Seven upgrades, applied.

Every fact in the graph carries metadata: a confidence score (how strong is the citation), a valid_from, an optional valid_until, and a last_verified date. Confidence isn't decorative — the rule engine reads it.

High-confidence facts (≥0.95) are load-bearing: the tax simulator, the rules engine, and the briefing generator all use them directly. Medium-confidence facts (0.4–0.94) get a hover-tooltip showing the source so the user can judge. Low-confidence facts (<0.4) get a coral “refresh requested” flag — the system pauses before relying on them.

Estate documents revise. Wills get rewritten when a child is born or a beneficiary dies. Trust deeds get amended when tax laws change. Powers of attorney expire and renew. The naive approach is to overwrite — but that loses history, and history is exactly what auditors and successor attorneys need.

With a supersedes edge in the alphabet, every revision creates a new node and links it to its predecessor. Old versions stay queryable; the “current” version is whichever has no outgoing supersedes edge.

Not all information is the same age, the same confidence, or the same lifespan. A raw passport PDF is one thing; the parsed fact “Dan is a Maltese citizen as of 2025” is another; the rule “use Maltese passport for BVI KYC, never Chinese” is yet another. They live at different levels of the stack.

The architecture already implements this implicitly — raw documents in raw/, sources in wiki/sources/, entities in wiki/entities/, principles in principles.md. v2 gives the tiers names and a promotion model: facts move up the stack as evidence confirms them.

Past ~200 entities, a single index file becomes too big for any LLM to scan in one pass. You need real retrieval. Three streams complement each other: BM25 nails exact-term recall (find every page that says “Hong Kong”), vector embeddings nail semantic recall (find pages about offshore jurisdictions even without the word), graph traversal nails relational recall (find every vehicle whose located_in is HK).

Reciprocal rank fusion merges the three lists. A result that ranks high in any one stream rises; a result that ranks high in all three rises further.

Karpathy's original wiki said: “good answers can be filed back as new pages.” v2 takes this further. Every completed work product — a scenario simulation, a cross-professional reconciliation, an override decision — is itself a piece of knowledge worth keeping. The crystallizer auto-extracts the conclusion, the entities involved, the source spans, the rationale, and files it as a typed Analysis node.

This closes the loop: the system's own work product becomes input to future reasoning. The PFIC reconciliation gets cited the next time a similar conflict arises. The S1 simulation outcome informs scenario-3 planning. Reasoning compounds instead of evaporating.

Our rule engine fires on writes. v2 expands that to five hook points:

On session start, load context relevant to recent activity — the principal walks into a briefing of what's open, what changed, what needs decision. On document ingest, the write-time rules we already have. On query, log the question + answer to the decision log; surface any conflict it touched. On session end, crystallize the session into typed Analysis nodes. On schedule, periodic background work: freshness check, lint pass, low-confidence refresh prompts.

The compartmentalisation we sketched on the Scopes page goes deeper than just “which entities does this professional see.” It applies per property per audience. A passport number is full-fidelity for USLP (they need it for KYC); the same fact is “Malta nationality, number redacted” for BY&S (they only need the jurisdiction for tax); for a public summary, it's masked to ****1400.

The graph stores the source of truth once. The filter at render time decides what each view exposes. The principal sees the full provenance trail, including which views received what, and when.