A corrector lifts an agent's output above its true competence. The system looks reliable (σ_corr) while raw skill (σ_raw) quietly decays — so trust calibrated on observed output is wrong. The masking index M* = σ_corr/σ_raw measures the gap.
per-node M* from analyze_pipeline() — bars sorted by severity
Computed by mso-core.js (pinned to minimal-oversight==0.1.2). At σ_skill=0.55, c=0.70 this reproduces the paper's Experiment 1 value M*=1.83 exactly. M* > 1 means the corrector is doing the work; the agent could be failing silently underneath.