Dealer gamma exposure in S&P 500 options is widely claimed to predict realised volatility, with effects concentrated in stressed environments. I test that claim out of sample and find the opposite: against a properly specified Corsi HAR baseline with a corporate credit control, dealer gamma carries forecasting information about overnight gap magnitude only in low-VIX regimes. On stressed days the baseline already absorbs the signal; on calm days it is actively miscalibrated, and dealer gamma restores forecasting power. Pooling the two regimes washes out the effect, which is why prior work that did not split by regime may have missed it. Two orthogonal volatility predictors (VIX3M/VIX and CBOE SKEW) carry weaker calm-regime signals of their own, but a joint specification analysis locates the surviving forecasting power in the dealer-gamma-by-calm-regime interaction term. The result is robust to stress threshold, training cutoff, tenor, sign convention, and a Holm correction across one hundred Clark and West tests. The result is robust to stress threshold, training cutoff, tenor, sign convention, and a Holm correction across one hundred Clark and West tests. At the deep upper tail (τ = 0.99), relevant for Basel FRTB Expected Shortfall calculations which integrate over the upper 2.5% of the loss distribution, the augmented model reduces calm-regime quantile loss by roughly fifteen percent. The economic implication is the inverse of the practitioner narrative: dealer gamma is most useful precisely where conventional volatility indicators are weakest.