Citations, parsers, and orchestration
A day of deepening citation extraction, simplifying agent architecture, and separating deterministic parsing from AI judgment.
The work continued on the same axis as the previous day: legal documents, citation extraction, and agent architecture. The difference was that today the conversation became more concrete.
Preserving structure
First, I adjusted the project creation path so the original repository state stayed preserved. Instead of changing something without a reference point, the idea was to clone or mirror the right structure and then compare what needed to change.
I also reviewed the backend folder and turned the analysis into a clearer prompt for recreating the structure. That is useful when several agents are trying to reach the same technical design.
Deterministic before agentic
The most important theme was the citation question. At high document volume, relying only on an extraction agent does not make sense. The stronger path starts with deterministic patterns, regex, parsing, normalization, and tests with real examples. Then AI can help with edge cases, explanations, and exception checks.
I also removed optional complexity from the orchestration. The architecture needed fewer conceptual layers and more clarity about where the orchestrator lived, what it called, and which step owned each responsibility.
The day ended with a practical decision: use AI where it expands the system, not where it replaces a parser that should be reliable.