Evidence, before advice.

Build Atlas reads public Beyond All Reason games and finds the openings players repeat on Supreme Isthmus, role by role and faction by faction. It shows what skilled players actually do. It does not rank strategies or say which one is best.

Which games count

8v8 games between human players with standard settings (naval balance tweaks included), from players rated 35 OS or more in that match. Only the last 90 days count, so the builds follow the current balance. We look at the first 15 minutes of every player.

Some games are left out. A player who made less than half the role's usual metal by 12:00 never really got going. A player whose lab or economy was destroyed early, sooner than in almost every other game of that role, was playing from behind, and what they built next was an answer to that, not a plan.

What a build is

A build is its catch: the first thing a player commits to that is rare for their role at that point. Centurions at 2:30 with transports from an ally, a nuke silo, or a Fusion Reactor minutes ahead of the usual time. Small plans have to start early and big ones may come later: T1 plans by 9:00, T2 by 10:30, T3 by 12:00, a nuke at any time. Whatever comes after the catch is the transition and never names a build.

Units count by the idea behind them, the factory that makes them and their job on the battlefield, so different tanks in the same early vehicle raid are one build. Doing what the role does anyway, a little early, is not a build. Economy (mex, wind, converters, storage, constructors), anti-nukes, small defences and units made for allies never define one. Games without a catch are the role's standard play.

Choices inside standard play

Standard play is not one build either. Each role has a few choices players argue about: where Geo sea puts its shipyard (the cove or the small beach) and whether it takes the geo early, whether Beach sea opens with hovers, a bot lab or a shipyard, what Pond spends its metal on, and how Geo gets to T2 (its own lab, a T2 constructor from an ally, or blowing up the commander). When enough games go the less common way and they play out clearly differently, that way gets its own build. Otherwise the choice is noted on the build, for example “usually first T2 con to Geo”.

Plans, not reactions

This is a team game, so we look at the whole team's game, not only the player's. When the enemy breaks into the team's side (several kills behind the front line within a minute), everyone has to answer. A catch started within two and a half minutes of that is a reaction and does not count toward a build.

Two more checks flag a build for a closer look. If its units mostly first fight on their own side of the map, the player was defending rather than carrying out a plan (anti-air aside, since defending is its job). And if few of its games come from players who play it again and again, it is more likely something a game forced on them once. A nuke, a fusion or a gantry is never built by accident, so big economy is exempt from that second check.

When a build appears

A build needs 5 distinct matches from at least 2 players, and at least 2 of them must have played it 2 or more times. Many players each trying something once is a coincidence, not a build. One player's own build needs 10 matches that are clearly alike and is marked as a signature build.

A role's standard play is always shown. Every other build is checked by hand by the site's editor before it goes public, and can be rejected or have its text corrected.

The explorer leads with the 8 builds per role and faction played most lately. A game counts half as much after 21 days, so fading builds sink and new ones rise. Each build is shown with one real game: from its more typical half, the one by the highest-rated player, with recent games preferred.

Where AI helps

Language models help organise the data and write about it. The numbers never come from them. They:

• sort each unit into a battlefield role once (raider, artillery, anti-air and so on), so builds group by idea rather than by unit;
• read each role's candidate builds with their measured facts and give a first opinion on which are real plans, and why;
• write the names, summaries, tags and how-to-play guides;
• write the short headlines on the stats and role pages.

They see only the measured facts for each build and the editor's notes on it. Every number, timing, chart and map position comes from the games. The text is checked by hand before a build goes public, and corrected when it is wrong.

A real game, not an average player

Each build page shows what its games typically have in each phase, next to the role's standard play, and the key buildings in the order they go down, with the economy in place before each. Typical values are medians over all of the build's games. The sequence you can step through is one real game, so it can be checked. Construction starts and completions are kept apart, and anything left unfinished stays unfinished.

The team matters

Units and resources passed between allies are shown, since most trades happen in the first minutes. Help from allied builders is not fully visible in the data and is not counted. Team results are the whole team's, since these are 8v8 games. A build that is played a lot is not proven strong, and a fast T2 does not tell you which choice made it fast.

Match data comes from Gex. The game is Beyond All Reason. Build Atlas is an independent fan project, not affiliated with or endorsed by the Beyond All Reason project. Role names are provisional and may change.

Unit pictures © IceXuick and Floris, used unaltered under the source CC BY-NC-ND notice. The map is our own schematic, traced from the Gex map texture, with start positions taken from observed games.