A self-directing system needs to decide when it is done. Quality rises with diminishing returns; the stopping rule sets where you cut. Stop too early and it is unfinished; never stop and it runs away. This is freeze ≠ finish drawn as a curve.
Quality follows q(t)=1−e^(−t/τ) — real diminishing returns. The MARGINAL rule stops exactly where the slope falls to ε: q′(t)=(1/τ)e^(−t/τ)=ε ⇒ t*=τ·ln(1/(τε)). The readout computes t* and confirms q′(t*)=ε to 3 decimals — a solved stopping condition, not a guess. Every rule reports the gap it leaves (1−q at the cut), so the cost of stopping is always on the table.
The quality curve is a chosen toy (single exponential); real tasks have knees, plateaus and cliffs. 'Done' is wherever you set ε or the budget — the instrument shows the trade, it does not know the true value of finishing your specific task.
This is the FRZ primitive itself: MARGINAL and SELF-DECLARE are freezes with a stated gap (the shaded 1−q), never 'finished'. NEVER is the runaway that can't stop. Freeze ≠ finish means the honest end is a cut with its remainder named — exactly what the shaded gap shows.