Target related hard-negative contexts

backend/tests/test_failure_driven_yolo_sampling.py

@@ -207,3 +207,35 @@ def test_region_cap_drops_only_repeats_and_preserves_every_unique_tile() -> None
     assert metadata["sampled_entries_by_region"]["flanders"] == 7
     assert metadata["dropped_region_repeat_count"] == 13
     assert metadata["sampled_entries_by_region"]["flanders"] / len(paths) <= .65
+
+
+def test_coastal_precision_failure_targets_port_and_dunes_negatives() -> None:
+    manifest = {"samples": [
+        {"sample_slug": "coastal-train", "split": "train", "region": "flanders", "context": "coastal-urban"},
+        {"sample_slug": "port-negative", "split": "train", "region": "flanders", "context": "port-hard-negative"},
+        {"sample_slug": "dunes-negative", "split": "train", "region": "flanders", "context": "dunes-negative"},
+        {"sample_slug": "coastal-cal", "split": "calibration", "region": "flanders", "context": "coastal-urban"},
+    ]}
+    summary = {"tiles": [
+        {"sample_slug": "coastal-train", "split": "train", "label_count": 2, "image_path": "/tmp/coastal.png"},
+        {"sample_slug": "port-negative", "split": "train", "label_count": 0, "image_path": "/tmp/port.png"},
+        {"sample_slug": "dunes-negative", "split": "train", "label_count": 0, "image_path": "/tmp/dunes.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": .3, "precision": .2, "recall": .4}},
+            "samples": {"coastal-cal": {"f1": .1, "precision": .05, "recall": .2}},
+        },
+    }
+
+    paths, metadata = MODULE.build_sampling(
+        summary=summary, manifest=manifest, assessment=assessment, max_region_share=1.0
+    )
+
+    assert paths.count(str(Path("/tmp/port.png").resolve())) == 6
+    assert paths.count(str(Path("/tmp/dunes.png").resolve())) == 6
+    assert "flanders:port-hard-negative" in metadata["targeted_negative_contexts"]
+    assert "flanders:dunes-negative" in metadata["targeted_negative_contexts"]

docs/BELGIUM_BUILDING_TRAINING_LOOP.md

@@ -113,6 +113,12 @@ The deterministic cap removes only repeated entries and retains every unique
 train tile at least once; manifests record pre-cap counts, final counts and the
 number of dropped repeats. This keeps a weak region prominent without turning
 the national detector into a single-region expert.
+Precision correction expands failed semantic contexts into related negative
+families: coastal urban failures target port/dunes negatives, industrial
+failures target industrial/rail/port negatives, and ribbon/rural/regional
+architecture failures target their governed farmland, forest or quarry
+counterparts. These diagnostic negative repeats are ordered ahead of generic
+repeats so the regional cap cannot discard them first.
 The checkpointed orchestrator invokes this builder after every rejected
 iteration, stores its checksum in `training-loop-state.json`, and uses the
 resulting dataset YAML for the next checkpoint. A restart resumes both the

docs/CODEX_EXECUTION_LOG.md

@@ -19,6 +19,11 @@
   sampling assigned 75.5% of entries to Flanders. The cap retains every unique
   tile, removes repeats only and writes pre/post regional counts into the
   checksummed sampling evidence.
+- Added semantic hard-negative families after the v31 calibration isolated
+  Oostende coastal/port precision as the dominant Flemish error. Coastal,
+  industrial, ribbon, rural and regional-architecture failures now target
+  related train-only negative contexts, with diagnostic repeats preserved
+  ahead of generic repeats when the regional cap applies.
 
 ## 2026-07-27 - Guest demo and product professionalization
 

scripts/build_failure_driven_yolo_sampling.py

@@ -12,6 +12,15 @@ from pathlib import Path
 from typing import Any
 
 
+PRECISION_NEGATIVE_CONTEXTS = {
+    "coastal-urban": {"port-hard-negative", "dunes-negative"},
+    "industrial": {"industrial-hard-negative", "rail-hard-negative", "port-hard-negative"},
+    "ribbon-development": {"farmland-hard-negative", "forest-hard-negative"},
+    "rural-town": {"farmland-hard-negative", "forest-hard-negative", "quarry-hard-negative"},
+    "regional-architecture": {"forest-hard-negative", "quarry-hard-negative"},
+}
+
+
 def file_sha256(path: Path) -> str:
     digest = hashlib.sha256()
     with path.open("rb") as stream:
@@ -94,6 +103,12 @@ def build_sampling(
             and metrics["precision"] < gates["min_region_precision"]
         ):
             weak_precision_contexts.add(key)
+    targeted_negative_contexts = set(weak_precision_contexts)
+    for region, context in weak_precision_contexts:
+        targeted_negative_contexts.update(
+            (region, related)
+            for related in PRECISION_NEGATIVE_CONTEXTS.get(context, set())
+        )
 
     base_paths_by_region: dict[str, list[str]] = {}
     extra_paths_by_region: dict[str, list[str]] = {}
@@ -120,12 +135,19 @@ def build_sampling(
         if tile["label_count"] == 0 and (background_failed or region in weak_precision_regions):
             repeat = (
                 context_negative_repeat
-                if context_key in weak_precision_contexts
+                if context_key in targeted_negative_contexts
                 else negative_repeat
             )
         path = str(Path(tile["image_path"]).resolve())
         base_paths_by_region.setdefault(region, []).append(path)
-        extra_paths_by_region.setdefault(region, []).extend([path] * (repeat - 1))
+        extras = extra_paths_by_region.setdefault(region, [])
+        repeated = [path] * (repeat - 1)
+        if tile["label_count"] == 0 and context_key in targeted_negative_contexts:
+            # Preserve the most diagnostic hard-negative repeats when the
+            # regional cap has to remove lower-priority repetition.
+            extras[:0] = repeated
+        else:
+            extras.extend(repeated)
         selected_samples.add(tile["sample_slug"])
 
     pre_cap_counts = Counter({
@@ -163,6 +185,9 @@ def build_sampling(
         "weak_precision_regions": sorted(weak_precision_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": [
+            f"{region}:{context}" for region, context in sorted(targeted_negative_contexts)
+        ],
         "background_gate_failed": background_failed,
         "positive_repeat": positive_repeat,
         "negative_repeat": negative_repeat,