Background: Digital assessment of neurodivergent adults is hampered by a fundamental tension: self-report, informant-report, and performance measures often disagree, yet existing platforms typically collapse these discrepancies into a single score or discard them as noise. In autism and related neurodivergences, such disagreement can reflect clinically meaningful phenomena, including masking, context-dependent trait expression, informant-context mismatch, and limited insight, rather than invalid data. Current systems rarely make their integration logic auditable or neuroaffirmative. Objective: This methods paper presents the SPECTRA Evidence Stack, a formally specified three-layer framework (Response Quality, Multi-source Convergence, and Confidence Synthesis) designed to integrate heterogeneous assessment signals transparently. The framework produces structured outputs, including quality flags, convergence profiles, reason codes, and next-step recommendations, without claiming diagnostic validity. Methods: We describe the architecture, formal definitions, and rule logic of each layer. Response Quality flags identify data conditions requiring caution without pathologizing the respondent. Multi-source Convergence assigns interpretable discrepancy profiles across self-report, informant-report, and performance data. Confidence Synthesis combines these artifacts into an evidence-confidence band, mandatory reason codes, and recommended actions. We report a formative computational evaluation using a four-condition ablation design on 250 synthetic records, plus three clinical demonstration vignettes. Results: The full-stack condition produced higher and less variable confidence scores (mean 0.64, SD 0.09, 95% CI 0.63-0.65) than the no-convergence (mean 0.50, SD 0.12, 95% CI 0.48-0.51; paired Cohen dz=2.57) and baseline conditions (mean 0.50, SD 0.31, 95% CI 0.46-0.54). Discordance-sensitive cases showed the largest separation between full-stack and ablated conditions, demonstrating that the framework treats discrepancy as structured signal. Demonstration vignettes illustrated masking-compatible divergence, informant-context mismatch, and sparse-evidence scenarios. Conclusions: The SPECTRA Evidence Stack provides a replicable, auditable, and neuroaffirmative architecture for integrating multi-source neurodivergence assessment data. It makes uncertainty visible, preserves discrepancy as interpretable evidence, and supports shared clinical decision-making. This formative methods paper does not establish diagnostic accuracy; a subsequent clinical validation study against independent adjudication is planned.