test(server): final smoke with the calibrated gates live

Still endpoint: 111 detections, 44 named, 42 correct, 95.5% precision, 2 wrong
(both unknown facings the local lane matches geometrically).

Stream, 20 s pan 12: 62 tracks labelled, 0 label changes, 0 unexplained, 17/27
facings named at 84% precision, 0 wrong, 0 unenrolled false accepts.

Both recordings committed under server-smoke/ and stream-gated/.

recognization/.gitignore

@@ -25,3 +25,14 @@ stock-*/identity_head.pt
 .gallery_cache_*.npz
 .demo-token
 config/local.env
+
+# Fetched images. Every set keeps its manifest.json (source URL, licence,
+# artist) and its results in git; the pixels are re-fetchable from those URLs
+# and are 59 MB for the Wikimedia shelves alone.
+data/external-shelves/*/*.jpg
+data/external-shelves/*/*.jpeg
+data/external-shelves/*/*.JPG
+data/external-shelves/*/*.png
+data/packshots/research/*/images/
+data/packshots/research/*/pages/
+data/packshots/research/*/*.jpg

recognization/code/app.py

@@ -1010,8 +1010,9 @@ def main() -> None:
     parser.add_argument("--live-method", type=str, default=None, choices=list(CLASSIFY_METHODS))
     parser.add_argument("--live-conf", type=float, default=None)
     parser.add_argument("--identity", type=str, default=None,
-                        choices=["matcher", "local", "gallery"],
-                        help="which mechanism names products (default local)")
+                        choices=["matcher", "local", "gallery", "van"],
+                        help="which mechanism names products (default local; "
+                             "van is the always-argmax baseline)")
     parser.add_argument("--live-inflight", type=int, default=None,
                         help="classifications in flight per session")
     parser.add_argument("--live-attempts", type=int, default=None,

recognization/code/bench_ocr.py

@@ -0,0 +1,42 @@
+"""OCR cost per crop, on production-shaped crops.
+
+The decision record claims ~18-26 ms per crop on Mac and re-measured 18.0 ms on
+40 production crops. This is that measurement, so the number is reproducible
+rather than remembered.
+
+    .venv/bin/python code/bench_ocr.py
+"""
+import json, statistics, sys, time
+sys.path.insert(0, 'code')
+import cv2
+from eval_identity import LocalLaneAdapter, crop_of
+
+lane = LocalLaneAdapter('stock-lopia-clean')
+labels = json.load(open('eval/labels/image6.json'))
+image = cv2.imread('add/image6.jpg')
+
+crops = []
+for facing in labels['facings']:
+    if facing.get('label') in (None, 'exclude'):
+        continue
+    crop = crop_of(image, facing['box'])
+    if crop is not None and crop.size:
+        crops.append(crop)
+    if len(crops) >= 40:
+        break
+
+# Warm the recognizer: the first call loads the Vision model.
+lane.text_lane.read(crops[0])
+
+times = []
+for crop in crops:
+    started = time.perf_counter()
+    lane.text_lane.read(crop)
+    times.append((time.perf_counter() - started) * 1000)
+
+times.sort()
+print(f'{len(times)} crops')
+print(f'  p50 {statistics.median(times):.1f} ms')
+print(f'  mean {statistics.mean(times):.1f} ms')
+print(f'  p95 {times[int(len(times) * 0.95)]:.1f} ms')
+print(f'  min {times[0]:.1f}  max {times[-1]:.1f}')

recognization/code/collect_shelf_images.py

@@ -129,7 +129,9 @@ def image_info(titles: list) -> dict:
 
 
 def main() -> None:
-    root = sys.argv[1] if len(sys.argv) > 1 else os.path.expanduser("~/dorayaki-data/shelf-images")
+    default = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
+                           "data", "external-shelves", "wikimedia-commons")
+    root = sys.argv[1] if len(sys.argv) > 1 else default
     os.makedirs(root, exist_ok=True)
 
     print("searching Commons...")
@@ -139,7 +141,7 @@ def main() -> None:
     info = image_info(sorted(found))
     print(f"  metadata for {len(info)} files")
 
-    kept, skipped = [], {"license": 0, "size": 0, "missing": 0}
+    kept, skipped = [], {"license": 0, "size": 0, "missing": 0, "not_a_photo": 0}
     for title, entry in sorted(info.items()):
         if not entry.get("url"):
             skipped["missing"] += 1
@@ -148,6 +150,12 @@ def main() -> None:
         if not any(token in licence for token in PERMISSIVE):
             skipped["license"] += 1
             continue
+        # A book scan is not a shelf. The search terms pull in .djvu volumes
+        # whose page 1 is a title page, and Commons serves those as JPEG
+        # thumbnails, so the extension check has to be on the source title.
+        if any(title.lower().endswith(ext) for ext in (".djvu", ".pdf", ".svg", ".tif", ".tiff")):
+            skipped["not_a_photo"] += 1
+            continue
         if (entry.get("width") or 0) < 1000 or (entry.get("height") or 0) < 700:
             skipped["size"] += 1
             continue
@@ -156,9 +164,13 @@ def main() -> None:
     print(f"  keeping {len(kept)} (skipped: {skipped})")
     manifest = []
     for index, (title, entry) in enumerate(kept):
-        stem = os.path.basename(urllib.parse.urlparse(entry["url"]).path).replace(" ", "_")
-        stem = urllib.parse.unquote(stem)[:60]
-        name = f"{index:03d}_{stem}"
+        # Truncate the stem, never the extension: a 60-character cut on the
+        # whole basename drops it, and the file then has no type for cv2 or
+        # for anyone reading the directory.
+        stem, extension = os.path.splitext(
+            os.path.basename(urllib.parse.urlparse(entry["url"]).path).replace(" ", "_")
+        )
+        name = f"{index:03d}_{urllib.parse.unquote(stem)[:60]}{extension.lower()}"
         path = os.path.join(root, name)
         if not os.path.exists(path):
             try:

recognization/code/engine.py

@@ -89,8 +89,11 @@ class EngineSettings:
     # calibrated open-set gate: measured on the corrected image6 labels it shows
     # 36 of 61 in-catalog facings at 95% precision with the text lane, against
     # 35 at 35% for Van's always-argmax matcher. `matcher` keeps Van's composite
-    # as a selectable baseline lane (E-1). `gallery` is fenced off: raw DINOv2
-    # is 12-21% top-1 on 7 SKUs and its recall@10 falls to 7/42 at 493 SKUs.
+    # as a selectable baseline lane (E-1). `van` is that same scorer with the
+    # gate removed - it answers its top SKU every time, which is Van's actual
+    # contract and the only way to reproduce the stream comparison in
+    # DECISION.md. `gallery` is fenced off: raw DINOv2 is 12-21% top-1 on 7
+    # SKUs and its recall@10 falls to 7/42 at 493 SKUs.
     identity_backend: str = "local"
     # Cosine similarity below which the gallery calls a facing unknown. Measured:
     # same-product pairs sit above 0.6, different products below 0.5.
@@ -469,6 +472,8 @@ class RecognitionEngine:
             return "cosine"
         if backend == "local":
             return "local"
+        if backend == "van":
+            return "van"
         return "inlier"
 
     def identity_text_lane(self) -> Any:

recognization/code/eval_distractors.py

@@ -0,0 +1,116 @@
+"""Open-world false-accept probe: unrelated shelf photos through the shipped lane.
+
+data/external-shelves/ holds Wikimedia Commons supermarket photos of shelves
+that share nothing with the LOPIA catalog. Every accept here is a false accept
+by construction, which is what makes the set useful: the image6 shelf can only
+measure false accepts on the products it happens to contain.
+
+The stored eval.json predates the open-set gate, so it is re-run rather than
+trusted.
+
+    .venv/bin/python code/eval_distractors.py --stock stock-lopia-clean \
+        data/external-shelves/wikimedia-commons
+"""
+
+from __future__ import annotations
+
+import argparse
+import glob
+import json
+import os
+import sys
+import time
+from typing import Any, Dict, List
+
+current_dir = os.path.dirname(os.path.abspath(__file__))
+if current_dir not in sys.path:
+    sys.path.insert(0, current_dir)
+
+import cv2  # noqa: E402
+
+from eval_identity import LOCAL_ACCEPT, LocalLaneAdapter  # noqa: E402
+
+
+def main(argv: List[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("directories", nargs="+")
+    parser.add_argument("--stock", default="stock-lopia-clean")
+    parser.add_argument("--write", action="store_true")
+    parser.add_argument("--limit", type=int, default=0, help="images per set (0 = all)")
+    args = parser.parse_args(argv)
+
+    from engine import build_default_engine, EngineSettings
+
+    engine = build_default_engine(EngineSettings(stock_dir=args.stock))
+    lane = LocalLaneAdapter(args.stock)
+    for directory in args.directories:
+        manifest_path = os.path.join(directory, "manifest.json")
+        try:
+            with open(manifest_path) as handle:
+                manifest = json.load(handle)
+        except (OSError, ValueError):
+            manifest = {"images": []}
+        meta = {row["file"]: row for row in manifest.get("images", [])}
+        rows: List[Dict[str, Any]] = []
+
+        paths = sorted(
+            path for path in glob.glob(os.path.join(directory, "*"))
+            if os.path.splitext(path)[1].lower() in (".jpg", ".jpeg", ".png")
+        )
+        if args.limit:
+            paths = paths[: args.limit]
+        for path in paths:
+            name = os.path.basename(path)
+            image = cv2.imread(path)
+            if image is None:
+                continue
+            started = time.perf_counter()
+            detections = engine.detect(image)
+            accepted = 0
+            for detection in detections:
+                box = detection.get("box") or detection.get("bbox")
+                if not box:
+                    continue
+                x1, y1, x2, y2 = [int(v) for v in box]
+                crop = image[max(0, y1):y2, max(0, x1):x2]
+                if crop.size == 0:
+                    continue
+                decision = lane.decide(crop, lane.rank(crop), accept=LOCAL_ACCEPT,
+                                       margin=0.0, scale="full")
+                if decision.get("best_sku"):
+                    accepted += 1
+            rows.append({
+                "file": name,
+                "title": meta.get(name, {}).get("title", ""),
+                "detections": len(detections),
+                "false_accepts": accepted,
+                "ms": round((time.perf_counter() - started) * 1000),
+            })
+
+        detections = sum(r["detections"] for r in rows)
+        accepts = sum(r["false_accepts"] for r in rows)
+        report = {
+            "dataset": os.path.basename(directory.rstrip("/")),
+            "catalog": args.stock,
+            "note": "Unrelated shelves: every accept is a false accept by construction.",
+            "images": len(rows),
+            "detections": detections,
+            "false_accepts": accepts,
+            "false_accept_rate": round(accepts / detections, 4) if detections else None,
+            "rows": rows,
+        }
+        print(f"{report['dataset']}: {len(rows)} images, {detections} detections, "
+              f"{accepts} false accepts ({report['false_accept_rate'] or 0:.2%})")
+        for row in sorted(rows, key=lambda r: -r["false_accepts"])[:5]:
+            if row["false_accepts"]:
+                print(f"    {row['false_accepts']:>3}  {row['file'][:55]}")
+        if args.write:
+            out = os.path.join(directory, "eval_result.json")
+            with open(out, "w") as handle:
+                json.dump(report, handle, ensure_ascii=False, indent=1)
+            print(f"    wrote {out}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

recognization/code/eval_external.py

@@ -0,0 +1,127 @@
+"""Score external-proxy image sets through the shipped identity lane.
+
+These sets are not Tier 2 (see data/external-eval/README.md). They exist to test
+the pipeline on data it had no part in building, so the one thing this script
+must never do is score an image that has since been enrolled - that is
+training-set accuracy wearing a held-out label.
+
+`leaked()` is the same check the shelf evaluator runs, and it is applied here
+for that reason: enrolling a packshot from an external set is a normal thing to
+do (it is how F-12 was closed), and it silently converts one of these rows into
+a leak unless something refuses it.
+
+    .venv/bin/python code/eval_external.py --stock stock-lopia-clean \
+        data/external-eval/kewpie data/external-eval/other-mayo
+"""
+
+from __future__ import annotations
+
+import argparse
+import glob
+import json
+import os
+import sys
+from typing import Any, Dict, List
+
+current_dir = os.path.dirname(os.path.abspath(__file__))
+if current_dir not in sys.path:
+    sys.path.insert(0, current_dir)
+
+import cv2  # noqa: E402
+
+from eval_identity import LOCAL_ACCEPT, LocalLaneAdapter, leaked  # noqa: E402
+
+
+def score_set(lane: Any, directory: str, stock_dir: str) -> Dict[str, Any]:
+    rows: List[Dict[str, Any]] = []
+    for path in sorted(glob.glob(os.path.join(directory, "*.jpg"))):
+        name = os.path.basename(path)
+        truth = "sku_" + name.split("_")[1]
+        image = cv2.imread(path)
+        if image is None:
+            rows.append({"file": name, "truth": truth, "verdict": "unreadable"})
+            continue
+        # The whole image is the crop: an external packshot is not a shelf facing.
+        box = (0, 0, image.shape[1], image.shape[0])
+        leak = leaked(image, box, stock_dir)
+        if leak:
+            rows.append({
+                "file": name, "truth": truth, "verdict": "leaked",
+                "leaked_into": leak,
+                "note": "enrolled in the catalog; scoring it would be training accuracy",
+            })
+            continue
+        decision = lane.decide(
+            image, lane.rank(image), accept=LOCAL_ACCEPT, margin=0.0, scale="full"
+        )
+        best = decision.get("best_sku")
+        verdict = "correct" if best == truth else ("abstain" if not best else "wrong")
+        rows.append({
+            "file": name, "truth": truth, "best": best,
+            "confidence": decision.get("confidence"),
+            "status": decision.get("status"),
+            "source": decision.get("source"),
+            "verdict": verdict,
+        })
+
+    scored = [r for r in rows if r["verdict"] in ("correct", "abstain", "wrong")]
+    counts = {v: sum(1 for r in scored if r["verdict"] == v)
+              for v in ("correct", "abstain", "wrong")}
+    # Per SKU, because a set total hides which products the lane can and cannot
+    # name. Every SKU here has images from two independent source domains, so a
+    # SKU that scores the same on both is not one lucky photograph.
+    by_sku: Dict[str, Dict[str, int]] = {}
+    for row in scored:
+        bucket = by_sku.setdefault(row["truth"], {"correct": 0, "abstain": 0, "wrong": 0})
+        bucket[row["verdict"]] += 1
+    return {
+        "dataset": os.path.basename(directory.rstrip("/")),
+        "catalog": stock_dir,
+        "note": "External proxy images, not Tier 2. Scored through the shipped "
+                "LocalLaneAdapter + DecisionPolicy.",
+        "rows": rows,
+        "scored": len(scored),
+        "leaked": sum(1 for r in rows if r["verdict"] == "leaked"),
+        "by_sku": by_sku,
+        **counts,
+    }
+
+
+def main(argv: List[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("directories", nargs="+", help="external set directories")
+    parser.add_argument("--stock", default="stock-lopia-clean")
+    parser.add_argument("--write", action="store_true",
+                        help="write eval_result.json into each set directory")
+    args = parser.parse_args(argv)
+
+    lane = LocalLaneAdapter(args.stock)
+    totals = {"correct": 0, "abstain": 0, "wrong": 0, "leaked": 0, "scored": 0}
+    for directory in args.directories:
+        report = score_set(lane, directory, args.stock)
+        for key in totals:
+            totals[key] += report[key]
+        print(f"{report['dataset']}: correct {report['correct']} | "
+              f"abstain {report['abstain']} | wrong {report['wrong']} | "
+              f"leaked {report['leaked']} of {report['scored']} scored")
+        for sku, bucket in sorted(report["by_sku"].items()):
+            print(f"    {sku:<10} correct {bucket['correct']}  "
+                  f"abstain {bucket['abstain']}  wrong {bucket['wrong']}")
+        for row in report["rows"]:
+            if row["verdict"] in ("wrong", "leaked"):
+                print(f"    {row['verdict'].upper():<7} {row['file']} "
+                      f"truth={row['truth']} best={row.get('best')} "
+                      f"{row.get('leaked_into', '')}")
+        if args.write:
+            out = os.path.join(directory, "eval_result.json")
+            with open(out, "w") as handle:
+                json.dump(report, handle, ensure_ascii=False, indent=1)
+            print(f"    wrote {out}")
+    print(f"\nTOTAL: {totals['scored']} scored, correct {totals['correct']}, "
+          f"abstain {totals['abstain']}, FALSE ACCEPTS {totals['wrong']}, "
+          f"leaked {totals['leaked']}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

recognization/code/eval_external_images.py

@@ -1,10 +1,14 @@
 """
 Run the pipeline over store photos it has never seen.
 
