diff --git a/docs/PLAN.md b/docs/PLAN.md
index 8b32d79..4a92eeb 100644
--- a/docs/PLAN.md
+++ b/docs/PLAN.md
@@ -1971,3 +1971,37 @@ Next: freeze shared features and normalization for detection warm-up, add AI-rev
 source-train instances and class-specific confusers, then use a predeclared acceptance
 rule requiring both detection improvement and preserved scene agreement. Do not extend
 the same run or use source test / Tao-Hsin test for iteration.
+
+### 10.26 GCS source-train intake and reviewed expansion — 2026-09-16
+
+GCS access restored. Added bounded, generation-pinned intake for `gs://studioa` from
+existing train cameras only: 18 clips / 453,678,490 bytes, nine Kaohsiung/Taichung cameras,
+two dates/time slots. Recorded generations, MD5, SHA-256 and extraction offsets. This
+bucket's timestamp format differs from studioa-recording; old UTC assumptions are not
+reused. Camera views were compared with legacy images. Initial systemd PATH failure
+was diagnosed from logs and resumed with the ffprobe path; pinned worker unchanged.
+
+36 extracted frames -> 16 visually selected frames after redundancy screening.
+SAM3 produced 504 tentative positives and 553 unresolved candidates; assistant inspected
+183 individual masks in context and accepted 98 observations. One phone proposal was
+corrected to a physical tablet. Depictions, device parts, mats, headphones, remotes,
+grouped packages and contaminated masks were deferred. Repeated observations are not
+claimed as distinct physical objects; unselected proposals do not create negatives.
+
+Instances v3 now has 27 train frames / nine cameras / 129 positives; source val remains
+six frames / 31 positives. Semantic extension contains 137 frames, Tao-Hsin fold
+62 train / 33 val / 26 test, with 16 source-test frames still excluded. All 121 prior
+semantic mask files and all 17 prior instance images/targets/negative masks are
+byte-identical; every prior split role is unchanged. No test inference occurred.
+
+59 related tests passed. GPU smoke covered all 27 train frames in 14 updates with finite
+loss/gradients; all 2,366,580 checked unknown logits had zero gradient. 297 frozen inputs
+verified, all three completed services exited 0, no checkpoint written. Tool commits:
+`c7ebd13`, `460213d`, `bafe929`. Unrelated `:memory:.ses` remains untouched.
+
+[Counts, visual gallery, methods and remaining gaps](STUDIOA_GCS_EXPANSION.md),
+[tracked report](reviews/studioa_gcs_expansion_20260916.json).
+Cardboard boxes remain at three train observations; class-specific hard negatives and
+stronger validation remain pending. Next: detector warm-up with shared scene features
+and normalization frozen, then predeclared source-val acceptance checks. Data expansion
+is complete for this batch, not evidence of a successful model upgrade.
diff --git a/docs/STUDIOA_GCS_EXPANSION.md b/docs/STUDIOA_GCS_EXPANSION.md
new file mode 100644
index 0000000..bd2fb22
--- /dev/null
+++ b/docs/STUDIOA_GCS_EXPANSION.md
@@ -0,0 +1,82 @@
+# StudioA GCS 補資料 — 2026-09-16
+
+已完成第一批 GCS 下載、抽幀、AI 覆核及訓練包擴充。偵測 train 從 **11 張／31 個
+實例／7 支鏡頭**增加到 **27 張／129 個實例／9 支鏡頭**。本輪沒有重訓或替換模型。
+
+## 來源與選樣
+
+`gs://studioa` 的現有來源訓練鏡頭：高雄 4、7、8；台中 3、4、5、8、9、11。
+依既有 Tao-Hsin fold 限制取得資料，沒有把 val/test 鏡頭改成 train。
+
+- 2026-09-10 約 12:00、2026-09-12 約 18:50，18 段影片，共 453,678,490 bytes。
+- 下載上限 1 GB；每段記錄 GCS generation、MD5、大小、SHA-256 及 ffprobe 資訊。
+- 每段抽取第 30、330 秒，共 36 幀；36 幀皆非既有影像的逐像素完全重複。
+- 助理比對既有相機視角及抽幀原圖，保留 16 張，另外 20 張視為近似或相鄰冗餘。
+- 此 bucket 的檔名格式不同於 `studioa-recording`，舊腳本的 UTC 轉換不可直接套用。
+  正式批次燒錄時間約為 12:00／12:05、18:50／18:55；另一次探測片段發現時差，
+  因此檔名只用於抽樣，不用作已校準的精確事件時間。
+
+下載由持久化 systemd 工作執行。初次服務未繼承 ffprobe 路徑而失敗；補上 PATH 後
+從同一份版本鎖定清單續跑，已下載影片核對 MD5 後重用，原始錯誤日誌保留。
+
+## AI 覆核與訓練資料
+
+SAM3 固定 revision `3c879f39826c281e95690f02c7821c4de09afae7`，使用本機快取。
+16 張影像產生 504 個初步正向候選與 553 個未決候選。助理從中逐項查看 183 個
+候選的原圖脈絡與隔離遮罩，接受 **98 個新增實例標註**；其他候選不作正負真值。
+這是影像內的實例觀測數，並非 98 件不同實物；跨日期可能再次拍到同一商品。
+
+排除廣告／包裝中的手機與人物、桌墊誤認平板、筆電螢幕誤認平板、耳機誤認喇叭、
+遙控器誤認手機、展示桌誤認椅子、多盒合併及遮罩污染。另將一個 phone 候選更正
+為畫面中可確認的實體平板，保留原提案類別與更正理由。仍有疑義的類別直接保留未知。
+
+| 偵測類別 | 原 train | 新增 | 現 train | 原 val，保持不變 |
+| --- | ---: | ---: | ---: | ---: |
+| 筆電 | 3 | 6 | 9 | 5 |
+| 手機 | 4 | 13 | 17 | 4 |
+| 平板 | 4 | 11 | 15 | 1 |
+| 盒裝商品 | 3 | 7 | 10 | 1 |
+| 紙箱 | 3 | 0 | 3 | 2 |
+| 喇叭 | 2 | 7 | 9 | 1 |
+| 海報 | 2 | 18 | 20 | 2 |
+| 消防設備 | 2 | 6 | 8 | 1 |
+| 椅子 | 2 | 13 | 15 | 5 |
+| 人物 | 6 | 17 | 23 | 9 |
+| 合計 | 31 | 98 | 129 | 31 |
+
+新影像只提供已覆核正向遮罩；其餘像素為 ignore 255，沒有補成背景。未選取的
+teacher proposals 保留在 `unreviewed_proposals` 供追查，不會自動升格成監督訊號。
+負樣本仍為原先六塊 train 空地板區域；本批沒有新增類別專屬的混淆負樣本。
+
+- 語意包：`runs/studioa_gcs_training_extension_20260916_v1/semantic/`，共 137 張。
+  Tao-Hsin fold 為 62 train／33 val／26 test，另 16 張來源 test 仍排除。
+- 實例包：`runs/studioa_partial_instances_20260916_v3/`，27 train／6 val。
+- 原本 **121 張語意遮罩逐檔完全相同**，所有 fold 指派不變。
+- 原本 **17 張實例影像、正框／類別 target、空白區 mask 逐檔完全相同**。
+- 本輪沒有執行 test 推論或計算新準確率。
+
+[16 張標註總覽](../runs/studioa_gcs_expansion_20260916_v1/reviewed_contact.jpg) ·
+[逐圖瀏覽](../runs/studioa_gcs_expansion_20260916_v1/reviewed_gallery.html) ·
+[物件覆核決定](reviews/studioa_gcs_object_decisions_20260916.json) ·
+[完整實例匯出決定](reviews/studioa_gcs_merged_instance_decisions_20260916.json)
+
+## 驗證與限制
+
+59 項相關測試通過，包含新來源時間解析、拒絕 val/test 鏡頭、來源 hash、AI 選樣、
+類別修正、未選候選忽略，以及擴充前後原標籤／分割保留；lint、型別及提交檢查通過。
+
+`runs/studioa_gcs_instance_smoke_20260916_v1/` 凍結 297 個檔案，GPU 對全數 27 張
+train 執行 14 次更新，512×896、batch 2、bf16。loss／梯度有限，檢查 2,366,580 個
+未知分類輸出，非零梯度數為 **0**。分類器更新範數 0.0140317，沒有寫出 checkpoint。
+下載、AI 標註及 GPU 檢查服務最後皆正常退出，exit 0。這只證明資料接線有效。
+
+資料量增加仍不能證明模型可靠。紙箱只有 3 個 train 標註；val 僅 31 個實例，
+多類只有 1 個。相同鏡頭／實物的跨日期觀測也不是獨立樣本。所有標註由 AI 產生及
+覆核，不能解讀為人工真值、獨立準確率、完整 3D 或顧客分析能力。
+
+下一步固定共享場景特徵及 normalization，暖身偵測 head，同時補紙箱與明確的
+類別混淆負樣本；之後才比較來源 val 的偵測改善與語意保持。不要直接延長前次
+已失敗的完全聯合訓練設定。
+
+[可追溯結果與 hash](reviews/studioa_gcs_expansion_20260916.json)。資料取得工具 commit
+`c7ebd13`，raw-media AI freeze `460213d`，覆核擴充與 GPU 檢查 `bafe929`。
diff --git a/docs/STUDIOA_PARTIAL_DETECTION.md b/docs/STUDIOA_PARTIAL_DETECTION.md
index 68f1c4d..20ffbba 100644
--- a/docs/STUDIOA_PARTIAL_DETECTION.md
+++ b/docs/STUDIOA_PARTIAL_DETECTION.md
@@ -182,3 +182,157 @@ normalization 漂移可能影響語意表現，但本輪尚未隔離因果，不
 型別 ratchet 與提交 hook 通過。645 個凍結輸入及 6 個輸出 hash 核對，checkpoint
 job 身分／有限 tensor／初始權重保留皆通過。另有低解析度 CPU 全流程檢查，
 只驗證程式連接，沒有把其分數當作模型效果。
+
+## 2026-09-16：固定場景的偵測頭暖身契約
+
+GCS 擴充後 instances v3 有 27 張 train／129 個物件觀測、6 張 val／31 個觀測。
+來源為 `studioa_gcs_training_extension_20260916_v1/semantic`；舊 val 與 test 分配
+不變。同一物品可能跨日期重複出現，129 不代表獨立物品數。
+
+`studioa_train.py prepare --detector-warmup --instances … --initial-checkpoint …`
+從 scene pilot v2 的 best 初始化，只讓 `det_head.*` 參數可訓練。所有其他模組
+固定 eval，包含 BatchNorm running statistics 與 dropout，並在每次儲存前驗證
+凍結 tensor 完全相同。模型 train/eval 切換及 checkpoint 重載均保留此契約。
+scene dataset 設為 `validation_only`，不建立它的 train loader；metadata 明記
+train_size=0。全部 33 張來源 val 的 float32 scene logits 在訓練前和選定模型
+載入後計算 SHA256，必須完全相同。續跑也重新對照原始 scene checkpoint。
+
+預先固定：60 epochs 上限、batch 2、lr 2e-4、26 steps warmup、20 輪無改善停止、
+bf16 training、deterministic、關閉 TF32／cuDNN benchmark、EMA 關閉。
+best 選擇最大 reviewed-positive recall（沿用 score >0.20／IoU ≥0.50），同分
+保留較早 epoch。第 0 輪仍參與選擇。完成後只有 recall 嚴格改善、覆核空白區
+誤報不增加且場景完全一致，才記為通過此次暖身；這不是部署驗收。
+未知區預測持續獨立報告，不視為正確或錯誤；本輪不讀取 test 進行推論。
+
+紙箱仍只有 3 個 train 觀測；負樣本仍僅已覆核空地板，尚無物品間類別混淆的
+負向監督。暖身結果只能回答「固定既有場景特徵，這批部分監督能否學出偵測」，
+不能回答真實整店精度、3D 尺寸或顧客行為是否正確。
+
+## 暖身結果：場景保留成功，偵測仍不足以使用
+
+`runs/studioa_detector_warmup_20260916_v1/` 完成，程式版本 `562532d`；
+第 40 輪因連續 20 輪無嚴格改善停止，共 520 次偵測 optimizer 更新。
+選定第 20 輪（260 次更新），systemd 正常退出，test 未推論。
+
+| 檢查項目 | 結果 |
+| --- | --- |
+| 已覆核正實例召回 | 初始 0/31 → 選定 7/31（22.58%） |
+| 命中類別 | person 5/9、phone 1/4、poster 1/2；其他七類 0 |
+| 覆核空白區誤報 | 0 → 0；僅覆蓋 51,761 個輸入像素 |
+| 未配對、真偽未定預測 | 593；六張影像每張均達 100 框上限 |
+| 凍結參數及 buffers | 194 個 tensors；每次儲存皆核對，best/last 與原模型完全相同 |
+| 33 張來源 val 場景 logits | 訓練前後 SHA256 完全相同 |
+| 場景 AI 標籤 mIoU | 40 輪均為 0.2506537344807518 |
+
+通過的是「固定場景後，偵測頭得到有限學習」這項暖身契約。重疊框與類別混淆
+仍明顯，593 個未知預測不能當作真陽性，也不能據此計算 precision。相比前次
+聯合 pilot，本輪同時改了資料量與訓練隔離，不能把改善全歸因於其中一項。
+本輪關閉 TF32 並固定 deterministic 設定，mIoU 與較早 pilot 的微小差異也不能
+當成場景改善；直接證據是相同設定下完整 logits 和原場景 tensors 不變。
+
+[六張逐圖對照](../runs/studioa_detector_warmup_20260916_v1/review/selected_val_review.jpg)
+左側是 AI 覆核子集，右側只顯示前 20 框；評估仍用固定上限 100 框。
+[逐框 JSON](../runs/studioa_detector_warmup_20260916_v1/review/selected_diagnostics.json)
+可完整重算並精確重現所有偵測驗證指標。
+[可追溯結果](reviews/studioa_detector_warmup_20260916.json) 記錄設定、檢查和 hashes。
+
+757 個凍結輸入、8 個正式輸出 hash 核對通過；best/last 無非有限 tensor，
+job 身分正確，原始 scene checkpoint 沒有改寫。主要測試批次 171 項通過，
+資料與契約追加批次 20 項通過（兩批有重疊）；lint、型別與提交 hook 通過。
+metadata 的 dirty 警示來自既有無關未追蹤檔，訓練程式使用已提交且逐檔凍結
+的 snapshot；語意資料記為 train_size=0、val_size=33。
+
+下一步優先補逐類混淆負樣本的資料契約：例如確認價牌不是手機，只否定 phone
+通道，不能將實際物品整片畫成十類共同背景。紙箱三個 train 觀測全部來自同一
+張影像與同一鏡頭，應由其他允許的 source-train 鏡頭補充。所有新增覆核由 AI
+執行，val/test 分配及 score／NMS 門檻保持固定。尚不推進 Stage 2–4 的效果宣稱。
+
+## 2026-09-16：逐類負樣本與第二支紙箱鏡頭
+
+instances v4 保留 v3 的 33 張影像、正框、空白區 masks、companions 與 split
+逐檔一致；新增 `0058-Kaohsiung-cam08` 中可分離的左側紙箱。相鄰兩箱合併的
+proposal 被拒絕，只接受單箱遮罩。train 現為 28 張／130 個觀測；紙箱 4 個、
+2 支鏡頭，仍十分稀少。val 保持 6 張／31 個正例，語意資料完全不改。
+
+AI 重新查看 62 張來源 train 總覽及候選原圖，在 5 張既有影像選定 19 個保守
+區域、47 項逐類負向決定：桌墊與遙控器不是手機、包裝印刷的平板不是實體平板、
+桌上型螢幕不是筆電、海報上的 HomePod 不是實體喇叭、木椅不是紙箱。
+這批資料重用已下載且分割合法的影像，不需要新增 GCS 下載。
+
+`class_negative_rects` 必須由綁定影像／companion hash 的 AI 覆核提供，且只准
+用於 train。匯出為逐類 0/1 masks；沒有明確否定的類別仍未知。dataset 對每個
+通道共同做縮放、裁切、翻轉與 padding，再組成 `[C,H,W]`，padding 255 忽略。
+FCOS 在對應 grid point 只加入被否定通道的 focal loss，不改 regression 或
+centerness。相同類別的正框在所有金字塔尺度都有優先權；匯出時也拒絕矛盾覆核。
+既有全類空白 mask 與部分驗證指標契約不變。無此欄位的舊資料維持原行為。
+
+[覆核決定](reviews/studioa_class_negative_decisions_20260916.json) ·
+[逐類負樣本總覽](../runs/studioa_confusers_20260916_v1/class_negative_contact.jpg) ·
+[新增單箱遮罩](../runs/studioa_confusers_20260916_v1/vlm-unknown-0016.jpg)。
+
+重訓沿用固定場景 warmup 設定與原始 scene 初始化，score／NMS／max_det 不調整。
+本輪完整資料變更與前版 7/31 比較；不把資料增補與 loss 修改當成單因素因果實驗。
+未知預測數減少不等於 precision 改善；僅通過隨機初始化基準也不等於超過前版。
+
+### 第二輪結果：不升級模型
+
+`runs/studioa_detector_warmup_20260916_v2/` 使用 commit `c274cb2`，第 25 輪
+觸發 20 輪無改善停止（350 次更新），選定第 5 輪（70 次更新）。
+
+| 同一來源 val、同一評估契約 | 前版 v1 | 本輪 v2 |
+| --- | ---: | ---: |
+| 覆核正例召回 | 7/31（22.58%） | 6/31（19.35%） |
+| person / phone / poster 命中 | 5 / 1 / 1 | 5 / 0 / 1 |
+| 覆核空白區誤報 | 0 | 0 |
+| 未配對、真偽未定預測 | 593 | 594 |
+| 場景輸出 | 與原模型一致 | 與原模型一致 |
+
+因此 **保留 v1 作為較佳實驗基準，v2 不升級**。worker `warmup.accepted=true`
+只代表超過第 0 輪隨機偵測頭，不代表超過 v1；明確決定另見
+[promotion_decision.json](../runs/studioa_detector_warmup_20260916_v2/promotion_decision.json)。
+兩版都未達部署條件；新程式與 AI 資料保留，沒有改寫任何舊 checkpoint。
+
+五張已訓練負樣本影像中的 3,543 個 grid/class 對，score >0.20 的數量由
+1,776 降至 15，說明選定模型在這些訓練區域的反應較低。這是訓練區診斷，
+且兩版選定 epoch 不同，不能當作驗證 precision 改善或單因素因果證明。
+val 六張影像仍全數達 100 框上限，尚未改善其他鏡頭的混淆。
+
+固定門檻下的解碼前手機診斷進一步區分：v2 四個 val 手機中，兩個完全沒有
+IoU ≥0.50 的原始候選框；另外兩個雖有定位候選，phone score 最高只有約
+0.024–0.025，低於原定 0.20。這四個漏檢不能單靠提高 max_det 解決。
+下一步先量化每類正負 loss 比例、限制逐類負項影響，再進行預先設定的比較，
+並補充小物件的定位正例；不利用這組 val 調門檻掩蓋問題。
+
+79 項測試、lint、型別、提交 hook 通過。GPU smoke 對 28 張 train 做 14 次
+更新，2,457,821 個未知分類輸出梯度均為 0。319 個 smoke 輸入、779 個訓練
+輸入及 8 個正式輸出 hashes 核對；best/last 有限、job 身分正確，194 個共享
+tensors 與原模型相同。33 張 scene logits SHA256 完全一致，所有 test 未推論。
+兩個 systemd 工作正常退出。儲存的逐框結果能完整重算官方偵測指標。
+
+[本輪逐圖對照](../runs/studioa_detector_warmup_20260916_v2/review/selected_val_review.jpg) ·
+[解碼前手機診斷](../runs/studioa_detector_warmup_20260916_v2/review/phone_candidate_probe.json) ·
+[完整可追溯結果](reviews/studioa_class_negative_warmup_20260916.json)。
+
+## 2026-09-16：逐類負項正規化的固定比較
+
+先對 28 張 source train、不增強、float32、逐張影像做分類 loss 與 logit 梯度
+分解。v1 模型的 phone 正向梯度 L1 合計約 0.0271，新增逐類負項約 9.631；
+其中一張沒有 phone 正例的桌面影像，610 個 phone 負向點貢獻約 7.168。
+這說明新增負項會受到覆核區面積與每張正例數影響，但不是原先 batch 2 增強
+訓練的梯度重播，也不是模型參數梯度；不能單憑比例就判定退步原因。
+130 個 train 正框均有可分配的特徵點；「無原始框達 IoU 0.50」和「無訓練
+指派點」是不同問題，仍需另外衡量小物件框回歸品質。
+
+只比較一個變因 `class_negative_normalization`：`sum` 保留原行為；
+`positive_budget` 將每個 batch/class 新增負項權重總和限制在
+`max(該類正向指派點數, 1)`，每個負點權重最多 1。已有正項、全類空白負項、
+regression、centerness 與未知通道均不改；缺少該類正例的 batch 仍可提供
+最多一個等效負向點。這限制的是監督權重總量，不是梯度範數。
+
+兩組從同一 scene checkpoint、同 seed、同 instances v4 開始；所有訓練與
+驗證條件沿用 v2。固定 score >0.20／IoU ≥0.50／NMS 0.6／每圖 100 框，
+不得看結果後調門檻。只有候選超過 v1 的 7/31、人物至少 5、手機至少 1、
+海報至少 1、覆核空白誤報不增加且 scene logits 完全一致，才更新實驗基準。
+本次只有這兩組，不依此 val 分數追加參數搜尋。
+
+[執行前比較計畫與診斷](reviews/studioa_negative_balance_plan_20260916.json)。
diff --git a/docs/reviews/studioa_class_negative_decisions_20260916.json b/docs/reviews/studioa_class_negative_decisions_20260916.json
new file mode 100644
index 0000000..f028958
--- /dev/null
+++ b/docs/reviews/studioa_class_negative_decisions_20260916.json
@@ -0,0 +1,1843 @@
+{
+  "reviewer_kind": "ai",
+  "reviewer": "assistant_visual_analysis",
+  "held_out": "Tao-Hsin",
+  "source_manifest_sha256": "0139eaccc00c6cf2fd841c7cfbd3a231c9b744a624b9c8ec95f82bc6e7a7b1ee",
+  "scope": "AI visual review of original source-train images only; per-class absence, other channels unknown. Existing positives and val unchanged; one additional unambiguous cardboard instance.",
+  "frames": [
+    {
+      "frame_id": "0050-Kaohsiung-cam04",
+      "image_sha256": "7b1e1be2cf234ab9d881fef37b49789f4f1b64eff655758b2e674d6819ee6bac",
+      "companion_sha256": "0423e2002e1c61c16d815172dd247c61c3f4a21d982f6c64d88ad962b16656e0",
+      "positives": [
+        {
+          "id": "ai-0005",
+          "entity": "laptop",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0007",
+          "entity": "cardboard_box",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0008",
+          "entity": "cardboard_box",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0010",
+          "entity": "cardboard_box",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0017",
+          "entity": "laptop",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            1650,
+            400,
+            1800,
+            650
+          ],
+          "reason": "Assistant inspected source image: empty floor patch, no visible object from the declared ten detection classes; not inferred from missing labels."
+        }
+      ]
+    },
+    {
+      "frame_id": "0030-Taichung-cam08",
+      "image_sha256": "e95523bd8f92ab2fbb7ed5fa672e25b7cf01d531ed94711a016572351d1b1ea4",
+      "companion_sha256": "e452f0356246f096cc9bcf3c8a16b403c24a87983900fb9fc509a76397a5d720",
+      "positives": [
+        {
+          "id": "ai-0003",
+          "entity": "tablet",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0004",
+          "entity": "tablet",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0005",
+          "entity": "tablet",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            1000,
+            960,
+            1500,
+            1050
+          ],
+          "reason": "Assistant inspected source image: empty floor patch, no visible object from the declared ten detection classes; not inferred from missing labels."
+        }
+      ]
+    },
+    {
+      "frame_id": "0112-Taichung-cam11",
+      "image_sha256": "3353277f188293dba730eefd1f3f42ddbd4e6b50ee87e38d4c954f22d8495355",
+      "companion_sha256": "051477d05602005a21aef67b3af839ec4956f1c3511be724e006eafb1ff30eb9",
+      "positives": [
+        {
+          "id": "ai-0003",
+          "entity": "speaker",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            1100,
+            650,
+            1250,
+            850
+          ],
+          "reason": "Assistant inspected source image: empty floor patch, no visible object from the declared ten detection classes; not inferred from missing labels."
+        }
+      ]
+    },
+    {
+      "frame_id": "0078-Taichung-cam09",
+      "image_sha256": "30918a1a840d0e54a4d88dd8c34bb2907a0b7931f25ce60f7f4e878b5528f6b2",
+      "companion_sha256": "bedfde8148dae9581b2fa7322caa9afb173a9448e394227f68887d27baa4ffc0",
+      "positives": [
+        {
+          "id": "ai-0002",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0003",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            1500,
+            850,
+            1800,
+            1000
+          ],
+          "reason": "Assistant inspected source image: empty floor patch, no visible object from the declared ten detection classes; not inferred from missing labels."
+        }
+      ]
+    },
+    {
+      "frame_id": "0025-Taichung-cam05",
+      "image_sha256": "08b398cb45b48e99021b07d5247539c8138805cfb043c7d86421c6bfd2fe24a4",
+      "companion_sha256": "9dd17564938c9007cd40050ce79aceaebfedbad2b9aa1f1970174557ac462e87",
+      "positives": [
+        {
+          "id": "ai-0009",
+          "entity": "phone",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            1730,
+            760,
+            1810,
+            930
+          ],
+          "reason": "Assistant inspected source image: empty floor patch, no visible object from the declared ten detection classes; not inferred from missing labels."
+        }
+      ]
+    },
+    {
+      "frame_id": "0031-Taichung-cam08",
+      "image_sha256": "ddd3f30a15b4b527c8dc5a025b7e9a5c5cf21561936b3e98fa17bbcd8d411b4c",
+      "companion_sha256": "51097af5eae2df30092a13a14afc6ebaf964454f62d0800d9291c31bbca40b53",
+      "positives": [
+        {
+          "id": "ai-0005",
+          "entity": "phone",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0007",
+          "entity": "phone",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0008",
+          "entity": "phone",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            200,
+            920,
+            450,
+            1030
+          ],
+          "reason": "Assistant inspected source image: empty floor patch, no visible object from the declared ten detection classes; not inferred from missing labels."
+        }
+      ]
+    },
+    {
+      "frame_id": "0008-Kaohsiung-cam07",
+      "image_sha256": "6acd71ce695fa11235c63cf3d83cea62bb39017bfbfd2eba47e198449ad48a28",
+      "companion_sha256": "901ae7a38b80cd42c036e8e6f594dc668096a8646132bb76102ce3ac494ad632",
+      "positives": [
+        {
+          "id": "ai-0028",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0002",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "0002-Kaohsiung-cam04",
+      "image_sha256": "a7339928d2e72363cefb35a8f6e25d8b1784ceaf747f87b815293dfdc53257bb",
+      "companion_sha256": "3dfce18c1f9d25ef717683209d6396964fcaa024ba1477777078ebf0de796436",
+      "positives": [
+        {
+          "id": "ai-0004",
+          "entity": "laptop",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0005",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0013",
+          "entity": "poster",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "0033-Taichung-cam09",
+      "image_sha256": "c5b1f42f5025b4479295913c3c41c9cfa324fbc5b4200bc6a2994efa52220b3d",
+      "companion_sha256": "276fbc3a765a4b33f48c00a1bf2edbb7dd9ed8fc5da0a71df6f3a85f56824137",
+      "positives": [
+        {
+          "id": "ai-0011",
+          "entity": "poster",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "0022-Taichung-cam04",
+      "image_sha256": "ad8f01fb6860896c9bbd1e3ff9bf86c10515013c23f4e8c2de87802fcea2d07b",
+      "companion_sha256": "f094e3b5cf5919a27ee9155700bc98cc57af081e5750f4c8ebe3f96da235691f",
+      "positives": [
+        {
+          "id": "ai-0010",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0013",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0003",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "0082-Taichung-cam11",
+      "image_sha256": "4c85c1f28791ebc81eb3f687cc55f913899381ac90cb6a16ef223aeea997df1d",
+      "companion_sha256": "26b1250e73b68449e4b501b9509702909c3021e44a506f1108e1f8cad4c36492",
+      "positives": [
+        {
+          "id": "ai-0010",
+          "entity": "tablet",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0003",
+          "entity": "speaker",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0002",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "0007-Kaohsiung-cam06",
+      "image_sha256": "8340acd371f970f294282695828b7fdd2d27e08b9529575f7f275290b38b50f8",
+      "companion_sha256": "3a26496740076a5e3a58e36c63fae01393ed39b344a66d0b0fe67440b2f1bc8f",
+      "positives": [
+        {
+          "id": "ai-0005",
+          "entity": "laptop",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0011",
+          "entity": "laptop",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0010",
+          "entity": "tablet",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0007",
+          "entity": "speaker",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0016",
+          "entity": "poster",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0002",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            150,
+            800,
+            400,
+            950
+          ],
+          "reason": "Assistant inspected original context with this rectangle: clear floor without a visible object from the ten detection classes."
+        }
+      ]
+    },
+    {
+      "frame_id": "0053-Kaohsiung-cam05",
+      "image_sha256": "68440d495ad4d190d7b68b046d64707339105fa7c5290b36b2c46abd3a1f4e8a",
+      "companion_sha256": "a9565f0ac21f98f711d0a0ff3c91608767fdc023a91557974fc41903ded832ab",
+      "positives": [
+        {
+          "id": "ai-0004",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0002",
+          "entity": "cardboard_box",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0006",
+          "entity": "cardboard_box",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0005",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            850,
+            500,
+            1250,
+            800
+          ],
+          "reason": "Assistant inspected original context with this rectangle: clear floor without a visible object from the ten detection classes."
+        }
+      ]
+    },
+    {
+      "frame_id": "0016-Taichung-cam01",
+      "image_sha256": "fb3086c1ff8ed11a485226d69bacacb33ea4d534d0708096c80526f348230784",
+      "companion_sha256": "b30aed64d3c57dba492dec341442706b77458556b8afb6991fe1b2f6741d7bee",
+      "positives": [
+        {
+          "id": "ai-0002",
+          "entity": "laptop",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0021",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0004",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0005",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            600,
+            700,
+            700,
+            900
+          ],
+          "reason": "Assistant inspected original context with this rectangle: clear floor without a visible object from the ten detection classes."
+        }
+      ]
+    },
+    {
+      "frame_id": "0076-Taichung-cam07",
+      "image_sha256": "a7e42db0b7ac8ebae37bc90da01b090b7de169b9cf0bf1b06088d049add3f83e",
+      "companion_sha256": "bf42c51b3bb9657e8bd99379eaabba2d1e53e4d4559bae34f2ceae1959b11dd0",
+      "positives": [
+        {
+          "id": "ai-0005",
+          "entity": "laptop",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0006",
+          "entity": "poster",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0001",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0004",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0002",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            900,
+            800,
+            1200,
+            1000
+          ],
+          "reason": "Assistant inspected original context with this rectangle: clear floor without a visible object from the ten detection classes."
+        }
+      ]
+    },
+    {
+      "frame_id": "0034-Taichung-cam10",
+      "image_sha256": "28c5c9e5a09f55a473bda61d0d84de4a8f679ebc3db43f892358a6a56d0c57fd",
+      "companion_sha256": "602aa819d8944d6dadc76bedecfa8b2609641066dd2dc6a046c52cbebf2846d3",
+      "positives": [
+        {
+          "id": "ai-0002",
+          "entity": "laptop",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0007",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0005",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0010",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "0013-Kaohsiung-cam11",
+      "image_sha256": "4c19e17cad632f5551e2280b514e76884e74658d62841354e3ddfcc1c73a6f19",
+      "companion_sha256": "06b6e65f9c04d603749965cb03a5fc7eb6ee2c8419638baeb6cb93bba69f5f0a",
+      "positives": [
+        {
+          "id": "vlm-unknown-0004",
+          "entity": "phone",
+          "reason": "Assistant inspected context and isolated mask: real tethered smartphone on foreground circular display table; excludes phone depicted on TV advertisement."
+        },
+        {
+          "id": "vlm-unknown-0006",
+          "entity": "phone",
+          "reason": "Assistant inspected context and isolated mask: real tethered smartphone on foreground circular display table; excludes phone depicted on TV advertisement."
+        },
+        {
+          "id": "vlm-unknown-0007",
+          "entity": "phone",
+          "reason": "Assistant inspected context and isolated mask: real tethered smartphone on foreground circular display table; excludes phone depicted on TV advertisement."
+        },
+        {
+          "id": "vlm-unknown-0008",
+          "entity": "phone",
+          "reason": "Assistant inspected context and isolated mask: real tethered smartphone on foreground circular display table; excludes phone depicted on TV advertisement."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Kaohsiung-cam04-t330",
+      "image_sha256": "a479c28dc3688ca122efa93ce4288707fffa4eb2031ca942d860fae2397c0add",
+      "companion_sha256": "8cbb99661c64303eae9bd97cf37953079b614235f7f39622385d8b70302abc89",
+      "positives": [
+        {
+          "id": "ai-0005",
+          "entity": "laptop",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 0
+        },
+        {
+          "id": "ai-0011",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 3
+        },
+        {
+          "id": "ai-0019",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 4
+        },
+        {
+          "id": "ai-0009",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 5
+        },
+        {
+          "id": "ai-0006",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 6
+        },
+        {
+          "id": "ai-0031",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 12
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 13
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 14
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Kaohsiung-cam07-t030",
+      "image_sha256": "43a4ec3d39087331280da0a4b868ffc47b98cebb4410a304a2f1d14adf9fdc9d",
+      "companion_sha256": "18d91aac7c1cbb4a2e29a8a6f3ef9df12f4e7ef96d967e94e7175ff509d06e53",
+      "positives": [
+        {
+          "id": "ai-0080",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 19
+        },
+        {
+          "id": "ai-0017",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 21
+        }
+      ],
+      "negative_rects": [],
+      "class_negative_rects": [
+        {
+          "entity": "phone",
+          "xyxy": [
+            290,
+            385,
+            465,
+            480
+          ],
+          "reason": "Printed device/case pictures on hanging retail packaging; no physical device."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            290,
+            385,
+            465,
+            480
+          ],
+          "reason": "Printed device/case pictures on hanging retail packaging; no physical device."
+        },
+        {
+          "entity": "laptop",
+          "xyxy": [
+            290,
+            385,
+            465,
+            480
+          ],
+          "reason": "Printed device/case pictures on hanging retail packaging; no physical device."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            665,
+            374,
+            818,
+            429
+          ],
+          "reason": "Printed tablet/case faces on sealed retail packs."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            665,
+            374,
+            818,
+            429
+          ],
+          "reason": "Printed tablet/case faces on sealed retail packs."
+        },
+        {
+          "entity": "laptop",
+          "xyxy": [
+            665,
+            374,
+            818,
+            429
+          ],
+          "reason": "Printed tablet/case faces on sealed retail packs."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            938,
+            250,
+            1050,
+            325
+          ],
+          "reason": "Device pictures on hanging white case packages."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            938,
+            250,
+            1050,
+            325
+          ],
+          "reason": "Device pictures on hanging white case packages."
+        },
+        {
+          "entity": "laptop",
+          "xyxy": [
+            938,
+            250,
+            1050,
+            325
+          ],
+          "reason": "Device pictures on hanging white case packages."
+        },
+        {
+          "entity": "person",
+          "xyxy": [
+            1375,
+            408,
+            1440,
+            483
+          ],
+          "reason": "Printed advertising figures on retail packaging, not a physical person."
+        }
+      ]
+    },
+    {
+      "frame_id": "gcs-20260910-Kaohsiung-cam08-t030",
+      "image_sha256": "facae65b929ba40c46bf9bb671e1b158c3e9edd392734e6cdea2b4fb69717462",
+      "companion_sha256": "97cf772c23ef3a6058ae6aac87f2cf207b5391d34336713254257157b839e66f",
+      "positives": [
+        {
+          "id": "ai-0023",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 27
+        },
+        {
+          "id": "ai-0025",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 28
+        },
+        {
+          "id": "ai-0018",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 29
+        },
+        {
+          "id": "ai-0042",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 31
+        }
+      ],
+      "negative_rects": [],
+      "class_negative_rects": [
+        {
+          "entity": "laptop",
+          "xyxy": [
+            880,
+            175,
+            960,
+            220
+          ],
+          "reason": "Desktop monitor on stand, distinct from tablet/laptop."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            880,
+            175,
+            960,
+            220
+          ],
+          "reason": "Desktop monitor on stand, distinct from tablet/laptop."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            880,
+            175,
+            960,
+            220
+          ],
+          "reason": "Desktop monitor on stand, distinct from tablet/laptop."
+        },
+        {
+          "entity": "laptop",
+          "xyxy": [
+            1092,
+            183,
+            1174,
+            235
+          ],
+          "reason": "Desktop monitor panel, not a portable device."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            1092,
+            183,
+            1174,
+            235
+          ],
+          "reason": "Desktop monitor panel, not a portable device."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            1092,
+            183,
+            1174,
+            235
+          ],
+          "reason": "Desktop monitor panel, not a portable device."
+        },
+        {
+          "entity": "laptop",
+          "xyxy": [
+            1374,
+            220,
+            1444,
+            260
+          ],
+          "reason": "Desktop monitor panel on wall-side display table."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            1374,
+            220,
+            1444,
+            260
+          ],
+          "reason": "Desktop monitor panel on wall-side display table."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            1374,
+            220,
+            1444,
+            260
+          ],
+          "reason": "Desktop monitor panel on wall-side display table."
+        }
+      ]
+    },
+    {
+      "frame_id": "gcs-20260910-Taichung-cam03-t030",
+      "image_sha256": "e4dc5e1e322ad6b43c802169a816c06fd0b1f6cb8c399e87d4215cd5056e1f82",
+      "companion_sha256": "5ed3466365e1b23c8810968530385e6203616aeb487dae6f0c77e31e10043c6a",
+      "positives": [
+        {
+          "id": "ai-0005",
+          "entity": "laptop",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 32
+        },
+        {
+          "id": "ai-0006",
+          "entity": "laptop",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 33
+        },
+        {
+          "id": "ai-0014",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 35
+        },
+        {
+          "id": "ai-0011",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 36
+        },
+        {
+          "id": "ai-0008",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 37
+        },
+        {
+          "id": "ai-0020",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 40
+        },
+        {
+          "id": "ai-0057",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 41
+        },
+        {
+          "id": "ai-0010",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 42
+        },
+        {
+          "id": "ai-0003",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 43
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Taichung-cam04-t330",
+      "image_sha256": "7daaff94e19ec7f2eb63738d779c358cbbe59406841f5499f8269376befbcad1",
+      "companion_sha256": "503fb3c85237741465e9b3d8c3702535e9651206267e325c69634c81540e62f8",
+      "positives": [
+        {
+          "id": "ai-0054",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 50
+        },
+        {
+          "id": "ai-0057",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 51
+        },
+        {
+          "id": "ai-0050",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 53
+        },
+        {
+          "id": "ai-0009",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 55
+        },
+        {
+          "id": "ai-0028",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 56
+        },
+        {
+          "id": "ai-0016",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 58
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 59
+        },
+        {
+          "id": "ai-0012",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 60
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Taichung-cam05-t030",
+      "image_sha256": "8a7a6e129561f819a598241bbe50107322192a61ca450510ae97c957672686ad",
+      "companion_sha256": "dd4ebfadebb5c9b9a4777c130b124aa1cac9cb3610d20783acf610056b056e4e",
+      "positives": [
+        {
+          "id": "ai-0036",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 64
+        },
+        {
+          "id": "ai-0005",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 66
+        },
+        {
+          "id": "ai-0061",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 68
+        },
+        {
+          "id": "ai-0014",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 69
+        },
+        {
+          "id": "ai-0011",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 70
+        },
+        {
+          "id": "ai-0009",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 71
+        },
+        {
+          "id": "ai-0055",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 73
+        }
+      ],
+      "negative_rects": [],
+      "class_negative_rects": [
+        {
+          "entity": "speaker",
+          "xyxy": [
+            1030,
+            515,
+            1100,
+            600
+          ],
+          "reason": "Printed HomePod advertising panel, not a physical speaker/device; excludes real speaker below left."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            1030,
+            515,
+            1100,
+            600
+          ],
+          "reason": "Printed HomePod advertising panel, not a physical speaker/device; excludes real speaker below left."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            1030,
+            515,
+            1100,
+            600
+          ],
+          "reason": "Printed HomePod advertising panel, not a physical speaker/device; excludes real speaker below left."
+        },
+        {
+          "entity": "laptop",
+          "xyxy": [
+            1095,
+            199,
+            1230,
+            275
+          ],
+          "reason": "Desktop monitor panel on stand, not a laptop or handheld device."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            1095,
+            199,
+            1230,
+            275
+          ],
+          "reason": "Desktop monitor panel on stand, not a laptop or handheld device."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            1095,
+            199,
+            1230,
+            275
+          ],
+          "reason": "Desktop monitor panel on stand, not a laptop or handheld device."
+        },
+        {
+          "entity": "laptop",
+          "xyxy": [
+            949,
+            242,
+            1030,
+            292
+          ],
+          "reason": "Second desktop monitor panel; keyboard and stand establish desktop form."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            949,
+            242,
+            1030,
+            292
+          ],
+          "reason": "Second desktop monitor panel; keyboard and stand establish desktop form."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            949,
+            242,
+            1030,
+            292
+          ],
+          "reason": "Second desktop monitor panel; keyboard and stand establish desktop form."
+        }
+      ]
+    },
+    {
+      "frame_id": "gcs-20260910-Taichung-cam08-t030",
+      "image_sha256": "8f15fd481aeb8f63aa50b25dfa79637535b45defcecae4b5f99b9a5e547b5daa",
+      "companion_sha256": "8a5a47e8f624dde013d5b8bfddfc02b5b6f309196da9a60c94325871d1c28c2d",
+      "positives": [
+        {
+          "id": "ai-0002",
+          "entity": "tablet",
+          "reason": "Assistant corrected phone proposal to tablet: full-size physical iPad with tether and stand on tabletop, compared with the surrounding counter and power sockets; not the adjacent black desk mats.",
+          "review_index": 76
+        }
+      ],
+      "negative_rects": [],
+      "class_negative_rects": [
+        {
+          "entity": "phone",
+          "xyxy": [
+            730,
+            280,
+            1010,
+            500
+          ],
+          "reason": "Black work mat with cable, no device inside reviewed rectangle."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            730,
+            280,
+            1010,
+            500
+          ],
+          "reason": "Black work mat with cable, no device inside reviewed rectangle."
+        },
+        {
+          "entity": "laptop",
+          "xyxy": [
+            730,
+            280,
+            1010,
+            500
+          ],
+          "reason": "Black work mat with cable, no device inside reviewed rectangle."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            1190,
+            280,
+            1450,
+            505
+          ],
+          "reason": "Black work mat is not a tablet or laptop screen."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            1190,
+            280,
+            1450,
+            505
+          ],
+          "reason": "Black work mat is not a tablet or laptop screen."
+        },
+        {
+          "entity": "laptop",
+          "xyxy": [
+            1190,
+            280,
+            1450,
+            505
+          ],
+          "reason": "Black work mat is not a tablet or laptop screen."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            405,
+            575,
+            505,
+            616
+          ],
+          "reason": "Remote control in visible cradle, not a phone."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            413,
+            646,
+            500,
+            687
+          ],
+          "reason": "Remote control with white label, not a phone."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            415,
+            701,
+            507,
+            740
+          ],
+          "reason": "Remote control in cradle, not a phone."
+        }
+      ]
+    },
+    {
+      "frame_id": "gcs-20260910-Taichung-cam09-t030",
+      "image_sha256": "c989bfe1f6a7998834c79588b000dd24f869b96bfc93c14055b9a2b8d2995828",
+      "companion_sha256": "5f7a89f5e2cc3d58f472a954b8fd41992c5615be016b97956911f9bac6aead07",
+      "positives": [
+        {
+          "id": "ai-0066",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 82
+        },
+        {
+          "id": "ai-0028",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 83
+        },
+        {
+          "id": "ai-0026",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 84
+        },
+        {
