Different coarse-grained descriptions of the same deterministic system can yield different measured changes. Using a component-count fine observer and an occupied-block coarse observer, we establish three empirical regularities in Conway’s Game of Life. Early component net change predicts the full-run normalized narrative gap between observers with partial correlation r = −0.848 (95% bootstrap CI [−0.862, −0.830], N = 1000), sharpening after density control. The predictive scale that best forecasts a future observable depends on which observable: time-averaged live-cell count peaks at an intermediate scale (B = 2) while occupied-block count peaks at its own native scale (B = 8), with non-overlapping bootstrap confidence intervals. A linear law linking embedded isolated cells to future component-level decline is stable across six size-by-density conditions (mean slope ≈ −1.52, CV ≈ 0.09). Two stronger mechanism claims—global mediation through component net change, and residual coupling beyond death and local neighbourhood dynamics—do not survive their pre-specified tests. All results are established computationally within GoL; generalisation beyond this system remains an open empirical question.