-    python code/eval_external_images.py [~/dorayaki-data/shelf-images]
+    python code/eval_external_images.py [data/external-shelves/wikimedia-commons]
 
 Reports detection and identity per image, and writes a JSON summary. See
 docs/EXTERNAL-EVAL.md for what the numbers mean.
+
+The images live in the repository (`data/external-shelves/`), not in a home
+directory: they are evaluation evidence, and evidence that only exists on one
+machine cannot be re-checked.
 """
 import json, os, sys, time
 sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
@@ -13,8 +17,9 @@ from engine import build_default_engine, EngineSettings
 from tags import TagService, TagSettings
 from reconcile import attach_tags_to_boxes
 
-ROOT = os.path.expanduser(sys.argv[1] if len(sys.argv) > 1 else "~/dorayaki-data/shelf-images")
-manifest = json.load(open(os.path.join(ROOT, "manifest.json")))
+DEFAULT_ROOT = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
+                            "data", "external-shelves", "wikimedia-commons")
+ROOT = os.path.expanduser(sys.argv[1] if len(sys.argv) > 1 else DEFAULT_ROOT)
 
 engine = build_default_engine(EngineSettings(stock_dir="stock-lopia"))
 catalog_by_sku = {p["sku_id"]: p for p in engine.catalog()}

recognization/code/eval_identity.py

@@ -78,6 +78,12 @@ LIVE_LONG_SIDE = 1280
 # the enrollment protocol's session manifest (DECISION §8); this is the cheap
 # backstop that runs without one.
 LEAK_SIGNATURE_MAD = 20.0
+# A signature hit is confirmed by correlation before it counts as a leak. A
+# genuine re-cut, re-encode or rescale of an enrolled view correlates at 0.99+;
+# two different photographs of the same product sit near 0.3. Measured on the
+# l09 pair that the signature alone misjudged (0.28) and on re-encoded and
+# rescaled copies of the enrolled view (0.999, 0.997).
+LEAK_CORRELATION = 0.90
 
 
 def load_labels(path: str) -> Dict[str, Any]:
@@ -219,6 +225,24 @@ def _catalog_templates(stock_dir: str) -> List[Tuple[str, str, Tuple[int, int],
     return cached
 
 
+def _correlation(left: Any, right: Any) -> float:
+    """Mean-subtracted correlation of two images, resizing `right` to `left`.
+
+    Scale- and re-encode-tolerant, unlike a pixel comparison: the enrolled view
+    and a re-saved or rescaled copy of it correlate at 0.99+, while two
+    different photographs of the same product sit near 0.3.
+    """
+    if right.shape[:2] != left.shape[:2]:
+        right = cv2.resize(right, (left.shape[1], left.shape[0]),
+                           interpolation=cv2.INTER_AREA)
+    a = cv2.cvtColor(left, cv2.COLOR_BGR2GRAY).astype(np.float32).ravel()
+    b = cv2.cvtColor(right, cv2.COLOR_BGR2GRAY).astype(np.float32).ravel()
+    a -= a.mean()
+    b -= b.mean()
+    denom = float(np.linalg.norm(a) * np.linalg.norm(b))
+    return float((a * b).sum() / denom) if denom else 0.0
+
+
 def leaked(image: Any, box: Sequence[int], stock_dir: str) -> Optional[str]:
     """Name of the catalog template this crop is, or None when it is held out.
 
@@ -238,9 +262,21 @@ def leaked(image: Any, box: Sequence[int], stock_dir: str) -> Optional[str]:
         return None
     crop_signatures = _signatures(crop)
     for stem, name, shape, signatures in _catalog_templates(stock_dir):
-        if min(float(np.abs(signature - crop_signature).mean())
-               for signature in signatures for crop_signature in crop_signatures) < LEAK_SIGNATURE_MAD:
-            return stem
+        close = min(float(np.abs(signature - crop_signature).mean())
+                    for signature in signatures for crop_signature in crop_signatures)
+        if close < LEAK_SIGNATURE_MAD:
+            # The signature is a 64x64 grayscale thumbnail, so it is coarse
+            # enough to fire on two different photos of the same product: the
+            # l09 rear label from a retailer page sits 19.5 from the enrolled
+            # front packshot, under the 20.0 line, while being a different
+            # photograph (resized MAD 26.3). Confirm before calling it a leak.
+            # A genuine re-cut or re-encode correlates at 0.99+; that pair
+            # correlates at 0.28, so the two signals together are decisive
+            # where either alone is not.
+            template = cv2.imread(os.path.join(stock_dir, name))
+            if template is None or _correlation(crop, template) >= LEAK_CORRELATION:
+                return stem
+            continue
         if shape != crop.shape[:2]:
             continue
         template = cv2.imread(os.path.join(stock_dir, name))
@@ -377,10 +413,14 @@ class LocalLaneAdapter(Lane):
 
     name = "local"
 
-    def __init__(self, stock_dir: str, ocr: Any = None) -> None:
+    def __init__(self, stock_dir: str, ocr: Any = None, use_text: bool = True) -> None:
         from identity_lanes import LocalLane, TextLane
 
         self.lane = LocalLane(stock_dir)
+        # E-3 asks what the text lane is worth, which is only answerable by
+        # running the same crops without it. Building the lane anyway keeps the
+        # comparison to one variable: the OCR cost is paid either way.
+        self.use_text = use_text
         self.text_lane = TextLane.from_catalog(
             _catalog_rows(stock_dir), stock_dir, ocr if ocr is not None else _ocr_engine()
         )
@@ -392,6 +432,10 @@ class LocalLaneAdapter(Lane):
                margin: float, scale: str) -> Dict[str, Any]:
         from identity_lanes import DecisionPolicy
 
+        if not self.use_text:
+            from identity import decide_identity
+
+            return decide_identity(dict(ranked), scale="local", min_score=accept)
         read = self.text_lane.read(crop)
         text_rank = sorted(
             self.text_lane.scores_from_text(read).items(),
@@ -605,6 +649,8 @@ def main() -> int:
     parser.add_argument("--lanes", default="local,matcher,gallery,hybrid-verify")
     parser.add_argument("--shortlist", type=int, default=5)
     parser.add_argument("--accept", type=float, default=None, help="override the lane's own floor")
+    parser.add_argument("--no-text", action="store_true",
+                        help="run the local lane without the text lane (E-3)")
     parser.add_argument("--margin", type=float, default=None)
     parser.add_argument("--live", action="store_true", help="also score at live resolution")
     parser.add_argument("--json", help="write the report here")
@@ -644,7 +690,7 @@ def main() -> int:
     if "van" in wanted:
         lanes["van"] = VanLane(args.stock_dir)
     if "local" in wanted:
-        lanes["local"] = LocalLaneAdapter(args.stock_dir)
+        lanes["local"] = LocalLaneAdapter(args.stock_dir, use_text=not args.no_text)
     if "gallery" in wanted or "hybrid-verify" in wanted:
         lanes["gallery"] = GalleryLane(args.stock_dir, catalog_skus)
     if "hybrid-verify" in wanted:

recognization/code/finetune_identity.py

@@ -292,7 +292,9 @@ def main() -> None:
     parser.add_argument("--epochs", type=int, default=60)
     parser.add_argument("--views", type=int, default=2, help="enrolment views per class")
     parser.add_argument("--augment", type=int, default=12, help="augmented copies per enrolment view")
-    parser.add_argument("--image", default=os.path.expanduser("~/dorayaki-data/shelf-images"))
+    parser.add_argument("--image", default=os.path.join(
+        os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
+        "data", "external-shelves", "wikimedia-commons"))
     parser.add_argument("--stock", default=None,
                         help="train on a store's enrolled views instead of the milk shelf, "
                              "and write <stock>/identity_head.pt for the gallery to load")

recognization/code/identity.py

@@ -30,8 +30,11 @@ DEFAULT_MIN_CONFIDENCE = 0.55
 
 # `local` scale (identity_lanes.LocalLane). The value is already a margin over
 # the runner-up, so the relative-margin test does not apply to it: a single
-# absolute line is the whole gate.
-LOCAL_ACCEPT = 2.0
+# absolute line is the whole gate. Kept equal to identity_lanes.LOCAL_ACCEPT,
+# which is what the engine's DecisionPolicy uses; this copy exists for the
+# evaluator's `--no-text` path, and two different lines would score the same
+# crops two ways.
+LOCAL_ACCEPT = 4.0
 
 
 def relative_confidence(top_score: float, runner_up_score: float) -> float:
@@ -248,6 +251,26 @@ def decide_identity(
             "message": message,
         }
 
+    if scale == "van":
+        # Van's contract: the top SKU, always answered. No gate, no abstention -
+        # this is the baseline the decision is measured against, and it exists
+        # so the stream comparison is reproducible rather than asserted.
+        if top_score <= 0.0:
+            level, message, best = IDENTITY_REJECTED, "No matching evidence", None
+        else:
+            level, message, best = IDENTITY_ACCEPTED, "Always-argmax baseline", top_sku
+        return {
+            "status": "success" if level == IDENTITY_ACCEPTED else level,
+            "method": method,
+            "best_sku": best,
+            "score": round(top_score, 2),
+            "confidence": confidence,
+            "identity_level": level,
+            "ranked_candidates": ranked,
+            "details": top_details,
+            "message": message,
+        }
+
     if top_score <= 0.0:
         level = IDENTITY_REJECTED
         message = "No matching evidence"

recognization/code/identity_lanes.py

@@ -28,7 +28,7 @@ import os
 import re
 import sys
 from dataclasses import dataclass
-from typing import Any, Dict, List, Optional, Sequence, Tuple
+from typing import Any, Dict, List, Optional, Sequence, Set, Tuple
 
 current_dir = os.path.dirname(os.path.abspath(__file__))
 if current_dir not in sys.path:
@@ -43,7 +43,14 @@ from rebuild_catalog import bigrams, normalize_text, signature_bigrams, sku_of
 # on the corrected image6 labels: 2.0 keeps 95% of shown labels correct at 59%
 # coverage, 1.0 keeps 87% at 66%. The stricter line is the default because a
 # wrong label on a shelf is worse than a missing one.
-LOCAL_ACCEPT = 2.0
+#
+# 4.0 since 2026-09-29: the line was tuned on image6 alone, where every crop is
+# a facing of a product the shelf actually holds. On 2297 crops from unrelated
+# Wikimedia shelves, 2.0 accepted 25 of them on geometry alone (scores 2.26-4.81
+# with no text overlap at all), a 1.1% far-unknown rate against G3's 1%. At 4.0
+# that falls to 8 (0.35%) while image6 keeps 42 of 48 correct labels at 95.5%
+# precision, which is the only region passing both G2 and G3 with coverage left.
+LOCAL_ACCEPT = 4.0
 # The text lane's own accept line: a bigram overlap this high, with this much
 # margin over the runner-up, names the row on its own. Measured: the text lane
 # picks the right SKU on 41-42 of the 55 facings where it reads anything.
@@ -51,12 +58,16 @@ TEXT_ACCEPT = 0.5
 TEXT_MARGIN = 0.15
 # How many of the read's bigrams the winning row must actually contain. The
 # ratio alone is not evidence: a two-pair read the row happens to contain scores
-# 1.0, and "サソース" named sku_l08 that way. Two pairs is the floor that keeps
-# every correct text answer on image6 (40 of 40) while dropping the wrong ones
-# from 6 to 2, taking shown-label precision from 87% to 95%. It is a floor on
-# evidence, not on the ratio, so a long read of shared brand copy still scores
-# low and is still refused by the margin.
-TEXT_MIN_SHARED = 2
+# 1.0, and "サソース" named sku_l08 that way. Two pairs was the old floor, and it
+# is too low once the lane answers on unrelated shelves: on 2297 crops from
+# Wikimedia milk, soy and snack shelves, 20 of 26 text-lane answers rested on
+# exactly two shared bigrams ("避くて おいしい" -> sku_l12 via おい/いし), an
+# 11.4% false accept rate against G3's 1%. Four is where the two sides separate
+# once the local line is also at 4.0: false accepts fall to 8 (0.35%) while
+# image6 keeps 42 of its 48 correct labels. It is a floor on evidence, not on
+# the ratio, so a long read of shared brand copy still scores low and is still
+# refused by the margin.
+TEXT_MIN_SHARED = 4
 # Reference pyramid heights. A ~1000 px packshot against a 110-250 px crop is a
 # 4-8x ratio, and ORB's 8x1.2 pyramid tops out at 3.6x, which is why ORB and
 # BRISK find zero inliers at live resolution. Scaling each reference to the
@@ -73,6 +84,15 @@ HUE_VALUE_MIN = 50
 # Quantity tokens: a number with its unit, in either script. Cut whole, because
 # the unit's own letters ("g", "ml", "%") are what survived a digits-only rule.
 _QUANTITY = re.compile(r"\d+(?:\.\d+)?\s*(?:KG|G|ML|L|%|個|本|枚|入|袋|缶|パック)")
+# A package's own size, in mass units. Only mass can contradict a package size:
+# a percentage ("カロリー 50% カット") is a claim about the contents, not the
+# package, and counting it made every low-calorie crop look like a 50 g jar.
+_MASS = re.compile(r"(\d+(?:\.\d+)?)\s*(KG|G)", re.IGNORECASE)
+# The range a retail package of this class falls in. Below it are fragments of a
+# longer number ("20" read out of "200g"); above it are two numbers run together
+# ("5210" out of "210g"). Both are OCR noise, not a different package, and
+# either would veto a correct answer if taken at face value.
+PACKAGE_MASS_RANGE = (100.0, 2000.0)
 # How many surviving bigrams a read needs before it may name a product.
 IDENTITY_BIGRAM_FLOOR = 3
 
@@ -516,6 +536,40 @@ class DecisionPolicy:
             return self.text_min_shared
         return len(bigrams(read) & signature_bigrams(signature))
 
+    def _contradicts_size(self, read: str, sku: str) -> bool:
+        """Does the read print a package size the row's name does not have?
+
+        The near-unknown case the text ranking cannot catch. A crop of a
+        Kewpie 450 g bag reads "マヨネーズ 450g", and every mayonnaise row in
+        the catalog explains "マヨネーズ" equally well, so the ranking ties at
+        1.0 and the strict-greater veto above never fires - the local lane's
+        answer stands. The size is the part that differs, and it is printed
+        plainly, so it is the part worth checking.
+
+        Only a mass in `PACKAGE_MASS_RANGE` counts. Outside it the read is
+        noise, not a package: "20" is a fragment of "200g" and "5210" is two
+        numbers run together, and vetoing on either would cost correct
+        answers. Measured on image6, this vetoes 3 false accepts at full
+        resolution and 1 at live, and no correct answer at either.
+        """
+        signature = self.signatures.get(sku)
+        if signature is None:
+            return False
+        read_masses = {
+            float(number) for number, _unit in _MASS.findall(normalize_text(read))
+        }
+        plausible = {
+            mass for mass in read_masses
+            if PACKAGE_MASS_RANGE[0] <= mass <= PACKAGE_MASS_RANGE[1]
+        }
+        if not plausible:
+            return False
+        name_masses = {
+            float(number) for number, _unit in _MASS.findall(normalize_text(signature))
+        }
+        return not (read_masses & name_masses)
+
+
     def decide(
         self,
         local_rank: Any,
@@ -550,6 +604,11 @@ class DecisionPolicy:
             and (text_top[1] - text_runner) >= self.text_margin
             and carries_identity(text_read or "")
             and self._shared_evidence(text_read or "", text_top[0]) >= self.text_min_shared
+            # A row the read contradicts on size cannot be named by that read
+            # either. Without this the veto above only moves the answer: the
+            # local lane's row is blocked and the text lane then names the same
+            # row, because "マヨネーズ" is what both of them matched.
+            and not self._contradicts_size(text_read or "", text_top[0])
             else None
         )
 
