Observe first, propose without dispatch

In shadow mode a human operates a physical booth while an AI receives the available normalized observations. The AI writes proposed commands to a separate proposal log. Enforce read-only status in the engine dispatcher, not merely through an actor configuration flag or disabled UI button.

Record proposal creation time, intended execution frame, the observation revision used, confidence, and model/configuration identity. Keep proposal events separate from applied events so they cannot affect the live state projection.

Afterward, fork a virtual scenario from a recorded checkpoint and inject selected proposals through normal validation. Compare it with a reconstruction of the observed human performance.

This is a counterfactual experiment, not an exact alternate history: hardware logs may omit audio, FX, control movements, beatgrid corrections, or precise timing. Require the same audio assets and enough state to initialize a meaningful run; disclose unknown controls and timing uncertainty. A human's later actions may also depend on choices that the AI branch changed.

The first shadow milestone requires passive adapter evidence, a tested no-dispatch boundary, reconstructable checkpoints, and a report of missing observations.