Sources & methods
What the numbers mean
BAR Fight turns public Beyond All Reason replays into player comparisons, map statistics and position rankings. These numbers describe recorded games. They do not tell the whole story of a player or a team.
Where the data comes from
Our main source is GEX, a community project that extracts measurements and unit events from public replays. We also retain our earlier Supreme Isthmus captures for measurements they recorded. Source-aware comparisons choose the available evidence without counting the same match twice.
The official replay archive is the source of public games. This is not a complete record of every BAR match. Imports can be delayed, captures can be incomplete, and private games are not included.
Games, players and missing data
A game is one match. A player game is one player's appearance in one match: an 8v8 match can contribute 16 player games. Account IDs keep a player's history together when their name changes.
Coverage is the percentage of selected player games with usable measurements. Missing values stay missing; they are not treated as zero. A game can support an economy chart while lacking the unit events needed for an opening chart. Each chart keeps its own denominator, so different coverage percentages can be correct.
Openings and build timings
Our featured map answers use players with a recorded replay rating of OS 40 or higher, including exactly 40. That is a reference group of experienced players, not proof every choice was good. The interactive Statistics tool lets you change it.
The first fighting unit means the first completed mobile fighting unit in the recorded event data. Commanders, builders, buildings and static defences are excluded. Units completed at the same recorded time are shown together.
Build timing uses the first recorded completion of the chosen unit or group for each selected player game. The typical time is the median among player games where it was observed. The middle half runs from the 25th to the 75th percentile. The percentage reaching a build uses all selected player games with usable unit events, including games where it never appeared. Games do not have to last 20 minutes to contribute.
Grouped constructors or economy buildings combine explicitly selected units with a similar purpose across factions. Their costs and abilities are not assumed equal. Receiving a unit and building it are different questions; transfers do not prove payment.
Map positions and game versions
We combine opposite starting spots only when the map layout supports that pairing. Unclear starts remain outside a named position. A starting spot gives context, but a player can change responsibilities during a match.
Map versions stay separate. Featured timing summaries can include several game patches; their labels say which filters were used. In Statistics, choose a game version to narrow the answer. A common timing is a description, not a deadline: pressure, teammates and enemy choices can make a different build appropriate.
How the ranking score works
All maps use the same current completed-unit model. It compares five-minute changes with reference players in the same map version, format, starting layout, game and engine version, settings and time interval. References can come from any recorded date and any faction. The player being scored is excluded and reference accounts receive equal weight.
The model combines four measurements:
| Measurement | Weight | Meaning |
|---|---|---|
| Combat balance | 45% | Normal metal value of finished enemy units destroyed, minus finished units lost to enemies. |
| Metal income | 20% | Recorded income, including reclaim and game-rule income. |
| Energy income | 10% | Recorded energy production, including energy that could not be stored. |
| Fighting units built | 25% | Normal metal value of finished fighting units and defences; not exact spending. |
The weighted result moves toward 50 when there are fewer recorded playing dates. The displayed uncertainty range shows variation when those dates are resampled; it does not measure every possible source of bias.
Ranking eligibility currently requires a latest recorded OS above 30, at least 5 scored games across 3 UTC dates. Established results require at least 10 scored games across 5 dates and at least 60% usable coverage among games long enough to score. Early results use the same calculation but receive no medal. Games under 8 minutes remain context only; scoring omits the last 3 minutes of each game and uses intervals through minute 30.
These weights are design choices, not measured percentages of victory. Win rate does not add score points. Final-attacker combat credit misses assists; income can be unhelpful, and completed units may never reach a useful fight. The score is not official OpenSkill, win probability or a share of team victory.
What we can and cannot infer
We can measure recorded production, timing, transfers and combat events. We cannot reliably infer intention or what a player saw. A build followed by an enemy response does not prove a deliberate counter. Winning after a purchase does not prove the purchase caused the win. Team pressure and events elsewhere on the map matter.
Public map summaries are refreshed offline and show their data dates. Rankings are a separate daily publication and can have a different date. Open the interactive tools to change the filters and inspect the supporting measurements.