@@ -562,6 +621,8 @@ class DecisionPolicy:
             # artwork, so the geometry is real and only the words differ.
             if text_top and text_top[0] != candidate and text_top[1] > local_text_floor(text, candidate):
                 answer = None
+            elif self._contradicts_size(text_read or "", candidate):
+                answer = None
             else:
                 answer, source = candidate, "local"
         if answer is None and text_names:

recognization/code/live.py

@@ -1301,6 +1301,10 @@ class LiveSession:
                 decision = DecisionPolicy(
                     signatures=text_lane.signatures if text_lane else None
                 ).decide(list(scores.items()), text_rank, text_read=text_read)
+            elif scale == "van":
+                # Van's contract on the stream: the aggregate's top SKU, always
+                # answered. Same scorer as `matcher`, no gate.
+                decision = decide_identity(mean_scores, method=method, scale="van")
             else:
                 decision = decide_identity(mean_scores, method=method, scale=scale)
             track.candidates = [

recognization/code/rebuild_catalog.py

@@ -170,6 +170,34 @@ def name_ocr_similarity(name: str, ocr_text: str) -> float:
     return len(name_bigrams & bigrams(ocr_text)) / max(1, len(name_bigrams))
 
 
+def best_orientation_text(ocr_engine: Any, image: np.ndarray, name: str = "") -> str:
+    """OCR text from whichever of the four right-angle rotations matches best.
+
+    A packshot's identifying text is not always horizontal: the やまや
+    めんたいマヨネーズ packshot reads 明太手屋さんが only at 270 degrees, and
+    the upright pass returns three stray characters. Checking one orientation
+    refused a correct image, which is the worst failure this check has - it
+    rejects good data while claiming to validate it.
+
+    The winner is the rotation that explains the most of the catalog name, not
+    the one that reads the most characters: the upright pass on that packshot
+    returns stray kana that concatenate longer than the real read, so length
+    picks the wrong one. With no name to compare against, length is the only
+    signal left.
+    """
+    rotations = (
+        image,
+        cv2.rotate(image, cv2.ROTATE_90_CLOCKWISE),
+        cv2.rotate(image, cv2.ROTATE_180),
+        cv2.rotate(image, cv2.ROTATE_90_COUNTERCLOCKWISE),
+    )
+    reads = [" ".join(line.text for line in ocr_engine.read(upscale(rotated)).lines)
+             for rotated in rotations]
+    if name:
+        return max(reads, key=lambda text: name_ocr_similarity(name, text))
+    return max(reads, key=lambda text: len(text.strip()))
+
+
 def collect(source: str) -> Dict[str, List[str]]:
     by_sku: Dict[str, List[str]] = {}
     for name in sorted(os.listdir(source)):
@@ -291,7 +319,7 @@ def audit(
                 continue
             if ocr_engine is not None and names is not None:
                 try:
-                    text = " ".join(line.text for line in ocr_engine.read(upscale(image)).lines)
+                    text = best_orientation_text(ocr_engine, image, names.get(sku, ""))
                 except OcrUnavailable as error:
                     print(f"OCR skipped for {name}: {error}", file=sys.stderr)
                 else:
@@ -317,6 +345,25 @@ def audit(
     return kept, refused
 
 
+def unattributed(source: str) -> List[str]:
+    """SKUs whose views have no entry in the source's ATTRIBUTION.json.
+
+    Reported, not refused. A view with no recorded source is still a usable
+    reference - the six that predate the attribution file are independent of
+    the evaluation photo and their OCR matches their catalog names - but the
+    provenance is gone, and provenance is what lets a later audit tell a
+    re-sourced packshot from a leaked shelf crop. Naming them keeps the gap
+    from growing quietly.
+    """
+    path = os.path.join(source, "ATTRIBUTION.json")
+    try:
+        with open(path) as handle:
+            recorded = json.load(handle)
+    except (OSError, ValueError):
+        recorded = {}
+    return sorted(sku for sku in collect(source) if sku not in recorded)
+
+
 def text_discriminated_siblings(
     kept: Dict[str, List[str]], source: str, names: Optional[Dict[str, str]] = None
 ) -> List[Dict[str, Any]]:
@@ -535,6 +582,8 @@ def main(argv: Optional[Sequence[str]] = None) -> int:
     # and vetoes it from the CSV alone, so it is a coverage gap, not a refusal.
     text_only = sorted(set(names) - set(kept))
     siblings = text_discriminated_siblings(kept, args.source, names)
+    small = below_pyramid_floor(kept, args.source)
+    no_source = unattributed(args.source)
     verdict = live_ready_verdict(kept, names, shelf_skus)
     report = {
         "source": args.source,
@@ -548,6 +597,7 @@ def main(argv: Optional[Sequence[str]] = None) -> int:
         "text_only_skus": text_only,
         "text_discriminated_siblings": siblings,
         "below_pyramid_floor": small,
+        "unattributed_skus": no_source,
         "completeness": verdict,
         "skus": {sku: views for sku, views in sorted(kept.items())},
         "refused": refused,
@@ -571,6 +621,9 @@ def main(argv: Optional[Sequence[str]] = None) -> int:
         for row in small:
             print(f"  {row['sku']:<10} {row['view']:<20} {row['height']}px tall - "
                   "scores are not comparable with the rest; re-source this view")
+    if no_source:
+        print(f"\nno recorded source ({len(no_source)}): {', '.join(no_source)} - "
+              "usable, but provenance is missing; re-sourcing would close it")
     if verdict["live_ready"] is None:
         print(f"\ncompleteness: unknown - {verdict['reason']}")
     elif verdict["live_ready"]:

recognization/code/test_engine.py

@@ -136,14 +136,19 @@ class TextLaneSignatureTests(unittest.TestCase):
         self.assertGreaterEqual(ranked[0][1], TEXT_ACCEPT)
         self.assertGreaterEqual(ranked[0][1] - ranked[1][1], TEXT_MARGIN)
 
-    def test_one_shared_pair_is_not_enough_to_name_a_row(self):
-        """A ratio of 1.0 on a single pair is not evidence.
+    def test_a_thin_read_is_not_enough_to_name_a_row(self):
+        """A ratio of 1.0 on a few pairs is not evidence.
 
         "XXイス" contains exactly one pair the catalog knows, and the row that
         contains it scores 1.0 with a full margin over every other row. On
-        image6 that shape named sku_l08 from the read "サソース". The floor
-        asks how much of the read the row actually explains, so one pair is
-        refused while the same read with two is accepted.
+        image6 that shape named sku_l08 from the read "サソース". The floor asks
+        how much of the read the row actually explains, so one pair and two
+        pairs are both refused while a real read of the name is accepted.
+
+        Two pairs used to be enough. It is not: on 2297 crops from unrelated
+        Wikimedia shelves, 20 of 26 text-lane answers rested on exactly two
+        shared bigrams - "避くて おいしい" named sku_l12 through おい/いし - for
+        an 11.4% false accept rate. The floor is 4.
         """
         from identity_lanes import DecisionPolicy, TextLane
 
@@ -153,16 +158,22 @@ class TextLaneSignatureTests(unittest.TestCase):
         ]
         lane = TextLane.from_catalog(rows, "no-such-dir", ocr_engine=None)
 
-        thin = sorted(lane.scores_from_text("XXイス").items(), key=lambda kv: kv[1], reverse=True)
-        self.assertEqual(thin[0][1], 1.0, "the ratio alone cannot see the difference")
-        self.assertIsNone(
-            DecisionPolicy(signatures=lane.signatures).decide([], thin, text_read="XXイス")["best_sku"]
-        )
+        for thin_read in ("XXイス", "ZZソース"):
+            thin = sorted(lane.scores_from_text(thin_read).items(),
+                          key=lambda kv: kv[1], reverse=True)
+            self.assertEqual(thin[0][1], 1.0, "the ratio alone cannot see the difference")
+            self.assertIsNone(
+                DecisionPolicy(signatures=lane.signatures).decide(
+                    [], thin, text_read=thin_read
+                )["best_sku"],
+                f"{thin_read!r} shares too few pairs to name a row",
+            )
 
-        enough = sorted(lane.scores_from_text("ZZソース").items(), key=lambda kv: kv[1], reverse=True)
+        enough = sorted(lane.scores_from_text("オイスター").items(),
+                        key=lambda kv: kv[1], reverse=True)
         self.assertEqual(
             DecisionPolicy(signatures=lane.signatures).decide(
-                [], enough, text_read="ZZソース"
+                [], enough, text_read="オイスター"
             )["best_sku"],
             "sku_l08",
         )
@@ -616,8 +627,11 @@ class DecisionPolicyTests(unittest.TestCase):
         the margin over the runner-up. Re-sorting by value names a different SKU
         than the lane chose - here sku_b has the larger value but sku_a is the
         lane's own first choice, and sku_a is what must be answered.
+
+        Both values clear LOCAL_ACCEPT, so the test is about the ordering and
+        not about the gate.
         """
-        decision = self.policy().decide([("sku_a", 3.0), ("sku_b", 5.0)], [])
+        decision = self.policy().decide([("sku_a", 4.5), ("sku_b", 5.0)], [])
         self.assertEqual(decision["best_sku"], "sku_a")
 
     def test_a_size_only_read_does_not_fill_in(self):
@@ -632,6 +646,56 @@ class DecisionPolicyTests(unittest.TestCase):
         )
         self.assertIsNone(decision["best_sku"])
 
+    def test_a_read_printing_another_size_vetoes_the_local_answer(self):
+        """The documented near-unknown: a Kewpie 450 g bag the catalog lacks.
+
+        The crop reads "マヨネーズ 450g", which every mayonnaise row explains
+        equally, so the text ranking ties at 1.0 and the strict-greater veto
+        cannot fire. The size is what differs, and the catalog's Kewpie rows
+        are 200 g, so the read contradicts the row the local lane chose.
+        """
+        from identity_lanes import DecisionPolicy
+
+        policy = DecisionPolicy(signatures={"sku_a": "キューピー ハーフマヨネーズ 400g"})
+        decision = policy.decide(
+            [("sku_a", 5.0)], [("sku_a", 1.0)], text_read="ピーマヨネーズ 450g"
+        )
+        self.assertIsNone(decision["best_sku"])
+
+    def test_a_read_agreeing_on_size_does_not_veto(self):
+        """The same read with the row's own size must still answer."""
+        from identity_lanes import DecisionPolicy
+
+        policy = DecisionPolicy(signatures={"sku_a": "キューピー ハーフマヨネーズ 400g"})
+        decision = policy.decide(
+            [("sku_a", 5.0)], [("sku_a", 1.0)], text_read="ピーマヨネーズ 400g"
+        )
+        self.assertEqual(decision["best_sku"], "sku_a")
+
+    def test_ocr_noise_around_a_size_does_not_veto(self):
+        """A fragment or a run-together number is noise, not a package size.
+
+        Live reads of a correct 200 g row came back as "20" and "5210"; both
+        are outside the range a package of this class falls in, and vetoing on
+        either would cost a correct answer.
+        """
+        from identity_lanes import DecisionPolicy
+
+        policy = DecisionPolicy(signatures={"sku_a": "キューピー ハーフマヨネーズ 200g"})
+        for read in ("アマニ油 20gs", "内臓脂訪を 滅与ず ッデュービーフィ5 210g", "n0gベ% 50"):
+            decision = policy.decide([("sku_a", 5.0)], [("sku_a", 1.0)], text_read=read)
+            self.assertEqual(decision["best_sku"], "sku_a", read)
+
+    def test_a_percentage_is_not_a_package_size(self):
+        """A low-calorie claim must not read as a 50 g package."""
+        from identity_lanes import DecisionPolicy
+
+        policy = DecisionPolicy(signatures={"sku_a": "キューピー ハーフマヨネーズ 400g"})
+        decision = policy.decide(
+            [("sku_a", 5.0)], [("sku_a", 1.0)], text_read="カロリー 50% カット 400g"
+        )
+        self.assertEqual(decision["best_sku"], "sku_a")
+
     def test_a_named_read_with_a_size_still_fills_in(self):
         decision = self.policy().decide([], [("sku_b", 0.9)], text_read="コクうま 65%")
         self.assertEqual(decision["best_sku"], "sku_b")

recognization/code/test_identity.py

@@ -50,3 +50,20 @@ class IdentityRankingTests(unittest.TestCase):
         self.assertEqual(ranked["identity_level"], IDENTITY_ACCEPTED)
         self.assertEqual(ranked["best_sku"], "sku_a")
         self.assertEqual(ranked["status"], "success")
+
+    def test_van_answers_what_the_gate_would_refuse(self):
+        # Van's contract is the baseline the decision is measured against, so it
+        # must answer exactly where the shipped gate abstains. A weak score and
+        # a close pair are both refused above; here both are named.
+        weak = decide_identity({"sku_a": 0.4, "sku_b": 0.0}, scale="van")
+        self.assertEqual(weak["identity_level"], IDENTITY_ACCEPTED)
+        self.assertEqual(weak["best_sku"], "sku_a")
+        close = decide_identity({"sku_a": 20.0, "sku_b": 19.5}, scale="van")
+        self.assertEqual(close["best_sku"], "sku_a")
+
+    def test_van_still_rejects_when_there_is_no_evidence(self):
+        # Always-argmax is not always-answer: with nothing to rank there is no
+        # argmax, and naming a SKU from an empty score set would be a fabrication.
+        ranked = decide_identity({}, scale="van")
+        self.assertEqual(ranked["identity_level"], IDENTITY_REJECTED)
+        self.assertIsNone(ranked["best_sku"])

recognization/code/test_rebuild_catalog.py

@@ -184,5 +184,95 @@ class BadShelfListFailsBeforeTheSwapTests(unittest.TestCase):
         self.assertFalse(os.path.exists(output), "the catalog was replaced before the check")
 