+          "id": "ai-0002",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 85
+        },
+        {
+          "id": "ai-0005",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 86
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Taichung-cam11-t330",
+      "image_sha256": "272d31978a93c47d1a1aaa04f8581cd0c47135abdd60aab94d0650ef49186d7d",
+      "companion_sha256": "b5d5d582256ff8a0ba2a0f87bedde01d97739e6e1eb9ba8ffdb1e885b9251965",
+      "positives": [
+        {
+          "id": "ai-0017",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 91
+        },
+        {
+          "id": "ai-0030",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 92
+        },
+        {
+          "id": "ai-0011",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 93
+        },
+        {
+          "id": "ai-0028",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 95
+        },
+        {
+          "id": "ai-0002",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 96
+        },
+        {
+          "id": "ai-0006",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 97
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 98
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 99
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260912-Kaohsiung-cam04-t030",
+      "image_sha256": "8f2455851cd3d7d609931dfe0c7681801df689eb20633bb42af56f06837d0730",
+      "companion_sha256": "e57e043fc7a2e8f7a9fce6db43bd41c0758fd9a997cfc803ed49aafdf71a8d43",
+      "positives": [
+        {
+          "id": "ai-0006",
+          "entity": "laptop",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 100
+        },
+        {
+          "id": "ai-0039",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 102
+        },
+        {
+          "id": "ai-0010",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 103
+        },
+        {
+          "id": "ai-0020",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 105
+        },
+        {
+          "id": "ai-0008",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 106
+        },
+        {
+          "id": "ai-0074",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 113
+        },
+        {
+          "id": "ai-0068",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 114
+        },
+        {
+          "id": "ai-0002",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 115
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 116
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260912-Taichung-cam03-t330",
+      "image_sha256": "6f696165a74ce3d833e6fc59a868213f3f261d6b904e2964553abd3107513c52",
+      "companion_sha256": "ffbc822f24b384d131a9e6255a56762405c5c9999e140ebcb54a53edb5ffda85",
+      "positives": [
+        {
+          "id": "ai-0011",
+          "entity": "laptop",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 117
+        },
+        {
+          "id": "ai-0010",
+          "entity": "laptop",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 118
+        },
+        {
+          "id": "ai-0017",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 120
+        },
+        {
+          "id": "ai-0019",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 121
+        },
+        {
+          "id": "ai-0015",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 125
+        },
+        {
+          "id": "ai-0005",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 126
+        },
+        {
+          "id": "ai-0006",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 127
+        },
+        {
+          "id": "ai-0008",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 129
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 130
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260912-Taichung-cam04-t030",
+      "image_sha256": "e528fceb02e0402a1b98840d1e13cd342b8e823ee2bd9f7367353ab9be0bb4cb",
+      "companion_sha256": "d21485f131d60f65e634d033b0a5b58c18c26e78dc4bd3edb89083f882076e8e",
+      "positives": [
+        {
+          "id": "ai-0016",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 139
+        },
+        {
+          "id": "ai-0033",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 140
+        },
+        {
+          "id": "ai-0020",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 141
+        },
+        {
+          "id": "ai-0038",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 142
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 143
+        },
+        {
+          "id": "ai-0003",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 144
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260912-Taichung-cam05-t330",
+      "image_sha256": "f0757ff6f43fadc72a211aee58b1132ac29afe242e35a07ef55c41a2b5bf4209",
+      "companion_sha256": "8b39a80107cfbf5bfd72aed896d486f1579b4a4939f7f72d54696dd98ab1c20e",
+      "positives": [
+        {
+          "id": "ai-0014",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 151
+        },
+        {
+          "id": "ai-0081",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 154
+        },
+        {
+          "id": "ai-0020",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 155
+        },
+        {
+          "id": "ai-0019",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 156
+        },
+        {
+          "id": "ai-0010",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 157
+        },
+        {
+          "id": "ai-0024",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 158
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 159
+        },
+        {
+          "id": "ai-0004",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 160
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260912-Taichung-cam08-t030",
+      "image_sha256": "04aae73f6f1184bb6b1fdd4f2f924243c68f6a423b794ca7d806e4fc03cad2ad",
+      "companion_sha256": "d9d462189ede768927dc8b63bf925e6bd141076a6adbc7ba9c2318505e54d331",
+      "positives": [
+        {
+          "id": "ai-0002",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 161
+        },
+        {
+          "id": "ai-0003",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 162
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 164
+        },
+        {
+          "id": "ai-0005",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 165
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260912-Taichung-cam08-t330",
+      "image_sha256": "c834c7362415b2fa11d32470166b44b0a80519929e258ab8b228bdb5a1a79578",
+      "companion_sha256": "2c5ec8ec5dfa4e68f5d6fa34e4cfd1ff0f3203233738a149bd975c8e6a0787cf",
+      "positives": [
+        {
+          "id": "ai-0007",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 167
+        },
+        {
+          "id": "ai-0001",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 168
+        },
+        {
+          "id": "ai-0002",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 169
+        },
+        {
+          "id": "ai-0005",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 173
+        },
+        {
+          "id": "ai-0004",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 174
+        }
+      ],
+      "negative_rects": [],
+      "class_negative_rects": [
+        {
+          "entity": "phone",
+          "xyxy": [
+            800,
+            270,
+            978,
+            475
+          ],
+          "reason": "Visible work mat and cable, excluding actual phone to its left."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            800,
+            270,
+            978,
+            475
+          ],
+          "reason": "Visible work mat and cable, excluding actual phone to its left."
+        },
+        {
+          "entity": "laptop",
+          "xyxy": [
+            800,
+            270,
+            978,
+            475
+          ],
+          "reason": "Visible work mat and cable, excluding actual phone to its left."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            760,
+            692,
+            935,
+            825
+          ],
+          "reason": "Unoccupied mat interior below the watch."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            760,
+            692,
+            935,
+            825
+          ],
+          "reason": "Unoccupied mat interior below the watch."
+        },
+        {
+          "entity": "laptop",
+          "xyxy": [
+            760,
+            692,
+            935,
+            825
+          ],
+          "reason": "Unoccupied mat interior below the watch."
+        },
+        {
+          "entity": "phone",
+          "xyxy": [
+            1300,
+            790,
+            1480,
+            832
+          ],
+          "reason": "Visible mat interior below umbrella, no device."
+        },
+        {
+          "entity": "tablet",
+          "xyxy": [
+            1300,
+            790,
+            1480,
+            832
+          ],
+          "reason": "Visible mat interior below umbrella, no device."
+        },
+        {
+          "entity": "laptop",
+          "xyxy": [
+            1300,
+            790,
+            1480,
+            832
+          ],
+          "reason": "Visible mat interior below umbrella, no device."
+        },
+        {
+          "entity": "cardboard_box",
+          "xyxy": [
+            1225,
+            85,
+            1390,
+            221
+          ],
+          "reason": "Wood-grain stool seat with cutout handle, not a cardboard carton."
+        }
+      ]
+    },
+    {
+      "frame_id": "gcs-20260912-Taichung-cam11-t330",
+      "image_sha256": "f3360b01d4e4d82e1fc8c552714e287efadaeedd015cc3ab0440ee01603ab19a",
+      "companion_sha256": "adba1791cd74f51dab852defec9e743ed3b20045454f100b59a93cb1617b370c",
+      "positives": [
+        {
+          "id": "ai-0003",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 176
+        },
+        {
+          "id": "ai-0016",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 178
+        },
+        {
+          "id": "ai-0000",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 179
+        },
+        {
+          "id": "ai-0011",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 180
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 181
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "0058-Kaohsiung-cam08",
+      "image_sha256": "58dbfa809ad611de31ccc2bd1e2ba07b00377974702513e0bde6cef7e8cacd06",
+      "companion_sha256": "5b183748ae1474ca2acefe6d2e92f666181b5d0e6af96309e1434745d3a8f066",
+      "positives": [
+        {
+          "id": "vlm-unknown-0016",
+          "entity": "cardboard_box",
+          "reason": "Assistant inspected original frame and isolated mask. Left plain brown closed shipping carton held by a person; mask separates this single carton from the adjacent one. Overrides earlier VLM abstention; combined two-carton proposal vlm-unknown-0013 rejected."
+        }
+      ],
+      "negative_rects": []
+    }
+  ],
+  "deferred_candidates": [
+    {
+      "frame_id": "0008-Kaohsiung-cam07",
+      "id": "ai-0035",
+      "reason": "Multiple packaged objects in one mask"
+    },
+    {
+      "frame_id": "0008-Kaohsiung-cam07",
+      "id": "ai-0036",
+      "reason": "Disconnected mask contaminates bounding box"
+    },
+    {
+      "frame_id": "0002-Kaohsiung-cam04",
+      "id": "ai-0007",
+      "reason": "Ambiguous payment terminal or calculator"
+    },
+    {
+      "frame_id": "0033-Taichung-cam09",
+      "id": "ai-0007",
+      "reason": "Multiple overlapping packages"
+    },
+    {
+      "frame_id": "0033-Taichung-cam09",
+      "id": "ai-0008",
+      "reason": "Multiple overlapping packages"
+    },
+    {
+      "frame_id": "0082-Taichung-cam11",
+      "id": "ai-0001",
+      "reason": "Mask includes floor or disconnected background"
+    },
+    {
+      "frame_id": "0007-Kaohsiung-cam06",
+      "id": "ai-0008",
+      "reason": "Printed price placard, not physical phone"
+    },
+    {
+      "frame_id": "0007-Kaohsiung-cam06",
+      "id": "ai-0018",
+      "reason": "Headphones, not speaker"
+    },
+    {
+      "frame_id": "0053-Kaohsiung-cam05",
+      "id": "ai-0003",
+      "reason": "Phone versus payment terminal unresolved"
+    },
+    {
+      "frame_id": "0016-Taichung-cam01",
+      "id": "ai-0014",
+      "reason": "Disconnected mask contaminates bounding box"
+    },
+    {
+      "frame_id": "0016-Taichung-cam01",
+      "id": "ai-0015",
+      "reason": "Printed placard, not boxed stock"
+    },
+    {
+      "frame_id": "0016-Taichung-cam01",
+      "id": "ai-0016",
+      "reason": "Printed placard, not boxed stock"
+    },
+    {
+      "frame_id": "0034-Taichung-cam10",
+      "id": "ai-0020",
+      "reason": "Laptop versus detachable tablet unresolved"
+    },
+    {
+      "frame_id": "0034-Taichung-cam10",
+      "id": "ai-0014",
+      "reason": "Physical phone versus printed price card unresolved"
+    },
+    {
+      "frame_id": "0013-Kaohsiung-cam11",
+      "id": "ai-0004",
+      "reason": "Phone depicted inside TV advertisement, not a physical scene instance"
+    }
+  ],
+  "parent_review_sha256": "15081f1fe3dddddbff13d505e3d0f4ed7b237cdcda33dace406f7fcf76e88de9",
+  "added_review_sha256": "9f4c858c78cc83a28f40769d859f8f373572f8b8ee1f5058e0297c3df2615b55"
+}
diff --git a/docs/reviews/studioa_class_negative_warmup_20260916.json b/docs/reviews/studioa_class_negative_warmup_20260916.json
new file mode 100644
index 0000000..d4b87ce
--- /dev/null
+++ b/docs/reviews/studioa_class_negative_warmup_20260916.json
@@ -0,0 +1,404 @@
+{
+  "status": "completed AI-reviewed data/loss extension; candidate not promoted",
+  "git_commit": "c274cb2fadb44df8d31e1731ff224a1b42ed842b",
+  "data": {
+    "root": "runs/studioa_partial_instances_20260916_v4",
+    "source_train_frames_visually_screened": 62,
+    "class_negative_frames": 5,
+    "reviewed_image_regions": 19,
+    "class_negative_decisions": 47,
+    "class_negative_pixels": {
+      "phone": 263044,
+      "tablet": 251789,
+      "laptop": 245839,
+      "person": 4875,
+      "speaker": 5950,
+      "cardboard_box": 22440
+    },
+    "train_frames": 28,
+    "train_positive_observations": 130,
+    "train_by_class": {
+      "laptop": 9,
+      "cardboard_box": 4,
+      "tablet": 15,
+      "speaker": 9,
+      "fire_equipment": 8,
+      "phone": 17,
+      "boxed_stock": 10,
+      "poster": 20,
+      "person": 23,
+      "chair": 15
+    },
+    "cardboard_train_cameras": 2,
+    "new_cardboard_frame": "0058-Kaohsiung-cam08",
+    "new_cardboard_source_id": "vlm-unknown-0016",
+    "combined_two_box_proposal_rejected": "vlm-unknown-0013",
+    "new_gcs_download_bytes": 0,
+    "old_images_targets_empty_masks_companions_preserved": 33,
+    "val_frames": 6,
+    "val_positive_observations": 31,
+    "semantic_data_unchanged": true,
+    "all_old_split_assignments_unchanged": true
+  },
+  "validation": {
+    "tests_passed": 79,
+    "ruff_ty_and_commit_hooks_passed": true,
+    "gpu_smoke_frames": 28,
+    "gpu_smoke_steps": 14,
+    "unknown_gradient_channels_checked": 2457821,
+    "unknown_gradient_nonzero": 0,
+    "smoke_frozen_files_verified": 319,
+    "training_frozen_files_verified": 779,
+    "training_outputs_verified": 8,
+    "best_last_shared_tensors_equal_original": true,
+    "checkpoint_tensors_finite": true,
+    "per_image_predictions_reproduce_official_metrics": true,
+    "both_services_exit_status": 0
+  },
+  "training": {
+    "run": "runs/studioa_detector_warmup_20260916_v2",
+    "last_epoch": 25,
+    "best_epoch": 5,
+    "optimizer_steps": 350,
+    "best_optimizer_steps": 70,
+    "early_stop_patience": 20,
+    "scene_invariance": {
+      "frozen_state_tensors": 194,
+      "frozen_state_equal": true,
+      "scene_before": {
+        "frames": 33,
+        "float32_logits_sha256": "72b0bc0688abbc2e2a8cbafb9275219300f36dc87ef620cb34b9ffda4acd8867"
+      },
+      "scene_after": {
+        "frames": 33,
+        "float32_logits_sha256": "72b0bc0688abbc2e2a8cbafb9275219300f36dc87ef620cb34b9ffda4acd8867"
+      },
+      "scene_logits_equal": true,
+      "baseline_recall": 0.0,
+      "selected_recall": 0.1935483870967742,
+      "baseline_empty_fp": 0.0,
+      "selected_empty_fp": 0.0,
+      "accepted": true
+    },
+    "worker_accepted_means": "random-initialization baseline only; see promotion decision for comparison to previous trained detector"
+  },
+  "comparison": {
+    "reference_run": "runs/studioa_detector_warmup_20260916_v1",
+    "candidate_run": "runs/studioa_detector_warmup_20260916_v2",
+    "reference_recall": 0.22580645161290322,
+    "candidate_recall": 0.1935483870967742,
+    "accepted_for_upgrade": false,
+    "reason": "Candidate recall 6/31 is below reference 7/31; zero empty-region FP and unchanged scene do not compensate.",
+    "retained_reference_checkpoint": "runs/studioa_detector_warmup_20260916_v1/model/best.pt",
+    "worker_acceptance_scope": "report.json warmup.accepted is relative to random detector epoch 0 only, not this reference comparison.",
+    "deployment_ready": false
+  },
+  "selected_validation": {
+    "matched": 6,
+    "support": 31,
+    "recall": 0.1935483870967742,
+    "empty_fp": 0,
+    "empty_pixels": 51761,
+    "unknown_predictions": 594,
+    "every_image_hits_100_prediction_cap": true,
+    "per_class": {
+      "laptop": {
+        "support": 5,
+        "matched": 0
+      },
+      "phone": {
+        "support": 4,
+        "matched": 0
+      },
+      "tablet": {
+        "support": 1,
+        "matched": 0
+      },
+      "boxed_stock": {
+        "support": 1,
+        "matched": 0
+      },
+      "cardboard_box": {
+        "support": 2,
+        "matched": 0
+      },
+      "speaker": {
+        "support": 1,
+        "matched": 0
+      },
+      "poster": {
+        "support": 2,
+        "matched": 1
+      },
+      "fire_equipment": {
+        "support": 1,
+        "matched": 0
+      },
+      "chair": {
+        "support": 5,
+        "matched": 0
+      },
+      "person": {
+        "support": 9,
+        "matched": 5
+      }
+    }
+  },
+  "train_negative_response_probe": {
+    "scope": "five training images; pre-NMS grid-class scores in explicitly reviewed negative regions; not held-out accuracy, not object false positive count",
+    "frames": [
+      "gcs-20260910-Kaohsiung-cam07-t030",
+      "gcs-20260910-Kaohsiung-cam08-t030",
+      "gcs-20260910-Taichung-cam05-t030",
+      "gcs-20260910-Taichung-cam08-t030",
+      "gcs-20260912-Taichung-cam08-t330"
+    ],
+    "results": {
+      "v1": {
+        "checkpoint_sha256": "8381ff3ab6d97a86d9c2896568867971991805288c9f847d6f09d3928aa2c3f5",
+        "reviewed_grid_class_pairs": 3543,
+        "score_over_020": 1776,
+        "mean_cls_times_centerness": 0.2296547035626476,
+        "per_class": {
+          "laptop": {
+            "grid_points": 1092,
+            "score_over_020": 357,
+            "mean_score": 0.17110549139742476
+          },
+          "phone": {
+            "grid_points": 1174,
+            "score_over_020": 763,
+            "mean_score": 0.2740117663944384
+          },
+          "tablet": {
+            "grid_points": 1119,
+            "score_over_020": 600,
+            "mean_score": 0.24573806704582116
+          },
+          "cardboard_box": {
+            "grid_points": 106,
+            "score_over_020": 15,
+            "mean_score": 0.11761251121829704
+          },
+          "speaker": {
+            "grid_points": 27,
+            "score_over_020": 20,
+            "mean_score": 0.3011369218842851
+          },
+          "person": {
+            "grid_points": 25,
+            "score_over_020": 21,
+            "mean_score": 0.3820433706045151
+          }
+        }
+      },
+      "v2": {
+        "checkpoint_sha256": "7c1d23f263bc1a1b18936c46e3226dba9c666c4585ac9ad931b5bb5560a9642a",
+        "reviewed_grid_class_pairs": 3543,
+        "score_over_020": 15,
+        "mean_cls_times_centerness": 0.030960009247288024,
+        "per_class": {
+          "laptop": {
+            "grid_points": 1092,
+            "score_over_020": 1,
+            "mean_score": 0.022852739354598955
+          },
+          "phone": {
+            "grid_points": 1174,
+            "score_over_020": 0,
+            "mean_score": 0.02824464823167733
+          },
+          "tablet": {
+            "grid_points": 1119,
+            "score_over_020": 1,
+            "mean_score": 0.03571344299077109
+          },
+          "cardboard_box": {
+            "grid_points": 106,
+            "score_over_020": 0,
+            "mean_score": 0.037260681531339324
+          },
+          "speaker": {
+            "grid_points": 27,
+            "score_over_020": 0,
+            "mean_score": 0.10272153511781383
+          },
+          "person": {
+            "grid_points": 25,
+            "score_over_020": 13,
+            "mean_score": 0.19561791867017747
+          }
+        }
+      }
+    }
+  },
+  "phone_failure_diagnosis": {
+    "scope": "source-val phone geometric diagnostics at existing score >0.20 and IoU >=0.50; no tuning or test inference; not one-to-one recall",
+    "results": {
+      "v1": [
+        {
+          "frame_id": "0013-Kaohsiung-cam11",
+          "box": [
+            540.86669921875,
+            399.26666259765625,
+            566.066650390625,
+            439.8666687011719
+          ],
+          "localized_grid_candidates": 11,
+          "localized_phone_score_over_020": 10,
+          "eligible_before_nms_and_cap": 5,
+          "retained": true,
+          "max_phone_score_on_localized_boxes": 0.40896254777908325,
+          "reason": "retained"
+        },
+        {
+          "frame_id": "0013-Kaohsiung-cam11",
+          "box": [
+            494.6666564941406,
+            403.0,
+            516.5999755859375,
+            445.0
+          ],
+          "localized_grid_candidates": 0,
+          "localized_phone_score_over_020": 0,
+          "eligible_before_nms_and_cap": 0,
+          "retained": false,
+          "max_phone_score_on_localized_boxes": null,
+          "reason": "no_box_localized_at_iou_050"
+        },
+        {
+          "frame_id": "0013-Kaohsiung-cam11",
+          "box": [
+            572.5999755859375,
+            375.0,
+            600.5999755859375,
+            412.79998779296875
+          ],
+          "localized_grid_candidates": 7,
+          "localized_phone_score_over_020": 6,
+          "eligible_before_nms_and_cap": 4,
+          "retained": false,
+          "max_phone_score_on_localized_boxes": 0.3057223856449127,
+          "reason": "removed_by_nms_or_top100"
+        },
+        {
+          "frame_id": "0013-Kaohsiung-cam11",
+          "box": [
+            457.3333435058594,
+            385.73333740234375,
+            482.0666809082031,
+            425.8666687011719
+          ],
+          "localized_grid_candidates": 0,
+          "localized_phone_score_over_020": 0,
+          "eligible_before_nms_and_cap": 0,
+          "retained": false,
+          "max_phone_score_on_localized_boxes": null,
+          "reason": "no_box_localized_at_iou_050"
+        }
+      ],
+      "v2": [
+        {
+          "frame_id": "0013-Kaohsiung-cam11",
+          "box": [
+            540.86669921875,
+            399.26666259765625,
+            566.066650390625,
+            439.8666687011719
+          ],
+          "localized_grid_candidates": 1,
+          "localized_phone_score_over_020": 0,
+          "eligible_before_nms_and_cap": 0,
+          "retained": false,
+          "max_phone_score_on_localized_boxes": 0.02521689049899578,
+          "reason": "phone_score_below_threshold"
+        },
+        {
+          "frame_id": "0013-Kaohsiung-cam11",
+          "box": [
+            494.6666564941406,
+            403.0,
+            516.5999755859375,
+            445.0
+          ],
+          "localized_grid_candidates": 0,
+          "localized_phone_score_over_020": 0,
+          "eligible_before_nms_and_cap": 0,
+          "retained": false,
+          "max_phone_score_on_localized_boxes": null,
+          "reason": "no_box_localized_at_iou_050"
+        },
+        {
+          "frame_id": "0013-Kaohsiung-cam11",
+          "box": [
+            572.5999755859375,
+            375.0,
+            600.5999755859375,
+            412.79998779296875
+          ],
+          "localized_grid_candidates": 5,
+          "localized_phone_score_over_020": 0,
+          "eligible_before_nms_and_cap": 0,
+          "retained": false,
+          "max_phone_score_on_localized_boxes": 0.023917578160762787,
+          "reason": "phone_score_below_threshold"
+        },
+        {
+          "frame_id": "0013-Kaohsiung-cam11",
+          "box": [
+            457.3333435058594,
+            385.73333740234375,
+            482.0666809082031,
+            425.8666687011719
+          ],
+          "localized_grid_candidates": 0,
+          "localized_phone_score_over_020": 0,
+          "eligible_before_nms_and_cap": 0,
+          "retained": false,
+          "max_phone_score_on_localized_boxes": null,
+          "reason": "no_box_localized_at_iou_050"
+        }
+      ]
+    }
+  },
+  "limits": [
+    "All labels and reviews are AI-produced; no independent accuracy claim.",
+    "Negative-region probe is on training images and different selected epochs, not evidence of held-out precision or isolated causal improvement.",
+    "V2 phone: two reviewed boxes have no raw candidate at IoU >=0.50; two have localized candidates but max phone score approximately 0.024-0.025, below unchanged 0.20 threshold. Raising max_det alone cannot repair those four misses.",
+    "Same validation set was used for checkpoint selection; it is not a fresh independent test."
+  ],
+  "next_step": "Measure per-class positive/negative loss balance and bound class-negative influence before another predeclared comparison; improve small-object positive localization support. Preserve fixed thresholds and original test isolation.",
+  "test_evaluated": false,
+  "independent_accuracy": false,
+  "deployment_ready": false,
+  "artifacts": {
+    "runs/studioa_detector_warmup_20260916_v2/job.json": "c042b7471bbc3f608725404948f53f670505913b7c816ad17aba3abd5d12f07e",
+    "runs/studioa_detector_warmup_20260916_v2/report.json": "bda6ba29c40ba2f71fc3228755aae58ebe3cb0bb7d315d839d83befde6adc228",
+    "runs/studioa_detector_warmup_20260916_v2/promotion_decision.json": "1224a16aa6f37c4f47f74abdec814554e7deac9bd20f49cdda3982e8cc0e3eba",
+    "runs/studioa_detector_warmup_20260916_v2/status.json": "8d5d32f512ef2b1d787f6cb713204c17daf773748812261a3dd736258adb33ef",
+    "runs/studioa_class_negative_smoke_20260916_v1/job.json": "ba72688fec6f61819486fafb4cb09f81feebb7d97dae5386bcad2180dae4947d",
+    "runs/studioa_class_negative_smoke_20260916_v1/report.json": "ba4e6894d3e0f767434549214fc9229b14330ce889e26ff283144babc12d58a8",
+    "runs/studioa_class_negative_smoke_20260916_v1/status.json": "3214ab26ed45b60457e62581fb64b8ee09bfc5f5b85a9ea6c6bbbca928062729",
+    "runs/studioa_partial_instances_20260916_v4/manifest.json": "e12339ea28db97a8a59c839ad095bf58d961f92af46b0758d6343ebc83475fc1",
+    "runs/studioa_partial_instances_20260916_v4/report.json": "2c17a48f9772030eee957fd24027830a711b768eaa64cf964408d646c6529dd6",
+    "docs/reviews/studioa_class_negative_decisions_20260916.json": "c647a5e13f533363af44465da22b424a8441c6b89e385703acac402b51b1e932",
+    "runs/studioa_confusers_20260916_v1/class_negative_contact.jpg": "9c5657ca1b5298a269336646fa67634e2694d64e847d881cb6336e50ce1ae55e",
+    "runs/studioa_confusers_20260916_v1/vlm-unknown-0016.jpg": "4c596e0c887859948e47fb2cc55c95fd1e64bf743364081b965f4535ba7198b1",
+    "runs/studioa_detector_warmup_20260916_v2/model/best.pt": "7c1d23f263bc1a1b18936c46e3226dba9c666c4585ac9ad931b5bb5560a9642a",
+    "runs/studioa_detector_warmup_20260916_v2/model/last.pt": "cc9213a6996d915bed67d1bdccf8dfb1b956719de150d509c427b8ecc10cef39",
+    "runs/studioa_detector_warmup_20260916_v2/validation.json": "6950b3e1e5ac11d506131bc7e33b6f40bd6e14051ce705dd4ce1f40ca58fdf97",
+    "runs/studioa_detector_warmup_20260916_v2/baseline_val.json": "445010c860040e87a677a984948f275e120f3f0f9bd6c85c070820efe7831962",
+    "runs/studioa_detector_warmup_20260916_v2/model/metrics.jsonl": "e9c2012e73bb114c4faea57117851039b26c649a05976a079aa64da60bc220ae",
+    "runs/studioa_detector_warmup_20260916_v2/validation_scene_preview.jpg": "e1798a2a38d0cab2eb1b06b75d18d29a8d66ad69cecb3082cc7c9bad25b78b57",
+    "runs/studioa_detector_warmup_20260916_v2/scene_before.json": "d7c9e18e014bb4e47b0a6acbae3e56ead9448a44e7dffe79097eeb4221d527a4",
+    "runs/studioa_detector_warmup_20260916_v2/warmup_evidence.json": "2e90f6e5b4e97f471fd2a99f1de143153801a87a1262024410fefeec30002c3c",
+    "runs/studioa_detector_warmup_20260916_v2/review/probe_phone_candidates.py": "3047e7c151d125504dd6c49a0cc8ac28ff80b043572fda8bf73ba58c01284161",
+    "runs/studioa_detector_warmup_20260916_v2/review/inspect_selected.py": "eae0fb3f94338de83abfe586b430b753a028abed44564355a95a82bdd5462e91",
+    "runs/studioa_detector_warmup_20260916_v2/review/probe_train_negatives.py": "c8de43945061d229f7ec3e30c3f10a8ee988014418d4c1f525b69302b496181d",
+    "runs/studioa_detector_warmup_20260916_v2/review/train_negative_probe.json": "a05f1b19ad42e34344ad341d05764a040f987ce13ecb3aaef931270bff260dc3",
+    "runs/studioa_detector_warmup_20260916_v2/review/phone_candidate_probe.json": "632416d4446caa8d1314b532dc2c49fdbf3d10810caa5d5f98df714ad904b454",
+    "runs/studioa_detector_warmup_20260916_v2/review/retained_prediction_error_analysis.json": "07e66d8c4e6431a144b7e90565d56da092fcc589216d8ec4ed2e85c4d36edd20",
+    "runs/studioa_detector_warmup_20260916_v2/review/selected_diagnostics.json": "cd3cf1c69d5de3e29339691a7e80a567a886a29a337da69b71d140c11b92c739",
+    "runs/studioa_detector_warmup_20260916_v2/review/selected_val_review.jpg": "14c2e2db2da562358270df87d313ed6f08a33c1f1ec9189c5796ce5bfcc48b17"
+  }
+}
diff --git a/docs/reviews/studioa_detector_warmup_20260916.json b/docs/reviews/studioa_detector_warmup_20260916.json
new file mode 100644
index 0000000..175db39
--- /dev/null
+++ b/docs/reviews/studioa_detector_warmup_20260916.json
@@ -0,0 +1,158 @@
+{
+  "status": "completed; frozen-scene detector warmup accepted; detector not deployment-ready",
+  "run": "runs/studioa_detector_warmup_20260916_v1",
+  "git_commit": "562532d9c72ead3e7c78ce162c7b66c9666701ae",
+  "job_sha256": "be9412eb64d286d129ef79c794c6da36b6df7d2b163851f12bc8f686ff66cbe6",
+  "training": {
+    "train_frames": 27,
+    "train_observations": 129,
+    "train_cameras": 9,
+    "scene_train_frames_used": 0,
+    "max_epochs": 60,
+    "last_epoch": 40,
+    "best_epoch": 20,
+    "optimizer_steps": 520,
+    "best_optimizer_steps": 260,
+    "stop_reason": "20 validations without strict recall improvement",
+    "systemd_unit": "studioa-detector-warmup-20260916-v1.service",
+    "service_result": "success",
+    "exit_status": 0
+  },
+  "scene_invariance": {
+    "frozen_state_tensors": 194,
+    "frozen_state_equal": true,
+    "scene_before": {
+      "frames": 33,
+      "float32_logits_sha256": "72b0bc0688abbc2e2a8cbafb9275219300f36dc87ef620cb34b9ffda4acd8867"
+    },
+    "scene_after": {
+      "frames": 33,
+      "float32_logits_sha256": "72b0bc0688abbc2e2a8cbafb9275219300f36dc87ef620cb34b9ffda4acd8867"
+    },
+    "scene_logits_equal": true,
+    "baseline_recall": 0.0,
+    "selected_recall": 0.22580645161290322,
+    "baseline_empty_fp": 0.0,
+    "selected_empty_fp": 0.0,
+    "accepted": true
+  },
+  "evaluation": {
+    "split": "source_val_only",
+    "frames": 6,
+    "reviewed_positives": 31,
+    "matched": 7,
+    "recall": 0.22580645161290322,
+    "reviewed_empty_false_alarms": 0,
+    "reviewed_empty_input_pixels": 51761,
+    "unmatched_truth_unknown_predictions": 593,
+    "each_frame_reached_max_det_100": true,
+    "per_class": {
+      "laptop": {
+        "train_observations": 9,
+        "train_camera_count": 2,
+        "val_support": 5,
+        "val_matched": 0
+      },
+      "phone": {
+        "train_observations": 17,
+        "train_camera_count": 4,
+        "val_support": 4,
+        "val_matched": 1
+      },
+      "tablet": {
+        "train_observations": 15,
+        "train_camera_count": 5,
+        "val_support": 1,
+        "val_matched": 0
+      },
+      "boxed_stock": {
+        "train_observations": 10,
+        "train_camera_count": 5,
+        "val_support": 1,
+        "val_matched": 0
+      },
+      "cardboard_box": {
+        "train_observations": 3,
+        "train_camera_count": 1,
+        "val_support": 2,
+        "val_matched": 0
+      },
+      "speaker": {
+        "train_observations": 9,
+        "train_camera_count": 3,
+        "val_support": 1,
+        "val_matched": 0
+      },
+      "poster": {
+        "train_observations": 20,
+        "train_camera_count": 7,
+        "val_support": 2,
+        "val_matched": 1
+      },
+      "fire_equipment": {
+        "train_observations": 8,
+        "train_camera_count": 2,
+        "val_support": 1,
+        "val_matched": 0
+      },
+      "chair": {
+        "train_observations": 15,
+        "train_camera_count": 5,
+        "val_support": 5,
+        "val_matched": 0
+      },
+      "person": {
+        "train_observations": 23,
+        "train_camera_count": 7,
+        "val_support": 9,
+        "val_matched": 5
+      }
+    }
+  },
+  "verification": {
+    "frozen_input_hashes_verified": 757,
+    "output_hashes_verified": 8,
+    "best_and_last_shared_state_equal_original": true,
+    "checkpoint_tensors_finite": true,
+    "scene_miou_all_40_epochs": 0.2506537344807518,
+    "saved_per_image_predictions_reproduce_all_detection_metrics": true,
+    "main_test_batch_passed": 171,
+    "data_contract_followup_test_batch_passed": 20,
+    "test_batches_overlap": true,
+    "changed_file_ruff_ty_and_commit_hooks_passed": true,
+    "test_evaluated": false
+  },
+  "interpretation": [
+    "Acceptance applies only to preserving scene outputs while improving partial validation recall, not general detection quality.",
+    "All six images reach the fixed 100-prediction cap; 593 unmatched predictions remain unknown under partial supervision and cannot define precision.",
+    "Seven classes have zero matched validation positives.",
+    "Cardboard has only three training observations, all in one frame from one camera.",
+    "Compare scene mIoU to this run baseline: deterministic TF32-disabled settings differ from earlier runs.",
+    "Data quantity and training isolation changed together relative to the joint pilot; no single-factor causal attribution."
+  ],
+  "next_step": [
+    "AI-review source-train confusers and add explicit per-class negative supervision; do not mark real objects as all-class background.",
+    "Seek cardboard in other allowed source-train cameras; preserve val/test assignments.",
+    "Keep score/NMS/max_det fixed; do not tune thresholds on this val split."
+  ],
+  "independent_accuracy": false,
+  "deployment_ready": false,
+  "artifacts": {
+    "runs/studioa_detector_warmup_20260916_v1/job.json": "be9412eb64d286d129ef79c794c6da36b6df7d2b163851f12bc8f686ff66cbe6",
+    "runs/studioa_detector_warmup_20260916_v1/report.json": "e4af368a4221673df7cf6440a52d9eae32bfcabba3f7bac52bb975a17a385921",
+    "runs/studioa_detector_warmup_20260916_v1/status.json": "8d448ee2341364ce4a3fe174d5dff529c8be0b2b29e3f458c12a13951544f614",
+    "runs/studioa_detector_warmup_20260916_v1/service.log": "a717ed6aa1976c5afabb27a61d6fa7e7f9e410c7bba8304d3572a88db9bd73b6",
+    "runs/studioa_detector_warmup_20260916_v1/config.json": "55080088b611308124460203c1a869c736eb0a6d332cc8977a9e084bc778689a",
+    "runs/studioa_detector_warmup_20260916_v1/review/inspect_selected.py": "ceab1f30d937928152d46b946d84a444231f29598f60d487988836d762fe0092",
+    "runs/studioa_detector_warmup_20260916_v1/review/selected_val_review.jpg": "8ce9cee0423fc81d708947c300d75c0e0d55c9e8103ebab0b3165bdbc1d5cd8c",
+    "runs/studioa_detector_warmup_20260916_v1/review/selected_diagnostics.json": "f51ee51464278bdd6193b71390018c7f98d6fa7551cdbffba7583c5f6444a9fa",
+    "runs/studioa_detector_warmup_20260916_v1/model/best.pt": "8381ff3ab6d97a86d9c2896568867971991805288c9f847d6f09d3928aa2c3f5",
+    "runs/studioa_detector_warmup_20260916_v1/model/last.pt": "c9337234b942d950591c9f2ab49b35c0976bed05294ab432918b033ed54d68a2",
+    "runs/studioa_detector_warmup_20260916_v1/validation.json": "a9e3b29410192b1f6bf8d49f583bd8b641f583c09f5b5558767cfd4a9171fa5a",
+    "runs/studioa_detector_warmup_20260916_v1/baseline_val.json": "445010c860040e87a677a984948f275e120f3f0f9bd6c85c070820efe7831962",
+    "runs/studioa_detector_warmup_20260916_v1/model/metrics.jsonl": "00b7e9492f55f7bd5ede40f3fcbba1a114d13e401c0d324126f234148026e846",
+    "runs/studioa_detector_warmup_20260916_v1/validation_scene_preview.jpg": "e1798a2a38d0cab2eb1b06b75d18d29a8d66ad69cecb3082cc7c9bad25b78b57",
+    "runs/studioa_detector_warmup_20260916_v1/scene_before.json": "d7c9e18e014bb4e47b0a6acbae3e56ead9448a44e7dffe79097eeb4221d527a4",
+    "runs/studioa_detector_warmup_20260916_v1/warmup_evidence.json": "93b264b01dfe739bd1ca301060719fb7dbb3dd89d1a33cc27c9a1a8975064641"
+  }
+}
diff --git a/docs/reviews/studioa_gcs_expansion_20260916.json b/docs/reviews/studioa_gcs_expansion_20260916.json
new file mode 100644
index 0000000..4d64d33
--- /dev/null
+++ b/docs/reviews/studioa_gcs_expansion_20260916.json
@@ -0,0 +1,119 @@
+{
+  "status": "completed AI-reviewed data expansion; not a trained model upgrade",
+  "intake": {
+    "bucket": "gs://studioa",
+    "source_stores": [
+      "Kaohsiung",
+      "Taichung"
+    ],
+    "source_train_cameras": 9,
+    "date_slots": [
+      "2026-09-10 12:00",
+      "2026-09-12 18:50"
+    ],
+    "filename_time_is_approximate": true,
+    "clips": 18,
+    "download_bytes": 453678490,
+    "download_budget_bytes": 1000000000,
+    "gcs_generation_and_md5_verified": true,
+    "candidate_frames": 36,
+    "selected_frames": 16,
+    "excluded_redundant_frames": 20,
+    "excluded_exact_pixel_duplicates": 0
+  },
+  "annotation": {
+    "teacher": "facebook/sam3",
+    "source_ai_entities": 504,
+    "source_unresolved_candidates": 553,
+    "candidate_masks_visually_inspected": 183,
+    "new_accepted_instance_observations": 98,
+    "manually_class_corrected_by_assistant": 1,
+    "unselected_proposals": "unknown; no negative/background claim",
+    "human_reviewed": false,
+    "unique_physical_object_count": "not measured; same physical object may recur across dates"
+  },
+  "instance_training_package": "runs/studioa_partial_instances_20260916_v3",
+  "semantic_training_package": "runs/studioa_gcs_training_extension_20260916_v1/semantic",
+  "counts": {
+    "train": {
+      "frames": 27,
+      "cameras": 9,
+      "instances": 129,
+      "by_class": {
+        "laptop": 9,
+        "cardboard_box": 3,
+        "tablet": 15,
+        "speaker": 9,
+        "fire_equipment": 8,
+        "phone": 17,
+        "boxed_stock": 10,
+        "poster": 20,
+        "person": 23,
+        "chair": 15
+      },
+      "negative_pixels": 198600
+    },
+    "val": {
+      "frames": 6,
+      "cameras": 6,
+      "instances": 31,
+      "by_class": {
+        "laptop": 5,
+        "tablet": 1,
+        "speaker": 1,
+        "poster": 2,
+        "person": 9,
+        "boxed_stock": 1,
+        "cardboard_box": 2,
+        "fire_equipment": 1,
+        "chair": 5,
+        "phone": 4
+      },
+      "negative_pixels": 237500
+    }
+  },
+  "preservation": {
+    "old_semantic_masks_byte_identical": 121,
+    "old_instance_images_targets_negative_masks_byte_identical": 17,
+    "all_prior_fold_assignments_unchanged": true,
+    "test_evaluated": false
+  },
+  "validation": {
+    "related_tests_passed": 59,
+    "intake_output_hashes_verified": 57,
+    "ai_report_output_hashes_verified": 17,
+    "smoke_frozen_files_verified": 297,
+    "gpu_steps": 14,
+    "gpu_frames": 27,
+    "unknown_logit_channels_checked": 2366580,
+    "unknown_nonzero_gradients": 0,
+    "classifier_update_norm": 0.014031676575541496,
+    "checkpoint_written": false,
+    "services_final_result": "success; exit 0",
+    "initial_intake_failure": "ffprobe missing from systemd PATH; resumed same pinned job with explicit PATH; logs retained"
+  },
+  "remaining_gaps": [
+    "Cardboard box still only 3 training observations; no new unambiguous boxes accepted.",
+    "Object-specific hard negatives have not been added; reviewed-empty regions remain the original floor patches.",
+    "Only 31 reviewed validation positives; this is not sufficient for deployment/generalization claims.",
+    "Same-camera repeated physical objects do not establish independent sample diversity."
+  ],
