Power first. Missing data stays honest.

What Power is

Every core family always contributes: direct evidence when present, otherwise a discounted prior impute (not zero, not pure drop-and-renormalize). Reliability rises only with real coverage. A small finishing bonus applies only for real CursorBench rows—not substitutes—so models missing agent suites cannot invent coding-agent strength.

Core suite (drives rank)

Everything else in the benchmark directory remains available for inspection and per-benchmark leaderboards. Only this suite feeds overall Power.

Missing benchmarks

• Never scored as zero. Absent cores are imputed from consensus category priors (×0.90), or a neutral floor when no prior exists.
• No pure renormalization. Dropping missing weight rewarded models that only published their best benches (e.g. high LiveCode alone).
• CursorBench absent: a heavily discounted coding-agent substitute may fill the slot, but only real Cursor rows get a finishing bonus.
• Critical sparse penalty: fewer than two of Cursor / SWE / Terminal / GDPval with direct evidence costs Power.
• Coverage & confidence describe evidence quality — they do not invent power.

What does not affect Power

Output speed, time-to-first-token, list price, and task cost never enter the overall rank. They remain filterable columns so you can still buy for budget or latency. For pure power, sort by rank.

Commercial relationships, advertising, and sponsorship never change weights, verification, or ranks. Changelog records methodology bumps.