 
-if __name__ == "__main__":
-    unittest.main()
+class OrientationTests(unittest.TestCase):
+    """A packshot whose identifying text is not horizontal.
+
+    The やまや めんたいマヨネーズ packshot reads its brand only when rotated;
+    the upright pass returns stray kana. Checking one orientation refused a
+    correct image, which is the worst failure this check has - it rejects good
+    data while claiming to validate it.
+    """
+    class FakeOcr:
+        """Reads a fixed string per rotation.
+
+        Rotations are identified by aspect ratio, not by exact shape: the
+        caller upscales before OCR, so the fake sees scaled dimensions. The
+        upright and 180-degree passes keep the source's orientation (wide),
+        the two 90-degree passes swap it (tall). A corner marker cannot tell
+        them apart, because `cv2.rotate` carries the corner value through.
+        """
+
+        name = "fake"
+
+        def __init__(self, by_orientation):
+            self.by_orientation = by_orientation
+
+        def read(self, image):
+            class Line:
+                def __init__(self, text):
+                    self.text = text
+
+            class Read:
+                def __init__(self, text):
+                    self.lines = [Line(part) for part in text.split("|") if part]
+
+            height, width = image.shape[:2]
+            key = "wide" if width > height else "tall"
+            return Read(self.by_orientation.get(key, ""))
+
+    def marked(self, height, width):
+        return np.zeros((height, width, 3), dtype="uint8")
+
+    def test_a_rotated_read_wins_when_it_matches_the_name(self):
+        """The upright pass reads noise; the rotated pass reads the product."""
+        from rebuild_catalog import best_orientation_text
+
+        engine = self.FakeOcr({"wide": "つくつ|た", "tall": "明太手屋さんが|マヨネーズ"})
+        text = best_orientation_text(engine, self.marked(40, 60), "やまや めんたいマヨネーズタイプ 200g")
+        self.assertIn("明太手屋さんが", text)
+
+    def test_the_upright_read_wins_when_it_matches(self):
+        from rebuild_catalog import best_orientation_text
+
+        engine = self.FakeOcr({"wide": "ピュアセレクト|マヨネーズ|450g", "tall": "ノイズ"})
+        text = best_orientation_text(engine, self.marked(40, 60), "ピュアセレクト マヨネーズ 450g")
+        self.assertIn("ピュアセレクト", text)
+
+    def test_without_a_name_the_longest_read_wins(self):
+        from rebuild_catalog import best_orientation_text
+
+        engine = self.FakeOcr({"wide": "ab", "tall": "abcdefgh"})
+        self.assertEqual(best_orientation_text(engine, self.marked(40, 60)), "abcdefgh")
+
+    def test_audit_keeps_a_view_whose_name_only_reads_rotated(self):
+        """The regression: `audit` must try every orientation before refusing.
+
+        This is the boundary that matters. `best_orientation_text` passing its
+        own unit test proves nothing about enrollment - the bug was that
+        `audit` called a single-orientation read, so a correct packshot whose
+        brand text runs vertically was refused as a name/OCR mismatch.
+        """
+        from rebuild_catalog import audit
+
+        root = tempfile.mkdtemp(prefix="orient-")
+        # Wider than tall, so the upright pass is the "wide" read (noise) and
+        # only a 90-degree rotation produces the "tall" read that matches.
+        cv2.imwrite(os.path.join(root, "sku_x1.jpg"), self.marked(300, 400))
+        engine = self.FakeOcr({"wide": "つくつ|た", "tall": "キユーピー|エッグケア|305g"})
+        kept, refused = audit(root, min_views=1, names={"sku_x1": "キューピー エッグケア 305g"},
+                              ocr_engine=engine)
+        self.assertEqual(kept, {"sku_x1": ["sku_x1.jpg"]}, refused)
+        self.assertEqual(refused, [])
+
+    def test_audit_still_refuses_a_view_the_name_never_matches(self):
+        """The check must keep its teeth: a wrong product is still refused."""
+        from rebuild_catalog import audit
+
+        root = tempfile.mkdtemp(prefix="orient-bad-")
+        cv2.imwrite(os.path.join(root, "sku_x1.jpg"), self.marked(400, 300))
+        engine = self.FakeOcr({"wide": "キューピー|マヨネーズ", "tall": "キューピー|マヨネーズ"})
+        kept, refused = audit(root, min_views=1, names={"sku_x1": "やまや めんたいマヨネーズタイプ 200g"},
+                              ocr_engine=engine)
+        self.assertEqual(kept, {})
+        self.assertEqual(len(refused), 1, refused)
+        self.assertIn("name/OCR", refused[0]["reason"])

recognization/data/README.md

@@ -0,0 +1,34 @@
+# Recognition data
+
+Everything the identity work needs to be reproducible: sourced packshots,
+evaluation runs, and the audits that produced them. Nothing here is ephemeral;
+a run that matters is written here, not to a temp directory.
+
+## `packshots/source/`
+
+Front packshots as downloaded, before enrollment. `ATTRIBUTION.json` records the
+source URL, the product as printed, and the manufacturer for each. `ocr-reads.json`
+holds the OCR text each image produced, so a name/OCR refusal can be re-checked
+without re-running Vision.
+
+`rebuild_catalog.py` reads this directory and writes a catalog:
+
+    .venv/bin/python code/rebuild_catalog.py --source data/packshots/source \
+        --output stock-lopia-clean --csv stock-lopia/stock.csv --min-views 1 --replace
+
+## `evaluations/`
+
+One directory per session. Each holds the raw artifacts a claim was computed
+from, so a number in DECISION.md can be traced back to the run that produced it.
+
+| Directory | Contents |
+|---|---|
+| `stream/` | WebSocket track dumps from `stream_client.py`, scored by `score_stream.py`. |
+| `still/` | `/v1/recognize` responses for the still-image path. |
+| `catalog-audit/` | `rebuild_catalog.py --json` reports. |
+| `scratch/` | Intermediate images (rotations, crops) kept because they document a finding. |
+
+Scoring a stream dump:
+
+    .venv/bin/python code/score_stream.py data/evaluations/<session>/stream/<file>.json \
+        --labels eval/labels/image6.json --stock-dir stock-lopia-clean

recognization/data/evaluations/2026-09-29-session/catalog-audit/enroll-2026-09-29.json

@@ -0,0 +1,76 @@
+{
+ "source": "data/packshots/source",
+ "output": "stock-lopia-clean",
+ "min_width": 120,
+ "min_height": 160,
+ "min_views": 1,
+ "duplicate_cosine": 0.98,
+ "min_name_ocr": 0.7,
+ "ocr": "vision",
+ "text_only_skus": [
+  "sku_l04",
+  "sku_l07",
+  "sku_l08",
+  "sku_l12"
+ ],
+ "text_discriminated_siblings": [
+  {
+   "sku": "sku_l09",
+   "view": "sku_l09.jpg",
+   "sibling_sku": "sku_l10",
+   "sibling_view": "sku_l10.jpg",
+   "cosine": 0.9817,
+   "name_similarity": 0.333
+  }
+ ],
+ "below_pyramid_floor": [],
+ "completeness": {
+  "live_ready": null,
+  "reason": "no --shelf-tags given; the completeness rule needs the shelf's SKU list",
+  "enrolled": 8,
+  "text_only": [
+   "sku_l04",
+   "sku_l07",
+   "sku_l08",
+   "sku_l12"
+  ]
+ },
+ "skus": {
+  "sku_l01": [
+   "sku_l01.jpg"
+  ],
+  "sku_l02": [
+   "sku_l02.jpg"
+  ],
+  "sku_l03": [
+   "sku_l03.jpg"
+  ],
+  "sku_l05": [
+   "sku_l05.jpg"
+  ],
+  "sku_l06": [
+   "sku_l06.jpg"
+  ],
+  "sku_l09": [
+   "sku_l09.jpg"
+  ],
+  "sku_l10": [
+   "sku_l10.jpg"
+  ],
+  "sku_l11": [
+   "sku_l11.jpg"
+  ]
+ },
+ "refused": [
+  {
+   "sku": "sku_l04",
+   "view": "sku_l04.jpg",
+   "reason": "name/OCR 0.00 below 0.70: 明太手屋さんが つくつ た コ"
+  },
+  {
+   "sku": "sku_l07",
+   "view": "sku_l07.jpg",
+   "reason": "name/OCR 0.57 below 0.70: ラッキー マヨネーズ 調味料（アミノ酸）不使用"
+  }
+ ]
+}

recognization/data/evaluations/2026-09-29-session/catalog-audit/eval-after-l10.json

@@ -0,0 +1,221 @@
+{
+ "image": "add/image6.jpg",
+ "stock_dir": "stock-lopia-clean",
+ "labels": "eval/labels/image6.json",
+ "provenance": {
+  "catalog_hash": "3f1653d9d962073af96121250006b9b6d86113556797bbee65de68b60e87a64d",
+  "labels_hash": "5aabf029921ced767d4e14a18e8bccb55812dd974852c56e0931a98fa10366cf",
+  "image_hash": "a08b89ba37d27c9d5a81bcd56bcd042e6baf54d2f182bbaaf47e97f14c9cd525",
+  "code_revision": "68061635d9fef3097609945bfd9ac1935791b398"
+ },
+ "facings": 111,
+ "leakage": {
+  "excluded": 0,
+  "labels": []
+ },
+ "hub_share_limit": 0.3,
+ "enrollment": {
+  "catalog_skus": [
+   "sku_l01",
+   "sku_l02",
+   "sku_l03",
+   "sku_l04",
+   "sku_l05",
+   "sku_l06",
+   "sku_l07",
+   "sku_l08",
+   "sku_l09",
+   "sku_l10",
+   "sku_l11",
+   "sku_l12"
+  ],
+  "enrolled_skus": [
+   "sku_l01",
+   "sku_l02",
+   "sku_l03",
+   "sku_l05",
+   "sku_l06",
+   "sku_l09",
+   "sku_l10",
+   "sku_l11"
+  ],
+  "unenrolled_skus": {
+   "sku_l07": 6,
+   "sku_l04": 1,
+   "sku_l08": 1
+  },
+  "facings_without_enrollment": 8,
+  "scored_facings": 109
+ },
+ "results": {
+  "full": [
+   {
+    "lane": "local",
+    "accept": 2.0,
+    "margin": 0.15,
+    "in_catalog": 61,
+    "out_catalog": 48,
+    "enrolled": 53,
+    "correct": 48,
+    "wrong": 9,
+    "rejected": 52,
+    "catalog_ceiling": 0.8689,
+    "identity_ceiling": 0.9245,
+    "decision_ceiling": 0.8113,
+    "end_to_end_coverage": 0.7869,
+    "false_accepts": 9,
+    "unknown_false_accepts": 8,
+    "unenrolled_false_accepts": 0,
+    "false_accept_rate": 0.1667,
+    "near_unknown_false_accept_rate": 0.0,
+    "precision": 0.8421,
+    "max_sku_share": 0.2281,
+    "hub_sku": "sku_l09",
+    "hub_alarm": false,
+    "histogram": {
+     "sku_l09": 13,
+     "sku_l11": 10,
+     "sku_l10": 9,
+     "sku_l01": 5,
+     "sku_l03": 5,
+     "sku_l07": 5,
+     "sku_l06": 4,
+     "sku_l02": 3,
+     "sku_l05": 3
+    },
+    "per_sku": {
+     "sku_l01": {
+      "correct": 5,
+      "total": 5
+     },
+     "sku_l02": {
+      "correct": 2,
+      "total": 4
+     },
+     "sku_l03": {
+      "correct": 4,
+      "total": 4
+     },
+     "sku_l04": {
+      "correct": 0,
+      "total": 1
+     },
+     "sku_l05": {
+      "correct": 3,
+      "total": 3
+     },
+     "sku_l06": {
+      "correct": 4,
+      "total": 4
+     },
+     "sku_l07": {
+      "correct": 5,
+      "total": 6
+     },
+     "sku_l08": {
+      "correct": 0,
+      "total": 1
+     },
+     "sku_l09": {
+      "correct": 11,
+      "total": 13
+     },
+     "sku_l10": {
+      "correct": 9,
+      "total": 11
+     },
+     "sku_l11": {
+      "correct": 5,
+      "total": 9
+     }
+    },
+    "ms_per_crop": 26.7
+   }
+  ],
+  "live": [
+   {
+    "lane": "local",
+    "accept": 2.0,
+    "margin": 0.15,
+    "in_catalog": 61,
+    "out_catalog": 48,
+    "enrolled": 53,
+    "correct": 41,
+    "wrong": 4,
+    "rejected": 64,
+    "catalog_ceiling": 0.8689,
+    "identity_ceiling": 0.8491,
+    "decision_ceiling": 0.6604,
+    "end_to_end_coverage": 0.6721,
+    "false_accepts": 4,
+    "unknown_false_accepts": 3,
+    "unenrolled_false_accepts": 0,
+    "false_accept_rate": 0.0625,
+    "near_unknown_false_accept_rate": 0.0,
+    "precision": 0.9111,
+    "max_sku_share": 0.2,
+    "hub_sku": "sku_l09",
+    "hub_alarm": false,
+    "histogram": {
+     "sku_l09": 9,
+     "sku_l10": 9,
+     "sku_l11": 7,
+     "sku_l07": 6,
+     "sku_l01": 4,
+     "sku_l03": 3,
+     "sku_l05": 3,
+     "sku_l06": 2,
+     "sku_l02": 1,
+     "sku_l04": 1
+    },
+    "per_sku": {
+     "sku_l01": {
+      "correct": 4,
+      "total": 5
+     },
+     "sku_l02": {
+      "correct": 1,
+      "total": 4
+     },
+     "sku_l03": {
+      "correct": 3,
+      "total": 4
+     },
+     "sku_l04": {
+      "correct": 1,
+      "total": 1
+     },
+     "sku_l05": {
+      "correct": 3,
+      "total": 3
+     },
+     "sku_l06": {
+      "correct": 2,
+      "total": 4
+     },
+     "sku_l07": {
+      "correct": 5,
+      "total": 6
+     },
+     "sku_l08": {
+      "correct": 0,
+      "total": 1
+     },
+     "sku_l09": {
+      "correct": 9,
+      "total": 13
+     },
+     "sku_l10": {
+      "correct": 9,
+      "total": 11
+     },
+     "sku_l11": {
+      "correct": 4,
+      "total": 9
+     }
+    },
+    "ms_per_crop": 20.9
+   }
+  ]
+ }
+}

recognization/data/evaluations/2026-09-29-session/catalog-audit/source-audit.json

@@ -0,0 +1,70 @@
+{
+ "source": "data/packshots/source",
+ "output": "/tmp/lopia-audit",
+ "min_width": 120,
+ "min_height": 160,
+ "min_views": 1,
+ "duplicate_cosine": 0.98,
+ "min_name_ocr": 0.7,
+ "ocr": "vision",
+ "text_only_skus": [
+  "sku_l04",
+  "sku_l07",
+  "sku_l08",
+  "sku_l09",
+  "sku_l10",
+  "sku_l12"
+ ],
+ "text_discriminated_siblings": [],
+ "below_pyramid_floor": [],
+ "completeness": {
+  "live_ready": null,
+  "reason": "no --shelf-tags given; the completeness rule needs the shelf's SKU list",
+  "enrolled": 6,
+  "text_only": [
+   "sku_l04",
+   "sku_l07",
+   "sku_l08",
+   "sku_l09",
+   "sku_l10",
+   "sku_l12"
+  ]
+ },
+ "skus": {
+  "sku_l01": [
+   "sku_l01.jpg"
+  ],
+  "sku_l02": [
+   "sku_l02.jpg"
+  ],
+  "sku_l03": [
+   "sku_l03.jpg"
+  ],
+  "sku_l05": [
+   "sku_l05.jpg"
+  ],
+  "sku_l06": [
+   "sku_l06.jpg"
+  ],
+  "sku_l11": [
+   "sku_l11.jpg"
+  ]
+ },
+ "refused": [
+  {
+   "sku": "sku_l04",
+   "view": "sku_l04.jpg",
+   "reason": "name/OCR 0.00 below 0.70: 明太手屋さんが つくつ た コ"
+  },
+  {
+   "sku": "sku_l07",
+   "view": "sku_l07.jpg",
+   "reason": "name/OCR 0.57 below 0.70: ラッキー マヨネーズ 調味料（アミノ酸）不使用"
+  },
+  {
+   "sku": "sku_l09",
+   "view": "sku_l09.jpg",
+   "reason": "name/OCR 0.05 below 0.70: キューピーマヨネーズ KEWPIE MAYONNAISE ® 450g"
+  }
+ ]
+}

recognization/data/evaluations/2026-09-29-session/scratch/fa_detail.py

@@ -0,0 +1,27 @@
+"""Which facings are falsely accepted, and as what."""
+import sys, cv2
+sys.path.insert(0, "code")
+from eval_identity import LocalLaneAdapter, load_labels, at_live_resolution, leaked, UNKNOWN, crop_of
+from identity_lanes import DecisionPolicy
+
+payload = load_labels("eval/labels/image6.json")
+image = cv2.imread("add/image6.jpg")
+facings = [f for f in payload["facings"] if f["label"] != "exclude"]
+facings = [f for f in facings if leaked(image, f["box"], "stock-lopia-clean") is None]
+enrolled = {"sku_l01","sku_l02","sku_l03","sku_l05","sku_l06","sku_l09","sku_l10","sku_l11"}
+
+for live in (False, True):
+    picture = at_live_resolution(image) if live else image
+    scaled = facings if not live else [
+        {**f, "box": [int(round(v * picture.shape[1] / image.shape[1])) for v in f["box"]]}
+        for f in facings]
+    lane = LocalLaneAdapter("stock-lopia-clean")
+    print(f"=== {'live' if live else 'full'} ===")
+    for f in scaled:
+        crop = crop_of(picture, f["box"])
+        if crop.size == 0: continue
+        ranked = lane.rank(crop)
+        d = lane.decide(crop, ranked, accept=2.0, margin=0.0, scale="local")
+        ans = d.get("best_sku") if d.get("identity_level") == "accepted" else None
+        if ans and ans != f["label"]:
+            print(f"  #{f['index']:>3} truth={f['label']:<9} answered={ans:<9} enrolled={f['label'] in enrolled}")

