How to read our numbers

Par in a pod is 25%, not 50%.

Four seats means an average deck wins a quarter of its games — so 55% in a four-player pod is more than double par, a strong result. Head-to-head is different: two seats, par 50%, and a 55–45 head-to-head is close to a coin flip. The seat count is always on the result; read the number against the right par.

Every seat plays the same policy.

By default the model gives every seat the same interaction layer, the same mulligan rules, the same targeting doctrine — equal policy. Your table is not like that: you pilot your own brew better than a stock precon pilots itself. When our number and your table disagree, equal policy is often the reason, and the results page says so. Skill is expressed as named policies (per-seat, under Advanced), never as a slider or a hidden number.

Lines are routes, not endings.

A game “ends” by combat, poison, combo, mill, commander damage, or attrition — the same vocabulary everywhere on this site. A line like a granted miracle is a route: the free-cast Eldrazi still wins by combat, and the trace on the results page attributes the win to the line that made it happen.

Every number carries its provenance.

Engine version (currently rev58), seed, field version, oracle snapshot date, and grammar revision ride on every result. Same inputs, same stamps, same number — anyone can reproduce it. A number without those stamps is not one of ours.

Below 500 scored games we claim no accuracy.

The calibration record is public and scores predictions against real logged games — but not until the corpus is large enough to score honestly. Until then, every model number is model-relative: a statement about the simulation, not a claim about your table.

Logged games validate. They never fit.

No parameter ever moves to shrink an error. Real games rank the misses, humans diagnose them, mechanics get built from card text, the engine version bumps, and new games judge the new engine. We publish the misses either way.