When a homeowner teaches the rover "that's a ravine, don't go within five feet, you'd suffer damage," they are giving it three things layered on top of each other:

A spatial fact — this specific area, here in this yard.
A semantic category — it's a ravine, a kind of thing.
A causal explanation — depth + irrecoverability = damage.

The third layer is what makes the rule portable. A 2D no-go zone teaches the rover to avoid that ravine. The causal explanation teaches it to avoid ravines as a category. A year later, during the overnight reflection cycle, when the rover encounters an unfamiliar depth discontinuity on the far side of the yard, it can reason "depth + irrecoverable → ravine pattern → apply five-foot buffer" without ever having been walked past that specific spot.

This is the actual differentiation. Imitation-learning approaches teach what to do. Pure geometric annotation teaches where the rules apply. Causal explanation teaches why — and why is what generalizes across yards the rover has never seen.

SLAM and occupancy grids do not go away. They are the floor: fast, dumb, reflexive, sensor-speed. When a depth discontinuity appears at 1.5 meters, the rover has 200 ms to brake — not three seconds to consult the VLM about whether this is a ravine or a sun-shadow.

The semantic map is the ceiling: meaning, names, categories, causal rules, learned from the training walk and the overnight reflection cycle.

The ceiling teaches the floor. A ravine annotation lives at the ceiling (semantic: depth + irrecoverable = damage) and gets projected down to the floor as a hardened "do not cross" polygon that the reflex layer enforces at sensor speeds. The cognitive layer does the meaning-making once; the reflex layer enforces the result fast, forever.

Mapped onto the four-tier architecture already published in ARCH-01: the ceiling lives at L4 (semantic reasoning) and gets re-read by the overnight reflection cycle. The projection happens at L3 (planner) at training time. The floor is enforced at L1 (reflex). The training walk is the moment all four layers get populated at once, top-down, by the homeowner.

Two persistence layers, both updated during the walk and re-coherent overnight:

The geometric map (floor) · Occupancy grid plus named-region polygons. Hardened constraints for the reflex layer: hard boundaries, no-go zones, distance buffers. Compiled down from the semantic layer at the end of the walk, and re-compiled during overnight reflection as understanding deepens.

The semantic memory (ceiling) · A structured store of named regions, categorical hazards, and causal rules in language. Each rule is stored with its full causal chain — not just the consequence. "Don't cross the road — traffic, would be hit, fatal" is stored as the whole chain, not just "don't cross road." The VLM consults this during planning. The overnight reflection cycle re-reads it, generalizes patterns ("multiple things called 'ravine' share depth + irrecoverable"), and writes new rules the next day's planner uses.

Most teach-by-demonstration approaches are one of three things, and each has a known ceiling:

Imitation learning — do what I do. Brittle to novel situations the demonstration didn't cover.

Pure geometric annotation — don't go where the polygon is. Doesn't generalize beyond the polygons drawn.

Pure language interface — tell me in English what to do. Hand-wavy without a shared visual reference.

This protocol is none of those. It is shared visual attention + gestural annotation + causal explanation + a confirmation loop. The homeowner and the rover are looking at the same pixels. The homeowner is drawing on those pixels with their finger. The homeowner is also explaining why the drawing matters. The rover proposes back what it understood. The exchange becomes the memory.