§7.4 · The Recovery Frontier
We shipped this policy believing it was intelligent. Then we measured it.
Net value as a function of churnAversion, the knob that says how much we trust our own churn term. Every point, and all three reference lines, is a measured mean over 20,000 episodes × 5 seeds.
- Shipped point
- churnAversion 1 · ₹1,24,54,217
- Best on this grid — vs shipped
- churnAversion 4 · +₹1,43,943
- Shipped point to Oracle
- ₹3,67,137
The y-axis does not start at zero — it is clipped to the measured range so the differences between arms are visible. Every level on it is labelled.
How to read it
The Oracle line is the ceiling: an agent that reads planted ground truth, priced for the same fatigue and churn as everyone else. It does not read the swept knob, so it is flat. The distance from the curve up to it is the entire prize available to any policy in this world.
The shipped operating point is not on the maximum of the curve we measured, and sits ₹1,43,943 below it. The best point on this grid is still ₹2,23,194 short of the Oracle. We did not optimise this. We measured it, found it was wrong, and are showing you the distance.
Selection discipline. This grid runs on seeds 1, 2, 3, 4, 5. Reading the argmax off it and shipping that would be selecting on the data that selected it — the exact overfitting §7.4 refuses. A candidate only ships after it survives seeds that were never used to choose it.
Second knob: escalation value gate
Spread across the whole gate sweep: ₹67,315 — against ₹18,46,263 across the churn-aversion sweep. The gate was the smaller problem. The shipped gate is the red dot.