sb-core — the SharpeBench scoring kernel

A pure, deterministic library that turns a set of agent trajectories (per-seed × per-window return series + decision traces) into a luck-robust, risk-adjusted score and leaderboard ranking.

Design invariants — these make a score reproducible for a pinned release and input:

• Pure. No I/O, no system clock, no ambient randomness. Any randomness (the significance bootstrap) takes an explicit seed argument.
• Deterministic. Plain f64 math, fixed reduction order, no parallel float sums. CI pins two committed golden fields on Linux, macOS, and Windows; that check does not cover every input or platform.
• No unsafe.

The headline idea: an agent does not rank on raw return. It ranks only if its edge survives (a) deflation for the number of agents tested ([deflated_sharpe]), (b) reliability across every seed×window ([mod@pass_k]), and (c) decision-process discipline ([process]). Raw return is reported but never the rank key — see [composite].