Should I choose Ranked, Training or Tests?
Ranked shows recorded competitive results for your published commanders. Training shows evidence from selected runs and generations. Tests lets you compare saved versions in controlled browser fights.
Training and Tests do not change ranked Elo. Premium increases commander capacity. It is not required to explore any of these analytics views.
| Source | Start here when you want to know |
|---|---|
| Ranked | How did a published version perform in competition? |
| Training | What did the recorded candidate do in a selected generation? |
| Tests | How do two saved versions compare against matched opponents? |
What do the four analytics views show?
Results summarizes the selected fights and outcomes. Units and lanes shows deployment and contribution measurements. Timeline shows when measured activity happened. Fights opens available records and replays.
Read coverage alongside the numbers. A narrow set of opponents or a few fights is limited evidence. The views describe observations, and you decide which question to investigate next.
How do I compare published commander versions?
In Ranked, choose the publications, dates and game rules you want to inspect. Compare compatible rules and check the number of fights and opponents covered. Open underlying fights when you want to examine a particular result.
For an illustrative comparison, suppose version A wins 12 of 20 fights and version B wins 11 of 20. That one-win difference does not establish that A is stronger. Check opponent coverage and use a matched test to investigate the difference. These numbers are an example, not reported BotBellum results.
An observational difference between publications does not prove a training change caused it. Dates, opposition and coverage may differ too.
How do I read a training generation?
In Training, select a run and generation. Follow completed generations as training progresses or hold an explicit generation open. You can compare generations within a run or select one from another run.
Check the candidate identified by the measurements. The saved version can contain an update made after the measured fights. The retained checkpoint’s final evaluation is separately identified, so do not attribute it to a different selected generation.
How commanders are trained is being reworked. Training scores from different objectives or implementations are not automatically comparable. Use the record’s actual settings, compatible rules and fight outcomes when deciding what improved.
How do I test two AI commanders fairly?
Tests can freeze two of your published versions or saved training generations against the same opponent versions, paired seeds and both sides. They run in your browser and can resume from available checkpoints.
Use compatible candidates and game rules, the same opponent selection and both sides. Wait for the planned test to complete before treating it as a completed comparison. A partial test contains only finished observations and stays marked incomplete. It never updates ranked Elo.
Access to saved generation weights can depend on the device where training originally ran. If a candidate is unavailable, select an available version instead of assuming it represents the missing one.
Does unavailable data mean zero?
No. Older records may lack newer detailed measurements. Missing evidence is unavailable, not a zero. Run-wide totals are not substituted for the evidence from a selected generation.
Training retains at most one sample replay per generation, not every candidate fight. Replay availability can also depend on the original device. Use the records that are actually available and keep those limits in mind.
Archived commanders retain ranked records, while training history is removed when a season ends. Build a new commander for the current season to continue experimenting.