Why Stellarch Runs Bots, and Bans Yours

We're building a card game where bots draft and fight at a high level. We will never knowingly let you run one.

That sounds like a contradiction. It isn't. Both halves come from the same rule the whole game is built on: it has to stay fair.

---

Fair warning, this one goes under the hood. If you only want the short version, here it is: the only bots in Stellarch are ours, they keep the game alive and balanced before launch, and we will never let a player run one to farm rewards. Everything below is the how and the why.

Here is where all of it lands. The play screen, where you pick a league, spend energy, and queue a match:

---

First, why a card game needs bots at all

Stellarch isn't open yet. We're pre-launch: the card pool and the combat engine are built, and the waitlist is filling. That creates two problems before a single player walks in, and one machine solves both.

Opponents before there's a crowd. A new game with empty matchmaking feels dead. If your first queue spins forever, you're gone. So we built an opponent that fields a credible, legal, league-appropriate team on demand. The showcase battles you can watch on our site right now were all drafted by this bot.

Balance before the doors open. You don't balance a card game on a spreadsheet. You balance it by playing it thousands of times and watching what wins too much. We don't have thousands of humans yet. We have bots playing hundreds of thousands of matches every hour against each other, and every match is a data point telling us which card is quietly broken.

A bot that plays the game well is also the fastest, cheapest balance lab you will ever build.

Keep that word in mind: our bot. It matters later.

---

Ten tiers, rookie to ruthless

The bot runs on a difficulty ladder from 1 to 10, and tier 1 is not "random."

Random is lazy and unteachable. Tier 1 makes recognizable mistakes: an archer parked in the front line, a commander in the wrong color, ignoring a ruleset that punishes the whole board. You can look at it and understand exactly what went wrong. That is the point. The floor teaches.

Tier 10 is the opposite, and it refuses to be predictable. If it always picked the single best lineup, you would memorize the counter in ten games. So at the top it samples from its strongest few options instead of locking to one.

---

How one pick actually happens

Drafting a 7-card team (a commander plus six monsters) is a search through a brutally large space. Here is the shape of it, minus the recipe.

Commander first. The commander sets the team's color and identity, so the bot weighs every legal commander before placing a single monster. Prune too early and you amputate entire strategies.

Then it builds, slot by slot, keeping the strongest partial teams at each step and letting them branch. It protects small teams on purpose: a lean two-card lineup gets to compete against a full six-card board on its own merits instead of being buried by raw stat totals.

It reads the situation. This is the part that matters most:

• Earthquake ruleset active? It reaches for flyers, because everything grounded bleeds every round.
• Facing a magic-heavy field? It brings the anti-magic answers.
• Already has a healer down? A tank gets more valuable, because now it can be kept alive.

It scores combos, not just cards. Two cards that win more often together than apart earn a bonus when they share a team. Nobody typed "these two synergize." The bot learned it from its own match history and started reaching for those pairs.

It places everyone right. Tanks forward, snipers and healers in the back where they survive, then a final pass that reshuffles the chosen cards into their best order.

At the higher tiers it goes further: it drafts a bracket of candidate teams, simulates them fighting each other, and crowns the survivor. Draft sixteen, let them brawl, keep the champion.

---

If you want to see the combat all this is drafting for, the rulesets and abilities are in the docs. Every match resolves from a seed, so the same teams always produce the exact same fight, which is also how we prove a result was real.

---

How it gets smarter: self-play and three generations of models

It trains the way the strongest game AIs do. It plays itself.

Every hour it builds a diverse field of teams and runs them against each other across thousands of matches. Win or lose, each match becomes a labeled example: this team, in this spot, beat that team. Feed enough of those in, and a model can predict how good a draft is before a punch is thrown.

We've moved through three generations of that model, each able to see something the last one couldn't:

• Logistic regression. The starting point. Fast, learns the weight of each factor on its own, but mostly blind to interactions ("card A is great only when card B is also present").
• Factorization machines. These learn pairwise interactions automatically, which is exactly what a game built on card combos needs.
• A deep cross network. The newest tier, able to model deeper, higher-order interactions between many factors at once.

