How scoring works: position, travel, and direction

One number, recomputed constantly

Everything the swarm does comes down to a single score it computes for each open property, for each tractor, on every GPS ping. The machine claims the highest-scoring property it can. Understanding the score is understanding the engine.

The two halves of the score

The score blends two weighted parts:

• A position part — how well-placed is this property? This is where direction control lives (outside-in, inside-out, or nearest-first, measured from the depot anchor) and where density cohesion adds its same-block and close-pair pull. It answers "is this the right property to be at, given where the fleet and the storm are?"
• A travel part — how quick is it to get there? This uses the real road travel time from OpenRouteService, so genuinely reachable properties beat ones that merely look near.

Conceptually: score = (direction weight × position) + (travel weight × travel). The two weights are what let the fleet both spread out and sweep — turn up position and it favors strategy and cohesion; turn up travel and it favors raw proximity.

The behaviors are all the same score

The behaviors described elsewhere in this section aren't separate systems bolted on — they are all expressed inside this one score:

• Start-spread adds a fan-out preference on the opening move.
• Density cohesion boosts properties on the block you're already sweeping.
• Anti-steal holds back a property a peer is already working beside.
• Direction tilts the position part toward or away from the depot.

Then the claim

Once the best property is chosen, the tractor claims it with the atomic database operation — so scoring picks the best property and claiming guarantees only one machine gets it. If the claim is lost to a peer, the engine simply re-scores and offers the next best property.