Skills
One cheap re-verifier call cuts wrong actions 79% — and the obvious fix, locking the entity, makes things 8.5x worse
"Binding drift" is when an agent picks the right entity at step 1 then silently binds to a different one later; measured over 200 workflows and 580 binding-scored steps across eight model backends, it hits 18% of eligible workflows naturally. The intuitive mitigation — persisting the first binding as an entity lock — amplified wrong actions 3.0x on average and 8.5x on Claude Opus 4.5, because it cements early mistakes. What worked instead was a single cheap second model call that re-reads the original instruction before acting, reducing wrong actions by 79% and landing within 1 point of the oracle upper bound.
↳ Follow the thread