recognization/data/evaluations/2026-09-29-session/scratch/fa_text.py

@@ -0,0 +1,54 @@
+"""What the falsely-accepted crops actually read, and what the text lane scores.
+
+The advisory says crops 82-97 are Kewpie Mayonnaise 450g - a product the
+catalog does not hold - and that no catalog row can therefore win the text
+ranking, so the veto at DecisionPolicy.decide never fires. This prints the
+reads and the scores that claim depends on.
+"""
+import csv
+import sys
+
+import cv2
+import numpy as np
+
+sys.path.insert(0, "code")
+from identity_lanes import TextLane, carries_identity  # noqa: E402
+from ocr import build_ocr_engine  # noqa: E402
+
+INDICES = [73, 82, 83, 84, 86, 94, 95, 96, 97]
+
+rows = []
+with open("stock-lopia-clean/stock.csv", encoding="utf-8") as handle:
+    for record in csv.reader(handle):
+        if len(record) >= 2 and record[0].startswith("sku_"):
+            rows.append({"sku_id": record[0], "product_name": record[1]})
+
+engine = build_ocr_engine("vision")
+lane = TextLane.from_catalog(rows, "stock-lopia-clean", ocr_engine=engine)
+
+labels = {f["index"]: f for f in __import__("json").load(open("eval/labels/image6.json"))["facings"]}
+image = cv2.imread("add/image6.jpg")
+
+tiles = []
+for index in INDICES:
+    facing = labels[index]
+    x1, y1, x2, y2 = facing["box"]
+    crop = image[y1:y2, x1:x2]
+    read = lane.read(crop)
+    scores = lane.scores_from_text(read)
+    top = sorted(scores.items(), key=lambda kv: -kv[1])[:3]
+    print(f"#{index:>3} truth={facing['label']:<9} carries={carries_identity(read)}")
+    print(f"      read: {read[:110]}")
+    print(f"      text top: {[(s, round(v, 3)) for s, v in top]}")
+
+    height, width = crop.shape[:2]
+    scale = 260.0 / max(1, height)
+    resized = cv2.resize(crop, (max(1, int(width * scale)), 260), interpolation=cv2.INTER_CUBIC)
+    canvas = np.full((300, 220, 3), 255, np.uint8)
+    canvas[30:290, : min(resized.shape[1], 220)] = resized[:, :220]
+    cv2.putText(canvas, str(index), (6, 22), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 0, 255), 2)
+    tiles.append(canvas)
+
+grid = np.vstack([np.hstack(tiles[:5]), np.hstack(tiles[5:])])
+cv2.imwrite("data/evaluations/2026-09-29-session/scratch/fa_tiles.jpg", grid)
+print("tile:", grid.shape)

recognization/data/evaluations/2026-09-29-session/scratch/size_gate.py

@@ -0,0 +1,71 @@
+"""Mass-only size contradiction, with full reads.
+
+The first pass counted every quantity token, including percentages, so a crop
+reading "カロリー 50% カット" looked like a 50 g package. Only mass units can
+contradict a package size, so this restricts to g/kg and prints the whole read
+for every accepted facing that carries one.
+"""
+import csv
+import json
+import re
+import sys
+
+import cv2
+
+sys.path.insert(0, "code")
+from eval_identity import LocalLaneAdapter, at_live_resolution, crop_of, leaked  # noqa: E402
+from identity_lanes import TextLane, normalize_text  # noqa: E402
+from ocr import build_ocr_engine  # noqa: E402
+
+MASS = re.compile(r"(\d+(?:\.\d+)?)\s*(KG|G)", re.IGNORECASE)
+# Package sizes this product class actually comes in. Below it are OCR fragments
+# of a longer number ("20" from 200); above it are concatenations ("5210").
+PLAUSIBLE = (100.0, 2000.0)
+
+
+def masses(text: str):
+    return {float(number) for number, _unit in MASS.findall(normalize_text(text))}
+
+
+rows = []
+with open("stock-lopia-clean/stock.csv", encoding="utf-8") as handle:
+    for record in csv.reader(handle):
+        if len(record) >= 2 and record[0].startswith("sku_"):
+            rows.append({"sku_id": record[0], "product_name": record[1]})
+names = {row["sku_id"]: row["product_name"] for row in rows}
+
+engine = build_ocr_engine("vision")
+text_lane = TextLane.from_catalog(rows, "stock-lopia-clean", ocr_engine=engine)
+local_lane = LocalLaneAdapter("stock-lopia-clean")
+
+payload = json.load(open("eval/labels/image6.json"))
+image = cv2.imread("add/image6.jpg")
+facings = [f for f in payload["facings"] if f["label"] != "exclude"]
+facings = [f for f in facings if leaked(image, f["box"], "stock-lopia-clean") is None]
+
+for live in (False, True):
+    picture = at_live_resolution(image) if live else image
+    scaled = facings if not live else [
+        {**f, "box": [int(round(v * picture.shape[1] / image.shape[1])) for v in f["box"]]}
+        for f in facings]
+    print(f"=== {'live' if live else 'full'} ===")
+    for facing in scaled:
+        crop = crop_of(picture, facing["box"])
+        if crop.size == 0:
+            continue
+        ranked = local_lane.rank(crop)
+        decision = local_lane.decide(crop, ranked, accept=2.0, margin=0.0, scale="local")
+        if decision.get("identity_level") != "accepted":
+            continue
+        answer = decision.get("best_sku")
+        read = text_lane.read(crop)
+        read_masses = masses(read)
+        name_masses = masses(names.get(answer, ""))
+        if not read_masses or not name_masses:
+            continue
+        plausible = {m for m in read_masses if PLAUSIBLE[0] <= m <= PLAUSIBLE[1]}
+        veto = bool(plausible) and not (read_masses & name_masses)
+        correct = "correct" if answer == facing["label"] else "WRONG"
+        print(f"  #{facing['index']:>3} {'VETO' if veto else 'keep'} {correct:<7} "
+              f"answer={answer:<9} read_m={sorted(read_masses)} name_m={sorted(name_masses)}")
+        print(f"        read: {read[:120]}")

recognization/data/evaluations/2026-09-29-session/server-smoke/final-summary.json

@@ -0,0 +1,28 @@
+{
+ "endpoint": "/v1/recognize",
+ "image": "add/image6.jpg",
+ "catalog": "stock-lopia-clean",
+ "gates": "LOCAL_ACCEPT=4.0 TEXT_MIN_SHARED=4",
+ "still": {
+  "detections": 111,
+  "named": 44,
+  "matched": 44,
+  "correct": 42,
+  "wrong": 2,
+  "precision": 0.955
+ },
+ "stream": {
+  "seconds": 20,
+  "pan": 12,
+  "tracks_seen": 165,
+  "tracks_labelled": 62,
+  "label_changes": 0,
+  "unexplained_label_changes": 0,
+  "facings_named": "17/27",
+  "precision": 0.84,
+  "wrong": 0,
+  "unenrolled_false_accepts": 0,
+  "unknown_false_accepts": 8
+ },
+ "note": "Final state with every catalog correction and both calibrated accept lines live."
+}

recognization/data/evaluations/2026-09-29-session/server-smoke/gate-summary.json

@@ -0,0 +1,24 @@
+{
+ "endpoint": "/v1/recognize",
+ "image": "add/image6.jpg",
+ "catalog": "stock-lopia-clean",
+ "gates": "LOCAL_ACCEPT=4.0 TEXT_MIN_SHARED=4",
+ "detections": 111,
+ "named": 44,
+ "matched": 44,
+ "correct": 42,
+ "wrong": 2,
+ "precision": 0.955,
+ "wrong_detail": [
+  "sku_l07 vs unknown",
+  "sku_l03 vs unknown"
+ ],
+ "previous": {
+  "gates": "2.0/2",
+  "named": 53,
+  "correct": 48,
+  "wrong": 5,
+  "precision": 0.906
+ },
+ "note": "Live still endpoint, box-IoU matched to the corrected labels. The stricter gates trade 9 names for 3 wrong labels."
+}

recognization/data/evaluations/2026-09-29-session/server-smoke/summary.json

@@ -0,0 +1,23 @@
+{
+ "endpoint": "/v1/recognize",
+ "image": "add/image6.jpg",
+ "catalog": "stock-lopia-clean",
+ "catalog_version": "b862c55f4623",
+ "detections": 111,
+ "named": 53,
+ "named_without_product_info": 0,
+ "view_suffixed_ids": [],
+ "elapsed_s": 2.38,
+ "by_sku": {
+  "sku_l10": 9,
+  "sku_l11": 5,
+  "sku_l01": 5,
+  "sku_l06": 4,
+  "sku_l09": 13,
+  "sku_l07": 6,
+  "sku_l02": 3,
+  "sku_l03": 5,
+  "sku_l05": 3
+ },
+ "note": "Final state: l06 current package enrolled (4/4 facings named, was 1/4 at live), l10/l11 400g, l12 renamed. l12 not named: no enrolled view."
+}

recognization/data/evaluations/2026-09-29-session/still/final_full.json

@@ -0,0 +1,214 @@
+{
+ "image": "add/image6.jpg",
+ "stock_dir": "stock-lopia-clean",
+ "labels": "eval/labels/image6.json",
+ "provenance": {
+  "catalog_hash": "10f55211435aa64c0ed4ca24a7b902ce4ed8a30299409680942b78ec9b3d08c8",
+  "labels_hash": "5aabf029921ced767d4e14a18e8bccb55812dd974852c56e0931a98fa10366cf",
+  "image_hash": "a08b89ba37d27c9d5a81bcd56bcd042e6baf54d2f182bbaaf47e97f14c9cd525",
+  "code_revision": "b6c87c45ea2aa3ed59f5764274341f3ffed30fe0"
+ },
+ "facings": 111,
+ "leakage": {
+  "excluded": 0,
+  "labels": []
+ },
+ "hub_share_limit": 0.3,
+ "enrollment": {
+  "catalog_skus": [
+   "sku_l01",
+   "sku_l02",
+   "sku_l03",
+   "sku_l04",
+   "sku_l05",
+   "sku_l06",
+   "sku_l07",
+   "sku_l08",
+   "sku_l09",
+   "sku_l10",
+   "sku_l11",
+   "sku_l12"
+  ],
+  "enrolled_skus": [
+   "sku_l01",
+   "sku_l02",
+   "sku_l03",
+   "sku_l05",
+   "sku_l06",
+   "sku_l09",
+   "sku_l11"
+  ],
+  "unenrolled_skus": {
+   "sku_l10": 11,
+   "sku_l07": 6,
+   "sku_l04": 1,
+   "sku_l08": 1
+  },
+  "facings_without_enrollment": 19,
+  "scored_facings": 109
+ },
+ "results": {
+  "full": [
+   {
+    "lane": "van",
+    "accept": 1.0,
+    "margin": 0.15,
+    "in_catalog": 61,
+    "out_catalog": 48,
+    "enrolled": 42,
+    "correct": 32,
+    "wrong": 77,
+    "rejected": 0,
+    "catalog_ceiling": 0.6885,
+    "identity_ceiling": 0.7619,
+    "decision_ceiling": 0.7619,
+    "end_to_end_coverage": 0.5246,
+    "false_accepts": 77,
+    "unknown_false_accepts": 48,
+    "unenrolled_false_accepts": 19,
+    "false_accept_rate": 1.0,
+    "near_unknown_false_accept_rate": 1.0,
+    "precision": 0.2936,
+    "max_sku_share": 0.2569,
+    "hub_sku": "sku_l09",
+    "hub_alarm": false,
+    "histogram": {
+     "sku_l09": 28,
+     "sku_l05": 19,
+     "sku_l02": 15,
+     "sku_l03": 14,
+     "sku_l06": 13,
+     "sku_l11": 12,
+     "sku_l01": 8
+    },
+    "per_sku": {
+     "sku_l01": {
+      "correct": 5,
+      "total": 5
+     },
+     "sku_l02": {
+      "correct": 4,
+      "total": 4
+     },
+     "sku_l03": {
+      "correct": 4,
+      "total": 4
+     },
+     "sku_l04": {
+      "correct": 0,
+      "total": 1
+     },
+     "sku_l05": {
+      "correct": 3,
+      "total": 3
+     },
+     "sku_l06": {
+      "correct": 0,
+      "total": 4
+     },
+     "sku_l07": {
+      "correct": 0,
+      "total": 6
+     },
+     "sku_l08": {
+      "correct": 0,
+      "total": 1
+     },
+     "sku_l09": {
+      "correct": 13,
+      "total": 13
+     },
+     "sku_l10": {
+      "correct": 0,
+      "total": 11
+     },
+     "sku_l11": {
+      "correct": 3,
+      "total": 9
+     }
+    },
+    "ms_per_crop": 15.4
+   },
+   {
+    "lane": "local",
+    "accept": 2.0,
+    "margin": 0.15,
+    "in_catalog": 61,
+    "out_catalog": 48,
+    "enrolled": 42,
+    "correct": 36,
+    "wrong": 5,
+    "rejected": 68,
+    "catalog_ceiling": 0.6885,
+    "identity_ceiling": 0.9286,
+    "decision_ceiling": 0.8571,
+    "end_to_end_coverage": 0.5902,
+    "false_accepts": 5,
+    "unknown_false_accepts": 4,
+    "unenrolled_false_accepts": 0,
+    "false_accept_rate": 0.0833,
+    "near_unknown_false_accept_rate": 0.0,
+    "precision": 0.878,
+    "max_sku_share": 0.2927,
+    "hub_sku": "sku_l09",
+    "hub_alarm": false,
+    "histogram": {
+     "sku_l09": 12,
+     "sku_l11": 8,
+     "sku_l03": 6,
+     "sku_l01": 5,
+     "sku_l07": 4,
+     "sku_l02": 3,
+     "sku_l05": 3
+    },
+    "per_sku": {
+     "sku_l01": {
+      "correct": 5,
+      "total": 5
+     },
+     "sku_l02": {
+      "correct": 2,
+      "total": 4
+     },
+     "sku_l03": {
+      "correct": 4,
+      "total": 4
+     },
+     "sku_l04": {
+      "correct": 0,
+      "total": 1
+     },
+     "sku_l05": {
+      "correct": 3,
+      "total": 3
+     },
+     "sku_l06": {
+      "correct": 0,
+      "total": 4
+     },
+     "sku_l07": {
+      "correct": 4,
+      "total": 6
+     },
+     "sku_l08": {
+      "correct": 0,
+      "total": 1
+     },
+     "sku_l09": {
+      "correct": 12,
+      "total": 13
+     },
+     "sku_l10": {
+      "correct": 0,
+      "total": 11
+     },
+     "sku_l11": {
+      "correct": 6,
+      "total": 9
+     }
+    },
+    "ms_per_crop": 26.6
+   }
+  ]
+ }
+}

