📈 #16 Per-year IC + structural break GROUNDED

Does the feature's edge stay the same sign and strength across time, or does it break/flip — and if it breaks, exactly when? Per-block IC + a sup-F mean-shift break with a block-bootstrap critical value (asymptotic CVs are invalid on short autocorrelated IC). All 4 gates ground, robust across 4 seeds.

4 / 4

gates · 4 seeds

1.0

sign-flip detect rate

0 blk

break-date MAE

18 / 20

terminal (17 G · 1 R)

The four gates

Gate (§7 row 16)	Result (4 seeds)
Admit stability	same-sign across 8 blocks · ICIR ~5.4 · min-block\\|IC\\| 0.72× median (≥.5)	PASS
Power (sign-flip) & break-date	detect 1.0 · break-date MAE 0.0 blocks (located exactly)	PASS
Spurious-break FPR	0.0–0.033 (block-bootstrap CV; asymptotic CV invalid here)	PASS
Dead-feature guard	0.0 — a dead feature can't hold same-sign → never vacuously "stable"	PASS

Two verify-before-report fixes. (1) The spurious-break FPR is controlled by a conservative 98th-percentile block-bootstrap critical value — the #10 conservative-calibration pattern; the asymptotic sup-F critical value over-rejects on the short autocorrelated per-block IC. (2) The dead-feature guard checks the full stability admit (same-sign & ICIR≥.5 & min≥.5×median), not ICIR alone — a dead feature (IC~0) cannot hold the same sign across 8 blocks, so it is flagged as precondition-unmet rather than being vacuously called "stable." This gate sits behind the realness instrument (the §7 precondition).