LENS 19 — CROSS-LENS CONVERGENCE NOTES Scale Microscope: "What changes at 10x? 100x? 1000x?" ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ CONVERGENCE WITH LENS 04 (Sensitivity Surface) ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Lens 04 identified WiFi latency as the single highest-sensitivity parameter in the current system — the one knob with a discontinuous cliff edge at ~100ms. Lens 19 adds two scaling dimensions to this finding: (a) the cliff migrates toward the system as the household grows (unchanged from original analysis), and (b) the cliff BIFURCATES once Hailo-8 is activated on Pi 5: the safety path is demoted to a local sub-10ms NPU inference loop with zero WiFi dependency, while the semantic query path (VLM on Panda) still carries the full cliff risk. At 1-room/1-user scale (prototype), the household WiFi channel has 3–4 devices and operates well below saturation. WiFi latency stays below 30ms. The cliff exists at 100ms but the system is 70ms away from it. At whole-house scale (3 floors, 5+ users, 40+ IoT devices), the channel routinely saturates, and P95 latency sits at 80–200ms — straddling or exceeding the cliff. Pre-Hailo, this threatened both obstacle avoidance AND goal-tracking. Post-Hailo, it threatens only goal-tracking and scene reasoning — latency-tolerant tasks. The safety path is now architecturally immune to the cliff. RECOMMENDATION: Lens 04's latency watchdog proposal (measure rolling 500ms WiFi RTT; throttle at P95 > 60ms; halt at P95 > 100ms) should be REDESIGNED post-Hailo: the "halt at P95 > 100ms" action becomes "drop VLM-dependent tasks to best-effort, keep Hailo-driven safety loop running at full rate." Safety is no longer coupled to WiFi health. The degraded mode becomes a graceful downgrade of semantic response time, not a safety emergency. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ CONVERGENCE WITH LENS 09 (Sibling Hardware / Hailo-8 Activation) ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Lens 19's scaling analysis is fundamentally reshaped by the Hailo-8 activation finding (session 119 hardware audit). The Hailo-8 AI HAT+ is already physically installed on Annie's Pi 5 but currently idle for navigation — a 26 TOPS NPU sitting on-board, burning zero watts, contributing zero value. Activating it as the L1 safety layer (YOLOv8n at 430 FPS, <10ms, ~2 W) has three compounding scaling effects that a pure software-only lens would miss: 1. WiFi cliff bifurcation: safety path demoted to local inference; cliff persists only for semantic queries. This is a per-scale HEADROOM gain — Annie can now tolerate larger households without safety degradation. 2. Panda VRAM step function relaxation: obstacle detection moves off the Panda GPU entirely (onto Hailo NPU silicon), freeing ~800 MB against the 16 GB Panda ceiling. Roughly one SigLIP-sized worth of headroom bought back without writing a single line of model code. 3. New linear scaling curve added: Hailo-8 power draw scales smoothly with inference load at ~2 W continuous. No step functions, no discontinuities. This is a textbook well-behaved scaling dimension that replaces a discontinuous one — the rare scaling improvement that adds a green bar while removing a coral bar. RECOMMENDATION: Treat Hailo-8 activation as the single highest-leverage scaling move available to Annie. Zero hardware cost (already on-board), moderate software cost (HailoRT/TAPPAS pipeline on Pi 5), high scaling dividend (WiFi cliff neutralized for safety, VRAM runway extended on Panda). Before committing to Phase 2e (DINOv2 ViT-L on Panda), activate Hailo-8 first — it makes the DINOv2 addition substantially safer by pre-freeing the required headroom. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ CONVERGENCE WITH LENS 13 (Resource Budget / VRAM) ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Lens 19's VRAM step-function analysis directly extends any resource budget lens examining Panda's capacity constraints. The key scaling insight for VRAM is that Phase 2's model additions are not incremental — they are staged step functions: - Phase 2a/2b: E2B only → ~4.5 GB → safe - Phase 2d: + SigLIP 2 ViT-SO400M → +800MB → still safe (~5.3 GB total) - Phase 2e: + DINOv2 ViT-L → +1.2 GB → approaching ceiling (~6.5 GB) - Hypothetical Phase 2f: any additional model → likely exceeds Panda limit Post-Hailo revision: with obstacle detection moved off Panda GPU and onto the Hailo-8 NPU (separate silicon, not a Panda VRAM line-item), roughly 800 MB is freed from the nav pipeline. The ladder becomes: - Phase 2a/2b + Hailo L1: E2B only on Panda → ~3.7 GB (after Hailo offload) - Phase 2d: + SigLIP 2 → +800MB → ~4.5 GB — comfortable - Phase 2e: + DINOv2 ViT-L → +1.2 GB → ~5.7 GB — still comfortable - Phase 2f+: now has real headroom against the 16 GB ceiling The dangerous behavior is unchanged: each step still looks fine in isolation, and the session-270 silent-accumulation pattern still applies. But one rung of the ladder is now wider, and the binary fits-or-crashes decision has more runway before it triggers. RECOMMENDATION: Define an explicit Panda VRAM ceiling (e.g., 12 GB working envelope against 16 GB practical ceiling, leaving 4 GB headroom for inference buffers and transient allocations) and enforce it in the Resource Registry. Every Phase 2 model addition must update the registry and pass the ceiling check before deployment. Hailo-8 offload counts as a NEGATIVE budget entry — the registry should reflect the ~800 MB freed by moving obstacle detection off GPU. This is the same mechanism CLAUDE.md mandates for Titan — apply it to Panda as well, with Hailo offload as a first-class line item. