Real RuVector, running in this tab

• VectorDB from @ruvector/wasm 0.2.1: HNSW index, cosine distance, 32 dimensions.
• Every experience is a real insert(vector, id, metadata).
• Every risk estimate is a real search(query, k=8). The memory panel shows the live count and the average time per search. Browsers round single timings to 0.1 ms, so only the average is meaningful.
• Power cycle uses saveToIndexedDB() and VectorDB.loadFromIndexedDB(). A page reload uses them too.
• Chapter 4 runs RuVector's SONA engine (built from source): each episode becomes a trajectory (startTrajectory, endTrajectory), forceLearn() clusters them with the ReasoningBank (k-means++), and findPatterns() returns the patterns. A second VectorDB stores the patterns. SONA's MicroLoRA and EWC++ are not used: their effect here was too small to matter.

The only JavaScript "learning" code

Risk = similarity-weighted average of the outcomes of the 8 neighbours that RuVector returns. Floor that resembles nothing in memory gets risk 0, so only real incidents create risk. The planner adds risk × 80 to a cell's cost, plus 25% of a neighbouring cell's risk as a safety margin for the robot's width.

A pass over a cell is stored as "safe" only if the robot drove it at operating speed (above 0.8 m/s) and was not recovering from an incident. Slow or recovering passes prove nothing, so they are not stored.

Simulated

• Robot physics: a kinematic model with surface grip, not a full physics engine.
• Perception: each floor cell yields a 32-d feature vector. Oil and polished floor look alike; cable ramps and drain grates look alike. The robot never sees surface names.
• Labels such as "motor oil" are for you, not for the robot.

Fair comparison

• Same layout, tasks, seed, physics, planner and time step on both sides.
• The baseline is the same robot with learning switched off, as deployed robots usually are.
• Headless evidence: npm run evidence runs 40 + 15 deliveries without rendering.