YOLO + semantic segmentation image-space complementarity

状态：development / CROSS_SOURCE_IMAGE_SPACE_SIGNAL_REPLICATED / CLASS_STABILITY_MIXED / NO_EFFECT_AUTHORITY

研究问题与版本

DUAL_LOOP_SEGMENTATION_COMPLEMENTARITY_R1 只回答：在同一 RGB frame 和同一 YOLO box union 上，固定 semantic-segmentation reference 是否产生可计算的、未被 YOLO 覆盖的 image-space class regions，以及这些区域的时间稳定性和主机成本。该 设计已在 Shiraz 与 Shanghai 两个 source 上以同一 host YOLO 合同复现；跨来源结果 仍是 Development image-space evidence，不是风险或融合效果。

本轮使用用户明确授权的固定 Development 诊断；两个 source 都是机制诊断输入， 不作 held-out Confirmation 或总体外推。它不读取中央阻塞 Agent 标签、risk、feedback 或 event 字段，不产生可通行性、风险、提醒、安全、Android 或生产结论。

稳定 Interface

从仓库根目录运行：

& E:\codex-tools\tools\venvs\blindassist-venv-export312\Scripts\python.exe `
  scripts\research\dual_loop_segmentation_complementarity\complementarity.py `
  --manifest artifacts.local\evidence\dual-loop-r1-unseen-natural-event-r0\rank2-shiraz\input-10hz-r1\manifest.jsonl `
  --trace artifacts.local\evidence\dual-loop-r1-unseen-natural-event-r0\rank2-shiraz\device-r1\baseline-output\trace.jsonl `
  --model artifacts.local\evidence\segmentation-candidate\sanpo-v3-pretrained-weighted-best-int8-20260713.tflite `
  --output artifacts.local\evidence\dual-loop-segmentation-complementarity-r1\report.json `
  --frames-output artifacts.local\evidence\dual-loop-segmentation-complementarity-r1\frames.jsonl `
  --threads 2

Inputs must have exact frame identity and image SHA matches. The runner uses every YOLO rectangle without confidence/NMS/risk filtering, projects boxes to the model output grid with clipped normalized coordinates, and uses raw segmentation argmax class masks. Missing, duplicate, reordered, or mismatched identities fail closed; no interpolation or nearest-frame repair is allowed.

若已有 RGB manifest 但没有匹配的 YOLO trace，可先用固定模型资产生成 Development-only host reference trace：

& E:\codex-tools\tools\venvs\blindassist-venv-export312\Scripts\python.exe `
  scripts\research\dual_loop_segmentation_complementarity\produce_host_trace.py `
  --manifest artifacts.local\evidence\dual-loop-r1-unseen-natural-event-r0\input-10hz-r1\manifest.jsonl `
  --model app\src\main\assets\yolo11n_fp16_320.tflite `
  --labels app\src\main\assets\coco_labels.txt `
  --output artifacts.local\evidence\dual-loop-segmentation-complementarity-r2-shanghai-host-yolo\trace.jsonl `
  --receipt artifacts.local\evidence\dual-loop-segmentation-complementarity-r2-shanghai-host-yolo\receipt.json `
  --progress artifacts.local\evidence\dual-loop-segmentation-complementarity-r2-shanghai-host-yolo\progress.json

该 trace 的 authority 是 DEVELOPMENT_HOST_REFERENCE_ONLY，必须披露 host LiteRT 与 QNN/device backend 的差异；它不能冒充手机输出或生产 parity。

输出

All outputs stay under artifacts.local/:

• report.json: frozen contract, input/model hashes, pairing, class-wise uncovered fractions, geometric union increment, temporal IoU/component summaries, runtime and stop checks;
• frames.jsonl: one paired-frame descriptive row, with no risk/feedback/event fields;
• progress.json: bounded progress receipt for long host execution.
• validation.json: independent recomputation receipt for frame count, ordering, class partition, union arithmetic, forbidden fields, and input hashes.
• Host detector preparation additionally writes trace.jsonl and receipt.json; the receipt binds the RGB manifest, YOLO asset/labels, backend, tensor shapes, thresholds and decoder contract. These are Development-only inputs to the complementarity runner.

The four argmax classes remain separate: walkable, boundary_step_curb, obstacle, and unknown_nonwalkable. Because the union of all four argmax masks covers the analysis grid by construction, the primary complementary interpretation is class-specific uncovered_fraction; union_increment is reported transparently as the geometric union quantity and is not called obstacle or risk discovery.

安全边界

• Observation unit is source_id + frame_id + image_sha256; frames are repeated observations, not independent samples.
• The current inputs are previously consumed Development sessions (burned), so all results are mechanism diagnostics only.
• No Android assets, production model, feedback path, default behavior, or safety authority is modified.
• A non-zero uncovered region is not a traversability, obstacle, event, or risk truth.
• This runner does not select among models or tune thresholds; the declared reference is fixed.

停止条件

Stop the current evidence version with NOT_EVALUABLE/failure if model interface or finite values fail, image identity pairing fails, any image hash or dimensions mismatch, or the segmentation output collapses to one class across the evaluated input. These conditions close only this candidate/evidence version, not the segmentation research question.

假设与规则质疑

The estimand is image-space only and uses session-first summaries. No p-value, event gate, effect gate, or frame-independent uncertainty claim is emitted. If later uncertainty is needed, source/session clustering must be retained.

失败资产复用

Reports, frame rows and host trace receipts may be reused as Development diagnostics, regression fixtures, or candidate failure records. They must not be relabeled as held-out confirmation, risk truth, device parity, or production evidence. Cross-source qualitative replication does not authorize a fusion operator; a later fusion design must freeze its own objective image-space unit and comparison rule.