Standard Retrieval-Augmented Generation (RAG) systems, including state-of-the-art multi-document synthesizers like Google’s NotebookLM,fail at discourse archaeology due to two fundamental problems:(1) Source biaslarger or semantically dominant texts disproportionatelyoccupy the context window, silencing minority voices despiteequal relevance; (2) Temporal conflationsystems cannot reconstructhow discourse evolves across document versions, treating temporalsnapshots as isolated units and destroying rhetorical trajectories.Together, these constitute not merely technical limitations but anepistemological failure: standard RAG conflates retrieval relevancewith discourse fidelity, mistaking semantic dominance for epistemicpriority.We present the Hermeneutic Reconstruction Engine (HRE), aconstraint-based architecture enforcing source parity through threelayers: (1) VIP Protection guarantees minimum representation(kmin = 3 chunks per document) at retrieval, preventing eliminationof small or semantically distant texts; (2) Essence Parity caps maximumrepresentation (kmax = 12 chunks) at synthesis, preventing dominanceby large texts while preserving quality through best-chunk selection;(3) Hermeneutic Enforcer validates claims via two-axis taxonomy(Type: Paraphrase/Meta/Inference × Validity: Supported/Exaggeration/Contradiction/Hallucination), enabling nuanced validationbeyond binary fact-checking.We systematically compare HRE against NotebookLM (standardand Deep Research modes) across four discourse archaeology tasks:political discourse analysis (Harvard bias reports), AI rhetorical deconstruction(X-Grok argumentation), creative hypothesis validation(Dante/IFS connection), and prompting robustness (Burckhardt interpretation).HRE achieves consistent advantages on structural fairnesstasks: +17% on political discourse (detecting hidden hierarchies despiteexplicit denial), +14% on rhetorical deconstruction (identifyingperformance vs. practice of skepticism), +10% on hermeneutic innovation(validating creative cross-temporal synthesis). On standard interpretationtasks, HRE and NotebookLM perform equivalently, demonstratingcomplementarity rather than competition.Recent large-scale evaluations corroborate this limitation systematically:RAGBench Labbé et al., 2025, evaluating 26 systems across4,484 examples, documents a 25% factuality drop in multi-documentsynthesis attributable to source bias—precisely the architectural gapHRE addresses through VIP Protection and Essence Parity.HRE demonstrates that discourse archaeologythe reconstructionof temporal, rhetorical, and linguistic evolution across heterogeneoustextsrequires architectural fairness constraints, not merely better embeddingsor prompts. Our work shows that proportional representationin multi-document synthesis is achievable through structural guaranteesoperating independently of model quality.
Grigori Pantijelew (Wed,) studied this question.