+  "next_step": "Freeze shared scene features and normalization for detector warm-up; add reviewed class-specific confusers and cardboard boxes; do not repeat the failed unrestricted joint recipe.",
+  "independent_accuracy": false,
+  "deployment_ready": false,
+  "artifacts": {
+    "runs/studioa_gcs_intake_20260916_v1/plan.json": "1b4d177c7b0b01f142395cd0cf325deee222e8224ee89876650cd128598033e9",
+    "runs/studioa_gcs_intake_20260916_v1/manifest.json": "a59a572a41e29dc42e6b8113c00f3a8a581abee341993aff964fe3a82fa59d0d",
+    "runs/studioa_gcs_intake_20260916_v1/report.json": "f925f7e6c20a89539432d18ff9d54aa093c425479b9398d696bab844fdf49f0a",
+    "runs/studioa_gcs_ai_20260916_v1/job.json": "2c48f8697a173d8bc4460d2960806d6e620ee98c5871a19ef4eff7006984b13d",
+    "runs/studioa_gcs_ai_20260916_v1/report.json": "68f3b1c535b01cb1c1dc7524539cd99cee2d988e2c3c7b12253dd98b7bad9d62",
+    "runs/studioa_gcs_training_extension_20260916_v1/extension_report.json": "55b66c5e6414f6001c366b2003a6ca132886a8fd4af91d19d223d48cf5519438",
+    "runs/studioa_gcs_training_extension_20260916_v1/semantic/manifest.json": "0139eaccc00c6cf2fd841c7cfbd3a231c9b744a624b9c8ec95f82bc6e7a7b1ee",
+    "runs/studioa_partial_instances_20260916_v3/manifest.json": "f74f99103a841bd4553d41cc8ce85e1257486de0cb065a108dbc077dfb643367",
+    "runs/studioa_partial_instances_20260916_v3/report.json": "43892b57c74c50a441952949c74bfde145a72fd5c4922743feea6ec199d342f7",
+    "runs/studioa_gcs_instance_smoke_20260916_v1/job.json": "d336b04d781e597f6a1edbbe0bf14fd0e75a11679e2691bb4eb7e4154aa69dfa",
+    "runs/studioa_gcs_instance_smoke_20260916_v1/report.json": "757be212a1abd9c1a8301e5fe02ba7f3eadfcfc337fa32de97043b8d1d411c55",
+    "runs/studioa_gcs_expansion_20260916_v1/reviewed_contact.jpg": "c9479bae04651c97b281d4347506f828d09e9b512813c0eeef23f011779c928c",
+    "runs/studioa_gcs_expansion_20260916_v1/reviewed_gallery.html": "215ff833266288f2e60c1131fdb3f77a6ada9e97d068d0e20337dc20e55cadb7"
+  }
+}
diff --git a/docs/reviews/studioa_gcs_merged_instance_decisions_20260916.json b/docs/reviews/studioa_gcs_merged_instance_decisions_20260916.json
new file mode 100644
index 0000000..df9b402
--- /dev/null
+++ b/docs/reviews/studioa_gcs_merged_instance_decisions_20260916.json
@@ -0,0 +1,1350 @@
+{
+  "reviewer_kind": "ai",
+  "reviewer": "assistant_visual_analysis",
+  "held_out": "Tao-Hsin",
+  "source_manifest_sha256": "0139eaccc00c6cf2fd841c7cfbd3a231c9b744a624b9c8ec95f82bc6e7a7b1ee",
+  "scope": "Existing reviewed train/val subset plus 16 AI-reviewed GCS source-train frames; unknown pixels are not negative. Original val positives/boxes/negative masks preserved.",
+  "frames": [
+    {
+      "frame_id": "0050-Kaohsiung-cam04",
+      "image_sha256": "7b1e1be2cf234ab9d881fef37b49789f4f1b64eff655758b2e674d6819ee6bac",
+      "companion_sha256": "0423e2002e1c61c16d815172dd247c61c3f4a21d982f6c64d88ad962b16656e0",
+      "positives": [
+        {
+          "id": "ai-0005",
+          "entity": "laptop",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0007",
+          "entity": "cardboard_box",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0008",
+          "entity": "cardboard_box",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0010",
+          "entity": "cardboard_box",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0017",
+          "entity": "laptop",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            1650,
+            400,
+            1800,
+            650
+          ],
+          "reason": "Assistant inspected source image: empty floor patch, no visible object from the declared ten detection classes; not inferred from missing labels."
+        }
+      ]
+    },
+    {
+      "frame_id": "0030-Taichung-cam08",
+      "image_sha256": "e95523bd8f92ab2fbb7ed5fa672e25b7cf01d531ed94711a016572351d1b1ea4",
+      "companion_sha256": "e452f0356246f096cc9bcf3c8a16b403c24a87983900fb9fc509a76397a5d720",
+      "positives": [
+        {
+          "id": "ai-0003",
+          "entity": "tablet",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0004",
+          "entity": "tablet",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0005",
+          "entity": "tablet",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            1000,
+            960,
+            1500,
+            1050
+          ],
+          "reason": "Assistant inspected source image: empty floor patch, no visible object from the declared ten detection classes; not inferred from missing labels."
+        }
+      ]
+    },
+    {
+      "frame_id": "0112-Taichung-cam11",
+      "image_sha256": "3353277f188293dba730eefd1f3f42ddbd4e6b50ee87e38d4c954f22d8495355",
+      "companion_sha256": "051477d05602005a21aef67b3af839ec4956f1c3511be724e006eafb1ff30eb9",
+      "positives": [
+        {
+          "id": "ai-0003",
+          "entity": "speaker",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            1100,
+            650,
+            1250,
+            850
+          ],
+          "reason": "Assistant inspected source image: empty floor patch, no visible object from the declared ten detection classes; not inferred from missing labels."
+        }
+      ]
+    },
+    {
+      "frame_id": "0078-Taichung-cam09",
+      "image_sha256": "30918a1a840d0e54a4d88dd8c34bb2907a0b7931f25ce60f7f4e878b5528f6b2",
+      "companion_sha256": "bedfde8148dae9581b2fa7322caa9afb173a9448e394227f68887d27baa4ffc0",
+      "positives": [
+        {
+          "id": "ai-0002",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0003",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            1500,
+            850,
+            1800,
+            1000
+          ],
+          "reason": "Assistant inspected source image: empty floor patch, no visible object from the declared ten detection classes; not inferred from missing labels."
+        }
+      ]
+    },
+    {
+      "frame_id": "0025-Taichung-cam05",
+      "image_sha256": "08b398cb45b48e99021b07d5247539c8138805cfb043c7d86421c6bfd2fe24a4",
+      "companion_sha256": "9dd17564938c9007cd40050ce79aceaebfedbad2b9aa1f1970174557ac462e87",
+      "positives": [
+        {
+          "id": "ai-0009",
+          "entity": "phone",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            1730,
+            760,
+            1810,
+            930
+          ],
+          "reason": "Assistant inspected source image: empty floor patch, no visible object from the declared ten detection classes; not inferred from missing labels."
+        }
+      ]
+    },
+    {
+      "frame_id": "0031-Taichung-cam08",
+      "image_sha256": "ddd3f30a15b4b527c8dc5a025b7e9a5c5cf21561936b3e98fa17bbcd8d411b4c",
+      "companion_sha256": "51097af5eae2df30092a13a14afc6ebaf964454f62d0800d9291c31bbca40b53",
+      "positives": [
+        {
+          "id": "ai-0005",
+          "entity": "phone",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0007",
+          "entity": "phone",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        },
+        {
+          "id": "ai-0008",
+          "entity": "phone",
+          "reason": "Assistant inspected original scene and isolated candidate mask; visible single object supports this class. Partial review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            200,
+            920,
+            450,
+            1030
+          ],
+          "reason": "Assistant inspected source image: empty floor patch, no visible object from the declared ten detection classes; not inferred from missing labels."
+        }
+      ]
+    },
+    {
+      "frame_id": "0008-Kaohsiung-cam07",
+      "image_sha256": "6acd71ce695fa11235c63cf3d83cea62bb39017bfbfd2eba47e198449ad48a28",
+      "companion_sha256": "901ae7a38b80cd42c036e8e6f594dc668096a8646132bb76102ce3ac494ad632",
+      "positives": [
+        {
+          "id": "ai-0028",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0002",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "0002-Kaohsiung-cam04",
+      "image_sha256": "a7339928d2e72363cefb35a8f6e25d8b1784ceaf747f87b815293dfdc53257bb",
+      "companion_sha256": "3dfce18c1f9d25ef717683209d6396964fcaa024ba1477777078ebf0de796436",
+      "positives": [
+        {
+          "id": "ai-0004",
+          "entity": "laptop",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0005",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0013",
+          "entity": "poster",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "0033-Taichung-cam09",
+      "image_sha256": "c5b1f42f5025b4479295913c3c41c9cfa324fbc5b4200bc6a2994efa52220b3d",
+      "companion_sha256": "276fbc3a765a4b33f48c00a1bf2edbb7dd9ed8fc5da0a71df6f3a85f56824137",
+      "positives": [
+        {
+          "id": "ai-0011",
+          "entity": "poster",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "0022-Taichung-cam04",
+      "image_sha256": "ad8f01fb6860896c9bbd1e3ff9bf86c10515013c23f4e8c2de87802fcea2d07b",
+      "companion_sha256": "f094e3b5cf5919a27ee9155700bc98cc57af081e5750f4c8ebe3f96da235691f",
+      "positives": [
+        {
+          "id": "ai-0010",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0013",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0003",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "0082-Taichung-cam11",
+      "image_sha256": "4c85c1f28791ebc81eb3f687cc55f913899381ac90cb6a16ef223aeea997df1d",
+      "companion_sha256": "26b1250e73b68449e4b501b9509702909c3021e44a506f1108e1f8cad4c36492",
+      "positives": [
+        {
+          "id": "ai-0010",
+          "entity": "tablet",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0003",
+          "entity": "speaker",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0002",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "0007-Kaohsiung-cam06",
+      "image_sha256": "8340acd371f970f294282695828b7fdd2d27e08b9529575f7f275290b38b50f8",
+      "companion_sha256": "3a26496740076a5e3a58e36c63fae01393ed39b344a66d0b0fe67440b2f1bc8f",
+      "positives": [
+        {
+          "id": "ai-0005",
+          "entity": "laptop",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0011",
+          "entity": "laptop",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0010",
+          "entity": "tablet",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0007",
+          "entity": "speaker",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0016",
+          "entity": "poster",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0002",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            150,
+            800,
+            400,
+            950
+          ],
+          "reason": "Assistant inspected original context with this rectangle: clear floor without a visible object from the ten detection classes."
+        }
+      ]
+    },
+    {
+      "frame_id": "0053-Kaohsiung-cam05",
+      "image_sha256": "68440d495ad4d190d7b68b046d64707339105fa7c5290b36b2c46abd3a1f4e8a",
+      "companion_sha256": "a9565f0ac21f98f711d0a0ff3c91608767fdc023a91557974fc41903ded832ab",
+      "positives": [
+        {
+          "id": "ai-0004",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0002",
+          "entity": "cardboard_box",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0006",
+          "entity": "cardboard_box",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0005",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            850,
+            500,
+            1250,
+            800
+          ],
+          "reason": "Assistant inspected original context with this rectangle: clear floor without a visible object from the ten detection classes."
+        }
+      ]
+    },
+    {
+      "frame_id": "0016-Taichung-cam01",
+      "image_sha256": "fb3086c1ff8ed11a485226d69bacacb33ea4d534d0708096c80526f348230784",
+      "companion_sha256": "b30aed64d3c57dba492dec341442706b77458556b8afb6991fe1b2f6741d7bee",
+      "positives": [
+        {
+          "id": "ai-0002",
+          "entity": "laptop",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0021",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0004",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0005",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            600,
+            700,
+            700,
+            900
+          ],
+          "reason": "Assistant inspected original context with this rectangle: clear floor without a visible object from the ten detection classes."
+        }
+      ]
+    },
+    {
+      "frame_id": "0076-Taichung-cam07",
+      "image_sha256": "a7e42db0b7ac8ebae37bc90da01b090b7de169b9cf0bf1b06088d049add3f83e",
+      "companion_sha256": "bf42c51b3bb9657e8bd99379eaabba2d1e53e4d4559bae34f2ceae1959b11dd0",
+      "positives": [
+        {
+          "id": "ai-0005",
+          "entity": "laptop",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0006",
+          "entity": "poster",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0001",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0004",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0002",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": [
+        {
+          "xyxy": [
+            900,
+            800,
+            1200,
+            1000
+          ],
+          "reason": "Assistant inspected original context with this rectangle: clear floor without a visible object from the ten detection classes."
+        }
+      ]
+    },
+    {
+      "frame_id": "0034-Taichung-cam10",
+      "image_sha256": "28c5c9e5a09f55a473bda61d0d84de4a8f679ebc3db43f892358a6a56d0c57fd",
+      "companion_sha256": "602aa819d8944d6dadc76bedecfa8b2609641066dd2dc6a046c52cbebf2846d3",
+      "positives": [
+        {
+          "id": "ai-0002",
+          "entity": "laptop",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0007",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0005",
+          "entity": "chair",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        },
+        {
+          "id": "ai-0010",
+          "entity": "person",
+          "reason": "Assistant inspected original context and isolated mask; a visible single physical object supports this class. Partial AI review only."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "0013-Kaohsiung-cam11",
+      "image_sha256": "4c19e17cad632f5551e2280b514e76884e74658d62841354e3ddfcc1c73a6f19",
+      "companion_sha256": "06b6e65f9c04d603749965cb03a5fc7eb6ee2c8419638baeb6cb93bba69f5f0a",
+      "positives": [
+        {
+          "id": "vlm-unknown-0004",
+          "entity": "phone",
+          "reason": "Assistant inspected context and isolated mask: real tethered smartphone on foreground circular display table; excludes phone depicted on TV advertisement."
+        },
+        {
+          "id": "vlm-unknown-0006",
+          "entity": "phone",
+          "reason": "Assistant inspected context and isolated mask: real tethered smartphone on foreground circular display table; excludes phone depicted on TV advertisement."
+        },
+        {
+          "id": "vlm-unknown-0007",
+          "entity": "phone",
+          "reason": "Assistant inspected context and isolated mask: real tethered smartphone on foreground circular display table; excludes phone depicted on TV advertisement."
+        },
+        {
+          "id": "vlm-unknown-0008",
+          "entity": "phone",
+          "reason": "Assistant inspected context and isolated mask: real tethered smartphone on foreground circular display table; excludes phone depicted on TV advertisement."
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Kaohsiung-cam04-t330",
+      "image_sha256": "a479c28dc3688ca122efa93ce4288707fffa4eb2031ca942d860fae2397c0add",
+      "companion_sha256": "8cbb99661c64303eae9bd97cf37953079b614235f7f39622385d8b70302abc89",
+      "positives": [
+        {
+          "id": "ai-0005",
+          "entity": "laptop",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 0
+        },
+        {
+          "id": "ai-0011",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 3
+        },
+        {
+          "id": "ai-0019",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 4
+        },
+        {
+          "id": "ai-0009",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 5
+        },
+        {
+          "id": "ai-0006",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 6
+        },
+        {
+          "id": "ai-0031",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 12
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 13
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 14
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Kaohsiung-cam07-t030",
+      "image_sha256": "43a4ec3d39087331280da0a4b868ffc47b98cebb4410a304a2f1d14adf9fdc9d",
+      "companion_sha256": "18d91aac7c1cbb4a2e29a8a6f3ef9df12f4e7ef96d967e94e7175ff509d06e53",
+      "positives": [
+        {
+          "id": "ai-0080",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 19
+        },
+        {
+          "id": "ai-0017",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 21
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Kaohsiung-cam08-t030",
+      "image_sha256": "facae65b929ba40c46bf9bb671e1b158c3e9edd392734e6cdea2b4fb69717462",
+      "companion_sha256": "97cf772c23ef3a6058ae6aac87f2cf207b5391d34336713254257157b839e66f",
+      "positives": [
+        {
+          "id": "ai-0023",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 27
+        },
+        {
+          "id": "ai-0025",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 28
+        },
+        {
+          "id": "ai-0018",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 29
+        },
+        {
+          "id": "ai-0042",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 31
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Taichung-cam03-t030",
+      "image_sha256": "e4dc5e1e322ad6b43c802169a816c06fd0b1f6cb8c399e87d4215cd5056e1f82",
+      "companion_sha256": "5ed3466365e1b23c8810968530385e6203616aeb487dae6f0c77e31e10043c6a",
+      "positives": [
+        {
+          "id": "ai-0005",
+          "entity": "laptop",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 32
+        },
+        {
+          "id": "ai-0006",
+          "entity": "laptop",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 33
+        },
+        {
+          "id": "ai-0014",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 35
+        },
+        {
+          "id": "ai-0011",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 36
+        },
+        {
+          "id": "ai-0008",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 37
+        },
+        {
+          "id": "ai-0020",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 40
+        },
+        {
+          "id": "ai-0057",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 41
+        },
+        {
+          "id": "ai-0010",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 42
+        },
+        {
+          "id": "ai-0003",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 43
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Taichung-cam04-t330",
+      "image_sha256": "7daaff94e19ec7f2eb63738d779c358cbbe59406841f5499f8269376befbcad1",
+      "companion_sha256": "503fb3c85237741465e9b3d8c3702535e9651206267e325c69634c81540e62f8",
+      "positives": [
+        {
+          "id": "ai-0054",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 50
+        },
+        {
+          "id": "ai-0057",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 51
+        },
+        {
+          "id": "ai-0050",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 53
+        },
+        {
+          "id": "ai-0009",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 55
+        },
+        {
+          "id": "ai-0028",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 56
+        },
+        {
+          "id": "ai-0016",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 58
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 59
+        },
+        {
+          "id": "ai-0012",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 60
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Taichung-cam05-t030",
+      "image_sha256": "8a7a6e129561f819a598241bbe50107322192a61ca450510ae97c957672686ad",
+      "companion_sha256": "dd4ebfadebb5c9b9a4777c130b124aa1cac9cb3610d20783acf610056b056e4e",
+      "positives": [
+        {
+          "id": "ai-0036",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 64
+        },
+        {
+          "id": "ai-0005",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 66
+        },
+        {
+          "id": "ai-0061",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 68
+        },
+        {
+          "id": "ai-0014",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 69
+        },
+        {
+          "id": "ai-0011",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 70
+        },
+        {
+          "id": "ai-0009",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 71
+        },
+        {
+          "id": "ai-0055",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 73
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Taichung-cam08-t030",
+      "image_sha256": "8f15fd481aeb8f63aa50b25dfa79637535b45defcecae4b5f99b9a5e547b5daa",
+      "companion_sha256": "8a5a47e8f624dde013d5b8bfddfc02b5b6f309196da9a60c94325871d1c28c2d",
+      "positives": [
+        {
+          "id": "ai-0002",
+          "entity": "tablet",
+          "reason": "Assistant corrected phone proposal to tablet: full-size physical iPad with tether and stand on tabletop, compared with the surrounding counter and power sockets; not the adjacent black desk mats.",
+          "review_index": 76
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Taichung-cam09-t030",
+      "image_sha256": "c989bfe1f6a7998834c79588b000dd24f869b96bfc93c14055b9a2b8d2995828",
+      "companion_sha256": "5f7a89f5e2cc3d58f472a954b8fd41992c5615be016b97956911f9bac6aead07",
+      "positives": [
+        {
+          "id": "ai-0066",
+          "entity": "boxed_stock",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 82
+        },
+        {
+          "id": "ai-0028",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 83
+        },
+        {
+          "id": "ai-0026",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 84
+        },
+        {
+          "id": "ai-0002",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 85
+        },
+        {
+          "id": "ai-0005",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 86
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260910-Taichung-cam11-t330",
+      "image_sha256": "272d31978a93c47d1a1aaa04f8581cd0c47135abdd60aab94d0650ef49186d7d",
+      "companion_sha256": "b5d5d582256ff8a0ba2a0f87bedde01d97739e6e1eb9ba8ffdb1e885b9251965",
+      "positives": [
+        {
+          "id": "ai-0017",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 91
+        },
+        {
+          "id": "ai-0030",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 92
+        },
+        {
+          "id": "ai-0011",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 93
+        },
+        {
+          "id": "ai-0028",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 95
+        },
+        {
+          "id": "ai-0002",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 96
+        },
+        {
+          "id": "ai-0006",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 97
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 98
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 99
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260912-Kaohsiung-cam04-t030",
+      "image_sha256": "8f2455851cd3d7d609931dfe0c7681801df689eb20633bb42af56f06837d0730",
+      "companion_sha256": "e57e043fc7a2e8f7a9fce6db43bd41c0758fd9a997cfc803ed49aafdf71a8d43",
+      "positives": [
+        {
+          "id": "ai-0006",
+          "entity": "laptop",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 100
+        },
+        {
+          "id": "ai-0039",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 102
+        },
+        {
+          "id": "ai-0010",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 103
+        },
+        {
+          "id": "ai-0020",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 105
+        },
+        {
+          "id": "ai-0008",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 106
+        },
+        {
+          "id": "ai-0074",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 113
+        },
+        {
+          "id": "ai-0068",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 114
+        },
+        {
+          "id": "ai-0002",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 115
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 116
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260912-Taichung-cam03-t330",
+      "image_sha256": "6f696165a74ce3d833e6fc59a868213f3f261d6b904e2964553abd3107513c52",
+      "companion_sha256": "ffbc822f24b384d131a9e6255a56762405c5c9999e140ebcb54a53edb5ffda85",
+      "positives": [
+        {
+          "id": "ai-0011",
+          "entity": "laptop",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 117
+        },
+        {
+          "id": "ai-0010",
+          "entity": "laptop",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 118
+        },
+        {
+          "id": "ai-0017",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 120
+        },
+        {
+          "id": "ai-0019",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 121
+        },
+        {
+          "id": "ai-0015",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 125
+        },
+        {
+          "id": "ai-0005",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 126
+        },
+        {
+          "id": "ai-0006",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 127
+        },
+        {
+          "id": "ai-0008",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 129
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 130
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260912-Taichung-cam04-t030",
+      "image_sha256": "e528fceb02e0402a1b98840d1e13cd342b8e823ee2bd9f7367353ab9be0bb4cb",
+      "companion_sha256": "d21485f131d60f65e634d033b0a5b58c18c26e78dc4bd3edb89083f882076e8e",
+      "positives": [
+        {
+          "id": "ai-0016",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 139
+        },
+        {
+          "id": "ai-0033",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 140
+        },
+        {
+          "id": "ai-0020",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 141
+        },
+        {
+          "id": "ai-0038",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 142
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 143
+        },
+        {
+          "id": "ai-0003",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 144
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260912-Taichung-cam05-t330",
+      "image_sha256": "f0757ff6f43fadc72a211aee58b1132ac29afe242e35a07ef55c41a2b5bf4209",
+      "companion_sha256": "8b39a80107cfbf5bfd72aed896d486f1579b4a4939f7f72d54696dd98ab1c20e",
+      "positives": [
+        {
+          "id": "ai-0014",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 151
+        },
+        {
+          "id": "ai-0081",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 154
+        },
+        {
+          "id": "ai-0020",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 155
+        },
+        {
+          "id": "ai-0019",
+          "entity": "fire_equipment",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 156
+        },
+        {
+          "id": "ai-0010",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 157
+        },
+        {
+          "id": "ai-0024",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 158
+        },
+        {
+          "id": "ai-0000",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 159
+        },
+        {
+          "id": "ai-0004",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 160
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260912-Taichung-cam08-t030",
+      "image_sha256": "04aae73f6f1184bb6b1fdd4f2f924243c68f6a423b794ca7d806e4fc03cad2ad",
+      "companion_sha256": "d9d462189ede768927dc8b63bf925e6bd141076a6adbc7ba9c2318505e54d331",
+      "positives": [
+        {
+          "id": "ai-0002",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 161
+        },
+        {
+          "id": "ai-0003",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 162
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 164
+        },
+        {
+          "id": "ai-0005",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 165
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260912-Taichung-cam08-t330",
+      "image_sha256": "c834c7362415b2fa11d32470166b44b0a80519929e258ab8b228bdb5a1a79578",
+      "companion_sha256": "2c5ec8ec5dfa4e68f5d6fa34e4cfd1ff0f3203233738a149bd975c8e6a0787cf",
+      "positives": [
+        {
+          "id": "ai-0007",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 167
+        },
+        {
+          "id": "ai-0001",
+          "entity": "phone",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 168
+        },
+        {
+          "id": "ai-0002",
+          "entity": "tablet",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 169
+        },
+        {
+          "id": "ai-0005",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 173
+        },
+        {
+          "id": "ai-0004",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 174
+        }
+      ],
+      "negative_rects": []
+    },
+    {
+      "frame_id": "gcs-20260912-Taichung-cam11-t330",
+      "image_sha256": "f3360b01d4e4d82e1fc8c552714e287efadaeedd015cc3ab0440ee01603ab19a",
+      "companion_sha256": "adba1791cd74f51dab852defec9e743ed3b20045454f100b59a93cb1617b370c",
+      "positives": [
+        {
+          "id": "ai-0003",
+          "entity": "speaker",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 176
+        },
+        {
+          "id": "ai-0016",
+          "entity": "poster",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 178
+        },
+        {
+          "id": "ai-0000",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 179
+        },
+        {
+          "id": "ai-0011",
+          "entity": "chair",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 180
+        },
+        {
+          "id": "ai-0001",
+          "entity": "person",
+          "reason": "Assistant inspected full camera context and isolated mask: one visible physical instance supports this class; partial AI review, not exhaustive labels.",
+          "review_index": 181
+        }
+      ],
+      "negative_rects": []
+    }
+  ],
+  "deferred_candidates": [
+    {
+      "frame_id": "0008-Kaohsiung-cam07",
+      "id": "ai-0035",
+      "reason": "Multiple packaged objects in one mask"
+    },
+    {
+      "frame_id": "0008-Kaohsiung-cam07",
+      "id": "ai-0036",
+      "reason": "Disconnected mask contaminates bounding box"
+    },
+    {
+      "frame_id": "0002-Kaohsiung-cam04",
+      "id": "ai-0007",
+      "reason": "Ambiguous payment terminal or calculator"
+    },
+    {
+      "frame_id": "0033-Taichung-cam09",
+      "id": "ai-0007",
+      "reason": "Multiple overlapping packages"
+    },
+    {
+      "frame_id": "0033-Taichung-cam09",
+      "id": "ai-0008",
+      "reason": "Multiple overlapping packages"
+    },
+    {
+      "frame_id": "0082-Taichung-cam11",
+      "id": "ai-0001",
+      "reason": "Mask includes floor or disconnected background"
+    },
+    {
+      "frame_id": "0007-Kaohsiung-cam06",
+      "id": "ai-0008",
+      "reason": "Printed price placard, not physical phone"
+    },
+    {
+      "frame_id": "0007-Kaohsiung-cam06",
+      "id": "ai-0018",
+      "reason": "Headphones, not speaker"
+    },
+    {
+      "frame_id": "0053-Kaohsiung-cam05",
+      "id": "ai-0003",
+      "reason": "Phone versus payment terminal unresolved"
+    },
+    {
+      "frame_id": "0016-Taichung-cam01",
+      "id": "ai-0014",
+      "reason": "Disconnected mask contaminates bounding box"
+    },
+    {
+      "frame_id": "0016-Taichung-cam01",
+      "id": "ai-0015",
+      "reason": "Printed placard, not boxed stock"
+    },
+    {
+      "frame_id": "0016-Taichung-cam01",
+      "id": "ai-0016",
+      "reason": "Printed placard, not boxed stock"
+    },
+    {
+      "frame_id": "0034-Taichung-cam10",
+      "id": "ai-0020",
+      "reason": "Laptop versus detachable tablet unresolved"
+    },
+    {
+      "frame_id": "0034-Taichung-cam10",
+      "id": "ai-0014",
+      "reason": "Physical phone versus printed price card unresolved"
+    },
+    {
+      "frame_id": "0013-Kaohsiung-cam11",
+      "id": "ai-0004",
+      "reason": "Phone depicted inside TV advertisement, not a physical scene instance"
+    }
+  ],
+  "parent_review_sha256": "8f7d9a491442c8482d5e392b7558c6c579d37609457c16d940013e6fa18cb59f",
+  "added_review_sha256": "9f4c858c78cc83a28f40769d859f8f373572f8b8ee1f5058e0297c3df2615b55"
+}
diff --git a/docs/reviews/studioa_negative_balance_plan_20260916.json b/docs/reviews/studioa_negative_balance_plan_20260916.json
new file mode 100644
index 0000000..880e4fb
--- /dev/null
+++ b/docs/reviews/studioa_negative_balance_plan_20260916.json
@@ -0,0 +1,202 @@
+{
+  "status": "predeclared before either comparison arm runs",
+  "question": "Does capping the added per-class negative mask mass preserve source-val positive recall while scene outputs remain fixed?",
+  "data": "runs/studioa_partial_instances_20260916_v4",
+  "data_manifest_sha256": "e12339ea28db97a8a59c839ad095bf58d961f92af46b0758d6343ebc83475fc1",
+  "initial_checkpoint": "runs/studioa_scene_pilot_20260916_v2/model/best.pt",
+  "arms": {
+    "control": "sum",
+    "candidate": "positive_budget"
+  },
+  "only_training_change": "model.heads.detection.loss.class_negative_normalization",
+  "normalization": "For each batch/class, newly supervised negative points exclude existing positive/empty supervision; multiply by min(1, max(positive_count,1)/max(new_negative_count,1)). Counts independent of logits. This bounds effective mask mass, not gradient norm.",
+  "fixed": {
+    "seed": 42,
+    "input_size": [
+      512,
+      896
+    ],
+    "batch_size": 2,
+    "lr": 0.0002,
+    "warmup_iters": 26,
+    "epochs": 60,
+    "early_stop_patience": 20,
+    "detector_only": true,
+    "scene_eval_frames": 33,
+    "detector_val_frames": 6,
+    "detector_val_positives": 31,
+    "score_strictly_greater_than": 0.2,
+    "iou_at_least": 0.5,
+    "nms": 0.6,
+    "max_det": 100,
+    "test_inference": false,
+    "no_threshold_sweep": true
+  },
+  "upgrade_criteria": {
+    "reference_run": "runs/studioa_detector_warmup_20260916_v1",
+    "total_matched_strictly_greater_than": 7,
+    "matched_person_at_least": 5,
+    "matched_phone_at_least": 1,
+    "matched_poster_at_least": 1,
+    "reviewed_empty_fp_at_most": 0,
+    "all_33_scene_logits_byte_identical": true,
+    "deployment_claim": false
+  },
+  "diagnostic_scope": "28 source-train images, deterministic no augmentation, single image normalization, float32; logit gradient L1 is not model parameter gradient or deployment accuracy",
+  "diagnostic_summary": {
+    "v1": {
+      "laptop": {
+        "positive": 0.012918037129566073,
+        "empty": 0.0003741828104466549,
+        "class_negative": 3.4755482971668243,
+        "positive_points": 442,
+        "class_negative_points": 1092
+      },
+      "phone": {
+        "positive": 0.027144994597620098,
+        "empty": 0.0011384378412913065,
+        "class_negative": 9.630970358848572,
+        "positive_points": 249,
+        "class_negative_points": 1174
+      },
+      "tablet": {
+        "positive": 0.009524137155267454,
+        "empty": 0.0015294674449251033,
+        "class_negative": 7.666859656572342,
+        "positive_points": 578,
+        "class_negative_points": 1119
+      },
+      "boxed_stock": {
+        "positive": 0.007861402924618233,
+        "empty": 0.0006191605880303541,
+        "class_negative": 0.0,
+        "positive_points": 290,
+        "class_negative_points": 0
+      },
+      "cardboard_box": {
+        "positive": 0.10530798323452473,
+        "empty": 0.00014953502613934688,
+        "class_negative": 0.1007998138666153,
+        "positive_points": 160,
+        "class_negative_points": 106
+      },
+      "speaker": {
+        "positive": 0.007214011835458223,
+        "empty": 0.00023646760564588476,
+        "class_negative": 0.13119319081306458,
+        "positive_points": 143,
+        "class_negative_points": 27
+      },
+      "poster": {
+        "positive": 0.07255605278413668,
+        "empty": 0.0036417511128092883,
+        "class_negative": 0.0,
+        "positive_points": 901,
+        "class_negative_points": 0
+      },
+      "fire_equipment": {
+        "positive": 0.010544058675591828,
+        "empty": 0.0023294030424949597,
+        "class_negative": 0.0,
+        "positive_points": 115,
+        "class_negative_points": 0
+      },
+      "chair": {
+        "positive": 0.024498611608578358,
+        "empty": 0.0010922043438768014,
+        "class_negative": 0.0,
+        "positive_points": 350,
+        "class_negative_points": 0
+      },
+      "person": {
+        "positive": 0.0603497300617164,
+        "empty": 0.008183579477190506,
+        "class_negative": 0.22365543246269226,
+        "positive_points": 720,
+        "class_negative_points": 25
+      }
+    },
+    "v2": {
+      "laptop": {
+        "positive": 0.21216800808906555,
+        "empty": 0.00010892083173530409,
+        "class_negative": 0.03900669398717582,
+        "positive_points": 442,
+        "class_negative_points": 1092
+      },
+      "phone": {
+        "positive": 0.5927745196968317,
+        "empty": 0.00010077359479510051,
+        "class_negative": 0.06103876978158951,
+        "positive_points": 249,
+        "class_negative_points": 1174
+      },
+      "tablet": {
+        "positive": 0.4205053150653839,
+        "empty": 0.0001715062435323489,
+        "class_negative": 0.10881199990399182,
+        "positive_points": 578,
+        "class_negative_points": 1119
+      },
+      "boxed_stock": {
+        "positive": 0.2133826856734231,
+        "empty": 0.0001516974407422822,
+        "class_negative": 0.0,
+        "positive_points": 290,
+        "class_negative_points": 0
+      },
+      "cardboard_box": {
+        "positive": 0.17267386615276337,
+        "empty": 0.00024281461901409784,
+        "class_negative": 0.0030626498628407717,
+        "positive_points": 160,
+        "class_negative_points": 106
+      },
+      "speaker": {
+        "positive": 0.25613438803702593,
+        "empty": 6.87338163061213e-05,
+        "class_negative": 0.011610515415668488,
+        "positive_points": 143,
+        "class_negative_points": 27
+      },
+      "poster": {
+        "positive": 0.4980439506471157,
+        "empty": 0.00047163556337181944,
+        "class_negative": 0.0,
+        "positive_points": 901,
+        "class_negative_points": 0
+      },
+      "fire_equipment": {
+        "positive": 0.11960204120259732,
+        "empty": 0.0004438365522219101,
+        "class_negative": 0.0,
+        "positive_points": 115,
+        "class_negative_points": 0
+      },
+      "chair": {
+        "positive": 0.2698142668232322,
+        "empty": 0.00014771780251976452,
+        "class_negative": 0.0,
+        "positive_points": 350,
+        "class_negative_points": 0
+      },
+      "person": {
+        "positive": 0.32032438833266497,
+        "empty": 0.0020131336023041513,
+        "class_negative": 0.07633097469806671,
+        "positive_points": 720,
+        "class_negative_points": 25
+      }
+    }
+  },
+  "diagnostic_limits": [
+    "Per-image float32 no-augmentation logit-gradient L1 is not the historical batch-2 augmented training gradient.",
+    "A well-fit positive can have a small focal gradient, so a large negative/positive ratio alone does not prove the cause of the validation regression.",
+    "All 130 train boxes have at least one assignable grid point in the unaugmented inputs; regression quality may still be insufficient."
+  ],
+  "diagnostic_artifacts": {
+    "runs/studioa_negative_balance_20260916_v1/diagnose.py": "8b7823c9ec02abe2cdda9cf852b2cedcf1ba14325c8cdc21f2dc44fda05bd371",
+    "runs/studioa_negative_balance_20260916_v1/loss_diagnosis.json": "3c5fe8da9fb69ae9f03a4d81828c5ff46417a66e9dd8500a560f5202ed30a53b",
+    "runs/studioa_negative_balance_20260916_v1/diagnose.log": "43a23f0fe14ec609efb9525229b6d76e1a842ec20155e9b23da5e188b160a551"
+  }
+}
diff --git a/scripts/campaign_site30k.py b/scripts/campaign_site30k.py
index 7acaf22..3ba4398 100644
--- a/scripts/campaign_site30k.py
+++ b/scripts/campaign_site30k.py
@@ -59,6 +59,7 @@
 
