From b8e40ba7a81ce59ecfc7b471a79b117e32c60f1b Mon Sep 17 00:00:00 2001
From: Jens <jens@localhost>
Date: Wed, 29 Jul 2026 23:10:19 +0200
Subject: [PATCH] Prioritize recall in failure-driven sampling

---
 .../test_failure_driven_yolo_sampling.py      | 30 ++++++++++++++++++-
 scripts/build_failure_driven_yolo_sampling.py | 16 +++++++++-
 2 files changed, 44 insertions(+), 2 deletions(-)

diff --git a/backend/tests/test_failure_driven_yolo_sampling.py b/backend/tests/test_failure_driven_yolo_sampling.py
index 71f395f6..19bdb0f8 100644
--- a/backend/tests/test_failure_driven_yolo_sampling.py
+++ b/backend/tests/test_failure_driven_yolo_sampling.py
@@ -163,11 +163,39 @@ def test_sampling_targets_failed_calibration_contexts_without_using_protected_ti
     )
 
     assert paths.count(str(Path("/tmp/industry-pos.png").resolve())) == 5
-    assert paths.count(str(Path("/tmp/industry-neg.png").resolve())) == 6
+    assert paths.count(str(Path("/tmp/industry-neg.png").resolve())) == 1
     assert paths.count(str(Path("/tmp/suburban-pos.png").resolve())) == 3
     assert not any("protected" in path for path in paths)
     assert metadata["weak_recall_contexts"] == ["flanders:industrial"]
     assert metadata["weak_precision_contexts"] == ["flanders:industrial"]
+    assert metadata["recall_dominant_regions"] == ["flanders"]
+
+
+def test_recall_dominance_does_not_suppress_negatives_when_background_gate_failed() -> None:
+    manifest = {"samples": [
+        {"sample_slug": "positive", "split": "train", "region": "flanders", "context": "industrial"},
+        {"sample_slug": "negative", "split": "train", "region": "flanders", "context": "industrial-hard-negative"},
+    ]}
+    summary = {"tiles": [
+        {"sample_slug": "positive", "split": "train", "label_count": 1, "image_path": "/tmp/positive.png"},
+        {"sample_slug": "negative", "split": "train", "label_count": 0, "image_path": "/tmp/negative.png"},
+    ]}
+    assessment = {
+        "status": "continue_training_loop",
+        "gates": {"min_region_f1": .45, "min_region_precision": .5, "min_region_recall": .4,
+                  "max_pure_empty_false_positives": 0},
+        "calibration": {"regions": {
+            "flanders": {"f1": .25, "precision": .4, "recall": .2},
+        }},
+        "background": {"pure_empty_false_positives": 1},
+    }
+
+    paths, metadata = MODULE.build_sampling(
+        summary=summary, manifest=manifest, assessment=assessment, max_region_share=1.0,
+    )
+
+    assert paths.count(str(Path("/tmp/negative.png").resolve())) == 4
+    assert metadata["recall_dominant_regions"] == []
 
 
 def test_region_cap_drops_only_repeats_and_preserves_every_unique_tile() -> None:
diff --git a/scripts/build_failure_driven_yolo_sampling.py b/scripts/build_failure_driven_yolo_sampling.py
index 343d762d..89e701ea 100644
--- a/scripts/build_failure_driven_yolo_sampling.py
+++ b/scripts/build_failure_driven_yolo_sampling.py
@@ -94,6 +94,15 @@ def build_sampling(
         and background["pure_empty_false_positives"]
         > gates["max_pure_empty_false_positives"]
     )
+    recall_dominant_regions = {
+        region
+        for region in weak_recall_regions & weak_precision_regions
+        if not background_failed
+        and (
+            regions[region]["recall"] / gates["min_region_recall"]
+            < regions[region]["precision"] / gates["min_region_precision"]
+        )
+    }
     weak_recall_contexts: set[tuple[str, str]] = set()
     weak_precision_contexts: set[tuple[str, str]] = set()
     for sample_slug, metrics in evaluation.get("samples", {}).items():
@@ -140,7 +149,11 @@ def build_sampling(
             # Precision-only correction still needs positive examples to avoid
             # shifting the classifier toward background and sacrificing recall.
             repeat = precision_positive_repeat
-        if tile["label_count"] == 0 and (background_failed or region in weak_precision_regions):
+        if (
+            tile["label_count"] == 0
+            and (background_failed or region in weak_precision_regions)
+            and region not in recall_dominant_regions
+        ):
             repeat = (
                 context_negative_repeat
                 if context_key in targeted_negative_contexts
@@ -197,6 +210,7 @@ def build_sampling(
         "failure_evidence_source": "test" if assessment.get("test") else "calibration",
         "weak_recall_regions": sorted(weak_recall_regions),
         "weak_precision_regions": sorted(weak_precision_regions),
+        "recall_dominant_regions": sorted(recall_dominant_regions),
         "weak_recall_contexts": [f"{region}:{context}" for region, context in sorted(weak_recall_contexts)],
         "weak_precision_contexts": [f"{region}:{context}" for region, context in sorted(weak_precision_contexts)],
         "targeted_negative_contexts": [