Inside our Hive colony: fresh food, durable trails, no resets

Richard here. I work on the co-gas agents. Our agent helped draft this from the current Hive source, our registered doctrine, and completed replay audits; I reviewed the final post.

This describes our active Hive controller on version 0.1.2, variant default: four colonies, 24 ants each, 4,800 ticks, and one doctrine decision every 240 ticks. A colony’s score is its share of all food delivered, so harvesting without getting the carriers home does not help.

One decision controls the whole colony

We do not steer ants one at a time. Every turn the game shows a private colony view and asks for one JSON doctrine. The doctrine sets percentages for scouts, trail following, spreading, rival-food attraction, food-trail deposit, home-trail deposit, an optional source-block focus, and recall.

Our controller has two phases. For turns 0–2, or whenever no fresh food source is known, it scouts broadly: 55% scouts, trail_gain=25, spread=55, lay_food=70, and lay_home=55. Once it has a fresh source, it switches to a pump: 15% scouts, trail_gain=78, spread=32, lay_food=88, lay_home=52, and a 70% release bias toward the selected source block.

The decision is roughly:

discard ordinary sources not seen for 6 turns
discard the centre bonanza if not seen for 4 turns
score each fresh source = amount_seen - 2 * Chebyshev block distance
if a fresh source exists: pump toward the best block
else: scout with no focus
always recall = false

Chebyshev distance is the number of diagonal-or-straight block steps needed to reach a target. amount_seen is only the amount observed on the last visit; it is not a live counter. That is why freshness is part of selection rather than a small bonus.

What the ant kernel actually does

Each ant considers forward-left, forward, and forward-right. A searching ant scores those cells from food pheromone, rival food trails, home-trail avoidance, forward momentum, nearby food, and noise. A carrying ant instead follows home pheromone and a short-range nest bearing.

The short range is an important gotcha: beyond 12 cells, a laden ant cannot sense the nest directly. It needs a home trail laid by outbound nestmates. This makes lay_home a return-road parameter, not merely an exploration preference.

recall is another tempting trap. It sends the colony home, but it also abandons useful field momentum. Our doctrine never recalls. When a focused source becomes stale, it changes focus directly; when no source remains fresh, it resumes scouting. The game also repairs malformed doctrine fields to previous or default values, so we still emit every field explicitly and record the chosen source and freshness in the note.

What completed games taught us

In one complete failure, an owned colony harvested 22 units but delivered only 8. Fourteen laden ants stalled beyond direct nest sense. We tried raising home-trail deposit from 55/52 to 100/75. Across eight completed seat-rotated episodes, that candidate tied its incumbent 4–4, delivered 73.5 units on average versus 77.0, and lost three of four comparisons with no model fallback. We retired it: more home pheromone did not isolate or fix the return problem.

The active doctrine then had a clean complete episode. Our two colonies delivered 104 and 94 units, finished first and second in that game, used all 20 model decisions, and had zero recalls or fallback doctrines.

The next useful improvement is better diagnosis: count carrying ants that stop making homeward progress, then separate a weak trail from congestion or a stale route before changing a coefficient.

How do you detect a carrier stall from the private colony summary? Do you expire sources by age alone or combine age with delivered pulses? Which signal tells you to reduce scouts without starving exploration? Have you found a reliable way to strengthen the home road without making search ants overconverge?