 import argparse
 import json
+import os
 import sys
 import time
 from collections import Counter, defaultdict
@@ -70,6 +71,7 @@
 from PIL import Image
 
 HERE = Path(__file__).resolve().parent
+ROOT = Path(os.environ.get("SYNCAI_ROOT", HERE.parent))
 sys.path.insert(0, str(HERE))
 
 # The teachers live in the wheel as of 2026-08-20 (52eb141 SAM 3, 556582b Grounding
@@ -158,13 +160,13 @@
 CONSENSUS = 0.9  # the measured setting from sam3_prelabel
 NMS_IOU = 0.55  # same as teachers.boxes.nms
 
-THIRD_OPINION_RUN = HERE.parent / "runs/hydranet_retail_security_b03_gdino"
+THIRD_OPINION_RUN = ROOT / "runs/hydranet_retail_security_b03_gdino"
 # b03 terrain head: 0 void, 1 floor, 2 wall, 3 column, 4 fixture, 5 person.
 # Veto map: site id -> the b03 id that counts as agreement. `display_table` and
 # `shelf` are absent on purpose -- the model cannot arbitrate a split it never saw.
 VETO_SITE_TO_B03 = {1: 1, 2: 2, 3: 3, 6: 5}
 
-PROGRESS_LOG = HERE.parent / "runs/site30k_qa/progress.log"
+PROGRESS_LOG = ROOT / "runs/site30k_qa/progress.log"
 
 
 def log_progress(msg: str) -> None:
@@ -351,9 +353,9 @@ def __init__(
             unproject,
         )
 
