Stabilize near-gate regional sampling

backend/tests/test_failure_driven_yolo_sampling.py

@@ -251,6 +251,33 @@ def test_region_cap_rotates_repeats_between_sampling_rounds() -> None:
     assert second_metadata["sampling_round"] == 2
 
 
+def test_precision_guard_band_keeps_near_gate_region_stabilized() -> None:
+    manifest = {"samples": [
+        {"sample_slug": "wa-positive", "split": "train", "region": "wallonia", "context": "rural-town"},
+        {"sample_slug": "wa-negative", "split": "train", "region": "wallonia", "context": "farmland-hard-negative"},
+    ]}
+    summary = {"tiles": [
+        {"sample_slug": "wa-positive", "split": "train", "label_count": 1, "image_path": "/tmp/wa-positive.png"},
+        {"sample_slug": "wa-negative", "split": "train", "label_count": 0, "image_path": "/tmp/wa-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": {
+            "wallonia": {"f1": .6, "precision": .52, "recall": .7},
+        }},
+    }
+
+    paths, metadata = MODULE.build_sampling(
+        summary=summary, manifest=manifest, assessment=assessment, max_region_share=1.0,
+    )
+
+    assert "wallonia" in metadata["weak_precision_regions"]
+    assert paths.count(str(Path("/tmp/wa-negative.png").resolve())) == 4
+    assert metadata["precision_guard_band"] == .03
+
+
 def test_coastal_precision_failure_targets_port_and_dunes_negatives() -> None:
     manifest = {"samples": [
         {"sample_slug": "coastal-train", "split": "train", "region": "flanders", "context": "coastal-urban"},

docs/CODEX_EXECUTION_LOG.md

@@ -11944,6 +11944,12 @@ Verified:
   unique train tile and all protected-split exclusions.
 - `py -3 -m pytest -q backend/tests/test_failure_driven_yolo_sampling.py backend/tests/test_belgium_training_loop.py`
   (`20 passed`).
+- Iteration 3 raised aggregate calibration F1 to `0.605`, passed Brussels and
+  every Wallonia gate, and isolated the remaining blocker to Flanders. When
+  iteration 4 removed Wallonia stabilization, Wallonia precision regressed
+  narrowly from `0.525` to `0.496`. Added a `0.03` precision/recall sampling
+  guard-band so a just-passing region retains stabilizing evidence while the
+  hard-failing region remains the primary target (`21 passed`).
 
 Open:
 

docs/TODO.md

@@ -1020,5 +1020,6 @@ This file now starts with the current implementation status. Older preparation/b
 - [x] Render and inspect a 56-tile contact sheet covering all 14 new AOIs before retraining.
 - [ ] Run the v43 calibration-first, failure-driven CUDA loop against the frozen regional release gates.
 - [x] Rotate region-capped repeat windows deterministically per loop round so persistent failures cannot reuse an identical training list indefinitely.
+- [x] Keep near-gate regional precision/recall stabilization inside a 0.03 sampling guard-band to prevent cross-region seesaw regressions.
 - [ ] Open protected test and pure-background evidence only after every calibration gate passes.
 - [ ] Queue final representative human sign-off only after all automated gates pass, then promote and redeploy the exact checksummed model.

scripts/build_failure_driven_yolo_sampling.py

@@ -51,6 +51,8 @@ def build_sampling(
     context_negative_repeat: int = 6,
     max_region_share: float = 0.65,
     sampling_round: int = 0,
+    precision_guard_band: float = 0.03,
+    recall_guard_band: float = 0.03,
 ) -> tuple[list[str], dict[str, Any]]:
     if assessment.get("status") != "continue_training_loop":
         raise ValueError("Failure-driven sampling requires a failed assessment")
@@ -66,6 +68,8 @@ def build_sampling(
         raise ValueError("max_region_share must be in (0, 1]")
     if sampling_round < 0:
         raise ValueError("sampling_round must be non-negative")
+    if precision_guard_band < 0 or recall_guard_band < 0:
+        raise ValueError("Guard bands must be non-negative")
 
     samples = {item["sample_slug"]: item for item in manifest["samples"]}
     gates = assessment["gates"]
@@ -77,12 +81,12 @@ def build_sampling(
         region
         for region, metrics in regions.items()
         if metrics["f1"] < gates["min_region_f1"]
-        or metrics["recall"] < gates["min_region_recall"]
+        or metrics["recall"] < gates["min_region_recall"] + recall_guard_band
     }
     weak_precision_regions = {
         region
         for region, metrics in regions.items()
-        if metrics["precision"] < gates["min_region_precision"]
+        if metrics["precision"] < gates["min_region_precision"] + precision_guard_band
     }
     background = assessment.get("background")
     background_failed = bool(
@@ -206,6 +210,8 @@ def build_sampling(
         "context_negative_repeat": context_negative_repeat,
         "max_region_share": max_region_share,
         "sampling_round": sampling_round,
+        "precision_guard_band": precision_guard_band,
+        "recall_guard_band": recall_guard_band,
         "source_train_tile_count": sum(
             1
             for tile in summary["tiles"]
@@ -235,6 +241,8 @@ def main() -> int:
     parser.add_argument("--context-negative-repeat", type=int, default=6)
     parser.add_argument("--max-region-share", type=float, default=0.65)
     parser.add_argument("--sampling-round", type=int)
+    parser.add_argument("--precision-guard-band", type=float, default=0.03)
+    parser.add_argument("--recall-guard-band", type=float, default=0.03)
     args = parser.parse_args()
 
     summary = json.loads(args.summary.read_text(encoding="utf-8"))
@@ -255,6 +263,8 @@ def main() -> int:
         context_negative_repeat=args.context_negative_repeat,
         max_region_share=args.max_region_share,
         sampling_round=sampling_round,
+        precision_guard_band=args.precision_guard_band,
+        recall_guard_band=args.recall_guard_band,
     )
     args.output_dir.mkdir(parents=True, exist_ok=True)
     train_list = args.output_dir / "train-failure-driven.txt"