Experiment complete. Here's the summary:

## Risultante

**Prime gaps are weakly blue noise** — the power spectral density confirms and constrains the 1/k anti-correlation law (ACF_1K_LAW).

**What's new:**
- **Low-frequency suppression is real** (z=-3.0 vs shuffle). Prime gaps have 28% less power at very low frequencies. Long-range anti-correlation destroys slow modes.
- **Spectral slope = +0.11** — slightly blue, a new spectral class. Not white, not pink, not 1/f.
- **Two-component structure**: a Poisson floor (white) + anti-correlated dip (logarithmic). The 1/k model captures the shape (r=0.73) but underpredicts amplitude by 2x — meaning it accounts for ~half the variance.
- **No Z/6Z spectral peaks** at f=1/6 or f=1/3. The modular arithmetic is invisible in frequency space.

**What constrains:**
- ACF_1K_LAW R² is 0.65 (weighted fit), not 0.998 (log-log). The previous method compressed deviations. The 1/k law is a good approximation, not an exact law.

**Consecutio**: the crossover frequency f_c ~ 0.05 (where anti-correlation dominates) should shift to lower f at larger primes — directly testable, and connects to POISSON_CONVERGENCE.

**Files**: `tools/exp_psd_prime_gaps.py` (reusable), `tools/data/reports/agent_20260409_0330.md`, `tools/data/psd_prime_gaps_results.json`, seme updated with PSD_BLUE_NOISE tension.