V33: Contrastive Self-Prediction

Status: COMPLETE. NEGATIVE.

Hypothesis: Predicting $\Delta_{\text{actual}} - \Delta_{\text{alternative}}$ forces counterfactual representation (rung 8). Standard prediction can be satisfied reactively; contrastive prediction requires representing "what would happen if."

Result: Contrastive loss destabilizes gradient learning. Mean $\intinfo$ = 0.054 ± 0.015 (late phase), significantly below V27 baseline (0.091). 0% HIGH, 30% MOD, 70% LOW across 10 seeds. Prediction MSE increases 1.5–18.7× over evolution in most seeds — the contrastive signal amplifies after drought cycles, decoupling the gradient from the viability signal. All three pre-registered predictions falsified.