The proliferation of deepfake technology has precipitated a critical trust crisis in digital multimedia forensics, calling into question the reliability of existing detection systems. Current models predominantly rely on softmax-normalized probabilities, which exhibit heightened vulnerability to adversarial perturbations and out-of-distribution (OOD) samples. To address this deficiency and provide courts with quantifiably reliable forensic evidence, this paper proposes a Dual-Threshold Reliability Assessment framework (DTRA) grounded in class-conditional feature space analysis. The DTRA framework quantifies epistemic uncertainty through Mahalanobis distance-based inconsistency scoring computed from deep feature representations. Departing from conventional single-threshold paradigms, we independently calibrate optimal decision intervals for authentic and forged classes on a held-out calibration set. The interval optimization is formulated as a risk-adjusted utility maximization problem that trades off empirical precision against effective sample coverage. Specifically, an interval search algorithm identifies the most reliable subrange of inconsistency scores for each class via an odds ratio-weighted utility function, eschewing the restrictive assumption of a zero lower bound. DTRA serves as a conservative safeguard: when sample evidence is ambiguous, it abstains rather than forces a prediction. While this conservatism reduces coverage, the predictions it retains are significantly more reliable, thereby reducing the risk of high-confidence misjudgments.