-        root = Path(calib_root) if calib_root is not None else HERE.parent / "runs/onboard01"
+        root = Path(calib_root) if calib_root is not None else ROOT / "runs/onboard01"
         calib = json.loads((root / f"{camera}.calib.json").read_text())
-        zones = json.loads((HERE.parent / f"runs/zones01/{camera}.zones.json").read_text())
+        zones = json.loads((ROOT / f"runs/zones01/{camera}.zones.json").read_text())
         if zones.get("units") != "m" or calib.get("scale") is None:
             raise SystemExit(f"{camera}: zones/calib not metric; v2 floor recipe needs both")
         poly = next(p for p in zones["proposals"] if p["name_suggestion"] == "walkable_floor")
@@ -650,13 +652,11 @@ def cmd_annotate(args) -> int:
                 geom_cache[camera] = GeomTeacher(camera, third)
             geom = geom_cache[camera]
             slot = Path(clip).stem.split("_")[1]
-            cand = HERE.parent / f"datasets/studioa_static/{camera}/plate_{slot}.png"
+            cand = ROOT / f"datasets/studioa_static/{camera}/plate_{slot}.png"
             if cand.exists():
                 plate_path = str(cand)  # the SAME clip's own median -- crispest diff
             else:
-                calib_p = json.loads(
-                    (HERE.parent / f"runs/onboard01/{camera}.calib.json").read_text()
-                )
+                calib_p = json.loads((ROOT / f"runs/onboard01/{camera}.calib.json").read_text())
                 plate_path = calib_p["plate_used"]
             plate_floor, dirty, plate_rgb = geom.plate_labeling(plate_path)
             # Per-pixel noise floor of |frame - plate|, static_plates style: a LOW
@@ -984,7 +984,7 @@ def cmd_floor_diag(args) -> int:
     report = {}
     for camera in args.cameras:
         geom = GeomTeacher(camera, third)
-        calib = json.loads((HERE.parent / f"runs/onboard01/{camera}.calib.json").read_text())
+        calib = json.loads((ROOT / f"runs/onboard01/{camera}.calib.json").read_text())
         plate = (
             Image.open(calib["plate_used"])
             .convert("RGB")
@@ -1408,7 +1408,7 @@ def build_parser() -> argparse.ArgumentParser:
     an.add_argument("--sample-fps", type=float, default=1.0)
     an.add_argument(
         "--split-json",
-        default=str(HERE.parent / "datasets/retail_objects_batch03/split.json"),
+        default=str(ROOT / "datasets/retail_objects_batch03/split.json"),
         help="camera split assignments to INHERIT (R1/R2)",
     )
     an.add_argument(
diff --git a/scripts/onboard_camera.py b/scripts/onboard_camera.py
index d7e712b..16d105f 100644
--- a/scripts/onboard_camera.py
+++ b/scripts/onboard_camera.py
@@ -75,6 +75,51 @@
 PLATE_MODEL_CKPT = for_terrain()
 K1_FLEET = pc.K1_FLEET  # the fleet lens; defined once in syncai_bev3d.plate_calibration
 VFOV_PRIMARY = 70.4  # likewise: pinned on cam01, a fleet assumption for the rest
+#: The one camera whose vfov and k1 were measured off its tile grid (calib01), as data.
+#: Until 2026-09-11 this was `camera == "Taichung-cam01"` inside the sweep, so a vfov
+#: measured on any other camera -- `floor_calibrate.py` writes them to a pins file --
+#: had nowhere to go: the primary row was the 70.4 constant whatever the instrument said.
+BUILTIN_PINS: dict[str, dict] = {
+    "Taichung-cam01": {
+        "vfov_deg": 70.4,
+        "source": "tile_grid_pinned",
+        "k1_source": "tile_grid_measured",
+    }
+}
+
+
+def load_pins(path: Path | None) -> dict[str, dict]:
+    """`floor_calibrate.py --pins-out` rows (`{camera: {vfov_deg, source, ...}}`) over
+    the built-in tile pin; a file pin for the same camera wins, being the later measurement."""
+    pins = {k: dict(v) for k, v in BUILTIN_PINS.items()}
+    if path is not None:
+        for cam, row in json.loads(Path(path).read_text()).items():
+            if "vfov_deg" not in row:
+                raise ValueError(f"{path}: pin for {cam} has no vfov_deg")
+            pins[cam] = {"source": "pinned", **row}
+    return pins
+
+
+def pin_for(camera: str, pins: dict[str, dict] | None) -> dict | None:
+    return (pins if pins is not None else BUILTIN_PINS).get(camera)
+
+
+def source_size_from_index(plates_root: Path, camera: str, slot: str) -> list[int] | None:
+    """The clip's native (w, h) that `static_plates.py` recorded for this slot, if any."""
+    index = Path(plates_root) / "index.json"
+    if not index.exists():
+        return None
+    slots = json.loads(index.read_text()).get("cameras", {}).get(camera, {}).get("slots", {})
+    wh = (slots.get(slot) or {}).get("source_wh")
+    return [int(wh[0]), int(wh[1])] if wh else None
+
+
+def primary_row(by_vfov: list[dict], vfov: float) -> dict | None:
+    """The sweep row at the camera's own vfov; None when that row has no fitted floor."""
+    row = next((r for r in by_vfov if abs(float(r["vfov_deg"]) - vfov) < 1e-9), None)
+    return row if row is not None and "pitch_deg" in row else None
+
+
 MIN_HEIGHTS = 10  # min samples for the person-height statistic to emit a number (task spec)
 DIRTY_PLATE_FRAC = 0.05  # person share above this marks a dirty plate (cam04's 8.6% is above)
 # calib01: the gap between the person-height scale and the tile-grid anchor (pose bias,
@@ -167,21 +212,31 @@ def onboard_one(
     meter: PlatePersonMeter | None,
     plates_root: Path,
     person_anns: Path = PERSON_ANNS,
+    utc_offset: int = pc.DEFAULT_UTC_OFFSET_HOURS,
+    pins: dict[str, dict] | None = None,
+    daytime: tuple[int, int] = pc.DEFAULT_DAYTIME_HOURS_LOCAL,
 ) -> dict:
     now = _dt.date.today().isoformat()
-    is_pinned = camera == "Taichung-cam01"  # tile grid: k1 and vfov both measured
+    pin = pin_for(camera, pins)
+    is_pinned = pin is not None
+    vfov_primary = float(pin["vfov_deg"]) if pin else VFOV_PRIMARY
+    k1_source = (pin or {}).get("k1_source", "fleet_hardware_assumed")
     flags: list[str] = []
     if not is_pinned:
-        flags += ["vfov_fleet_assumed", "k1_fleet_assumed"]
+        flags.append("vfov_fleet_assumed")
+    if k1_source == "fleet_hardware_assumed":
+        flags.append("k1_fleet_assumed")
+    # the sweep always contains the camera's own vfov, so the primary row exists
+    vfovs = sorted(set(vfovs) | {vfov_primary})
 
     result: dict = {
         "schema": SCHEMA,
         "camera": camera,
         "generated": now,
-        "vfov_assumed_deg": VFOV_PRIMARY,
-        "vfov_source": "tile_grid_pinned" if is_pinned else "fleet_hardware_assumed",
+        "vfov_assumed_deg": vfov_primary,
+        "vfov_source": pin["source"] if pin else "fleet_hardware_assumed",
         "k1_division_model": k1,
-        "k1_source": "tile_grid_measured" if is_pinned else "fleet_hardware_assumed",
+        "k1_source": k1_source,
         # calib02's visual priors (door height / floor tiles) cannot be automated: the
         # fields are kept, the values left empty, and the SOP's manual step back-fills
         # them. This is "not measured", not "measured out as null".
@@ -217,12 +272,17 @@ def onboard_one(
         )
         return result
 
-    slot = pc.pick_daytime_slot(cam_dir)
+    slot = pc.pick_daytime_slot(cam_dir, utc_offset, daytime)
     plate_path = cam_dir / f"plate_{slot}.png"
     rgb = np.asarray(Image.open(plate_path).convert("RGB"))
     h, w = rgb.shape[:2]
     result.update(
-        {"plate_used": str(plate_path), "plate_slot_utc": slot, "frame_hw_px": [h, w]}
+        {
+            "plate_used": str(plate_path),
+            "plate_slot_utc": slot,
+            "frame_hw_px": [h, w],
+            "source_size_px": source_size_from_index(plates_root, camera, slot),
+        }
     )
 
     # Plate dirty region (the raw plate, not the undistorted one -- stable_infer
@@ -273,8 +333,8 @@ def onboard_one(
     result["by_vfov"] = by_vfov
 
     ok_rows = [r for r in by_vfov if "pitch_deg" in r]
-    primary = next((r for r in by_vfov if r["vfov_deg"] == VFOV_PRIMARY), None)
-    if primary is None or "pitch_deg" not in primary:
+    primary = primary_row(by_vfov, vfov_primary)
+    if primary is None:
         primary = ok_rows[0] if ok_rows else None
         if primary is not None:
             flags.append("primary_vfov_failed_using_fallback")
@@ -542,6 +602,13 @@ def main(argv=None) -> int:
     ap.add_argument("--plates-root", type=Path, default=pc.PLATES)
     ap.add_argument("--k1", type=float, default=K1_FLEET)
     ap.add_argument("--vfovs", default="55,70.4,85")
+    ap.add_argument(
+        "--vfov-pins",
+        type=Path,
+        default=None,
+        help="per-camera measured vfov, `floor_calibrate.py --pins-out`'s file; the "
+        "camera's row becomes its primary and joins the sweep",
+    )
     ap.add_argument("--skip-person-frac", action="store_true")
     ap.add_argument("--cameras-json", type=Path, default=CAMERAS_JSON)
     ap.add_argument(
@@ -555,6 +622,9 @@ def main(argv=None) -> int:
 
     fleet = selling_floor_cameras(args.cameras_json)
     cameras = args.camera or fleet
+    site = json.loads(Path(args.cameras_json).read_text())
+    utc_offset = pc.utc_offset_hours(site)
+    daytime = pc.daytime_hours_local(site)
     args.out.mkdir(parents=True, exist_ok=True)
 
     if args.report_only:
@@ -566,11 +636,22 @@ def main(argv=None) -> int:
     if not args.skip_person_frac:
         meter = PlatePersonMeter(PLATE_MODEL_CONFIG, PLATE_MODEL_CKPT)
     vfovs = [float(x) for x in args.vfovs.split(",")]
+    pins = load_pins(args.vfov_pins)
 
     for i, cam in enumerate(cameras, 1):
         print(f"[{i}/{len(cameras)}] {cam}")
         try:
-            result = onboard_one(cam, vfovs, args.k1, meter, args.plates_root, args.person_anns)
+            result = onboard_one(
+                cam,
+                vfovs,
+                args.k1,
+                meter,
+                args.plates_root,
+                args.person_anns,
+                utc_offset,
+                pins,
+                daytime=daytime,
+            )
         except Exception:
             traceback.print_exc()
             result = {
diff --git a/scripts/propose_zones.py b/scripts/propose_zones.py
index 758a274..d33a6bd 100644
--- a/scripts/propose_zones.py
+++ b/scripts/propose_zones.py
@@ -46,6 +46,7 @@
 import hashlib
 import json
 import math
+import os
 import sys
 from pathlib import Path
 
@@ -54,6 +55,7 @@
 from scipy import ndimage
 
 HERE = Path(__file__).resolve().parent
+ROOT = Path(os.environ.get("SYNCAI_ROOT", HERE.parent))
 sys.path.insert(0, str(HERE))
 
 from syncai_bev3d.floorplan import (  # noqa: E402
diff --git a/scripts/static_plates.py b/scripts/static_plates.py
index 4e9f3ae..ae409a9 100644
--- a/scripts/static_plates.py
+++ b/scripts/static_plates.py
@@ -73,6 +73,8 @@
 import numpy as np
 from PIL import Image
 
+from syncai_bev3d.plate_calibration import daytime_hours_local, utc_offset_hours
+
 # How many of a pixel's OWN noise floors it must move to count as dynamic.
 #
 # **Per pixel, not per frame, and that is the whole point.** A single floor for the frame
@@ -198,6 +200,32 @@ def plate_and_mask(frames: np.ndarray) -> tuple[np.ndarray, np.ndarray, dict]:
     )
 
 
+def source_size(path: Path) -> tuple[int, int] | None:
+    """The clip's native (width, height), before `WORK_W x WORK_H`: what a consumer of the
+    calibration will receive boxes in. None when ffprobe cannot say."""
+    proc = subprocess.run(
+        [
+            "ffprobe",
+            "-v",
+            "error",
+            "-select_streams",
+            "v:0",
+            "-show_entries",
+            "stream=width,height",
+            "-of",
+            "csv=p=0",
+            str(path),
+        ],
+        capture_output=True,
+        text=True,
+    )
+    try:
+        w, h = proc.stdout.strip().split(",")[:2]
+        return int(w), int(h)
+    except ValueError:
+        return None
+
+
 def slot_of(path: Path) -> str:
     """`archive_20260816-113024_...` -> `20260816-113024`, which is UTC. See the header."""
     return path.stem.split("_")[1]
@@ -216,7 +244,10 @@ def main(argv: list[str] | None = None) -> int:
     )
     args = ap.parse_args(argv)
 
-    roles = json.loads((args.cameras or args.root / "cameras.json").read_text())["cameras"]
+    site = json.loads((args.cameras or args.root / "cameras.json").read_text())
+    roles = site["cameras"]
+    utc_offset = utc_offset_hours(site)
+    daytime = daytime_hours_local(site)
     names = args.only or sorted(
         k for k, v in roles.items() if args.include_dead or v.get("role") != "dead"
     )
@@ -244,7 +275,11 @@ def main(argv: list[str] | None = None) -> int:
             Image.fromarray((static * 255).astype(np.uint8)).save(
                 cam_out / f"static_{slot}.png"
             )
-            slots[slot] = {**stats, "role": roles.get(cam, {}).get("role")}
+            slots[slot] = {
+                **stats,
+                "role": roles.get(cam, {}).get("role"),
+                "source_wh": source_size(clip),
+            }
             every = static if every is None else (every & static)
         if every is not None:
             Image.fromarray((every * 255).astype(np.uint8)).save(cam_out / "static_all.png")
@@ -264,7 +299,9 @@ def main(argv: list[str] | None = None) -> int:
         json.dumps(
             {
                 "measured": "static plates per camera per time slot",
-                "note": "slot keys are UTC; the burned-in timestamp is store-local (+8)",
+                "note": "slot keys are UTC; the burned-in timestamp is store-local",
+                "utc_offset_hours": utc_offset,
+                "daytime_hours_local": list(daytime),
                 "sample_fps": SAMPLE_FPS,
                 "work_size": [WORK_W, WORK_H],
                 "dynamic_mult": DYNAMIC_MULT,
diff --git a/src/syncai_bev3d/commissioning.py b/src/syncai_bev3d/commissioning.py
index 6d36172..bf0652d 100644
--- a/src/syncai_bev3d/commissioning.py
+++ b/src/syncai_bev3d/commissioning.py
@@ -81,6 +81,11 @@ def from_onboard_calib(path: str | Path) -> CameraFile:
             radius_px=math.hypot(h, w) / 2.0,
         ),
         plate_file=raw.get("plate_used"),
+        source_size_px=(
+            (int(raw["source_size_px"][0]), int(raw["source_size_px"][1]))
+            if raw.get("source_size_px")
+            else None
+        ),
         commissioned_at=raw.get("generated"),
         teachers=_teachers_of(raw),
     )
diff --git a/src/syncai_bev3d/plate_calibration.py b/src/syncai_bev3d/plate_calibration.py
index 46777ef..cec3703 100644
--- a/src/syncai_bev3d/plate_calibration.py
+++ b/src/syncai_bev3d/plate_calibration.py
@@ -79,13 +79,48 @@
 # plate handling
 
 
-def pick_daytime_slot(cam_dir: Path) -> str:
-    """Brightest plate among the daytime slots (slot keys are UTC; store-local is +8)."""
+#: The offset every site carried until 2026-09-11, when a US-Central factory's plates
+#: passed the daytime gate by coincidence (its 10 UTC slot read as 18 local under +8).
+#: A site states its own offset in `cameras.json`; this is only what an old file means.
+DEFAULT_UTC_OFFSET_HOURS = 8
+
+
+def utc_offset_hours(cameras_json: dict) -> int:
+    """The site's `utc_offset_hours` from a `cameras.json` payload, +8 when it says nothing."""
+    return int(cameras_json.get("utc_offset_hours", DEFAULT_UTC_OFFSET_HOURS))
+
+
+#: A shop's lit hours. A factory on night shifts is lit at 05:00 local (FTI, 2026-09-16:
+#: its only capture fell there and the gate refused it), so a site states its own window.
+DEFAULT_DAYTIME_HOURS_LOCAL = (8, 18)
+
+
+def daytime_hours_local(cameras_json: dict) -> tuple[int, int]:
+    """The site's `daytime_hours_local` `[start, end]` from `cameras.json`, 08-18 when absent.
+
+    Inclusive local hours; `start > end` is a window across midnight (`[20, 6]`)."""
+    start, end = cameras_json.get("daytime_hours_local", DEFAULT_DAYTIME_HOURS_LOCAL)
+    return int(start), int(end)
+
+
+def in_daytime(hour_local: int, daytime: tuple[int, int] = DEFAULT_DAYTIME_HOURS_LOCAL) -> bool:
+    start, end = daytime
+    if start <= end:
+        return start <= hour_local <= end
+    return hour_local >= start or hour_local <= end
+
+
+def pick_daytime_slot(
+    cam_dir: Path,
+    utc_offset: int = DEFAULT_UTC_OFFSET_HOURS,
+    daytime: tuple[int, int] = DEFAULT_DAYTIME_HOURS_LOCAL,
+) -> str:
+    """Brightest plate among the daytime slots (slot keys are UTC; local = UTC + offset)."""
     best, best_luma = None, -1.0
     for p in sorted(cam_dir.glob("plate_*.png")):
         slot = p.stem.split("_", 1)[1]
-        hour_local = (int(slot[9:11]) + 8) % 24
-        if not (8 <= hour_local <= 18):
+        hour_local = (int(slot[9:11]) + utc_offset) % 24
+        if not in_daytime(hour_local, daytime):
             continue
         luma = float(np.asarray(Image.open(p).convert("L"), dtype=float).mean())
         if luma > best_luma:
diff --git a/src/syncai_bev3d/rulers.py b/src/syncai_bev3d/rulers.py
index 1506375..326c265 100644
--- a/src/syncai_bev3d/rulers.py
+++ b/src/syncai_bev3d/rulers.py
@@ -64,10 +64,27 @@ class Ruler:
     factor: float  # multiply the camera's current metres by this
     sigma: float  # fractional uncertainty of `factor`
     note: str = ""
-
-
-def person_ruler(calib: dict) -> Ruler:
-    """The calibration's own person-prior scale: factor 1, uncertainty from its record."""
+    #: Whether the reference length came from outside the camera's own metres. Every
+    #: ruler *reads* through those metres; what separates a witness from an echo is where
+    #: its reference came from. The 1.70 m prior and a known object are anchored. A
+    #: catalogue tile chosen because the store's cameras agree on it is not: two cameras
+    #: through one over-scaled depth model agree with each other, not with the floor --
+    #: FTI's 600 mm raised floor read 0.85 / 0.74 m and the catalogue picked 0.80
+    #: (2026-09-10). A store-median counter is relative by construction.
+    anchored: bool = True
+
+
+def person_ruler(calib: dict) -> Ruler | None:
+    """The calibration's own person-prior scale: factor 1, uncertainty from its record.
+
+    None when the calibration says its scale was never measured -- `unmeasured`, or a
+    bootstrap on the depth model's raw metres: factor 1 there would anchor the other
+    rulers to the very reading they are meant to check. A record with no `scale_source`
+    at all is read as measured; absence is not a declaration.
+    """
+    source = str(calib.get("scale_source") or "")
+    if source.startswith("unmeasured") or "bootstrap" in source:
+        return None
     u = calib.get("uncertainty") or {}
     stat = float(u.get("scale_frac_stat_person_mad") or 0.09)
     sys_ = float(u.get("scale_frac_sys_person_prior") or 0.11)
@@ -93,11 +110,19 @@ def standard_tile(periods_m: dict[str, float]) -> tuple[float | None, float]:
     return best, best_err
 
 
-def tile_ruler(period_m: float | None, tile_m: float | None) -> Ruler | None:
+def tile_ruler(
+    period_m: float | None, tile_m: float | None, *, anchored: bool = True
+) -> Ruler | None:
+    """`anchored=False` when `tile_m` is `standard_tile`'s consensus pick rather than a
+    size known independently of the cameras' readings."""
     if not period_m or not tile_m:
         return None
     return Ruler(
-        "tile", tile_m / period_m, TILE_SIGMA, f"pitch {period_m:.2f} m vs {tile_m:.2f} m"
+        "tile",
+        tile_m / period_m,
+        TILE_SIGMA,
+        f"pitch {period_m:.2f} m vs {tile_m:.2f} m",
+        anchored=anchored,
     )
 
 
@@ -108,7 +133,11 @@ def table_ruler(table_h_m: float | None, store_table_h_m: list[float]) -> Ruler
         return None
     ref = float(np.median(store_table_h_m))
     return Ruler(
-        "table", ref / table_h_m, TABLE_SIGMA, f"{table_h_m:.2f} m vs store {ref:.2f} m"
+        "table",
+        ref / table_h_m,
+        TABLE_SIGMA,
+        f"{table_h_m:.2f} m vs store {ref:.2f} m",
+        anchored=False,
     )
 
 
@@ -124,6 +153,17 @@ def combine(rulers: list[Ruler]) -> Verdict:
     rs = tuple(r for r in rulers if r is not None)
     if not rs:
         return Verdict(1.0, False, "no ruler", ())
+    if not any(r.anchored for r in rs):
+        # A verdict here would be an echo: the rulers can agree with each other and all be
+        # wrong by the same factor. Say so rather than "within 10%".
+        names = ", ".join(r.name for r in rs)
+        return Verdict(
+            1.0,
+            False,
+            f"unanchored: {names} take their reference from the same metres they read; "
+            "needs a person prior or a size known independently",
+            rs,
+        )
     logs = np.array([math.log(r.factor) for r in rs])
     w = np.array([1.0 / (r.sigma**2) for r in rs])
     order = np.argsort(logs)
diff --git a/src/syncai_hydranet/analytics/world.py b/src/syncai_hydranet/analytics/world.py
index 31479c8..1027bc6 100644
--- a/src/syncai_hydranet/analytics/world.py
+++ b/src/syncai_hydranet/analytics/world.py
@@ -353,6 +353,24 @@ def as_rows(frame: WorldFrame) -> list[dict]:
     return rows
 
 
+def panel_box(
+    box, source_size_px: tuple[int, int], panel_size_px: tuple[int, int]
+) -> tuple[float, float, float, float]:
+    """A source-frame box drawn on a panel of another size: scaled by each axis' ratio.
+
+    `demo_video.py` drew every track box with `/ 2.0` -- a 1920-wide source over a 960-wide
+    panel, written as a constant. On a 1280-wide clip the boxes sat at two thirds of the
+    people; on a 3840-wide one they drew at twice their size (FTI, 2026-09-11). The blur
+    boxes beside them were scaled by the probed source size and were right, which is what
+    made the two disagree on the same panel. The ratio is a property of the two frames,
+    computed here and nowhere else.
+    """
+    x0, y0, x1, y1 = (float(v) for v in box)
+    sx = panel_size_px[0] / source_size_px[0]
+    sy = panel_size_px[1] / source_size_px[1]
+    return (x0 * sx, y0 * sy, x1 * sx, y1 * sy)
+
+
 def _finite(value: float | None) -> float | None:
     if value is None:
         return None
diff --git a/src/syncai_hydranet/config_schema.py b/src/syncai_hydranet/config_schema.py
index 2d24d92..db6eab1 100644
--- a/src/syncai_hydranet/config_schema.py
+++ b/src/syncai_hydranet/config_schema.py
@@ -66,6 +66,7 @@ class ConfigError(ValueError):
 }
 
 MODEL = {
+    "detection_only_training": Spec((bool,)),
     "backbone": Spec((dict,), required=True),
     "neck": Spec((dict,), required=True),
     "heads": Spec((dict,), required=True),
@@ -161,6 +162,7 @@ class ConfigError(ValueError):
         "cls_weight": Spec(NUMBER),
         "reg_weight": Spec(NUMBER),
         "centerness_weight": Spec(NUMBER),
+        "class_negative_normalization": Spec((str,), choices=("sum", "positive_budget")),
     },
     "pose_p3": {},
     "depth_fpn": {
@@ -199,6 +201,7 @@ class ConfigError(ValueError):
 }
 
 DATASET = {
+    "validation_only": Spec((bool,)),
     "name": Spec((str,), required=True),
     "type": Spec(
         (str,),
@@ -685,6 +688,8 @@ def _check_datasets(rep: _Report, dcfg: dict, head_names: set[str]) -> None:
         _check_section(rep, ds, DATASET, path)
         if not isinstance(ds, dict):
             continue
+        if ds.get("validation_only") and not ds.get("split_val"):
+            rep.errors.append(f"{path}: validation_only requires split_val")
         name = ds.get("name")
         if name in seen:
             rep.errors.append(f"{path}.name: {name!r} is used by an earlier dataset")
diff --git a/src/syncai_hydranet/data/datasets.py b/src/syncai_hydranet/data/datasets.py
index 44d9928..b9ed1ed 100644
--- a/src/syncai_hydranet/data/datasets.py
+++ b/src/syncai_hydranet/data/datasets.py
@@ -552,6 +552,8 @@ def split_leaks(datasets: list[dict]) -> list[tuple[str, str, str, list[str]]]:
     out: list[tuple[str, str, str, list[str]]] = []
     segs = [d for d in datasets if d.get("root")]
     for a in segs:
+        if a.get("validation_only", False):
+            continue
         trained = _configured_cameras(a, a.get("split_train", "train"))
         if not trained:
             continue
diff --git a/src/syncai_hydranet/data/studioa_instances.py b/src/syncai_hydranet/data/studioa_instances.py
index 73a5abc..d364cb4 100644
--- a/src/syncai_hydranet/data/studioa_instances.py
+++ b/src/syncai_hydranet/data/studioa_instances.py
@@ -8,6 +8,7 @@
 from pathlib import Path
 
 import numpy as np
+import torch
 from PIL import Image
 from torch.utils.data import Dataset
 
@@ -32,6 +33,38 @@
 )
 
 
+def class_negative_regions(item: dict, size, boxes, labels, split: str) -> dict:
+    """Only explicit, image-bound AI decisions can negate an individual class."""
+    regions = item.get("class_negative_rects", [])
+    if regions and split != "train":
+        raise ValueError("class negative additions are train-only")
+    w, h = size
+    masks = {}
+    for region in regions:
+        entity, rect = region.get("entity"), region.get("xyxy", [])
+        if (
+            entity not in CLASSES
+            or not region.get("reason")
+            or len(rect) != 4
+            or any(type(v) is not int for v in rect)
+        ):
+            raise ValueError("class negative region needs class, integer bounds and reason")
+        x0, y0, x1, y1 = rect
+        if not (0 <= x0 < x1 <= w and 0 <= y0 < y1 <= h):
+            raise ValueError("class negative region outside image")
+        label = CLASSES.index(entity)
+        for box, positive_label in zip(boxes, labels, strict=True):
+            if (
+                label == positive_label
+                and min(x1, box[2]) > max(x0, box[0])
+                and min(y1, box[3]) > max(y0, box[1])
+            ):
+                raise ValueError("class negative contradicts a reviewed positive box")
+        mask = masks.setdefault(entity, np.zeros((h, w), dtype=np.uint8))
+        mask[y0:y1, x0:x1] = 1
+    return masks
+
+
 def export_instances(source: Path, reviews: Path, out: Path) -> dict:
     """Export only source train/val frames explicitly inspected by the AI reviewer."""
     manifest = check_supervision(source)
@@ -90,7 +123,8 @@ def export_instances(source: Path, reviews: Path, out: Path) -> dict:
             negative[y0:y1, x0:x1] = 1
         if (occupied & (negative == 1)).any():
             raise ValueError("reviewed empty region overlaps a reviewed object")
-        prepared.append((frame, boxes, labels, identities, negative))
+        class_negatives = class_negative_regions(item, (w, h), boxes, labels, assignments[fid])
+        prepared.append((frame, boxes, labels, identities, negative, class_negatives))
     if not prepared:
         raise ValueError("empty instance review")
     out.mkdir(parents=True, exist_ok=False)
@@ -100,7 +134,7 @@ def export_instances(source: Path, reviews: Path, out: Path) -> dict:
     shutil.copyfile(__file__, out / "producer.py")
     frames = []
     counts = Counter(dict.fromkeys(CLASSES, 0))
-    for frame, boxes, labels, identities, negative in prepared:
+    for frame, boxes, labels, identities, negative, class_negatives in prepared:
         fid = frame["id"]
         shutil.copyfile(source / frame["image"], out / frame["image"])
         shutil.copyfile(source / frame["companion"], out / frame["companion"])
@@ -109,6 +143,11 @@ def export_instances(source: Path, reviews: Path, out: Path) -> dict:
             out / target_name, {"boxes": boxes, "labels": labels, "source_ids": identities}
         )
         Image.fromarray(negative).save(out / negative_name)
+        class_files = {}
+        for entity, mask in class_negatives.items():
+            name = f"annotations/{fid}.negative-{entity}.png"
+            Image.fromarray(mask).save(out / name)
+            class_files[entity] = name
         counts.update(CLASSES[i] for i in labels)
         frames.append(
             {
@@ -129,6 +168,8 @@ def export_instances(source: Path, reviews: Path, out: Path) -> dict:
                 "targets": target_name,
                 "negative_mask": negative_name,
                 "negative_pixels": int(negative.sum()),
+                "class_negative_masks": class_files,
+                "class_negative_pixels": {k: int(v.sum()) for k, v in class_negatives.items()},
                 "instances": len(labels),
             }
         )
@@ -142,10 +183,12 @@ def export_instances(source: Path, reviews: Path, out: Path) -> dict:
         "review_sha256": digest(reviews),
         "instances_by_class": dict(counts),
         "policy": (
-            "positive assigned channel only; negatives only in AI-reviewed empty regions; "
+            "positive assigned channel; negatives in AI-reviewed empty regions or explicit "
+            "per-class reviewed absence; "
             "unknown and padding ignored"
         ),
         "exhaustive_labels": False,
+        "class_negative_supervision": any(row[-1] for row in prepared),
         "independent_accuracy": False,
     }
     write_json(out / "manifest.json", result)
@@ -184,6 +227,22 @@ def check_instances(root: Path) -> dict:
             f["store"], f["original_split"], m["held_out"]
         ):
             raise ValueError("instance split role changed")
+        for entity, name in f.get("class_negative_masks", {}).items():
+            if (
+                split != "train"
+                or entity not in CLASSES
+                or not m.get("class_negative_supervision")
+            ):
+                raise ValueError("invalid class negative split/class contract")
+            if name not in report["outputs"]:
+                raise ValueError("unbound class negative mask")
+            with Image.open(_relative(root, name)) as image:
+                mask = np.array(image)
+            if (
+                mask.shape != tuple(reversed(f["image_size_px"]))
+                or not np.isin(mask, [0, 1]).all()
+            ):
+                raise ValueError("invalid class negative mask shape/values")
         for mapping, key in ((cameras, f["camera"]), (pixels, f["pixel_sha256"])):
             if mapping.setdefault(key, split) != split:
                 raise ValueError("instance camera/content split leak")
@@ -212,6 +271,7 @@ def __init__(
         ):
             raise ValueError("invalid instance fold/split/augmentation")
         self.root = root
+        self.class_negative_supervision = bool(m.get("class_negative_supervision", False))
         if partial_eval not in (None, "reviewed_regions_v1"):
             raise ValueError("unsupported partial evaluation protocol")
         self.partial_detection_evaluation = partial_eval
@@ -236,14 +296,30 @@ def __getitem__(self, index):
             rgb = image.convert("RGB")
         with Image.open(self.root / f["negative_mask"]) as image:
             negative = np.array(image)
+        masks = {"det_negative_mask": negative}
+        if self.class_negative_supervision:
+            for entity in CLASSES:
+                name = f.get("class_negative_masks", {}).get(entity)
+                if name:
+                    with Image.open(self.root / name) as image:
+                        masks[f"negative_{entity}"] = np.array(image)
+                else:
+                    masks[f"negative_{entity}"] = np.zeros_like(negative)
         sample = self.transform(
             Sample(
                 image=rgb,
-                masks={"det_negative_mask": negative},
+                masks=masks,
                 boxes=np.asarray(targets["boxes"], dtype=np.float32).reshape(-1, 4),
                 labels=np.asarray(targets["labels"], dtype=np.int64),
             )
         )
+        if self.class_negative_supervision:
+            sample["masks"]["det_class_negative_mask"] = torch.stack(
+                [
+                    sample["masks"].pop(f"negative_{entity}").to(torch.uint8)
+                    for entity in CLASSES
+                ]
+            )
         return {
             "image": sample["image"],
             "supervises": self.supervises,
diff --git a/src/syncai_hydranet/engine/trainer.py b/src/syncai_hydranet/engine/trainer.py
index 670e6d2..9c1a8ab 100644
--- a/src/syncai_hydranet/engine/trainer.py
+++ b/src/syncai_hydranet/engine/trainer.py
@@ -81,7 +81,12 @@ def _build_datasets(dcfg, input_size):
     aug = dcfg.get("augment")
     train_sets, val_sets, names, ratios, val_names = [], [], [], [], []
     for ds in dcfg["datasets"]:
-        train_sets.append(build_dataset(ds, input_size, "train", letterbox=lb, augment=aug))
+        if not ds.get("validation_only", False):
+            train_sets.append(build_dataset(ds, input_size, "train", letterbox=lb, augment=aug))
+            names.append(ds["name"])
+            ratios.append(float(ds.get("sample_ratio", 1.0)))
+        elif not ds.get("split_val"):
+            raise ValueError("validation_only requires split_val")
         # A dataset may contribute training signal without joining checkpoint selection:
         # omit `split_val` and it is trained on but never validated on.
         #
@@ -96,8 +101,8 @@ def _build_datasets(dcfg, input_size):
             # Validation never augments, so it takes no augment argument.
             val_sets.append(build_dataset(ds, input_size, "val", letterbox=lb))
             val_names.append(ds["name"])
-        names.append(ds["name"])
-        ratios.append(float(ds.get("sample_ratio", 1.0)))
+    if not train_sets:
+        raise ValueError("at least one training dataset is required")
     if not val_sets:
         raise ValueError(
             "no dataset declares split_val, so nothing can select a checkpoint; "
@@ -447,6 +452,7 @@ def _write_run_meta(self, cfg, names, train_sets, dcfg, val_size, total_iters) -
         to the data behind it.
         """
         n_params = sum(p.numel() for p in self.model.parameters())
+        train_sizes = {name: len(ds) for name, ds in zip(names, train_sets, strict=True)}
         meta = write_run_meta(
             self.out_dir,
             cfg,
@@ -456,16 +462,17 @@ def _write_run_meta(self, cfg, names, train_sets, dcfg, val_size, total_iters) -
             parameters=n_params,
             datasets=[
                 {
-                    "name": n,
-                    "train_size": len(t),
+                    "name": ds["name"],
+                    "train_size": train_sizes.get(ds["name"], 0),
+                    "validation_only": bool(ds.get("validation_only", False)),
                     # None, not 0: "not validated on" and "validated on nothing" are
                     # different facts and the run meta should not blur them.
-                    "val_size": val_size.get(n),
+                    "val_size": val_size.get(ds["name"]),
                     # Datasets live outside git; without this, "which data produced
                     # this checkpoint" has no answer six months later.
                     **fingerprint_dataset(ds),
                 }
-                for n, t, ds in zip(names, train_sets, dcfg["datasets"], strict=True)
+                for ds in dcfg["datasets"]
             ],
         )
         _log_code_version(meta["git"], self.out_dir, self.logger)
diff --git a/src/syncai_hydranet/geometry/camera_json.py b/src/syncai_hydranet/geometry/camera_json.py
index 5a3d400..ff760d0 100644
--- a/src/syncai_hydranet/geometry/camera_json.py
+++ b/src/syncai_hydranet/geometry/camera_json.py
@@ -119,6 +119,12 @@ class CameraFile:
     mask_ignore: int = IGNORE
     plate_file: str | None = None
     plate_sha256: str | None = None
+    # (width, height) of the stream the boxes will arrive in. `image_size_px` is the frame
+    # the intrinsics were fitted at -- half the plate's source by convention -- and every
+    # consumer that rescales boxes into it assumed the source was 1920x1080. An FTI
+    # camera commissioned from a 1280x720 stream and another from 3840x2160 both returned
+    # metres, wrong by 3/2 and 1/2, and nothing said so. None: not recorded (a v3 file).
+    source_size_px: tuple[int, int] | None = None
     commissioned_at: str | None = None  # ISO 8601, stamped by the writer
     # Which teacher models produced this file, as `{model_id: revision}`. The plate is
     # hashed (`plate_sha256`) and the models that read it were not, so two camera.jsons
@@ -192,6 +198,11 @@ def load(cls, path: str | Path) -> CameraFile:
             mask_files=dict(raw.get("mask_files", {})),
             mask_ignore=raw.get("mask_ignore", IGNORE),
             plate_file=raw.get("plate_file"),
+            source_size_px=(
+                (int(raw["source_size_px"][0]), int(raw["source_size_px"][1]))
+                if raw.get("source_size_px")
+                else None
+            ),
             plate_sha256=raw.get("plate_sha256"),
             commissioned_at=raw.get("commissioned_at"),
             teachers=raw.get("teachers"),
@@ -255,6 +266,15 @@ def validate(self) -> None:
         w, h = self.image_size_px
         if w <= 0 or h <= 0:
             problems.append(f"image_size_px {self.image_size_px} is not a size")
+        if self.source_size_px is not None:
+            sw, sh = self.source_size_px
+            if sw <= 0 or sh <= 0:
+                problems.append(f"source_size_px {self.source_size_px} is not a size")
+            elif w > 0 and h > 0 and abs((sw / sh) / (w / h) - 1) > 0.01:
+                problems.append(
+                    f"source_size_px {sw}x{sh} and image_size_px {w}x{h} differ in aspect; "
+                    "a box scaled by width alone lands on the wrong floor pixel"
+                )
         if self.appearance_thr is not None and not self.appearance_thr > 0:
             problems.append(
                 f"appearance_thr {self.appearance_thr!r} is not a positive distance. "
diff --git a/src/syncai_hydranet/models/heads/registry.py b/src/syncai_hydranet/models/heads/registry.py
index 003a701..deceac2 100644
--- a/src/syncai_hydranet/models/heads/registry.py
+++ b/src/syncai_hydranet/models/heads/registry.py
@@ -153,6 +153,7 @@ def loss(self, out: dict, targets: dict) -> tuple[torch.Tensor, dict]:
             # and trains unchanged.
             class_mask=targets.get("det_class_mask"),
             negative_mask=targets.get("det_negative_mask"),
+            class_negative_mask=targets.get("det_class_negative_mask"),
         )
 
     def supervised_by(self, _targets: dict) -> bool:
diff --git a/src/syncai_hydranet/models/hydranet.py b/src/syncai_hydranet/models/hydranet.py
index f99b3bf..00e2185 100644
--- a/src/syncai_hydranet/models/hydranet.py
+++ b/src/syncai_hydranet/models/hydranet.py
@@ -93,6 +93,9 @@ def __init__(self, cfg):
                     cls_weight=lcfg.get("cls_weight", 1.0),
                     reg_weight=lcfg.get("reg_weight", 1.0),
                     centerness_weight=lcfg.get("centerness_weight", 1.0),
+                    class_negative_normalization=lcfg.get(
+                        "class_negative_normalization", "sum"
+                    ),
                 )
             elif hcfg["type"] == "depth_fpn":
                 self.depth_heads[name] = build_depth_head(hcfg, ch)
@@ -116,6 +119,23 @@ def __init__(self, cfg):
         else:
             self.balancer = FixedWeighting(mcfg.get("fixed_weights", {}))
 
