From 962cdaab5fcf19bea99eed8a4f618477690dc54e Mon Sep 17 00:00:00 2001
From: kiyo9w <ngokapikapi@gmail.com>
Date: Tue, 29 Sep 2026 03:39:33 +0900
Subject: [PATCH] test(tier3): run the production pipeline over both reusable
 clips

The two clips were staged but never exercised. One frame per second through the
production detector and the shipped local lane:

- The 1080p supermarket pan: 227 crops, 0 accepted. Correct open-set behaviour -
  the frame is German pickle jars, not catalog products.
- The 360p mayonnaise bay: 1623 crops, 24 accepted (1.5%), across seven SKUs.
  The clip visibly contains Kewpie products, so those are not automatically
  false accepts, but at 640x360 the labels are too soft for per-SKU ground
  truth.

Recorded as an accept-rate probe with false_accepts and accuracy left null
rather than estimated. Tier 3 still needs an operator-recorded walk.
---
 .../tier3-sources/eval_result.json            | 46 +++++++++++++++++++
 1 file changed, 46 insertions(+)
 create mode 100644 recognization/data/external-eval/tier3-sources/eval_result.json

diff --git a/recognization/data/external-eval/tier3-sources/eval_result.json b/recognization/data/external-eval/tier3-sources/eval_result.json
new file mode 100644
index 0000000..14e43d9
--- /dev/null
+++ b/recognization/data/external-eval/tier3-sources/eval_result.json
@@ -0,0 +1,46 @@
+{
+ "dataset": "tier3-sources",
+ "tier": "tier3_proxy_candidate",
+ "note": "Not Tier 3: one frame per second was extracted from two reusable public clips and run through the production detector plus LocalLaneAdapter. No per-facing SKU ground truth exists, so this is an open-world accept-rate probe, not an accuracy score.",
+ "method": {
+  "frame_sample_hz": 1,
+  "frame_long_side": 1280,
+  "detector": "production detector",
+  "identity": "stock-lopia-clean production local lane"
+ },
+ "clips": [
+  {
+   "file": "media/hanzomon-supermarket-pan-132-152.mp4",
+   "frames": 20,
+   "detected_crops": 227,
+   "accepted": 0,
+   "accept_rate": 0.0,
+   "interpretation": "Correct open-set behaviour. Pixel inspection of frame 10 shows German Kuhne pickle jars, not catalog products."
+  },
+  {
+   "file": "media/kat-in-japan-mayonnaise-aisle-pan-694-720.mp4",
+   "frames": 26,
+   "detected_crops": 1623,
+   "accepted": 24,
+   "accept_rate": 0.0148,
+   "accepted_by_sku": {
+    "sku_l05": 13,
+    "sku_l03": 4,
+    "sku_l04": 2,
+    "sku_l08": 2,
+    "sku_l01": 1,
+    "sku_l09": 1,
+    "sku_l10": 1
+   },
+   "interpretation": "The clip visibly contains Kewpie and adjacent mayonnaise products, so accepted crops are not automatically false accepts. At 640x360 the labels are too soft for exact per-SKU ground truth. The 1.5% accept rate is therefore a triage signal only; Tier 3 still needs an operator-recorded walk with labels."
+  }
+ ],
+ "summary": {
+  "frames": 46,
+  "detected_crops": 1850,
+  "accepted": 24,
+  "accept_rate": 0.013,
+  "false_accepts": null,
+  "accuracy": null
+ }
+}