LENS 06: Second-Order Effects — Cross-Lens Convergence Notes
============================================================

## Strongest convergences

### Lens 16 (Build the Map to Remember)
This is the clearest convergence in the entire analysis. Lens 16's insight that "we build the map to navigate but it becomes load-bearing memory" is directly confirmed by the second-order cascade here: semantic map annotation (Phase 2c) starts as a navigation aid (Annie knows she is in the kitchen → she navigates correctly) but within months becomes a household spatial memory service. Mom asks about the house; she does not ask about navigation. The map's value as navigation aid is eclipsed by its value as witness. Lens 06 provides the causal chain that Lens 16 names but does not fully trace: VLM labels → SLAM grid cells → room emergence → voice query possibility → Mom discovers the use case → spatial witness role becomes load-bearing → map corruption feels like Annie amnesia.

Design implication shared by both lenses: the map must be versioned, timestamped at the observation level, and serialised independently of the robot chassis. "Annie remembers the house after the chassis breaks" is a feature, not a side effect.

### Lens 20 (Multi-modal convergence / proactive care)
The three-system composition identified here — Context Engine episodic memory + semantic map spatial memory + SER emotional state — is exactly the convergence Lens 20 describes. Lens 06 names the specific mechanism: these three systems were not designed to compose, but composition falls out naturally because all three write to Titan's shared context. The second-order effect is that proactive care (Annie suggesting where Mom's glasses are before being asked) emerges without any new feature development. It requires only that the semantic map's observations be queryable by the Tier 1 planner alongside Context Engine entries.

Risk flagged: this composition is not tested. Convergence could produce hallucinated confidence ("Annie says the glasses are on the nightstand" when the observation is 6 hours old). Observation staleness must be a first-class concept in the semantic map schema.

### Lens 21 (Voice-to-ESTOP gap / Mom's safety)
Three specific intersections:

1. Faster navigation (third-order consequence of obstacle confidence accumulation) worsens the ESTOP gap. If Annie accelerates in known-clear rooms (second-order: confidence accumulation → speed increase), the window between Mom appearing and collision risk shrinks. The 10 Hz "person" VLM obstacle label was designed for 1 m/s navigation. At higher speed, the same latency covers more distance. The ESTOP threshold may need to become speed-dependent: tighter at high confidence/high speed, looser at cautious mode.

2. Spatial witness role means Mom interacts with Annie more, including in Annie's path. Increased interaction frequency increases the statistical likelihood of a close-approach incident. The ESTOP gap matters more when Annie and Mom are in the same rooms more often.

3. Mom asking "Annie, what's in the kitchen?" while Annie is navigating creates a dual-attention demand on Tier 1. The LLM answering a spatial query and planning a route simultaneously (both via Titan) could introduce latency spikes into the planning tier. Tier 3/4 reactive navigation must remain completely independent of Tier 1 load; this is already the design but must be explicitly verified under concurrent query load.

### Lens 11 (Adversarial view / open-source race)
The third-order effect "open-source race to zero" is the Lens 11 adversarial argument applied to this specific architecture. Multi-query VLM + SLAM + semantic map is not proprietary. The research explicitly describes off-the-shelf components (VLMaps, OK-Robot, AnyLoc). Within 12-18 months, commodity home robots will ship this stack. Lens 06 identifies the true moat: not the architecture but the household-specific accumulated map, the family's trust relationship, and the deep integration with Context Engine's episodic memory. These cannot be replicated by a new robot entering the home.

Strategic implication: Annie's Phase 2 should prioritise depth of integration (semantic map ↔ Context Engine ↔ voice queries) over breadth of VLM capabilities. A thin semantic map that answers "where are my glasses?" reliably is more defensible than a comprehensive VLM pipeline that does not connect to conversational memory.