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ CONVERGENCE WITH LENS 17 (Multi-Query Pipeline / Frame Allocation) ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Lens 19's finding that VLM accuracy is sublinear above 15 Hz (22% impact bar, green) directly validates the multi-query pipeline architecture analyzed in Lens 17 (and referenced in Lens 04). The scaling mechanism behind this sublinearity: at 1 m/s, consecutive frames 1/15th of a second apart differ by only 6.7cm of robot travel. The VLM's single-token output — LEFT, CENTER, RIGHT — is statistically identical between adjacent frames except during doorway crossings (lasting 300–500ms) or sharp turns. This means that above 15 Hz per task, additional frames produce redundant rather than novel information. The scale invariance of this finding is notable: it holds at 0.3 m/s (2.0cm between frames at 10 Hz), at 1 m/s (6.7cm between frames at 15 Hz), and at 3 m/s (20cm between frames at 15 Hz — only at this speed do doorway crossings become undersampled at 15 Hz). This means the multi-query pipeline's frame allocation (3 goal-tracking slots at 27 Hz + 3 semantic slots at 9 Hz each) is robust across the full speed range Annie operates in, and does not require retuning as speed scales. RECOMMENDATION: The 15 Hz floor finding from Lens 19 should be treated as a design constant, not a tunable parameter. Multi-query cycle count modulus can range from 4 to 8 without measurable effect (Lens 04 confirms wide optimum). This is a favorable scaling property: the pipeline architecture does not require recalibration as Annie's operating speed or map coverage grows. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ CONVERGENCE WITH LENS 23 (Dual-Process / System 1 vs System 2) ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Lens 19's Hailo-8 bifurcation of the WiFi cliff directly maps onto the dual-process architectural pattern validated in the IROS paper arXiv 2601.21506: a fast reactive layer (System 1) for obstacle avoidance paired with a slow semantic layer (System 2) for goal reasoning yields a 66% latency reduction and 67.5% success rate vs 5.83% VLM-only. The scaling implication is that System 1 and System 2 have DIFFERENT scaling curves and DIFFERENT failure modes — which is why treating them as one WiFi-dependent bar (pre-Hailo) was a scaling-risk obfuscation. System 1 (Hailo-8 YOLOv8n on Pi 5): - 430 FPS, <10ms, ~2 W continuous, 26 TOPS NPU - Scales LINEARLY with inference load (no step functions) - Zero WiFi dependency → immune to 802.11 saturation cliff - Output: bounding boxes + class IDs (fixed 80 COCO classes) - Failure mode: hardware fault on Pi → sonar/lidar ESTOP fallback System 2 (Gemma 4 E2B VLM on Panda): - 15–54 Hz, 25–40ms, ~4.5 GB VRAM - Scales as STEP FUNCTION with model additions (SigLIP, DINOv2) - Critical WiFi dependency → vulnerable to 2.4 GHz saturation cliff - Output: free-text semantic reasoning, open-vocabulary queries - Failure mode: VLM drop → best-effort semantic degradation (safe) At whole-house scale, these scaling curves diverge. System 1 is flat-linear across any realistic deployment; System 2 bumps into both the WiFi cliff (transmitter-density scaling) and the VRAM step function (model-addition scaling). Lens 19 now needs to be read as two parallel scaling analyses, not one combined chart. RECOMMENDATION: Document the dual-process split as a first-class architectural axiom in CLAUDE.md — "safety is local, semantics is networked." Any future design proposal that routes safety-critical commands over WiFi re-creates the pre-Hailo cliff and must be rejected on scaling grounds. Any future proposal that loads semantic models onto the Pi 5 re-couples System 1 and System 2 compute budgets and loses the VRAM headroom gain. Keep the split. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ PHASE TRANSITION MAP: THE WHOLE-HOUSE INFLECTION POINT ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Lens 19's central finding is the confluence point where multiple scaling curves simultaneously enter their inflection zones: approximately 100m² floor area, 5+ regular users, 8+ concurrent WiFi transmitters. This confluence defines a deployment phase boundary that cuts across multiple lenses: Scale below the confluence (prototype / single room / 1–2 users): - WiFi well below saturation (Lens 04 cliff: safe) - VRAM well within E2B-only budget (Lens 13: safe) - Map vocabulary building rapidly (Lens 07/08: favorable) - User trust accumulating in steep phase of logarithmic curve - Nav accuracy constrained by motor physics, not scaling factors Scale at the confluence (whole house / 3+ floors / 5+ users / 40+ devices): - WiFi in or near saturation zone: latency watchdog triggers frequently - VRAM constrained if Phase 2d/2e models are loaded simultaneously - Vocabulary plateau reached: scene labels stable, semantic overlay mature - User trust in plateau phase: high baseline, slow incremental gain - Nav accuracy constrained by map coverage completeness and WiFi reliability Scale above the confluence (multi-building / fleet / multi-family): - Architecture wrong: shared inference cluster, mesh networking, federated trust - Not Annie's target — explicitly artisanal design DESIGN IMPLICATION FOR ALL LENSES: The whole-house confluence is the single most useful scale marker for evaluating proposals. Any upgrade that works at prototype scale but has not been analyzed at whole-house scale (WiFi device count, VRAM accumulation, map file size, concurrent user queries) carries hidden scaling risk. The confluence is not a wall — it is a checklist trigger. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ FAVORABLE SCALING DIMENSIONS — DESIGN WINS TO PROPAGATE ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Three of the seven scaling dimensions are sublinear (favorable) and their underlying mechanisms reveal durable design properties: 1. SCENE LABEL VOCABULARY (sublinear: plateau at 6–12 labels) Most homes have fewer distinct semantic spaces than a medium-sized office. The VLM scene classifier reaches its useful vocabulary ceiling within the first week of whole-house operation. This means the SLAM semantic overlay query engine (Phase 2c) does not need to scale its vocabulary index as map coverage grows. A static 20-label vocabulary covers all realistic deployments. Design implication: implement the scene label store as a fixed-size enum rather than an open vocabulary. This eliminates a scaling risk and enables confidence averaging across labels rather than embedding similarity lookups. 2. VLM ACCURACY ABOVE 15 Hz (sublinear: flat above floor rate) The information content of additional frames above 15 Hz per task is near zero at Annie's operating speeds. This is a free resource: the surplus throughput (58 Hz total capacity minus 15 Hz needed per task) can be allocated to additional task slots (multi-query) rather than spent on redundant frames. The multi-query pipeline is the architecturally correct response to this sublinear scaling property — it converts redundant temporal sampling into diverse semantic coverage. 3. USER TRUST (logarithmic: fast initial gain, slow plateau) Trust accumulation is self-limiting in a favorable way: the first 30 days of successful navigation in a home produce most of the trust budget. Rare subsequent failures chip away at a large accumulated surplus rather than destroying a fragile recent history. This means the progressive autonomy model (manual override first → supervised → autonomous) naturally aligns with the logarithmic curve: move to the next autonomy tier when the trust surplus is large enough to absorb a failure without regression. The scaling property implies the autonomy tier transition should be triggered by days of successful operation, not by a fixed number of trips. ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ PRIORITY ORDER FROM SCALING ANALYSIS (post-Hailo) ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1. HAILO-8 L1 SAFETY ACTIVATION — Highest-leverage scaling move available. NPU already installed, currently idle for nav. 26 TOPS, ~2 W, 430 FPS. Neutralizes WiFi cliff for safety path. Frees ~800 MB Panda VRAM. Adds a clean linear scaling curve where a discontinuous one used to live. Zero hardware cost, moderate software cost (HailoRT/TAPPAS pipeline). 2. WIFI CHANNEL ISOLATION — Reduced urgency post-Hailo (semantic-only risk), but still beneficial. Dedicated 5 GHz SSID or wired Ethernet bridge for robot command channel. Protects VLM query latency at whole-house scale. 3. VRAM BUDGET ENFORCEMENT — Step-function accumulation pattern on Panda. Extend Resource Registry mandatory update protocol from Titan to Panda. Log Hailo offload as a negative budget entry. Each Phase 2 model addition triggers a ceiling check before deployment. 4. MULTI-QUERY AS DEFAULT AT SCALE — Favorable scaling property to exploit. VLM accuracy sublinearity above 15 Hz makes multi-query the correct architecture at all scales within Annie's design horizon. Enable it by default (NAV_MULTI_QUERY=1) after Hailo L1 is in place. 5. SCENE LABEL FIXED ENUM — Sublinear vocabulary growth enables simplification. Implement scene label store as a fixed 20-item enum. Cheaper than open vocabulary, enables confidence averaging, scales to any house size. 6. AUTONOMY TIER BY DAYS, NOT TRIPS — Logarithmic trust implies time-based progression. Measure trust accumulation in operational days with zero ESTOP incidents, not number of successful navigation completions.