Case Wiki — Compounding Knowledge Base | AI Harness

Architecture

Three-Layer Knowledge Architecture

Autonomous Living Dossier — raw sources remain immutable ground truth; the institutional knowledge base is synthesized, linked, and audited continuously by sovereign AI.

Layer 1

Raw Sources

PDFs, DOCX, XLSX and statutory records in the case vault. Immutable — the sovereign pipeline extracts evidence but never modifies underlying records. This constitutes the verifiable ground truth.

Vault Ground Truth • Read-Only Verified Evidence

Layer 2

Autonomous Dossier

Agentic-maintained knowledge records. Source abstracts, entity cross-references, statutory concept indexes, synthesised Q&A insights, and live catalog indices. Automatically refreshed upon ingestion.

dossier/ • entities/ • insights/ • Agent Maintained

Layer 3

Schema & Governance

Institutional schema protocols and agent operational directives governing document taxonomy, milestone verification, statutory timeline tracking, and dynamic entity graph boundaries.

Governance Protocols • Rule-Bound Compliance

Living Dossier Knowledge Hierarchy

dossier/ ├── INDEX.md # Autonomous master catalog: structured cross-references and verified status ├── audit_log.md # Append-only cryptographic ledger: ingestion timestamps and reasoning audits ├── abstracts/ # Auto-extracted multi-page evidential briefing records │ ├── Resolution_Plan_Summary.md │ └── Statutory_Filing_Brief.md ├── insights/ # Curated legal-forensic synthesis (cross-referenced across evidential exhibits) │ ├── statutory_eligibility_analysis.md │ └── voting_rights_recalculation.md ├── precedents/ # Case jurisprudence and ratio decidendi alignment entries └── HEALTH_REPORT.md # Neural audit report: contradiction alerts, evidential gap flags

Operations

Three Core Dossier Operations

IngestAuto-Brief

Source Ingest → Evidential Brief

After an exhibit is indexed, the neural pipeline generates a comprehensive structured brief covering verified entities, legal instruments, financial quantum, and risk exceptions. This is immediately mapped into the master index.

Auto-creates dossier record Updates master index

QuerySynthesis Filing

Neural Answer → Evidential Insight

Every agent deliberation output supports instant archival to the Living Dossier. When an analysis cites multiple primary filings, the system flags it for inclusion in the case record with full evidential provenance and audit metadata.

Instant archival Dual-indexed in real time

AuditContinuous Verification

Case Dossier Health Audit

A periodic background neural verification pass across all dossier records and entity registries. Detects conflicts and surfaces structured cards in the Action Inbox — ensuring the dossier actively self-audits against inconsistencies.

Contradiction flags Inbox cards

Lint Operation — 8 Audit Checks

Contradictions

Mismatched amounts between claims_registry.md and avoidance_ledger.md entries across different filings.

Stale Claims

case_facts.md fields that are blank [UNKNOWN] but may be derivable from already-ingested documents via RAG.

Orphan Wiki Pages

Wiki pages with no inbound references from other pages or from the INDEX.md catalog — unlinked knowledge islands.

Missing Concept Pages

Important concepts mentioned in source summaries but lacking their own dedicated wiki page.

Orphan Chunks

Document vectors with zero retrieval count — knowledge that has never been surfaced in any RAG query and may be improperly tagged.

Statutory Milestone Gaps

CIRP T0→T330 milestones in timeline.md that lack a corresponding ingested filing (e.g., Day 30 CoC notice has no meeting notice document).

Missing Cross-References

References from one wiki page pointing to another that has no corresponding page — broken internal links in the knowledge graph.

Suggested Investigations

Data gaps that could be filled with additional document ingestion — new sources suggested by the LLM based on patterns in existing wiki pages.