### Lens 20 (WiFi brownout trust damage) — NEW, directly invoked by Hailo cascade
The positive cascade identified in this lens (Hailo-8 activation → L1 safety leaves WiFi path → Mom's trust stabilises → usage up → memory richer) is the constructive counter-move to Lens 20's warning that 2-second freezes during WiFi brownouts damage the trust curve measurably. Lens 06 names the specific sequence: trust-curve stabilisation is not a separate workstream — it is a second-order consequence of moving obstacle detection off WiFi. One hardware activation addresses a Lens 20 harm without any Lens 20-specific engineering. Convergence is strong and actionable.

## Weaker / one-directional convergences

### Lens 04 (WiFi cliff / latency)
Panda-Titan split for semantic vs. reactive tasks means Tier 1 planning crosses WiFi. Lens 04's finding (VLM rate insensitive above 15 Hz) suggests the Titan strategic planner can tolerate 4-8 ms WiFi round-trip without degradation. However, the concurrent spatial query scenario (Mom asks Annie a question while navigating) could create Titan load spikes that are invisible to Lens 04's single-task framing. Cross-lens risk: Lens 04 evaluates VLM rate in isolation; Lens 06 identifies concurrent multi-client Titan load as an emergent scenario not covered by that evaluation.

New cross-lens risk surfaced by the Hailo cascade: activating Hailo adds HailoRT + TAPPAS + HEF compile toolchain + Hailo firmware updates + Pi-kernel driver compatibility as new failure modes. Lens 04's sensitivity model must be extended to include firmware drift on the L1 safety path — a regression in Hailo firmware or a kernel bump that breaks the driver would silently collapse L1 back onto WiFi-dependent VLM. Lens 04 should own "what breaks if Hailo is silently not running?" as a first-class sensitivity question.

### Lens 21 — Stakeholder (Hailo second-order additions)
Beyond the ESTOP-gap intersections already listed, the Hailo cascade adds a stakeholder-facing argument: "Annie stops even with no WiFi." That sentence is persuasive to Mom in a way that "we have tiered fallbacks" is not. Lens 21's consent narrative is cheaper to earn once the safety story is no longer "trust the network." Conversely, the new operational cost (maintenance load from HailoRT/TAPPAS) lands on Rajesh, not Mom — asymmetric distribution of cascade benefit vs. cascade cost between the two primary stakeholders. Lens 21 should track this asymmetry.

### Lens 03 (llama-server embedding blocker)
The practical blocker identified in Lens 03 (llama-server does not cleanly expose intermediate VLM embeddings for multimodal inputs) directly gates Phase 2d (embedding extraction + place memory). Lens 06's third-order "home historian" and "map-as-identity" effects only materialise if Phase 2d ships. Lens 03 identified deploying SigLIP 2 ViT-SO400M (~800 MB VRAM on Panda) as the workaround. Lens 06 adds urgency: the place memory capability is architecturally upstream of the most profound second-order effects (spatial witness + home historian). Unblocking Lens 03 is not a capability enhancement; it is the prerequisite for the social transformation described here.

## Key design questions surfaced by cross-lens analysis

1. OBSERVATION STALENESS: Semantic map entries need timestamps. Annie must qualify spatial answers with age-of-knowledge. Without this, the social trust gap (Lens 06) and the proactive care composition error (Lens 20) both materialise.

2. CONCURRENT TITAN LOAD: Spatial query handling + Tier 1 nav planning should be tested under concurrent load. Tier 3/4 must be verified to stay responsive when Titan is busy.

3. SPEED-DEPENDENT ESTOP THRESHOLD: If obstacle-confidence speedup ships (Phase 2b), the lidar ESTOP threshold and "person" VLM detection sensitivity should be reviewed against the higher speed regime (Lens 21 intersection).

4. MAP SERIALISATION INDEPENDENCE: The semantic map must be serialised to Titan storage independently of the robot chassis (chassis ID, not robot identity). Chassis failure ≠ map loss.

5. CONSENT GRANULARITY: Room-level opt-out ("don't record in the bedroom") should be a first-class config option before Phase 2c ships. The SLAM grid already has room-boundary labels; use those as access control zones.