recognization/data/evaluations/2026-09-29-session/stream/tr_e.json

@@ -0,0 +1,554 @@
+{
+  "windows": [
+    [
+      0,
+      640,
+      1280,
+      1360
+    ],
+    [
+      12,
+      640,
+      1292,
+      1360
+    ],
+    [
+      24,
+      640,
+      1304,
+      1360
+    ],
+    [
+      36,
+      640,
+      1316,
+      1360
+    ],
+    [
+      48,
+      640,
+      1328,
+      1360
+    ],
+    [
+      60,
+      640,
+      1340,
+      1360
+    ],
+    [
+      72,
+      640,
+      1352,
+      1360
+    ],
+    [
+      84,
+      640,
+      1364,
+      1360
+    ],
+    [
+      96,
+      640,
+      1376,
+      1360
+    ],
+    [
+      108,
+      640,
+      1388,
+      1360
+    ],
+    [
+      120,
+      640,
+      1400,
+      1360
+    ],
+    [
+      132,
+      640,
+      1412,
+      1360
+    ],
+    [
+      144,
+      640,
+      1424,
+      1360
+    ],
+    [
+      156,
+      640,
+      1436,
+      1360
+    ],
+    [
+      168,
+      640,
+      1448,
+      1360
+    ],
+    [
+      180,
+      640,
+      1460,
+      1360
+    ],
+    [
+      192,
+      640,
+      1472,
+      1360
+    ],
+    [
+      204,
+      640,
+      1484,
+      1360
+    ],
+    [
+      216,
+      640,
+      1496,
+      1360
+    ],
+    [
+      220,
+      640,
+      1500,
+      1360
+    ],
+    [
+      216,
+      640,
+      1496,
+      1360
+    ],
+    [
+      204,
+      640,
+      1484,
+      1360
+    ],
+    [
+      192,
+      640,
+      1472,
+      1360
+    ],
+    [
+      180,
+      640,
+      1460,
+      1360
+    ],
+    [
+      168,
+      640,
+      1448,
+      1360
+    ],
+    [
+      156,
+      640,
+      1436,
+      1360
+    ],
+    [
+      144,
+      640,
+      1424,
+      1360
+    ],
+    [
+      132,
+      640,
+      1412,
+      1360
+    ],
+    [
+      120,
+      640,
+      1400,
+      1360
+    ],
+    [
+      108,
+      640,
+      1388,
+      1360
+    ],
+    [
+      96,
+      640,
+      1376,
+      1360
+    ],
+    [
+      84,
+      640,
+      1364,
+      1360
+    ],
+    [
+      72,
+      640,
+      1352,
+      1360
+    ],
+    [
+      60,
+      640,
+      1340,
+      1360
+    ],
+    [
+      48,
+      640,
+      1328,
+      1360
+    ],
+    [
+      36,
+      640,
+      1316,
+      1360
+    ],
+    [
+      24,
+      640,
+      1304,
+      1360
+    ],
+    [
+      12,
+      640,
+      1292,
+      1360
+    ],
+    [
+      0,
+      640,
+      1280,
+      1360
+    ],
+    [
+      12,
+      640,
+      1292,
+      1360
+    ],
+    [
+      24,
+      640,
+      1304,
+      1360
+    ],
+    [
+      36,
+      640,
+      1316,
+      1360
+    ],
+    [
+      48,
+      640,
+      1328,
+      1360
+    ],
+    [
+      60,
+      640,
+      1340,
+      1360
+    ],
+    [
+      72,
+      640,
+      1352,
+      1360
+    ],
+    [
+      84,
+      640,
+      1364,
+      1360
+    ],
+    [
+      96,
+      640,
+      1376,
+      1360
+    ],
+    [
+      108,
+      640,
+      1388,
+      1360
+    ],
+    [
+      120,
+      640,
+      1400,
+      1360
+    ],
+    [
+      132,
+      640,
+      1412,
+      1360
+    ],
+    [
+      144,
+      640,
+      1424,
+      1360
+    ],
+    [
+      156,
+      640,
+      1436,
+      1360
+    ],
+    [
+      168,
+      640,
+      1448,
+      1360
+    ],
+    [
+      180,
+      640,
+      1460,
+      1360
+    ],
+    [
+      192,
+      640,
+      1472,
+      1360
+    ],
+    [
+      204,
+      640,
+      1484,
+      1360
+    ],
+    [
+      216,
+      640,
+      1496,
+      1360
+    ],
+    [
+      220,
+      640,
+      1500,
+      1360
+    ],
+    [
+      216,
+      640,
+      1496,
+      1360
+    ],
+    [
+      204,
+      640,
+      1484,
+      1360
+    ],
+    [
+      192,
+      640,
+      1472,
+      1360
+    ],
+    [
+      180,
+      640,
+      1460,
+      1360
+    ],
+    [
+      168,
+      640,
+      1448,
+      1360
+    ],
+    [
+      156,
+      640,
+      1436,
+      1360
+    ],
+    [
+      144,
+      640,
+      1424,
+      1360
+    ],
+    [
+      132,
+      640,
+      1412,
+      1360
+    ],
+    [
+      120,
+      640,
+      1400,
+      1360
+    ],
+    [
+      108,
+      640,
+      1388,
+      1360
+    ],
+    [
+      96,
+      640,
+      1376,
+      1360
+    ],
+    [
+      84,
+      640,
+      1364,
+      1360
+    ],
+    [
+      72,
+      640,
+      1352,
+      1360
+    ],
+    [
+      60,
+      640,
+      1340,
+      1360
+    ],
+    [
+      48,
+      640,
+      1328,
+      1360
+    ],
+    [
+      36,
+      640,
+      1316,
+      1360
+    ],
+    [
+      24,
+      640,
+      1304,
+      1360
+    ],
+    [
+      12,
+      640,
+      1292,
+      1360
+    ],
+    [
+      0,
+      640,
+      1280,
+      1360
+    ],
+    [
+      12,
+      640,
+      1292,
+      1360
+    ],
+    [
+      24,
+      640,
+      1304,
+      1360
+    ],
+    [
+      36,
+      640,
+      1316,
+      1360
+    ],
+    [
+      48,
+      640,
+      1328,
+      1360
+    ],
+    [
+      60,
+      640,
+      1340,
+      1360
+    ],
+    [
+      72,
+      640,
+      1352,
+      1360
+    ],
+    [
+      84,
+      640,
+      1364,
+      1360
+    ],
+    [
+      96,
+      640,
+      1376,
+      1360
+    ],
+    [
+      108,
+      640,
+      1388,
+      1360
+    ],
+    [
+      120,
+      640,
+      1400,
+      1360
+    ],
+    [
+      132,
+      640,
+      1412,
+      1360
+    ],
+    [
+      144,
+      640,
+      1424,
+      1360
+    ],
+    [
+      156,
+      640,
+      1436,
+      1360
+    ],
+    [
+      168,
+      640,
+      1448,
+      1360
+    ]
+  ],
+  "source": "add/image6.jpg",
+  "pace": "send",
+  "seconds": 8.1,
+  "tracks": []
+}

recognization/data/evaluations/2026-09-29-session/stream/tracks_local_result.json

@@ -0,0 +1,721 @@
+{
+  "source": "add/image6.jpg",
+  "pace": "result",
+  "seconds": 12.0,
+  "tracks": [
+    {
+      "track_id": 1,
+      "sku_id": "sku_l09",
+      "label_box": [
+        580,
+        122,
+        666,
+        332
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 108,
+      "inherited": true,
+      "restarts": 2
+    },
+    {
+      "track_id": 2,
+      "sku_id": "sku_l09",
+      "label_box": [
+        659,
+        760,
+        746,
+        972
+      ],
+      "label_frame_id": 71,
+      "result_frame_id": 81,
+      "inherited": true,
+      "restarts": 2
+    },
+    {
+      "track_id": 4,
+      "sku_id": "sku_l09",
+      "label_box": [
+        586,
+        119,
+        674,
+        333
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 115,
+      "inherited": true,
+      "restarts": 5
+    },
+    {
+      "track_id": 5,
+      "sku_id": "sku_l09",
+      "label_box": [
+        512,
+        755,
+        604,
+        972
+      ],
+      "label_frame_id": 98,
+      "result_frame_id": 114,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 6,
+      "sku_id": "sku_l09",
+      "label_box": [
+        937,
+        764,
+        1036,
+        975
+      ],
+      "label_frame_id": 16,
+      "result_frame_id": 25,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 8,
+      "sku_id": "sku_l09",
+      "label_box": [
+        291,
+        754,
+        389,
+        971
+      ],
+      "label_frame_id": 107,
+      "result_frame_id": 114,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 9,
+      "sku_id": "sku_l09",
+      "label_box": [
+        459,
+        1068,
+        555,
+        1251
+      ],
+      "label_frame_id": 67,
+      "result_frame_id": 89,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 11,
+      "sku_id": "sku_l09",
+      "label_box": [
+        688,
+        1057,
+        789,
+        1142
+      ],
+      "label_frame_id": 50,
+      "result_frame_id": 59,
+      "inherited": false,
+      "restarts": 1
+    },
+    {
+      "track_id": 12,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1254,
+        1053,
+        1334,
+        1244
+      ],
+      "label_frame_id": 61,
+      "result_frame_id": 69,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 13,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1079,
+        762,
+        1176,
+        973
+      ],
+      "label_frame_id": 89,
+      "result_frame_id": 115,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 14,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1277,
+        1065,
+        1345,
+        1239
+      ],
+      "label_frame_id": 52,
+      "result_frame_id": 63,
+      "inherited": true,
+      "restarts": 2
+    },
+    {
+      "track_id": 15,
+      "sku_id": "sku_l09",
+      "label_box": [
+        798,
+        765,
+        966,
+        971
+      ],
+      "label_frame_id": 83,
+      "result_frame_id": 114,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 16,
+      "sku_id": "sku_l09",
+      "label_box": [
+        430,
+        1070,
+        543,
+        1250
+      ],
+      "label_frame_id": 99,
+      "result_frame_id": 113,
+      "inherited": false,
+      "restarts": 1
+    },
+    {
+      "track_id": 17,
+      "sku_id": "sku_l10",
+      "label_box": [
+        878,
+        1146,
+        970,
+        1245
+      ],
+      "label_frame_id": 90,
+      "result_frame_id": 106,
+      "inherited": false,
+      "restarts": 0
+    },
+    {
+      "track_id": 18,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1128,
+        763,
+        1271,
+        973
+      ],
+      "label_frame_id": 99,
+      "result_frame_id": 112,
+      "inherited": false,
+      "restarts": 10
+    },
+    {
+      "track_id": 19,
+      "sku_id": "sku_l09",
+      "label_box": [
+        680,
+        1058,
+        780,
+        1142
+      ],
+      "label_frame_id": 97,
+      "result_frame_id": 110,
+      "inherited": false,
+      "restarts": 2
+    },
+    {
+      "track_id": 21,
+      "sku_id": "sku_l10",
+      "label_box": [
+        673,
+        1146,
+        773,
+        1247
+      ],
+      "label_frame_id": 98,
+      "result_frame_id": 111,
+      "inherited": false,
+      "restarts": 0
+    },
+    {
+      "track_id": 22,
+      "sku_id": "sku_l09",
+      "label_box": [
+        287,
+        775,
+        347,
+        969
+      ],
+      "label_frame_id": 2,
+      "result_frame_id": 18,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 23,
+      "sku_id": "sku_l10",
+      "label_box": [
+        917,
+        1146,
+        999,
+        1244
+      ],
+      "label_frame_id": 98,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 25,
+      "sku_id": "sku_l09",
+      "label_box": [
+        488,
+        1078,
+        570,
+        1248
+      ],
+      "label_frame_id": 33,
+      "result_frame_id": 53,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 26,
+      "sku_id": "sku_l09",
+      "label_box": [
+        198,
+        755,
+        258,
+        972
+      ],
+      "label_frame_id": 2,
+      "result_frame_id": 15,
+      "inherited": false,
+      "restarts": 0
+    },
+    {
+      "track_id": 27,
+      "sku_id": "sku_l09",
+      "label_box": [
+        484,
+        1062,
+        572,
+        1248
+      ],
+      "label_frame_id": 89,
+      "result_frame_id": 97,
+      "inherited": true,
+      "restarts": 2
+    },
+    {
+      "track_id": 29,
+      "sku_id": "sku_l09",
+      "label_box": [
+        879,
+        1055,
+        972,
+        1141
+      ],
+      "label_frame_id": 57,
+      "result_frame_id": 70,
+      "inherited": true,
+      "restarts": 4
+    },
+    {
+      "track_id": 32,
+      "sku_id": "sku_l10",
+      "label_box": [
+        67,
+        422,
+        248,
+        613
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 47,
+      "inherited": true,
+      "restarts": 2
+    },
+    {
+      "track_id": 35,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1117,
+        763,
+        1179,
+        973
+      ],
+      "label_frame_id": 85,
+      "result_frame_id": 113,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 58,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1328,
+        1065,
+        1379,
+        1238
+      ],
+      "label_frame_id": 20,
+      "result_frame_id": 23,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 59,
+      "sku_id": "sku_l09",
+      "label_box": [
+        951,
+        124,
+        1090,
+        218
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 104,
+      "inherited": true,
+      "restarts": 5
+    },
+    {
+      "track_id": 114,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1191,
+        1054,
+        1271,
+        1245
+      ],
+      "label_frame_id": 99,
+      "result_frame_id": 111,
+      "inherited": false,
+      "restarts": 3
+    },
+    {
+      "track_id": 127,
+      "sku_id": "sku_l09",
+      "label_box": [
+        243,
+        755,
+        298,
+        972
+      ],
+      "label_frame_id": 35,
+      "result_frame_id": 40,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 135,
+      "sku_id": "sku_l09",
+      "label_box": [
+        521,
+        118,
+        608,
+        330
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 115,
+      "inherited": true,
+      "restarts": 4
+    },
+    {
+      "track_id": 136,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1072,
+        1054,
+        1147,
+        1245
+      ],
+      "label_frame_id": 53,
+      "result_frame_id": 59,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 163,
+      "sku_id": "sku_l09",
+      "label_box": [
+        125,
+        752,
+        191,
+        971
+      ],
+      "label_frame_id": 32,
+      "result_frame_id": 48,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 178,
+      "sku_id": "sku_l10",
+      "label_box": [
+        156,
+        419,
+        307,
+        616
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 41,
+      "inherited": true,
+      "restarts": 5
+    },
+    {
+      "track_id": 262,
+      "sku_id": "sku_l09",
+      "label_box": [
+        79,
+        113,
+        152,
+        331
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 74,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 269,
+      "sku_id": "sku_l09",
+      "label_box": [
+        197,
+        755,
+        257,
+        973
+      ],
+      "label_frame_id": 48,
+      "result_frame_id": 57,
+      "inherited": false,
+      "restarts": 1
+    },
+    {
+      "track_id": 295,
+      "sku_id": "sku_l09",
+      "label_box": [
+        84,
+        1059,
+        179,
+        1148
+      ],
+      "label_frame_id": 78,
+      "result_frame_id": 90,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 302,
+      "sku_id": "sku_l09",
+      "label_box": [
+        941,
+        410,
+        1009,
+        605
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 50,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 315,
+      "sku_id": "sku_l09",
+      "label_box": [
+        764,
+        414,
+        848,
+        505
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 51,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 331,
+      "sku_id": "sku_l09",
+      "label_box": [
+        399,
+        1057,
+        532,
+        1255
+      ],
+      "label_frame_id": 105,
+      "result_frame_id": 115,
+      "inherited": true,
+      "restarts": 9
+    },
+    {
+      "track_id": 345,
+      "sku_id": "sku_l09",
+      "label_box": [
+        763,
+        415,
+        854,
+        499
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 115,
+      "inherited": true,
+      "restarts": 3
+    },
+    {
+      "track_id": 353,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1119,
+        1053,
+        1180,
+        1245
+      ],
+      "label_frame_id": 99,
+      "result_frame_id": 113,
+      "inherited": false,
+      "restarts": 2
+    },
+    {
+      "track_id": 370,
+      "sku_id": "sku_l09",
+      "label_box": [
+        83,
+        117,
+        129,
+        333
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 69,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 378,
+      "sku_id": "sku_l09",
+      "label_box": [
+        576,
+        758,
+        667,
+        971
+      ],
+      "label_frame_id": 106,
+      "result_frame_id": 114,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 404,
+      "sku_id": "sku_l09",
+      "label_box": [
+        68,
+        753,
+        131,
+        971
+      ],
+      "label_frame_id": 78,
+      "result_frame_id": 85,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 416,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1089,
+        416,
+        1147,
+        604
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 83,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 499,
+      "sku_id": "sku_l09",
+      "label_box": [
+        443,
+        767,
+        510,
+        968
+      ],
+      "label_frame_id": 95,
+      "result_frame_id": 97,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 507,
+      "sku_id": "sku_l09",
+      "label_box": [
+        149,
+        757,
+        208,
+        971
+      ],
+      "label_frame_id": 84,
+      "result_frame_id": 90,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 552,
+      "sku_id": "sku_l09",
+      "label_box": [
+        757,
+        413,
+        849,
+        507
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 105,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 582,
+      "sku_id": "sku_l09",
+      "label_box": [
+        701,
+        413,
+        842,
+        505
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 102,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 613,
+      "sku_id": "sku_l09",
+      "label_box": [
+        477,
+        767,
+        537,
+        969
+      ],
+      "label_frame_id": 104,
+      "result_frame_id": 108,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 673,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1087,
+        1051,
+        1143,
+        1245
+      ],
+      "label_frame_id": 108,
+      "result_frame_id": 113,
+      "inherited": true,
+      "restarts": 0
+    }
+  ]
+}

