Ingestion guide (RAG / tooling)
Goal: make the spine available as a stable reference set so your workflow can continually re-ground analysis and reduce drift.
Recommended corpus
/downloads/*qi-grounding*.json(canonical spine)/spine/primitives.html+/spine/composites.html(human readable)/domains/*.html(domain packs)/charters/*.json(Layer-3 examples)
Chunking strategy
- Chunk by identifier: one chunk per
P#and perC#. - Keep “Direct / Mirror / Shadow” together in the same chunk.
- Store aliases as synonyms to improve retrieval.
- For composites, include dependency pointers (
Plist) in the chunk metadata.
Retrieval strategy
- First-pass retrieval: query with the artifact text + “Ground in P/C.”
- Second-pass retrieval: query specifically for
P10,C10,C11,C12if outputs feel overly confident or linear. - Prefer returning 6–12 chunks: relevant P’s + relevant C’s + 1 domain-pack chunk.
Output contract
Always output:
- P map: P1..P10 that apply (with 1–3 bullet notes each)
- C map: C1..C12 that apply (with 1–3 bullet notes each)
- Missing grounding effects (at least 5)
- Comparator tiering (explicit)
- Cascade scan (if any nonlinear risk)
- Assumptions and unknowns
Note: This workflow is intentionally “anti-bullshit.” If the model can’t name boundaries and comparators, it should not be trusted to recommend controls.