A weighted search runs while other paradigms show assigned diagrams
Read the explanation
The saved grid has twelve rows and twelve columns, giving one hundred forty four cells. The default density slider value twenty eight is an obstacle count, not twenty eight percent. It therefore blocks twenty eight of one hundred forty four cells, about nineteen point four percent, leaving one hundred sixteen unblocked. At two pixels per cell the bars measure two hundred eighty eight, fifty six and two hundred thirty two. Start and goal are excluded from obstacles. A seeded generator repeats the same maze for the same count, so reset does not choose a fresh random maze. The local two dimensional canvas draws actual cells and explored search positions. Search ranks frontier nodes by distance already traveled plus weighted Manhattan distance. From row zero column zero to row eleven column eleven, the direct Manhattan distance is twenty two moves. At weight one point five the starting heuristic is thirty three; at weight three it is sixty six. At five pixels per unit the bars measure one hundred ten, one hundred sixty five and three hundred thirty. A higher weight changes ordering and can sacrifice shortest path guarantees. The label calls every found result optimal, but that wording does not prove optimality. Path length counts stored coordinates including the start, so twenty two moves would produce twenty three listed path positions. Switching to a neural diagram draws layers of three, five, four and two nodes: fourteen circles. Transformer mode draws a six by six heatmap, thirty six cells, with colors from a sine function rather than computed query and key attention. At eight pixels per count neural nodes measure one hundred twelve, attention cells two hundred eighty eight and their difference twenty two one hundred seventy six. Generic modes assign latency one point two milliseconds and token or iteration labels, while search always displays three point four milliseconds. These are not measured performance. Native controls and JSON digest remain local; export path found tests nonempty path, even when a search ended blocked.