The computer commanders are meant to feel active — they build, research, expand from the first cycle, and commit real attacks. No invisible cheats; difficulty is a set of honest, tunable knobs.
Each cycle, every AI commander runs the same five-phase routine:
Forgiving. Slower expansion, smaller attacks, over-defends, and frequently skips optimal actions. Built for new players.
Fair and beatable. Expands and attacks like a reasonable human opponent, without early runaway snowballing.
Aggressive and efficient. Punishes mistakes, expands from cycle zero, and hunts homeworlds — with no hidden bonuses.
Difficulty tiers are kept in a strict order — easy < normal < hard — across every knob, so the curve is always monotonic.
Difficulty is not a script — it's this centralized set of parameters. Higher aggression, lower reserves, and earlier strike timing make Hard genuinely harder without ever letting it see hidden information.
| Parameter | Easy | Normal | Hard | What it does |
|---|---|---|---|---|
| Build efficiency | 0.78 | 0.92 | 1.00 | Chance per cycle it acts on its build plan |
| Expansion aggression | 0.60 | 0.80 | 0.97 | Chance per world per cycle to launch expansion |
| Attack aggression | 0.50 | 0.72 | 0.92 | Chance per world per cycle to attack when viable |
| Defense reserve | 0.50 | 0.38 | 0.25 | Fraction of ships always kept home |
| Max attack commit | 0.45 | 0.60 | 0.80 | Max fraction of a stack sent in one attack |
| Min ships before attack | 9 | 6 | 4 | Force required before attacking |
| Attack advantage req. | 1.50× | 1.22× | 1.05× | Attacker/defender ratio needed to strike |
| Earliest homeworld strike | cy 20 | cy 10 | cy 4 | No homeworld assault before this cycle |
| Imperfection chance | 0.25 | 0.10 | 0.03 | Chance to skip a noncritical optimal action |
| Tech preference | 0.70 | 0.85 | 1.00 | Value placed on Tech Labs vs. extra Factories |
The "imperfection chance" is what makes lower difficulties feel human rather than broken — an Easy AI occasionally passes on the optimal move instead of playing flawlessly-but-slowly.
A single planet's attacking stack is capped by garrison attrition, so a plateaued homeworld can't be cracked by one world alone. The AI handles this with a separate, lower siege advantage threshold (Easy 2.2× · Normal 1.7× · Hard 1.4×) that measures the combined available ships across all in-range worlds against the target's defenders. When that pooled force is enough, it commits a coordinated multi-world assault — the same tactic a strong human uses to break a fortified homeworld.
From a 16-seed simulation harness against a competent human baseline. Games are shown in cycles, with an approximate wall-clock at the default cycle timer.
| Map (planets) | Easy | Normal | Hard | ≈ minutes |
|---|---|---|---|---|
| Tiny (12) | 46 cy | 33 cy | 27 cy | ~13–23 min (30s timer) |
| Medium (35) | 60 cy | 53 cy | 43 cy | ~32–45 min (45s timer) |
| Large (60) | 64 cy | 65 cy | 48 cy | ~48–65 min (60s timer) |
100% of simulated games now resolve — earlier builds routinely ran past 300 cycles with no winner. Human win rate also falls as difficulty rises (on Tiny: Easy 3/6 → Normal 1/6 → Hard 0/6), confirming the curve is real rather than cosmetic.
Hard is aggressive and efficient, but it plays by the same rules you do — it sees the same fogged map, spends the same gold, and fights with the same combat math. There are no resource bonuses or vision cheats; the challenge comes entirely from tighter, earlier, better-committed decisions.
An Adaptive difficulty is present as a documented stub — currently identical to Normal — reserved for a future opponent that reads the match state and tunes itself to the player. It's labeled honestly rather than pretending to be something it isn't yet.