Post-Foundry
Crosswalk
Subject, Included, Relevant, Excluded — the identity-first retrieval system.
SIRE is deterministic identity metadata embedded in oracle corpus frontmatter. The Crosswalk processes the full oracle library to map coverage, overlaps, conflicts, and routing paths so multi-oracle governance can run deterministically.
Without it
Without a global authority map, multi-oracle retrieval degrades to best-match search. Overlapping sources create contradictions, gaps go undetected, and routing decisions happen by accident instead of by authority.
Position in the platform
- Oracle Foundry (Foundation): Produces oracles with SIRE tags and chunk embeddings
- SIRE Crosswalk (Post-Foundry): Builds global authority graph, certifies routability, and flags gaps/conflicts
- Prompt Compiler (L0): Consumes routing paths for multi-oracle activation and authority-aware prompting
- Claim Ledger (L1-L4): Uses authority map for gate boundary enforcement and attribution context
- Forensics Lab (L5): Uses crosswalk in Phase 2B to diagnose routing failure vs source gap
How SIRE enforces at runtime
Only excluded enforces. Subject, Included, and Relevant inform discovery but never veto. After hybrid retrieval returns candidate chunks, the exclusion gate purges any chunk whose content matches an excluded term.
- Query hits pgvector + tsvector hybrid search, returns candidate chunks ranked by semantic + lexical similarity
- Chunks grouped by sire_subject, subjects ranked by mean relevance score
- For each winning subject, load its sire_excluded array
- Walk candidate pool — disqualify any chunk containing an excluded term (case-insensitive, word-boundary match)
- Only non-excluded chunks passed to LLM synthesis
Relationship discovery model
- Layer 1 — Declared relationships: Matches Excluded-to-Included concepts across all oracles, validates Relevant references exist, and detects Subject overlap that triggers conflict resolution.
- Layer 2 — Discovered relationships: Computes cross-oracle embedding similarity (cosine threshold 0.80) to find semantically overlapping coverage the SIRE tags missed.
- Registry merge: Classifies confirmed (tag + embedding), declared-only, discovered-only, and conflict relationships into one unified authority registry with O(1) concept lookup.
