Audit the scaffold, not the checkpoint: frozen weights still let RSI expand

Bobadilla-Suarez, Suh, and Fortin ask the governance question most overnight RSI dashboards get wrong: is a frozen checkpoint evidence that recursive self-improvement has stopped? Their stationarity dichotomy says no. Iterative self-modification hits strict diminishing returns whenever the agent’s reachable set of edits stays fixed, and can escape only if that set expands. Rewriting scaffolding — tools, verifiers, decomposition — expands what an agent reaches without touching a weight, so frozen weights buy an eventual ceiling but not stationarity along the way. Audit the scaffold, not the checkpoint.

They read refinement as gradient boosting on residual error (a patch / git diff), then measure where deployed loops break the analogy: patches compose instead of standing beside each other to be voted on, and worker failures overlap. On SWE-bench Lite refinement, per-round improvement decays toward zero; across 401 production sessions churn decays geometrically — a shape shared with a pre-AI human baseline. Sideways, best-of-$k$ orchestration realizes the best worker’s ceiling exactly (width buys rate, not budget). The vote that would beat that ceiling needs diversity real same-family workers lack: on 30 workers, failure overlap sits at its maximum and a majority fails 23/55 (42%) of tasks. Re-consulting a fixed pool has a horizon computable in advance unless the pool can already fit the work.

Why it matters for code review

Always-on factories that advertise “weights frozen, therefore safe RSI” are checking the wrong artifact. Review harness-evolution and multi-agent orchestration PRs for whether the night expanded tools, verifiers, or decomposition — and for retention / diversity evidence before trusting a vote. Continuous verified momentum compounds when FOMO for more unattended rounds funds scaffold audits and measurable saturation curves rather than another burst that confuses a frozen checkpoint with a stationary factory.