LENS 25 CROSS-LENS CONNECTIONS Blind Spot Scan — "What's invisible because of where you're standing?" Session 119 validated this lens with the Hailo-8 idle-NPU discovery. The cross-lens implications below are updated to reflect that finding. --- LENS 03 (Inventory / What Do We Own?) — DIRECT DEPENDENCY, preventive control The session-119 Hailo-8 discovery is the canonical case for why Lens 03 must precede every architecture review. Lens 25's process-blind-spot ("the audit pattern never asked what we own") is literally a recommendation to embed Lens 03 as a checklist item in every audit. When Lens 03 is run with rigor, three owned-but-undrawn compute substrates surface: 1. Hailo-8 AI HAT+ — 26 TOPS NPU on the Pi 5, idle for nav, YOLOv8n at 430 FPS, <10 ms, no WiFi. 2. Beast — second DGX Spark (GB10, 128 GB unified memory), always-on, idle workload. 3. Orin NX 16GB — 100 TOPS, reserved for the next robot, available for ahead-of-time experiments. Lens 25 documents the cost of skipping Lens 03; Lens 03 is the preventive control. The proposed one-line amendment to every future review — "list every powered device in the house; explain why each is or isn't in the diagram" — is literally Lens 03 operationalized as an audit gate. --- LENS 04 (Sensitivity Surface / WiFi Cliff) — REINFORCED and PARTIALLY RESOLVED Lens 04 identified WiFi latency as the single cliff-edge parameter: below 80 ms stable, above 100 ms system collapse. Lens 25 previously extended Lens 04 with load-shedding (removes the whole inference infra, not just latency). Session 119 adds a resolution path: with the Hailo-8 NPU running L1 obstacle detection locally, loss of WiFi becomes graceful degradation to "safe local wander" rather than a hard brick. The EMA filter tuning Lens 04 worked on only matters when the system is running; Hailo-L1 is what keeps Annie safe when WiFi is gone. The cliff is not removed — but the failure mode below it is transformed from "brick" to "degraded but safe." --- LENS 10 (Compute Geography / Where Does the Work Happen?) — HARD UPDATE Lens 10's current geography assumes a single control loop: "Pi sensors → WiFi → Panda VLM → WiFi → Pi actuators." Session 119 adds a parallel local geography: "Pi camera → Hailo-8 NPU (same board, 26 TOPS) → Pi actuators" — a zero-WiFi, sub-10 ms local loop for L1 safety detection. Lens 10 must be rewritten to show dual-process geography, matching the IROS arXiv 2601.21506 pattern: L1 — Fast reactive (Hailo-8, on-Pi) — 30+ Hz, <10 ms, safety L2 — Semantic nav (Panda RTX 5070 Ti VLM) — 15–27 Hz, WiFi-bound, grounding L3 — Multi-query scene (Panda VLM) — 5–9 Hz, WiFi-bound, classification L4 — Strategy/planning (Titan Gemma 4 26B) — 1–2 Hz, WiFi-bound, decomposition 66% latency reduction and 67.5% success vs 5.83% VLM-only is the paper's validation. --- LENS 21 (Economic Logic / Why This Architecture Survives) — TENSION, requires rider Lens 21 assumes architectures that survive do so because they are efficient. The Hailo-idle case contradicts this: the architecture survived for months not because WiFi-round-trip-to-Panda was efficient, but because the Hailo's existence was not visible to the decision. Economic selection operates only over the visible option set; blind spots distort the fitness landscape. Lens 21 needs an explicit rider: efficiency arguments only apply to architectures that were considered. Invisible options do not compete, so the architecture we see is not necessarily the economically optimal one — it is the economically optimal one within the blind-spot-filtered option space. --- LENS 24 (Counterfactual / What Would This Look Like if Designed Today?) — DIRECT FEED A counterfactual Annie designed today with full hardware inventory would: - Run L1 obstacle detection on Hailo-8 (430 FPS, <10 ms, no WiFi). - Reserve the Panda VLM for semantic reasoning only (L2/L3). - Use Beast (idle 2nd DGX Spark) as an always-on eval/inference partner or shadow. - Treat Orin NX 16GB as the compute substrate for the next robot, with Isaac ROS as the framework. The gap between the current drawn system and the fresh-design system is the debt that Lens 25 surfaces. Lens 24 should cite this as the concrete debt catalog. --- SYNTHESIS: THE LENS CHAIN THAT SURFACES BLIND SPOTS Session 119 demonstrated that Lens 25 is most valuable when it runs inside a chain: Lens 03 (inventory) → Lens 25 (blind spots) → Lens 10 (geography) → Lens 24 (counterfactual) The multi-agent research convergence pattern from session 119 (hardware-inventory agent plus IROS-paper-validation agent, with the Hailo-8 surfacing as the cross-cutting find) is the operational template. Each agent saw a fragment; convergence produced the insight. No single lens caught it alone — the lens chain did. The highest-leverage process change is simple: before running the 26-lens catalog, run Lens 03 first, explicitly, and ensure every powered device in the house appears in the inventory. Then the other 25 lenses have a complete option space to evaluate.