recognization/data/evaluations/2026-09-29-session/stream/tracks_local_send.json

@@ -0,0 +1,791 @@
+{
+  "source": "add/image6.jpg",
+  "pace": "send",
+  "seconds": 12.0,
+  "tracks": [
+    {
+      "track_id": 1,
+      "sku_id": "sku_l09",
+      "label_box": [
+        527,
+        122,
+        613,
+        331
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 104,
+      "inherited": true,
+      "restarts": 2
+    },
+    {
+      "track_id": 2,
+      "sku_id": "sku_l09",
+      "label_box": [
+        696,
+        760,
+        783,
+        972
+      ],
+      "label_frame_id": 70,
+      "result_frame_id": 81,
+      "inherited": true,
+      "restarts": 2
+    },
+    {
+      "track_id": 4,
+      "sku_id": "sku_l09",
+      "label_box": [
+        494,
+        120,
+        582,
+        332
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 108,
+      "inherited": true,
+      "restarts": 5
+    },
+    {
+      "track_id": 5,
+      "sku_id": "sku_l09",
+      "label_box": [
+        390,
+        757,
+        482,
+        971
+      ],
+      "label_frame_id": 106,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 6,
+      "sku_id": "sku_l09",
+      "label_box": [
+        949,
+        763,
+        1049,
+        973
+      ],
+      "label_frame_id": 16,
+      "result_frame_id": 25,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 8,
+      "sku_id": "sku_l09",
+      "label_box": [
+        399,
+        754,
+        497,
+        971
+      ],
+      "label_frame_id": 98,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 9,
+      "sku_id": "sku_l09",
+      "label_box": [
+        441,
+        1071,
+        536,
+        1250
+      ],
+      "label_frame_id": 74,
+      "result_frame_id": 89,
+      "inherited": false,
+      "restarts": 0
+    },
+    {
+      "track_id": 12,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1194,
+        1053,
+        1274,
+        1244
+      ],
+      "label_frame_id": 50,
+      "result_frame_id": 69,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 13,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1127,
+        763,
+        1224,
+        973
+      ],
+      "label_frame_id": 97,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 14,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1277,
+        1065,
+        1345,
+        1239
+      ],
+      "label_frame_id": 52,
+      "result_frame_id": 63,
+      "inherited": true,
+      "restarts": 2
+    },
+    {
+      "track_id": 15,
+      "sku_id": "sku_l09",
+      "label_box": [
+        805,
+        764,
+        976,
+        972
+      ],
+      "label_frame_id": 89,
+      "result_frame_id": 109,
+      "inherited": false,
+      "restarts": 0
+    },
+    {
+      "track_id": 16,
+      "sku_id": "sku_l09",
+      "label_box": [
+        583,
+        1069,
+        683,
+        1251
+      ],
+      "label_frame_id": 99,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 17,
+      "sku_id": "sku_l10",
+      "label_box": [
+        867,
+        1147,
+        958,
+        1245
+      ],
+      "label_frame_id": 98,
+      "result_frame_id": 109,
+      "inherited": false,
+      "restarts": 0
+    },
+    {
+      "track_id": 18,
+      "sku_id": "sku_l09",
+      "label_box": [
+        926,
+        123,
+        1068,
+        332
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 102,
+      "inherited": true,
+      "restarts": 9
+    },
+    {
+      "track_id": 19,
+      "sku_id": "sku_l09",
+      "label_box": [
+        694,
+        1057,
+        794,
+        1142
+      ],
+      "label_frame_id": 108,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 2
+    },
+    {
+      "track_id": 20,
+      "sku_id": "sku_l10",
+      "label_box": [
+        211,
+        1128,
+        285,
+        1260
+      ],
+      "label_frame_id": 6,
+      "result_frame_id": 17,
+      "inherited": false,
+      "restarts": 2
+    },
+    {
+      "track_id": 21,
+      "sku_id": "sku_l10",
+      "label_box": [
+        848,
+        1145,
+        948,
+        1246
+      ],
+      "label_frame_id": 98,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 22,
+      "sku_id": "sku_l09",
+      "label_box": [
+        287,
+        775,
+        347,
+        969
+      ],
+      "label_frame_id": 2,
+      "result_frame_id": 18,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 23,
+      "sku_id": "sku_l10",
+      "label_box": [
+        869,
+        1144,
+        952,
+        1244
+      ],
+      "label_frame_id": 104,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 25,
+      "sku_id": "sku_l09",
+      "label_box": [
+        488,
+        1078,
+        570,
+        1248
+      ],
+      "label_frame_id": 33,
+      "result_frame_id": 53,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 26,
+      "sku_id": "sku_l09",
+      "label_box": [
+        198,
+        755,
+        258,
+        972
+      ],
+      "label_frame_id": 2,
+      "result_frame_id": 14,
+      "inherited": false,
+      "restarts": 0
+    },
+    {
+      "track_id": 27,
+      "sku_id": "sku_l09",
+      "label_box": [
+        522,
+        1067,
+        610,
+        1248
+      ],
+      "label_frame_id": 87,
+      "result_frame_id": 92,
+      "inherited": true,
+      "restarts": 2
+    },
+    {
+      "track_id": 29,
+      "sku_id": "sku_l09",
+      "label_box": [
+        947,
+        1055,
+        1038,
+        1139
+      ],
+      "label_frame_id": 58,
+      "result_frame_id": 73,
+      "inherited": true,
+      "restarts": 4
+    },
+    {
+      "track_id": 32,
+      "sku_id": "sku_l10",
+      "label_box": [
+        67,
+        422,
+        248,
+        613
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 47,
+      "inherited": true,
+      "restarts": 2
+    },
+    {
+      "track_id": 35,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1057,
+        764,
+        1118,
+        972
+      ],
+      "label_frame_id": 75,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 58,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1264,
+        1065,
+        1315,
+        1238
+      ],
+      "label_frame_id": 14,
+      "result_frame_id": 23,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 59,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1092,
+        126,
+        1229,
+        218
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 5
+    },
+    {
+      "track_id": 114,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1191,
+        1054,
+        1271,
+        1245
+      ],
+      "label_frame_id": 99,
+      "result_frame_id": 108,
+      "inherited": false,
+      "restarts": 3
+    },
+    {
+      "track_id": 127,
+      "sku_id": "sku_l09",
+      "label_box": [
+        354,
+        756,
+        409,
+        973
+      ],
+      "label_frame_id": 26,
+      "result_frame_id": 40,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 135,
+      "sku_id": "sku_l09",
+      "label_box": [
+        658,
+        760,
+        744,
+        972
+      ],
+      "label_frame_id": 92,
+      "result_frame_id": 104,
+      "inherited": true,
+      "restarts": 3
+    },
+    {
+      "track_id": 136,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1076,
+        1053,
+        1152,
+        1245
+      ],
+      "label_frame_id": 54,
+      "result_frame_id": 59,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 163,
+      "sku_id": "sku_l09",
+      "label_box": [
+        111,
+        754,
+        177,
+        970
+      ],
+      "label_frame_id": 32,
+      "result_frame_id": 50,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 178,
+      "sku_id": "sku_l10",
+      "label_box": [
+        156,
+        419,
+        307,
+        616
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 41,
+      "inherited": true,
+      "restarts": 5
+    },
+    {
+      "track_id": 262,
+      "sku_id": "sku_l09",
+      "label_box": [
+        449,
+        763,
+        522,
+        968
+      ],
+      "label_frame_id": 44,
+      "result_frame_id": 55,
+      "inherited": false,
+      "restarts": 0
+    },
+    {
+      "track_id": 269,
+      "sku_id": "sku_l09",
+      "label_box": [
+        113,
+        758,
+        172,
+        972
+      ],
+      "label_frame_id": 46,
+      "result_frame_id": 55,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 295,
+      "sku_id": "sku_l09",
+      "label_box": [
+        84,
+        1059,
+        179,
+        1148
+      ],
+      "label_frame_id": 78,
+      "result_frame_id": 90,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 302,
+      "sku_id": "sku_l09",
+      "label_box": [
+        929,
+        414,
+        995,
+        603
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 51,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 315,
+      "sku_id": "sku_l09",
+      "label_box": [
+        757,
+        413,
+        849,
+        507
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 67,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 331,
+      "sku_id": "sku_l09",
+      "label_box": [
+        422,
+        1057,
+        555,
+        1255
+      ],
+      "label_frame_id": 100,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 9
+    },
+    {
+      "track_id": 345,
+      "sku_id": "sku_l09",
+      "label_box": [
+        835,
+        1055,
+        927,
+        1140
+      ],
+      "label_frame_id": 107,
+      "result_frame_id": 111,
+      "inherited": true,
+      "restarts": 2
+    },
+    {
+      "track_id": 353,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1119,
+        1053,
+        1180,
+        1245
+      ],
+      "label_frame_id": 99,
+      "result_frame_id": 111,
+      "inherited": false,
+      "restarts": 2
+    },
+    {
+      "track_id": 370,
+      "sku_id": "sku_l09",
+      "label_box": [
+        48,
+        118,
+        91,
+        331
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 66,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 378,
+      "sku_id": "sku_l09",
+      "label_box": [
+        696,
+        758,
+        787,
+        971
+      ],
+      "label_frame_id": 97,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 404,
+      "sku_id": "sku_l09",
+      "label_box": [
+        45,
+        112,
+        83,
+        331
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 71,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 416,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1089,
+        416,
+        1147,
+        604
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 83,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 476,
+      "sku_id": "sku_l10",
+      "label_box": [
+        63,
+        519,
+        138,
+        614
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 80,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 499,
+      "sku_id": "sku_l09",
+      "label_box": [
+        79,
+        113,
+        152,
+        331
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 507,
+      "sku_id": "sku_l09",
+      "label_box": [
+        128,
+        114,
+        187,
+        331
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 85,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 536,
+      "sku_id": "sku_l09",
+      "label_box": [
+        619,
+        1075,
+        689,
+        1248
+      ],
+      "label_frame_id": 100,
+      "result_frame_id": 108,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 552,
+      "sku_id": "sku_l09",
+      "label_box": [
+        757,
+        413,
+        849,
+        507
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 105,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 568,
+      "sku_id": "sku_l10",
+      "label_box": [
+        45,
+        512,
+        143,
+        619
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 90,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 582,
+      "sku_id": "sku_l09",
+      "label_box": [
+        684,
+        413,
+        826,
+        506
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 102,
+      "inherited": true,
+      "restarts": 1
+    },
+    {
+      "track_id": 606,
+      "sku_id": "sku_l09",
+      "label_box": [
+        204,
+        758,
+        247,
+        971
+      ],
+      "label_frame_id": 102,
+      "result_frame_id": 112,
+      "inherited": false,
+      "restarts": 1
+    },
+    {
+      "track_id": 613,
+      "sku_id": "sku_l09",
+      "label_box": [
+        453,
+        773,
+        510,
+        969
+      ],
+      "label_frame_id": 100,
+      "result_frame_id": 111,
+      "inherited": false,
+      "restarts": 0
+    },
+    {
+      "track_id": 653,
+      "sku_id": "sku_l10",
+      "label_box": [
+        60,
+        510,
+        163,
+        622
+      ],
+      "label_frame_id": 0,
+      "result_frame_id": 104,
+      "inherited": true,
+      "restarts": 0
+    },
+    {
+      "track_id": 673,
+      "sku_id": "sku_l09",
+      "label_box": [
+        1087,
+        1051,
+        1143,
+        1245
+      ],
+      "label_frame_id": 108,
+      "result_frame_id": 112,
+      "inherited": true,
+      "restarts": 0
+    }
+  ]
+}

recognization/data/external-eval/README.md

@@ -0,0 +1,60 @@
+# External evidence (2026-09-29)
+
+Independent-source media used to stress the identity pipeline. **None of this is
+Tier 2.** DECISION.md's Tier 2 is a second labelled store visit; these are
+`external_proxy` artifacts, and every manifest here says so.
+
+| Directory | What it holds |
+|---|---|
+| `other-mayo/` | 8 accepted images for `sku_l04`, `sku_l07`, `sku_l09` from 6 source domains, plus the scored result |
+| `l12-reconciliation/` | What `sku_l12` actually is, from the shelf tag and package pixels |
+| `kewpie/` | 12 accepted images for six SKUs from 6 source domains, plus the scored result. One is now enrolled, so it is excluded from scoring. |
+| `tier3-sources/` | Two reusable pan clips and 11 rejected candidates for a future Tier-3 replay |
+
+## Leak checks
+
+Two, both in `code/eval_external.py`, both run before any number is reported:
+
+- **Against the evaluation photo.** Every image is template-matched against
+  `add/image6.jpg`. The highest score across all 20 is 0.675 (`sku_l07_1`), and
+  the 0.95+ band that identifies a crop of the evaluation photo is empty.
+- **Against the catalog.** An image that has since been enrolled is refused
+  rather than scored. `kewpie/sku_l06_1.jpg` is one: it is now the enrolled l06
+  view, so scoring it would be training accuracy wearing a held-out label.
+
+## Result: 19 scored, 0 false accepts
+
+    .venv/bin/python code/eval_external.py --stock stock-lopia-clean \
+        data/external-eval/kewpie data/external-eval/other-mayo
+
+| Set | Scored | Correct | Abstain | False accepts | Leaked |
+|---|---|---|---|---|---|
+| `kewpie/` | 11 | 8 | 3 | **0** | 1 |
+| `other-mayo/` | 8 | 4 | 4 | **0** | 0 |
+| **Total** | **19** | **12** | **7** | **0** | **1** |
+
+Two things this establishes that the image6 shelf cannot:
+
+- **`sku_l07` is answered with no enrolled image.** Both l07 images accept
+  through the text lane alone: `sku_l07_1` reads ラッキー フヨネーズ, and
+  `sku_l07_2` reads the rear label's 製造者：丸和油脂株式会社 with 内容量：200g.
+  The catalog row is 丸和油脂 ラッキーマヨネーズ 200g, so the manufacturer and
+  the size together name it. This is the text lane's coverage claim - that a
+  product with no reference image can still be named - holding on data that was
+  not used to build it.
+- **The l04 abstentions are correct behaviour, not a miss.** l04 has no
+  enrolled image either, and its reads are weak: `sku_l04_2` reads only
+  コと旨 タ 味 (all rows score 0.0), and `sku_l04_4` reads the rear label with
+  やまや garbled to まや, scoring 0.379 - below the accept line. Abstaining on
+  a weak read is the designed outcome.
+
+`sku_l07_2` accepts with a negative local score (-0.55) on text evidence alone.
+That is the lane doing its job, but it is worth naming: the accept rests on two
+matched tokens (manufacturer + size), not on geometry.
+
+## What is still missing
+
+No external source gives per-SKU identity for the whole catalog, and no
+reusable clip combines the target category with shelf-distance text
+resolution. The Tier-2 and Tier-3 gates in DECISION.md still need the
+operator's own recorded walks.

