ResearchPack Injection

Transitioning from parametric reasoning to evidence-grounded synthesis by binding sealed ResearchPacks to the model context. S7 is where the model "opens the books."

Prompt Blueprint Architecture
Layer 01: System Instructions
"You are a legal research assistant..."
Layer 02: ResearchPack (Injected)
Full Opinion + Contextual Snippets
~15k - 60k characters
Layer 03: S6 Baseline Draft
Preliminary Closed-Book analysis
Layer 04: IRAC Task Rules
JSON schema + Citation enforcement
Technical Binding

Every model run at S7 is reproducible. ResearchPacks are sealed at build time and bound to the prompt at runtime using deterministic variable injection. This eliminates RAG-delivery variance—the model receives the exact same character strings in every identical test run.

// s7_open_book_synthesis.py
variables = {
"research_packet_text": pack.render_prompt_text(),
"s6_draft": s6_result.parsed["irac"]
}
The Citation Protocol

Grounding is not optional. To pass the subsequent Stage 08 Integrity Gate, the S7 engine enforces four mandatory citation rules. Failure to cite the ResearchPack identifiers results in a baseline failure.

Enforcement Rules
  • Every material claim must cite a Document ID from the pack (e.g., [DOC1]).
  • Literal U.S. Reports citations must be inclusive (Name, Volume, Page).
  • Conflicts between documents must be surfaced and reasoned about.