Essential Data Carried Across Steps:
-
High-confidence entities, relationships, and classifications (threshold defined in Step 9), ensuring continuity of validated knowledge.
-
Unresolved contradictions, preserved to support future disambiguation and learning.
-
Temporal anchors that help in temporal disambiguation, co-reference, and anchoring.
-
Mutation-specific focus parameters (e.g., 2184insA domain flags), which help tune domain-specific inference pathways.
-
Keymaker’s internal selection heuristics, which evolve adaptively through Q&A iterations and reflect emergent prioritization strategies.
Discarded Data:
-
Intermediate parsing artifacts are discarded to reduce storage overhead and eliminate noise that no longer contributes to downstream decision-making.
-
Unvalidated transient hypotheses unless explicitly annotated by Vault, ensuring only intentional ambiguity persists.
-
Low-relevance entity mentions (< 0.2 salience) to maintain semantic sharpness and graph compactness.
-
Retention flags for high-complexity or controversial cases where intermediate parsing might provide valuable retrospection or error analysis capabilities.