Separate two-booster evaluation scores from frozen coreset construction

NOTICE.md

@@ -49,6 +49,13 @@ pipeline applies them:
    are recorded output and were not rewritten.
 6. **Corrections after the first build.** LI-Small HT weights were corrected,
    and typology decoding was aligned with the released integer-to-label mapping.
+7. **Evaluation scorer update (2026-09-14).** `ensemble_probs_coreset.npy`
+   now averages the temporally trained LightGBM+GFP and XGBoost+GFP models.
+   The original three-model construction scores are preserved in
+   `construction_probs_coreset.npy`; they include GCPAL+GFP trained with a
+   random split that overlaps the temporal test set. `scoring_metadata.json`
+   records both roles. Targets, graph text, labels, and inclusion weights
+   are unchanged by this update.
 
 ## What is not CDLA
 

README.md

@@ -230,7 +230,9 @@ Files under `extras/<configuration>/` follow the evaluation table's row order.
 | `typologies.npy` | Typology IDs; −1 means unannotated |
 | `ht_weights.npy` | HT importance weights |
 | `ht_subset_indices.npy` | Row positions in the full temporal test split |
-| `ensemble_probs_coreset.npy` | Ensemble illicit probabilities |
+| `ensemble_probs_coreset.npy` | Primary LightGBM + XGBoost ensemble probabilities |
+| `construction_probs_coreset.npy` | Original three-scorer probabilities used to construct the frozen subset |
+| `scoring_metadata.json` | Model members, seeds, thresholds, and provenance for both score arrays |
 | `features_gfp.npz` | 79 graph features per case, feature names, and subset indices |
 | `case_index.parquet` | Case identifiers, labels, and metadata without graph text; its `weight` column corresponds to `ht_weight` |
 
@@ -251,14 +253,18 @@ ds = load_dataset(
 
 </details>
 
-## Evaluated models
+## Evaluation and construction models
 
-The paper evaluates **three supervised baselines and seven LLMs**. LLM task
-metrics use five inference seeds under zero-shot and few-shot prompting on the
-same compact set. The detailed reasoning audit uses the available seed-42 traces.
+The primary ML reference averages **LightGBM+GFP and XGBoost+GFP**, both trained
+with temporal supervision. The original construction ensemble also included
+GCPAL+GFP. The frozen targets and their sampling weights are preserved.
+
+The paper evaluates **seven LLMs** with five inference seeds under zero-shot and
+few-shot prompting on the same compact set. The detailed reasoning audit uses
+the available seed-42 traces.
 
 <details>
-<summary>View all baseline models</summary>
+<summary>View the model list and construction provenance</summary>
 
 GFP stands for Graph Feature Preprocessor.
 
@@ -266,7 +272,7 @@ GFP stands for Graph Feature Preprocessor.
 |---|---|
 | Supervised | LightGBM + GFP |
 | Supervised | XGBoost + GFP |
-| Supervised | Graph Contrastive Pre-training for Anti-money Laundering + GFP (GCPAL+GFP) |
+| Construction scorer | Graph Contrastive Pre-training for Anti-money Laundering + GFP (GCPAL+GFP) |
 | LLM | OpenAI GPT-OSS-20B |
 | LLM | OpenAI GPT-OSS-120B |
 | LLM | NVIDIA Nemotron-3-Nano-30B-A3B |
@@ -276,8 +282,9 @@ GFP stands for Graph Feature Preprocessor.
 | LLM | Qwen3.5-397B-A17B |
 
 GCPAL's random fine-tuning split overlaps roughly 60% of the temporal test
-edges; the ensemble includes this member. Released thresholds are test-selected
-operating points. See the
+edges. Its predictions are retained for construction provenance. The primary
+evaluation ensemble excludes GCPAL and uses the original test-selected
+operating points (0.80 for HI-Small; 0.48 for LI-Small). See the
 [baseline results](https://github.com/nat-nischw/AMLworld-Compact#baselines)
 and paper for the evaluation protocol and interpretation.
 

extras/HI-Small/construction_probs_coreset.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35fa27b0eb390326d9fbcf7a3b468957fec379821aa17fae9eebbbd4d41b9f9b
+size 30152

extras/HI-Small/ensemble_probs_coreset.npy

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:35fa27b0eb390326d9fbcf7a3b468957fec379821aa17fae9eebbbd4d41b9f9b
+oid sha256:4817e24e8ba31255304935b5442685238584e9adce26ea3a86c241c8dee010f0
 size 30152

extras/HI-Small/scoring_metadata.json

@@ -0,0 +1,39 @@
+{
+  "dataset": "HI-Small",
+  "evaluation": {
+    "file": "ensemble_probs_coreset.npy",
+    "members": [
+      "LightGBM+GFP",
+      "XGBoost+GFP"
+    ],
+    "seeds": [
+      42,
+      123,
+      456,
+      789,
+      1011
+    ],
+    "aggregation": "arithmetic mean over members and seeds",
+    "threshold": 0.8,
+    "threshold_provenance": "inherited from original construction; not retuned",
+    "training_split": "temporal 60/20/20"
+  },
+  "construction": {
+    "file": "construction_probs_coreset.npy",
+    "members": [
+      "LightGBM+GFP",
+      "XGBoost+GFP",
+      "GCPAL+GFP"
+    ],
+    "seeds": [
+      42,
+      123,
+      456,
+      789,
+      1011
+    ],
+    "threshold": 0.8,
+    "frozen": true,
+    "note": "GCPAL used random-split fine-tuning and temporal inference; these historical scores determine the existing strata, targets, and HT weights only."
+  }
+}

extras/LI-Small/construction_probs_coreset.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:476e57d6965faae193cadbf79e03bf3f1581875e6af954e60fc3ee0de532c21c
+size 18272

extras/LI-Small/ensemble_probs_coreset.npy

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:476e57d6965faae193cadbf79e03bf3f1581875e6af954e60fc3ee0de532c21c
+oid sha256:3f200a4422f2c38f372489565237e2038e000474c6b05138f2d69f48d885ddaf
 size 18272

extras/LI-Small/scoring_metadata.json

@@ -0,0 +1,39 @@
+{
+  "dataset": "LI-Small",
+  "evaluation": {
+    "file": "ensemble_probs_coreset.npy",
+    "members": [
+      "LightGBM+GFP",
+      "XGBoost+GFP"
+    ],
+    "seeds": [
+      42,
+      123,
+      456,
+      789,
+      1011
+    ],
+    "aggregation": "arithmetic mean over members and seeds",
+    "threshold": 0.48,
+    "threshold_provenance": "inherited from original construction; not retuned",
+    "training_split": "temporal 60/20/20"
+  },
+  "construction": {
+    "file": "construction_probs_coreset.npy",
+    "members": [
+      "LightGBM+GFP",
+      "XGBoost+GFP",
+      "GCPAL+GFP"
+    ],
+    "seeds": [
+      42,
+      123,
+      456,
+      789,
+      1011
+    ],
+    "threshold": 0.48,
+    "frozen": true,
+    "note": "GCPAL used random-split fine-tuning and temporal inference; these historical scores determine the existing strata, targets, and HT weights only."
+  }
+}