Each new model only gets promoted if it beats the last one on data it has never seen. No vibes. It earns the slot or it doesn't serve.

---

Proving it's actually smart

A model can look great on paper and still draft like a clown. So we don't trust the score, we test the behavior.

There is a standing battery of checks the bot has to pass, and they are blunt on purpose. Put Earthquake on the board: does the team actually carry a flyer? It must. Cap the mana low: does it ever overspend? It can't. Ban a color: does a banned card sneak in? Never. Drop it into a magic field: does it bring the counter? It should. Any red light, and the build does not ship.

The model is allowed to be wrong. The behavior is not.

Then there is the difficulty curve itself. It isn't enough that tier 7 feels harder than tier 5. We make every tier prove it, running the ladder against a fixed top-tier opponent and confirming the win rates climb in order. A slider nobody verified is just a number. Ours has to earn its slope.

---

Now the line: why you can't bring your own bot

Here is the fair question. We just spent this whole post bragging about our bots. So why ban yours?

Because these two things are not the same thing, and the difference is the entire game.

Our bots take nothing out. They have no wallet. They earn no rewards. They are labeled as bots, so you always know when you're facing one. They exist to give you a fight and to keep the cards balanced, and that is all.

A player farm-bot takes everything out. Picture someone running the same kind of drafting brain we built, except across fifty accounts, around the clock, pointed at one thing: harvesting rewards while they sleep. That isn't playing. It's mining. And it does three ugly things:

• It drains the reward pool that's supposed to go to people who actually showed up.
• It turns the ladder into a fight against scripts instead of people.
• It makes the real winner "whoever automated hardest and rented the most accounts," which is pay-to-win wearing a different coat.

This genre has watched that exact movie before. A great core loop, hollowed out from the inside, until the people who came to play are grinding against the people who came to extract.

We're not doing that. The only bots in Stellarch are ours, and they farm nothing.
It may be an ongoing war, but we'll always be willing to fight that battle. That's simply the only reasonable choice when trying to build a fun, competitive game.

That is the same rule, pointed both ways. We build bots so the game is alive and fair before launch. We ban farm-bots so it stays fair after. Your edge is supposed to come from a sharp draft and knowing which cards matter, not from a script that never sleeps.

Fair, not free. Cards cost money and a deeper collection is a real edge, but knowledge beats raw spend, and no machine gets to grind that away from you.

---

Before you go, a peek at what's next

While the bots train, I'm building the guild foundry right now. Name your colony, lock a clan tag, pick your crest:

Work in progress. Guilds, vaults, and brawls are the social endgame.

---

What it means when the airlock opens

When you sit down, your first matches won't be swings at empty air. They will feel like real opponents with real plans, at a tier that fits you.

The meta you walk into will already have taken millions of bot games' worth of punishment, so nothing should be obviously, miserably broken on day one. The low tiers will make legible mistakes you can punish and learn from. The top tier will keep you honest by refusing to play the same hand twice. And every reward on the ladder will have gone to a human who earned it.

---

One real question for you

We're still deciding where the bots sit at launch. Should there be a boss tier above 10, something deliberately, unfairly strong for players who want to get wrecked on purpose? Or does the ladder topping out at "near-optimal but unpredictable" feel right?

Tell me in the comments. This one isn't decided, and your answer actually moves it.

---

Get in early

• Join the Discord and watch the build happen → https://discord.gg/K3geA9hD28

---

We tune these bots in the open, every week. The next dispatch goes into what they taught us about balance: which cards were secretly too strong, and what we did about it.

If you want to be in the room when the airlock opens, the closed alpha runs off the waitlist.

---

Enlist & stay in the loop

• Join the closed-alpha waitlist → https://stellarch.io/en/#waitlist
• Follow Hive account@stellarch on Hive → @stellarch
• Join the Discord → https://discord.gg/K3geA9hD28

---

Latest dispatch from Stellarch

• Meet Stellarch: A Card Game, Game First

---

Stellarch is a deterministic trading-card game on Hive: draft a 7-card team, outplay your opponent, climb the ranked ladder. Browse the cards · Read the docs · See the leaderboard