+        self.detection_only_training = bool(mcfg.get("detection_only_training", False))
+        if self.detection_only_training:
+            if self.det_head is None:
+                raise ValueError("detection_only_training requires a detection head")
+            for name, parameter in self.named_parameters():
+                parameter.requires_grad_(name.startswith("det_head."))
+            self.train(self.training)
+
+    def train(self, mode: bool = True):
+        """A frozen scene also needs frozen normalization and disabled dropout."""
+        super().train(mode)
+        if self.detection_only_training:
+            for name, module in self.named_children():
+                if name != "det_head":
+                    module.eval()
+        return self
+
     def forward(self, images: torch.Tensor) -> dict:
         """Pure convolution graph. This is exactly what gets exported to ONNX."""
         feats = self.neck(self.backbone(images))
diff --git a/src/syncai_hydranet/models/losses.py b/src/syncai_hydranet/models/losses.py
index a4e2242..cba05c0 100644
--- a/src/syncai_hydranet/models/losses.py
+++ b/src/syncai_hydranet/models/losses.py
@@ -146,11 +146,36 @@ def giou_loss(pred_ltrb: torch.Tensor, target_ltrb: torch.Tensor) -> torch.Tenso
     return (1.0 - giou).sum()
 
 
+def class_negative_weights(selected, positive, existing):
+    """Cap added negative mass per batch/class; large reviewed areas are not more labels.
+
+    Positive and existing empty supervision retain their original weights. For a
+    class absent from this batch, allow one effective negative point so negative-only
+    examples still teach absence. Counts and scales are detached supervision, not a
+    function of model confidence. This bounds mask mass, not the gradient norm.
+    """
+    novel = selected & (existing == 0)
+    budget = positive.float().sum(dim=(0, 1)).clamp(min=1)
+    count = novel.sum(dim=(0, 1)).clamp(min=1)
+    scale = (budget / count).clamp(max=1)
+    return novel.float() * scale[None, None, :]
+
+
 class FCOSLoss(nn.Module):
-    def __init__(self, num_classes: int, cls_weight=1.0, reg_weight=1.0, centerness_weight=1.0):
+    def __init__(
+        self,
+        num_classes: int,
+        cls_weight=1.0,
+        reg_weight=1.0,
+        centerness_weight=1.0,
+        class_negative_normalization="sum",
+    ):
         super().__init__()
         self.num_classes = num_classes
         self.w = (cls_weight, reg_weight, centerness_weight)
+        if class_negative_normalization not in ("sum", "positive_budget"):
+            raise ValueError("unsupported class_negative_normalization")
+        self.class_negative_normalization = class_negative_normalization
 
     def forward(
         self,
@@ -162,6 +187,7 @@ def forward(
         labels_list,
         class_mask=None,
         negative_mask=None,
+        class_negative_mask=None,
     ):
         """``class_mask`` is [B, C] or [C]: which channels this batch's dataset can label.
 
@@ -170,6 +196,8 @@ def forward(
         Only the assigned positive channel is supervised at a labelled box; other
         channels and unlabelled locations are unknown, not false negatives. None
         preserves exhaustive-box training, including its ordinary background loss.
+        ``class_negative_mask`` is optional [B,C,H,W] reviewed absence per class;
+        it never turns the other channels into background.
         """
         device = cls_out[0].device
         shapes = [c.shape[-2:] for c in cls_out]
@@ -189,6 +217,8 @@ def forward(
         # is why it survived every seg-only run.
         onehot[pos] = F.one_hot(cls_t[pos], self.num_classes).to(onehot.dtype)
         partial_mask = None
+        if class_negative_mask is not None and negative_mask is None:
+            raise ValueError("class_negative_mask requires partial negative_mask")
         if negative_mask is not None:
             if negative_mask.ndim != 3 or negative_mask.shape[0] != flat_cls.shape[0]:
                 raise ValueError("partial detection negative_mask must be [B,H,W]")
@@ -216,6 +246,34 @@ def forward(
                     ).any(dim=1)
                     negative[b] &= ~in_box
             partial_mask = torch.maximum(onehot, negative[..., None].to(onehot.dtype))
+            if class_negative_mask is not None:
+                expected = (flat_cls.shape[0], self.num_classes, h, w)
+                if tuple(class_negative_mask.shape) != expected:
+                    raise ValueError("class_negative_mask must be [B,C,H,W] aligned with image")
+                cm = class_negative_mask.to(device)
+                if not ((cm == 0) | (cm == 1) | (cm == 255)).all():
+                    raise ValueError("class_negative_mask supports only 0/1/255")
+                selected = (
+                    cm[:, :, xy[:, 1].clamp(max=h - 1), xy[:, 0].clamp(max=w - 1)] == 1
+                ).permute(0, 2, 1)
+                selected &= inside[None, :, None]
+                for b, (boxes, labels) in enumerate(zip(boxes_list, labels_list, strict=True)):
+                    # Reviewed positives take precedence on every pyramid level,
+                    # including levels to which their regression was not assigned.
+                    for box, label in zip(boxes, labels, strict=True):
+                        inside_box = (
+                            (points[:, 0] >= box[0])
+                            & (points[:, 0] <= box[2])
+                            & (points[:, 1] >= box[1])
+                            & (points[:, 1] <= box[3])
+                        )
+                        selected[b, inside_box, label] = False
+                weight = (
+                    class_negative_weights(selected, onehot, partial_mask)
+                    if self.class_negative_normalization == "positive_budget"
+                    else selected.to(onehot.dtype)
+                )
+                partial_mask = torch.maximum(partial_mask, weight)
         if class_mask is not None:
             # [B, C] -> [B, 1, C] against flat_cls's [B, points, C]; a [C] mask
             # broadcasts as it is. Cast rather than assume: under autocast flat_cls is
diff --git a/src/syncai_hydranet/paths.py b/src/syncai_hydranet/paths.py
new file mode 100644
index 0000000..5b054cd
--- /dev/null
+++ b/src/syncai_hydranet/paths.py
@@ -0,0 +1,22 @@
+"""Where the tree is: one answer for every tool that reads `runs/` and `datasets/`.
+
+Every commissioning tool derived its root from its own file, so the chain could only ever
+run against the checkout it lived in. Commissioning a second site (FTI, 2026-09-10) meant
+copying `src/ tools/ scripts/` into a checkout-shaped directory and symlinking the
+datasets the hardcoded paths named. `scene_mesh.py` already read `SYNCAI_ROOT`; now the
+tools do too, through this one function, so a second site is an environment variable
+rather than a second copy of the code.
+"""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+
+ENV = "SYNCAI_ROOT"
+
+
+def repo_root(derived: Path) -> Path:
+    """`$SYNCAI_ROOT` when set, else the root the caller derived from its own location."""
+    override = os.environ.get(ENV)
+    return Path(override).resolve() if override else derived
diff --git a/src/syncai_hydranet/shipped.py b/src/syncai_hydranet/shipped.py
index ef05ceb..f81cbf7 100644
--- a/src/syncai_hydranet/shipped.py
+++ b/src/syncai_hydranet/shipped.py
@@ -44,7 +44,9 @@
 
 from pathlib import Path
 
-REPO = Path(__file__).resolve().parents[2]
+from syncai_hydranet.paths import repo_root
+
+REPO = repo_root(Path(__file__).resolve().parents[2])
 
 #: The run every tool ships from. A timestamped directory rather than a stable name on
 #: purpose: the name says when the weights were trained, and promoting a new run is an
diff --git a/tests/test_camera_json.py b/tests/test_camera_json.py
index b9965c2..10f2d70 100644
--- a/tests/test_camera_json.py
+++ b/tests/test_camera_json.py
@@ -285,3 +285,21 @@ def test_ground_points_rejects_a_policy_it_does_not_have():
 
     with pytest.raises(ValueError, match="above_horizon must be"):
         a_camera_file().ground_points(np.array([[960.0, 900.0]]), above_horizon="warn")
+
+
+def test_the_source_frame_survives_a_round_trip_and_is_absent_from_older_files(tmp_path):
+    """The boxes' own frame is part of the contract; a v3 file simply does not know it."""
+    cf = dataclasses.replace(a_camera_file(), source_size_px=(1280, 720))
+    path = tmp_path / "c.json"
+    cf.save(path)
+    back = CameraFile.load(path)
+    assert back.source_size_px == (1280, 720)
+    raw = json.loads(path.read_text())
+    del raw["source_size_px"]
+    path.write_text(json.dumps(raw))
+    assert CameraFile.load(path).source_size_px is None
+
+
+def test_a_source_frame_of_another_aspect_is_refused():
+    with pytest.raises(ValueError, match="aspect"):
+        dataclasses.replace(a_camera_file(), source_size_px=(1280, 960)).validate()
diff --git a/tests/test_commissioning.py b/tests/test_commissioning.py
index e3fd72e..82502b6 100644
--- a/tests/test_commissioning.py
+++ b/tests/test_commissioning.py
@@ -206,3 +206,11 @@ def test_height_only_update_refuses_another_camera(tmp_path):
     cam = _commissioned(tmp_path)
     with pytest.raises(ValueError, match="camera_id"):
         regeometry_from_calib(cam, write(tmp_path, camera="Another-camera"))
+
+
+def test_the_streams_native_size_reaches_the_contract_when_the_scan_recorded_it(tmp_path):
+    """Recorded by static_plates from the clip, carried by the calib, kept by camera.json;
+    a scan that did not record it converts as before."""
+    cf = from_onboard_calib(write(tmp_path, source_size_px=[1280, 720]))
+    assert cf.source_size_px == (1280, 720)
+    assert from_onboard_calib(write(tmp_path)).source_size_px is None
diff --git a/tests/test_detector_warmup.py b/tests/test_detector_warmup.py
new file mode 100644
index 0000000..65e03fb
--- /dev/null
+++ b/tests/test_detector_warmup.py
@@ -0,0 +1,82 @@
+"""A detector update must not mutate any scene weights, buffers or logits."""
+
+import importlib.util
+from pathlib import Path
+
+import pytest
+import torch
+
+from syncai_hydranet.models.hydranet import HydraNet
+
+
+@pytest.fixture
+def worker():
+    path = Path(__file__).resolve().parents[1] / "tools/annotation/studioa_train.py"
+    spec = importlib.util.spec_from_file_location("warmup_worker_test", path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def config():
+    return {
+        "model": {
+            "backbone": {"name": "resnet18", "pretrained": False},
+            "neck": {"name": "fpn", "out_channels": 16, "num_levels": 5},
+            "heads": {
+                "scene": {
+                    "type": "semantic_fpn",
+                    "num_classes": 3,
+                    "channels": 16,
+                    "dropout": 0.5,
+                },
+                "detection": {"type": "fcos", "num_classes": 2, "channels": 16, "num_convs": 1},
+            },
+            "loss_balancing": "fixed",
+            "detection_only_training": True,
+        }
+    }
+
+
+def test_detector_updates_preserve_scene_through_modes_and_reload(worker):
+    torch.set_num_threads(2)
+    model = HydraNet(config())
+    reference = worker.frozen_state(model)
+    before_head = model.det_head.cls_pred.weight.detach().clone()
+    trainable = [p for p in model.parameters() if p.requires_grad]
+    assert {id(p) for p in trainable} == {id(p) for p in model.det_head.parameters()}
+    optimizer = torch.optim.AdamW(trainable, lr=1e-3)
+    images = torch.randn(2, 3, 64, 96)
+    loaders = [("studioa", [{"image": images}])]
+    before = worker.scene_signature(model, loaders, "cpu")
+    for _ in range(3):
+        model.eval().train()
+        assert model.det_head.training
+        output = model(images)
+        targets = {
+            "boxes": [torch.tensor([[16.0, 16.0, 48.0, 48.0]])] * 2,
+            "labels": [torch.tensor([1])] * 2,
+            "det_negative_mask": torch.zeros(2, 64, 96, dtype=torch.long),
+        }
+        loss, _ = model.compute_losses(output, targets, ["detection"])
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+        worker.assert_frozen(model, reference)
+    assert not torch.equal(before_head, model.det_head.cls_pred.weight)
+    restored = HydraNet(config())
+    restored.load_state_dict(model.state_dict())
+    restored.eval().train()
+    worker.assert_frozen(restored, reference)
+    assert worker.scene_signature(restored, loaders, "cpu") == before
+    with torch.no_grad():
+        restored.neck.state_dict()[next(iter(restored.neck.state_dict()))].add_(1)
+    with pytest.raises(ValueError, match="frozen scene state changed"):
+        worker.assert_frozen(restored, reference)
+
+
+def test_warmup_without_detection_rejected():
+    cfg = config()
+    del cfg["model"]["heads"]["detection"]
+    with pytest.raises(ValueError, match="requires a detection head"):
+        HydraNet(cfg)
diff --git a/tests/test_onboard_camera.py b/tests/test_onboard_camera.py
new file mode 100644
index 0000000..ada2a56
--- /dev/null
+++ b/tests/test_onboard_camera.py
@@ -0,0 +1,81 @@
+"""A measured vfov reaches the calibration: the pin is data, the primary row follows it.
+
+`onboard_camera.py` pinned exactly one camera by name and swept the 70.4 constant as
+the primary for every other; `floor_calibrate.py`'s measured vfov had no consumer.
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import json
+import sys
+from pathlib import Path
+
+import pytest
+
+REPO = Path(__file__).resolve().parent.parent
+
+
+def _load():
+    name = "_onboard_camera"
+    spec = importlib.util.spec_from_file_location(name, REPO / "scripts" / "onboard_camera.py")
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[name] = module
+    spec.loader.exec_module(module)
+    return module
+
+
+ob = _load()
+
+
+def test_the_tile_pinned_camera_is_data_not_a_name_test():
+    pin = ob.pin_for("Taichung-cam01", None)
+    assert pin["vfov_deg"] == 70.4 and pin["source"] == "tile_grid_pinned"
+    assert ob.pin_for("Taichung-cam04", None) is None
+
+
+def test_a_pins_file_adds_measured_cameras_and_a_later_pin_wins(tmp_path):
+    path = tmp_path / "pins.json"
+    path.write_text(
+        json.dumps(
+            {
+                "FTI-MDF-cam821": {"vfov_deg": 62.75, "source": "floor_orthogonality"},
+                "Taichung-cam01": {"vfov_deg": 69.0, "source": "floor_orthogonality"},
+            }
+        )
+    )
+    pins = ob.load_pins(path)
+    assert pins["FTI-MDF-cam821"]["vfov_deg"] == 62.75
+    assert pins["Taichung-cam01"]["vfov_deg"] == 69.0  # the measurement over the built-in
+    assert ob.pin_for("Taichung-cam04", pins) is None
+
+
+def test_a_pin_without_a_vfov_is_refused(tmp_path):
+    path = tmp_path / "pins.json"
+    path.write_text(json.dumps({"X-cam01": {"source": "typo"}}))
+    with pytest.raises(ValueError, match="vfov_deg"):
+        ob.load_pins(path)
+
+
+def test_the_primary_row_is_the_cameras_own_vfov_and_a_failed_row_is_none():
+    rows = [
+        {"vfov_deg": 55.0, "pitch_deg": 34.9},
+        {"vfov_deg": 62.75, "pitch_deg": 39.1},
+        {"vfov_deg": 70.4, "failed": "no plausible floor plane"},
+    ]
+    assert ob.primary_row(rows, 62.75)["pitch_deg"] == 39.1
+    assert ob.primary_row(rows, 70.4) is None
+    assert ob.primary_row(rows, 85.0) is None
+
+
+def test_the_plate_indexs_source_size_reaches_the_calib(tmp_path):
+    root = tmp_path / "plates"
+    root.mkdir()
+    (root / "index.json").write_text(
+        json.dumps(
+            {"cameras": {"X-cam01": {"slots": {"20260910-150000": {"source_wh": [1280, 720]}}}}}
+        )
+    )
+    assert ob.source_size_from_index(root, "X-cam01", "20260910-150000") == [1280, 720]
+    assert ob.source_size_from_index(root, "X-cam01", "20260910-160000") is None
+    assert ob.source_size_from_index(tmp_path / "nowhere", "X-cam01", "s") is None
diff --git a/tests/test_partial_detection.py b/tests/test_partial_detection.py
index f234a4b..990d443 100644
--- a/tests/test_partial_detection.py
+++ b/tests/test_partial_detection.py
@@ -7,6 +7,76 @@
 from syncai_hydranet.models.losses import FCOSLoss
 
 
+def test_class_negative_budget_bounds_mass_and_excludes_existing_supervision():
+    from syncai_hydranet.models.losses import class_negative_weights
+
+    positive = torch.zeros(2, 20, 3)
+    positive[0, :3, 0] = 1
+    existing = positive.clone()
+    existing[:, -2:, :] = 1  # existing empty negatives must not spend the added budget
+    selected = torch.ones_like(positive, dtype=torch.bool)
+    weight = class_negative_weights(selected, positive, existing)
+    torch.testing.assert_close(weight.sum((0, 1)), torch.tensor([3.0, 1.0, 1.0]))
+    assert (weight[existing > 0] == 0).all()
+    assert not weight.requires_grad
+
+
+def test_negative_only_area_replication_keeps_bounded_loss_and_total_gradient():
+    results = []
+    for area in (8, 16):
+        head, cls, reg, ctr, _, _, negative = batch()
+        negative.zero_()
+        cm = torch.zeros(1, 3, 32, 32, dtype=torch.uint8)
+        cm[:, 0, :area, :area] = 1
+        loss, _ = FCOSLoss(3, class_negative_normalization="positive_budget")(
+            head,
+            cls,
+            reg,
+            ctr,
+            [torch.zeros(0, 4)],
+            [torch.zeros(0, dtype=torch.long)],
+            negative_mask=negative,
+            class_negative_mask=cm,
+        )
+        loss.backward()
+        results.append((loss.detach(), sum(x.grad.sum() for x in cls)))
+        assert all((x.grad[:, 1:] == 0).all() for x in cls)
+        assert all((x.grad == 0).all() for x in reg + ctr)
+    torch.testing.assert_close(results[0], results[1])
+
+
+def test_budget_preserves_positive_and_empty_gradients_and_only_scales_added_negatives():
+    gradients = []
+    for mode in ("sum", "positive_budget"):
+        head, cls, reg, ctr, boxes, labels, negative = batch()
+        cm = torch.zeros(1, 3, 32, 32, dtype=torch.uint8)
+        cm[:, 0] = 1
+        cm[:, 1, 8:24, 8:24] = 1  # contradictory same-class signal remains protected
+        cm[:, 2, 24:] = 255
+        FCOSLoss(3, class_negative_normalization=mode)(
+            head,
+            cls,
+            reg,
+            ctr,
+            boxes,
+            labels,
+            negative_mask=negative,
+            class_negative_mask=cm,
+        )[0].backward()
+        gradients.append(([x.grad.clone() for x in cls], [x.grad.clone() for x in reg + ctr]))
+    a, b = gradients
+    torch.testing.assert_close(a[1], b[1])
+    for x, y in zip(a[0], b[0], strict=True):
+        torch.testing.assert_close(x[:, 1:], y[:, 1:])
+    torch.testing.assert_close(a[0][0][:, :, 0, 0], b[0][0][:, :, 0, 0])
+    assert 0 < b[0][0][0, 0, 0, 1] < a[0][0][0, 0, 0, 1]
+
+
+def test_bad_class_negative_normalization_rejected():
+    with pytest.raises(ValueError, match="normalization"):
+        FCOSLoss(3, class_negative_normalization="invalid")
+
+
 def batch():
     head = FCOSHead(8, 3, in_levels=[0, 1], channels=8, num_convs=1, strides=[8, 16])
     cls = [
@@ -44,6 +114,55 @@ def test_partial_focal_unknown_channels_padding_and_regression_gradients():
     assert (cls[1].grad == 0).all()  # small box out of second level's range
 
 
+def test_class_negative_only_affects_reviewed_channel_and_protects_positive_all_levels():
+    head, cls, reg, ctr, boxes, labels, negative = batch()
+    negative.zero_()
+    per_class = torch.zeros(1, 3, 32, 32, dtype=torch.uint8)
+    per_class[:, 0, :8, 8:16] = 1
+    per_class[:, 0, 8:16, 8:16] = 1  # explicit wrong-class negative on another positive
+    per_class[:, 1, 8:24, 8:24] = 1  # contradictory positive is protected at every level
+    per_class[:, 2, 24:] = 255  # padding is not a negative
+    FCOSLoss(3)(
+        head,
+        cls,
+        reg,
+        ctr,
+        boxes,
+        labels,
+        negative_mask=negative,
+        class_negative_mask=per_class,
+    )[0].backward()
+    g = cls[0].grad[0]
+    assert g[0, 0, 1] > 0
+    assert (g[1:, 0, 1] == 0).all()
+    assert g[1, 1, 1] < 0
+    assert g[0, 1, 1] > 0 and g[2, 1, 1] == 0
+    assert (g[:, 3, :] == 0).all()
+    assert (cls[1].grad[:, 1:] == 0).all()
+    assert cls[1].grad[0, 0, 0, 0] > 0
+    assert (reg[0].grad[:, :, 0, :] == 0).all()
+    assert (ctr[0].grad[:, :, 0, :] == 0).all()
+
+
+@pytest.mark.parametrize(
+    "bad", [torch.zeros(1, 32, 32), torch.zeros(1, 2, 32, 32), torch.full((1, 3, 32, 32), 2)]
+)
+def test_invalid_class_negatives_rejected(bad):
+    head, cls, reg, ctr, boxes, labels, negative = batch()
+    with pytest.raises(ValueError, match="class_negative_mask"):
+        FCOSLoss(3)(
+            head, cls, reg, ctr, boxes, labels, negative_mask=negative, class_negative_mask=bad
+        )
+
+
+def test_class_negative_cannot_turn_exhaustive_training_into_partial():
+    head, cls, reg, ctr, boxes, labels, _ = batch()
+    with pytest.raises(ValueError, match="requires partial"):
+        FCOSLoss(3)(
+            head, cls, reg, ctr, boxes, labels, class_negative_mask=torch.zeros(1, 3, 32, 32)
+        )
+
+
 def test_box_never_becomes_negative_on_an_unassigned_level():
     head, cls, reg, ctr, boxes, labels, negative = batch()
     negative.fill_(1)
diff --git a/tests/test_partial_detection_eval.py b/tests/test_partial_detection_eval.py
index 41de294..d93f7f2 100644
--- a/tests/test_partial_detection_eval.py
+++ b/tests/test_partial_detection_eval.py
@@ -121,6 +121,10 @@ def test_joint_config_never_declares_test_and_warm_start_checks_taxonomy(tmp_pat
 