recognization/data/external-eval/kewpie/eval_result.json

@@ -0,0 +1,150 @@
+{
+ "dataset": "kewpie",
+ "catalog": "stock-lopia-clean",
+ "note": "External proxy images, not Tier 2. Scored through the shipped LocalLaneAdapter + DecisionPolicy.",
+ "rows": [
+  {
+   "file": "sku_l01_1.jpg",
+   "truth": "sku_l01",
+   "best": "sku_l01",
+   "confidence": 13.1618,
+   "status": "success",
+   "source": "local",
+   "verdict": "correct"
+  },
+  {
+   "file": "sku_l01_2.jpg",
+   "truth": "sku_l01",
+   "best": "sku_l01",
+   "confidence": 20.986,
+   "status": "success",
+   "source": "local",
+   "verdict": "correct"
+  },
+  {
+   "file": "sku_l02_1.jpg",
+   "truth": "sku_l02",
+   "best": "sku_l02",
+   "confidence": 10.4288,
+   "status": "success",
+   "source": "text",
+   "verdict": "correct"
+  },
+  {
+   "file": "sku_l02_2.jpg",
+   "truth": "sku_l02",
+   "best": "sku_l02",
+   "confidence": 7.542,
+   "status": "success",
+   "source": "local",
+   "verdict": "correct"
+  },
+  {
+   "file": "sku_l03_1.jpg",
+   "truth": "sku_l03",
+   "best": "sku_l03",
+   "confidence": 14.3581,
+   "status": "success",
+   "source": "local",
+   "verdict": "correct"
+  },
+  {
+   "file": "sku_l03_2.jpg",
+   "truth": "sku_l03",
+   "best": "sku_l03",
+   "confidence": 48.9767,
+   "status": "success",
+   "source": "local",
+   "verdict": "correct"
+  },
+  {
+   "file": "sku_l05_1.jpg",
+   "truth": "sku_l05",
+   "best": null,
+   "confidence": 0.0,
+   "status": "unknown",
+   "source": null,
+   "verdict": "abstain"
+  },
+  {
+   "file": "sku_l05_2.jpg",
+   "truth": "sku_l05",
+   "best": "sku_l05",
+   "confidence": -3.9674,
+   "status": "success",
+   "source": "text",
+   "verdict": "correct"
+  },
+  {
+   "file": "sku_l06_1.jpg",
+   "truth": "sku_l06",
+   "verdict": "leaked",
+   "leaked_into": "sku_l06",
+   "note": "enrolled in the catalog; scoring it would be training accuracy"
+  },
+  {
+   "file": "sku_l06_2.jpg",
+   "truth": "sku_l06",
+   "best": "sku_l06",
+   "confidence": 294.8985,
+   "status": "success",
+   "source": "local",
+   "verdict": "correct"
+  },
+  {
+   "file": "sku_l11_1.jpg",
+   "truth": "sku_l11",
+   "best": null,
+   "confidence": -29.9175,
+   "status": "unknown",
+   "source": null,
+   "verdict": "abstain"
+  },
+  {
+   "file": "sku_l11_2.jpg",
+   "truth": "sku_l11",
+   "best": null,
+   "confidence": 14.1552,
+   "status": "ambiguous",
+   "source": null,
+   "verdict": "abstain"
+  }
+ ],
+ "scored": 11,
+ "leaked": 1,
+ "by_sku": {
+  "sku_l01": {
+   "correct": 2,
+   "abstain": 0,
+   "wrong": 0
+  },
+  "sku_l02": {
+   "correct": 2,
+   "abstain": 0,
+   "wrong": 0
+  },
+  "sku_l03": {
+   "correct": 2,
+   "abstain": 0,
+   "wrong": 0
+  },
+  "sku_l05": {
+   "correct": 1,
+   "abstain": 1,
+   "wrong": 0
+  },
+  "sku_l06": {
+   "correct": 1,
+   "abstain": 0,
+   "wrong": 0
+  },
+  "sku_l11": {
+   "correct": 0,
+   "abstain": 2,
+   "wrong": 0
+  }
+ },
+ "correct": 8,
+ "abstain": 3,
+ "wrong": 0
+}

recognization/data/external-eval/kewpie/manifest.json

@@ -0,0 +1,191 @@
+{
+  "dataset_type": "external_proxy",
+  "overall_note": "Independent-source Internet media for recognition stress testing. This is an external_proxy set: it is not held-out store-visit Tier 2 data and does not satisfy the DECISION.md Tier-2 capture protocol. Every accepted file was visually inspected for the exact Kewpie product and package size. Official Kewpie front packshots already enrolled in stock-lopia-clean were excluded. The two sku_l06 files are current-package retailer images showing 内臓脂肪を減らす キユーピーフィッテ; customer-listing photos showing the superseded 内臓脂肪ケア wrapper were inspected and rejected.",
+  "license_note": "All accepted files come from commercial retailer or customer marketplace pages without an explicit reuse grant. They are marked restricted-research for internal evaluation only and should not be redistributed.",
+  "coverage": {
+    "accepted_images": 12,
+    "accepted_skus": [
+      "sku_l01",
+      "sku_l02",
+      "sku_l03",
+      "sku_l05",
+      "sku_l06",
+      "sku_l11"
+    ],
+    "two_images_per_sku": [
+      "sku_l01",
+      "sku_l02",
+      "sku_l03",
+      "sku_l05",
+      "sku_l06",
+      "sku_l11"
+    ],
+    "two_source_domains_per_sku": [
+      "sku_l01",
+      "sku_l02",
+      "sku_l03",
+      "sku_l05",
+      "sku_l06",
+      "sku_l11"
+    ],
+    "not_found": []
+  },
+  "images": [
+    {
+      "file": "sku_l01_1.jpg",
+      "sku": "sku_l01",
+      "product": "キユーピー エッグケア 305g",
+      "source_page": "https://paypayfleamarket.yahoo.co.jp/item/z686464976",
+      "source_title": "キユーピー エッグケア 卵を使っていないマヨネーズタイプ 305g",
+      "media_url": "https://auctions.c.yimg.jp/images.auctions.yahoo.co.jp/image/dr000/auc0209/users/ad8016fc0a2ac99427f4bd970e286c7709594046/i-img1200x1200-1789785070735v2rkwx.jpg",
+      "width": 1200,
+      "height": 1200,
+      "view_type": "customer_front",
+      "license_status": "restricted-research",
+      "notes": "Customer-taken marketplace photo on a household surface. The front package visibly reads エッグケア and 305g; exact title corroborates the variant. Not the official Kewpie enrollment packshot. Source domain: paypayfleamarket.yahoo.co.jp; media host: auctions.c.yimg.jp."
+    },
+    {
+      "file": "sku_l01_2.jpg",
+      "sku": "sku_l01",
+      "product": "キユーピー エッグケア 305g",
+      "source_page": "https://www.amazon.co.jp/dp/B0BSVY8NNH",
+      "source_title": "キユーピー エッグケア (卵不使用) マヨネーズ タイプ 305g",
+      "media_url": "https://m.media-amazon.com/images/I/71lzmm5n5cL._AC_SL1500_.jpg",
+      "width": 1500,
+      "height": 1500,
+      "view_type": "retailer_in_use",
+      "license_status": "restricted-research",
+      "notes": "Retailer alternate image with the exact 305g package beside prepared potato salad; package text and size are visible. Not the enrolled plain front packshot. Source/media domain: amazon.co.jp / media-amazon.com."
+    },
+    {
+      "file": "sku_l02_1.jpg",
+      "sku": "sku_l02",
+      "product": "キユーピー アマニ油マヨネーズ 200g",
+      "source_page": "https://paypayfleamarket.yahoo.co.jp/item/z599237268",
+      "source_title": "キユーピー アマニ油マヨネーズ 200g 2個セット 血圧が高めの方に",
+      "media_url": "https://auctions.c.yimg.jp/images.auctions.yahoo.co.jp/image/dr000/auc0204/users/f571f9a659b398818c6c54e9862bf90f0e2d78ec/i-img900x1200-1777250065888ef77yp.jpg",
+      "width": 900,
+      "height": 1200,
+      "view_type": "customer_front_pair",
+      "license_status": "restricted-research",
+      "notes": "Customer-taken two-pack photo. Both fronts visibly read 血圧ケア アマニ油マヨネーズ and 200g; exact title corroborates size. Source domain: paypayfleamarket.yahoo.co.jp; media host: auctions.c.yimg.jp."
+    },
+    {
+      "file": "sku_l02_2.jpg",
+      "sku": "sku_l02",
+      "product": "キユーピー アマニ油マヨネーズ 200g",
+      "source_page": "https://www.amazon.co.jp/dp/B01F8L6KSQ",
+      "source_title": "キユーピー アマニ油マヨネーズ 200g [機能性表示食品]",
+      "media_url": "https://m.media-amazon.com/images/I/71s+50hXsPL._AC_SL1005_.jpg",
+      "width": 1005,
+      "height": 937,
+      "view_type": "back_label",
+      "license_status": "restricted-research",
+      "notes": "Retailer alternate rear-label view from the exact 200g listing. The unfolded back panel shows the アマニ油マヨネーズ label and 200g content declaration. Not a front enrollment image. Source/media domain: amazon.co.jp / media-amazon.com."
+    },
+    {
+      "file": "sku_l03_1.jpg",
+      "sku": "sku_l03",
+      "product": "キユーピー からしマヨネーズ 200g",
+      "source_page": "https://paypayfleamarket.yahoo.co.jp/item/z624032610",
+      "source_title": "キユーピー からしマヨネーズ 200g 4本セット",
+      "media_url": "https://auctions.c.yimg.jp/images.auctions.yahoo.co.jp/image/dr000/auc0206/users/198e53fdfc1edde3b4933f53fd4f0c02c3346cf5/i-img1200x714-1781002147108o12vd1.jpg",
+      "width": 1200,
+      "height": 714,
+      "view_type": "customer_front_bundle",
+      "license_status": "restricted-research",
+      "notes": "Customer-taken four-pack photo with visible からしマヨネーズ fronts; listing title fixes each unit at 200g. Household lighting and grouped packages differ from enrollment. Source domain: paypayfleamarket.yahoo.co.jp; media host: auctions.c.yimg.jp."
+    },
+    {
+      "file": "sku_l03_2.jpg",
+      "sku": "sku_l03",
+      "product": "キユーピー からしマヨネーズ 200g",
+      "source_page": "https://www.amazon.co.jp/dp/B0069K29EK",
+      "source_title": "キユーピー からしマヨネーズ 200g",
+      "media_url": "https://m.media-amazon.com/images/I/81ZODR6dv8L._AC_SL1500_.jpg",
+      "width": 1500,
+      "height": 1500,
+      "view_type": "retailer_in_use",
+      "license_status": "restricted-research",
+      "notes": "Retailer alternate image showing the exact 200g からしマヨネーズ package beside a prepared dish; the package front and 200g declaration are legible. Not the enrolled plain front packshot. Source/media domain: amazon.co.jp / media-amazon.com."
+    },
+    {
+      "file": "sku_l05_1.jpg",
+      "sku": "sku_l05",
+      "product": "キユーピー ゼロ ノンコレステロール 310g",
+      "source_page": "https://www.amazon.co.jp/dp/B0CBPWDXZJ",
+      "source_title": "キユーピー ゼロ ノンコレステロール マヨネーズ タイプ 310g",
+      "media_url": "https://m.media-amazon.com/images/I/81ykD7vVmIL._AC_SL1500_.jpg",
+      "width": 1305,
+      "height": 1500,
+      "view_type": "retailer_in_use_collage",
+      "license_status": "restricted-research",
+      "notes": "Retailer alternate usage collage with the exact キユーピーゼロ 310g package among prepared dishes. The composition differs materially from the enrolled front packshot. Source/media domain: amazon.co.jp / media-amazon.com."
+    },
+    {
+      "file": "sku_l05_2.jpg",
+      "sku": "sku_l05",
+      "product": "キユーピー ゼロ ノンコレステロール 310g",
+      "source_page": "https://store.shopping.yahoo.co.jp/eight-cat-store/s-b0cnsbqcly-20260717.html",
+      "source_title": "キユーピー ゼロ ノンコレステロール マヨネーズ タイプ 310g ×4個",
+      "media_url": "https://item-shopping.c.yimg.jp/i/n/eight-cat-store_s-b0cnsbqcly-20260717_4_d_20260717075244",
+      "width": 900,
+      "height": 900,
+      "view_type": "retailer_in_use_collage",
+      "license_status": "restricted-research",
+      "notes": "Independent retailer-hosted usage collage with the exact キユーピーゼロ 310g package shown among prepared dishes; the package front and 310g declaration are legible. Chosen over the same listing's plain front image so the set stays non-enrollment. Source domain: store.shopping.yahoo.co.jp; media host: item-shopping.c.yimg.jp."
+    },
+    {
+      "file": "sku_l06_1.jpg",
+      "sku": "sku_l06",
+      "product": "内臓脂肪を減らす キユーピーフィッテ 210g",
+      "source_page": "https://www.amazon.co.jp/dp/B08V2676LK",
+      "source_title": "キユーピー フィッテ 210g",
+      "media_url": "https://m.media-amazon.com/images/I/81zckoUh+vL._AC_SL1500_.jpg",
+      "width": 654,
+      "height": 1500,
+      "view_type": "retailer_current_package_front",
+      "license_status": "restricted-research",
+      "notes": "F-12 current-package evidence. The wrapper visibly says 内臓脂肪を減らす キユーピーフィッテ and 210g; the product page repeats 内臓脂肪を減らす and contains no 内臓脂肪ケア wording. This is not the superseded enrolled package. Source/media domain: amazon.co.jp / media-amazon.com."
+    },
+    {
+      "file": "sku_l06_2.jpg",
+      "sku": "sku_l06",
+      "product": "内臓脂肪を減らす キユーピーフィッテ 210g",
+      "source_page": "https://store.shopping.yahoo.co.jp/ladydrugheartshop/4901577083600.html",
+      "source_title": "キユーピー フィッテ 210g×5個",
+      "media_url": "https://item-shopping.c.yimg.jp/i/n/ladydrugheartshop_4901577083600_i_20260902221452",
+      "width": 500,
+      "height": 500,
+      "view_type": "retailer_current_package_front",
+      "license_status": "restricted-research",
+      "notes": "Second-domain F-12 corroboration. The current wrapper visibly says 内臓脂肪を減らす and 210g. Three PayPay Flea Market listings were also inspected, but all showed the superseded 内臓脂肪ケア wrapper and were rejected. Source domain: store.shopping.yahoo.co.jp; media host: item-shopping.c.yimg.jp."
+    },
+    {
+      "file": "sku_l11_1.jpg",
+      "sku": "sku_l11",
+      "product": "キユーピー ハーフ 400g",
+      "source_page": "https://paypayfleamarket.yahoo.co.jp/item/z692700538",
+      "source_title": "キユーピー ハーフ マヨネーズ 400g 2本セット カロリー50%カット",
+      "media_url": "https://auctions.c.yimg.jp/images.auctions.yahoo.co.jp/image/dr000/auc0209/users/67aace4c38f98ea5a750a86a39665a4c7bc9898d/i-img1200x1200-1790465979767g4gnwe.jpg",
+      "width": 1200,
+      "height": 1200,
+      "view_type": "customer_front_pair",
+      "license_status": "restricted-research",
+      "notes": "Customer-taken two-pack photo. Both fronts visibly show キユーピーハーフ; the title identifies each as 400g. Recorded as 400g, not the stock CSV's 450g wording. Source domain: paypayfleamarket.yahoo.co.jp; media host: auctions.c.yimg.jp."
+    },
+    {
+      "file": "sku_l11_2.jpg",
+      "sku": "sku_l11",
+      "product": "キユーピー ハーフ 400g",
+      "source_page": "https://www.amazon.co.jp/dp/B007XT4FIO",
+      "source_title": "キユーピー ハーフ 400g",
+      "media_url": "https://m.media-amazon.com/images/I/81u8+jlddyL._AC_SL1500_.jpg",
+      "width": 1406,
+      "height": 1500,
+      "view_type": "retailer_marketing_composite",
+      "license_status": "restricted-research",
+      "notes": "Retailer alternate marketing composite with the exact 400g package and product messaging. Not the plain official Kewpie front packshot. Recorded as 400g, matching the pictured package and listing title. Source/media domain: amazon.co.jp / media-amazon.com."
+    }
+  ]
+}