ChronoRAG-G: Slot-Grounded Temporal Evidence Resolution With Ordered Obligations Over a Graph-Structured Context Corpus

Temporal questions fail when retrieval loses one entity, metric, or period even though the corpus contains every required fact. Pooled queries favor dominant passages, while early conflict resolution can collapse revisions, guidance, or historical values that remain valid for different periods. ChronoRAG-G converts a question into ordered evidence obligations, where each obligation names one required answer fact, and preserves a stable slot identity for that fact through retrieval, allocation, and grounding. Its Graph Temporal Context Corpus (GTCC) stores linked temporal records, stable identities, provenance, and unresolved alternatives. The graph structure organizes relational and temporal context, while the frozen path expands underfilled obligations through additional dense retrieval rather than unrestricted graph traversal. Temporal and semantic compatibility checks, slot-preserving allocation, structured generation, and a single bounded correction step then produce the answer. On 1,005 questions over 480 updated earnings-call transcripts, ChronoRAG-G reached 80.70% audited answer accuracy (95% Wilson confidence interval (CI): 78.14%–83.02%), 95.92% strict trace-derived slot coverage (94.40%–97.04%), 98.66% micro required-slot coverage (98.25%–98.98%), and 97.31% hybrid evidence resolution (96.12%–98.15%). Strict coverage requires every required fact to resolve per question; micro coverage counts resolved facts across all required facts; hybrid resolution counts an accepted answer or a complete evidence trace. Complete traces still accompanied 167 rejected answers, locating the main remaining failure in answer construction. ChronoRAG-G treats temporal inconsistency as evidence to organize, not noise to erase before the question defines which state matters.