     cfg = worker.joint_config(copy.deepcopy(original), tmp_path, "Tao-Hsin")
     assert all("split_test" not in ds for ds in cfg["data"]["datasets"])
+    warmup = worker.detector_warmup_config(copy.deepcopy(cfg))
+    assert warmup["model"]["detection_only_training"]
+    assert warmup["data"]["datasets"][0]["validation_only"]
+    assert warmup["train"]["primary_metric"] == worker.DET_METRIC
     model = HydraNet(cfg)
     worker.warm_start(model, {"cfg": original, "model": source.state_dict()}, cfg)
     torch.testing.assert_close(
diff --git a/tests/test_plate_calibration.py b/tests/test_plate_calibration.py
index 23db7c5..d7eee59 100644
--- a/tests/test_plate_calibration.py
+++ b/tests/test_plate_calibration.py
@@ -622,3 +622,48 @@ def test_a_depth_frame_with_nothing_in_it_yields_no_candidates():
     assert floor_candidates(empty, cam, inlier_m=0.03) == []
     plane, residual, rows = choose_floor([])
     assert plane is None and residual is None and rows == []
+
+
+def _plate(dir_, slot, luma):
+    from PIL import Image
+
+    Image.new("L", (4, 4), luma).save(dir_ / f"plate_{slot}.png")
+
+
+def test_the_daytime_gate_uses_the_sites_offset_not_a_constant(tmp_path):
+    """A US-Central site (UTC-5) recorded at 15:00 UTC is 10:00 local -- daytime -- and the
+    same plates read under +8 pick another slot. The old constant refused the 15 UTC slot
+    (23 local) and passed a 10 UTC slot by luck (18 local)."""
+    from syncai_bev3d.plate_calibration import pick_daytime_slot
+
+    _plate(tmp_path, "20260910-150000", 200)
+    _plate(tmp_path, "20260910-030000", 250)  # 22:00 local at -5: brighter, but night
+    assert pick_daytime_slot(tmp_path, utc_offset=-5) == "20260910-150000"
+    assert pick_daytime_slot(tmp_path, utc_offset=8) == "20260910-030000"
+
+
+def test_a_cameras_json_without_an_offset_still_means_the_original_site():
+    from syncai_bev3d.plate_calibration import DEFAULT_UTC_OFFSET_HOURS, utc_offset_hours
+
+    assert utc_offset_hours({"cameras": {}}) == DEFAULT_UTC_OFFSET_HOURS == 8
+    assert utc_offset_hours({"utc_offset_hours": -5, "cameras": {}}) == -5
+
+
+def test_a_night_shift_site_states_its_own_lit_hours(tmp_path):
+    """FTI's only capture is 05:22 local (UTC-6, slot 11 UTC): lit, and refused by a shop's
+    08-18 window. The site's cameras.json widens the window; a window across midnight
+    is stated as [start, end] with start > end."""
+    from syncai_bev3d.plate_calibration import (
+        DEFAULT_DAYTIME_HOURS_LOCAL,
+        daytime_hours_local,
+        in_daytime,
+        pick_daytime_slot,
+    )
+
+    _plate(tmp_path, "20260910-112151", 149)
+    with pytest.raises(SystemExit, match="no daytime plate"):
+        pick_daytime_slot(tmp_path, utc_offset=-6)
+    assert pick_daytime_slot(tmp_path, utc_offset=-6, daytime=(4, 23)) == "20260910-112151"
+    assert daytime_hours_local({"cameras": {}}) == DEFAULT_DAYTIME_HOURS_LOCAL == (8, 18)
+    assert daytime_hours_local({"daytime_hours_local": [4, 23], "cameras": {}}) == (4, 23)
+    assert in_daytime(23, (20, 6)) and in_daytime(2, (20, 6)) and not in_daytime(12, (20, 6))
diff --git a/tests/test_rulers.py b/tests/test_rulers.py
index 3768a4d..ee29eb1 100644
--- a/tests/test_rulers.py
+++ b/tests/test_rulers.py
@@ -85,3 +85,32 @@ def test_two_rulers_on_opposite_sides_do_not_apply():
 def test_no_ruler_keeps():
     v = combine([])
     assert v.factor == 1.0 and not v.apply
+
+
+def test_a_bootstrap_calibration_has_no_person_ruler_to_anchor_to():
+    """Factor 1 on a scale that was never measured would anchor every other ruler to the
+    reading they exist to check."""
+    assert person_ruler({"scale_source": "unmeasured"}) is None
+    assert person_ruler({"scale_source": "dav2_metric_indoor_raw_bootstrap_unverified"}) is None
+    assert person_ruler({"scale_source": "person_height_median_vs_1.7m_prior_n38"}) is not None
+
+
+def test_unanchored_rulers_agreeing_is_not_a_verdict():
+    """FTI MDF room, 2026-09-10: a 600 mm raised floor read 0.85 and 0.74 m through DA-V2's
+    raw metres, the catalogue picked 0.80, and the two cameras 'agreed'. The consensus tile
+    and the store-median table both take their reference from the readings themselves."""
+    tile = tile_ruler(0.85, 0.80, anchored=False)
+    table = table_ruler(0.94, [0.94, 0.90, 0.92])
+    v = combine([tile, table])
+    assert not v.apply
+    assert v.factor == 1.0
+    assert "unanchored" in v.reason
+
+
+def test_an_anchored_ruler_restores_the_two_witness_rule():
+    """With the person prior present, an unanchored tile can still be the second witness."""
+    person = Ruler("person", 1.0, 0.14)
+    tile = tile_ruler(0.85, 0.60, anchored=False)  # x0.706, a real size this time
+    known = Ruler("folding_table", 0.74 / 0.94, 0.05, "0.94 m read vs 0.74 m")
+    v = combine([person, tile, known])
+    assert v.apply and abs(v.factor - 0.706) < 0.1
diff --git a/tests/test_stage0_baseline.py b/tests/test_stage0_baseline.py
new file mode 100644
index 0000000..9163135
--- /dev/null
+++ b/tests/test_stage0_baseline.py
@@ -0,0 +1,56 @@
+"""The Stage 0 freeze carries every file `capture_inputs` identifies, code included.
+
+`freeze()` copied `src/`, `scene_mesh.py`, `pyproject.toml` and `uv.lock` by name while
+`capture_inputs` also listed `rebuild_geometry.py`; the identity check at the end of the
+freeze then failed on every camera (FTI-SMT-cam912, 2026-09-16), and the runner never
+reached a scene. The copy now walks the identity record, so the two cannot drift apart.
+"""
+
+from __future__ import annotations
+
+import math
+import runpy
+from pathlib import Path
+
+import numpy as np
+from PIL import Image
+
+from syncai_bev3d.scene_audit import capture_inputs
+from syncai_hydranet.geometry.camera_json import CameraFile
+from syncai_hydranet.geometry.ground import Camera, GroundPlane
+
+TOOL = Path(__file__).resolve().parents[1] / "tools/commissioning/stage0_baseline.py"
+
+
+def _site(root: Path) -> None:
+    folder = root / "runs/commission01/sample/masks"
+    folder.mkdir(parents=True)
+    cache = root / "runs/site30k_qa/geometry_cache/sample.npz"
+    cache.parent.mkdir(parents=True)
+    np.savez(cache, horiz=np.ones((48, 64)))
+    CameraFile(
+        camera_id="sample",
+        image_size_px=(64, 48),
+        camera=Camera(fx=42, fy=42, cx=32, cy=24),
+        plane=GroundPlane(height=2.8, pitch=math.radians(25)),
+        mask_files={"objects": "sample/masks/objects.png"},
+    ).save(root / "runs/commission01/sample.camera.json")
+    Image.new("L", (64, 48)).save(folder / "objects.png")
+    (root / "src/syncai_bev3d").mkdir(parents=True)
+    (root / "src/syncai_bev3d/scene.py").write_text("SCENE = 1\n")
+    for name in ("scene_mesh.py", "rebuild_geometry.py"):
+        (root / "tools/commissioning" / name).parent.mkdir(parents=True, exist_ok=True)
+        (root / "tools/commissioning" / name).write_text(f"# {name}\n")
+    (root / "pyproject.toml").write_text("[project]\nname = 'sample'\n")
+    (root / "uv.lock").write_text("version = 1\n")
+
+
+def test_the_snapshot_identifies_as_the_source_it_was_frozen_from(tmp_path):
+    freeze = runpy.run_path(str(TOOL))["freeze"]
+    root, out = tmp_path / "site", tmp_path / "out"
+    _site(root)
+    out.mkdir()
+    snapshot = freeze(root, out, ["sample"])
+    assert (snapshot / "tools/commissioning/rebuild_geometry.py").is_file()
+    assert capture_inputs(snapshot, "sample") == capture_inputs(root, "sample")
+    assert (out / "frozen-inputs.json").is_file()
diff --git a/tests/test_stage0_commission.py b/tests/test_stage0_commission.py
new file mode 100644
index 0000000..92c77a1
--- /dev/null
+++ b/tests/test_stage0_commission.py
@@ -0,0 +1,104 @@
+"""The one-camera Stage 0 chain: the README's steps, the bootstrap rule, the site root."""
+
+from __future__ import annotations
+
+import json
+import runpy
+from pathlib import Path
+
+import pytest
+
+TOOL = Path(__file__).resolve().parents[1] / "tools/commissioning/stage0_commission.py"
+
+
+@pytest.fixture(scope="module")
+def tool():
+    return runpy.run_path(str(TOOL))
+
+
+def test_the_slot_is_the_archive_name_or_the_files_utc_mtime(tool, tmp_path):
+    named = tmp_path / "FTI-SMT-cam912_archive_20260910-112151_fti.mp4"
+    named.write_bytes(b"")
+    assert tool["slot_from_clip"](named) == "20260910-112151"
+    plain = tmp_path / "nvr9_ch12.mp4"
+    plain.write_bytes(b"")
+    assert len(tool["slot_from_clip"](plain)) == len("20260910-112151")
+
+
+def _calib(tmp_path, **fields):
+    p = tmp_path / "cam.calib.json"
+    d = {
+        "pitch_deg": 39.6,
+        "height_dav2_raw_m": 3.1,
+        "scale_source": "unmeasured",
+        "flags": ["scale_unmeasured"],
+    }
+    d.update(fields)
+    p.write_text(json.dumps(d))
+    return p
+
+
+def test_no_floor_is_a_refusal_and_bootstrap_needs_permission(tool, tmp_path):
+    with pytest.raises(SystemExit, match="2"):
+        tool["bootstrap_unmeasured"](_calib(tmp_path, pitch_deg=None), allow=True)
+    with pytest.raises(SystemExit, match="3"):
+        tool["bootstrap_unmeasured"](_calib(tmp_path), allow=False)
+
+
+def test_the_bootstrap_is_flagged_and_a_measured_scale_is_left_alone(tool, tmp_path):
+    p = _calib(tmp_path)
+    assert tool["bootstrap_unmeasured"](p, allow=True).startswith(
+        "dav2_metric_indoor_raw_bootstrap"
+    )
+    d = json.loads(p.read_text())
+    assert d["scale"] == 1.0 and d["height_m"] == 3.1
+    assert tool["BOOTSTRAP_FLAG"] in d["flags"] and "scale_unmeasured" not in d["flags"]
+    p = _calib(tmp_path, scale_source="person_height_median_vs_1.7m_prior_n38", flags=[])
+    assert (
+        tool["bootstrap_unmeasured"](p, allow=False) == "person_height_median_vs_1.7m_prior_n38"
+    )
+    assert "scale" not in json.loads(p.read_text())
+
+
+def test_the_plan_is_the_readme_chain_ending_in_a_freeze(tool, tmp_path):
+    steps = tool["plan"](
+        tmp_path,
+        "cam",
+        Path("datasets/fti_clips"),
+        Path("datasets/fti_static"),
+        tmp_path / "out",
+    )
+    labels = [label for label, _ in steps]
+    assert labels == [
+        "plate",
+        "person boxes",
+        "onboard",
+        "zones",
+        "review bundle",
+        "masks",
+        "extras",
+        "depth completion",
+        "scene3d",
+        "scene overlay",
+        "scene mesh",
+        "freeze",
+    ]
+    freeze = dict(steps)["freeze"]
+    bundle = dict(steps)["review bundle"]
+    assert bundle[-1] == "runs/commission_review/cam_batch"
+    stamped = dict(tool["plan"](tmp_path, "cam", Path("c"), Path("p"), None, "20260910-102225"))
+    assert stamped["review bundle"][-1] == "runs/commission_review/cam_20260910-102225"
+    assert freeze[1].endswith("stage0_baseline.py") and str(tmp_path) in freeze
+    assert len(tool["plan"](tmp_path, "cam", Path("c"), Path("p"), None)) == 11
+
+
+def test_dry_run_prints_the_chain_without_touching_the_tree(
+    tool, tmp_path, capsys, monkeypatch
+):
+    clip = tmp_path / "nvr9_ch12.mp4"
+    clip.write_bytes(b"")
+    monkeypatch.setitem(tool, "ROOT", tmp_path)
+    assert tool["main"](["cam", "--clip", str(clip), "--dry-run"]) == 0
+    out = capsys.readouterr().out
+    assert out.count("[") == 11 and "[scene mesh]" in out
+    assert not (tmp_path / "datasets").exists()
diff --git a/tests/test_studioa_supervision.py b/tests/test_studioa_supervision.py
index 16969cd..b225ef8 100644
--- a/tests/test_studioa_supervision.py
+++ b/tests/test_studioa_supervision.py
@@ -133,6 +133,9 @@ def test_reviewed_instances_export_transform_factory_and_boundaries(tmp_path):
                     {"id": data["entities"][0]["id"], "entity": "phone", "reason": "fixture"}
                 ],
                 "negative_rects": [{"xyxy": [32, 0, 40, 8], "reason": "fixture empty patch"}],
+                "class_negative_rects": [
+                    {"entity": "laptop", "xyxy": [0, 0, 8, 8], "reason": "phone is not laptop"}
+                ],
             }
         ],
     }
@@ -159,6 +162,10 @@ def test_reviewed_instances_export_transform_factory_and_boundaries(tmp_path):
     assert sample["targets"]["boxes"].tolist() == [[10.0, 0.0, 40.0, 24.0]]
     assert (sample["targets"]["det_negative_mask"][:8, :8] == 1).all()
     assert sample["targets"]["labels"].tolist() == [1]
+    class_mask = sample["targets"]["det_class_negative_mask"]
+    assert class_mask.shape == (10, 32, 40)
+    assert (class_mask[0, :8, 32:] == 1).all()  # flipped with image and boxes
+    assert (class_mask[1:] == 0).all()
     assert fingerprint_dataset(cfg)["splits"]["train"]["frames"] == [f["id"]]
     assert split_leaks([cfg]) == []
     with pytest.raises(ValueError, match="unchanged splits"):
@@ -176,6 +183,20 @@ def test_reviewed_instances_export_transform_factory_and_boundaries(tmp_path):
         check_instances(out)
 
 
+def test_class_negative_review_rejects_contradiction_and_val():
+    from syncai_hydranet.data.studioa_instances import class_negative_regions
+
+    item = {
+        "class_negative_rects": [{"entity": "phone", "xyxy": [1, 1, 4, 4], "reason": "fixture"}]
+    }
+    with pytest.raises(ValueError, match="contradicts"):
+        class_negative_regions(item, (10, 10), [[0, 0, 5, 5]], [1], "train")
+    with pytest.raises(ValueError, match="train-only"):
+        class_negative_regions(item, (10, 10), [], [], "val")
+    masks = class_negative_regions(item, (10, 10), [[0, 0, 5, 5]], [0], "train")
+    assert list(masks) == ["phone"] and masks["phone"].sum() == 9
+
+
 def test_real_export_reader_preserves_folds_source_labels_and_ignore(tmp_path):
     source, out = tmp_path / "source", tmp_path / "out"
     frames = source_package(source)
diff --git a/tests/test_tools_honour_syncai_root.py b/tests/test_tools_honour_syncai_root.py
new file mode 100644
index 0000000..e898210
--- /dev/null
+++ b/tests/test_tools_honour_syncai_root.py
@@ -0,0 +1,39 @@
+"""Every tool in the commissioning chain takes its tree root from `repo_root()`.
+
+A tool whose root is only `Path(__file__)...` runs against the checkout it lives in and
+nowhere else; commissioning a second site then means a second copy of the code
+(FTI, 2026-09-10). `SYNCAI_ROOT` is the one override, read in one place.
+"""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+import pytest
+
+REPO = Path(__file__).resolve().parents[1]
+FILES = sorted(
+    list((REPO / "tools" / "commissioning").glob("*.py"))
+    + list((REPO / "tools" / "site30k").glob("*.py"))
+    + [REPO / "scripts" / "propose_zones.py", REPO / "scripts" / "campaign_site30k.py"]
+)
+DERIVED = re.compile(r"^(?:ROOT|REPO|_REPO)\s*=\s*Path\(__file__\)", re.M)  # no override
+
+
+@pytest.mark.parametrize("path", FILES, ids=lambda p: p.name)
+def test_the_root_is_read_through_repo_root(path):
+    text = path.read_text()
+    assert not DERIVED.search(text), (
+        f"{path.relative_to(REPO)} derives its root from its own file with no override; "
+        'write ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))'
+    )
+
+
+def test_the_override_is_the_environment(monkeypatch, tmp_path):
+    from syncai_hydranet.paths import repo_root
+
+    monkeypatch.delenv("SYNCAI_ROOT", raising=False)
+    assert repo_root(tmp_path) == tmp_path
+    monkeypatch.setenv("SYNCAI_ROOT", str(tmp_path / "site2"))
+    assert repo_root(tmp_path) == (tmp_path / "site2").resolve()
diff --git a/tests/test_trainer.py b/tests/test_trainer.py
index c041b84..f38ef8d 100644
--- a/tests/test_trainer.py
+++ b/tests/test_trainer.py
@@ -96,6 +96,25 @@ def _cfg(out_dir, data_root, **train):
     )
 
 
+def test_validation_only_dataset_never_opens_train_and_is_recorded(tmp_path, data_root):
+    cfg = _cfg(tmp_path / "validation_only_run", data_root)
+    validation = dict(cfg["data"]["datasets"][0])
+    validation.update(name="scene_guard", validation_only=True, split_train="must_not_open")
+    cfg["data"]["datasets"].insert(0, validation)
+    trainer = Trainer(cfg)
+    try:
+        assert trainer.steps_per_epoch == 2
+        assert [name for name, _ in trainer.val_sets] == ["scene_guard", "tiny"]
+        meta = json.loads((trainer.out_dir / "meta.json").read_text())
+        assert meta["datasets"][0]["train_size"] == 0
+        assert meta["datasets"][0]["validation_only"]
+        assert meta["datasets"][0]["val_size"] == 2
+        assert meta["datasets"][1]["train_size"] == 4
+    finally:
+        if trainer.tb:
+            trainer.tb.close()
+
+
 @pytest.fixture(scope="module")
 def finished_run(tmp_path_factory, data_root):
     out = tmp_path_factory.mktemp("run") / "exp"
diff --git a/tests/test_world_frame.py b/tests/test_world_frame.py
index 8a5f609..5141e68 100644
--- a/tests/test_world_frame.py
+++ b/tests/test_world_frame.py
@@ -504,3 +504,14 @@ def test_a_canvas_region_without_a_source_size_is_refused():
     track, region = _canvas_track((0.5, 3.0))
     with pytest.raises(ValueError, match="without source_size_px"):
         world_frame([track], _half_res_camera_file(), 0, name="person", canvas_region=region)
+
+
+def test_a_track_box_is_drawn_by_the_panel_over_source_ratio_not_a_constant():
+    """`demo_video` drew boxes with `/ 2.0`: right for 1920x1080, 2/3 off on 1280x720 and
+    2x on 3840x2160. The ratio is per axis and per frame pair."""
+    from syncai_hydranet.analytics.world import panel_box
+
+    box = (640.0, 360.0, 700.0, 500.0)
+    assert panel_box(box, (1920, 1080), (960, 540)) == (320.0, 180.0, 350.0, 250.0)
+    assert panel_box(box, (1280, 720), (960, 540)) == (480.0, 270.0, 525.0, 375.0)
+    assert panel_box(box, (3840, 2160), (960, 540)) == (160.0, 90.0, 175.0, 125.0)
diff --git a/tools/annotation/studioa_instances.py b/tools/annotation/studioa_instances.py
index 140e35b..8f9673d 100644
--- a/tools/annotation/studioa_instances.py
+++ b/tools/annotation/studioa_instances.py
@@ -186,10 +186,17 @@ def stop(_signum, _frame):
                         & (points[:, None, 1] >= boxes[None, :, 1])
                         & (points[:, None, 1] <= boxes[None, :, 3])
                     ).any(dim=1)
-                    unknown = ~known_empty & ~in_any_box
+                    unknown = (
+                        (~known_empty & ~in_any_box)[:, None].expand(-1, len(CLASSES)).clone()
+                    )
+                    class_negative = targets.get("det_class_negative_mask")
+                    if class_negative is not None:
+                        reviewed = (class_negative[b, :, safe_y, safe_x] == 1).T
+                        reviewed &= ((x < w) & (y < h))[:, None]
+                        unknown &= ~reviewed
                     if torch.count_nonzero(grads[b, unknown]):
                         raise ValueError("unreviewed region produced background gradient")
-                    unknown_channels += int(unknown.sum()) * len(CLASSES)
+                    unknown_channels += int(unknown.sum())
                 optimizer.step()
                 steps.append(
                     {
diff --git a/tools/annotation/studioa_train.py b/tools/annotation/studioa_train.py
index d89f804..7afaf21 100644
--- a/tools/annotation/studioa_train.py
+++ b/tools/annotation/studioa_train.py
@@ -5,6 +5,7 @@
 
 import argparse
 import fcntl
+import hashlib
 import json
 import os
 import shutil
@@ -23,6 +24,80 @@
 from syncai_hydranet.data.studioa_supervision import CLASSES, check_supervision
 
 ROOT = Path(__file__).resolve().parents[2]
+DET_METRIC = "partial_det/studioa_instances/recall50_at_020"
+EMPTY_FP = "partial_det/studioa_instances/empty_fp"
+
+
+def detector_warmup_config(config: dict) -> dict:
+    config["experiment"] = "studioa_detector_warmup"
+    config["model"]["detection_only_training"] = True
+    config["model"]["fixed_weights"]["detection"] = 1.0
+    config["data"]["datasets"][0]["validation_only"] = True
+    config["train"].update(
+        epochs=60,
+        lr=2e-4,
+        warmup_iters=26,
+        early_stop_patience=20,
+        primary_metric=DET_METRIC,
+        deterministic=True,
+        cudnn_benchmark=False,
+        tf32=False,
+    )
+    check_config(config)
+    return config
+
+
+def frozen_state(model) -> dict:
+    return {
+        name: value.detach().cpu().clone()
+        for name, value in model.state_dict().items()
+        if not name.startswith("det_head.")
+    }
+
+
+def assert_frozen(model, reference: dict) -> None:
+    import torch
+
+    current = frozen_state(model)
+    if current.keys() != reference.keys():
+        raise ValueError("frozen scene state keys changed")
+    for name, value in reference.items():
+        if not torch.equal(current[name], value):
+            raise ValueError(f"frozen scene state changed: {name}")
+    for name, module in model.named_children():
+        if name != "det_head" and any(m.training for m in module.modules()):
+            raise ValueError(f"frozen scene module in training mode: {name}")
+
+
+def scene_signature(model, loaders, device) -> dict:
+    """Hash every source-validation scene logit, not just a preview or argmax."""
+    import torch
+
+    from syncai_hydranet.engine.evaluator import model_memory_format
+
+    was_training = model.training
+    model.eval()
+    signature = hashlib.sha256()
+    frames = 0
+    try:
+        with torch.inference_mode():
+            for name, loader in loaders:
+                if name != "studioa":
+                    continue
+                for batch in loader:
+                    images = (
+                        batch["image"]
+                        .to(device)
+                        .contiguous(memory_format=model_memory_format(model))
+                    )
+                    logits = model(images)["scene"].contiguous().cpu()
+                    signature.update(logits.numpy().tobytes())
+                    frames += len(images)
+    finally:
+        model.train(was_training)
+    if not frames:
+        raise ValueError("scene invariance requires source validation images")
+    return {"frames": frames, "float32_logits_sha256": signature.hexdigest()}
 
 
 def pilot_config(data: Path, out: Path, manifest: dict, held_out: str) -> dict:
@@ -155,6 +230,8 @@ def prepare(
     held_out: str,
     instances: Path | None = None,
     initial_checkpoint: Path | None = None,
+    detector_warmup: bool = False,
+    class_negative_normalization: str = "sum",
 ) -> None:
     from syncai_hydranet.utils.visualize import terrain_palette
 
@@ -164,6 +241,12 @@ def prepare(
         raise ValueError("pilot requires positive train support for every class")
     if (instances is None) != (initial_checkpoint is None):
         raise ValueError("joint pilot requires both instances and initial checkpoint")
+    if detector_warmup and instances is None:
+        raise ValueError("detector warmup requires instances and initial checkpoint")
+    if class_negative_normalization not in ("sum", "positive_budget"):
+        raise ValueError("unsupported class negative normalization")
+    if class_negative_normalization != "sum" and not detector_warmup:
+        raise ValueError("class negative normalization comparison requires detector warmup")
     instance_counts = {}
     if instances is not None:
         from collections import Counter
@@ -217,6 +300,12 @@ def prepare(
     config = pilot_config(out / "data", out, manifest, held_out)
     if instances is not None:
         config = joint_config(config, out / "instances", held_out)
+    if detector_warmup:
+        config = detector_warmup_config(config)
+        config["model"]["heads"]["detection"]["loss"] = {
+            "class_negative_normalization": class_negative_normalization
+        }
+        check_config(config)
     write_json(out / "config.json", config)
     git = subprocess.run(
         ["git", "rev-parse", "HEAD"], cwd=ROOT, capture_output=True, text=True, check=True
@@ -238,6 +327,8 @@ def prepare(
             if instances is not None
             else "studioa.semantic-pilot.v1",
             "joint": instances is not None,
+            "detector_warmup": detector_warmup,
+            "class_negative_normalization": class_negative_normalization,
             "instance_counts": instance_counts,
             "git_commit": git,
             "held_out": held_out,
@@ -245,7 +336,17 @@ def prepare(
             "source_manifest_sha256": digest(source / "manifest.json"),
             "classes": CLASSES,
             "class_weights": "sqrt median/train frequency clipped 0.25..4",
-            "selection": "source-camera val mIoU on retained AI pixels only",
+            "selection": (
+                "maximum reviewed-positive source-val recall at fixed score >0.20 / IoU >=0.50"
+                if detector_warmup
+                else "source-camera val mIoU on retained AI pixels only"
+            ),
+            "acceptance": (
+                "strict recall improvement; no increase in reviewed-empty FP; frozen scene "
+                "parameters/buffers and all source-val float32 logits byte-identical"
+                if detector_warmup
+                else None
+            ),
             "evaluation": (
                 "source val only; partial recall and empty-region alarms; no test evaluation"
                 if instances is not None
@@ -301,6 +402,8 @@ def stopped(_signal, _frame):
             raise RuntimeError("CUDA unavailable; refusing accidental CPU pilot training")
         torch.set_num_threads(4)
         cfg = json.loads((out / "config.json").read_text())
+        warmup = job.get("detector_warmup", False)
+        reference = None
 
         class PilotTrainer(Trainer):
             def state_dict(self, epoch):
@@ -309,24 +412,44 @@ def state_dict(self, epoch):
                 return result
 
             def record_epoch(self, epoch, metrics):
+                if warmup:
+                    assert reference is not None
+                    assert_frozen(self.model, reference)
                 improved = super().record_epoch(epoch, metrics)
                 status(
                     "training",
                     epoch=epoch,
                     global_step=self.global_step,
                     validation_teacher_miou=metrics["scene_mIoU"],
-                    best_validation_teacher_miou=self.best_metric,
+                    primary_metric=self.primary_metric,
+                    best_primary_metric=self.best_metric,
                 )
                 return improved
 
         last = out / "model/last.pt"
         trainer = PilotTrainer(cfg, resuming=last.exists())
+        if warmup:
+            warm_start(trainer.model, load_checkpoint(out / "initial.pt"), cfg)
+            reference = frozen_state(trainer.model)
+            initial_signature = scene_signature(
+                trainer.model, trainer.val_loaders, trainer.device
+            )
+            before_path = out / "scene_before.json"
+            if (
+                before_path.exists()
+                and json.loads(before_path.read_text()) != initial_signature
+            ):
+                raise ValueError("resumed scene reference differs from original run")
+            write_json(before_path, initial_signature)
         if last.exists():
             if load_checkpoint(last).get("studioa_job_sha256") != job_hash:
                 raise ValueError("foreign pilot checkpoint")
             trainer.load(str(last), resume=True)
+            if warmup:
+                assert reference is not None
+                assert_frozen(trainer.model, reference)
         else:
-            if job.get("joint"):
+            if job.get("joint") and not warmup:
                 warm_start(trainer.model, load_checkpoint(out / "initial.pt"), cfg)
             status("baseline_validation", gpu=torch.cuda.get_device_name(0))
             baseline = evaluate(
@@ -354,6 +477,24 @@ def record_epoch(self, epoch, metrics):
         if best.get("studioa_job_sha256") != job_hash:
             raise ValueError("foreign best checkpoint")
         trainer.model.load_state_dict(best["model"])
+        warmup_evidence = {}
+        if warmup:
+            assert reference is not None
+            assert_frozen(trainer.model, reference)
+            final_signature = scene_signature(
+                trainer.model, trainer.val_loaders, trainer.device
+            )
+            if final_signature != initial_signature:
+                raise ValueError(
+                    "source validation scene logits changed during detector warmup"
+                )
+            warmup_evidence = {
+                "frozen_state_tensors": len(reference),
+                "frozen_state_equal": True,
+                "scene_before": initial_signature,
+                "scene_after": final_signature,
+                "scene_logits_equal": True,
+            }
         joint = job.get("joint", False)
         status(
             "source_validation" if joint else "held_out_evaluation", best_epoch=best["epoch"]
@@ -376,19 +517,43 @@ def record_epoch(self, epoch, metrics):
         )
         evaluation_name = "validation" if joint else "test"
         write_json(out / f"{evaluation_name}.json", test)
+        if warmup:
+            baseline = json.loads((out / "baseline_val.json").read_text())
+            warmup_evidence.update(
+                baseline_recall=baseline[DET_METRIC],
+                selected_recall=test[DET_METRIC],
+                baseline_empty_fp=baseline[EMPTY_FP],
+                selected_empty_fp=test[EMPTY_FP],
+                accepted=(
+                    test[DET_METRIC] > baseline[DET_METRIC]
+                    and test[EMPTY_FP] <= baseline[EMPTY_FP]
+                ),
+            )
+            write_json(out / "warmup_evidence.json", warmup_evidence)
         for head, (images, predictions, targets) in samples.items():
             grid = prediction_grid(images, predictions, targets, terrain_palette(list(CLASSES)))
             Image.fromarray(grid).save(out / f"{evaluation_name}_{head}_preview.jpg")
         verify(out)
         report = {
             "status": "completed",
-            "kind": "AI partial joint pilot" if joint else "AI partial-semantic pilot",
+            "kind": (
+                "AI partial detector warmup"
+                if warmup
+                else "AI partial joint pilot"
+                if joint
+                else "AI partial-semantic pilot"
+            ),
             "job_sha256": job_hash,
             "git_commit": job["git_commit"],
             "counts": job["counts"],
             "best_epoch": best["epoch"],
             "last_epoch": load_checkpoint(last)["epoch"],
-            "best_validation_teacher_miou": best["best_metric"],
+            "primary_metric": trainer.primary_metric,
+            "best_primary_metric": best["best_metric"],
+            "best_validation_teacher_miou": test["scene_mIoU"]
+            if joint
+            else best["best_metric"],
+            "warmup": warmup_evidence,
             "test_evaluated": not joint,
             f"{evaluation_name}_teacher_miou": test["scene_mIoU"],
             "instance_counts": job.get("instance_counts", {}),
@@ -403,6 +568,8 @@ def record_epoch(self, epoch, metrics):
                     out / "baseline_val.json",
                     out / "model/metrics.jsonl",
                     *out.glob(f"{evaluation_name}_*_preview.jpg"),
+                    *out.glob("scene_before.json"),
+                    *out.glob("warmup_evidence.json"),
                 ]
             },
         }
@@ -428,6 +595,10 @@ def main():
     parser.add_argument("--source", type=Path)
     parser.add_argument("--instances", type=Path)
     parser.add_argument("--initial-checkpoint", type=Path)
+    parser.add_argument("--detector-warmup", action="store_true")
+    parser.add_argument(
+        "--class-negative-normalization", choices=("sum", "positive_budget"), default="sum"
+    )
     parser.add_argument(
         "--held-out", default="Tao-Hsin", choices=("Tao-Hsin", "Taichung", "Kaohsiung")
     )
@@ -435,7 +606,15 @@ def main():
     if args.action == "prepare":
         if args.source is None:
             parser.error("prepare requires --source")
-        prepare(args.source, args.out, args.held_out, args.instances, args.initial_checkpoint)
+        prepare(
+            args.source,
+            args.out,
+            args.held_out,
+            args.instances,
+            args.initial_checkpoint,
+            args.detector_warmup,
+            args.class_negative_normalization,
+        )
     else:
         out = args.out.resolve()
         target = out / "snapshot/tools/annotation/studioa_train.py"
diff --git a/tools/commissioning/appearance_calibrate.py b/tools/commissioning/appearance_calibrate.py
index 2a85db7..df3365f 100644
--- a/tools/commissioning/appearance_calibrate.py
+++ b/tools/commissioning/appearance_calibrate.py
@@ -36,6 +36,7 @@
 
 import argparse
 import json
+import os
 import sys
 from itertools import pairwise
 from pathlib import Path
@@ -44,7 +45,7 @@
 import torch
 from PIL import Image
 
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 sys.path.insert(0, str(ROOT / "src"))
 
 from syncai_hydranet.analytics.appearance import torso_histogram  # noqa: E402
diff --git a/tools/commissioning/cluster_rules.py b/tools/commissioning/cluster_rules.py
index edb4a3c..aa46f7a 100644
--- a/tools/commissioning/cluster_rules.py
+++ b/tools/commissioning/cluster_rules.py
@@ -25,6 +25,7 @@
 import argparse
 import importlib.util
 import json
+import os
 from pathlib import Path
 
 import numpy as np
@@ -35,7 +36,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 
 spec = importlib.util.spec_from_file_location("recipe", str(ROOT / "tools/site30k/recipe.py"))
 R = importlib.util.module_from_spec(spec)
diff --git a/tools/commissioning/demo_gif.py b/tools/commissioning/demo_gif.py
index 317aee4..fa6a6ae 100644
--- a/tools/commissioning/demo_gif.py
+++ b/tools/commissioning/demo_gif.py
@@ -45,6 +45,7 @@
 
 import argparse
 import json
+import os
 import subprocess
 import sys
 from pathlib import Path
@@ -72,7 +73,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 
 # The run the tools ship from, named once in `syncai_hydranet.shipped`. Six files
 # used to carry their own copy of this string and the best run was in none of them.
diff --git a/tools/commissioning/demo_video.py b/tools/commissioning/demo_video.py
index 8db73d6..5159e72 100644
--- a/tools/commissioning/demo_video.py
+++ b/tools/commissioning/demo_video.py
@@ -93,6 +93,7 @@
     track_staff,
 )
 from syncai_hydranet.analytics.tracker import Tracker
+from syncai_hydranet.analytics.world import panel_box
 from syncai_hydranet.config import load_config
 from syncai_hydranet.data.video import frames as decode_frames
 from syncai_hydranet.data.video import probe as probe_video
@@ -113,7 +114,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 
 
 def _display_verdict(track):
@@ -128,6 +129,19 @@ def _display_verdict(track):
 RUN = SHIPPED_RUN
 
