text\":\"Computation fires when a 4-digit year token is being read as a complete unit after a date phrase.\",\"cues\":[\"4-digit year after month/day\",\"year preceded by date preposition\",\"year as sentence-final timestamp\"],\"exclusions\":[\"century reference with 'th'\",\"partial year awaiting more digits\",\"year mid-list without date context\"]},{\"text\":\"Computation fires at the final digit of a year that completes a full date expression.\",\"cues\":[\"last digit of 4-digit year\",\"preceded by month name or number\",\"year closes a date phrase\"],\"exclusions\":[\"year in a list of years\",\"year not preceded by date\",\"incomplete year span\"]},{\"text\":\"Computation fires when a year token is the sole temporal anchor of its clause, not part of a year list.\",\"cues\":[\"single year in clause\",\"year tied to event verb\",\"no consecutive year tokens\"],\"exclusions\":[\"year in year-list construction\",\"year followed by another year\",\"century with suffix\"]}]}",
+      "round_index": 0,
+      "text": "Computation fires when a 4-digit year token is being read as a complete unit after a date phrase."
+    },
+    "semantics_valid": true,
+    "test_auc": 0.7667236328125,
+    "validation_auc": 0.9127197265625
+  },
+  "status": "SEMANTICS_RETAINED",
+  "support": [
+    2523,
+    2413
+  ],
+  "top_k": 2
+}