lessons-bench

Overview ¶

Command lessons-bench runs the active experiment: does the seamark lessons hook change what a headless agent writes? Two arms differing in exactly one bit (the hook wired or absent), a fresh synthetic or pinned-public worktree per trial, a frozen task, and a deterministic judge. See internal/bench for the design and its guardrails.

Paid runs spend len(selected arms) x -trials agent sessions. The "all" instance selector is restricted to no-agent preflight/dry-run modes. Run at release cadence, never in per-PR CI.