 
+def _check_source_size(cf, src_w: int, src_h: int, clip) -> None:
+    """Boxes are rescaled by the probed size, so a different stream is not a scaling
+    error any more -- it is a different lens crop, and the calibration cannot say which."""
+    if cf.source_size_px and tuple(cf.source_size_px) != (src_w, src_h):
+        cw, ch = cf.source_size_px
+        print(
+            f"  {cf.camera_id}: {Path(str(clip)).name} is {src_w}x{src_h}; the camera was "
+            f"commissioned from a {cw}x{ch} stream. Same aspect keeps the metres; a crop "
+            "does not, and only the NVR knows which this is.",
+            flush=True,
+        )
+
+
 def _records_pass(
     args, clip, src_w, src_h, size, model, device, cf, person_label, rng, staff_model
 ):
@@ -299,6 +313,7 @@ def _render_in_chunks(args, camera, clip, cf, bounds, staff_model) -> int:
     # The decoded size the recorded boxes are in, so the replay converts to calibrated
     # pixels the same way the recording pass did.
     src_w, src_h, _ = probe_video(str(clip))
+    _check_source_size(cf, src_w, src_h, clip)
     t0 = time.time()
     provenance = capture(ROOT, camera, clip, RUN, args.checkpoint)
     per = math.ceil(args.frames / args.workers)
@@ -706,6 +721,7 @@ def main() -> int:
         ]
     n = n_det = n_fp = n_placed = n_outside = n_blur = n_posed = 0
     src_w, src_h, _ = probe_video(str(clip))
+    _check_source_size(cf, src_w, src_h, clip)
     # The plate is this camera's own empty shop, named by its camera.json. Missing is a
     # refusal rather than a silent single-instrument run: the whole argument for two
     # instruments is that neither is trusted alone.
@@ -855,7 +871,7 @@ def main() -> int:
         moving = 0
         for t in tracks:
             seen_ids.add(t.track_id)
-            bx = np.asarray(t.box, float) / 2.0
+            bx = panel_box(t.box, (src_w, src_h), view_img.size)
             box_col = track_colour(t, None if staff_model is None else _display_verdict)
             d.rectangle(list(bx), outline=box_col, width=2)
             d.text((bx[0] + 3, bx[1] + 2), f"#{t.track_id}", fill=box_col)
diff --git a/tools/commissioning/depth_complete.py b/tools/commissioning/depth_complete.py
index 75658ca..d72e294 100644
--- a/tools/commissioning/depth_complete.py
+++ b/tools/commissioning/depth_complete.py
@@ -18,6 +18,7 @@
 Usage: uv run python tools/commissioning/depth_complete.py <camera> [...]
 """
 
+import os
 import sys
 from pathlib import Path
 
@@ -32,7 +33,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 IGNORE = 255
 TALL_M = 1.35
 FLAT_M = 0.08
diff --git a/tools/commissioning/extras_pass.py b/tools/commissioning/extras_pass.py
index d5d8287..253cbc4 100644
--- a/tools/commissioning/extras_pass.py
+++ b/tools/commissioning/extras_pass.py
@@ -16,6 +16,7 @@
 """
 
 import dataclasses
+import os
 import sys
 from pathlib import Path
 
@@ -34,7 +35,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 W, H = 1920, 1080
 MIN_PX = 2000  # doors and the generic product union
 MIN_SUB_PX = 150  # a phone on a table is small and real
diff --git a/tools/commissioning/facing_review.py b/tools/commissioning/facing_review.py
index 4c3e1d8..1a774aa 100644
--- a/tools/commissioning/facing_review.py
+++ b/tools/commissioning/facing_review.py
@@ -7,6 +7,7 @@
 
 import argparse
 import json
+import os
 from pathlib import Path
 
 import numpy as np
@@ -15,7 +16,7 @@
 from syncai_bev3d.object_instances import image_digest, load_instances
 from syncai_hydranet.geometry.camera_json import CameraFile
 
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 
 
 def main():
diff --git a/tools/commissioning/floor_calibrate.py b/tools/commissioning/floor_calibrate.py
index 2b512c0..def27b8 100644
--- a/tools/commissioning/floor_calibrate.py
+++ b/tools/commissioning/floor_calibrate.py
@@ -23,6 +23,7 @@
 import itertools
 import json
 import math
+import os
 from pathlib import Path
 
 import numpy as np
@@ -33,7 +34,7 @@
 from syncai_hydranet.geometry.camera_json import CameraFile
 from syncai_hydranet.geometry.ground import Camera, pixel_to_ground, undistort_points
 
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 OUT = ROOT / "runs/commission_review/floor_calib"
 # The sweep. 45-95 covers every lens the fleet census found (PLAN 7.12).
 VFOV_MIN, VFOV_MAX, VFOV_STEP = 45.0, 95.0, 2.5
diff --git a/tools/commissioning/footprints_from_masks.py b/tools/commissioning/footprints_from_masks.py
index ec373d2..be7cbdd 100644
--- a/tools/commissioning/footprints_from_masks.py
+++ b/tools/commissioning/footprints_from_masks.py
@@ -86,6 +86,7 @@
 
 import argparse
 import json
+import os
 from pathlib import Path
 
 import numpy as np
@@ -98,7 +99,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 COMMISSIONED = ROOT / "runs/commission01"
 OUT = ROOT / "runs/footprints01"
 
diff --git a/tools/commissioning/fp_polygons.py b/tools/commissioning/fp_polygons.py
index b8f8ab1..35c1258 100644
--- a/tools/commissioning/fp_polygons.py
+++ b/tools/commissioning/fp_polygons.py
@@ -13,6 +13,7 @@
 """
 
 import json
+import os
 import sys
 from collections import defaultdict
 from pathlib import Path
@@ -28,7 +29,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 ANN_DIR = ROOT / "datasets/site30k_v1/annotations"
 GRAY_MAX = 0.35
 CELL_PX = 64
diff --git a/tools/commissioning/geometry_bench.py b/tools/commissioning/geometry_bench.py
index f5d4c98..040646d 100644
--- a/tools/commissioning/geometry_bench.py
+++ b/tools/commissioning/geometry_bench.py
@@ -56,6 +56,7 @@
 import argparse
 import json
 import math
+import os
 import sys
 from dataclasses import dataclass, field
 from pathlib import Path
@@ -72,7 +73,7 @@
     unproject,
 )
 
-ROOT = Path(__file__).resolve().parent.parent.parent
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parent.parent.parent))
 
 
 FRAME_H, FRAME_W = 1080, 1920
diff --git a/tools/commissioning/glazing_pass.py b/tools/commissioning/glazing_pass.py
index 49de62e..97f5c45 100644
--- a/tools/commissioning/glazing_pass.py
+++ b/tools/commissioning/glazing_pass.py
@@ -9,6 +9,7 @@
 
 import argparse
 import json
+import os
 import shutil
 from pathlib import Path
 
@@ -24,7 +25,7 @@
 from syncai_hydranet.models.glazing import GlazingEncoder, GlazingHead
 from syncai_hydranet.utils.visualize import preprocess
 
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 
 
 def main():
diff --git a/tools/commissioning/heads_video.py b/tools/commissioning/heads_video.py
index b8c5e8e..7149db0 100644
--- a/tools/commissioning/heads_video.py
+++ b/tools/commissioning/heads_video.py
@@ -76,7 +76,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 
 
 def _display_verdict(track):
diff --git a/tools/commissioning/heatmap3d.py b/tools/commissioning/heatmap3d.py
index 2130746..ddef1e6 100644
--- a/tools/commissioning/heatmap3d.py
+++ b/tools/commissioning/heatmap3d.py
@@ -15,6 +15,7 @@
 
 import argparse
 import json
+import os
 import sys
 import time
 from pathlib import Path
@@ -38,7 +39,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 
 
 def main() -> int:
diff --git a/tools/commissioning/map_anything_eval.py b/tools/commissioning/map_anything_eval.py
index b36cfcd..78866d9 100755
--- a/tools/commissioning/map_anything_eval.py
+++ b/tools/commissioning/map_anything_eval.py
@@ -59,11 +59,12 @@
 import argparse
 import json
 import math
+import os
 import sys
 from dataclasses import dataclass, field
 from pathlib import Path
 
-REPO = Path(__file__).resolve().parent.parent.parent
+REPO = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parent.parent.parent))
 
 # PLAN 7.19: pinned by tile grid on this camera and assumed on every other.
 ANCHOR_CAMERA = "Taichung-cam01"
diff --git a/tools/commissioning/masks_diagnose.py b/tools/commissioning/masks_diagnose.py
index 42acc02..f4ab94d 100644
--- a/tools/commissioning/masks_diagnose.py
+++ b/tools/commissioning/masks_diagnose.py
@@ -25,6 +25,7 @@
 import argparse
 import importlib.util
 import json
+import os
 from pathlib import Path
 
 import numpy as np
@@ -34,7 +35,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 
 spec = importlib.util.spec_from_file_location("recipe", str(ROOT / "tools/site30k/recipe.py"))
 R = importlib.util.module_from_spec(spec)
diff --git a/tools/commissioning/masks_pass.py b/tools/commissioning/masks_pass.py
index 2787e88..0b81a01 100644
--- a/tools/commissioning/masks_pass.py
+++ b/tools/commissioning/masks_pass.py
@@ -8,6 +8,7 @@
 
 import argparse
 import importlib.util
+import os
 from pathlib import Path
 
 import numpy as np
@@ -18,7 +19,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 
 spec = importlib.util.spec_from_file_location("recipe", str(ROOT / "tools/site30k/recipe.py"))
 R = importlib.util.module_from_spec(spec)
diff --git a/tools/commissioning/measure_floor.py b/tools/commissioning/measure_floor.py
index c16465c..05cf2c1 100644
--- a/tools/commissioning/measure_floor.py
+++ b/tools/commissioning/measure_floor.py
@@ -5,13 +5,14 @@
 
 import argparse
 import json
+import os
 from pathlib import Path
 
 from PIL import Image
 
 from syncai_hydranet.geometry.camera_json import CameraFile
 
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 PAGE = Path(__file__).with_suffix(".html").read_text()
 
 
diff --git a/tools/commissioning/objects_pass.py b/tools/commissioning/objects_pass.py
index da168a5..62215e3 100644
--- a/tools/commissioning/objects_pass.py
+++ b/tools/commissioning/objects_pass.py
@@ -5,6 +5,7 @@
 """
 
 import argparse
+import os
 from pathlib import Path
 
 from syncai_bev3d.object_instances import OBJECT_PROMPTS
@@ -12,7 +13,7 @@
 from syncai_bev3d.teachers.scene_objects import run_camera
 from syncai_hydranet.utils.device import pick_device
 
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 
 
 def main():
diff --git a/tools/commissioning/scale_rulers.py b/tools/commissioning/scale_rulers.py
index 27a1ab2..ff9698e 100644
--- a/tools/commissioning/scale_rulers.py
+++ b/tools/commissioning/scale_rulers.py
@@ -19,6 +19,7 @@
 import argparse
 import datetime as dt
 import json
+import os
 import shutil
 import subprocess
 import sys
@@ -29,7 +30,7 @@
 from syncai_bev3d import floor_axis, rulers, scene_mesh
 from syncai_hydranet.geometry.ground import pixel_to_ground, undistort_points
 
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 VERDICTS = ROOT / "runs/commission_review/scale"
 
 
@@ -85,7 +86,8 @@ def verdicts(cameras: list[str]) -> dict[str, rulers.Verdict]:
             r = read[c]
             rs = [
                 rulers.person_ruler(r["calib"]),
-                rulers.tile_ruler(r["period"], tile),
+                # the catalogue size is the cameras' consensus, not an independent fact
+                rulers.tile_ruler(r["period"], tile, anchored=False),
                 rulers.table_ruler(r["table_h"], tables),
             ]
             v = rulers.combine([x for x in rs if x is not None])
diff --git a/tools/commissioning/scene3d.py b/tools/commissioning/scene3d.py
index 8a79ab6..d1379be 100644
--- a/tools/commissioning/scene3d.py
+++ b/tools/commissioning/scene3d.py
@@ -13,6 +13,7 @@
 where that projection is exact.
 """
 
+import os
 import sys
 from pathlib import Path
 
@@ -29,7 +30,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 CLASSES = {1: "floor", 2: "wall", 3: "column", 4: "display_table", 5: "display_shelf"}
 PALETTE = np.array(
     [
diff --git a/tools/commissioning/scene_mesh.py b/tools/commissioning/scene_mesh.py
index f3c5713..2efa285 100644
--- a/tools/commissioning/scene_mesh.py
+++ b/tools/commissioning/scene_mesh.py
@@ -11,6 +11,7 @@
 
 import argparse
 import json
+import os
 import shutil
 import tempfile
 from pathlib import Path
@@ -29,7 +30,7 @@
     render,
 )
 
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 
 
 def export_glb(camera, items, *, out=None):
diff --git a/tools/commissioning/scene_overlay.py b/tools/commissioning/scene_overlay.py
index 300510c..e3f81ce 100644
--- a/tools/commissioning/scene_overlay.py
+++ b/tools/commissioning/scene_overlay.py
@@ -12,6 +12,7 @@
 """
 
 import argparse
+import os
 from pathlib import Path
 
 import numpy as np
@@ -25,7 +26,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 
 
 def project(verts: np.ndarray, cf: CameraFile, scale: float) -> tuple[np.ndarray, np.ndarray]:
diff --git a/tools/commissioning/service_zones.py b/tools/commissioning/service_zones.py
index 567f067..fbfaeff 100644
--- a/tools/commissioning/service_zones.py
+++ b/tools/commissioning/service_zones.py
@@ -52,6 +52,7 @@
 import argparse
 import dataclasses
 import json
+import os
 from pathlib import Path
 
 import numpy as np
@@ -71,7 +72,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 COMMISSIONED = ROOT / "runs/commission01"
 OUT = ROOT / "runs/service_zones01"
 
diff --git a/tools/commissioning/social_card.py b/tools/commissioning/social_card.py
index f2d9ef1..a055bf3 100644
--- a/tools/commissioning/social_card.py
+++ b/tools/commissioning/social_card.py
@@ -21,6 +21,7 @@
 from __future__ import annotations
 
 import argparse
+import os
 from pathlib import Path
 
 import numpy as np
@@ -32,7 +33,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 CARD = (1280, 640)
 PANEL_W = 660  # the scene half; the text needs the rest at thumbnail size
 CAPTION_H = 62  # `scene_mesh.render`'s own caption band, rendered then cut
diff --git a/tools/commissioning/stage0_baseline.py b/tools/commissioning/stage0_baseline.py
index 2aea32b..da2e531 100644
--- a/tools/commissioning/stage0_baseline.py
+++ b/tools/commissioning/stage0_baseline.py
@@ -62,25 +62,27 @@ def freeze(root, out, cameras):
     shutil.copytree(
         root / "src", snapshot / "src", ignore=shutil.ignore_patterns("__pycache__", "*.pyc")
     )
-    tool = Path("tools/commissioning/scene_mesh.py")
-    (snapshot / tool).parent.mkdir(parents=True)
-    shutil.copy2(root / tool, snapshot / tool)
-    shutil.copy2(Path(__file__), snapshot / "tools/commissioning/stage0_baseline.py")
-    for name in ("pyproject.toml", "uv.lock"):
-        shutil.copy2(root / name, snapshot / name)
+    runner = Path("tools/commissioning/stage0_baseline.py")
+    (snapshot / runner).parent.mkdir(parents=True, exist_ok=True)
+    shutil.copy2(Path(__file__), snapshot / runner)
     inputs = {camera: capture_inputs(root, camera) for camera in cameras}
+    # Everything capture_inputs identifies travels with the snapshot -- the inputs and
+    # the code alike -- so the two identity records can only differ when a file changed
+    # under the copy, never because this list and that one drifted apart.
     for record in inputs.values():
-        for name, digest in record["inputs"].items():
-            if digest is None:
-                continue
-            relative = Path(name)
-            if relative.is_absolute() or ".." in relative.parts:
-                raise ValueError(f"snapshot requires checkout-relative input: {name}")
-            target = snapshot / relative
-            target.parent.mkdir(parents=True, exist_ok=True)
-            shutil.copy2(root / relative, target)
-            if sha256(target) != digest:
-                raise RuntimeError(f"input changed while copying: {name}")
+        for section in ("inputs", "code"):
+            for name, digest in record[section].items():
+                if digest is None:
+                    continue
+                relative = Path(name)
+                if relative.is_absolute() or ".." in relative.parts:
+                    raise ValueError(f"snapshot requires checkout-relative input: {name}")
+                target = snapshot / relative
+                if not target.exists():
+                    target.parent.mkdir(parents=True, exist_ok=True)
+                    shutil.copy2(root / relative, target)
+                if sha256(target) != digest:
+                    raise RuntimeError(f"input changed while copying: {name}")
     for camera in cameras:
         if capture_inputs(snapshot, camera) != inputs[camera]:
             raise RuntimeError(f"snapshot is incomplete or source changed: {camera}")
diff --git a/tools/commissioning/stage0_commission.py b/tools/commissioning/stage0_commission.py
new file mode 100644
index 0000000..0c3eeda
--- /dev/null
+++ b/tools/commissioning/stage0_commission.py
@@ -0,0 +1,263 @@
+"""Stage 0 for one camera from one clip: the README chain as one command.
+
+    uv run python tools/commissioning/stage0_commission.py <camera> --clip <mp4> \\
+        [--note "nvr 10.12.31.9 ch12"] [--allow-bootstrap] \\
+        [--freeze-out runs/stage0_<cam>] [--dry-run]
+
+Until 2026-09-16 the chain for a second site lived in a shell script under `runs/` of a
+copied checkout (FTI, `batch_one.sh`), with the bootstrap rule as inline Python nobody
+could test. The steps here are the README's, in the README's order, against the tree
+`SYNCAI_ROOT` names: static plate, person boxes, onboard, zones, the camera.json
+contract, the review bundle, masks, extras, depth completion, the three scene renders,
+and finally `stage0_baseline.py`, which freezes the result as a review candidate.
+
+`--allow-bootstrap` is the only decision the tool takes on its own, and it says so in
+the calib: a camera with no person corpus gets scale 1.0 on DA-V2's raw metres, flagged
+`scale_bootstrap_dav2_raw_unverified`, so the tile ruler has a walkable mask to read
+through. Nothing downstream may present those metres as measured.
+"""
+
+from __future__ import annotations
+
+import argparse
+import datetime as _dt
+import json
+import os
+import re
+import subprocess
+import sys
+from pathlib import Path
+
+from syncai_hydranet.paths import repo_root
+
+ROOT = repo_root(Path(__file__).resolve().parents[2])
+SLOT = re.compile(r"_(\d{8}-\d{6})_")
+BOOTSTRAP_FLAG = "scale_bootstrap_dav2_raw_unverified"
+
+
+def slot_from_clip(clip: Path) -> str:
+    """The `_YYYYMMDD-HHMMSS_` slot in an archive name, else the file's mtime in UTC."""
+    m = SLOT.search(clip.name)
+    if m:
+        return m.group(1)
+    stamp = _dt.datetime.fromtimestamp(clip.stat().st_mtime, tz=_dt.UTC)
+    return stamp.strftime("%Y%m%d-%H%M%S")
+
+
+def register_camera(cameras_json: Path, camera: str, note: str) -> None:
+    site = json.loads(cameras_json.read_text())
+    site["cameras"].setdefault(camera, {"role": "factory", "note": note})
+    cameras_json.write_text(json.dumps(site, indent=1, ensure_ascii=False) + "\n")
+
+
+def bootstrap_unmeasured(calib_path: Path, allow: bool) -> str:
+    """Return the calib's scale source; substitute the bootstrap when there is none.
+
+    No floor plane at any vfov is a refusal (exit 2): the view shows too little floor to
+    commission. An unmeasured scale with `allow` False is also a refusal (exit 3): the
+    caller has to say it wants raw metres."""
+    d = json.loads(calib_path.read_text())
+    if d.get("height_dav2_raw_m") is None or d.get("pitch_deg") is None:
+        raise SystemExit(2)
+    source = str(d.get("scale_source", "unmeasured"))
+    if not source.startswith("unmeasured"):
+        return source
+    if not allow:
+        raise SystemExit(3)
+    d.update(
+        {
+            "height_m": d["height_dav2_raw_m"],
+            "scale": 1.0,
+            "height_source": "dav2_plane_height_raw_BOOTSTRAP_unverified",
+            "scale_source": "dav2_metric_indoor_raw_bootstrap_unverified",
+        }
+    )
+    d["flags"] = [f for f in d.get("flags", []) if f != "scale_unmeasured"] + [BOOTSTRAP_FLAG]
+    calib_path.write_text(json.dumps(d, indent=1) + "\n")
+    return d["scale_source"]
+
+
+def plan(
+    root: Path,
+    camera: str,
+    clips_root: Path,
+    plates_root: Path,
+    freeze_out: Path | None,
+    slot: str = "",
+):
+    """The chain's commands, each relative to `root`, in the README's order.
+
+    The review bundle refuses to overwrite, so a rerun writes beside the last one, named
+    by the clip's slot rather than deleting what a person may have looked at."""
+    py = sys.executable
+    anns = f"datasets/fti_person_gdino/{camera}/annotations/instances_all.json"
+    steps = [
+        (
+            "plate",
+            [
+                py,
+                "scripts/static_plates.py",
+                "--root",
+                str(clips_root),
+                "--out",
+                str(plates_root),
+                "--only",
+                camera,
+            ],
+        ),
+        (
+            "person boxes",
+            [
+                py,
+                "scripts/gdino_person_boxes.py",
+                "--out",
+                f"datasets/fti_person_gdino/{camera}",
+                "--frames",
+                "24",
+                "--train-thr",
+                "0.35",
+                "--clips",
+                *sorted(str(p) for p in (root / clips_root / camera).glob("archive_*.mp4")),
+            ],
+        ),
+        (
+            "onboard",
+            [
+                py,
+                "scripts/onboard_camera.py",
+                "--camera",
+                camera,
+                "--out",
+                "runs/onboard01",
+                "--plates-root",
+                str(plates_root),
+                "--cameras-json",
+                str(clips_root / "cameras.json"),
+                "--person-anns",
+                anns,
+                "--skip-person-frac",
+            ],
+        ),
+        (
+            "zones",
+            [
+                py,
+                "scripts/propose_zones.py",
+                "--cameras",
+                camera,
+                "--calib-dir",
+                "runs/onboard01",
+                "--out",
+                "runs/zones01",
+            ],
+        ),
+        (
+            "review bundle",
+            [
+                py,
+                "tools/commissioning/commission_camera.py",
+                f"runs/onboard01/{camera}.calib.json",
+                "--out",
+                f"runs/commission_review/{camera}_{slot or 'batch'}",
+            ],
+        ),
+        (
+            "masks",
+            [
+                py,
+                "tools/commissioning/masks_pass.py",
+                camera,
+                "--plates-root",
+                str(plates_root),
+                "--out-root",
+                "runs/commission01",
+                "--calib-root",
+                "runs/onboard01",
+            ],
+        ),
+        ("extras", [py, "tools/commissioning/extras_pass.py", camera]),
+        ("depth completion", [py, "tools/commissioning/depth_complete.py", camera]),
+        ("scene3d", [py, "tools/commissioning/scene3d.py", camera]),
+        ("scene overlay", [py, "tools/commissioning/scene_overlay.py", camera]),
+        ("scene mesh", [py, "tools/commissioning/scene_mesh.py", camera]),
+    ]
+    if freeze_out is not None:
+        steps.append(
+            (
+                "freeze",
+                [
+                    py,
+                    "tools/commissioning/stage0_baseline.py",
+                    "--root",
+                    str(root),
+                    "--out",
+                    str(freeze_out),
+                    camera,
+                ],
+            )
+        )
+    return steps
+
+
+def main(argv=None):
+    ap = argparse.ArgumentParser(
+        description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    ap.add_argument("camera")
+    ap.add_argument("--clip", type=Path, required=True, help="one recorded clip of the camera")
+    ap.add_argument("--note", default="", help="where the stream came from, for cameras.json")
+    ap.add_argument("--clips-root", type=Path, default=Path("datasets/fti_clips"))
+    ap.add_argument("--plates-root", type=Path, default=Path("datasets/fti_static"))
+    ap.add_argument(
+        "--allow-bootstrap",
+        action="store_true",
+        help="scale 1.0 on raw DA-V2 metres when no person box; flagged in the calib",
+    )
+    ap.add_argument(
+        "--freeze-out",
+        type=Path,
+        help="stage0_baseline output directory; omit to stop at the renders",
+    )
+    ap.add_argument("--dry-run", action="store_true", help="print the commands and stop")
+    args = ap.parse_args(argv)
+    root = ROOT
+    clips_root, plates_root = args.clips_root, args.plates_root
+    clip = args.clip.resolve()
+    if not clip.is_file():
+        ap.error(f"clip not found: {clip}")
+    link_dir = root / clips_root / args.camera
+    slot = slot_from_clip(clip)
+    link = link_dir / f"archive_{slot}_fti.mp4"
+    if args.dry_run:
+        print(f"link {link} -> {clip}")
+        for label, argv_ in plan(
+            root, args.camera, clips_root, plates_root, args.freeze_out, slot
+        ):
+            print(f"[{label}] {' '.join(argv_)}")
+        return 0
+    link_dir.mkdir(parents=True, exist_ok=True)
+    if link.is_symlink() or link.exists():
+        link.unlink()
+    link.symlink_to(clip)
+    register_camera(root / clips_root / "cameras.json", args.camera, args.note)
+    env = dict(os.environ, SYNCAI_ROOT=str(root), PYTHONPATH=str(root / "src"))
+    for label, argv_ in plan(root, args.camera, clips_root, plates_root, args.freeze_out, slot):
+        print(f"== {args.camera} {label}", flush=True)
+        subprocess.run(["nice", "-n", "10", *argv_], cwd=root, env=env, check=True)
+        if label == "onboard":
+            calib = root / f"runs/onboard01/{args.camera}.calib.json"
+            source = bootstrap_unmeasured(calib, args.allow_bootstrap)
+            print(f"   scale source: {source}", flush=True)
+        if label == "zones":
+            from syncai_bev3d.commissioning import from_onboard_calib
+
+            (root / "runs/commission01").mkdir(parents=True, exist_ok=True)
+            from_onboard_calib(root / f"runs/onboard01/{args.camera}.calib.json").save(
+                root / f"runs/commission01/{args.camera}.camera.json"
+            )
+    print(f"STAGE0_DONE {args.camera}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/tools/commissioning/stage0_train.py b/tools/commissioning/stage0_train.py
index ccdcfe1..7cec76b 100644
--- a/tools/commissioning/stage0_train.py
+++ b/tools/commissioning/stage0_train.py
@@ -37,7 +37,7 @@
 from syncai_hydranet.data.label_maps import get_scheme
 from syncai_hydranet.data.store_split import STORES, camera_store, fold_split, validate_fold
 
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 SOURCES = {
     "retail_objects_batch02": "retail_surfaces_from_objects",
     "retail_objects_batch03": "retail_surfaces_from_objects",
diff --git a/tools/commissioning/tile_ruler.py b/tools/commissioning/tile_ruler.py
index cc04fdf..97010e6 100644
--- a/tools/commissioning/tile_ruler.py
+++ b/tools/commissioning/tile_ruler.py
@@ -21,6 +21,7 @@
 """
 
 import argparse
+import os
 from pathlib import Path
 
 import numpy as np
@@ -29,7 +30,7 @@
 from syncai_bev3d.rulers import write_scaled_root
 from syncai_hydranet.geometry.ground import pixel_to_ground, undistort_points
 
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 # Below this autocorrelation the pitch is not trusted to rescale anything.
 STRENGTH_MIN = 0.35
 
diff --git a/tools/commissioning/zone_draw.py b/tools/commissioning/zone_draw.py
index e771c6c..27777bb 100644
--- a/tools/commissioning/zone_draw.py
+++ b/tools/commissioning/zone_draw.py
@@ -51,6 +51,7 @@
 import base64
 import dataclasses
 import json
+import os
 from pathlib import Path
 
 import numpy as np
@@ -61,7 +62,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 COMMISSIONED = ROOT / "runs/commission01"
 OUT = ROOT / "runs/zone_draw01"
 DRAWABLE_KINDS = sorted(ZONE_KINDS - {"walkable"})
diff --git a/tools/commissioning/zones_apply.py b/tools/commissioning/zones_apply.py
index de5a130..7ea2c8b 100644
--- a/tools/commissioning/zones_apply.py
+++ b/tools/commissioning/zones_apply.py
@@ -23,6 +23,7 @@
 
 import dataclasses
 import json
+import os
 import sys
 from pathlib import Path
 
@@ -35,7 +36,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 ZONES_HUMAN = ROOT / "tools/site30k/zones.json"
 ZONES_PROPOSED = ROOT / "runs/zones01"
 CLASS_FILES = {
diff --git a/tools/commissioning/zones_confirm.py b/tools/commissioning/zones_confirm.py
index 49dac87..ee98bb3 100644
--- a/tools/commissioning/zones_confirm.py
+++ b/tools/commissioning/zones_confirm.py
@@ -62,6 +62,7 @@
 import argparse
 import dataclasses
 import json
+import os
 from pathlib import Path
 
 import numpy as np
@@ -74,7 +75,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 PROPOSED = ROOT / "runs/zones01"
 COMMISSIONED = ROOT / "runs/commission01"
 OUT = ROOT / "runs/zones_confirm01"
diff --git a/tools/site30k/boxes.py b/tools/site30k/boxes.py
index cd4915a..0bf3b40 100644
--- a/tools/site30k/boxes.py
+++ b/tools/site30k/boxes.py
@@ -19,6 +19,7 @@
 
 import importlib.util
 import json
+import os
 import sys
 from pathlib import Path
 
@@ -30,7 +31,7 @@
 # machine but the one it was typed on, and `sys.path` is the one place where that
 # fails before anything else can report it. `parents[2]` is the repo root --
 # tools/site30k/<file>.py.
-_REPO = Path(__file__).resolve().parents[2]
+_REPO = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 sys.path.insert(0, str(_REPO / "scripts"))
 spec = importlib.util.spec_from_file_location(
     "campaign", str(_REPO / "scripts/campaign_site30k.py")
diff --git a/tools/site30k/plan.py b/tools/site30k/plan.py
index e08cdf9..e34124c 100644
--- a/tools/site30k/plan.py
+++ b/tools/site30k/plan.py
@@ -33,6 +33,7 @@
 
 import argparse
 import json
+import os
 import re
 from collections import defaultdict
 from concurrent.futures import ThreadPoolExecutor
@@ -48,7 +49,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 PULL = ROOT / "datasets/studioa_pull_site30k"
 CALIB = ROOT / "runs/onboard01"
 # Filenames are UTC; the store is UTC+8. 23:00-06:00 local is the night tranche.
diff --git a/tools/site30k/recipe.py b/tools/site30k/recipe.py
index 5a2b53d..1113517 100644
--- a/tools/site30k/recipe.py
+++ b/tools/site30k/recipe.py
@@ -55,7 +55,7 @@
 # machine but the one it was typed on, and `sys.path` is the one place where that
 # fails before anything else can report it. `parents[2]` is the repo root --
 # tools/site30k/<file>.py.
-_REPO = Path(__file__).resolve().parents[2]
+_REPO = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 sys.path.insert(0, str(_REPO / "scripts"))
 import importlib.util  # noqa: E402
 
@@ -82,7 +82,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 # The pilot ran off datasets/studioa_clips; the campaign runs off the phase-2 pull. Both
 # are the same layout (<root>/<camera>/<stem>.mp4), so the root is a setting rather than
 # a fork in the code, and the plates follow it so a campaign plate never lands in the
diff --git a/tools/site30k/run_campaign.py b/tools/site30k/run_campaign.py
index 0e16fc4..d59e2b7 100644
--- a/tools/site30k/run_campaign.py
+++ b/tools/site30k/run_campaign.py
@@ -31,7 +31,7 @@
 # as an absolute path, so a second checkout ran against the first one's `runs/` and
 # any machine but this one failed at import with a path and no reason. Two levels up
 # from `tools/<group>/<tool>.py`, and `tests/test_no_absolute_sys_path.py` keeps it so.
-ROOT = Path(__file__).resolve().parents[2]
+ROOT = Path(os.environ.get("SYNCAI_ROOT", Path(__file__).resolve().parents[2]))
 RECIPE = ROOT / "tools/site30k/recipe.py"
 PYTHON = ROOT / ".venv/bin/python"
 PULL = "datasets/studioa_pull_site30k"