Large language models produce outputs that mix valid knowledge, plausible inference, and hallucination with no native mechanism to distinguish them. Current approaches address this piecemeal: confidence calibration handles uncertainty estimation, retrieval-augmented generation handles factual grounding, and curriculum learning handles training order - but no unified framework connects these interventions across the full LLM stack. We present Darshana, a six-component interpretation framework derived from classical Indian philosophical schools (Darshanas), mapped to specific engineering layers from training architecture through output synthesis. Each component generates testable engineering primitives: epistemic self-classification from Yoga's five cognitive modes (Vritti), selective tool routing from Nyaya's four knowledge sources (Pramana), intent-aware query rewriting from Mimamsa's six interpretation principles (Linga), structured knowledge organization from Vaisheshika's seven-category ontology (Padartha), and multi-source synthesis from Vedanta's unification criteria. In controlled experiments across 4,400+ generations, all five orchestration-layer components outperformed equal-sophistication generic controls (71–93% pairwise win rates, Sonnet judge; 60–67% on cross-judge validation by GPT-4o, confirming the advantage with estimated same-model bias of 15–20pp). At the training layer, Yoga's progressive mastery principle, implemented as stage-gated supervised fine-tuning, produced 60–62% win rates against reverse-ordered curricula across three model architectures (Qwen, LLaMA, Phi), all statistically significant. Components placed at incorrect layers failed dramatically (0% win rate for Mimamsa at runtime vs. 82% as query rewriter), confirming that the framework's value lies in principled layer assignment, not in any individual technique. We report both successes and systematic failures - including LoRA equivalence to prompting, multi-agent degradation, and DPO failure - as evidence that the framework generates falsifiable predictions. To our knowledge, this is the first unified framework mapping Indian epistemology to the full LLM engineering stack with controlled experimental validation. Code and data: https://github.com/rishi-manglesh/darshana_llm and https://github.com/rishi-manglesh/vedic_llm