4 · When the JSON ontology is no longer enough — graduating to a KG

The signal that JSON has hit its ceiling

JSON-first is right until it isn't. Graduate from a JSON ontology to a graph-backed one when at least two of these hold:

1. Multi-hop questions are now the product. 'Show every open claim whose repair workshop also appears in a fraud ring connected to this policyholder' — customer → policy → claim → workshop → fraud pattern. JSON can store those links; it can't traverse them.
2. Cross-system entity resolution is costing real hours. The same policyholder is cust_4821 in claims, P-99213 in billing, and a name string in the fraud tool. Reconciling them by hand, repeatedly, is the tax a shared identity (IRIs) removes.
3. Regulators want relationship evidence, not flat logs. 'Prove this claim was never linked to a flagged workshop' is a path query, awkward over event logs.
4. The rule set outgrows maintainable JSON branching and needs semantic reuse / inference (e.g. automatic hagelschaden ⊑ elementarschaden rollups everywhere).

If none of these bite, stay JSON-first — a graph you don't query multi-hop is pure operational overhead.

What the graph adds that JSON genuinely cannot

Be precise about what you're buying, so the upgrade is a decision, not fashion:

• Traversal as a first-class operation. Variable-length path queries (customer → … → fraudRing) are native; in JSON they're N+1 lookups you hand-stitch.
• Inference. Declare hagelschaden rdfs:subClassOf elementarschaden once and every query, report and constraint inherits it. In JSON you maintain the rollup in every place by hand.
• Global identity. An IRI like …/policy/AB-483920 is the same node no matter which system wrote it, so resolution happens by design, not by nightly reconciliation.
• Constraints that travel with the data (SHACL). The claim shape ships with the graph instead of living in one validating service.

Crucially, phase 2 keeps the JSON ontology — the lexical normalisation and STT hotwords still do the Swiss-German accuracy work. The graph sits downstream of extraction, on the already-canonicalised entities. You're adding a reasoning layer, not replacing the accuracy layer.